Merge #4473

4473: Bring back changes from v1.7.0 to main r=curquiza a=curquiza



Co-authored-by: ManyTheFish <many@meilisearch.com>
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
Co-authored-by: Many the fish <many@meilisearch.com>
Co-authored-by: Tamo <tamo@meilisearch.com>
Co-authored-by: meili-bors[bot] <89034592+meili-bors[bot]@users.noreply.github.com>

.cargo/config.toml

@@ -1,2 +1,2 @@
 [alias]
-xtask = "run --package xtask --"
+xtask = "run --release --package xtask --"

.github/workflows/bench-manual.yml

@@ -0,0 +1,30 @@
+name: Bench (manual)
+
+on:
+    workflow_dispatch:
+        inputs:
+            workload:
+                description: 'The path to the workloads to execute (workloads/...)'
+                required: true
+                default: 'workloads/movies.json'
+
+env:
+    WORKLOAD_NAME: ${{ github.event.inputs.workload }}
+
+jobs:
+    benchmarks:
+        name: Run and upload benchmarks
+        runs-on: benchmarks
+        timeout-minutes: 180 # 3h
+        steps:
+            - uses: actions/checkout@v3
+            - uses: actions-rs/toolchain@v1
+              with:
+                profile: minimal
+                toolchain: stable
+                override: true
+
+            - name: Run benchmarks - workload ${WORKLOAD_NAME} - branch ${{ github.ref }} - commit ${{ github.sha }}
+              run: |
+               cargo xtask bench --api-key "${{ secrets.BENCHMARK_API_KEY }}" --dashboard-url "${{ vars.BENCHMARK_DASHBOARD_URL }}" --reason "Manual [Run #${{ github.run_id }}](https://github.com/meilisearch/meilisearch/actions/runs/${{ github.run_id }})" -- ${WORKLOAD_NAME}
+

.github/workflows/bench-pr.yml

@@ -0,0 +1,46 @@
+name: Bench (PR)
+on:
+    issue_comment:
+        types: [created]
+
+permissions:
+    issues: write
+
+env:
+    GH_TOKEN: ${{ secrets.MEILI_BOT_GH_PAT }}
+
+jobs:
+    run-benchmarks-on-comment:
+      if: startsWith(github.event.comment.body, '/bench')
+      name: Run and upload benchmarks
+      runs-on: benchmarks
+      timeout-minutes: 180 # 3h
+      steps:
+        - name: Check for Command
+          id: command
+          uses: xt0rted/slash-command-action@v2
+          with:
+              command: bench
+              reaction-type: "rocket"
+              repo-token: ${{ env.GH_TOKEN }}
+
+        - uses: xt0rted/pull-request-comment-branch@v2
+          id: comment-branch
+          with:
+            repo_token: ${{ env.GH_TOKEN }}
+
+        - uses: actions/checkout@v3
+          if: success()
+          with:
+            fetch-depth: 0 # fetch full history to be able to get main commit sha
+            ref: ${{ steps.comment-branch.outputs.head_ref }}
+
+        - uses: actions-rs/toolchain@v1
+          with:
+            profile: minimal
+            toolchain: stable
+            override: true
+
+        - name: Run benchmarks on PR ${{ github.event.issue.id }}
+          run: |
+            cargo xtask bench --api-key "${{ secrets.BENCHMARK_API_KEY }}" --dashboard-url "${{ vars.BENCHMARK_DASHBOARD_URL }}" --reason "[Comment](${{ github.event.comment.url }}) on [#${{github.event.issue.id}}](${{ github.event.issue.url }})" -- ${{ steps.command.outputs.command-arguments }}

.github/workflows/bench-push-indexing.yml

@@ -0,0 +1,25 @@
+name: Indexing bench (push)
+
+on:
+    push:
+        branches:
+            - main
+
+jobs:
+    benchmarks:
+        name: Run and upload benchmarks
+        runs-on: benchmarks
+        timeout-minutes: 180 # 3h
+        steps:
+          - uses: actions/checkout@v3
+          - uses: actions-rs/toolchain@v1
+            with:
+              profile: minimal
+              toolchain: stable
+              override: true
+
+          # Run benchmarks
+          - name: Run benchmarks - Dataset ${BENCH_NAME} - Branch main - Commit ${{ github.sha }}
+            run: |
+              cargo xtask bench --api-key "${{ secrets.BENCHMARK_API_KEY }}" --dashboard-url "${{ vars.BENCHMARK_DASHBOARD_URL }}" --reason "Push on `main` [Run #${{ github.run_id }}](https://github.com/meilisearch/meilisearch/actions/runs/${{ github.run_id }})" -- workloads/*.json
+

.github/workflows/test-suite.yml

@@ -31,17 +31,10 @@ jobs:
           apt-get update && apt-get install -y curl
           apt-get install build-essential -y
       - name: Setup test with Rust stable
-        if: github.event_name != 'schedule'
         uses: actions-rs/toolchain@v1
         with:
           toolchain: stable
           override: true
-      - name: Setup test with Rust nightly
-        if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
-        uses: actions-rs/toolchain@v1
-        with:
-          toolchain: nightly
-          override: true
       - name: Cache dependencies
         uses: Swatinem/rust-cache@v2.7.1
       - name: Run cargo check without any default features

.gitignore

@@ -9,6 +9,8 @@
 /data.ms
 /snapshots
 /dumps
+/bench
+/_xtask_benchmark.ms
 
 # Snapshots
 ## ... large

Cargo.toml

@@ -17,7 +17,7 @@ members = [
     "benchmarks",
     "fuzzers",
     "tracing-trace",
-    "xtask",
+    "xtask", "build-info",
 ]
 
 [workspace.package]

Dockerfile

@@ -8,7 +8,7 @@ WORKDIR /
 ARG     COMMIT_SHA
 ARG     COMMIT_DATE
 ARG     GIT_TAG
-ENV     VERGEN_GIT_SHA=${COMMIT_SHA} VERGEN_GIT_COMMIT_TIMESTAMP=${COMMIT_DATE} VERGEN_GIT_SEMVER_LIGHTWEIGHT=${GIT_TAG}
+ENV     VERGEN_GIT_SHA=${COMMIT_SHA} VERGEN_GIT_COMMIT_TIMESTAMP=${COMMIT_DATE} VERGEN_GIT_DESCRIBE=${GIT_TAG}
 ENV     RUSTFLAGS="-C target-feature=-crt-static"
 
 COPY    . .

build-info/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "build-info"
+version.workspace = true
+authors.workspace = true
+description.workspace = true
+homepage.workspace = true
+readme.workspace = true
+edition.workspace = true
+license.workspace = true
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+time = { version = "0.3.34", features = ["parsing"] }
+
+[build-dependencies]
+anyhow = "1.0.80"
+vergen-git2 = "1.0.0-beta.2"

build-info/build.rs

@@ -0,0 +1,22 @@
+fn main() {
+    if let Err(err) = emit_git_variables() {
+        println!("cargo:warning=vergen: {}", err);
+    }
+}
+
+fn emit_git_variables() -> anyhow::Result<()> {
+    // Note: any code that needs VERGEN_ environment variables should take care to define them manually in the Dockerfile and pass them
+    // in the corresponding GitHub workflow (publish_docker.yml).
+    // This is due to the Dockerfile building the binary outside of the git directory.
+    let mut builder = vergen_git2::Git2Builder::default();
+
+    builder.branch(true);
+    builder.commit_timestamp(true);
+    builder.commit_message(true);
+    builder.describe(true, true, None);
+    builder.sha(false);
+
+    let git2 = builder.build()?;
+
+    vergen_git2::Emitter::default().fail_on_error().add_instructions(&git2)?.emit()
+}

build-info/src/lib.rs

@@ -0,0 +1,203 @@
+use time::format_description::well_known::Iso8601;
+
+#[derive(Debug, Clone)]
+pub struct BuildInfo {
+    pub branch: Option<&'static str>,
+    pub describe: Option<DescribeResult>,
+    pub commit_sha1: Option<&'static str>,
+    pub commit_msg: Option<&'static str>,
+    pub commit_timestamp: Option<time::OffsetDateTime>,
+}
+
+impl BuildInfo {
+    pub fn from_build() -> Self {
+        let branch: Option<&'static str> = option_env!("VERGEN_GIT_BRANCH");
+        let describe = DescribeResult::from_build();
+        let commit_sha1 = option_env!("VERGEN_GIT_SHA");
+        let commit_msg = option_env!("VERGEN_GIT_COMMIT_MESSAGE");
+        let commit_timestamp = option_env!("VERGEN_GIT_COMMIT_TIMESTAMP");
+
+        let commit_timestamp = commit_timestamp.and_then(|commit_timestamp| {
+            time::OffsetDateTime::parse(commit_timestamp, &Iso8601::DEFAULT).ok()
+        });
+
+        Self { branch, describe, commit_sha1, commit_msg, commit_timestamp }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum DescribeResult {
+    Prototype { name: &'static str },
+    Release { version: &'static str, major: u64, minor: u64, patch: u64 },
+    Prerelease { version: &'static str, major: u64, minor: u64, patch: u64, rc: u64 },
+    NotATag { describe: &'static str },
+}
+
+impl DescribeResult {
+    pub fn new(describe: &'static str) -> Self {
+        if let Some(name) = prototype_name(describe) {
+            Self::Prototype { name }
+        } else if let Some(release) = release_version(describe) {
+            release
+        } else if let Some(prerelease) = prerelease_version(describe) {
+            prerelease
+        } else {
+            Self::NotATag { describe }
+        }
+    }
+
+    pub fn from_build() -> Option<Self> {
+        let describe: &'static str = option_env!("VERGEN_GIT_DESCRIBE")?;
+        Some(Self::new(describe))
+    }
+
+    pub fn as_tag(&self) -> Option<&'static str> {
+        match self {
+            DescribeResult::Prototype { name } => Some(name),
+            DescribeResult::Release { version, .. } => Some(version),
+            DescribeResult::Prerelease { version, .. } => Some(version),
+            DescribeResult::NotATag { describe: _ } => None,
+        }
+    }
+
+    pub fn as_prototype(&self) -> Option<&'static str> {
+        match self {
+            DescribeResult::Prototype { name } => Some(name),
+            DescribeResult::Release { .. }
+            | DescribeResult::Prerelease { .. }
+            | DescribeResult::NotATag { .. } => None,
+        }
+    }
+}
+
+/// Parses the input as a prototype name.
+///
+/// Returns `Some(prototype_name)` if the following conditions are met on this value:
+///
+/// 1. starts with `prototype-`,
+/// 2. ends with `-<some_number>`,
+/// 3. does not end with `<some_number>-<some_number>`.
+///
+/// Otherwise, returns `None`.
+fn prototype_name(describe: &'static str) -> Option<&'static str> {
+    if !describe.starts_with("prototype-") {
+        return None;
+    }
+
+    let mut rsplit_prototype = describe.rsplit('-');
+    // last component MUST be a number
+    rsplit_prototype.next()?.parse::<u64>().ok()?;
+    // before than last component SHALL NOT be a number
+    rsplit_prototype.next()?.parse::<u64>().err()?;
+
+    Some(describe)
+}
+
+fn release_version(describe: &'static str) -> Option<DescribeResult> {
+    if !describe.starts_with('v') {
+        return None;
+    }
+
+    // full release version don't contain a `-`
+    if describe.contains('-') {
+        return None;
+    }
+
+    // full release version parse as vX.Y.Z, with X, Y, Z numbers.
+    let mut dots = describe[1..].split('.');
+    let major: u64 = dots.next()?.parse().ok()?;
+    let minor: u64 = dots.next()?.parse().ok()?;
+    let patch: u64 = dots.next()?.parse().ok()?;
+
+    if dots.next().is_some() {
+        return None;
+    }
+
+    Some(DescribeResult::Release { version: describe, major, minor, patch })
+}
+
+fn prerelease_version(describe: &'static str) -> Option<DescribeResult> {
+    // prerelease version is in the shape vM.N.P-rc.C
+    let mut hyphen = describe.rsplit('-');
+    let prerelease = hyphen.next()?;
+    if !prerelease.starts_with("rc.") {
+        return None;
+    }
+
+    let rc: u64 = prerelease[3..].parse().ok()?;
+
+    let release = hyphen.next()?;
+
+    let DescribeResult::Release { version: _, major, minor, patch } = release_version(release)?
+    else {
+        return None;
+    };
+
+    Some(DescribeResult::Prerelease { version: describe, major, minor, patch, rc })
+}
+
+#[cfg(test)]
+mod test {
+    use super::DescribeResult;
+
+    fn assert_not_a_tag(describe: &'static str) {
+        assert_eq!(DescribeResult::NotATag { describe }, DescribeResult::new(describe))
+    }
+
+    fn assert_proto(describe: &'static str) {
+        assert_eq!(DescribeResult::Prototype { name: describe }, DescribeResult::new(describe))
+    }
+
+    fn assert_release(describe: &'static str, major: u64, minor: u64, patch: u64) {
+        assert_eq!(
+            DescribeResult::Release { version: describe, major, minor, patch },
+            DescribeResult::new(describe)
+        )
+    }
+
+    fn assert_prerelease(describe: &'static str, major: u64, minor: u64, patch: u64, rc: u64) {
+        assert_eq!(
+            DescribeResult::Prerelease { version: describe, major, minor, patch, rc },
+            DescribeResult::new(describe)
+        )
+    }
+
+    #[test]
+    fn not_a_tag() {
+        assert_not_a_tag("whatever-fuzzy");
+        assert_not_a_tag("whatever-fuzzy-5-ggg-dirty");
+        assert_not_a_tag("whatever-fuzzy-120-ggg-dirty");
+
+        // technically a tag, but not a proto nor a version, so not parsed as a tag
+        assert_not_a_tag("whatever");
+
+        // dirty version
+        assert_not_a_tag("v1.7.0-1-ggga-dirty");
+        assert_not_a_tag("v1.7.0-rc.1-1-ggga-dirty");
+
+        // after version
+        assert_not_a_tag("v1.7.0-1-ggga");
+        assert_not_a_tag("v1.7.0-rc.1-1-ggga");
+
+        // after proto
+        assert_not_a_tag("protoype-tag-0-1-ggga");
+        assert_not_a_tag("protoype-tag-0-1-ggga-dirty");
+    }
+
+    #[test]
+    fn prototype() {
+        assert_proto("prototype-tag-0");
+        assert_proto("prototype-tag-10");
+        assert_proto("prototype-long-name-tag-10");
+    }
+
+    #[test]
+    fn release() {
+        assert_release("v1.7.2", 1, 7, 2);
+    }
+
+    #[test]
+    fn prerelease() {
+        assert_prerelease("v1.7.2-rc.3", 1, 7, 2, 3);
+    }
+}

dump/src/reader/compat/v2_to_v3.rs

@@ -1,4 +1,3 @@
-use std::convert::TryInto;
 use std::str::FromStr;
 
 use time::OffsetDateTime;

file-store/Cargo.toml

@@ -13,6 +13,7 @@ license.workspace = true
 [dependencies]
 tempfile = "3.9.0"
 thiserror = "1.0.56"
+tracing = "0.1.40"
 uuid = { version = "1.6.1", features = ["serde", "v4"] }
 
 [dev-dependencies]

file-store/src/lib.rs

@@ -61,7 +61,13 @@ impl FileStore {
     /// Returns the file corresponding to the requested uuid.
     pub fn get_update(&self, uuid: Uuid) -> Result<StdFile> {
         let path = self.get_update_path(uuid);
-        let file = StdFile::open(path)?;
+        let file = match StdFile::open(path) {
+            Ok(file) => file,
+            Err(e) => {
+                tracing::error!("Can't access update file {uuid}: {e}");
+                return Err(e.into());
+            }
+        };
         Ok(file)
     }
 
@@ -96,8 +102,12 @@ impl FileStore {
 
     pub fn delete(&self, uuid: Uuid) -> Result<()> {
         let path = self.path.join(uuid.to_string());
-        std::fs::remove_file(path)?;
+        if let Err(e) = std::fs::remove_file(path) {
-        Ok(())
+            tracing::error!("Can't delete file {uuid}: {e}");
+            Err(e.into())
+        } else {
+            Ok(())
+        }
     }
 
     /// List the Uuids of the files in the FileStore

index-scheduler/Cargo.toml

@@ -23,6 +23,7 @@ meilisearch-auth = { path = "../meilisearch-auth" }
 meilisearch-types = { path = "../meilisearch-types" }
 page_size = "0.5.0"
 puffin = { version = "0.16.0", features = ["serialization"] }
+rayon = "1.8.1"
 roaring = { version = "0.10.2", features = ["serde"] }
 serde = { version = "1.0.195", features = ["derive"] }
 serde_json = { version = "1.0.111", features = ["preserve_order"] }

index-scheduler/src/batch.rs

@@ -142,22 +142,28 @@ pub(crate) enum IndexOperation {
 
 impl Batch {
     /// Return the task ids associated with this batch.
-    pub fn ids(&self) -> Vec<TaskId> {
+    pub fn ids(&self) -> RoaringBitmap {
         match self {
             Batch::TaskCancelation { task, .. }
             | Batch::Dump(task)
             | Batch::IndexCreation { task, .. }
-            | Batch::IndexUpdate { task, .. } => vec![task.uid],
+            | Batch::IndexUpdate { task, .. } => {
+                RoaringBitmap::from_sorted_iter(std::iter::once(task.uid)).unwrap()
+            }
             Batch::SnapshotCreation(tasks)
             | Batch::TaskDeletions(tasks)
-            | Batch::IndexDeletion { tasks, .. } => tasks.iter().map(|task| task.uid).collect(),
+            | Batch::IndexDeletion { tasks, .. } => {
+                RoaringBitmap::from_iter(tasks.iter().map(|task| task.uid))
+            }
             Batch::IndexOperation { op, .. } => match op {
                 IndexOperation::DocumentOperation { tasks, .. }
                 | IndexOperation::Settings { tasks, .. }
                 | IndexOperation::DocumentClear { tasks, .. } => {
-                    tasks.iter().map(|task| task.uid).collect()
+                    RoaringBitmap::from_iter(tasks.iter().map(|task| task.uid))
+                }
+                IndexOperation::IndexDocumentDeletionByFilter { task, .. } => {
+                    RoaringBitmap::from_sorted_iter(std::iter::once(task.uid)).unwrap()
                 }
-                IndexOperation::IndexDocumentDeletionByFilter { task, .. } => vec![task.uid],
                 IndexOperation::SettingsAndDocumentOperation {
                     document_import_tasks: tasks,
                     settings_tasks: other,
@@ -167,9 +173,11 @@ impl Batch {
                     cleared_tasks: tasks,
                     settings_tasks: other,
                     ..
-                } => tasks.iter().chain(other).map(|task| task.uid).collect(),
+                } => RoaringBitmap::from_iter(tasks.iter().chain(other).map(|task| task.uid)),
             },
-            Batch::IndexSwap { task } => vec![task.uid],
+            Batch::IndexSwap { task } => {
+                RoaringBitmap::from_sorted_iter(std::iter::once(task.uid)).unwrap()
+            }
         }
     }
 

index-scheduler/src/features.rs

@@ -48,7 +48,7 @@ impl RoFeatures {
             Ok(())
         } else {
             Err(FeatureNotEnabledError {
-                disabled_action: "getting logs through the `/logs/stream` route",
+                disabled_action: "Modifying logs through the `/logs/*` routes",
                 feature: "logs route",
                 issue_link: "https://github.com/orgs/meilisearch/discussions/721",
             }

index-scheduler/src/lib.rs

@@ -37,8 +37,8 @@ use std::fs::File;
 use std::io::{self, BufReader, Read};
 use std::ops::{Bound, RangeBounds};
 use std::path::{Path, PathBuf};
-use std::sync::atomic::AtomicBool;
+use std::sync::atomic::Ordering::{self, Relaxed};
-use std::sync::atomic::Ordering::Relaxed;
+use std::sync::atomic::{AtomicBool, AtomicU32};
 use std::sync::{Arc, RwLock};
 use std::time::Duration;
 
@@ -60,6 +60,8 @@ use meilisearch_types::milli::{self, CboRoaringBitmapCodec, Index, RoaringBitmap
 use meilisearch_types::task_view::TaskView;
 use meilisearch_types::tasks::{Kind, KindWithContent, Status, Task};
 use puffin::FrameView;
+use rayon::current_num_threads;
+use rayon::prelude::{IntoParallelIterator, ParallelIterator};
 use roaring::RoaringBitmap;
 use synchronoise::SignalEvent;
 use time::format_description::well_known::Rfc3339;
@@ -1170,15 +1172,13 @@ impl IndexScheduler {
         drop(rtxn);
 
         // 1. store the starting date with the bitmap of processing tasks.
-        let mut ids = batch.ids();
+        let ids = batch.ids();
-        ids.sort_unstable();
         let processed_tasks = ids.len();
-        let processing_tasks = RoaringBitmap::from_sorted_iter(ids.iter().copied()).unwrap();
         let started_at = OffsetDateTime::now_utc();
 
         // We reset the must_stop flag to be sure that we don't stop processing tasks
         self.must_stop_processing.reset();
-        self.processing_tasks.write().unwrap().start_processing_at(started_at, processing_tasks);
+        self.processing_tasks.write().unwrap().start_processing_at(started_at, ids.clone());
 
         #[cfg(test)]
         self.breakpoint(Breakpoint::BatchCreated);
@@ -1207,6 +1207,9 @@ impl IndexScheduler {
                 #[cfg(test)]
                 self.breakpoint(Breakpoint::ProcessBatchSucceeded);
 
+                let mut success = 0;
+                let mut failure = 0;
+
                 #[allow(unused_variables)]
                 for (i, mut task) in tasks.into_iter().enumerate() {
                     task.started_at = Some(started_at);
@@ -1219,13 +1222,15 @@ impl IndexScheduler {
                         },
                     )?;
 
+                    match task.error {
+                        Some(_) => failure += 1,
+                        None => success += 1,
+                    }
+
                     self.update_task(&mut wtxn, &task)
                         .map_err(|e| Error::TaskDatabaseUpdate(Box::new(e)))?;
-                    if let Err(e) = self.delete_persisted_task_data(&task) {
-                        tracing::error!("Failure to delete the content files associated with task {}. Error: {e}", task.uid);
-                    }
                 }
-                tracing::info!("A batch of tasks was successfully completed.");
+                tracing::info!("A batch of tasks was successfully completed with {success} successful tasks and {failure} failed tasks.");
             }
             // If we have an abortion error we must stop the tick here and re-schedule tasks.
             Err(Error::Milli(milli::Error::InternalError(
@@ -1236,6 +1241,7 @@ impl IndexScheduler {
                 self.breakpoint(Breakpoint::AbortedIndexation);
                 wtxn.abort();
 
+                tracing::info!("A batch of tasks was aborted.");
                 // We make sure that we don't call `stop_processing` on the `processing_tasks`,
                 // this is because we want to let the next tick call `create_next_batch` and keep
                 // the `started_at` date times and `processings` of the current processing tasks.
@@ -1257,6 +1263,8 @@ impl IndexScheduler {
                 self.index_mapper.resize_index(&wtxn, &index_uid)?;
                 wtxn.abort();
 
+                tracing::info!("The max database size was reached. Resizing the index.");
+
                 return Ok(TickOutcome::TickAgain(0));
             }
             // In case of a failure we must get back and patch all the tasks with the error.
@@ -1264,7 +1272,7 @@ impl IndexScheduler {
                 #[cfg(test)]
                 self.breakpoint(Breakpoint::ProcessBatchFailed);
                 let error: ResponseError = err.into();
-                for id in ids {
+                for id in ids.iter() {
                     let mut task = self
                         .get_task(&wtxn, id)
                         .map_err(|e| Error::TaskDatabaseUpdate(Box::new(e)))?
@@ -1278,9 +1286,8 @@ impl IndexScheduler {
                     #[cfg(test)]
                     self.maybe_fail(tests::FailureLocation::UpdatingTaskAfterProcessBatchFailure)?;
 
-                    if let Err(e) = self.delete_persisted_task_data(&task) {
+                    tracing::info!("Batch failed {}", error);
-                        tracing::error!("Failure to delete the content files associated with task {}. Error: {e}", task.uid);
+
-                    }
                     self.update_task(&mut wtxn, &task)
                         .map_err(|e| Error::TaskDatabaseUpdate(Box::new(e)))?;
                 }
@@ -1294,6 +1301,28 @@ impl IndexScheduler {
 
         wtxn.commit().map_err(Error::HeedTransaction)?;
 
+        // Once the tasks are commited, we should delete all the update files associated ASAP to avoid leaking files in case of a restart
+        tracing::debug!("Deleting the upadate files");
+
+        //We take one read transaction **per thread**. Then, every thread is going to pull out new IDs from the roaring bitmap with the help of an atomic shared index into the bitmap
+        let idx = AtomicU32::new(0);
+        (0..current_num_threads()).into_par_iter().try_for_each(|_| -> Result<()> {
+            let rtxn = self.read_txn()?;
+            while let Some(id) = ids.select(idx.fetch_add(1, Ordering::Relaxed)) {
+                let task = self
+                    .get_task(&rtxn, id)
+                    .map_err(|e| Error::TaskDatabaseUpdate(Box::new(e)))?
+                    .ok_or(Error::CorruptedTaskQueue)?;
+                if let Err(e) = self.delete_persisted_task_data(&task) {
+                    tracing::error!(
+                        "Failure to delete the content files associated with task {}. Error: {e}",
+                        task.uid
+                    );
+                }
+            }
+            Ok(())
+        })?;
+
         // We shouldn't crash the tick function if we can't send data to the webhook.
         let _ = self.notify_webhook(&processed);
 
@@ -1366,7 +1395,9 @@ impl IndexScheduler {
 
             // let reader = GzEncoder::new(BufReader::new(task_reader), Compression::default());
             let reader = GzEncoder::new(BufReader::new(task_reader), Compression::default());
-            let request = ureq::post(url).set("Content-Encoding", "gzip");
+            let request = ureq::post(url)
+                .set("Content-Encoding", "gzip")
+                .set("Content-Type", "application/x-ndjson");
             let request = match &self.webhook_authorization_header {
                 Some(header) => request.set("Authorization", header),
                 None => request,
@@ -1706,7 +1737,7 @@ pub enum TickOutcome {
     /// The scheduler should immediately attempt another `tick`.
     ///
     /// The `usize` field contains the number of processed tasks.
-    TickAgain(usize),
+    TickAgain(u64),
     /// The scheduler should wait for an external signal before attempting another `tick`.
     WaitForSignal,
 }

index-scheduler/src/uuid_codec.rs

@@ -1,5 +1,4 @@
 use std::borrow::Cow;
-use std::convert::TryInto;
 
 use meilisearch_types::heed::{BoxedError, BytesDecode, BytesEncode};
 use uuid::Uuid;

meilisearch-auth/src/store.rs

@@ -1,7 +1,6 @@
 use std::borrow::Cow;
 use std::cmp::Reverse;
 use std::collections::HashSet;
-use std::convert::{TryFrom, TryInto};
 use std::fs::create_dir_all;
 use std::path::Path;
 use std::result::Result as StdResult;

meilisearch-types/Cargo.toml

@@ -54,3 +54,5 @@ thai = ["milli/thai"]
 greek = ["milli/greek"]
 # allow khmer specialized tokenization
 khmer = ["milli/khmer"]
+# allow vietnamese specialized tokenization
+vietnamese = ["milli/vietnamese"]

meilisearch/Cargo.toml

@@ -104,9 +104,10 @@ serde_urlencoded = "0.7.1"
 termcolor = "1.4.1"
 url = { version = "2.5.0", features = ["serde"] }
 tracing = "0.1.40"
-tracing-subscriber = "0.3.18"
+tracing-subscriber = { version = "0.3.18", features = ["json"] }
 tracing-trace = { version = "0.1.0", path = "../tracing-trace" }
 tracing-actix-web = "0.7.9"
+build-info = { version = "1.7.0", path = "../build-info" }
 
 [dev-dependencies]
 actix-rt = "2.9.0"
@@ -131,7 +132,6 @@ reqwest = { version = "0.11.23", features = [
 sha-1 = { version = "0.10.1", optional = true }
 static-files = { version = "0.2.3", optional = true }
 tempfile = { version = "3.9.0", optional = true }
-vergen = { version = "7.5.1", default-features = false, features = ["git"] }
 zip = { version = "0.6.6", optional = true }
 
 [features]
@@ -154,6 +154,7 @@ japanese = ["meilisearch-types/japanese"]
 thai = ["meilisearch-types/thai"]
 greek = ["meilisearch-types/greek"]
 khmer = ["meilisearch-types/khmer"]
+vietnamese = ["meilisearch-types/vietnamese"]
 
 [package.metadata.mini-dashboard]
 assets-url = "https://github.com/meilisearch/mini-dashboard/releases/download/v0.2.13/build.zip"

meilisearch/build.rs

@@ -1,17 +1,4 @@
-use vergen::{vergen, Config, SemverKind};
-
 fn main() {
-    // Note: any code that needs VERGEN_ environment variables should take care to define them manually in the Dockerfile and pass them
-    // in the corresponding GitHub workflow (publish_docker.yml).
-    // This is due to the Dockerfile building the binary outside of the git directory.
-    let mut config = Config::default();
-    // allow using non-annotated tags
-    *config.git_mut().semver_kind_mut() = SemverKind::Lightweight;
-
-    if let Err(e) = vergen(config) {
-        println!("cargo:warning=vergen: {}", e);
-    }
-
     #[cfg(feature = "mini-dashboard")]
     mini_dashboard::setup_mini_dashboard().expect("Could not load the mini-dashboard assets");
 }

meilisearch/src/analytics/segment_analytics.rs

@@ -28,7 +28,9 @@ use super::{
     config_user_id_path, DocumentDeletionKind, DocumentFetchKind, MEILISEARCH_CONFIG_PATH,
 };
 use crate::analytics::Analytics;
-use crate::option::{default_http_addr, IndexerOpts, MaxMemory, MaxThreads, ScheduleSnapshot};
+use crate::option::{
+    default_http_addr, IndexerOpts, LogMode, MaxMemory, MaxThreads, ScheduleSnapshot,
+};
 use crate::routes::indexes::documents::UpdateDocumentsQuery;
 use crate::routes::indexes::facet_search::FacetSearchQuery;
 use crate::routes::tasks::TasksFilterQuery;
@@ -250,10 +252,12 @@ impl super::Analytics for SegmentAnalytics {
 struct Infos {
     env: String,
     experimental_enable_metrics: bool,
+    experimental_logs_mode: LogMode,
     experimental_replication_parameters: bool,
     experimental_enable_logs_route: bool,
     experimental_reduce_indexing_memory_usage: bool,
     experimental_max_number_of_batched_tasks: usize,
+    gpu_enabled: bool,
     db_path: bool,
     import_dump: bool,
     dump_dir: bool,
@@ -289,6 +293,7 @@ impl From<Opt> for Infos {
         let Opt {
             db_path,
             experimental_enable_metrics,
+            experimental_logs_mode,
             experimental_replication_parameters,
             experimental_enable_logs_route,
             experimental_reduce_indexing_memory_usage,
@@ -337,9 +342,11 @@ impl From<Opt> for Infos {
         Self {
             env,
             experimental_enable_metrics,
+            experimental_logs_mode,
             experimental_replication_parameters,
             experimental_enable_logs_route,
             experimental_reduce_indexing_memory_usage,
+            gpu_enabled: meilisearch_types::milli::vector::is_cuda_enabled(),
             db_path: db_path != PathBuf::from("./data.ms"),
             import_dump: import_dump.is_some(),
             dump_dir: dump_dir != PathBuf::from("dumps/"),
@@ -466,7 +473,9 @@ impl Segment {
             create_all_stats(index_scheduler.into(), auth_controller.into(), &AuthFilter::default())
         {
             // Replace the version number with the prototype name if any.
-            let version = if let Some(prototype) = crate::prototype_name() {
+            let version = if let Some(prototype) = build_info::DescribeResult::from_build()
+                .and_then(|describe| describe.as_prototype())
+            {
                 prototype
             } else {
                 env!("CARGO_PKG_VERSION")

meilisearch/src/lib.rs

@@ -97,11 +97,25 @@ pub type LogRouteType = tracing_subscriber::filter::Filtered<
     tracing_subscriber::Registry,
 >;
 
