Merge e6295c9c5f02df14134322b831c7cd329578a4f1 into 42257eec53e66d6360d3ca61fc1b6106646cc56b

5272: Fix Batches Deletion and flaky tests r=irevoire a=Kerollmops

- This issue fixes #5263 by removing the batches from the date and time databases.
- It also introduces a new `enqueued_at` field in the batch object to quickly retrieve them in the `batches.enqueued_at` database
- Finally, it probably fixes all the flaky tests of the batches: https://github.com/meilisearch/meilisearch/issues/5256

Co-authored-by: Kerollmops <clement@meilisearch.com>
Co-authored-by: Tamo <tamo@meilisearch.com>

crates/index-scheduler/src/insta_snapshot.rs

@@ -1,7 +1,7 @@
 use std::collections::BTreeSet;
 use std::fmt::Write;
 
-use meilisearch_types::batches::Batch;
+use meilisearch_types::batches::{Batch, BatchEnqueuedAt};
 use meilisearch_types::heed::types::{SerdeBincode, SerdeJson, Str};
 use meilisearch_types::heed::{Database, RoTxn};
 use meilisearch_types::milli::{CboRoaringBitmapCodec, RoaringBitmapCodec, BEU32};
@@ -341,10 +341,14 @@ pub fn snapshot_canceled_by(rtxn: &RoTxn, db: Database<BEU32, RoaringBitmapCodec
 
 pub fn snapshot_batch(batch: &Batch) -> String {
     let mut snap = String::new();
-    let Batch { uid, details, stats, started_at, finished_at, progress: _ } = batch;
+    let Batch { uid, details, stats, started_at, finished_at, progress: _, enqueued_at } = batch;
     if let Some(finished_at) = finished_at {
         assert!(finished_at > started_at);
     }
+    let BatchEnqueuedAt { earliest, oldest } = enqueued_at.unwrap();
+    assert!(*started_at > earliest);
+    assert!(earliest >= oldest);
+
     snap.push('{');
     snap.push_str(&format!("uid: {uid}, "));
     snap.push_str(&format!("details: {}, ", serde_json::to_string(details).unwrap()));

crates/index-scheduler/src/queue/batches.rs

@@ -12,8 +12,8 @@ use time::OffsetDateTime;
 use super::{Query, Queue};
 use crate::processing::ProcessingTasks;
 use crate::utils::{
-    insert_task_datetime, keep_ids_within_datetimes, map_bound, remove_task_datetime,
-    ProcessingBatch,
+    insert_task_datetime, keep_ids_within_datetimes, map_bound,
+    remove_n_tasks_datetime_earlier_than, remove_task_datetime, ProcessingBatch,
 };
 use crate::{Error, Result, BEI128};
 