+pub type SubscriberForSecondLayer = tracing_subscriber::layer::Layered<
+    tracing_subscriber::reload::Layer<LogRouteType, tracing_subscriber::Registry>,
+    tracing_subscriber::Registry,
+>;
+
+pub type LogStderrHandle =
+    tracing_subscriber::reload::Handle<LogStderrType, SubscriberForSecondLayer>;
+
+pub type LogStderrType = tracing_subscriber::filter::Filtered<
+    Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>,
+    Targets,
+    SubscriberForSecondLayer,
+>;
+
 pub fn create_app(
     index_scheduler: Data<IndexScheduler>,
     auth_controller: Data<AuthController>,
     opt: Opt,
-    logs: LogRouteHandle,
+    logs: (LogRouteHandle, LogStderrHandle),
     analytics: Arc<dyn Analytics>,
     enable_dashboard: bool,
 ) -> actix_web::App<
@@ -412,6 +426,9 @@ fn import_dump(
         let reader = BufReader::new(file);
         let reader = DocumentsBatchReader::from_reader(reader)?;
 
+        let embedder_configs = index.embedding_configs(&wtxn)?;
+        let embedders = index_scheduler.embedders(embedder_configs)?;
+
         let builder = milli::update::IndexDocuments::new(
             &mut wtxn,
             &index,
@@ -424,6 +441,8 @@ fn import_dump(
             || false,
         )?;
 
+        let builder = builder.with_embedders(embedders);
+
         let (builder, user_result) = builder.add_documents(reader)?;
         let user_result = user_result?;
         tracing::info!(documents_found = user_result, "{} documents found.", user_result);
@@ -447,7 +466,7 @@ pub fn configure_data(
     index_scheduler: Data<IndexScheduler>,
     auth: Data<AuthController>,
     opt: &Opt,
-    logs: LogRouteHandle,
+    (logs_route, logs_stderr): (LogRouteHandle, LogStderrHandle),
     analytics: Arc<dyn Analytics>,
 ) {
     let http_payload_size_limit = opt.http_payload_size_limit.get_bytes() as usize;
@@ -455,7 +474,8 @@ pub fn configure_data(
         .app_data(index_scheduler)
         .app_data(auth)
         .app_data(web::Data::from(analytics))
-        .app_data(web::Data::new(logs))
+        .app_data(web::Data::new(logs_route))
+        .app_data(web::Data::new(logs_stderr))
         .app_data(web::Data::new(opt.clone()))
         .app_data(
             web::JsonConfig::default()
@@ -516,30 +536,3 @@ pub fn dashboard(config: &mut web::ServiceConfig, enable_frontend: bool) {
 pub fn dashboard(config: &mut web::ServiceConfig, _enable_frontend: bool) {
     config.service(web::resource("/").route(web::get().to(routes::running)));
 }
-
-/// Parses the output of
-/// [`VERGEN_GIT_SEMVER_LIGHTWEIGHT`](https://docs.rs/vergen/latest/vergen/struct.Git.html#instructions)
-///  as a prototype name.
-///
-/// Returns `Some(prototype_name)` if the following conditions are met on this value:
-///
-/// 1. starts with `prototype-`,
-/// 2. ends with `-<some_number>`,
-/// 3. does not end with `<some_number>-<some_number>`.
-///
-/// Otherwise, returns `None`.
-pub fn prototype_name() -> Option<&'static str> {
-    let prototype: &'static str = option_env!("VERGEN_GIT_SEMVER_LIGHTWEIGHT")?;
-
-    if !prototype.starts_with("prototype-") {
-        return None;
-    }
-
-    let mut rsplit_prototype = prototype.rsplit('-');
-    // last component MUST be a number
-    rsplit_prototype.next()?.parse::<u64>().ok()?;
-    // before than last component SHALL NOT be a number
-    rsplit_prototype.next()?.parse::<u64>().err()?;
-
-    Some(prototype)
-}

meilisearch/src/main.rs

@@ -1,5 +1,5 @@
 use std::env;
-use std::io::{stderr, Write};
+use std::io::{stderr, LineWriter, Write};
 use std::path::PathBuf;
 use std::str::FromStr;
 use std::sync::Arc;
@@ -10,8 +10,10 @@ use actix_web::HttpServer;
 use index_scheduler::IndexScheduler;
 use is_terminal::IsTerminal;
 use meilisearch::analytics::Analytics;
+use meilisearch::option::LogMode;
 use meilisearch::{
-    analytics, create_app, prototype_name, setup_meilisearch, LogRouteHandle, LogRouteType, Opt,
+    analytics, create_app, setup_meilisearch, LogRouteHandle, LogRouteType, LogStderrHandle,
+    LogStderrType, Opt, SubscriberForSecondLayer,
 };
 use meilisearch_auth::{generate_master_key, AuthController, MASTER_KEY_MIN_SIZE};
 use mimalloc::MiMalloc;
@@ -23,28 +25,44 @@ use tracing_subscriber::Layer;
 #[global_allocator]
 static ALLOC: MiMalloc = MiMalloc;
 
-fn default_layer() -> LogRouteType {
+fn default_log_route_layer() -> LogRouteType {
     None.with_filter(tracing_subscriber::filter::Targets::new().with_target("", LevelFilter::OFF))
 }
 
+fn default_log_stderr_layer(opt: &Opt) -> LogStderrType {
+    let layer = tracing_subscriber::fmt::layer()
+        .with_writer(|| LineWriter::new(std::io::stderr()))
+        .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE);
+
+    let layer = match opt.experimental_logs_mode {
+        LogMode::Human => Box::new(layer)
+            as Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>,
+        LogMode::Json => Box::new(layer.json())
+            as Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>,
+    };
+
+    layer.with_filter(
+        tracing_subscriber::filter::Targets::new()
+            .with_target("", LevelFilter::from_str(&opt.log_level.to_string()).unwrap()),
+    )
+}
+
 /// does all the setup before meilisearch is launched
-fn setup(opt: &Opt) -> anyhow::Result<LogRouteHandle> {
+fn setup(opt: &Opt) -> anyhow::Result<(LogRouteHandle, LogStderrHandle)> {
-    let (route_layer, route_layer_handle) = tracing_subscriber::reload::Layer::new(default_layer());
+    let (route_layer, route_layer_handle) =
+        tracing_subscriber::reload::Layer::new(default_log_route_layer());
     let route_layer: tracing_subscriber::reload::Layer<_, _> = route_layer;
 
-    let subscriber = tracing_subscriber::registry().with(route_layer).with(
+    let (stderr_layer, stderr_layer_handle) =
-        tracing_subscriber::fmt::layer()
+        tracing_subscriber::reload::Layer::new(default_log_stderr_layer(opt));
-            .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE)
+    let route_layer: tracing_subscriber::reload::Layer<_, _> = route_layer;
-            .with_filter(
+
-                tracing_subscriber::filter::LevelFilter::from_str(&opt.log_level.to_string())
+    let subscriber = tracing_subscriber::registry().with(route_layer).with(stderr_layer);
-                    .unwrap(),
-            ),
-    );
 
     // set the subscriber as the default for the application
     tracing::subscriber::set_global_default(subscriber).unwrap();
 
-    Ok(route_layer_handle)
+    Ok((route_layer_handle, stderr_layer_handle))
 }
 
 fn on_panic(info: &std::panic::PanicInfo) {
@@ -110,7 +128,7 @@ async fn run_http(
     index_scheduler: Arc<IndexScheduler>,
     auth_controller: Arc<AuthController>,
     opt: Opt,
-    logs: LogRouteHandle,
+    logs: (LogRouteHandle, LogStderrHandle),
     analytics: Arc<dyn Analytics>,
 ) -> anyhow::Result<()> {
     let enable_dashboard = &opt.env == "development";
@@ -145,8 +163,8 @@ pub fn print_launch_resume(
     analytics: Arc<dyn Analytics>,
     config_read_from: Option<PathBuf>,
 ) {
-    let commit_sha = option_env!("VERGEN_GIT_SHA").unwrap_or("unknown");
+    let build_info = build_info::BuildInfo::from_build();
-    let commit_date = option_env!("VERGEN_GIT_COMMIT_TIMESTAMP").unwrap_or("unknown");
+
     let protocol =
         if opt.ssl_cert_path.is_some() && opt.ssl_key_path.is_some() { "https" } else { "http" };
     let ascii_name = r#"
@@ -171,10 +189,18 @@ pub fn print_launch_resume(
     eprintln!("Database path:\t\t{:?}", opt.db_path);
     eprintln!("Server listening on:\t\"{}://{}\"", protocol, opt.http_addr);
     eprintln!("Environment:\t\t{:?}", opt.env);
-    eprintln!("Commit SHA:\t\t{:?}", commit_sha.to_string());
+    eprintln!("Commit SHA:\t\t{:?}", build_info.commit_sha1.unwrap_or("unknown"));
-    eprintln!("Commit date:\t\t{:?}", commit_date.to_string());
+    eprintln!(
+        "Commit date:\t\t{:?}",
+        build_info
+            .commit_timestamp
+            .and_then(|commit_timestamp| commit_timestamp
+                .format(&time::format_description::well_known::Rfc3339)
+                .ok())
+            .unwrap_or("unknown".into())
+    );
     eprintln!("Package version:\t{:?}", env!("CARGO_PKG_VERSION").to_string());
-    if let Some(prototype) = prototype_name() {
+    if let Some(prototype) = build_info.describe.and_then(|describe| describe.as_prototype()) {
         eprintln!("Prototype:\t\t{:?}", prototype);
     }
 

meilisearch/src/option.rs

@@ -1,4 +1,3 @@
-use std::convert::TryFrom;
 use std::env::VarError;
 use std::ffi::OsStr;
 use std::fmt::Display;
@@ -51,6 +50,7 @@ const MEILI_IGNORE_MISSING_DUMP: &str = "MEILI_IGNORE_MISSING_DUMP";
 const MEILI_IGNORE_DUMP_IF_DB_EXISTS: &str = "MEILI_IGNORE_DUMP_IF_DB_EXISTS";
 const MEILI_DUMP_DIR: &str = "MEILI_DUMP_DIR";
 const MEILI_LOG_LEVEL: &str = "MEILI_LOG_LEVEL";
+const MEILI_EXPERIMENTAL_LOGS_MODE: &str = "MEILI_EXPERIMENTAL_LOGS_MODE";
 const MEILI_EXPERIMENTAL_REPLICATION_PARAMETERS: &str = "MEILI_EXPERIMENTAL_REPLICATION_PARAMETERS";
 const MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE: &str = "MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE";
 const MEILI_EXPERIMENTAL_ENABLE_METRICS: &str = "MEILI_EXPERIMENTAL_ENABLE_METRICS";
@@ -80,6 +80,39 @@ const DEFAULT_LOG_EVERY_N: usize = 100_000;
 pub const INDEX_SIZE: u64 = 2 * 1024 * 1024 * 1024 * 1024; // 2 TiB
 pub const TASK_DB_SIZE: u64 = 20 * 1024 * 1024 * 1024; // 20 GiB
 
+#[derive(Debug, Default, Clone, Copy, Serialize, Deserialize)]
+#[serde(rename_all = "UPPERCASE")]
+pub enum LogMode {
+    #[default]
+    Human,
+    Json,
+}
+
+impl Display for LogMode {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            LogMode::Human => Display::fmt("HUMAN", f),
+            LogMode::Json => Display::fmt("JSON", f),
+        }
+    }
+}
+
+impl FromStr for LogMode {
+    type Err = LogModeError;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.trim().to_lowercase().as_str() {
+            "human" => Ok(LogMode::Human),
+            "json" => Ok(LogMode::Json),
+            _ => Err(LogModeError(s.to_owned())),
+        }
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+#[error("Unsupported log mode level `{0}`. Supported values are `HUMAN` and `JSON`.")]
+pub struct LogModeError(String);
+
 #[derive(Debug, Default, Clone, Copy, Serialize, Deserialize)]
 #[serde(rename_all = "UPPERCASE")]
 pub enum LogLevel {
@@ -311,9 +344,16 @@ pub struct Opt {
     #[serde(default)]
     pub experimental_enable_metrics: bool,
 
+    /// Experimental logs mode feature. For more information, see: <https://github.com/orgs/meilisearch/discussions/723>
+    ///
+    /// Change the mode of the logs on the console.
+    #[clap(long, env = MEILI_EXPERIMENTAL_LOGS_MODE, default_value_t)]
+    #[serde(default)]
+    pub experimental_logs_mode: LogMode,
+
     /// Experimental logs route feature. For more information, see: <https://github.com/orgs/meilisearch/discussions/721>
     ///
-    /// Enables the log route on the `POST /logs/stream` endpoint and the `DELETE /logs/stream` to stop receiving logs.
+    /// Enables the log routes on the `POST /logs/stream`, `POST /logs/stderr` endpoints, and the `DELETE /logs/stream` to stop receiving logs.
     #[clap(long, env = MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE)]
     #[serde(default)]
     pub experimental_enable_logs_route: bool,
@@ -433,6 +473,7 @@ impl Opt {
             #[cfg(feature = "analytics")]
             no_analytics,
             experimental_enable_metrics,
+            experimental_logs_mode,
             experimental_enable_logs_route,
             experimental_replication_parameters,
             experimental_reduce_indexing_memory_usage,
@@ -491,6 +532,10 @@ impl Opt {
             MEILI_EXPERIMENTAL_ENABLE_METRICS,
             experimental_enable_metrics.to_string(),
         );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_LOGS_MODE,
+            experimental_logs_mode.to_string(),
+        );
         export_to_env_if_not_present(
             MEILI_EXPERIMENTAL_REPLICATION_PARAMETERS,
             experimental_replication_parameters.to_string(),

meilisearch/src/routes/api_key.rs

@@ -10,7 +10,7 @@ use meilisearch_types::deserr::query_params::Param;
 use meilisearch_types::deserr::{DeserrJsonError, DeserrQueryParamError};
 use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::error::{Code, ResponseError};
-use meilisearch_types::keys::{Action, CreateApiKey, Key, PatchApiKey};
+use meilisearch_types::keys::{CreateApiKey, Key, PatchApiKey};
 use serde::{Deserialize, Serialize};
 use time::OffsetDateTime;
 use uuid::Uuid;

meilisearch/src/routes/logs.rs

@@ -22,14 +22,15 @@ use crate::error::MeilisearchHttpError;
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
-use crate::LogRouteHandle;
+use crate::{LogRouteHandle, LogStderrHandle};
 
 pub fn configure(cfg: &mut web::ServiceConfig) {
     cfg.service(
         web::resource("stream")
             .route(web::post().to(SeqHandler(get_logs)))
             .route(web::delete().to(SeqHandler(cancel_logs))),
-    );
+    )
+    .service(web::resource("stderr").route(web::post().to(SeqHandler(update_stderr_target))));
 }
 
 #[derive(Debug, Default, Clone, Copy, Deserr, PartialEq, Eq)]
@@ -37,6 +38,7 @@ pub fn configure(cfg: &mut web::ServiceConfig) {
 pub enum LogMode {
     #[default]
     Human,
+    Json,
     Profile,
 }
 
@@ -165,7 +167,18 @@ fn make_layer<
 
             let fmt_layer = tracing_subscriber::fmt::layer()
                 .with_writer(move || LogWriter { sender: sender.clone() })
-                .with_span_events(tracing_subscriber::fmt::format::FmtSpan::ACTIVE);
+                .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE);
+
+            let stream = byte_stream(receiver, guard);
+            (Box::new(fmt_layer) as Box<dyn Layer<S> + Send + Sync>, Box::pin(stream))
+        }
+        LogMode::Json => {
+            let (sender, receiver) = tokio::sync::mpsc::unbounded_channel();
+
+            let fmt_layer = tracing_subscriber::fmt::layer()
+                .with_writer(move || LogWriter { sender: sender.clone() })
+                .json()
+                .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE);
 
             let stream = byte_stream(receiver, guard);
             (Box::new(fmt_layer) as Box<dyn Layer<S> + Send + Sync>, Box::pin(stream))
@@ -279,3 +292,27 @@ pub async fn cancel_logs(
 
     Ok(HttpResponse::NoContent().finish())
 }
+
+#[derive(Debug, Deserr)]
+#[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
+pub struct UpdateStderrLogs {
+    #[deserr(default = "info".parse().unwrap(), try_from(&String) = MyTargets::from_str -> DeserrJsonError<BadRequest>)]
+    target: MyTargets,
+}
+
+pub async fn update_stderr_target(
+    index_scheduler: GuardedData<ActionPolicy<{ actions::METRICS_GET }>, Data<IndexScheduler>>,
+    logs: Data<LogStderrHandle>,
+    body: AwebJson<UpdateStderrLogs, DeserrJsonError>,
+) -> Result<HttpResponse, ResponseError> {
+    index_scheduler.features().check_logs_route()?;
+
+    let opt = body.into_inner();
+
+    logs.modify(|layer| {
+        *layer.filter_mut() = opt.target.0.clone();
+    })
+    .unwrap();
+
+    Ok(HttpResponse::NoContent().finish())
+}

meilisearch/src/routes/mod.rs

@@ -359,12 +359,18 @@ async fn get_version(
 ) -> HttpResponse {
     analytics.publish("Version Seen".to_string(), json!(null), Some(&req));
 
-    let commit_sha = option_env!("VERGEN_GIT_SHA").unwrap_or("unknown");
+    let build_info = build_info::BuildInfo::from_build();
-    let commit_date = option_env!("VERGEN_GIT_COMMIT_TIMESTAMP").unwrap_or("unknown");
 
     HttpResponse::Ok().json(VersionResponse {
-        commit_sha: commit_sha.to_string(),
+        commit_sha: build_info.commit_sha1.unwrap_or("unknown").to_string(),
-        commit_date: commit_date.to_string(),
+        commit_date: build_info
+            .commit_timestamp
+            .and_then(|commit_timestamp| {
+                commit_timestamp
+                    .format(&time::format_description::well_known::Iso8601::DEFAULT)
+                    .ok()
+            })
+            .unwrap_or("unknown".into()),
         pkg_version: env!("CARGO_PKG_VERSION").to_string(),
     })
 }

meilisearch/tests/common/server.rs

@@ -9,7 +9,7 @@ use actix_web::http::StatusCode;
 use byte_unit::{Byte, ByteUnit};
 use clap::Parser;
 use meilisearch::option::{IndexerOpts, MaxMemory, Opt};
-use meilisearch::{analytics, create_app, setup_meilisearch};
+use meilisearch::{analytics, create_app, setup_meilisearch, SubscriberForSecondLayer};
 use once_cell::sync::Lazy;
 use tempfile::TempDir;
 use tokio::time::sleep;
@@ -87,12 +87,20 @@ impl Server {
             tracing_subscriber::reload::Layer::new(None.with_filter(
                 tracing_subscriber::filter::Targets::new().with_target("", LevelFilter::OFF),
             ));
+        let (_stderr_layer, stderr_layer_handle) = tracing_subscriber::reload::Layer::new(
+            (Box::new(
+                tracing_subscriber::fmt::layer()
+                    .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE),
+            )
+                as Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>)
+                .with_filter(tracing_subscriber::filter::Targets::new()),
+        );
 
         actix_web::test::init_service(create_app(
             self.service.index_scheduler.clone().into(),
             self.service.auth.clone().into(),
             self.service.options.clone(),
-            route_layer_handle,
+            (route_layer_handle, stderr_layer_handle),
             analytics::MockAnalytics::new(&self.service.options),
             true,
         ))

meilisearch/tests/common/service.rs

@@ -5,7 +5,7 @@ use actix_web::http::StatusCode;
 use actix_web::test;
 use actix_web::test::TestRequest;
 use index_scheduler::IndexScheduler;
-use meilisearch::{analytics, create_app, Opt};
+use meilisearch::{analytics, create_app, Opt, SubscriberForSecondLayer};
 use meilisearch_auth::AuthController;
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::Layer;
@@ -111,12 +111,20 @@ impl Service {
             tracing_subscriber::reload::Layer::new(None.with_filter(
                 tracing_subscriber::filter::Targets::new().with_target("", LevelFilter::OFF),
             ));
+        let (_stderr_layer, stderr_layer_handle) = tracing_subscriber::reload::Layer::new(
+            (Box::new(
+                tracing_subscriber::fmt::layer()
+                    .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE),
+            )
+                as Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>)
+                .with_filter(tracing_subscriber::filter::Targets::new()),
+        );
 
         let app = test::init_service(create_app(
             self.index_scheduler.clone().into(),
             self.auth.clone().into(),
             self.options.clone(),
-            route_layer_handle,
+            (route_layer_handle, stderr_layer_handle),
             analytics::MockAnalytics::new(&self.options),
             true,
         ))

meilisearch/tests/documents/update_documents.rs

@@ -1,4 +1,4 @@
-use meili_snap::snapshot;
+use meili_snap::{json_string, snapshot};
 
 use crate::common::encoder::Encoder;
 use crate::common::{GetAllDocumentsOptions, Server};
@@ -209,3 +209,93 @@ async fn error_update_documents_missing_document_id() {
         "https://docs.meilisearch.com/errors#missing_document_id"
     );
 }
+
+#[actix_rt::test]
+async fn update_faceted_document() {
+    let server = Server::new().await;
+    let index = server.index("test");
+
+    let (response, code) = index
+        .update_settings(json!({
+            "rankingRules": ["facet:asc"],
+        }))
+        .await;
+    assert_eq!("202", code.as_str(), "{:?}", response);
+    index.wait_task(0).await;
+
+    let documents: Vec<_> = (0..1000)
+        .map(|id| {
+            json!({
+                "doc_id": id,
+                "facet": (id/3),
+            })
+        })
+        .collect();
+
+    let (_response, code) = index.add_documents(documents.into(), None).await;
+    assert_eq!(code, 202);
+
+    index.wait_task(1).await;
+
+    let documents = json!([
+        {
+            "doc_id": 9,
+            "facet": 1.5,
+        }
+    ]);
+
+    let (response, code) = index.update_documents(documents, None).await;
+    assert_eq!(code, 202, "response: {}", response);
+
+    index.wait_task(2).await;
+
+    index
+        .search(json!({"limit": 10}), |response, code| {
+            snapshot!(code, @"200 OK");
+            snapshot!(json_string!(response["hits"]), @r###"
+            [
+              {
+                "doc_id": 0,
+                "facet": 0
+              },
+              {
+                "doc_id": 1,
+                "facet": 0
+              },
+              {
+                "doc_id": 2,
+                "facet": 0
+              },
+              {
+                "doc_id": 3,
+                "facet": 1
+              },
+              {
+                "doc_id": 4,
+                "facet": 1
+              },
+              {
+                "doc_id": 5,
+                "facet": 1
+              },
+              {
+                "doc_id": 9,
+                "facet": 1.5
+              },
+              {
+                "doc_id": 6,
+                "facet": 2
+              },
+              {
+                "doc_id": 7,
+                "facet": 2
+              },
+              {
+                "doc_id": 8,
+                "facet": 2
+              }
+            ]
+            "###);
+        })
+        .await;
+}

meilisearch/tests/logs/error.rs

@@ -89,7 +89,7 @@ async fn logs_stream_bad_mode() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(response, @r###"
     {
-      "message": "Unknown value `tamo` at `.mode`: expected one of `human`, `profile`",
+      "message": "Unknown value `tamo` at `.mode`: expected one of `human`, `json`, `profile`",
       "code": "bad_request",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#bad_request"
@@ -146,7 +146,7 @@ async fn logs_stream_bad_profile_memory() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(response, @r###"
     {
-      "message": "Unknown value `fmt` at `.mode`: expected one of `human`, `profile`",
+      "message": "Unknown value `fmt` at `.mode`: expected one of `human`, `json`, `profile`",
       "code": "bad_request",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#bad_request"
@@ -162,7 +162,7 @@ async fn logs_stream_without_enabling_the_route() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(response, @r###"
     {
-      "message": "getting logs through the `/logs/stream` route requires enabling the `logs route` experimental feature. See https://github.com/orgs/meilisearch/discussions/721",
+      "message": "Modifying logs through the `/logs/*` routes requires enabling the `logs route` experimental feature. See https://github.com/orgs/meilisearch/discussions/721",
       "code": "feature_not_enabled",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
@@ -173,7 +173,18 @@ async fn logs_stream_without_enabling_the_route() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(response, @r###"
     {
-      "message": "getting logs through the `/logs/stream` route requires enabling the `logs route` experimental feature. See https://github.com/orgs/meilisearch/discussions/721",
+      "message": "Modifying logs through the `/logs/*` routes requires enabling the `logs route` experimental feature. See https://github.com/orgs/meilisearch/discussions/721",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+
+    let (response, code) = server.service.post("/logs/stderr", json!({})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(response, @r###"
+    {
+      "message": "Modifying logs through the `/logs/*` routes requires enabling the `logs route` experimental feature. See https://github.com/orgs/meilisearch/discussions/721",
       "code": "feature_not_enabled",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#feature_not_enabled"

meilisearch/tests/logs/mod.rs

@@ -5,7 +5,7 @@ use std::str::FromStr;
 
 use actix_web::http::header::ContentType;
 use meili_snap::snapshot;
-use meilisearch::{analytics, create_app, Opt};
+use meilisearch::{analytics, create_app, Opt, SubscriberForSecondLayer};
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::Layer;
@@ -27,18 +27,25 @@ async fn basic_test_log_stream_route() {
         tracing_subscriber::reload::Layer::new(None.with_filter(
             tracing_subscriber::filter::Targets::new().with_target("", LevelFilter::OFF),
         ));
+    let (_stderr_layer, stderr_layer_handle) = tracing_subscriber::reload::Layer::new(
+        (Box::new(
+            tracing_subscriber::fmt::layer()
+                .with_span_events(tracing_subscriber::fmt::format::FmtSpan::CLOSE),
+        ) as Box<dyn tracing_subscriber::Layer<SubscriberForSecondLayer> + Send + Sync>)
+            .with_filter(tracing_subscriber::filter::Targets::new()),
+    );
 
     let subscriber = tracing_subscriber::registry().with(route_layer).with(
         tracing_subscriber::fmt::layer()
             .with_span_events(tracing_subscriber::fmt::format::FmtSpan::ACTIVE)
-            .with_filter(tracing_subscriber::filter::LevelFilter::from_str("INFO").unwrap()),
+            .with_filter(tracing_subscriber::filter::LevelFilter::from_str("OFF").unwrap()),
     );
 
     let app = actix_web::test::init_service(create_app(
         server.service.index_scheduler.clone().into(),
         server.service.auth.clone().into(),
         server.service.options.clone(),
-        route_layer_handle,
+        (route_layer_handle, stderr_layer_handle),
         analytics::MockAnalytics::new(&server.service.options),
         true,
     ))

meilisearch/tests/tasks/webhook.rs

@@ -7,7 +7,7 @@ use std::sync::Arc;
 use actix_http::body::MessageBody;
 use actix_web::dev::{ServiceFactory, ServiceResponse};
 use actix_web::web::{Bytes, Data};
-use actix_web::{post, App, HttpResponse, HttpServer};
+use actix_web::{post, App, HttpRequest, HttpResponse, HttpServer};
 use meili_snap::{json_string, snapshot};
 use meilisearch::Opt;
 use tokio::sync::mpsc;
@@ -17,7 +17,17 @@ use crate::common::{default_settings, Server};
 use crate::json;
 
 #[post("/")]
-async fn forward_body(sender: Data<mpsc::UnboundedSender<Vec<u8>>>, body: Bytes) -> HttpResponse {
+async fn forward_body(
+    req: HttpRequest,
+    sender: Data<mpsc::UnboundedSender<Vec<u8>>>,
+    body: Bytes,
+) -> HttpResponse {
+    let headers = req.headers();
+    assert_eq!(headers.get("content-type").unwrap(), "application/x-ndjson");
+    assert_eq!(headers.get("transfer-encoding").unwrap(), "chunked");
+    assert_eq!(headers.get("accept-encoding").unwrap(), "gzip");
+    assert_eq!(headers.get("content-encoding").unwrap(), "gzip");
+
     let body = body.to_vec();
     sender.send(body).unwrap();
     HttpResponse::Ok().into()

meilitool/src/uuid_codec.rs

@@ -1,5 +1,4 @@
 use std::borrow::Cow;
-use std::convert::TryInto;
 
 use meilisearch_types::heed::{BoxedError, BytesDecode, BytesEncode};
 use uuid::Uuid;

milli/Cargo.toml

@@ -17,7 +17,7 @@ bincode = "1.3.3"
 bstr = "1.9.0"
 bytemuck = { version = "1.14.0", features = ["extern_crate_alloc"] }
 byteorder = "1.5.0"
-charabia = { version = "0.8.5", default-features = false }
+charabia = { version = "0.8.7", default-features = false }
 concat-arrays = "0.1.2"
 crossbeam-channel = "0.5.11"
 deserr = "0.6.1"
@@ -70,13 +70,13 @@ itertools = "0.11.0"
 # profiling
 puffin = "0.16.0"
 
-# logging
-logging_timer = "1.1.0"
 csv = "1.3.0"
 candle-core = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
 candle-transformers = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
 candle-nn = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
-tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.14.1", version = "0.14.1", default_features = false, features = ["onig"] }
+tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.14.1", version = "0.14.1", default_features = false, features = [
+    "onig",
+] }
 hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls", default_features = false, features = [
     "online",
 ] }
@@ -102,7 +102,16 @@ meili-snap = { path = "../meili-snap" }
 rand = { version = "0.8.5", features = ["small_rng"] }
 
 [features]
-all-tokenizations = ["charabia/chinese", "charabia/hebrew", "charabia/japanese", "charabia/thai", "charabia/korean", "charabia/greek", "charabia/khmer"]
+all-tokenizations = [
+    "charabia/chinese",
+    "charabia/hebrew",
+    "charabia/japanese",
+    "charabia/thai",
+    "charabia/korean",
+    "charabia/greek",
+    "charabia/khmer",
+    "charabia/vietnamese",
+]
 
 # Use POSIX semaphores instead of SysV semaphores in LMDB
 # For more information on this feature, see heed's Cargo.toml
@@ -130,5 +139,7 @@ greek = ["charabia/greek"]
 # allow khmer specialized tokenization
 khmer = ["charabia/khmer"]
 
+vietnamese = ["charabia/vietnamese"]
+
 # allow CUDA support, see <https://github.com/meilisearch/meilisearch/issues/4306>
 cuda = ["candle-core/cuda"]

milli/src/search/new/bucket_sort.rs

@@ -15,7 +15,7 @@ pub struct BucketSortOutput {
 
 // TODO: would probably be good to regroup some of these inside of a struct?
 #[allow(clippy::too_many_arguments)]
-#[logging_timer::time]
+#[tracing::instrument(level = "trace", skip_all, target = "search::bucket_sort")]
 pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
     ctx: &mut SearchContext<'ctx>,
     mut ranking_rules: Vec<BoxRankingRule<'ctx, Q>>,

milli/src/search/new/mod.rs

@@ -191,7 +191,7 @@ fn resolve_maximally_reduced_query_graph(
     Ok(docids)
 }
 
-#[logging_timer::time]
+#[tracing::instrument(level = "trace", skip_all, target = "search")]
 fn resolve_universe(
     ctx: &mut SearchContext,
     initial_universe: &RoaringBitmap,
@@ -557,7 +557,7 @@ pub fn execute_vector_search(
 }
 
 #[allow(clippy::too_many_arguments)]
-#[logging_timer::time]
+#[tracing::instrument(level = "trace", skip_all, target = "search")]
 pub fn execute_search(
     ctx: &mut SearchContext,
     query: Option<&str>,
@@ -577,6 +577,9 @@ pub fn execute_search(
 
     let mut located_query_terms = None;
     let query_terms = if let Some(query) = query {
+        let span = tracing::trace_span!(target: "search::tokens", "tokenizer_builder");
+        let entered = span.enter();
+
         // We make sure that the analyzer is aware of the stop words
         // this ensures that the query builder is able to properly remove them.
         let mut tokbuilder = TokenizerBuilder::new();
@@ -605,7 +608,12 @@ pub fn execute_search(
         }
 
         let tokenizer = tokbuilder.build();
+        drop(entered);
+
+        let span = tracing::trace_span!(target: "search::tokens", "tokenize");
+        let entered = span.enter();
         let tokens = tokenizer.tokenize(query);
+        drop(entered);
 
         let query_terms = located_query_terms_from_tokens(ctx, tokens, words_limit)?;
         if query_terms.is_empty() {

milli/src/search/new/query_term/compute_derivations.rs

@@ -6,9 +6,10 @@ use fst::automaton::Str;
 use fst::{Automaton, IntoStreamer, Streamer};
 use heed::types::DecodeIgnore;
 
-use super::*;
+use super::{OneTypoTerm, Phrase, QueryTerm, ZeroTypoTerm};
 use crate::search::fst_utils::{Complement, Intersection, StartsWith, Union};
-use crate::search::new::query_term::TwoTypoTerm;
+use crate::search::new::interner::{DedupInterner, Interned};
+use crate::search::new::query_term::{Lazy, TwoTypoTerm};
 use crate::search::new::{limits, SearchContext};
 use crate::search::{build_dfa, get_first};
 use crate::{Result, MAX_WORD_LENGTH};

milli/src/search/new/query_term/mod.rs

@@ -7,7 +7,6 @@ use std::collections::BTreeSet;
 use std::iter::FromIterator;
 use std::ops::RangeInclusive;
 
-use compute_derivations::partially_initialized_term_from_word;
 use either::Either;
 pub use ntypo_subset::NTypoTermSubset;
 pub use parse_query::{located_query_terms_from_tokens, make_ngram, number_of_typos_allowed};

milli/src/search/new/query_term/parse_query.rs

@@ -1,11 +1,15 @@
+use std::collections::BTreeSet;
+
 use charabia::normalizer::NormalizedTokenIter;
 use charabia::{SeparatorKind, TokenKind};
 
-use super::*;
+use super::compute_derivations::partially_initialized_term_from_word;
+use super::{LocatedQueryTerm, ZeroTypoTerm};
+use crate::search::new::query_term::{Lazy, Phrase, QueryTerm};
 use crate::{Result, SearchContext, MAX_WORD_LENGTH};
 
 /// Convert the tokenised search query into a list of located query terms.
-#[logging_timer::time]
+#[tracing::instrument(level = "trace", skip_all, target = "search::query")]
 pub fn located_query_terms_from_tokens(
     ctx: &mut SearchContext,
     query: NormalizedTokenIter,
@@ -225,7 +229,7 @@ pub fn make_ngram(
 }
 
 struct PhraseBuilder {
-    words: Vec<Option<Interned<String>>>,
+    words: Vec<Option<crate::search::new::Interned<String>>>,
     start: u16,
     end: u16,
 }

milli/src/update/facet/bulk.rs

@@ -1,7 +1,7 @@
 use std::fs::File;
 use std::io::BufReader;
 
-use grenad::CompressionType;
+use grenad::{CompressionType, Merger};
 use heed::types::Bytes;
 use heed::{BytesDecode, BytesEncode, Error, PutFlags, RoTxn, RwTxn};
 use roaring::RoaringBitmap;
@@ -14,6 +14,7 @@ use crate::heed_codec::facet::{
 use crate::heed_codec::BytesRefCodec;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd};
 use crate::update::index_documents::{create_writer, valid_lmdb_key, writer_into_reader};
+use crate::update::MergeFn;
 use crate::{CboRoaringBitmapCodec, CboRoaringBitmapLenCodec, FieldId, Index, Result};
 
 /// Algorithm to insert elememts into the `facet_id_(string/f64)_docids` databases
@@ -28,7 +29,7 @@ pub struct FacetsUpdateBulk<'i> {
     facet_type: FacetType,
     field_ids: Vec<FieldId>,
     // None if level 0 does not need to be updated
-    delta_data: Option<grenad::Reader<BufReader<File>>>,
+    delta_data: Option<Merger<BufReader<File>, MergeFn>>,
 }
 
 impl<'i> FacetsUpdateBulk<'i> {
@@ -36,7 +37,7 @@ impl<'i> FacetsUpdateBulk<'i> {
         index: &'i Index,
         field_ids: Vec<FieldId>,
         facet_type: FacetType,
-        delta_data: grenad::Reader<BufReader<File>>,
+        delta_data: Merger<BufReader<File>, MergeFn>,
         group_size: u8,
         min_level_size: u8,
     ) -> FacetsUpdateBulk<'i> {
@@ -65,7 +66,7 @@ impl<'i> FacetsUpdateBulk<'i> {
         }
     }
 
-    #[logging_timer::time("FacetsUpdateBulk::{}")]
+    #[tracing::instrument(level = "trace", skip_all, target = "indexing::facets::bulk")]
     pub fn execute(self, wtxn: &mut heed::RwTxn) -> Result<()> {
         let Self { index, field_ids, group_size, min_level_size, facet_type, delta_data } = self;
 
@@ -89,7 +90,7 @@ impl<'i> FacetsUpdateBulk<'i> {
 /// Implementation of `FacetsUpdateBulk` that is independent of milli's `Index` type
 pub(crate) struct FacetsUpdateBulkInner<R: std::io::Read + std::io::Seek> {
     pub db: heed::Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
-    pub delta_data: Option<grenad::Reader<R>>,
+    pub delta_data: Option<Merger<R, MergeFn>>,
     pub group_size: u8,
     pub min_level_size: u8,
 }
@@ -129,8 +130,8 @@ impl<R: std::io::Read + std::io::Seek> FacetsUpdateBulkInner<R> {
         if self.db.is_empty(wtxn)? {
             let mut buffer = Vec::new();
             let mut database = self.db.iter_mut(wtxn)?.remap_types::<Bytes, Bytes>();
-            let mut cursor = delta_data.into_cursor()?;
+            let mut iter = delta_data.into_stream_merger_iter()?;
-            while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
                 if !valid_lmdb_key(key) {
                     continue;
                 }
@@ -154,8 +155,8 @@ impl<R: std::io::Read + std::io::Seek> FacetsUpdateBulkInner<R> {
             let mut buffer = Vec::new();
             let database = self.db.remap_types::<Bytes, Bytes>();
 
-            let mut cursor = delta_data.into_cursor()?;
+            let mut iter = delta_data.into_stream_merger_iter()?;
-            while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
                 if !valid_lmdb_key(key) {
                     continue;
                 }

milli/src/update/facet/incremental.rs

@@ -1,6 +1,7 @@
 use std::fs::File;
 use std::io::BufReader;
 
+use grenad::Merger;
 use heed::types::{Bytes, DecodeIgnore};
 use heed::{BytesDecode, Error, RoTxn, RwTxn};
 use obkv::KvReader;
@@ -14,31 +15,56 @@ use crate::heed_codec::BytesRefCodec;
 use crate::search::facet::get_highest_level;
 use crate::update::del_add::DelAdd;
 use crate::update::index_documents::valid_lmdb_key;
+use crate::update::MergeFn;
 use crate::{CboRoaringBitmapCodec, Index, Result};
 
-enum InsertionResult {
+/// Enum used as a return value for the facet incremental indexing.
+///
+/// - `ModificationResult::InPlace` means that modifying the `facet_value` into the `level` did not have
+/// an effect on the number of keys in that level. Therefore, it did not increase the number of children
+/// of the parent node.
+///
+/// - `ModificationResult::Insert` means that modifying the `facet_value` into the `level` resulted
+/// in the addition of a new key in that level, and that therefore the number of children
+/// of the parent node should be incremented.
+///
+/// - `ModificationResult::Remove` means that modifying the `facet_value` into the `level` resulted in a change in the
+/// number of keys in the level. For example, removing a document id from the facet value `3` could
+/// cause it to have no corresponding document in level 0 anymore, and therefore the key was deleted
+/// entirely. In that case, `ModificationResult::Remove` is returned. The parent of the deleted key must
+/// then adjust its group size. If its group size falls to 0, then it will need to be deleted as well.
+///
+/// - `ModificationResult::Reduce/Expand` means that modifying the `facet_value` into the `level` resulted in a change in the
+/// bounds of the keys of the level. For example, removing a document id from the facet value
+/// `3` might have caused the facet value `3` to have no corresponding document in level 0. Therefore,
+/// in level 1, the key with the left bound `3` had to be changed to the next facet value (e.g. 4).
+/// In that case `ModificationResult::Reduce` is returned. The parent of the reduced key may need to adjust
+/// its left bound as well.
+///
+/// - `ModificationResult::Nothing` means that modifying the `facet_value` didn't have any impact into the `level`.
+/// This case is reachable when a document id is removed from a sub-level node but is still present in another one.
+/// For example, removing `2` from a document containing `2` and `3`, the document id will removed form the `level 0` but should remain in the group node [1..4] in `level 1`.
+enum ModificationResult {
     InPlace,
     Expand,
     Insert,
-}
-enum DeletionResult {
-    InPlace,
     Reduce { next: Option<Vec<u8>> },
     Remove { next: Option<Vec<u8>> },
+    Nothing,
 }
 
 /// Algorithm to incrementally insert and delete elememts into the
 /// `facet_id_(string/f64)_docids` databases.
 pub struct FacetsUpdateIncremental {
     inner: FacetsUpdateIncrementalInner,
-    delta_data: grenad::Reader<BufReader<File>>,
+    delta_data: Merger<BufReader<File>, MergeFn>,
 }
 
 impl FacetsUpdateIncremental {
     pub fn new(
         index: &Index,
         facet_type: FacetType,
-        delta_data: grenad::Reader<BufReader<File>>,
+        delta_data: Merger<BufReader<File>, MergeFn>,
         group_size: u8,
         min_level_size: u8,
         max_group_size: u8,
@@ -61,34 +87,59 @@ impl FacetsUpdateIncremental {
         }
     }
 
+    #[tracing::instrument(level = "trace", skip_all, target = "indexing::facets::incremental")]
     pub fn execute(self, wtxn: &mut RwTxn) -> crate::Result<()> {
-        let mut cursor = self.delta_data.into_cursor()?;
+        let mut current_field_id = None;
-        while let Some((key, value)) = cursor.move_on_next()? {
+        let mut facet_level_may_be_updated = false;
+        let mut iter = self.delta_data.into_stream_merger_iter()?;
+        while let Some((key, value)) = iter.next()? {
             if !valid_lmdb_key(key) {
                 continue;
             }
+
             let key = FacetGroupKeyCodec::<BytesRefCodec>::bytes_decode(key)
                 .map_err(heed::Error::Encoding)?;
-            let value = KvReader::new(value);
 
+            if facet_level_may_be_updated
+                && current_field_id.map_or(false, |fid| fid != key.field_id)
+            {
+                // Only add or remove a level after making all the field modifications.
+                self.inner.add_or_delete_level(wtxn, current_field_id.unwrap())?;
+                facet_level_may_be_updated = false;
+            }
+            current_field_id = Some(key.field_id);
+
+            let value = KvReader::new(value);
             let docids_to_delete = value
                 .get(DelAdd::Deletion)
                 .map(CboRoaringBitmapCodec::bytes_decode)
-                .map(|o| o.map_err(heed::Error::Encoding));
+                .map(|o| o.map_err(heed::Error::Encoding))
+                .transpose()?;
 
             let docids_to_add = value
                 .get(DelAdd::Addition)
                 .map(CboRoaringBitmapCodec::bytes_decode)
-                .map(|o| o.map_err(heed::Error::Encoding));
+                .map(|o| o.map_err(heed::Error::Encoding))
+                .transpose()?;
 
-            if let Some(docids_to_delete) = docids_to_delete {
+            let level_size_changed = self.inner.modify(
-                let docids_to_delete = docids_to_delete?;
+                wtxn,
-                self.inner.delete(wtxn, key.field_id, key.left_bound, &docids_to_delete)?;
+                key.field_id,
+                key.left_bound,
+                docids_to_add.as_ref(),
+                docids_to_delete.as_ref(),
+            )?;
+
+            if level_size_changed {
+                // if a node has been added or removed from the highest level,
+                // we may have to update the facet level.
+                facet_level_may_be_updated = true;
             }
+        }
 
-            if let Some(docids_to_add) = docids_to_add {
+        if let Some(field_id) = current_field_id {
-                let docids_to_add = docids_to_add?;
+            if facet_level_may_be_updated {
-                self.inner.insert(wtxn, key.field_id, key.left_bound, &docids_to_add)?;
+                self.inner.add_or_delete_level(wtxn, field_id)?;
             }
         }
 
@@ -162,138 +213,78 @@ impl FacetsUpdateIncrementalInner {
     ///
     /// ## Return
     /// See documentation of `insert_in_level`
-    fn insert_in_level_0(
+    fn modify_in_level_0(
         &self,
         txn: &mut RwTxn,
         field_id: u16,
         facet_value: &[u8],
-        docids: &RoaringBitmap,
+        add_docids: Option<&RoaringBitmap>,
-    ) -> Result<InsertionResult> {
+        del_docids: Option<&RoaringBitmap>,
+    ) -> Result<ModificationResult> {
         let key = FacetGroupKey { field_id, level: 0, left_bound: facet_value };
-        let value = FacetGroupValue { bitmap: docids.clone(), size: 1 };
 
-        let mut level0_prefix = vec![];
+        let old_value = self.db.get(txn, &key)?;
-        level0_prefix.extend_from_slice(&field_id.to_be_bytes());
+        match (old_value, add_docids, del_docids) {
-        level0_prefix.push(0);
+            // Addition + deletion on an existing value
-
+            (Some(FacetGroupValue { bitmap, .. }), Some(add_docids), Some(del_docids)) => {
-        let mut iter =
+                let value = FacetGroupValue { bitmap: (bitmap - del_docids) | add_docids, size: 1 };
-            self.db.remap_types::<Bytes, DecodeIgnore>().prefix_iter(txn, &level0_prefix)?;
+                self.db.put(txn, &key, &value)?;
-
+                Ok(ModificationResult::InPlace)
-        if iter.next().is_none() {
+            }
-            drop(iter);
+            // Addition on an existing value
-            self.db.put(txn, &key, &value)?;
+            (Some(FacetGroupValue { bitmap, .. }), Some(add_docids), None) => {
-            Ok(InsertionResult::Insert)
+                let value = FacetGroupValue { bitmap: bitmap | add_docids, size: 1 };
-        } else {
+                self.db.put(txn, &key, &value)?;
-            drop(iter);
+                Ok(ModificationResult::InPlace)
-            let old_value = self.db.get(txn, &key)?;
+            }
-            match old_value {
+            // Addition of a new value (ignore deletion)
-                Some(mut updated_value) => {
+            (None, Some(add_docids), _) => {
-                    // now merge the two
+                let value = FacetGroupValue { bitmap: add_docids.clone(), size: 1 };
-                    updated_value.bitmap |= value.bitmap;
+                self.db.put(txn, &key, &value)?;
-                    self.db.put(txn, &key, &updated_value)?;
+                Ok(ModificationResult::Insert)
-                    Ok(InsertionResult::InPlace)
+            }
-                }
+            // Deletion on an existing value, fully delete the key if the resulted value is empty.
-                None => {
+            (Some(FacetGroupValue { mut bitmap, .. }), None, Some(del_docids)) => {
+                bitmap -= del_docids;
+                if bitmap.is_empty() {
+                    // Full deletion
+                    let mut next_key = None;
+                    if let Some((next, _)) =
+                        self.db.remap_data_type::<DecodeIgnore>().get_greater_than(txn, &key)?
+                    {
+                        if next.field_id == field_id && next.level == 0 {
+                            next_key = Some(next.left_bound.to_vec());
+                        }
+                    }
+                    self.db.delete(txn, &key)?;
+                    Ok(ModificationResult::Remove { next: next_key })
+                } else {
+                    // Partial deletion
+                    let value = FacetGroupValue { bitmap, size: 1 };
                     self.db.put(txn, &key, &value)?;
-                    Ok(InsertionResult::Insert)
+                    Ok(ModificationResult::InPlace)
                 }
             }
+            // Otherwise do nothing (None + no addition + deletion == Some + no addition + no deletion == Nothing),
+            // may be unreachable at some point.
+            (None, None, _) | (Some(_), None, None) => Ok(ModificationResult::Nothing),
         }
     }
 
-    /// Insert the given facet value  and corresponding document ids in all the levels of the database up to the given `level`.
+    /// Split a level node into two balanced nodes.
-    /// This function works recursively.
     ///
-    /// ## Return
+    /// # Return
-    /// Returns the effect of adding the facet value to the database on the given `level`.
+    /// Returns `ModificationResult::Insert` if the split is successful.
-    ///
+    fn split_group(
-    /// - `InsertionResult::InPlace` means that inserting the `facet_value` into the `level` did not have
-    /// an effect on the number of keys in that level. Therefore, it did not increase the number of children
-    /// of the parent node.
-    ///
-    /// - `InsertionResult::Insert` means that inserting the `facet_value` into the `level` resulted
-    /// in the addition of a new key in that level, and that therefore the number of children
-    /// of the parent node should be incremented.
-    fn insert_in_level(
         &self,
         txn: &mut RwTxn,
         field_id: u16,
         level: u8,
-        facet_value: &[u8],
+        insertion_key: FacetGroupKey<Vec<u8>>,
-        docids: &RoaringBitmap,
+        insertion_value: FacetGroupValue,
-    ) -> Result<InsertionResult> {
+    ) -> Result<ModificationResult> {
-        if level == 0 {
+        let size_left = insertion_value.size / 2;
-            return self.insert_in_level_0(txn, field_id, facet_value, docids);
+        let size_right = insertion_value.size - size_left;
-        }
-
-        let max_group_size = self.max_group_size;
-
-        let result = self.insert_in_level(txn, field_id, level - 1, facet_value, docids)?;
-        // level below inserted an element
-
-        let (insertion_key, insertion_value) =
-            self.find_insertion_key_value(field_id, level, facet_value, txn)?;
-
-        match result {
-            // because we know that we inserted in place, the facet_value is not a new one
-            // thus it doesn't extend a group, and thus the insertion key computed above is
-            // still correct
-            InsertionResult::InPlace => {
-                let mut updated_value = insertion_value;
-                updated_value.bitmap |= docids;
-                self.db.put(txn, &insertion_key.as_ref(), &updated_value)?;
-
-                return Ok(InsertionResult::InPlace);
-            }
-            InsertionResult::Expand => {}
-            InsertionResult::Insert => {}
-        }
-
-        // Here we know that inserting the facet value in the level below resulted in the creation
-        // of a new key. Therefore, it may be the case that we need to modify the left bound of the
-        // insertion key (see documentation of `find_insertion_key_value` for an example of when that
-        // could happen).
-        let (insertion_key, insertion_key_was_modified) = {
-            let mut new_insertion_key = insertion_key.clone();
-            let mut key_should_be_modified = false;
-
-            if facet_value < insertion_key.left_bound.as_slice() {
-                new_insertion_key.left_bound = facet_value.to_vec();
-                key_should_be_modified = true;
-            }
-            if key_should_be_modified {
-                let is_deleted = self.db.delete(txn, &insertion_key.as_ref())?;
-                assert!(is_deleted);
-                self.db.put(txn, &new_insertion_key.as_ref(), &insertion_value)?;
-            }
-            (new_insertion_key, key_should_be_modified)
-        };
-        // Now we know that the insertion key contains the `facet_value`.
-
-        // We still need to update the insertion value by:
-        // 1. Incrementing the number of children (since the recursive call returned `InsertionResult::Insert`)
-        // 2. Merge the previous docids with the new one
-        let mut updated_value = insertion_value;
-
-        if matches!(result, InsertionResult::Insert) {
-            updated_value.size += 1;
-        }
-
-        if updated_value.size < max_group_size {
-            updated_value.bitmap |= docids;
-            self.db.put(txn, &insertion_key.as_ref(), &updated_value)?;
-            if insertion_key_was_modified {
-                return Ok(InsertionResult::Expand);
-            } else {
-                return Ok(InsertionResult::InPlace);
-            }
-        }
-
-        // We've increased the group size of the value and realised it has become greater than or equal to `max_group_size`
-        // Therefore it must be split into two nodes.
-
-        let size_left = updated_value.size / 2;
-        let size_right = updated_value.size - size_left;
 
         let level_below = level - 1;
 
@@ -347,34 +338,228 @@ impl FacetsUpdateIncrementalInner {
         self.db.put(txn, &group_left.0.as_ref(), &group_left.1)?;
         self.db.put(txn, &group_right.0.as_ref(), &group_right.1)?;
 
-        Ok(InsertionResult::Insert)
+        Ok(ModificationResult::Insert)
     }
 
-    /// Insert the given facet value and corresponding document ids in the database.
+    /// Remove the docids still present in the related sub-level nodes from the del_docids.
-    pub fn insert(
+    ///
+    /// This process is needed to avoid removing docids from a group node where the docid is present in several sub-nodes.
+    fn trim_del_docids<'a>(
+        &self,
+        txn: &mut RwTxn,
+        field_id: u16,
+        level: u8,
+        insertion_key: &FacetGroupKey<Vec<u8>>,
+        insertion_value_size: usize,
+        del_docids: &'a RoaringBitmap,
+    ) -> Result<std::borrow::Cow<'a, RoaringBitmap>> {
+        let level_below = level - 1;
+
+        let start_key = FacetGroupKey {
+            field_id,
+            level: level_below,
+            left_bound: insertion_key.left_bound.as_slice(),
+        };
+
+        let mut del_docids = std::borrow::Cow::Borrowed(del_docids);
+        let iter = self.db.range(txn, &(start_key..))?.take(insertion_value_size);
+        for next in iter {
+            let (_, value) = next?;
+            // if a sublevel bitmap as common docids with del_docids,
+            // then these docids shouldn't be removed and so, remove them from the deletion list.
+            if !value.bitmap.is_disjoint(&del_docids) {
+                *del_docids.to_mut() -= value.bitmap;
+            }
+        }
+
+        Ok(del_docids)
+    }
+
+    /// Modify the given facet value and corresponding document ids in all the levels of the database up to the given `level`.
+    /// This function works recursively.
+    ///
+    /// ## Return
+    /// Returns the effect of modifying the facet value to the database on the given `level`.
+    ///
+    fn modify_in_level(
+        &self,
+        txn: &mut RwTxn,
+        field_id: u16,
+        level: u8,
+        facet_value: &[u8],
+        add_docids: Option<&RoaringBitmap>,
+        del_docids: Option<&RoaringBitmap>,
+    ) -> Result<ModificationResult> {
+        if level == 0 {
+            return self.modify_in_level_0(txn, field_id, facet_value, add_docids, del_docids);
+        }
+
+        let result =
+            self.modify_in_level(txn, field_id, level - 1, facet_value, add_docids, del_docids)?;
+        // level below inserted an element
+
+        if let ModificationResult::Nothing = result {
+            // if the previous level has not been modified,
+            // early return ModificationResult::Nothing.
+            return Ok(ModificationResult::Nothing);
+        }
+
+        let (insertion_key, insertion_value) =
+            self.find_insertion_key_value(field_id, level, facet_value, txn)?;
+        let insertion_value_size = insertion_value.size as usize;
+
+        let mut insertion_value_was_modified = false;
+        let mut updated_value = insertion_value;
+
+        if let ModificationResult::Insert = result {
+            // if a key has been inserted in the sub-level raise the value size.
+            updated_value.size += 1;
+            insertion_value_was_modified = true;
+        } else if let ModificationResult::Remove { .. } = result {
+            if updated_value.size <= 1 {
+                // if the only remaining node is the one to delete,
+                // delete the key instead and early return.
+                let is_deleted = self.db.delete(txn, &insertion_key.as_ref())?;
+                assert!(is_deleted);
+                return Ok(result);
+            } else {
+                // Reduce the value size
+                updated_value.size -= 1;
+                insertion_value_was_modified = true;
+            }
+        }
+
+        let (insertion_key, insertion_key_modification) =
+            if let ModificationResult::InPlace = result {
+                (insertion_key, ModificationResult::InPlace)
+            } else {
+                // Inserting or deleting the facet value in the level below resulted in the creation
+                // of a new key. Therefore, it may be the case that we need to modify the left bound of the
+                // insertion key (see documentation of `find_insertion_key_value` for an example of when that
+                // could happen).
+                let mut new_insertion_key = insertion_key.clone();
+                let mut key_modification = ModificationResult::InPlace;
+
+                if let ModificationResult::Remove { next } | ModificationResult::Reduce { next } =
+                    result
+                {
+                    // if the deleted facet_value is the left_bound of the current node,
+                    // the left_bound should be updated reducing the current node.
+                    let reduced_range = facet_value == insertion_key.left_bound;
+                    if reduced_range {
+                        new_insertion_key.left_bound = next.clone().unwrap();
+                        key_modification = ModificationResult::Reduce { next };
+                    }
+                } else if facet_value < insertion_key.left_bound.as_slice() {
+                    // if the added facet_value is the under the left_bound of the current node,
+                    // the left_bound should be updated expanding the current node.
+                    new_insertion_key.left_bound = facet_value.to_vec();
+                    key_modification = ModificationResult::Expand;
+                }
+
+                if matches!(
+                    key_modification,
+                    ModificationResult::Expand | ModificationResult::Reduce { .. }
+                ) {
+                    // if the node should be updated, delete it, it will be recreated using a new key later.
+                    let is_deleted = self.db.delete(txn, &insertion_key.as_ref())?;
+                    assert!(is_deleted);
+                }
+                (new_insertion_key, key_modification)
+            };
+
+        if updated_value.size < self.max_group_size {
+            // If there are docids to delete, trim them avoiding unexpected removal.
+            if let Some(del_docids) = del_docids
+                .map(|ids| {
+                    self.trim_del_docids(
+                        txn,
+                        field_id,
+                        level,
+                        &insertion_key,
+                        insertion_value_size,
+                        ids,
+                    )
+                })
+                .transpose()?
+                .filter(|ids| !ids.is_empty())
+            {
+                updated_value.bitmap -= &*del_docids;
+                insertion_value_was_modified = true;
+            }
+
+            if let Some(add_docids) = add_docids {
+                updated_value.bitmap |= add_docids;
+                insertion_value_was_modified = true;
+            }
+
+            if insertion_value_was_modified
+                || matches!(
+                    insertion_key_modification,
+                    ModificationResult::Expand | ModificationResult::Reduce { .. }
+                )
+            {
+                // if any modification occured, insert it in the database.
+                self.db.put(txn, &insertion_key.as_ref(), &updated_value)?;
+                Ok(insertion_key_modification)
+            } else {
+                // this case is reachable when a docid is removed from a sub-level node but is still present in another one.
+                // For instance, a document containing 2 and 3, if 2 is removed, the docid should remain in the group node [1..4].
+                Ok(ModificationResult::Nothing)
+            }
+        } else {
+            // We've increased the group size of the value and realised it has become greater than or equal to `max_group_size`
+            // Therefore it must be split into two nodes.
+            self.split_group(txn, field_id, level, insertion_key, updated_value)
+        }
+    }
+
+    /// Modify the given facet value and corresponding document ids in the database.
+    /// If no more document ids correspond to the facet value, delete it completely.
+    ///
+    /// ## Return
+    /// Returns `true` if some tree-nodes of the highest level have been removed or added implying a potential
+    /// addition or deletion of a facet level.
+    /// Otherwise returns `false` if the tree-nodes have been modified in place.
+    pub fn modify(
         &self,
         txn: &mut RwTxn,
         field_id: u16,
         facet_value: &[u8],
-        docids: &RoaringBitmap,
+        add_docids: Option<&RoaringBitmap>,
-    ) -> Result<()> {
+        del_docids: Option<&RoaringBitmap>,
-        if docids.is_empty() {
+    ) -> Result<bool> {
-            return Ok(());
+        if add_docids.map_or(true, RoaringBitmap::is_empty)
+            && del_docids.map_or(true, RoaringBitmap::is_empty)
+        {
+            return Ok(false);
         }
-        let group_size = self.group_size;
 
         let highest_level = get_highest_level(txn, self.db, field_id)?;
 
-        let result = self.insert_in_level(txn, field_id, highest_level, facet_value, docids)?;
+        let result = self.modify_in_level(
+            txn,
+            field_id,
+            highest_level,
+            facet_value,
+            add_docids,
+            del_docids,
+        )?;
         match result {
-            InsertionResult::InPlace => return Ok(()),
+            ModificationResult::InPlace
-            InsertionResult::Expand => return Ok(()),
+            | ModificationResult::Expand
-            InsertionResult::Insert => {}
+            | ModificationResult::Nothing
+            | ModificationResult::Reduce { .. } => Ok(false),
+            ModificationResult::Insert | ModificationResult::Remove { .. } => Ok(true),
         }
+    }
 
-        // Here we check whether the highest level has exceeded `min_level_size` * `self.group_size`.
+    /// Check whether the highest level has exceeded `min_level_size` * `self.group_size`.
-        // If it has, we must build an addition level above it.
+    /// If it has, we must build an addition level above it.
-
+    /// Then check whether the highest level is under `min_level_size`.
+    /// If it has, we must remove the complete level.
+    pub(crate) fn add_or_delete_level(&self, txn: &mut RwTxn, field_id: u16) -> Result<()> {
+        let highest_level = get_highest_level(txn, self.db, field_id)?;
         let mut highest_level_prefix = vec![];
         highest_level_prefix.extend_from_slice(&field_id.to_be_bytes());
         highest_level_prefix.push(highest_level);
@@ -382,14 +567,48 @@ impl FacetsUpdateIncrementalInner {
         let size_highest_level =
             self.db.remap_types::<Bytes, Bytes>().prefix_iter(txn, &highest_level_prefix)?.count();
 
-        if size_highest_level < self.group_size as usize * self.min_level_size as usize {
+        if size_highest_level >= self.group_size as usize * self.min_level_size as usize {
-            return Ok(());
+            self.add_level(txn, field_id, highest_level, &highest_level_prefix, size_highest_level)
+        } else if size_highest_level < self.min_level_size as usize && highest_level != 0 {
+            self.delete_level(txn, &highest_level_prefix)
+        } else {
+            Ok(())
         }
+    }
 
+    /// Delete a level.
+    fn delete_level(&self, txn: &mut RwTxn, highest_level_prefix: &[u8]) -> Result<()> {
+        let mut to_delete = vec![];
+        let mut iter =
+            self.db.remap_types::<Bytes, Bytes>().prefix_iter(txn, highest_level_prefix)?;
+        for el in iter.by_ref() {
+            let (k, _) = el?;
+            to_delete.push(
+                FacetGroupKeyCodec::<BytesRefCodec>::bytes_decode(k)
+                    .map_err(Error::Encoding)?
+                    .into_owned(),
+            );
+        }
+        drop(iter);
+        for k in to_delete {
+            self.db.delete(txn, &k.as_ref())?;
+        }
+        Ok(())
+    }
+
+    /// Build an additional level for the field id.
+    fn add_level(
+        &self,
+        txn: &mut RwTxn,
+        field_id: u16,
+        highest_level: u8,
+        highest_level_prefix: &[u8],
+        size_highest_level: usize,
+    ) -> Result<()> {
         let mut groups_iter = self
             .db
             .remap_types::<Bytes, FacetGroupValueCodec>()
-            .prefix_iter(txn, &highest_level_prefix)?;
+            .prefix_iter(txn, highest_level_prefix)?;
 
         let nbr_new_groups = size_highest_level / self.group_size as usize;
         let nbr_leftover_elements = size_highest_level % self.group_size as usize;
@@ -398,7 +617,7 @@ impl FacetsUpdateIncrementalInner {
         for _ in 0..nbr_new_groups {
             let mut first_key = None;
             let mut values = RoaringBitmap::new();
-            for _ in 0..group_size {
+            for _ in 0..self.group_size {
                 let (key_bytes, value_i) = groups_iter.next().unwrap()?;
                 let key_i = FacetGroupKeyCodec::<BytesRefCodec>::bytes_decode(key_bytes)
                     .map_err(Error::Encoding)?;
@@ -413,7 +632,7 @@ impl FacetsUpdateIncrementalInner {
                 level: highest_level + 1,
                 left_bound: first_key.unwrap().left_bound,
             };
-            let value = FacetGroupValue { size: group_size, bitmap: values };
+            let value = FacetGroupValue { size: self.group_size, bitmap: values };
             to_add.push((key.into_owned(), value));
         }
         // now we add the rest of the level, in case its size is > group_size * min_level_size
@@ -448,173 +667,6 @@ impl FacetsUpdateIncrementalInner {
         }
         Ok(())
     }
-
-    /// Delete the given document id from the given facet value in the database, from level 0 to the
-    /// the given level.
-    ///
-    /// ## Return
-    /// Returns the effect of removing the document id from the database on the given `level`.
-    ///
-    /// - `DeletionResult::InPlace` means that deleting the document id did not have
-    /// an effect on the keys in that level.
-    ///
-    /// - `DeletionResult::Reduce` means that deleting the document id resulted in a change in the
-    /// number of keys in the level. For example, removing a document id from the facet value `3` could
-    /// cause it to have no corresponding document in level 0 anymore, and therefore the key was deleted
-    /// entirely. In that case, `DeletionResult::Remove` is returned. The parent of the deleted key must
-    /// then adjust its group size. If its group size falls to 0, then it will need to be deleted as well.
-    ///
-    /// - `DeletionResult::Reduce` means that deleting the document id resulted in a change in the
-    /// bounds of the keys of the level. For example, removing a document id from the facet value
-    /// `3` might have caused the facet value `3` to have no corresponding document in level 0. Therefore,
-    /// in level 1, the key with the left bound `3` had to be changed to the next facet value (e.g. 4).
-    /// In that case `DeletionResult::Reduce` is returned. The parent of the reduced key may need to adjust
-    /// its left bound as well.
-    fn delete_in_level(
-        &self,
-        txn: &mut RwTxn,
-        field_id: u16,
-        level: u8,
-        facet_value: &[u8],
-        docids: &RoaringBitmap,
-    ) -> Result<DeletionResult> {
-        if level == 0 {
-            return self.delete_in_level_0(txn, field_id, facet_value, docids);
-        }
-        let (deletion_key, mut bitmap) =
-            self.find_insertion_key_value(field_id, level, facet_value, txn)?;
-
-        let result = self.delete_in_level(txn, field_id, level - 1, facet_value, docids)?;
-
-        let mut decrease_size = false;
-        let next_key = match result {
-            DeletionResult::InPlace => {
-                bitmap.bitmap -= docids;
-                self.db.put(txn, &deletion_key.as_ref(), &bitmap)?;
-                return Ok(DeletionResult::InPlace);
-            }
-            DeletionResult::Reduce { next } => next,
-            DeletionResult::Remove { next } => {
-                decrease_size = true;
-                next
-            }
-        };
-        // If either DeletionResult::Reduce or DeletionResult::Remove was returned,
-        // then we may need to adjust the left_bound of the deletion key.
-
-        // If DeletionResult::Remove was returned, then we need to decrease the group
-        // size of the deletion key.
-        let mut updated_value = bitmap;
-        if decrease_size {
-            updated_value.size -= 1;
-        }
-
-        if updated_value.size == 0 {
-            self.db.delete(txn, &deletion_key.as_ref())?;
-            Ok(DeletionResult::Remove { next: next_key })
-        } else {
-            let mut updated_deletion_key = deletion_key.clone();
-            let reduced_range = facet_value == deletion_key.left_bound;
-            if reduced_range {
-                updated_deletion_key.left_bound = next_key.clone().unwrap();
-            }
-            updated_value.bitmap -= docids;
-            let _ = self.db.delete(txn, &deletion_key.as_ref())?;
-            self.db.put(txn, &updated_deletion_key.as_ref(), &updated_value)?;
-            if reduced_range {
-                Ok(DeletionResult::Reduce { next: next_key })
-            } else {
-                Ok(DeletionResult::InPlace)
-            }
-        }
-    }
-
-    fn delete_in_level_0(
-        &self,
-        txn: &mut RwTxn,
-        field_id: u16,
-        facet_value: &[u8],
-        docids: &RoaringBitmap,
-    ) -> Result<DeletionResult> {
-        let key = FacetGroupKey { field_id, level: 0, left_bound: facet_value };
-        let mut bitmap = self.db.get(txn, &key)?.unwrap().bitmap;
-        bitmap -= docids;
-
-        if bitmap.is_empty() {
-            let mut next_key = None;
-            if let Some((next, _)) =
-                self.db.remap_data_type::<DecodeIgnore>().get_greater_than(txn, &key)?
-            {
-                if next.field_id == field_id && next.level == 0 {
-                    next_key = Some(next.left_bound.to_vec());
-                }
-            }
-            self.db.delete(txn, &key)?;
-            Ok(DeletionResult::Remove { next: next_key })
-        } else {
-            self.db.put(txn, &key, &FacetGroupValue { size: 1, bitmap })?;
-            Ok(DeletionResult::InPlace)
-        }
-    }
-
-    pub fn delete(
-        &self,
-        txn: &mut RwTxn,
-        field_id: u16,
-        facet_value: &[u8],
-        docids: &RoaringBitmap,
-    ) -> Result<()> {
-        if self
-            .db
-            .remap_data_type::<DecodeIgnore>()
-            .get(txn, &FacetGroupKey { field_id, level: 0, left_bound: facet_value })?
-            .is_none()
-        {
-            return Ok(());
-        }
-        let highest_level = get_highest_level(txn, self.db, field_id)?;
-
-        let result = self.delete_in_level(txn, field_id, highest_level, facet_value, docids)?;
-        match result {
-            DeletionResult::InPlace => return Ok(()),
-            DeletionResult::Reduce { .. } => return Ok(()),
-            DeletionResult::Remove { .. } => {}
-        }
-
-        // if we either removed a key from the highest level, its size may have fallen
-        // below `min_level_size`, in which case we need to remove the entire level
-
-        let mut highest_level_prefix = vec![];
-        highest_level_prefix.extend_from_slice(&field_id.to_be_bytes());
-        highest_level_prefix.push(highest_level);
-
-        if highest_level == 0
-            || self
-                .db
-                .remap_types::<Bytes, Bytes>()
-                .prefix_iter(txn, &highest_level_prefix)?
-                .count()
-                >= self.min_level_size as usize
-        {
-            return Ok(());
-        }
-        let mut to_delete = vec![];
-        let mut iter =
-            self.db.remap_types::<Bytes, Bytes>().prefix_iter(txn, &highest_level_prefix)?;
-        for el in iter.by_ref() {
-            let (k, _) = el?;
-            to_delete.push(
-                FacetGroupKeyCodec::<BytesRefCodec>::bytes_decode(k)
-                    .map_err(Error::Encoding)?
-                    .into_owned(),
-            );
-        }
-        drop(iter);
-        for k in to_delete {
-            self.db.delete(txn, &k.as_ref())?;
-        }
-        Ok(())
-    }
 }
 
 impl<'a> FacetGroupKey<&'a [u8]> {

milli/src/update/facet/mod.rs

@@ -79,12 +79,9 @@ pub const FACET_MIN_LEVEL_SIZE: u8 = 5;
 use std::collections::BTreeSet;
 use std::fs::File;
 use std::io::BufReader;
-use std::iter::FromIterator;
 
-use charabia::normalizer::{Normalize, NormalizerOption};
+use grenad::Merger;
-use grenad::{CompressionType, SortAlgorithm};
+use heed::types::{Bytes, DecodeIgnore};
-use heed::types::{Bytes, DecodeIgnore, SerdeJson};
-use heed::BytesEncode;
 use time::OffsetDateTime;
 use tracing::debug;
 
@@ -93,9 +90,9 @@ use super::FacetsUpdateBulk;
 use crate::facet::FacetType;
 use crate::heed_codec::facet::{FacetGroupKey, FacetGroupKeyCodec, FacetGroupValueCodec};
 use crate::heed_codec::BytesRefCodec;
-use crate::update::index_documents::create_sorter;
+use crate::update::del_add::{DelAdd, KvReaderDelAdd};
-use crate::update::merge_btreeset_string;
+use crate::update::MergeFn;
-use crate::{BEU16StrCodec, Index, Result, MAX_FACET_VALUE_LENGTH};
+use crate::{try_split_array_at, FieldId, Index, Result};
 
 pub mod bulk;
 pub mod incremental;
@@ -108,16 +105,20 @@ pub struct FacetsUpdate<'i> {
     index: &'i Index,
     database: heed::Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
     facet_type: FacetType,
-    delta_data: grenad::Reader<BufReader<File>>,
+    delta_data: Merger<BufReader<File>, MergeFn>,
+    normalized_delta_data: Option<Merger<BufReader<File>, MergeFn>>,
     group_size: u8,
     max_group_size: u8,
     min_level_size: u8,
+    data_size: u64,
 }
 impl<'i> FacetsUpdate<'i> {
     pub fn new(
         index: &'i Index,
         facet_type: FacetType,
-        delta_data: grenad::Reader<BufReader<File>>,
+        delta_data: Merger<BufReader<File>, MergeFn>,
+        normalized_delta_data: Option<Merger<BufReader<File>, MergeFn>>,
+        data_size: u64,
     ) -> Self {
         let database = match facet_type {
             FacetType::String => {
@@ -135,18 +136,20 @@ impl<'i> FacetsUpdate<'i> {
             min_level_size: FACET_MIN_LEVEL_SIZE,
             facet_type,
             delta_data,
+            normalized_delta_data,
+            data_size,
         }
     }
 
     pub fn execute(self, wtxn: &mut heed::RwTxn) -> Result<()> {
-        if self.delta_data.is_empty() {
+        if self.data_size == 0 {
             return Ok(());
         }
         debug!("Computing and writing the facet values levels docids into LMDB on disk...");
         self.index.set_updated_at(wtxn, &OffsetDateTime::now_utc())?;
 
         // See self::comparison_bench::benchmark_facet_indexing
-        if self.delta_data.len() >= (self.database.len(wtxn)? / 50) {
+        if self.data_size >= (self.database.len(wtxn)? / 500) {
             let field_ids =
                 self.index.faceted_fields_ids(wtxn)?.iter().copied().collect::<Vec<_>>();
             let bulk_update = FacetsUpdateBulk::new(
@@ -170,96 +173,110 @@ impl<'i> FacetsUpdate<'i> {
             incremental_update.execute(wtxn)?;
         }
 
-        // We clear the list of normalized-for-search facets
+        match self.normalized_delta_data {
-        // and the previous FSTs to compute everything from scratch
+            Some(data) => index_facet_search(wtxn, data, self.index),
-        self.index.facet_id_normalized_string_strings.clear(wtxn)?;
+            None => Ok(()),
-        self.index.facet_id_string_fst.clear(wtxn)?;
-
-        // As we can't use the same write transaction to read and write in two different databases
-        // we must create a temporary sorter that we will write into LMDB afterward.
-        // As multiple unnormalized facet values can become the same normalized facet value
-        // we must merge them together.
-        let mut sorter = create_sorter(
-            SortAlgorithm::Unstable,
-            merge_btreeset_string,
-            CompressionType::None,
-            None,
-            None,
-            None,
-        );
-
-        // We iterate on the list of original, semi-normalized, facet values
-        // and normalize them for search, inserting them in LMDB in any given order.
-        let options = NormalizerOption { lossy: true, ..Default::default() };
-        let database = self.index.facet_id_string_docids.remap_data_type::<DecodeIgnore>();
-        for result in database.iter(wtxn)? {
-            let (facet_group_key, ()) = result?;
-            if let FacetGroupKey { field_id, level: 0, left_bound } = facet_group_key {
-                let mut normalized_facet = left_bound.normalize(&options);
-                let normalized_truncated_facet: String;
-                if normalized_facet.len() > MAX_FACET_VALUE_LENGTH {
-                    normalized_truncated_facet = normalized_facet
-                        .char_indices()
-                        .take_while(|(idx, _)| *idx < MAX_FACET_VALUE_LENGTH)
-                        .map(|(_, c)| c)
-                        .collect();
-                    normalized_facet = normalized_truncated_facet.into();
-                }
-                let set = BTreeSet::from_iter(std::iter::once(left_bound));
-                let key = (field_id, normalized_facet.as_ref());
-                let key = BEU16StrCodec::bytes_encode(&key).map_err(heed::Error::Encoding)?;
-                let val = SerdeJson::bytes_encode(&set).map_err(heed::Error::Encoding)?;
-                sorter.insert(key, val)?;
-            }
         }
-
-        // In this loop we don't need to take care of merging bitmaps
-        // as the grenad sorter already merged them for us.
-        let mut merger_iter = sorter.into_stream_merger_iter()?;
-        while let Some((key_bytes, btreeset_bytes)) = merger_iter.next()? {
-            self.index.facet_id_normalized_string_strings.remap_types::<Bytes, Bytes>().put(
-                wtxn,
-                key_bytes,
-                btreeset_bytes,
-            )?;
-        }
-
-        // We compute one FST by string facet
-        let mut text_fsts = vec![];
-        let mut current_fst: Option<(u16, fst::SetBuilder<Vec<u8>>)> = None;
-        let database =
-            self.index.facet_id_normalized_string_strings.remap_data_type::<DecodeIgnore>();
-        for result in database.iter(wtxn)? {
-            let ((field_id, normalized_facet), _) = result?;
-            current_fst = match current_fst.take() {
-                Some((fid, fst_builder)) if fid != field_id => {
-                    let fst = fst_builder.into_set();
-                    text_fsts.push((fid, fst));
-                    Some((field_id, fst::SetBuilder::memory()))
-                }
-                Some((field_id, fst_builder)) => Some((field_id, fst_builder)),
-                None => Some((field_id, fst::SetBuilder::memory())),
-            };
-
-            if let Some((_, fst_builder)) = current_fst.as_mut() {
-                fst_builder.insert(normalized_facet)?;
-            }
-        }
-
-        if let Some((field_id, fst_builder)) = current_fst {
-            let fst = fst_builder.into_set();
-            text_fsts.push((field_id, fst));
-        }
-
-        // We write those FSTs in LMDB now
-        for (field_id, fst) in text_fsts {
-            self.index.facet_id_string_fst.put(wtxn, &field_id, &fst)?;
-        }
-
-        Ok(())
     }
 }
 
+fn index_facet_search(
+    wtxn: &mut heed::RwTxn,
+    normalized_delta_data: Merger<BufReader<File>, MergeFn>,
+    index: &Index,
+) -> Result<()> {
+    let mut iter = normalized_delta_data.into_stream_merger_iter()?;
+    while let Some((key_bytes, delta_bytes)) = iter.next()? {
+        let deladd_reader = KvReaderDelAdd::new(delta_bytes);
+
+        let database_set = index
+            .facet_id_normalized_string_strings
+            .remap_key_type::<Bytes>()
+            .get(wtxn, key_bytes)?
+            .unwrap_or_default();
+
+        let add_set = deladd_reader
+            .get(DelAdd::Addition)
+            .and_then(|bytes| serde_json::from_slice::<BTreeSet<String>>(bytes).ok())
+            .unwrap_or_default();
+
+        let del_set = match deladd_reader
+            .get(DelAdd::Deletion)
+            .and_then(|bytes| serde_json::from_slice::<BTreeSet<String>>(bytes).ok())
+        {
+            Some(del_set) => {
+                let (field_id_bytes, _) = try_split_array_at(key_bytes).unwrap();
+                let field_id = FieldId::from_be_bytes(field_id_bytes);
+                let mut set = BTreeSet::new();
+                for facet in del_set {
+                    let key = FacetGroupKey { field_id, level: 0, left_bound: facet.as_str() };
+                    // Check if the referenced value doesn't exist anymore before deleting it.
+                    if index
+                        .facet_id_string_docids
+                        .remap_data_type::<DecodeIgnore>()
+                        .get(wtxn, &key)?
+                        .is_none()
+                    {
+                        set.insert(facet);
+                    }
+                }
+                set
+            }
+            None => BTreeSet::new(),
+        };
+
+        let set: BTreeSet<_> =
+            database_set.difference(&del_set).chain(add_set.iter()).cloned().collect();
+
+        if set.is_empty() {
+            index
+                .facet_id_normalized_string_strings
+                .remap_key_type::<Bytes>()
+                .delete(wtxn, key_bytes)?;
+        } else {
+            index
+                .facet_id_normalized_string_strings
+                .remap_key_type::<Bytes>()
+                .put(wtxn, key_bytes, &set)?;
+        }
+    }
+
+    // We clear the FST of normalized-for-search to compute everything from scratch.
+    index.facet_id_string_fst.clear(wtxn)?;
+    // We compute one FST by string facet
+    let mut text_fsts = vec![];
+    let mut current_fst: Option<(u16, fst::SetBuilder<Vec<u8>>)> = None;
+    let database = index.facet_id_normalized_string_strings.remap_data_type::<DecodeIgnore>();
+    for result in database.iter(wtxn)? {
+        let ((field_id, normalized_facet), _) = result?;
+        current_fst = match current_fst.take() {
+            Some((fid, fst_builder)) if fid != field_id => {
+                let fst = fst_builder.into_set();
+                text_fsts.push((fid, fst));
+                Some((field_id, fst::SetBuilder::memory()))
+            }
+            Some((field_id, fst_builder)) => Some((field_id, fst_builder)),
+            None => Some((field_id, fst::SetBuilder::memory())),
+        };
+
+        if let Some((_, fst_builder)) = current_fst.as_mut() {
+            fst_builder.insert(normalized_facet)?;
+        }
+    }
+
+    if let Some((field_id, fst_builder)) = current_fst {
+        let fst = fst_builder.into_set();
+        text_fsts.push((field_id, fst));
+    }
+
+    // We write those FSTs in LMDB now
+    for (field_id, fst) in text_fsts {
+        index.facet_id_string_fst.put(wtxn, &field_id, &fst)?;
+    }
+
+    Ok(())
+}
+
 #[cfg(test)]
 pub(crate) mod test_helpers {
     use std::cell::Cell;
@@ -268,6 +285,7 @@ pub(crate) mod test_helpers {
     use std::marker::PhantomData;
     use std::rc::Rc;
 
+    use grenad::MergerBuilder;
     use heed::types::Bytes;
     use heed::{BytesDecode, BytesEncode, Env, RoTxn, RwTxn};
     use roaring::RoaringBitmap;
@@ -280,7 +298,8 @@ pub(crate) mod test_helpers {
     use crate::search::facet::get_highest_level;
     use crate::snapshot_tests::display_bitmap;
     use crate::update::del_add::{DelAdd, KvWriterDelAdd};
-    use crate::update::FacetsUpdateIncrementalInner;
+    use crate::update::index_documents::merge_deladd_cbo_roaring_bitmaps;
+    use crate::update::{FacetsUpdateIncrementalInner, MergeFn};
     use crate::CboRoaringBitmapCodec;
 
     /// Utility function to generate a string whose position in a lexicographically
@@ -410,7 +429,8 @@ pub(crate) mod test_helpers {
                 max_group_size: self.max_group_size.get(),
             };
             let key_bytes = BoundCodec::bytes_encode(key).unwrap();
-            update.insert(wtxn, field_id, &key_bytes, docids).unwrap();
+            update.modify(wtxn, field_id, &key_bytes, Some(docids), None).unwrap();
+            update.add_or_delete_level(wtxn, field_id).unwrap();
         }
         pub fn delete_single_docid<'a>(
             &self,
@@ -436,7 +456,8 @@ pub(crate) mod test_helpers {
                 max_group_size: self.max_group_size.get(),
             };
             let key_bytes = BoundCodec::bytes_encode(key).unwrap();
-            update.delete(wtxn, field_id, &key_bytes, docids).unwrap();
+            update.modify(wtxn, field_id, &key_bytes, None, Some(docids)).unwrap();
+            update.add_or_delete_level(wtxn, field_id).unwrap();
         }
 
         pub fn bulk_insert<'a, 'b>(
@@ -463,10 +484,13 @@ pub(crate) mod test_helpers {
             }
             writer.finish().unwrap();
             let reader = grenad::Reader::new(std::io::Cursor::new(new_data)).unwrap();
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            builder.push(reader.into_cursor().unwrap());
+            let merger = builder.build();
 
             let update = FacetsUpdateBulkInner {
                 db: self.content,
-                delta_data: Some(reader),
+                delta_data: Some(merger),
                 group_size: self.group_size.get(),
                 min_level_size: self.min_level_size.get(),
             };

milli/src/update/index_documents/extract/extract_docid_word_positions.rs

@@ -26,7 +26,7 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
     obkv_documents: grenad::Reader<R>,
     indexer: GrenadParameters,
     searchable_fields: &Option<HashSet<FieldId>>,
-    stop_words: Option<&fst::Set<&[u8]>>,
+    stop_words: Option<&fst::Set<Vec<u8>>>,
     allowed_separators: Option<&[&str]>,
     dictionary: Option<&[&str]>,
     max_positions_per_attributes: Option<u32>,
@@ -181,11 +181,11 @@ fn searchable_fields_changed(
 
 /// Factorize tokenizer building.
 fn tokenizer_builder<'a>(
-    stop_words: Option<&'a fst::Set<&[u8]>>,
+    stop_words: Option<&'a fst::Set<Vec<u8>>>,
     allowed_separators: Option<&'a [&str]>,
     dictionary: Option<&'a [&str]>,
     script_language: Option<&'a HashMap<Script, Vec<Language>>>,
-) -> TokenizerBuilder<'a, &'a [u8]> {
+) -> TokenizerBuilder<'a, Vec<u8>> {
     let mut tokenizer_builder = TokenizerBuilder::new();
     if let Some(stop_words) = stop_words {
         tokenizer_builder.stop_words(stop_words);
@@ -211,7 +211,7 @@ fn lang_safe_tokens_from_document<'a>(
     obkv: &KvReader<FieldId>,
     searchable_fields: &Option<HashSet<FieldId>>,
     tokenizer: &Tokenizer,
-    stop_words: Option<&fst::Set<&[u8]>>,
+    stop_words: Option<&fst::Set<Vec<u8>>>,
     allowed_separators: Option<&[&str]>,
     dictionary: Option<&[&str]>,
     max_positions_per_attributes: u32,

milli/src/update/index_documents/extract/extract_facet_string_docids.rs

@@ -1,15 +1,21 @@
+use std::collections::BTreeSet;
 use std::fs::File;
 use std::io::BufReader;
+use std::iter::FromIterator;
 use std::{io, str};
 
+use charabia::normalizer::{Normalize, NormalizerOption};
+use heed::types::SerdeJson;
 use heed::BytesEncode;
 
 use super::helpers::{create_sorter, sorter_into_reader, try_split_array_at, GrenadParameters};
 use crate::heed_codec::facet::{FacetGroupKey, FacetGroupKeyCodec};
-use crate::heed_codec::StrRefCodec;
+use crate::heed_codec::{BEU16StrCodec, StrRefCodec};
-use crate::update::del_add::{KvReaderDelAdd, KvWriterDelAdd};
+use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
-use crate::update::index_documents::helpers::merge_deladd_cbo_roaring_bitmaps;
+use crate::update::index_documents::helpers::{
-use crate::{FieldId, Result};
+    merge_deladd_btreeset_string, merge_deladd_cbo_roaring_bitmaps,
+};
+use crate::{FieldId, Result, MAX_FACET_VALUE_LENGTH};
 
 /// Extracts the facet string and the documents ids where this facet string appear.
 ///
@@ -19,10 +25,11 @@ use crate::{FieldId, Result};
 pub fn extract_facet_string_docids<R: io::Read + io::Seek>(
     docid_fid_facet_string: grenad::Reader<R>,
     indexer: GrenadParameters,
-) -> Result<grenad::Reader<BufReader<File>>> {
+) -> Result<(grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>)> {
     puffin::profile_function!();
 
     let max_memory = indexer.max_memory_by_thread();
+    let options = NormalizerOption { lossy: true, ..Default::default() };
 
     let mut facet_string_docids_sorter = create_sorter(
         grenad::SortAlgorithm::Stable,
@@ -30,12 +37,30 @@ pub fn extract_facet_string_docids<R: io::Read + io::Seek>(
         indexer.chunk_compression_type,
         indexer.chunk_compression_level,
         indexer.max_nb_chunks,
-        max_memory,
+        max_memory.map(|m| m / 2),
+    );
+
+    let mut normalized_facet_string_docids_sorter = create_sorter(
+        grenad::SortAlgorithm::Stable,
+        merge_deladd_btreeset_string,
+        indexer.chunk_compression_type,
+        indexer.chunk_compression_level,
+        indexer.max_nb_chunks,
+        max_memory.map(|m| m / 2),
     );
 
     let mut buffer = Vec::new();
     let mut cursor = docid_fid_facet_string.into_cursor()?;
     while let Some((key, deladd_original_value_bytes)) = cursor.move_on_next()? {
+        let deladd_reader = KvReaderDelAdd::new(deladd_original_value_bytes);
+
+        // nothing to do if we delete and re-add the value.
+        if deladd_reader.get(DelAdd::Deletion).is_some()
+            && deladd_reader.get(DelAdd::Addition).is_some()
+        {
+            continue;
+        }
+
         let (field_id_bytes, bytes) = try_split_array_at(key).unwrap();
         let field_id = FieldId::from_be_bytes(field_id_bytes);
 
@@ -44,17 +69,46 @@ pub fn extract_facet_string_docids<R: io::Read + io::Seek>(
         let document_id = u32::from_be_bytes(document_id_bytes);
 
         let normalized_value = str::from_utf8(normalized_value_bytes)?;
+
+        // Facet search normalization
+        {
+            let mut hyper_normalized_value = normalized_value.normalize(&options);
+            let normalized_truncated_facet: String;
+            if hyper_normalized_value.len() > MAX_FACET_VALUE_LENGTH {
+                normalized_truncated_facet = hyper_normalized_value
+                    .char_indices()
+                    .take_while(|(idx, _)| *idx < MAX_FACET_VALUE_LENGTH)
+                    .map(|(_, c)| c)
+                    .collect();
+                hyper_normalized_value = normalized_truncated_facet.into();
+            }
+            let set = BTreeSet::from_iter(std::iter::once(normalized_value));
+
+            buffer.clear();
+            let mut obkv = KvWriterDelAdd::new(&mut buffer);
+            for (deladd_key, _) in deladd_reader.iter() {
+                let val = SerdeJson::bytes_encode(&set).map_err(heed::Error::Encoding)?;
+                obkv.insert(deladd_key, val)?;
+            }
+            obkv.finish()?;
+
+            let key = (field_id, hyper_normalized_value.as_ref());
+            let key_bytes = BEU16StrCodec::bytes_encode(&key).map_err(heed::Error::Encoding)?;
+            normalized_facet_string_docids_sorter.insert(key_bytes, &buffer)?;
+        }
+
         let key = FacetGroupKey { field_id, level: 0, left_bound: normalized_value };
         let key_bytes = FacetGroupKeyCodec::<StrRefCodec>::bytes_encode(&key).unwrap();
 
         buffer.clear();
         let mut obkv = KvWriterDelAdd::new(&mut buffer);
-        for (deladd_key, _) in KvReaderDelAdd::new(deladd_original_value_bytes).iter() {
+        for (deladd_key, _) in deladd_reader.iter() {
             obkv.insert(deladd_key, document_id.to_ne_bytes())?;
         }
         obkv.finish()?;
         facet_string_docids_sorter.insert(&key_bytes, &buffer)?;
     }
 
-    sorter_into_reader(facet_string_docids_sorter, indexer)
+    let normalized = sorter_into_reader(normalized_facet_string_docids_sorter, indexer)?;
+    sorter_into_reader(facet_string_docids_sorter, indexer).map(|s| (s, normalized))
 }

milli/src/update/index_documents/extract/extract_vector_points.rs

@@ -257,6 +257,7 @@ fn push_vectors_diff(
     key_buffer: &mut Vec<u8>,
     delta: VectorStateDelta,
 ) -> Result<()> {
+    puffin::profile_function!();
     let (must_remove, prompt, (mut del_vectors, mut add_vectors)) = delta.into_values();
     if must_remove {
         key_buffer.truncate(TRUNCATE_SIZE);
@@ -332,13 +333,14 @@ fn extract_vectors(
     }
 }
 
-#[logging_timer::time]
+#[tracing::instrument(level = "trace", skip_all, target = "indexing::extract")]
 pub fn extract_embeddings<R: io::Read + io::Seek>(
     // docid, prompt
     prompt_reader: grenad::Reader<R>,
     indexer: GrenadParameters,
     embedder: Arc<Embedder>,
 ) -> Result<grenad::Reader<BufReader<File>>> {
+    puffin::profile_function!();
     let n_chunks = embedder.chunk_count_hint(); // chunk level parallelism
     let n_vectors_per_chunk = embedder.prompt_count_in_chunk_hint(); // number of vectors in a single chunk
 

milli/src/update/index_documents/extract/mod.rs

@@ -15,7 +15,6 @@ use std::io::BufReader;
 
 use crossbeam_channel::Sender;
 use rayon::prelude::*;
-use tracing::debug;
 
 use self::extract_docid_word_positions::extract_docid_word_positions;
 use self::extract_facet_number_docids::extract_facet_number_docids;
@@ -29,10 +28,7 @@ use self::extract_vector_points::{
 use self::extract_word_docids::extract_word_docids;
 use self::extract_word_pair_proximity_docids::extract_word_pair_proximity_docids;
 use self::extract_word_position_docids::extract_word_position_docids;
-use super::helpers::{
+use super::helpers::{as_cloneable_grenad, CursorClonableMmap, GrenadParameters};
-    as_cloneable_grenad, merge_deladd_cbo_roaring_bitmaps, CursorClonableMmap, GrenadParameters,
-    MergeFn, MergeableReader,
-};
 use super::{helpers, TypedChunk};
 use crate::proximity::ProximityPrecision;
 use crate::vector::EmbeddingConfigs;
@@ -52,7 +48,7 @@ pub(crate) fn data_from_obkv_documents(
     primary_key_id: FieldId,
     geo_fields_ids: Option<(FieldId, FieldId)>,
     field_id_map: FieldsIdsMap,
-    stop_words: Option<fst::Set<&[u8]>>,
+    stop_words: Option<fst::Set<Vec<u8>>>,
     allowed_separators: Option<&[&str]>,
     dictionary: Option<&[&str]>,
     max_positions_per_attributes: Option<u32>,
@@ -62,227 +58,170 @@ pub(crate) fn data_from_obkv_documents(
 ) -> Result<()> {
     puffin::profile_function!();
 
-    original_obkv_chunks
+    let (original_pipeline_result, flattened_pipeline_result): (Result<_>, Result<_>) = rayon::join(
-        .par_bridge()
+        || {
-        .map(|original_documents_chunk| {
+            original_obkv_chunks
-            send_original_documents_data(
+                .par_bridge()
-                original_documents_chunk,
+                .map(|original_documents_chunk| {
-                indexer,
+                    send_original_documents_data(
-                lmdb_writer_sx.clone(),
+                        original_documents_chunk,
-                field_id_map.clone(),
+                        indexer,
-                embedders.clone(),
+                        lmdb_writer_sx.clone(),
-            )
+                        field_id_map.clone(),
-        })
+                        embedders.clone(),
-        .collect::<Result<()>>()?;
+                    )
-
+                })
-    #[allow(clippy::type_complexity)]
+                .collect::<Result<()>>()
-    let result: Result<(Vec<_>, (Vec<_>, (Vec<_>, (Vec<_>, (Vec<_>, Vec<_>)))))> =
+        },
-        flattened_obkv_chunks
+        || {
-            .par_bridge()
+            flattened_obkv_chunks
-            .map(|flattened_obkv_chunks| {
+                .par_bridge()
-                send_and_extract_flattened_documents_data(
+                .map(|flattened_obkv_chunks| {
-                    flattened_obkv_chunks,
+                    send_and_extract_flattened_documents_data(
-                    indexer,
+                        flattened_obkv_chunks,
-                    lmdb_writer_sx.clone(),
+                        indexer,
-                    &searchable_fields,
+                        lmdb_writer_sx.clone(),
-                    &faceted_fields,
+                        &searchable_fields,
-                    primary_key_id,
+                        &faceted_fields,
-                    geo_fields_ids,
+                        primary_key_id,
-                    &stop_words,
+                        geo_fields_ids,
-                    &allowed_separators,
+                        &stop_words,
-                    &dictionary,
+                        &allowed_separators,
-                    max_positions_per_attributes,
+                        &dictionary,
-                )
+                        max_positions_per_attributes,
-            })
+                    )
-            .collect();
+                })
-
+                .map(|result| {
-    let (
+                    if let Ok((
-        docid_word_positions_chunks,
+                        ref docid_word_positions_chunk,
-        (
+                        (ref fid_docid_facet_numbers_chunk, ref fid_docid_facet_strings_chunk),
-            fid_docid_facet_numbers_chunks,
+                    )) = result
-            (
+                    {
-                fid_docid_facet_strings_chunks,
+                        run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>(
-                (
+                            docid_word_positions_chunk.clone(),
-                    facet_is_null_docids_chunks,
+                            indexer,
-                    (facet_is_empty_docids_chunks, facet_exists_docids_chunks),
+                            lmdb_writer_sx.clone(),
-                ),
+                            extract_fid_word_count_docids,
-            ),
+                            TypedChunk::FieldIdWordCountDocids,
-        ),
+                            "field-id-wordcount-docids",
-    ) = result?;
+                        );
-
+
-    // merge facet_exists_docids and send them as a typed chunk
+                        let exact_attributes = exact_attributes.clone();
-    {
+                        run_extraction_task::<
-        let lmdb_writer_sx = lmdb_writer_sx.clone();
+                            _,
-        rayon::spawn(move || {
+                            _,
-            debug!(database = "facet-id-exists-docids", "merge");
+                            (
-            match facet_exists_docids_chunks.merge(merge_deladd_cbo_roaring_bitmaps, &indexer) {
+                                grenad::Reader<BufReader<File>>,
-                Ok(reader) => {
+                                grenad::Reader<BufReader<File>>,
-                    let _ = lmdb_writer_sx.send(Ok(TypedChunk::FieldIdFacetExistsDocids(reader)));
+                                grenad::Reader<BufReader<File>>,
-                }
+                            ),
-                Err(e) => {
+                        >(
-                    let _ = lmdb_writer_sx.send(Err(e));
+                            docid_word_positions_chunk.clone(),
-                }
+                            indexer,
-            }
+                            lmdb_writer_sx.clone(),
-        });
+                            move |doc_word_pos, indexer| {
-    }
+                                extract_word_docids(doc_word_pos, indexer, &exact_attributes)
-
+                            },
-    // merge facet_is_null_docids and send them as a typed chunk
+                            |(
-    {
+                                word_docids_reader,
-        let lmdb_writer_sx = lmdb_writer_sx.clone();
+                                exact_word_docids_reader,
-        rayon::spawn(move || {
+                                word_fid_docids_reader,
-            debug!(database = "facet-id-is-null-docids", "merge");
+                            )| {
-            match facet_is_null_docids_chunks.merge(merge_deladd_cbo_roaring_bitmaps, &indexer) {
+                                TypedChunk::WordDocids {
-                Ok(reader) => {
+                                    word_docids_reader,
-                    let _ = lmdb_writer_sx.send(Ok(TypedChunk::FieldIdFacetIsNullDocids(reader)));
+                                    exact_word_docids_reader,
-                }
+                                    word_fid_docids_reader,
-                Err(e) => {
+                                }
-                    let _ = lmdb_writer_sx.send(Err(e));
+                            },
-                }
+                            "word-docids",
-            }
+                        );
-        });
+
-    }
+                        run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>(
-
+                            docid_word_positions_chunk.clone(),
-    // merge facet_is_empty_docids and send them as a typed chunk
+                            indexer,
-    {
+                            lmdb_writer_sx.clone(),
-        let lmdb_writer_sx = lmdb_writer_sx.clone();
+                            extract_word_position_docids,
-        rayon::spawn(move || {
+                            TypedChunk::WordPositionDocids,
-            debug!(database = "facet-id-is-empty-docids", "merge");
+                            "word-position-docids",
-            match facet_is_empty_docids_chunks.merge(merge_deladd_cbo_roaring_bitmaps, &indexer) {
+                        );
-                Ok(reader) => {
+
-                    let _ = lmdb_writer_sx.send(Ok(TypedChunk::FieldIdFacetIsEmptyDocids(reader)));
+                        run_extraction_task::<
-                }
+                            _,
-                Err(e) => {
+                            _,
-                    let _ = lmdb_writer_sx.send(Err(e));
+                            (grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>),
-                }
+                        >(
-            }
+                            fid_docid_facet_strings_chunk.clone(),
-        });
+                            indexer,
-    }
+                            lmdb_writer_sx.clone(),
-
+                            extract_facet_string_docids,
-    if proximity_precision == ProximityPrecision::ByWord {
+                            TypedChunk::FieldIdFacetStringDocids,
-        spawn_extraction_task::<_, _, Vec<grenad::Reader<BufReader<File>>>>(
+                            "field-id-facet-string-docids",
-            docid_word_positions_chunks.clone(),
+                        );
-            indexer,
+
-            lmdb_writer_sx.clone(),
+                        run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>(
-            extract_word_pair_proximity_docids,
+                            fid_docid_facet_numbers_chunk.clone(),
-            merge_deladd_cbo_roaring_bitmaps,
+                            indexer,
-            TypedChunk::WordPairProximityDocids,
+                            lmdb_writer_sx.clone(),
-            "word-pair-proximity-docids",
+                            extract_facet_number_docids,
-        );
+                            TypedChunk::FieldIdFacetNumberDocids,
-    }
+                            "field-id-facet-number-docids",
-
+                        );
-    spawn_extraction_task::<_, _, Vec<grenad::Reader<BufReader<File>>>>(
+
-        docid_word_positions_chunks.clone(),
+                        if proximity_precision == ProximityPrecision::ByWord {
-        indexer,
+                            run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>(
-        lmdb_writer_sx.clone(),
+                                docid_word_positions_chunk.clone(),
-        extract_fid_word_count_docids,
+                                indexer,
-        merge_deladd_cbo_roaring_bitmaps,
+                                lmdb_writer_sx.clone(),
-        TypedChunk::FieldIdWordCountDocids,
+                                extract_word_pair_proximity_docids,
-        "field-id-wordcount-docids",
+                                TypedChunk::WordPairProximityDocids,
-    );
+                                "word-pair-proximity-docids",
-
+                            );
-    spawn_extraction_task::<
+                        }
-        _,
+                    }
-        _,
+
-        Vec<(
+                    Ok(())
-            grenad::Reader<BufReader<File>>,
+                })
-            grenad::Reader<BufReader<File>>,
+                .collect::<Result<()>>()
-            grenad::Reader<BufReader<File>>,
-        )>,
-    >(
-        docid_word_positions_chunks.clone(),
-        indexer,
-        lmdb_writer_sx.clone(),
-        move |doc_word_pos, indexer| extract_word_docids(doc_word_pos, indexer, &exact_attributes),
-        merge_deladd_cbo_roaring_bitmaps,
-        |(word_docids_reader, exact_word_docids_reader, word_fid_docids_reader)| {
-            TypedChunk::WordDocids {
-                word_docids_reader,
-                exact_word_docids_reader,
-                word_fid_docids_reader,
-            }
         },
-        "word-docids",
     );
 
-    spawn_extraction_task::<_, _, Vec<grenad::Reader<BufReader<File>>>>(
+    original_pipeline_result.and(flattened_pipeline_result)
-        docid_word_positions_chunks.clone(),
-        indexer,
-        lmdb_writer_sx.clone(),
-        extract_word_position_docids,
-        merge_deladd_cbo_roaring_bitmaps,
-        TypedChunk::WordPositionDocids,
-        "word-position-docids",
-    );
-
-    spawn_extraction_task::<_, _, Vec<grenad::Reader<BufReader<File>>>>(
-        fid_docid_facet_strings_chunks,
-        indexer,
-        lmdb_writer_sx.clone(),
-        extract_facet_string_docids,
-        merge_deladd_cbo_roaring_bitmaps,
-        TypedChunk::FieldIdFacetStringDocids,
-        "field-id-facet-string-docids",
-    );
-
-    spawn_extraction_task::<_, _, Vec<grenad::Reader<BufReader<File>>>>(
-        fid_docid_facet_numbers_chunks,
-        indexer,
-        lmdb_writer_sx,
-        extract_facet_number_docids,
-        merge_deladd_cbo_roaring_bitmaps,
-        TypedChunk::FieldIdFacetNumberDocids,
-        "field-id-facet-number-docids",
-    );
-
-    Ok(())
 }
 
 /// Spawn a new task to extract data for a specific DB using extract_fn.
 /// Generated grenad chunks are merged using the merge_fn.
 /// The result of merged chunks is serialized as TypedChunk using the serialize_fn
 /// and sent into lmdb_writer_sx.
-fn spawn_extraction_task<FE, FS, M>(
+fn run_extraction_task<FE, FS, M>(
-    chunks: Vec<grenad::Reader<CursorClonableMmap>>,
+    chunk: grenad::Reader<CursorClonableMmap>,
     indexer: GrenadParameters,
     lmdb_writer_sx: Sender<Result<TypedChunk>>,
     extract_fn: FE,
-    merge_fn: MergeFn,
     serialize_fn: FS,
     name: &'static str,
 ) where
-    FE: Fn(grenad::Reader<CursorClonableMmap>, GrenadParameters) -> Result<M::Output>
+    FE: Fn(grenad::Reader<CursorClonableMmap>, GrenadParameters) -> Result<M>
         + Sync
         + Send
         + 'static,
-    FS: Fn(M::Output) -> TypedChunk + Sync + Send + 'static,
+    FS: Fn(M) -> TypedChunk + Sync + Send + 'static,
-    M: MergeableReader + FromParallelIterator<M::Output> + Send + 'static,
+    M: Send,
-    M::Output: Send,
 {
     let current_span = tracing::Span::current();
 
     rayon::spawn(move || {
-        let child_span =
+        let child_span = tracing::trace_span!(target: "indexing::extract::details", parent: &current_span, "extract_multiple_chunks");
-            tracing::trace_span!(target: "", parent: &current_span, "extract_multiple_chunks");
         let _entered = child_span.enter();
-        puffin::profile_scope!("extract_multiple_chunksdexing::details, ", name);
+        puffin::profile_scope!("extract_multiple_chunks", name);
-        let chunks: Result<M> =
+        match extract_fn(chunk, indexer) {
-            chunks.into_par_iter().map(|chunk| extract_fn(chunk, indexer)).collect();
+            Ok(chunk) => {
-        let current_span = tracing::Span::current();
+                let _ = lmdb_writer_sx.send(Ok(serialize_fn(chunk)));
-
-        rayon::spawn(move || match chunks {
-            Ok(chunks) => {
-                let child_span = tracing::trace_span!(target: "", parent: &current_span, "merge_multiple_chunks");
-                let _entered = child_span.enter();
-                debug!(database = name, "merge");
-                puffin::profile_scope!("merge_multiple_chunks", name);
-                let reader = chunks.merge(merge_fn, &indexer);
-                let _ = lmdb_writer_sx.send(reader.map(serialize_fn));
             }
             Err(e) => {
                 let _ = lmdb_writer_sx.send(Err(e));
             }
-        })
+        }
-    });
+    })
 }
 
 /// Extract chunked data and send it into lmdb_writer_sx sender:
@@ -340,7 +279,7 @@ fn send_original_documents_data(
     });
 
     // TODO: create a custom internal error
-    lmdb_writer_sx.send(Ok(TypedChunk::Documents(original_documents_chunk))).unwrap();
+    let _ = lmdb_writer_sx.send(Ok(TypedChunk::Documents(original_documents_chunk)));
     Ok(())
 }
 
@@ -360,22 +299,13 @@ fn send_and_extract_flattened_documents_data(
     faceted_fields: &HashSet<FieldId>,
     primary_key_id: FieldId,
     geo_fields_ids: Option<(FieldId, FieldId)>,
-    stop_words: &Option<fst::Set<&[u8]>>,
+    stop_words: &Option<fst::Set<Vec<u8>>>,
     allowed_separators: &Option<&[&str]>,
     dictionary: &Option<&[&str]>,
     max_positions_per_attributes: Option<u32>,
 ) -> Result<(
     grenad::Reader<CursorClonableMmap>,
-    (
+    (grenad::Reader<CursorClonableMmap>, grenad::Reader<CursorClonableMmap>),
-        grenad::Reader<CursorClonableMmap>,
-        (
-            grenad::Reader<CursorClonableMmap>,
-            (
-                grenad::Reader<BufReader<File>>,
-                (grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>),
-            ),
-        ),
-    ),
 )> {
     let flattened_documents_chunk =
         flattened_documents_chunk.and_then(|c| unsafe { as_cloneable_grenad(&c) })?;
@@ -446,16 +376,17 @@ fn send_and_extract_flattened_documents_data(
                     fid_docid_facet_strings_chunk.clone(),
                 )));
 
-                Ok((
+                let _ = lmdb_writer_sx
-                    fid_docid_facet_numbers_chunk,
+                    .send(Ok(TypedChunk::FieldIdFacetIsNullDocids(fid_facet_is_null_docids_chunk)));
-                    (
+
-                        fid_docid_facet_strings_chunk,
+                let _ = lmdb_writer_sx.send(Ok(TypedChunk::FieldIdFacetIsEmptyDocids(
-                        (
+                    fid_facet_is_empty_docids_chunk,
-                            fid_facet_is_null_docids_chunk,
+                )));
-                            (fid_facet_is_empty_docids_chunk, fid_facet_exists_docids_chunk),
+
-                        ),
+                let _ = lmdb_writer_sx
-                    ),
+                    .send(Ok(TypedChunk::FieldIdFacetExistsDocids(fid_facet_exists_docids_chunk)));
-                ))
+
+                Ok((fid_docid_facet_numbers_chunk, fid_docid_facet_strings_chunk))
             },
         );
 

milli/src/update/index_documents/helpers/grenad_helpers.rs

@@ -90,90 +90,6 @@ pub unsafe fn as_cloneable_grenad(
     Ok(reader)
 }
 
-pub trait MergeableReader
-where
-    Self: Sized,
-{
-    type Output;
-
-    fn merge(self, merge_fn: MergeFn, indexer: &GrenadParameters) -> Result<Self::Output>;
-}
-
-impl MergeableReader for Vec<grenad::Reader<BufReader<File>>> {
-    type Output = grenad::Reader<BufReader<File>>;
-
-    fn merge(self, merge_fn: MergeFn, params: &GrenadParameters) -> Result<Self::Output> {
-        let mut merger = MergerBuilder::new(merge_fn);
-        self.into_iter().try_for_each(|r| merger.push(r))?;
-        merger.finish(params)
-    }
-}
-
-impl MergeableReader for Vec<(grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>)> {
-    type Output = (grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>);
-
-    fn merge(self, merge_fn: MergeFn, params: &GrenadParameters) -> Result<Self::Output> {
-        let mut m1 = MergerBuilder::new(merge_fn);
-        let mut m2 = MergerBuilder::new(merge_fn);
-        for (r1, r2) in self.into_iter() {
-            m1.push(r1)?;
-            m2.push(r2)?;
-        }
-        Ok((m1.finish(params)?, m2.finish(params)?))
-    }
-}
-
-impl MergeableReader
-    for Vec<(
-        grenad::Reader<BufReader<File>>,
-        grenad::Reader<BufReader<File>>,
-        grenad::Reader<BufReader<File>>,
-    )>
-{
-    type Output = (
-        grenad::Reader<BufReader<File>>,
-        grenad::Reader<BufReader<File>>,
-        grenad::Reader<BufReader<File>>,
-    );
-
-    fn merge(self, merge_fn: MergeFn, params: &GrenadParameters) -> Result<Self::Output> {
-        let mut m1 = MergerBuilder::new(merge_fn);
-        let mut m2 = MergerBuilder::new(merge_fn);
-        let mut m3 = MergerBuilder::new(merge_fn);
-        for (r1, r2, r3) in self.into_iter() {
-            m1.push(r1)?;
-            m2.push(r2)?;
-            m3.push(r3)?;
-        }
-        Ok((m1.finish(params)?, m2.finish(params)?, m3.finish(params)?))
-    }
-}
-
-struct MergerBuilder<R>(grenad::MergerBuilder<R, MergeFn>);
-
-impl<R: io::Read + io::Seek> MergerBuilder<R> {
-    fn new(merge_fn: MergeFn) -> Self {
-        Self(grenad::MergerBuilder::new(merge_fn))
-    }
-
-    fn push(&mut self, reader: grenad::Reader<R>) -> Result<()> {
-        self.0.push(reader.into_cursor()?);
-        Ok(())
-    }
-
-    fn finish(self, params: &GrenadParameters) -> Result<grenad::Reader<BufReader<File>>> {
-        let merger = self.0.build();
-        let mut writer = create_writer(
-            params.chunk_compression_type,
-            params.chunk_compression_level,
-            tempfile::tempfile()?,
-        );
-        merger.write_into_stream_writer(&mut writer)?;
-
-        writer_into_reader(writer)
-    }
-}
-
 #[derive(Debug, Clone, Copy)]
 pub struct GrenadParameters {
     pub chunk_compression_type: CompressionType,

milli/src/update/index_documents/helpers/merge_functions.rs

@@ -35,27 +35,6 @@ pub fn merge_roaring_bitmaps<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Resul
     }
 }
 
-pub fn merge_btreeset_string<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
-    if values.len() == 1 {
-        Ok(values[0].clone())
-    } else {
-        // TODO improve the perf by using a `#[borrow] Cow<str>`.
-        let strings: BTreeSet<String> = values
-            .iter()
-            .map(AsRef::as_ref)
-            .map(serde_json::from_slice::<BTreeSet<String>>)
-            .map(StdResult::unwrap)
-            .reduce(|mut current, new| {
-                for x in new {
-                    current.insert(x);
-                }
-                current
-            })
-            .unwrap();
-        Ok(Cow::Owned(serde_json::to_vec(&strings).unwrap()))
-    }
-}
-
 pub fn keep_first<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
     Ok(values[0].clone())
 }
@@ -243,3 +222,40 @@ pub fn merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap<'a>(
         buffer,
     )?)
 }
+
+/// Do a union of BtreeSet on both sides of a DelAdd obkv
+/// separately and outputs a new DelAdd with both unions.
+pub fn merge_deladd_btreeset_string<'a>(
+    _key: &[u8],
+    values: &[Cow<'a, [u8]>],
+) -> Result<Cow<'a, [u8]>> {
+    if values.len() == 1 {
+        Ok(values[0].clone())
+    } else {
+        // Retrieve the bitmaps from both sides
+        let mut del_set = BTreeSet::new();
+        let mut add_set = BTreeSet::new();
+        for value in values {
+            let obkv = KvReaderDelAdd::new(value);
+            if let Some(bytes) = obkv.get(DelAdd::Deletion) {
+                let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
+                for value in set {
+                    del_set.insert(value);
+                }
+            }
+            if let Some(bytes) = obkv.get(DelAdd::Addition) {
+                let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
+                for value in set {
+                    add_set.insert(value);
+                }
+            }
+        }
+
+        let mut output_deladd_obkv = KvWriterDelAdd::memory();
+        let del = serde_json::to_vec(&del_set).unwrap();
+        output_deladd_obkv.insert(DelAdd::Deletion, &del)?;
+        let add = serde_json::to_vec(&add_set).unwrap();
+        output_deladd_obkv.insert(DelAdd::Addition, &add)?;
+        output_deladd_obkv.into_inner().map(Cow::from).map_err(Into::into)
+    }
+}

milli/src/update/index_documents/helpers/mod.rs

@@ -10,10 +10,10 @@ use fst::{IntoStreamer, Streamer};
 pub use grenad_helpers::{
     as_cloneable_grenad, create_sorter, create_writer, grenad_obkv_into_chunks,
     merge_ignore_values, sorter_into_reader, write_sorter_into_database, writer_into_reader,
-    GrenadParameters, MergeableReader,
+    GrenadParameters,
 };
 pub use merge_functions::{
-    keep_first, keep_latest_obkv, merge_btreeset_string, merge_cbo_roaring_bitmaps,
+    keep_first, keep_latest_obkv, merge_cbo_roaring_bitmaps, merge_deladd_btreeset_string,
     merge_deladd_cbo_roaring_bitmaps, merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
     merge_roaring_bitmaps, obkvs_keep_last_addition_merge_deletions,
     obkvs_merge_additions_and_deletions, MergeFn,

milli/src/update/index_documents/mod.rs

@@ -5,29 +5,29 @@ mod transform;
 mod typed_chunk;
 
 use std::collections::{HashMap, HashSet};
-use std::io::{Cursor, Read, Seek};
+use std::io::{Read, Seek};
 use std::iter::FromIterator;
 use std::num::NonZeroU32;
 use std::result::Result as StdResult;
 
 use crossbeam_channel::{Receiver, Sender};
+use grenad::{Merger, MergerBuilder};
 use heed::types::Str;
 use heed::Database;
 use rand::SeedableRng;
 use roaring::RoaringBitmap;
 use serde::{Deserialize, Serialize};
 use slice_group_by::GroupBy;
-use tracing::debug_span;
+use tracing::debug;
-use typed_chunk::{write_typed_chunk_into_index, TypedChunk};
+use typed_chunk::{write_typed_chunk_into_index, ChunkAccumulator, TypedChunk};
 
 use self::enrich::enrich_documents_batch;
 pub use self::enrich::{extract_finite_float_from_value, DocumentId};
 pub use self::helpers::{
     as_cloneable_grenad, create_sorter, create_writer, fst_stream_into_hashset,
-    fst_stream_into_vec, merge_btreeset_string, merge_cbo_roaring_bitmaps,
+    fst_stream_into_vec, merge_cbo_roaring_bitmaps, merge_deladd_cbo_roaring_bitmaps,
-    merge_deladd_cbo_roaring_bitmaps, merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
+    merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, merge_roaring_bitmaps,
-    merge_roaring_bitmaps, valid_lmdb_key, write_sorter_into_database, writer_into_reader,
+    valid_lmdb_key, write_sorter_into_database, writer_into_reader, MergeFn,
-    ClonableMmap, MergeFn,
 };
 use self::helpers::{grenad_obkv_into_chunks, GrenadParameters};
 pub use self::transform::{Transform, TransformOutput};
@@ -95,8 +95,8 @@ pub struct IndexDocumentsConfig {
 
 impl<'t, 'i, 'a, FP, FA> IndexDocuments<'t, 'i, 'a, FP, FA>
 where
-    FP: Fn(UpdateIndexingStep) + Sync,
+    FP: Fn(UpdateIndexingStep) + Sync + Send,
-    FA: Fn() -> bool + Sync,
+    FA: Fn() -> bool + Sync + Send,
 {
     pub fn new(
         wtxn: &'t mut heed::RwTxn<'i>,
@@ -284,7 +284,7 @@ where
     #[tracing::instrument(
         level = "trace",
         skip_all,
-        target = "profile::indexing::details",
+        target = "indexing::details",
         name = "index_documents_raw"
     )]
     pub fn execute_raw(self, output: TransformOutput) -> Result<u64>
@@ -326,9 +326,6 @@ where
             }
         };
 
-        let original_documents = grenad::Reader::new(original_documents)?;
-        let flattened_documents = grenad::Reader::new(flattened_documents)?;
-
         // create LMDB writer channel
         let (lmdb_writer_sx, lmdb_writer_rx): (
             Sender<Result<TypedChunk>>,
@@ -367,11 +364,7 @@ where
 
         let stop_words = self.index.stop_words(self.wtxn)?;
         let separators = self.index.allowed_separators(self.wtxn)?;
-        let separators: Option<Vec<_>> =
-            separators.as_ref().map(|x| x.iter().map(String::as_str).collect());
         let dictionary = self.index.dictionary(self.wtxn)?;
-        let dictionary: Option<Vec<_>> =
-            dictionary.as_ref().map(|x| x.iter().map(String::as_str).collect());
         let exact_attributes = self.index.exact_attributes_ids(self.wtxn)?;
         let proximity_precision = self.index.proximity_precision(self.wtxn)?.unwrap_or_default();
 
@@ -381,141 +374,204 @@ where
             max_memory: self.indexer_config.max_memory,
             max_nb_chunks: self.indexer_config.max_nb_chunks, // default value, may be chosen.
         };
-        let documents_chunk_size =
+        let documents_chunk_size = match self.indexer_config.documents_chunk_size {
-            self.indexer_config.documents_chunk_size.unwrap_or(1024 * 1024 * 4); // 4MiB
+            Some(chunk_size) => chunk_size,
+            None => {
+                let default_chunk_size = 1024 * 1024 * 4; // 4MiB
+                let min_chunk_size = 1024 * 512; // 512KiB
+
+                // compute the chunk size from the number of available threads and the inputed data size.
+                let total_size = flattened_documents.metadata().map(|m| m.len());
+                let current_num_threads = pool.current_num_threads();
+                // if we have more than 2 thread, create a number of chunk equal to 3/4 threads count
+                let chunk_count = if current_num_threads > 2 {
+                    (current_num_threads * 3 / 4).max(2)
+                } else {
+                    current_num_threads
+                };
+                total_size
+                    .map_or(default_chunk_size, |size| (size as usize) / chunk_count)
+                    .max(min_chunk_size)
+            }
+        };
+
+        let original_documents = grenad::Reader::new(original_documents)?;
+        let flattened_documents = grenad::Reader::new(flattened_documents)?;
+
         let max_positions_per_attributes = self.indexer_config.max_positions_per_attributes;
 
         let cloned_embedder = self.embedders.clone();
 
+        let mut final_documents_ids = RoaringBitmap::new();
+        let mut databases_seen = 0;
+        let mut word_position_docids = None;
+        let mut word_fid_docids = None;
+        let mut word_docids = None;
+        let mut exact_word_docids = None;
+        let mut chunk_accumulator = ChunkAccumulator::default();
+        let mut dimension = HashMap::new();
+        let stop_words = stop_words.map(|sw| sw.map_data(Vec::from).unwrap());
+
         let current_span = tracing::Span::current();
 
         // Run extraction pipeline in parallel.
         pool.install(|| {
-            let child_span = tracing::trace_span!(target: "indexing::details", parent: &current_span, "extract_and_send_grenad_chunks");
+            rayon::spawn(move || {
+                let child_span = tracing::trace_span!(target: "indexing::details", parent: &current_span, "extract_and_send_grenad_chunks");
             let _enter = child_span.enter();
             puffin::profile_scope!("extract_and_send_grenad_chunks");
-            // split obkv file into several chunks
+                // split obkv file into several chunks
-            let original_chunk_iter =
+                let original_chunk_iter =
-                grenad_obkv_into_chunks(original_documents, pool_params, documents_chunk_size);
+                    grenad_obkv_into_chunks(original_documents, pool_params, documents_chunk_size);
 
-            // split obkv file into several chunks
+                // split obkv file into several chunks
-            let flattened_chunk_iter =
+                let flattened_chunk_iter =
-                grenad_obkv_into_chunks(flattened_documents, pool_params, documents_chunk_size);
+                    grenad_obkv_into_chunks(flattened_documents, pool_params, documents_chunk_size);
 
-            let result = original_chunk_iter.and_then(|original_chunk| {
+                let separators: Option<Vec<_>> =
-                let flattened_chunk = flattened_chunk_iter?;
+                    separators.as_ref().map(|x| x.iter().map(String::as_str).collect());
-                // extract all databases from the chunked obkv douments
+                let dictionary: Option<Vec<_>> =
-                extract::data_from_obkv_documents(
+                    dictionary.as_ref().map(|x| x.iter().map(String::as_str).collect());
-                    original_chunk,
+                let result = original_chunk_iter.and_then(|original_chunk| {
-                    flattened_chunk,
+                    let flattened_chunk = flattened_chunk_iter?;
-                    pool_params,
+                    // extract all databases from the chunked obkv douments
-                    lmdb_writer_sx.clone(),
+                    extract::data_from_obkv_documents(
-                    searchable_fields,
+                        original_chunk,
-                    faceted_fields,
+                        flattened_chunk,
-                    primary_key_id,
+                        pool_params,
-                    geo_fields_ids,
+                        lmdb_writer_sx.clone(),
-                    field_id_map,
+                        searchable_fields,
-                    stop_words,
+                        faceted_fields,
-                    separators.as_deref(),
+                        primary_key_id,
-                    dictionary.as_deref(),
+                        geo_fields_ids,
-                    max_positions_per_attributes,
+                        field_id_map,
-                    exact_attributes,
+                        stop_words,
-                    proximity_precision,
+                        separators.as_deref(),
-                    cloned_embedder,
+                        dictionary.as_deref(),
-                )
+                        max_positions_per_attributes,
+                        exact_attributes,
+                        proximity_precision,
+                        cloned_embedder,
+                    )
+                });
+
+                if let Err(e) = result {
+                    let _ = lmdb_writer_sx.send(Err(e));
+                }
+
+                // needs to be dropped to avoid channel waiting lock.
+                drop(lmdb_writer_sx);
             });
 
-            if let Err(e) = result {
+            (self.progress)(UpdateIndexingStep::MergeDataIntoFinalDatabase {
-                let _ = lmdb_writer_sx.send(Err(e));
+                databases_seen,
-            }
+                total_databases: TOTAL_POSTING_DATABASE_COUNT,
+            });
 
-            // needs to be dropped to avoid channel waiting lock.
+            loop {
-            drop(lmdb_writer_sx);
+                if (self.should_abort)() {
-        });
+                    return Err(Error::InternalError(InternalError::AbortedIndexation));
+                }
 
-        let index_is_empty = self.index.number_of_documents(self.wtxn)? == 0;
+                match lmdb_writer_rx.clone().recv_timeout(std::time::Duration::from_millis(500)) {
-        let mut final_documents_ids = RoaringBitmap::new();
+                    Err(status) => {
+                        if let Some(typed_chunks) = chunk_accumulator.pop_longest() {
+                            let (docids, is_merged_database) =
+                                write_typed_chunk_into_index(typed_chunks, self.index, self.wtxn)?;
+                            if !docids.is_empty() {
+                                final_documents_ids |= docids;
+                                let documents_seen_count = final_documents_ids.len();
+                                (self.progress)(UpdateIndexingStep::IndexDocuments {
+                                    documents_seen: documents_seen_count as usize,
+                                    total_documents: documents_count,
+                                });
+                                debug!(documents = documents_seen_count, total = documents_count, "Seen");
+                            }
+                            if is_merged_database {
+                                databases_seen += 1;
+                                (self.progress)(UpdateIndexingStep::MergeDataIntoFinalDatabase {
+                                    databases_seen,
+                                    total_databases: TOTAL_POSTING_DATABASE_COUNT,
+                                });
+                            }
+                        // If no more chunk remains in the chunk accumulator and the channel is disconected, break.
+                        } else if status == crossbeam_channel::RecvTimeoutError::Disconnected {
+                            break;
+                        } else {
+                            rayon::yield_now();
+                        }
+                    }
+                    Ok(result) => {
+                        let typed_chunk = match result? {
+                            TypedChunk::WordDocids {
+                                word_docids_reader,
+                                exact_word_docids_reader,
+                                word_fid_docids_reader,
+                            } => {
+                                let cloneable_chunk =
+                                    unsafe { as_cloneable_grenad(&word_docids_reader)? };
+                                let word_docids = word_docids.get_or_insert_with(|| {
+                                    MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn)
+                                });
+                                word_docids.push(cloneable_chunk.into_cursor()?);
+                                let cloneable_chunk =
+                                    unsafe { as_cloneable_grenad(&exact_word_docids_reader)? };
+                                let exact_word_docids =
+                                    exact_word_docids.get_or_insert_with(|| {
+                                        MergerBuilder::new(
+                                            merge_deladd_cbo_roaring_bitmaps as MergeFn,
+                                        )
+                                    });
+                                exact_word_docids.push(cloneable_chunk.into_cursor()?);
+                                let cloneable_chunk =
+                                    unsafe { as_cloneable_grenad(&word_fid_docids_reader)? };
+                                let word_fid_docids = word_fid_docids.get_or_insert_with(|| {
+                                    MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn)
+                                });
+                                word_fid_docids.push(cloneable_chunk.into_cursor()?);
+                                TypedChunk::WordDocids {
+                                    word_docids_reader,
+                                    exact_word_docids_reader,
+                                    word_fid_docids_reader,
+                                }
+                            }
+                            TypedChunk::WordPositionDocids(chunk) => {
+                                let cloneable_chunk = unsafe { as_cloneable_grenad(&chunk)? };
+                                let word_position_docids =
+                                    word_position_docids.get_or_insert_with(|| {
+                                        MergerBuilder::new(
+                                            merge_deladd_cbo_roaring_bitmaps as MergeFn,
+                                        )
+                                    });
+                                word_position_docids.push(cloneable_chunk.into_cursor()?);
+                                TypedChunk::WordPositionDocids(chunk)
+                            }
+                            TypedChunk::VectorPoints {
+                                expected_dimension,
+                                remove_vectors,
+                                embeddings,
+                                manual_vectors,
+                                embedder_name,
+                            } => {
+                                dimension.insert(embedder_name.clone(), expected_dimension);
+                                TypedChunk::VectorPoints {
+                                    remove_vectors,
+                                    embeddings,
+                                    expected_dimension,
+                                    manual_vectors,
+                                    embedder_name,
+                                }
+                            }
+                            otherwise => otherwise,
+                        };
 
-        let mut databases_seen = 0;
+                        chunk_accumulator.insert(typed_chunk);
-        (self.progress)(UpdateIndexingStep::MergeDataIntoFinalDatabase {
-            databases_seen,
-            total_databases: TOTAL_POSTING_DATABASE_COUNT,
-        });
-
-        let mut word_position_docids = None;
-        let mut word_fid_docids = None;
-        let mut word_docids = None;
-        let mut exact_word_docids = None;
-
-        let mut dimension = HashMap::new();
-
-        for result in lmdb_writer_rx {
-            if (self.should_abort)() {
-                return Err(Error::InternalError(InternalError::AbortedIndexation));
-            }
-
-            let typed_chunk = match result? {
-                TypedChunk::WordDocids {
-                    word_docids_reader,
-                    exact_word_docids_reader,
-                    word_fid_docids_reader,
-                } => {
-                    let cloneable_chunk = unsafe { as_cloneable_grenad(&word_docids_reader)? };
-                    word_docids = Some(cloneable_chunk);
-                    let cloneable_chunk =
-                        unsafe { as_cloneable_grenad(&exact_word_docids_reader)? };
-                    exact_word_docids = Some(cloneable_chunk);
-                    let cloneable_chunk = unsafe { as_cloneable_grenad(&word_fid_docids_reader)? };
-                    word_fid_docids = Some(cloneable_chunk);
-                    TypedChunk::WordDocids {
-                        word_docids_reader,
-                        exact_word_docids_reader,
-                        word_fid_docids_reader,
                     }
                 }
-                TypedChunk::WordPositionDocids(chunk) => {
+            }
-                    let cloneable_chunk = unsafe { as_cloneable_grenad(&chunk)? };
-                    word_position_docids = Some(cloneable_chunk);
-                    TypedChunk::WordPositionDocids(chunk)
-                }
-                TypedChunk::VectorPoints {
-                    expected_dimension,
-                    remove_vectors,
-                    embeddings,
-                    manual_vectors,
-                    embedder_name,
-                } => {
-                    dimension.insert(embedder_name.clone(), expected_dimension);
-                    TypedChunk::VectorPoints {
-                        remove_vectors,
-                        embeddings,
-                        expected_dimension,
-                        manual_vectors,
-                        embedder_name,
-                    }
-                }
-                otherwise => otherwise,
-            };
 
-            let (docids, is_merged_database) =
+            Ok(())
-                write_typed_chunk_into_index(typed_chunk, self.index, self.wtxn, index_is_empty)?;
+        })?;
-            if !docids.is_empty() {
-                final_documents_ids |= docids;
-                let documents_seen_count = final_documents_ids.len();
-                (self.progress)(UpdateIndexingStep::IndexDocuments {
-                    documents_seen: documents_seen_count as usize,
-                    total_documents: documents_count,
-                });
-                debug_span!("Seen", documents = documents_seen_count, total = documents_count);
-            }
-            if is_merged_database {
-                databases_seen += 1;
-                (self.progress)(UpdateIndexingStep::MergeDataIntoFinalDatabase {
-                    databases_seen,
-                    total_databases: TOTAL_POSTING_DATABASE_COUNT,
-                });
-            }
-        }
 
         // We write the field distribution into the main database
         self.index.put_field_distribution(self.wtxn, &field_distribution)?;
@@ -548,10 +604,10 @@ where
         }
 
         self.execute_prefix_databases(
-            word_docids,
+            word_docids.map(MergerBuilder::build),
-            exact_word_docids,
+            exact_word_docids.map(MergerBuilder::build),
-            word_position_docids,
+            word_position_docids.map(MergerBuilder::build),
-            word_fid_docids,
+            word_fid_docids.map(MergerBuilder::build),
         )?;
 
         Ok(number_of_documents)
@@ -565,10 +621,10 @@ where
     )]
     pub fn execute_prefix_databases(
         self,
-        word_docids: Option<grenad::Reader<CursorClonableMmap>>,
+        word_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        exact_word_docids: Option<grenad::Reader<CursorClonableMmap>>,
+        exact_word_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        word_position_docids: Option<grenad::Reader<CursorClonableMmap>>,
+        word_position_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        word_fid_docids: Option<grenad::Reader<CursorClonableMmap>>,
+        word_fid_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
     ) -> Result<()>
     where
         FP: Fn(UpdateIndexingStep) + Sync,
@@ -751,7 +807,7 @@ where
 )]
 fn execute_word_prefix_docids(
     txn: &mut heed::RwTxn,
-    reader: grenad::Reader<Cursor<ClonableMmap>>,
+    merger: Merger<CursorClonableMmap, MergeFn>,
     word_docids_db: Database<Str, CboRoaringBitmapCodec>,
     word_prefix_docids_db: Database<Str, CboRoaringBitmapCodec>,
     indexer_config: &IndexerConfig,
@@ -761,13 +817,12 @@ fn execute_word_prefix_docids(
 ) -> Result<()> {
     puffin::profile_function!();
 
-    let cursor = reader.into_cursor()?;
     let mut builder = WordPrefixDocids::new(txn, word_docids_db, word_prefix_docids_db);
     builder.chunk_compression_type = indexer_config.chunk_compression_type;
     builder.chunk_compression_level = indexer_config.chunk_compression_level;
     builder.max_nb_chunks = indexer_config.max_nb_chunks;
     builder.max_memory = indexer_config.max_memory;
-    builder.execute(cursor, new_prefix_fst_words, common_prefix_fst_words, del_prefix_fst_words)?;
+    builder.execute(merger, new_prefix_fst_words, common_prefix_fst_words, del_prefix_fst_words)?;
     Ok(())
 }
 

milli/src/update/index_documents/typed_chunk.rs

@@ -5,27 +5,64 @@ use std::io::{self, BufReader};
 
 use bytemuck::allocation::pod_collect_to_vec;
 use charabia::{Language, Script};
-use grenad::MergerBuilder;
+use grenad::{Merger, MergerBuilder};
 use heed::types::Bytes;
-use heed::{PutFlags, RwTxn};
+use heed::RwTxn;
 use obkv::{KvReader, KvWriter};
 use roaring::RoaringBitmap;
 
 use super::helpers::{
-    self, merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, merge_ignore_values,
+    self, keep_first, merge_deladd_btreeset_string, merge_deladd_cbo_roaring_bitmaps,
-    valid_lmdb_key, CursorClonableMmap,
+    merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, merge_ignore_values, valid_lmdb_key,
+    CursorClonableMmap,
 };
-use super::{ClonableMmap, MergeFn};
+use super::MergeFn;
 use crate::external_documents_ids::{DocumentOperation, DocumentOperationKind};
 use crate::facet::FacetType;
 use crate::index::db_name::DOCUMENTS;
 use crate::update::del_add::{deladd_serialize_add_side, DelAdd, KvReaderDelAdd};
 use crate::update::facet::FacetsUpdate;
-use crate::update::index_documents::helpers::{as_cloneable_grenad, try_split_array_at};
+use crate::update::index_documents::helpers::{
+    as_cloneable_grenad, keep_latest_obkv, try_split_array_at,
+};
 use crate::{
     lat_lng_to_xyz, DocumentId, FieldId, GeoPoint, Index, InternalError, Result, SerializationError,
 };
 
+/// This struct accumulates and group the TypedChunks
+/// and is able to give the biggest accumulated group to index them all together
+/// with a merger.
+#[derive(Default)]
+pub(crate) struct ChunkAccumulator {
+    inner: Vec<Vec<TypedChunk>>,
+}
+
+impl ChunkAccumulator {
+    pub fn pop_longest(&mut self) -> Option<Vec<TypedChunk>> {
+        match self.inner.iter().max_by_key(|v| v.len()) {
+            Some(left) => {
+                let position = self.inner.iter().position(|right| left.len() == right.len());
+                position.map(|p| self.inner.remove(p)).filter(|v| !v.is_empty())
+            }
+            None => None,
+        }
+    }
+
+    pub fn insert(&mut self, chunk: TypedChunk) {
+        match self
+            .inner
+            .iter()
+            .position(|right| right.first().map_or(false, |right| chunk.mergeable_with(right)))
+        {
+            Some(position) => {
+                let v = self.inner.get_mut(position).unwrap();
+                v.push(chunk);
+            }
+            None => self.inner.push(vec![chunk]),
+        }
+    }
+}
+
 pub(crate) enum TypedChunk {
     FieldIdDocidFacetStrings(grenad::Reader<CursorClonableMmap>),
     FieldIdDocidFacetNumbers(grenad::Reader<CursorClonableMmap>),
@@ -38,7 +75,7 @@ pub(crate) enum TypedChunk {
     },
     WordPositionDocids(grenad::Reader<BufReader<File>>),
     WordPairProximityDocids(grenad::Reader<BufReader<File>>),
-    FieldIdFacetStringDocids(grenad::Reader<BufReader<File>>),
+    FieldIdFacetStringDocids((grenad::Reader<BufReader<File>>, grenad::Reader<BufReader<File>>)),
     FieldIdFacetNumberDocids(grenad::Reader<BufReader<File>>),
     FieldIdFacetExistsDocids(grenad::Reader<BufReader<File>>),
     FieldIdFacetIsNullDocids(grenad::Reader<BufReader<File>>),
@@ -54,6 +91,33 @@ pub(crate) enum TypedChunk {
     ScriptLanguageDocids(HashMap<(Script, Language), (RoaringBitmap, RoaringBitmap)>),
 }
 
+impl TypedChunk {
+    fn mergeable_with(&self, other: &Self) -> bool {
+        use TypedChunk::*;
+        match (self, other) {
+            (FieldIdDocidFacetStrings(_), FieldIdDocidFacetStrings(_))
+            | (FieldIdDocidFacetNumbers(_), FieldIdDocidFacetNumbers(_))
+            | (Documents(_), Documents(_))
+            | (FieldIdWordCountDocids(_), FieldIdWordCountDocids(_))
+            | (WordDocids { .. }, WordDocids { .. })
+            | (WordPositionDocids(_), WordPositionDocids(_))
+            | (WordPairProximityDocids(_), WordPairProximityDocids(_))
+            | (FieldIdFacetStringDocids(_), FieldIdFacetStringDocids(_))
+            | (FieldIdFacetNumberDocids(_), FieldIdFacetNumberDocids(_))
+            | (FieldIdFacetExistsDocids(_), FieldIdFacetExistsDocids(_))
+            | (FieldIdFacetIsNullDocids(_), FieldIdFacetIsNullDocids(_))
+            | (FieldIdFacetIsEmptyDocids(_), FieldIdFacetIsEmptyDocids(_))
+            | (GeoPoints(_), GeoPoints(_))
+            | (ScriptLanguageDocids(_), ScriptLanguageDocids(_)) => true,
+            (
+                VectorPoints { embedder_name: left, expected_dimension: left_dim, .. },
+                VectorPoints { embedder_name: right, expected_dimension: right_dim, .. },
+            ) => left == right && left_dim == right_dim,
+            _ => false,
+        }
+    }
+}
+
 impl TypedChunk {
     pub fn to_debug_string(&self) -> String {
         match self {
@@ -85,7 +149,7 @@ impl TypedChunk {
             TypedChunk::WordPairProximityDocids(grenad) => {
                 format!("WordPairProximityDocids {{ number_of_entries: {} }}", grenad.len())
             }
-            TypedChunk::FieldIdFacetStringDocids(grenad) => {
+            TypedChunk::FieldIdFacetStringDocids((grenad, _)) => {
                 format!("FieldIdFacetStringDocids {{ number_of_entries: {} }}", grenad.len())
             }
             TypedChunk::FieldIdFacetNumberDocids(grenad) => {
@@ -117,23 +181,32 @@ impl TypedChunk {
 /// Return new documents seen.
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::write_db")]
 pub(crate) fn write_typed_chunk_into_index(
-    typed_chunk: TypedChunk,
+    typed_chunks: Vec<TypedChunk>,
     index: &Index,
     wtxn: &mut RwTxn,
-    index_is_empty: bool,
 ) -> Result<(RoaringBitmap, bool)> {
-    puffin::profile_function!(typed_chunk.to_debug_string());
+    puffin::profile_function!(typed_chunks[0].to_debug_string());
 
     let mut is_merged_database = false;
-    match typed_chunk {
+    match typed_chunks[0] {
-        TypedChunk::Documents(obkv_documents_iter) => {
+        TypedChunk::Documents(_) => {
             let span = tracing::trace_span!(target: "indexing::write_db", "documents");
             let _entered = span.enter();
+
+            let mut builder = MergerBuilder::new(keep_latest_obkv as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::Documents(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
             let mut operations: Vec<DocumentOperation> = Default::default();
 
             let mut docids = index.documents_ids(wtxn)?;
-            let mut cursor = obkv_documents_iter.into_cursor()?;
+            let mut iter = merger.into_stream_merger_iter()?;
-            while let Some((key, reader)) = cursor.move_on_next()? {
+            while let Some((key, reader)) = iter.next()? {
                 let mut writer: KvWriter<_, FieldId> = KvWriter::memory();
                 let reader: KvReader<FieldId> = KvReader::new(reader);
 
@@ -174,59 +247,91 @@ pub(crate) fn write_typed_chunk_into_index(
             external_documents_docids.apply(wtxn, operations)?;
             index.put_documents_ids(wtxn, &docids)?;
         }
-        TypedChunk::FieldIdWordCountDocids(fid_word_count_docids_iter) => {
+        TypedChunk::FieldIdWordCountDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_word_count_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                fid_word_count_docids_iter,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdWordCountDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.field_id_word_count_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::WordDocids {
+        TypedChunk::WordDocids { .. } => {
-            word_docids_reader,
-            exact_word_docids_reader,
-            word_fid_docids_reader,
-        } => {
             let span = tracing::trace_span!(target: "indexing::write_db", "word_docids");
             let _entered = span.enter();
-            let word_docids_iter = unsafe { as_cloneable_grenad(&word_docids_reader) }?;
+
-            append_entries_into_database(
+            let mut word_docids_builder =
-                word_docids_iter.clone(),
+                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut exact_word_docids_builder =
+                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut word_fid_docids_builder =
+                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut fst_merger_builder = MergerBuilder::new(merge_ignore_values as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::WordDocids {
+                    word_docids_reader,
+                    exact_word_docids_reader,
+                    word_fid_docids_reader,
+                } = typed_chunk
+                else {
+                    unreachable!();
+                };
+                let clonable_word_docids = unsafe { as_cloneable_grenad(&word_docids_reader) }?;
+                let clonable_exact_word_docids =
+                    unsafe { as_cloneable_grenad(&exact_word_docids_reader) }?;
+
+                word_docids_builder.push(word_docids_reader.into_cursor()?);
+                exact_word_docids_builder.push(exact_word_docids_reader.into_cursor()?);
+                word_fid_docids_builder.push(word_fid_docids_reader.into_cursor()?);
+                fst_merger_builder.push(clonable_word_docids.into_cursor()?);
+                fst_merger_builder.push(clonable_exact_word_docids.into_cursor()?);
+            }
+
+            let word_docids_merger = word_docids_builder.build();
+            write_entries_into_database(
+                word_docids_merger,
                 &index.word_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
 
-            let exact_word_docids_iter = unsafe { as_cloneable_grenad(&exact_word_docids_reader) }?;
+            let exact_word_docids_merger = exact_word_docids_builder.build();
-            append_entries_into_database(
+            write_entries_into_database(
-                exact_word_docids_iter.clone(),
+                exact_word_docids_merger,
                 &index.exact_word_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
 
-            let word_fid_docids_iter = unsafe { as_cloneable_grenad(&word_fid_docids_reader) }?;
+            let word_fid_docids_merger = word_fid_docids_builder.build();
-            append_entries_into_database(
+            write_entries_into_database(
-                word_fid_docids_iter,
+                word_fid_docids_merger,
                 &index.word_fid_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
 
             // create fst from word docids
-            let fst = merge_word_docids_reader_into_fst(word_docids_iter, exact_word_docids_iter)?;
+            let fst_merger = fst_merger_builder.build();
+            let fst = merge_word_docids_reader_into_fst(fst_merger)?;
             let db_fst = index.words_fst(wtxn)?;
 
             // merge new fst with database fst
@@ -237,98 +342,202 @@ pub(crate) fn write_typed_chunk_into_index(
             index.put_words_fst(wtxn, &fst)?;
             is_merged_database = true;
         }
-        TypedChunk::WordPositionDocids(word_position_docids_iter) => {
+        TypedChunk::WordPositionDocids(_) => {
             let span = tracing::trace_span!(target: "indexing::write_db", "word_position_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                word_position_docids_iter,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::WordPositionDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.word_position_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdFacetNumberDocids(facet_id_number_docids_iter) => {
+        TypedChunk::FieldIdFacetNumberDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db","field_id_facet_number_docids");
             let _entered = span.enter();
-            let indexer = FacetsUpdate::new(index, FacetType::Number, facet_id_number_docids_iter);
+
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut data_size = 0;
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdFacetNumberDocids(facet_id_number_docids) = typed_chunk
+                else {
+                    unreachable!();
+                };
+
+                data_size += facet_id_number_docids.len();
+                builder.push(facet_id_number_docids.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            let indexer = FacetsUpdate::new(index, FacetType::Number, merger, None, data_size);
             indexer.execute(wtxn)?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdFacetStringDocids(facet_id_string_docids_iter) => {
+        TypedChunk::FieldIdFacetStringDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_facet_string_docids");
             let _entered = span.enter();
-            let indexer = FacetsUpdate::new(index, FacetType::String, facet_id_string_docids_iter);
+
+            let mut facet_id_string_builder =
+                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut normalized_facet_id_string_builder =
+                MergerBuilder::new(merge_deladd_btreeset_string as MergeFn);
+            let mut data_size = 0;
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdFacetStringDocids((
+                    facet_id_string_docids,
+                    normalized_facet_id_string_docids,
+                )) = typed_chunk
+                else {
+                    unreachable!();
+                };
+
+                data_size += facet_id_string_docids.len();
+                facet_id_string_builder.push(facet_id_string_docids.into_cursor()?);
+                normalized_facet_id_string_builder
+                    .push(normalized_facet_id_string_docids.into_cursor()?);
+            }
+            let facet_id_string_merger = facet_id_string_builder.build();
+            let normalized_facet_id_string_merger = normalized_facet_id_string_builder.build();
+
+            let indexer = FacetsUpdate::new(
+                index,
+                FacetType::String,
+                facet_id_string_merger,
+                Some(normalized_facet_id_string_merger),
+                data_size,
+            );
             indexer.execute(wtxn)?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdFacetExistsDocids(facet_id_exists_docids) => {
+        TypedChunk::FieldIdFacetExistsDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_facet_exists_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                facet_id_exists_docids,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdFacetExistsDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.facet_id_exists_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdFacetIsNullDocids(facet_id_is_null_docids) => {
+        TypedChunk::FieldIdFacetIsNullDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_facet_is_null_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                facet_id_is_null_docids,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdFacetIsNullDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.facet_id_is_null_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdFacetIsEmptyDocids(facet_id_is_empty_docids) => {
+        TypedChunk::FieldIdFacetIsEmptyDocids(_) => {
-            let span = tracing::trace_span!(target: "profile::indexing::write_db", "field_id_facet_is_empty_docids");
+            let span = tracing::trace_span!(target: "indexing::write_db", "field_id_facet_is_empty_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                facet_id_is_empty_docids,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdFacetIsEmptyDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.facet_id_is_empty_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::WordPairProximityDocids(word_pair_proximity_docids_iter) => {
+        TypedChunk::WordPairProximityDocids(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "word_pair_proximity_docids");
             let _entered = span.enter();
-            append_entries_into_database(
+
-                word_pair_proximity_docids_iter,
+            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::WordPairProximityDocids(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
+            write_entries_into_database(
+                merger,
                 &index.word_pair_proximity_docids,
                 wtxn,
-                index_is_empty,
                 deladd_serialize_add_side,
                 merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
             )?;
             is_merged_database = true;
         }
-        TypedChunk::FieldIdDocidFacetNumbers(fid_docid_facet_number) => {
+        TypedChunk::FieldIdDocidFacetNumbers(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_docid_facet_numbers");
             let _entered = span.enter();
+
+            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdDocidFacetNumbers(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
             let index_fid_docid_facet_numbers =
                 index.field_id_docid_facet_f64s.remap_types::<Bytes, Bytes>();
-            let mut cursor = fid_docid_facet_number.into_cursor()?;
+            let mut iter = merger.into_stream_merger_iter()?;
-            while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
                 let reader = KvReaderDelAdd::new(value);
                 if valid_lmdb_key(key) {
                     match (reader.get(DelAdd::Deletion), reader.get(DelAdd::Addition)) {
@@ -344,14 +553,25 @@ pub(crate) fn write_typed_chunk_into_index(
                 }
             }
         }
-        TypedChunk::FieldIdDocidFacetStrings(fid_docid_facet_string) => {
+        TypedChunk::FieldIdDocidFacetStrings(_) => {
             let span =
                 tracing::trace_span!(target: "indexing::write_db", "field_id_docid_facet_strings");
             let _entered = span.enter();
+
+            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::FieldIdDocidFacetStrings(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
             let index_fid_docid_facet_strings =
                 index.field_id_docid_facet_strings.remap_types::<Bytes, Bytes>();
-            let mut cursor = fid_docid_facet_string.into_cursor()?;
+            let mut iter = merger.into_stream_merger_iter()?;
-            while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
                 let reader = KvReaderDelAdd::new(value);
                 if valid_lmdb_key(key) {
                     match (reader.get(DelAdd::Deletion), reader.get(DelAdd::Addition)) {
@@ -367,14 +587,25 @@ pub(crate) fn write_typed_chunk_into_index(
                 }
             }
         }
-        TypedChunk::GeoPoints(geo_points) => {
+        TypedChunk::GeoPoints(_) => {
             let span = tracing::trace_span!(target: "indexing::write_db", "geo_points");
             let _entered = span.enter();
+
+            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            for typed_chunk in typed_chunks {
+                let TypedChunk::GeoPoints(chunk) = typed_chunk else {
+                    unreachable!();
+                };
+
+                builder.push(chunk.into_cursor()?);
+            }
+            let merger = builder.build();
+
             let mut rtree = index.geo_rtree(wtxn)?.unwrap_or_default();
             let mut geo_faceted_docids = index.geo_faceted_documents_ids(wtxn)?;
 
-            let mut cursor = geo_points.into_cursor()?;
+            let mut iter = merger.into_stream_merger_iter()?;
-            while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
                 // convert the key back to a u32 (4 bytes)
                 let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();
 
@@ -393,15 +624,38 @@ pub(crate) fn write_typed_chunk_into_index(
             index.put_geo_rtree(wtxn, &rtree)?;
             index.put_geo_faceted_documents_ids(wtxn, &geo_faceted_docids)?;
         }
-        TypedChunk::VectorPoints {
+        TypedChunk::VectorPoints { .. } => {
-            remove_vectors,
-            manual_vectors,
-            embeddings,
-            expected_dimension,
-            embedder_name,
-        } => {
             let span = tracing::trace_span!(target: "indexing::write_db", "vector_points");
             let _entered = span.enter();
+
+            let mut remove_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut manual_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut embeddings_builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut params = None;
+            for typed_chunk in typed_chunks {
+                let TypedChunk::VectorPoints {
+                    remove_vectors,
+                    manual_vectors,
+                    embeddings,
+                    expected_dimension,
+                    embedder_name,
+                } = typed_chunk
+                else {
+                    unreachable!();
+                };
+
+                params = Some((expected_dimension, embedder_name));
+
+                remove_vectors_builder.push(remove_vectors.into_cursor()?);
+                manual_vectors_builder.push(manual_vectors.into_cursor()?);
+                if let Some(embeddings) = embeddings {
+                    embeddings_builder.push(embeddings.into_cursor()?);
+                }
+            }
+
+            // typed chunks has always at least 1 chunk.
+            let Some((expected_dimension, embedder_name)) = params else { unreachable!() };
+
             let embedder_index = index.embedder_category_id.get(wtxn, &embedder_name)?.ok_or(
                 InternalError::DatabaseMissingEntry { db_name: "embedder_category_id", key: None },
             )?;
@@ -419,8 +673,9 @@ pub(crate) fn write_typed_chunk_into_index(
             let writers = writers?;
 
             // remove vectors for docids we want them removed
-            let mut cursor = remove_vectors.into_cursor()?;
+            let merger = remove_vectors_builder.build();
-            while let Some((key, _)) = cursor.move_on_next()? {
+            let mut iter = merger.into_stream_merger_iter()?;
+            while let Some((key, _)) = iter.next()? {
                 let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();
 
                 for writer in &writers {
@@ -432,40 +687,39 @@ pub(crate) fn write_typed_chunk_into_index(
             }
 
             // add generated embeddings
-            if let Some(embeddings) = embeddings {
+            let merger = embeddings_builder.build();
-                let mut cursor = embeddings.into_cursor()?;
+            let mut iter = merger.into_stream_merger_iter()?;
-                while let Some((key, value)) = cursor.move_on_next()? {
+            while let Some((key, value)) = iter.next()? {
-                    let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();
+                let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();
-                    let data = pod_collect_to_vec(value);
+                let data = pod_collect_to_vec(value);
-                    // it is a code error to have embeddings and not expected_dimension
+                // it is a code error to have embeddings and not expected_dimension
-                    let embeddings =
+                let embeddings = crate::vector::Embeddings::from_inner(data, expected_dimension)
-                        crate::vector::Embeddings::from_inner(data, expected_dimension)
+                    // code error if we somehow got the wrong dimension
-                            // code error if we somehow got the wrong dimension
+                    .unwrap();
-                            .unwrap();
 
-                    if embeddings.embedding_count() > usize::from(u8::MAX) {
+                if embeddings.embedding_count() > usize::from(u8::MAX) {
-                        let external_docid = if let Ok(Some(Ok(index))) = index
+                    let external_docid = if let Ok(Some(Ok(index))) = index
-                            .external_id_of(wtxn, std::iter::once(docid))
+                        .external_id_of(wtxn, std::iter::once(docid))
-                            .map(|it| it.into_iter().next())
+                        .map(|it| it.into_iter().next())
-                        {
+                    {
-                            index
+                        index
-                        } else {
+                    } else {
-                            format!("internal docid={docid}")
+                        format!("internal docid={docid}")
-                        };
+                    };
-                        return Err(crate::Error::UserError(crate::UserError::TooManyVectors(
+                    return Err(crate::Error::UserError(crate::UserError::TooManyVectors(
-                            external_docid,
+                        external_docid,
-                            embeddings.embedding_count(),
+                        embeddings.embedding_count(),
-                        )));
+                    )));
-                    }
+                }
-                    for (embedding, writer) in embeddings.iter().zip(&writers) {
+                for (embedding, writer) in embeddings.iter().zip(&writers) {
-                        writer.add_item(wtxn, docid, embedding)?;
+                    writer.add_item(wtxn, docid, embedding)?;
-                    }
                 }
             }
 
             // perform the manual diff
-            let mut cursor = manual_vectors.into_cursor()?;
+            let merger = manual_vectors_builder.build();
-            while let Some((key, value)) = cursor.move_on_next()? {
+            let mut iter = merger.into_stream_merger_iter()?;
+            while let Some((key, value)) = iter.next()? {
                 // convert the key back to a u32 (4 bytes)
                 let (left, _index) = try_split_array_at(key).unwrap();
                 let docid = DocumentId::from_be_bytes(left);
@@ -519,26 +773,30 @@ pub(crate) fn write_typed_chunk_into_index(
 
             tracing::debug!("Finished vector chunk for {}", embedder_name);
         }
-        TypedChunk::ScriptLanguageDocids(sl_map) => {
+        TypedChunk::ScriptLanguageDocids(_) => {
             let span = tracing::trace_span!(target: "indexing::write_db", "script_language_docids");
             let _entered = span.enter();
-            for (key, (deletion, addition)) in sl_map {
-                let mut db_key_exists = false;
-                let final_value = match index.script_language_docids.get(wtxn, &key)? {
-                    Some(db_values) => {
-                        db_key_exists = true;
-                        (db_values - deletion) | addition
-                    }
-                    None => addition,
-                };
 
-                if final_value.is_empty() {
+            for typed_chunk in typed_chunks {
-                    // If the database entry exists, delete it.
+                let TypedChunk::ScriptLanguageDocids(sl_map) = typed_chunk else { unreachable!() };
-                    if db_key_exists {
+                for (key, (deletion, addition)) in sl_map {
-                        index.script_language_docids.delete(wtxn, &key)?;
+                    let mut db_key_exists = false;
+                    let final_value = match index.script_language_docids.get(wtxn, &key)? {
+                        Some(db_values) => {
+                            db_key_exists = true;
+                            (db_values - deletion) | addition
+                        }
+                        None => addition,
+                    };
+
+                    if final_value.is_empty() {
+                        // If the database entry exists, delete it.
+                        if db_key_exists {
+                            index.script_language_docids.delete(wtxn, &key)?;
+                        }
+                    } else {
+                        index.script_language_docids.put(wtxn, &key, &final_value)?;
                     }
-                } else {
-                    index.script_language_docids.put(wtxn, &key, &final_value)?;
                 }
             }
         }
@@ -557,13 +815,9 @@ fn extract_geo_point(value: &[u8], docid: DocumentId) -> GeoPoint {
 }
 
 fn merge_word_docids_reader_into_fst(
-    word_docids_iter: grenad::Reader<io::Cursor<ClonableMmap>>,
+    merger: Merger<CursorClonableMmap, MergeFn>,
-    exact_word_docids_iter: grenad::Reader<io::Cursor<ClonableMmap>>,
 ) -> Result<fst::Set<Vec<u8>>> {
-    let mut merger_builder = MergerBuilder::new(merge_ignore_values as MergeFn);
+    let mut iter = merger.into_stream_merger_iter()?;
-    merger_builder.push(word_docids_iter.into_cursor()?);
-    merger_builder.push(exact_word_docids_iter.into_cursor()?);
-    let mut iter = merger_builder.build().into_stream_merger_iter()?;
     let mut builder = fst::SetBuilder::memory();
 
     while let Some((k, _)) = iter.next()? {
@@ -577,10 +831,9 @@ fn merge_word_docids_reader_into_fst(
 /// merge_values function is used if an entry already exist in the database.
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::write_db")]
 fn write_entries_into_database<R, K, V, FS, FM>(
-    data: grenad::Reader<R>,
+    merger: Merger<R, MergeFn>,
     database: &heed::Database<K, V>,
     wtxn: &mut RwTxn,
-    index_is_empty: bool,
     serialize_value: FS,
     merge_values: FM,
 ) -> Result<()>
@@ -589,22 +842,17 @@ where
     FS: for<'a> Fn(&'a [u8], &'a mut Vec<u8>) -> Result<&'a [u8]>,
     FM: for<'a> Fn(&[u8], &[u8], &'a mut Vec<u8>) -> Result<Option<&'a [u8]>>,
 {
-    puffin::profile_function!(format!("number of entries: {}", data.len()));
+    puffin::profile_function!();
-
     let mut buffer = Vec::new();
     let database = database.remap_types::<Bytes, Bytes>();
 
-    let mut cursor = data.into_cursor()?;
+    let mut iter = merger.into_stream_merger_iter()?;
-    while let Some((key, value)) = cursor.move_on_next()? {
+    while let Some((key, value)) = iter.next()? {
         if valid_lmdb_key(key) {
             buffer.clear();
-            let value = if index_is_empty {
+            let value = match database.get(wtxn, key)? {
-                Some(serialize_value(value, &mut buffer)?)
+                Some(prev_value) => merge_values(value, prev_value, &mut buffer)?,
-            } else {
+                None => Some(serialize_value(value, &mut buffer)?),
-                match database.get(wtxn, key)? {
-                    Some(prev_value) => merge_values(value, prev_value, &mut buffer)?,
-                    None => Some(serialize_value(value, &mut buffer)?),
-                }
             };
             match value {
                 Some(value) => database.put(wtxn, key, value)?,
@@ -614,62 +862,5 @@ where
             }
         }
     }
-
-    Ok(())
-}
-
-/// Write provided entries in database using serialize_value function.
-/// merge_values function is used if an entry already exist in the database.
-/// All provided entries must be ordered.
-/// If the index is not empty, write_entries_into_database is called instead.
-#[tracing::instrument(level = "trace", skip_all, target = "indexing::write_db")]
-fn append_entries_into_database<R, K, V, FS, FM>(
-    data: grenad::Reader<R>,
-    database: &heed::Database<K, V>,
-    wtxn: &mut RwTxn,
-    index_is_empty: bool,
-    serialize_value: FS,
-    merge_values: FM,
-) -> Result<()>
-where
-    R: io::Read + io::Seek,
-    FS: for<'a> Fn(&'a [u8], &'a mut Vec<u8>) -> Result<&'a [u8]>,
-    FM: for<'a> Fn(&[u8], &[u8], &'a mut Vec<u8>) -> Result<Option<&'a [u8]>>,
-    K: for<'a> heed::BytesDecode<'a>,
-{
-    puffin::profile_function!(format!("number of entries: {}", data.len()));
-
-    if !index_is_empty {
-        return write_entries_into_database(
-            data,
-            database,
-            wtxn,
-            false,
-            serialize_value,
-            merge_values,
-        );
-    }
-
-    let mut buffer = Vec::new();
-    let mut database = database.iter_mut(wtxn)?.remap_types::<Bytes, Bytes>();
-
-    let mut cursor = data.into_cursor()?;
-    while let Some((key, value)) = cursor.move_on_next()? {
-        if valid_lmdb_key(key) {
-            debug_assert!(
-                K::bytes_decode(key).is_ok(),
-                "Couldn't decode key with the database decoder, key length: {} - key bytes: {:x?}",
-                key.len(),
-                &key
-            );
-            buffer.clear();
-            let value = serialize_value(value, &mut buffer)?;
-            unsafe {
-                // safety: We do not keep a reference to anything that lives inside the database
-                database.put_current_with_options::<Bytes>(PutFlags::APPEND, key, value)?
-            };
-        }
-    }
-
     Ok(())
 }

milli/src/update/mod.rs

@@ -3,9 +3,8 @@ pub use self::clear_documents::ClearDocuments;
 pub use self::facet::bulk::FacetsUpdateBulk;
 pub use self::facet::incremental::FacetsUpdateIncrementalInner;
 pub use self::index_documents::{
-    merge_btreeset_string, merge_cbo_roaring_bitmaps, merge_roaring_bitmaps,
+    merge_cbo_roaring_bitmaps, merge_roaring_bitmaps, DocumentAdditionResult, DocumentId,
-    DocumentAdditionResult, DocumentId, IndexDocuments, IndexDocumentsConfig, IndexDocumentsMethod,
+    IndexDocuments, IndexDocumentsConfig, IndexDocumentsMethod, MergeFn,
-    MergeFn,
 };
 pub use self::indexer_config::IndexerConfig;
 pub use self::settings::{validate_embedding_settings, Setting, Settings};

milli/src/update/word_prefix_docids.rs

@@ -47,7 +47,7 @@ impl<'t, 'i> WordPrefixDocids<'t, 'i> {
     )]
     pub fn execute(
         self,
-        mut new_word_docids_iter: grenad::ReaderCursor<CursorClonableMmap>,
+        new_word_docids: grenad::Merger<CursorClonableMmap, MergeFn>,
         new_prefix_fst_words: &[String],
         common_prefix_fst_words: &[&[String]],
         del_prefix_fst_words: &HashSet<Vec<u8>>,
@@ -68,7 +68,8 @@ impl<'t, 'i> WordPrefixDocids<'t, 'i> {
         if !common_prefix_fst_words.is_empty() {
             let mut current_prefixes: Option<&&[String]> = None;
             let mut prefixes_cache = HashMap::new();
-            while let Some((word, data)) = new_word_docids_iter.move_on_next()? {
+            let mut new_word_docids_iter = new_word_docids.into_stream_merger_iter()?;
+            while let Some((word, data)) = new_word_docids_iter.next()? {
                 current_prefixes = match current_prefixes.take() {
                     Some(prefixes) if word.starts_with(prefixes[0].as_bytes()) => Some(prefixes),
                     _otherwise => {

milli/src/update/words_prefix_integer_docids.rs

@@ -52,7 +52,7 @@ impl<'t, 'i> WordPrefixIntegerDocids<'t, 'i> {
     )]
     pub fn execute(
         self,
-        new_word_integer_docids: grenad::Reader<CursorClonableMmap>,
+        new_word_integer_docids: grenad::Merger<CursorClonableMmap, MergeFn>,
         new_prefix_fst_words: &[String],
         common_prefix_fst_words: &[&[String]],
         del_prefix_fst_words: &HashSet<Vec<u8>>,
@@ -69,14 +69,14 @@ impl<'t, 'i> WordPrefixIntegerDocids<'t, 'i> {
             self.max_memory,
         );
 
-        let mut new_word_integer_docids_iter = new_word_integer_docids.into_cursor()?;
-
         if !common_prefix_fst_words.is_empty() {
             // We fetch all the new common prefixes between the previous and new prefix fst.
             let mut buffer = Vec::new();
             let mut current_prefixes: Option<&&[String]> = None;
             let mut prefixes_cache = HashMap::new();
-            while let Some((key, data)) = new_word_integer_docids_iter.move_on_next()? {
+            let mut new_word_integer_docids_iter =
+                new_word_integer_docids.into_stream_merger_iter()?;
+            while let Some((key, data)) = new_word_integer_docids_iter.next()? {
                 let (word, pos) =
                     StrBEU16Codec::bytes_decode(key).map_err(heed::Error::Decoding)?;
 

milli/src/vector/error.rs

@@ -59,8 +59,8 @@ pub enum EmbedErrorKind {
     OpenAiAuth(OpenAiError),
     #[error("sent too many requests to OpenAI: {0}")]
     OpenAiTooManyRequests(OpenAiError),
-    #[error("received internal error from OpenAI: {0}")]
+    #[error("received internal error from OpenAI: {0:?}")]
-    OpenAiInternalServerError(OpenAiError),
+    OpenAiInternalServerError(Option<OpenAiError>),
     #[error("sent too many tokens in a request to OpenAI: {0}")]
     OpenAiTooManyTokens(OpenAiError),
     #[error("received unhandled HTTP status code {0} from OpenAI")]
@@ -106,7 +106,7 @@ impl EmbedError {
         Self { kind: EmbedErrorKind::OpenAiTooManyRequests(inner), fault: FaultSource::Runtime }
     }
 
-    pub(crate) fn openai_internal_server_error(inner: OpenAiError) -> EmbedError {
+    pub(crate) fn openai_internal_server_error(inner: Option<OpenAiError>) -> EmbedError {
         Self { kind: EmbedErrorKind::OpenAiInternalServerError(inner), fault: FaultSource::Runtime }
     }
 

milli/src/vector/mod.rs

@@ -261,3 +261,7 @@ impl DistributionShift {
         score
     }
 }
+
+pub const fn is_cuda_enabled() -> bool {
+    cfg!(feature = "cuda")
+}

milli/src/vector/openai.rs

@@ -178,6 +178,8 @@ impl Embedder {
                     retry.into_duration(attempt)
                 }
             }?;
+
+            let retry_duration = retry_duration.min(std::time::Duration::from_secs(60)); // don't wait more than a minute
             tracing::warn!(
                 "Attempt #{}, retrying after {}ms.",
                 attempt,
@@ -220,24 +222,12 @@ impl Embedder {
                         error_response.error,
                     )));
                 }
-                StatusCode::INTERNAL_SERVER_ERROR => {
+                StatusCode::INTERNAL_SERVER_ERROR
-                    let error_response: OpenAiErrorResponse = response
+                | StatusCode::BAD_GATEWAY
-                        .json()
+                | StatusCode::SERVICE_UNAVAILABLE => {
-                        .await
+                    let error_response: Result<OpenAiErrorResponse, _> = response.json().await;
-                        .map_err(EmbedError::openai_unexpected)
-                        .map_err(Retry::retry_later)?;
                     return Err(Retry::retry_later(EmbedError::openai_internal_server_error(
-                        error_response.error,
+                        error_response.ok().map(|error_response| error_response.error),
-                    )));
-                }
-                StatusCode::SERVICE_UNAVAILABLE => {
-                    let error_response: OpenAiErrorResponse = response
-                        .json()
-                        .await
-                        .map_err(EmbedError::openai_unexpected)
-                        .map_err(Retry::retry_later)?;
-                    return Err(Retry::retry_later(EmbedError::openai_internal_server_error(
-                        error_response.error,
                     )));
                 }
                 StatusCode::BAD_REQUEST => {
@@ -248,14 +238,14 @@ impl Embedder {
                         .map_err(EmbedError::openai_unexpected)
                         .map_err(Retry::retry_later)?;
 
-                    tracing::warn!("OpenAI: input was too long, retrying on tokenized version. For best performance, limit the size of your prompt.");
+                    tracing::warn!("OpenAI: received `BAD_REQUEST`. Input was maybe too long, retrying on tokenized version. For best performance, limit the size of your prompt.");
 
                     return Err(Retry::retry_tokenized(EmbedError::openai_too_many_tokens(
                         error_response.error,
                     )));
                 }
                 code => {
-                    return Err(Retry::give_up(EmbedError::openai_unhandled_status_code(
+                    return Err(Retry::retry_later(EmbedError::openai_unhandled_status_code(
                         code.as_u16(),
                     )));
                 }

tracing-trace/src/processor/span_stats.rs

@@ -1,4 +1,5 @@
 use std::collections::{BTreeMap, HashMap};
+use std::ops::Range;
 use std::time::Duration;
 
 use serde::{Deserialize, Serialize};
@@ -16,6 +17,51 @@ enum SpanStatus {
 pub struct CallStats {
     pub call_count: usize,
     pub time: u64,
+    pub self_time: u64,
+}
+
+#[derive(Debug, Default)]
+pub struct SelfTime {
+    child_ranges: Vec<Range<Duration>>,
+}
+
+impl SelfTime {
+    pub fn new() -> Self {
+        Default::default()
+    }
+
+    pub fn add_child_range(&mut self, child_range: Range<Duration>) {
+        self.child_ranges.push(child_range)
+    }
+
+    pub fn self_duration(&mut self, self_range: Range<Duration>) -> Duration {
+        if self.child_ranges.is_empty() {
+            return self_range.end - self_range.start;
+        }
+
+        // by sorting child ranges by their start time,
+        // we make sure that no child will start before the last one we visited.
+        self.child_ranges
+            .sort_by(|left, right| left.start.cmp(&right.start).then(left.end.cmp(&right.end)));
+        // self duration computed by adding all the segments where the span is not executing a child
+        let mut self_duration = Duration::from_nanos(0);
+
+        // last point in time where we are certain that this span was not executing a child.
+        let mut committed_point = self_range.start;
+
+        for child_range in &self.child_ranges {
+            if child_range.start > committed_point {
+                // we add to the self duration the point between the end of the latest span and the beginning of the next span
+                self_duration += child_range.start - committed_point;
+            }
+            if committed_point < child_range.end {
+                // then we set ourselves to the end of the latest span
+                committed_point = child_range.end;
+            }
+        }
+
+        self_duration
+    }
 }
 
 pub fn to_call_stats<R: std::io::Read>(
@@ -23,6 +69,9 @@ pub fn to_call_stats<R: std::io::Read>(
 ) -> Result<BTreeMap<String, CallStats>, Error> {
     let mut calls = HashMap::new();
     let mut spans = HashMap::new();
+    let mut last_point = Duration::from_nanos(0);
+    let mut first_point = None;
+    let mut total_self_time = SelfTime::new();
     for entry in trace {
         let entry = entry?;
         match entry {
@@ -31,10 +80,11 @@ pub fn to_call_stats<R: std::io::Read>(
             }
             Entry::NewThread(_) => {}
             Entry::NewSpan(span) => {
-                spans.insert(span.id, (span, SpanStatus::Outside));
+                spans.insert(span.id, (span, SpanStatus::Outside, SelfTime::new()));
             }
             Entry::SpanEnter(SpanEnter { id, time, memory: _ }) => {
-                let (_, status) = spans.get_mut(&id).unwrap();
+                first_point.get_or_insert(time);
+                let (_, status, _) = spans.get_mut(&id).unwrap();
 
                 let SpanStatus::Outside = status else {
                     continue;
@@ -43,18 +93,32 @@ pub fn to_call_stats<R: std::io::Read>(
                 *status = SpanStatus::Inside(time);
             }
             Entry::SpanExit(SpanExit { id, time: end, memory: _ }) => {
-                let (span, status) = spans.get_mut(&id).unwrap();
+                let (span, status, self_time) = spans.get_mut(&id).unwrap();
 
                 let SpanStatus::Inside(begin) = status else {
                     continue;
                 };
                 let begin = *begin;
 
+                if last_point < end {
+                    last_point = end;
+                }
+
                 *status = SpanStatus::Outside;
 
+                let self_range = begin..end;
+
+                let self_duration = self_time.self_duration(self_range.clone());
+                *self_time = SelfTime::new();
+
                 let span = *span;
+                if let Some(parent_id) = span.parent_id {
+                    let (_, _, parent_self_time) = spans.get_mut(&parent_id).unwrap();
+                    parent_self_time.add_child_range(self_range.clone())
+                }
+                total_self_time.add_child_range(self_range);
                 let (_, call_list) = calls.get_mut(&span.call_id).unwrap();
-                call_list.push(end - begin);
+                call_list.push((end - begin, self_duration));
             }
             Entry::SpanClose(SpanClose { id, time: _ }) => {
                 spans.remove(&id);
@@ -63,17 +127,31 @@ pub fn to_call_stats<R: std::io::Read>(
         }
     }
 
+    let total_self_time = first_point
+        .map(|first_point| (first_point, total_self_time.self_duration(first_point..last_point)));
+
     Ok(calls
         .into_iter()
         .map(|(_, (call_site, calls))| (site_to_string(call_site), calls_to_stats(calls)))
+        .chain(total_self_time.map(|(first_point, total_self_time)| {
+            (
+                "::meta::total".to_string(),
+                CallStats {
+                    call_count: 1,
+                    time: (last_point - first_point).as_nanos() as u64,
+                    self_time: total_self_time.as_nanos() as u64,
+                },
+            )
+        }))
         .collect())
 }
 
 fn site_to_string(call_site: NewCallsite) -> String {
     format!("{}::{}", call_site.target, call_site.name)
 }
-fn calls_to_stats(calls: Vec<Duration>) -> CallStats {
+fn calls_to_stats(calls: Vec<(Duration, Duration)>) -> CallStats {
     let nb = calls.len();
-    let sum: Duration = calls.iter().sum();
+    let sum: Duration = calls.iter().map(|(total, _)| total).sum();
-    CallStats { call_count: nb, time: sum.as_nanos() as u64 }
+    let self_sum: Duration = calls.iter().map(|(_, self_duration)| self_duration).sum();
+    CallStats { call_count: nb, time: sum.as_nanos() as u64, self_time: self_sum.as_nanos() as u64 }
 }

workloads/hackernews.json

@@ -0,0 +1,164 @@
+{
+  "name": "hackernews.ndjson_1M",
+  "run_count": 3,
+  "extra_cli_args": [],
+  "assets": {
+    "hackernews-100_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-100_000.ndjson",
+      "sha256": "60ecd23485d560edbd90d9ca31f0e6dba1455422f2a44e402600fbb5f7f1b213"
+    },
+    "hackernews-200_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-200_000.ndjson",
+      "sha256": "785b0271fdb47cba574fab617d5d332276b835c05dd86e4a95251cf7892a1685"
+    },
+    "hackernews-300_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-300_000.ndjson",
+      "sha256": "de73c7154652eddfaf69cdc3b2f824d5c452f095f40a20a1c97bb1b5c4d80ab2"
+    },
+    "hackernews-400_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-400_000.ndjson",
+      "sha256": "c1b00a24689110f366447e434c201c086d6f456d54ed1c4995894102794d8fe7"
+    },
+    "hackernews-500_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-500_000.ndjson",
+      "sha256": "ae98f9dbef8193d750e3e2dbb6a91648941a1edca5f6e82c143e7996f4840083"
+    },
+    "hackernews-600_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-600_000.ndjson",
+      "sha256": "b495fdc72c4a944801f786400f22076ab99186bee9699f67cbab2f21f5b74dbe"
+    },
+    "hackernews-700_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-700_000.ndjson",
+      "sha256": "4b2c63974f3dabaa4954e3d4598b48324d03c522321ac05b0d583f36cb78a28b"
+    },
+    "hackernews-800_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-800_000.ndjson",
+      "sha256": "cb7b6afe0e6caa1be111be256821bc63b0771b2a0e1fad95af7aaeeffd7ba546"
+    },
+    "hackernews-900_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-900_000.ndjson",
+      "sha256": "e1154ddcd398f1c867758a93db5bcb21a07b9e55530c188a2917fdef332d3ba9"
+    },
+    "hackernews-1_000_000.ndjson": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-1_000_000.ndjson",
+      "sha256": "27e25efd0b68b159b8b21350d9af76938710cb29ce0393fa71b41c4f3c630ffe"
+    }
+  },
+  "commands": [
+    {
+      "route": "indexes/movies/settings",
+      "method": "PATCH",
+      "body": {
+        "inline": {
+          "displayedAttributes": [
+            "title",
+            "by",
+            "score",
+            "time"
+          ],
+          "searchableAttributes": [
+            "title"
+          ],
+          "filterableAttributes": [
+            "by"
+          ],
+          "sortableAttributes": [
+            "score",
+            "time"
+          ]
+        }
+      },
+      "synchronous": "DontWait"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-100_000.ndjson"
+      },
+      "synchronous": "WaitForTask"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-200_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-300_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-400_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-500_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-600_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-700_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-800_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-900_000.ndjson"
+      },
+      "synchronous": "WaitForResponse"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "hackernews-1_000_000.ndjson"
+      },
+      "synchronous": "WaitForTask"
+    }
+  ]
+}

workloads/movies-nothreads.json

@@ -0,0 +1,44 @@
+{
+  "name": "movies.json,no-threads",
+  "run_count": 2,
+  "extra_cli_args": [
+    "--max-indexing-threads=1"
+  ],
+  "assets": {
+    "movies.json": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/movies.json",
+      "sha256": "5b6e4cb660bc20327776e8a33ea197b43d9ec84856710ead1cc87ab24df77de1"
+    }
+  },
+  "commands": [
+    {
+      "route": "indexes/movies/settings",
+      "method": "PATCH",
+      "body": {
+        "inline": {
+          "searchableAttributes": [
+            "title",
+            "overview"
+          ],
+          "filterableAttributes": [
+            "genres",
+            "release_date"
+          ],
+          "sortableAttributes": [
+            "release_date"
+          ]
+        }
+      },
+      "synchronous": "DontWait"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "movies.json"
+      },
+      "synchronous": "WaitForTask"
+    }
+  ]
+}

workloads/movies.json

@@ -0,0 +1,42 @@
+{
+  "name": "movies.json",
+  "run_count": 10,
+  "extra_cli_args": [],
+  "assets": {
+    "movies.json": {
+      "local_location": null,
+      "remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/movies.json",
+      "sha256": "5b6e4cb660bc20327776e8a33ea197b43d9ec84856710ead1cc87ab24df77de1"
+    }
+  },
+  "commands": [
+    {
+      "route": "indexes/movies/settings",
+      "method": "PATCH",
+      "body": {
+        "inline": {
+          "searchableAttributes": [
+            "title",
+            "overview"
+          ],
+          "filterableAttributes": [
+            "genres",
+            "release_date"
+          ],
+          "sortableAttributes": [
+            "release_date"
+          ]
+        }
+      },
+      "synchronous": "DontWait"
+    },
+    {
+      "route": "indexes/movies/documents",
+      "method": "POST",
+      "body": {
+        "asset": "movies.json"
+      },
+      "synchronous": "WaitForTask"
+    }
+  ]
+}

xtask/Cargo.toml

@@ -11,5 +11,34 @@ license.workspace = true
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
+anyhow = "1.0.79"
+build-info = { version = "1.7.0", path = "../build-info" }
 cargo_metadata = "0.18.1"
 clap = { version = "4.4.14", features = ["derive"] }
+futures-core = "0.3.30"
+futures-util = "0.3.30"
+reqwest = { version = "0.11.23", features = [
+    "stream",
+    "json",
+    "rustls-tls",
+], default_features = false }
+serde = { version = "1.0.195", features = ["derive"] }
+serde_json = "1.0.111"
+sha2 = "0.10.8"
+sysinfo = "0.30.5"
+time = { version = "0.3.32", features = [
+    "serde",
+    "serde-human-readable",
+    "macros",
+] }
+tokio = { version = "1.35.1", features = [
+    "rt",
+    "net",
+    "time",
+    "process",
+    "signal",
+] }
+tracing = "0.1.40"
+tracing-subscriber = "0.3.18"
+tracing-trace = { version = "0.1.0", path = "../tracing-trace" }
+uuid = { version = "1.7.0", features = ["v7", "serde"] }

xtask/src/bench/assets.rs

@@ -0,0 +1,250 @@
+use std::collections::BTreeMap;
+use std::io::{Read as _, Seek as _, Write as _};
+
+use anyhow::{bail, Context};
+use futures_util::TryStreamExt as _;
+use serde::Deserialize;
+use sha2::Digest;
+
+use super::client::Client;
+
+#[derive(Deserialize, Clone)]
+pub struct Asset {
+    pub local_location: Option<String>,
+    pub remote_location: Option<String>,
+    #[serde(default)]
+    pub format: AssetFormat,
+    pub sha256: Option<String>,
+}
+
+#[derive(Deserialize, Default, Copy, Clone)]
+pub enum AssetFormat {
+    #[default]
+    Auto,
+    Json,
+    NdJson,
+    Raw,
+}
+
+impl AssetFormat {
+    pub fn to_content_type(self, filename: &str) -> &'static str {
+        match self {
+            AssetFormat::Auto => Self::auto_detect(filename).to_content_type(filename),
+            AssetFormat::Json => "application/json",
+            AssetFormat::NdJson => "application/x-ndjson",
+            AssetFormat::Raw => "application/octet-stream",
+        }
+    }
+
+    fn auto_detect(filename: &str) -> Self {
+        let path = std::path::Path::new(filename);
+        match path.extension().and_then(|extension| extension.to_str()) {
+            Some(extension) if extension.eq_ignore_ascii_case("json") => Self::Json,
+            Some(extension) if extension.eq_ignore_ascii_case("ndjson") => Self::NdJson,
+            extension => {
+                tracing::warn!(asset = filename, ?extension, "asset has format `Auto`, but extension was not recognized. Specify `Raw` format to suppress this warning.");
+                AssetFormat::Raw
+            }
+        }
+    }
+}
+
+pub fn fetch_asset(
+    name: &str,
+    assets: &BTreeMap<String, Asset>,
+    asset_folder: &str,
+) -> anyhow::Result<(std::fs::File, AssetFormat)> {
+    let asset =
+        assets.get(name).with_context(|| format!("could not find asset with name '{name}'"))?;
+    let filename = if let Some(local_filename) = &asset.local_location {
+        local_filename.clone()
+    } else {
+        format!("{asset_folder}/{name}")
+    };
+
+    Ok((
+        std::fs::File::open(&filename)
+            .with_context(|| format!("could not open asset '{name}' at '{filename}'"))?,
+        asset.format,
+    ))
+}
+
+#[tracing::instrument(skip(client, assets), fields(asset_count = assets.len()))]
+pub async fn fetch_assets(
+    client: &Client,
+    assets: &BTreeMap<String, Asset>,
+    asset_folder: &str,
+) -> anyhow::Result<()> {
+    let mut download_tasks = tokio::task::JoinSet::new();
+    for (name, asset) in assets {
+        // trying local
+        if let Some(local) = &asset.local_location {
+            match std::fs::File::open(local) {
+                Ok(file) => {
+                    if check_sha256(name, asset, file)? {
+                        continue;
+                    } else {
+                        tracing::warn!(asset = name, file = local, "found local resource for asset but hash differed, skipping to asset store");
+                    }
+                }
+                Err(error) => match error.kind() {
+                    std::io::ErrorKind::NotFound => { /* file does not exist, go to remote, no need for logs */
+                    }
+                    _ => tracing::warn!(
+                        error = &error as &dyn std::error::Error,
+                        "error checking local resource, skipping to asset store"
+                    ),
+                },
+            }
+        }
+
+        // checking asset store
+        let store_filename = format!("{}/{}", asset_folder, name);
+
+        match std::fs::File::open(&store_filename) {
+            Ok(file) => {
+                if check_sha256(name, asset, file)? {
+                    continue;
+                } else {
+                    tracing::warn!(asset = name, file = store_filename, "found resource for asset in asset store, but hash differed, skipping to remote method");
+                }
+            }
+            Err(error) => match error.kind() {
+                std::io::ErrorKind::NotFound => { /* file does not exist, go to remote, no need for logs */
+                }
+                _ => tracing::warn!(
+                    error = &error as &dyn std::error::Error,
+                    "error checking resource in store, skipping to remote method"
+                ),
+            },
+        }
+
+        // downloading remote
+        match &asset.remote_location {
+            Some(location) => {
+                std::fs::create_dir_all(asset_folder).with_context(|| format!("could not create asset folder at {asset_folder}"))?;
+                download_tasks.spawn({
+                    let client = client.clone();
+                    let name = name.to_string();
+                    let location = location.to_string();
+                    let store_filename = store_filename.clone();
+                    let asset = asset.clone();
+                    download_asset(client, name, asset, location, store_filename)});
+            },
+            None => bail!("asset {name} has no remote location, but was not found locally or in the asset store"),
+        }
+    }
+
+    while let Some(res) = download_tasks.join_next().await {
+        res.context("download task panicked")?.context("download task failed")?;
+    }
+
+    Ok(())
+}
+
+fn check_sha256(name: &str, asset: &Asset, mut file: std::fs::File) -> anyhow::Result<bool> {
+    let mut bytes = Vec::new();
+    file.read_to_end(&mut bytes).with_context(|| format!("hashing file for asset {name}"))?;
+    let mut file_hash = sha2::Sha256::new();
+    file_hash.update(&bytes);
+    let file_hash = file_hash.finalize();
+    let file_hash = format!("{:x}", file_hash);
+    tracing::debug!(hash = file_hash, "hashed local file");
+
+    Ok(match &asset.sha256 {
+        Some(hash) => {
+            tracing::debug!(hash, "hash from workload");
+            if hash.to_ascii_lowercase() == file_hash {
+                true
+            } else {
+                tracing::warn!(
+                    file_hash,
+                    asset_hash = hash.to_ascii_lowercase(),
+                    "hashes don't match"
+                );
+                false
+            }
+        }
+        None => {
+            tracing::warn!(sha256 = file_hash, "Skipping hash for asset {name} that doesn't have one. Please add it to workload file");
+            true
+        }
+    })
+}
+
+#[tracing::instrument(skip(client, asset, name), fields(asset = name))]
+async fn download_asset(
+    client: Client,
+    name: String,
+    asset: Asset,
+    src: String,
+    dest_filename: String,
+) -> anyhow::Result<()> {
+    let context = || format!("failure downloading asset {name} from {src}");
+
+    let response = client.get(&src).send().await.with_context(context)?;
+
+    let file = std::fs::File::options()
+        .create(true)
+        .truncate(true)
+        .write(true)
+        .read(true)
+        .open(&dest_filename)
+        .with_context(|| format!("creating destination file {dest_filename}"))
+        .with_context(context)?;
+
+    let mut dest = std::io::BufWriter::new(
+        file.try_clone().context("cloning I/O handle").with_context(context)?,
+    );
+
+    let total_len: Option<u64> = response
+        .headers()
+        .get(reqwest::header::CONTENT_LENGTH)
+        .and_then(|value| value.to_str().ok())
+        .and_then(|value| value.parse().ok());
+
+    let progress = tokio::spawn({
+        let name = name.clone();
+        async move {
+            loop {
+                match file.metadata().context("could not get file metadata") {
+                    Ok(metadata) => {
+                        let len = metadata.len();
+                        tracing::info!(
+                            asset = name,
+                            downloaded_bytes = len,
+                            total_bytes = total_len,
+                            "asset download in progress"
+                        );
+                    }
+                    Err(error) => {
+                        tracing::warn!(%error, "could not get file metadata");
+                    }
+                }
+                tokio::time::sleep(std::time::Duration::from_secs(60)).await;
+            }
+        }
+    });
+
+    let writing_context = || format!("while writing to destination file at {dest_filename}");
+
+    let mut response = response.bytes_stream();
+
+    while let Some(bytes) =
+        response.try_next().await.context("while downloading file").with_context(context)?
+    {
+        dest.write_all(&bytes).with_context(writing_context).with_context(context)?;
+    }
+
+    progress.abort();
+
+    let mut file = dest.into_inner().with_context(writing_context).with_context(context)?;
+
+    file.rewind().context("while rewinding asset file")?;
+
+    if !check_sha256(&name, &asset, file)? {
+        bail!("asset '{name}': sha256 mismatch for file {dest_filename} downloaded from {src}")
+    }
+
+    Ok(())
+}

xtask/src/bench/client.rs

@@ -0,0 +1,80 @@
+use anyhow::Context;
+use serde::Deserialize;
+
+#[derive(Debug, Clone)]
+pub struct Client {
+    base_url: Option<String>,
+    client: reqwest::Client,
+}
+
+impl Client {
+    pub fn new(
+        base_url: Option<String>,
+        api_key: Option<&str>,
+        timeout: Option<std::time::Duration>,
+    ) -> anyhow::Result<Self> {
+        let mut headers = reqwest::header::HeaderMap::new();
+        if let Some(api_key) = api_key {
+            headers.append(
+                reqwest::header::AUTHORIZATION,
+                reqwest::header::HeaderValue::from_str(&format!("Bearer {api_key}"))
+                    .context("Invalid authorization header")?,
+            );
+        }
+
+        let client = reqwest::ClientBuilder::new().default_headers(headers);
+        let client = if let Some(timeout) = timeout { client.timeout(timeout) } else { client };
+        let client = client.build()?;
+        Ok(Self { base_url, client })
+    }
+
+    pub fn request(&self, method: reqwest::Method, route: &str) -> reqwest::RequestBuilder {
+        if let Some(base_url) = &self.base_url {
+            if route.is_empty() {
+                self.client.request(method, base_url)
+            } else {
+                self.client.request(method, format!("{}/{}", base_url, route))
+            }
+        } else {
+            self.client.request(method, route)
+        }
+    }
+
+    pub fn get(&self, route: &str) -> reqwest::RequestBuilder {
+        self.request(reqwest::Method::GET, route)
+    }
+
+    pub fn put(&self, route: &str) -> reqwest::RequestBuilder {
+        self.request(reqwest::Method::PUT, route)
+    }
+
+    pub fn post(&self, route: &str) -> reqwest::RequestBuilder {
+        self.request(reqwest::Method::POST, route)
+    }
+
+    pub fn delete(&self, route: &str) -> reqwest::RequestBuilder {
+        self.request(reqwest::Method::DELETE, route)
+    }
+}
+
+#[derive(Debug, Clone, Copy, Deserialize)]
+#[serde(rename_all = "SCREAMING_SNAKE_CASE")]
+pub enum Method {
+    Get,
+    Post,
+    Patch,
+    Delete,
+    Put,
+}
+
+impl From<Method> for reqwest::Method {
+    fn from(value: Method) -> Self {
+        match value {
+            Method::Get => Self::GET,
+            Method::Post => Self::POST,
+            Method::Patch => Self::PATCH,
+            Method::Delete => Self::DELETE,
+            Method::Put => Self::PUT,
+        }
+    }
+}

xtask/src/bench/command.rs

@@ -0,0 +1,194 @@
+use std::collections::BTreeMap;
+use std::fmt::Display;
+use std::io::Read as _;
+
+use anyhow::{bail, Context as _};
+use serde::Deserialize;
+
+use super::assets::{fetch_asset, Asset};
+use super::client::{Client, Method};
+
+#[derive(Clone, Deserialize)]
+pub struct Command {
+    pub route: String,
+    pub method: Method,
+    #[serde(default)]
+    pub body: Body,
+    #[serde(default)]
+    pub synchronous: SyncMode,
+}
+
+#[derive(Default, Clone, Deserialize)]
+#[serde(untagged)]
+pub enum Body {
+    Inline {
+        inline: serde_json::Value,
+    },
+    Asset {
+        asset: String,
+    },
+    #[default]
+    Empty,
+}
+
+impl Body {
+    pub fn get(
+        self,
+        assets: &BTreeMap<String, Asset>,
+        asset_folder: &str,
+    ) -> anyhow::Result<Option<(Vec<u8>, &'static str)>> {
+        Ok(match self {
+            Body::Inline { inline: body } => Some((
+                serde_json::to_vec(&body)
+                    .context("serializing to bytes")
+                    .context("while getting inline body")?,
+                "application/json",
+            )),
+            Body::Asset { asset: name } => Some({
+                let context = || format!("while getting body from asset '{name}'");
+                let (mut file, format) =
+                    fetch_asset(&name, assets, asset_folder).with_context(context)?;
+                let mut buf = Vec::new();
+                file.read_to_end(&mut buf).with_context(context)?;
+                (buf, format.to_content_type(&name))
+            }),
+            Body::Empty => None,
+        })
+    }
+}
+
+impl Display for Command {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?} {} ({:?})", self.method, self.route, self.synchronous)
+    }
+}
+
+#[derive(Default, Debug, Clone, Copy, Deserialize)]
+pub enum SyncMode {
+    DontWait,
+    #[default]
+    WaitForResponse,
+    WaitForTask,
+}
+
+pub async fn run_batch(
+    client: &Client,
+    batch: &[Command],
+    assets: &BTreeMap<String, Asset>,
+    asset_folder: &str,
+) -> anyhow::Result<()> {
+    let [.., last] = batch else { return Ok(()) };
+    let sync = last.synchronous;
+
+    let mut tasks = tokio::task::JoinSet::new();
+
+    for command in batch {
+        // FIXME: you probably don't want to copy assets everytime here
+        tasks.spawn({
+            let client = client.clone();
+            let command = command.clone();
+            let assets = assets.clone();
+            let asset_folder = asset_folder.to_owned();
+
+            async move { run(client, command, &assets, &asset_folder).await }
+        });
+    }
+
+    while let Some(result) = tasks.join_next().await {
+        result
+            .context("panicked while executing command")?
+            .context("error while executing command")?;
+    }
+
+    match sync {
+        SyncMode::DontWait => {}
+        SyncMode::WaitForResponse => {}
+        SyncMode::WaitForTask => wait_for_tasks(client).await?,
+    }
+
+    Ok(())
+}
+
+async fn wait_for_tasks(client: &Client) -> anyhow::Result<()> {
+    loop {
+        let response = client
+            .get("tasks?statuses=enqueued,processing")
+            .send()
+            .await
+            .context("could not wait for tasks")?;
+        let response: serde_json::Value = response
+            .json()
+            .await
+            .context("could not deserialize response to JSON")
+            .context("could not wait for tasks")?;
+        match response.get("total") {
+            Some(serde_json::Value::Number(number)) => {
+                let number = number.as_u64().with_context(|| {
+                    format!("waiting for tasks: could not parse 'total' as integer, got {}", number)
+                })?;
+                if number == 0 {
+                    break;
+                } else {
+                    tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                    continue;
+                }
+            }
+            Some(thing_else) => {
+                bail!(format!(
+                    "waiting for tasks: could not parse 'total' as a number, got '{thing_else}'"
+                ))
+            }
+            None => {
+                bail!(format!(
+                    "waiting for tasks: expected response to contain 'total', got '{response}'"
+                ))
+            }
+        }
+    }
+    Ok(())
+}
+
+#[tracing::instrument(skip(client, command, assets, asset_folder), fields(command = %command))]
+pub async fn run(
+    client: Client,
+    mut command: Command,
+    assets: &BTreeMap<String, Asset>,
+    asset_folder: &str,
+) -> anyhow::Result<()> {
+    // memtake the body here to leave an empty body in its place, so that command is not partially moved-out
+    let body = std::mem::take(&mut command.body)
+        .get(assets, asset_folder)
+        .with_context(|| format!("while getting body for command {command}"))?;
+
+    let request = client.request(command.method.into(), &command.route);
+
+    let request = if let Some((body, content_type)) = body {
+        request.body(body).header(reqwest::header::CONTENT_TYPE, content_type)
+    } else {
+        request
+    };
+
+    let response =
+        request.send().await.with_context(|| format!("error sending command: {}", command))?;
+
+    let code = response.status();
+    if code.is_client_error() {
+        tracing::error!(%command, %code, "error in workload file");
+        let response: serde_json::Value = response
+            .json()
+            .await
+            .context("could not deserialize response as JSON")
+            .context("parsing error in workload file when sending command")?;
+        bail!("error in workload file: server responded with error code {code} and '{response}'")
+    } else if code.is_server_error() {
+        tracing::error!(%command, %code, "server error");
+        let response: serde_json::Value = response
+            .json()
+            .await
+            .context("could not deserialize response as JSON")
+            .context("parsing server error when sending command")?;
+        bail!("server error: server responded with error code {code} and '{response}'")
+    }
+
+    Ok(())
+}

xtask/src/bench/dashboard.rs

@@ -0,0 +1,167 @@
+use std::collections::BTreeMap;
+
+use anyhow::{bail, Context};
+use serde_json::json;
+use tokio::signal::ctrl_c;
+use tokio::task::AbortHandle;
+use tracing_trace::processor::span_stats::CallStats;
+use uuid::Uuid;
+
+use super::client::Client;
+use super::env_info;
+use super::workload::Workload;
+
+pub async fn cancel_on_ctrl_c(
+    invocation_uuid: Uuid,
+    dashboard_client: Client,
+    abort_handle: AbortHandle,
+) {
+    tracing::info!("press Ctrl-C to cancel the invocation");
+    match ctrl_c().await {
+        Ok(()) => {
+            tracing::info!(%invocation_uuid, "received Ctrl-C, cancelling invocation");
+            mark_as_failed(dashboard_client, invocation_uuid, None).await;
+            abort_handle.abort();
+        }
+        Err(error) => tracing::warn!(
+            error = &error as &dyn std::error::Error,
+            "failed to listen to Ctrl-C signal, invocation won't be canceled on Ctrl-C"
+        ),
+    }
+}
+
+pub async fn mark_as_failed(
+    dashboard_client: Client,
+    invocation_uuid: Uuid,
+    failure_reason: Option<String>,
+) {
+    let response = dashboard_client
+        .post("cancel-invocation")
+        .json(&json!({
+            "invocation_uuid": invocation_uuid,
+            "failure_reason": failure_reason,
+        }))
+        .send()
+        .await;
+    let response = match response {
+        Ok(response) => response,
+        Err(response_error) => {
+            tracing::error!(error = &response_error as &dyn std::error::Error, %invocation_uuid, "could not mark invocation as failed");
+            return;
+        }
+    };
+
+    if !response.status().is_success() {
+        tracing::error!(
+            %invocation_uuid,
+            "could not mark invocation as failed: {}",
+            response.text().await.unwrap()
+        );
+        return;
+    }
+    tracing::warn!(%invocation_uuid, "marked invocation as failed or canceled");
+}
+
+pub async fn send_machine_info(
+    dashboard_client: &Client,
+    env: &env_info::Environment,
+) -> anyhow::Result<()> {
+    let response = dashboard_client
+        .put("machine")
+        .json(&json!({"hostname": env.hostname}))
+        .send()
+        .await
+        .context("sending machine information")?;
+    if !response.status().is_success() {
+        bail!(
+            "could not send machine information: {} {}",
+            response.status(),
+            response.text().await.unwrap_or_else(|_| "unknown".into())
+        );
+    }
+    Ok(())
+}
+
+pub async fn create_invocation(
+    dashboard_client: &Client,
+    build_info: build_info::BuildInfo,
+    commit_message: &str,
+    env: env_info::Environment,
+    max_workloads: usize,
+    reason: Option<&str>,
+) -> anyhow::Result<Uuid> {
+    let response = dashboard_client
+        .put("invocation")
+        .json(&json!({
+            "commit": {
+                "sha1": build_info.commit_sha1,
+                "message": commit_message,
+                "commit_date": build_info.commit_timestamp,
+                "branch": build_info.branch,
+                "tag": build_info.describe.and_then(|describe| describe.as_tag()),
+            },
+            "machine_hostname": env.hostname,
+            "max_workloads": max_workloads,
+            "reason": reason
+        }))
+        .send()
+        .await
+        .context("sending invocation")?;
+    if !response.status().is_success() {
+        bail!(
+            "could not send new invocation: {}",
+            response.text().await.unwrap_or_else(|_| "unknown".into())
+        );
+    }
+    let invocation_uuid: Uuid =
+        response.json().await.context("could not deserialize invocation response as JSON")?;
+    Ok(invocation_uuid)
+}
+
+pub async fn create_workload(
+    dashboard_client: &Client,
+    invocation_uuid: Uuid,
+    workload: &Workload,
+) -> anyhow::Result<Uuid> {
+    let response = dashboard_client
+        .put("workload")
+        .json(&json!({
+            "invocation_uuid": invocation_uuid,
+            "name": &workload.name,
+            "max_runs": workload.run_count,
+        }))
+        .send()
+        .await
+        .context("could not create new workload")?;
+
+    if !response.status().is_success() {
+        bail!("creating new workload failed: {}", response.text().await.unwrap())
+    }
+
+    let workload_uuid: Uuid =
+        response.json().await.context("could not deserialize JSON as UUID")?;
+    Ok(workload_uuid)
+}
+
+pub async fn create_run(
+    dashboard_client: Client,
+    workload_uuid: Uuid,
+    report: &BTreeMap<String, CallStats>,
+) -> anyhow::Result<()> {
+    let response = dashboard_client
+        .put("run")
+        .json(&json!({
+            "workload_uuid": workload_uuid,
+            "data": report
+        }))
+        .send()
+        .await
+        .context("sending new run")?;
+    if !response.status().is_success() {
+        bail!(
+            "sending new run failed: {}",
+            response.text().await.unwrap_or_else(|_| "unknown".into())
+        )
+    }
+    Ok(())
+}

xtask/src/bench/env_info.rs

@@ -0,0 +1,75 @@
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct Environment {
+    pub hostname: Option<String>,
+    pub cpu: String,
+
+    /// Advertised or nominal clock speed in Hertz.
+    pub clock_speed: u64,
+
+    /// Total number of bytes of memory provided by the system. */
+    pub memory: u64,
+    pub os_type: String,
+    pub software: Vec<VersionInfo>,
+
+    pub user_name: String,
+
+    /// Is set true when the data was gathered by a manual run,
+    /// possibly on a developer machine, instead of the usual benchmark server.
+    pub manual_run: bool,
+}
+
+impl Environment {
+    pub fn generate_from_current_config() -> Self {
+        use sysinfo::System;
+
+        let unknown_string = String::from("Unknown");
+        let mut system = System::new();
+        system.refresh_cpu();
+        system.refresh_cpu_frequency();
+        system.refresh_memory();
+
+        let (cpu, frequency) = match system.cpus().first() {
+            Some(cpu) => (
+                format!("{} @ {:.2}GHz", cpu.brand(), cpu.frequency() as f64 / 1000.0),
+                cpu.frequency() * 1_000_000,
+            ),
+            None => (unknown_string.clone(), 0),
+        };
+
+        let mut software = Vec::new();
+        if let Some(distribution) = System::name() {
+            software
+                .push(VersionInfo { name: distribution, version: String::from("distribution") });
+        }
+        if let Some(kernel) = System::kernel_version() {
+            software.push(VersionInfo { name: kernel, version: String::from("kernel") });
+        }
+        if let Some(os) = System::os_version() {
+            software.push(VersionInfo { name: os, version: String::from("kernel-release") });
+        }
+        if let Some(arch) = System::cpu_arch() {
+            software.push(VersionInfo { name: arch, version: String::from("arch") });
+        }
+
+        Self {
+            hostname: System::host_name(),
+            cpu,
+            clock_speed: frequency,
+            memory: system.total_memory(),
+            os_type: System::long_os_version().unwrap_or(unknown_string.clone()),
+            user_name: System::name().unwrap_or(unknown_string.clone()),
+            manual_run: false,
+            software,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct VersionInfo {
+    pub name: String,
+    pub version: String,
+}

xtask/src/bench/meili_process.rs

@@ -0,0 +1,112 @@
+use std::collections::BTreeMap;
+
+use anyhow::{bail, Context as _};
+
+use super::assets::Asset;
+use super::client::Client;
+use super::workload::Workload;
+
+pub async fn kill(mut meilisearch: tokio::process::Child) {
+    if let Err(error) = meilisearch.kill().await {
+        tracing::warn!(
+            error = &error as &dyn std::error::Error,
+            "while terminating Meilisearch server"
+        )
+    }
+}
+
+#[tracing::instrument]
+pub async fn build() -> anyhow::Result<()> {
+    let mut command = tokio::process::Command::new("cargo");
+    command.arg("build").arg("--release").arg("-p").arg("meilisearch");
+
+    command.kill_on_drop(true);
+
+    let mut builder = command.spawn().context("error building Meilisearch")?;
+
+    if !builder.wait().await.context("could not build Meilisearch")?.success() {
+        bail!("failed building Meilisearch")
+    }
+
+    Ok(())
+}
+
+#[tracing::instrument(skip(client, master_key, workload), fields(workload = workload.name))]
+pub async fn start(
+    client: &Client,
+    master_key: Option<&str>,
+    workload: &Workload,
+    asset_folder: &str,
+) -> anyhow::Result<tokio::process::Child> {
+    let mut command = tokio::process::Command::new("cargo");
+    command
+        .arg("run")
+        .arg("--release")
+        .arg("-p")
+        .arg("meilisearch")
+        .arg("--bin")
+        .arg("meilisearch")
+        .arg("--");
+
+    command.arg("--db-path").arg("./_xtask_benchmark.ms");
+    if let Some(master_key) = master_key {
+        command.arg("--master-key").arg(master_key);
+    }
+    command.arg("--experimental-enable-logs-route");
+
+    for extra_arg in workload.extra_cli_args.iter() {
+        command.arg(extra_arg);
+    }
+
+    command.kill_on_drop(true);
+
+    let mut meilisearch = command.spawn().context("Error starting Meilisearch")?;
+
+    wait_for_health(client, &mut meilisearch, &workload.assets, asset_folder).await?;
+
+    Ok(meilisearch)
+}
+
+async fn wait_for_health(
+    client: &Client,
+    meilisearch: &mut tokio::process::Child,
+    assets: &BTreeMap<String, Asset>,
+    asset_folder: &str,
+) -> anyhow::Result<()> {
+    for i in 0..100 {
+        let res = super::command::run(client.clone(), health_command(), assets, asset_folder).await;
+        if res.is_ok() {
+            // check that this is actually the current Meilisearch instance that answered us
+            if let Some(exit_code) =
+                meilisearch.try_wait().context("cannot check Meilisearch server process status")?
+            {
+                tracing::error!("Got an health response from a different process");
+                bail!("Meilisearch server exited early with code {exit_code}");
+            }
+
+            return Ok(());
+        }
+        tokio::time::sleep(std::time::Duration::from_millis(500)).await;
+        // check whether the Meilisearch instance exited early (cut the wait)
+        if let Some(exit_code) =
+            meilisearch.try_wait().context("cannot check Meilisearch server process status")?
+        {
+            bail!("Meilisearch server exited early with code {exit_code}");
+        }
+        tracing::debug!(attempt = i, "Waiting for Meilisearch to go up");
+    }
+    bail!("meilisearch is not responding")
+}
+
+fn health_command() -> super::command::Command {
+    super::command::Command {
+        route: "/health".into(),
+        method: super::client::Method::Get,
+        body: Default::default(),
+        synchronous: super::command::SyncMode::WaitForResponse,
+    }
+}
+
+pub fn delete_db() {
+    let _ = std::fs::remove_dir_all("./_xtask_benchmark.ms");
+}

xtask/src/bench/mod.rs

@@ -0,0 +1,203 @@
+mod assets;
+mod client;
+mod command;
+mod dashboard;
+mod env_info;
+mod meili_process;
+mod workload;
+
+use std::path::PathBuf;
+
+use anyhow::Context;
+use clap::Parser;
+use tracing_subscriber::fmt::format::FmtSpan;
+use tracing_subscriber::layer::SubscriberExt;
+use tracing_subscriber::Layer;
+
+use self::client::Client;
+use self::workload::Workload;
+
+pub fn default_http_addr() -> String {
+    "127.0.0.1:7700".to_string()
+}
+pub fn default_report_folder() -> String {
+    "./bench/reports/".into()
+}
+
+pub fn default_asset_folder() -> String {
+    "./bench/assets/".into()
+}
+
+pub fn default_log_filter() -> String {
+    "info".into()
+}
+
+pub fn default_dashboard_url() -> String {
+    "http://localhost:9001".into()
+}
+
+/// Run benchmarks from a workload
+#[derive(Parser, Debug)]
+pub struct BenchDeriveArgs {
+    /// Filename of the workload file, pass multiple filenames
+    /// to run multiple workloads in the specified order.
+    ///
+    /// Each workload run will get its own report file.
+    #[arg(value_name = "WORKLOAD_FILE", last = false)]
+    workload_file: Vec<PathBuf>,
+
+    /// URL of the dashboard.
+    #[arg(long, default_value_t = default_dashboard_url())]
+    dashboard_url: String,
+
+    /// Directory to output reports.
+    #[arg(long, default_value_t = default_report_folder())]
+    report_folder: String,
+
+    /// Directory to store the remote assets.
+    #[arg(long, default_value_t = default_asset_folder())]
+    asset_folder: String,
+
+    /// Log directives
+    #[arg(short, long, default_value_t = default_log_filter())]
+    log_filter: String,
+
+    /// Benchmark dashboard API key
+    #[arg(long)]
+    api_key: Option<String>,
+
+    /// Meilisearch master keys
+    #[arg(long)]
+    master_key: Option<String>,
+
+    /// Authentication bearer for fetching assets
+    #[arg(long)]
+    assets_key: Option<String>,
+
+    /// Reason for the benchmark invocation
+    #[arg(short, long)]
+    reason: Option<String>,
+}
+
+pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
+    // setup logs
+    let filter: tracing_subscriber::filter::Targets =
+        args.log_filter.parse().context("invalid --log-filter")?;
+
+    let subscriber = tracing_subscriber::registry().with(
+        tracing_subscriber::fmt::layer()
+            .with_span_events(FmtSpan::NEW | FmtSpan::CLOSE)
+            .with_filter(filter),
+    );
+    tracing::subscriber::set_global_default(subscriber).context("could not setup logging")?;
+
+    // fetch environment and build info
+    let env = env_info::Environment::generate_from_current_config();
+    let build_info = build_info::BuildInfo::from_build();
+
+    // tokio runtime
+    let rt = tokio::runtime::Builder::new_current_thread().enable_io().enable_time().build()?;
+    let _scope = rt.enter();
+
+    // setup clients
+    let assets_client =
+        Client::new(None, args.assets_key.as_deref(), Some(std::time::Duration::from_secs(3600)))?; // 1h
+
+    let dashboard_client = Client::new(
+        Some(format!("{}/api/v1", args.dashboard_url)),
+        args.api_key.as_deref(),
+        Some(std::time::Duration::from_secs(60)),
+    )?;
+
+    // reporting uses its own client because keeping the stream open to wait for entries
+    // blocks any other requests
+    // Also we don't want any pesky timeout because we don't know how much time it will take to recover the full trace
+    let logs_client = Client::new(
+        Some("http://127.0.0.1:7700/logs/stream".into()),
+        args.master_key.as_deref(),
+        None,
+    )?;
+
+    let meili_client = Client::new(
+        Some("http://127.0.0.1:7700".into()),
+        args.master_key.as_deref(),
+        Some(std::time::Duration::from_secs(60)),
+    )?;
+
+    // enter runtime
+
+    rt.block_on(async {
+        dashboard::send_machine_info(&dashboard_client, &env).await?;
+
+        let commit_message = build_info.commit_msg.context("missing commit message")?.split('\n').next().unwrap();
+        let max_workloads = args.workload_file.len();
+        let reason: Option<&str> = args.reason.as_deref();
+        let invocation_uuid = dashboard::create_invocation(&dashboard_client, build_info, commit_message, env, max_workloads, reason).await?;
+
+        tracing::info!(workload_count = args.workload_file.len(), "handling workload files");
+
+        // main task
+        let workload_runs = tokio::spawn(
+            {
+                let dashboard_client = dashboard_client.clone();
+                async move {
+            for workload_file in args.workload_file.iter() {
+                let workload: Workload = serde_json::from_reader(
+                    std::fs::File::open(workload_file)
+                        .with_context(|| format!("error opening {}", workload_file.display()))?,
+                )
+                .with_context(|| format!("error parsing {} as JSON", workload_file.display()))?;
+
+                workload::execute(
+                    &assets_client,
+                    &dashboard_client,
+                    &logs_client,
+                    &meili_client,
+                    invocation_uuid,
+                    args.master_key.as_deref(),
+                    workload,
+                    &args,
+                )
+                .await?;
+            }
+            Ok::<(), anyhow::Error>(())
+        }});
+
+        // handle ctrl-c
+        let abort_handle = workload_runs.abort_handle();
+        tokio::spawn({
+            let dashboard_client = dashboard_client.clone();
+            dashboard::cancel_on_ctrl_c(invocation_uuid, dashboard_client, abort_handle)
+        });
+
+        // wait for the end of the main task, handle result
+        match workload_runs.await {
+            Ok(Ok(_)) => {
+                tracing::info!("Success");
+                Ok::<(), anyhow::Error>(())
+            }
+            Ok(Err(error)) => {
+                tracing::error!(%invocation_uuid, error = %error, "invocation failed, attempting to report the failure to dashboard");
+                dashboard::mark_as_failed(dashboard_client, invocation_uuid, Some(error.to_string())).await;
+                tracing::warn!(%invocation_uuid, "invocation marked as failed following error");
+                Err(error)
+            },
+            Err(join_error) => {
+                match join_error.try_into_panic() {
+                    Ok(panic) => {
+                        tracing::error!("invocation panicked, attempting to report the failure to dashboard");
+                        dashboard::mark_as_failed(dashboard_client, invocation_uuid, Some("Panicked".into())).await;
+                        std::panic::resume_unwind(panic)
+                    }
+                    Err(_) => {
+                        tracing::warn!("task was canceled");
+                        Ok(())
+                    }
+                }
+            },
+        }
+
+    })?;
+
+    Ok(())
+}

xtask/src/bench/workload.rs

@@ -0,0 +1,262 @@
+use std::collections::BTreeMap;
+use std::fs::File;
+use std::io::{Seek as _, Write as _};
+
+use anyhow::{bail, Context as _};
+use futures_util::TryStreamExt as _;
+use serde::Deserialize;
+use serde_json::json;
+use tokio::task::JoinHandle;
+use uuid::Uuid;
+
+use super::assets::Asset;
+use super::client::Client;
+use super::command::SyncMode;
+use super::BenchDeriveArgs;
+use crate::bench::{assets, dashboard, meili_process};
+
+#[derive(Deserialize)]
+pub struct Workload {
+    pub name: String,
+    pub run_count: u16,
+    pub extra_cli_args: Vec<String>,
+    pub assets: BTreeMap<String, Asset>,
+    pub commands: Vec<super::command::Command>,
+}
+
+async fn run_commands(
+    dashboard_client: &Client,
+    logs_client: &Client,
+    meili_client: &Client,
+    workload_uuid: Uuid,
+    workload: &Workload,
+    args: &BenchDeriveArgs,
+    run_number: u16,
+) -> anyhow::Result<JoinHandle<anyhow::Result<File>>> {
+    let report_folder = &args.report_folder;
+    let workload_name = &workload.name;
+
+    std::fs::create_dir_all(report_folder)
+        .with_context(|| format!("could not create report directory at {report_folder}"))?;
+
+    let trace_filename = format!("{report_folder}/{workload_name}-{run_number}-trace.json");
+    let report_filename = format!("{report_folder}/{workload_name}-{run_number}-report.json");
+
+    let report_handle = start_report(logs_client, trace_filename).await?;
+
+    for batch in workload
+        .commands
+        .as_slice()
+        .split_inclusive(|command| !matches!(command.synchronous, SyncMode::DontWait))
+    {
+        super::command::run_batch(meili_client, batch, &workload.assets, &args.asset_folder)
+            .await?;
+    }
+
+    let processor =
+        stop_report(dashboard_client, logs_client, workload_uuid, report_filename, report_handle)
+            .await?;
+
+    Ok(processor)
+}
+
+#[allow(clippy::too_many_arguments)] // not best code quality, but this is a benchmark runner
+#[tracing::instrument(skip(assets_client, dashboard_client, logs_client, meili_client, workload, master_key, args), fields(workload = workload.name))]
+pub async fn execute(
+    assets_client: &Client,
+    dashboard_client: &Client,
+    logs_client: &Client,
+    meili_client: &Client,
+    invocation_uuid: Uuid,
+    master_key: Option<&str>,
+    workload: Workload,
+    args: &BenchDeriveArgs,
+) -> anyhow::Result<()> {
+    assets::fetch_assets(assets_client, &workload.assets, &args.asset_folder).await?;
+
+    let workload_uuid =
+        dashboard::create_workload(dashboard_client, invocation_uuid, &workload).await?;
+
+    let mut tasks = Vec::new();
+
+    for i in 0..workload.run_count {
+        tasks.push(
+            execute_run(
+                dashboard_client,
+                logs_client,
+                meili_client,
+                workload_uuid,
+                master_key,
+                &workload,
+                args,
+                i,
+            )
+            .await?,
+        );
+    }
+
+    let mut reports = Vec::with_capacity(workload.run_count as usize);
+
+    for task in tasks {
+        reports.push(
+            task.await
+                .context("task panicked while processing report")?
+                .context("task failed while processing report")?,
+        );
+    }
+
+    tracing::info!(workload = workload.name, "Successful workload");
+
+    Ok(())
+}
+
+#[allow(clippy::too_many_arguments)] // not best code quality, but this is a benchmark runner
+#[tracing::instrument(skip(dashboard_client, logs_client, meili_client, workload, master_key, args), fields(workload = %workload.name))]
+async fn execute_run(
+    dashboard_client: &Client,
+    logs_client: &Client,
+    meili_client: &Client,
+    workload_uuid: Uuid,
+    master_key: Option<&str>,
+    workload: &Workload,
+    args: &BenchDeriveArgs,
+    run_number: u16,
+) -> anyhow::Result<tokio::task::JoinHandle<anyhow::Result<std::fs::File>>> {
+    meili_process::delete_db();
+
+    meili_process::build().await?;
+    let meilisearch =
+        meili_process::start(meili_client, master_key, workload, &args.asset_folder).await?;
+
+    let processor = run_commands(
+        dashboard_client,
+        logs_client,
+        meili_client,
+        workload_uuid,
+        workload,
+        args,
+        run_number,
+    )
+    .await?;
+
+    meili_process::kill(meilisearch).await;
+
+    tracing::info!(run_number, "Successful run");
+
+    Ok(processor)
+}
+
+async fn start_report(
+    logs_client: &Client,
+    filename: String,
+) -> anyhow::Result<tokio::task::JoinHandle<anyhow::Result<std::fs::File>>> {
+    let report_file = std::fs::File::options()
+        .create(true)
+        .truncate(true)
+        .write(true)
+        .read(true)
+        .open(&filename)
+        .with_context(|| format!("could not create file at {filename}"))?;
+    let mut report_file = std::io::BufWriter::new(report_file);
+
+    let response = logs_client
+        .post("")
+        .json(&json!({
+            "mode": "profile",
+            "target": "indexing::=trace"
+        }))
+        .send()
+        .await
+        .context("failed to start report")?;
+
+    let code = response.status();
+    if code.is_client_error() {
+        tracing::error!(%code, "request error when trying to start report");
+        let response: serde_json::Value = response
+            .json()
+            .await
+            .context("could not deserialize response as JSON")
+            .context("response error when trying to start report")?;
+        bail!(
+            "request error when trying to start report: server responded with error code {code} and '{response}'"
+        )
+    } else if code.is_server_error() {
+        tracing::error!(%code, "server error when trying to start report");
+        let response: serde_json::Value = response
+            .json()
+            .await
+            .context("could not deserialize response as JSON")
+            .context("response error trying to start report")?;
+        bail!("server error when trying to start report: server responded with error code {code} and '{response}'")
+    }
+
+    Ok(tokio::task::spawn(async move {
+        let mut stream = response.bytes_stream();
+        while let Some(bytes) = stream.try_next().await.context("while waiting for report")? {
+            report_file
+                .write_all(&bytes)
+                .with_context(|| format!("while writing report to {filename}"))?;
+        }
+        report_file.into_inner().with_context(|| format!("while writing report to {filename}"))
+    }))
+}
+
+async fn stop_report(
+    dashboard_client: &Client,
+    logs_client: &Client,
+    workload_uuid: Uuid,
+    filename: String,
+    report_handle: tokio::task::JoinHandle<anyhow::Result<std::fs::File>>,
+) -> anyhow::Result<tokio::task::JoinHandle<anyhow::Result<std::fs::File>>> {
+    let response = logs_client.delete("").send().await.context("while stopping report")?;
+    if !response.status().is_success() {
+        bail!("received HTTP {} while stopping report", response.status())
+    }
+
+    let mut file = tokio::time::timeout(std::time::Duration::from_secs(1000), report_handle)
+        .await
+        .context("while waiting for the end of the report")?
+        .context("report writing task panicked")?
+        .context("while writing report")?;
+
+    file.rewind().context("while rewinding report file")?;
+
+    let process_handle = tokio::task::spawn({
+        let dashboard_client = dashboard_client.clone();
+        async move {
+            let span = tracing::info_span!("processing trace to report", filename);
+            let _guard = span.enter();
+            let report = tracing_trace::processor::span_stats::to_call_stats(
+                tracing_trace::TraceReader::new(std::io::BufReader::new(file)),
+            )
+            .context("could not convert trace to report")?;
+            let context = || format!("writing report to {filename}");
+
+            dashboard::create_run(dashboard_client, workload_uuid, &report).await?;
+
+            let mut output_file = std::io::BufWriter::new(
+                std::fs::File::options()
+                    .create(true)
+                    .truncate(true)
+                    .write(true)
+                    .read(true)
+                    .open(&filename)
+                    .with_context(context)?,
+            );
+
+            for (key, value) in report {
+                serde_json::to_writer(&mut output_file, &json!({key: value}))
+                    .context("serializing span stat")?;
+                writeln!(&mut output_file).with_context(context)?;
+            }
+            output_file.flush().with_context(context)?;
+            let mut output_file = output_file.into_inner().with_context(context)?;
+
+            output_file.rewind().context("could not rewind output_file").with_context(context)?;
+
+            Ok(output_file)
+        }
+    });
+
+    Ok(process_handle)
+}

xtask/src/lib.rs

@@ -0,0 +1 @@
+pub mod bench;

xtask/src/main.rs

@@ -1,6 +1,7 @@
 use std::collections::HashSet;
 
 use clap::Parser;
+use xtask::bench::BenchDeriveArgs;
 
 /// List features available in the workspace
 #[derive(Parser, Debug)]
@@ -17,13 +18,16 @@ struct ListFeaturesDeriveArgs {
 #[command(bin_name = "cargo xtask")]
 enum Command {
     ListFeatures(ListFeaturesDeriveArgs),
+    Bench(BenchDeriveArgs),
 }
 
-fn main() {
+fn main() -> anyhow::Result<()> {
     let args = Command::parse();
     match args {
         Command::ListFeatures(args) => list_features(args),
+        Command::Bench(args) => xtask::bench::run(args)?,
     }
+    Ok(())
 }
 
 fn list_features(args: ListFeaturesDeriveArgs) {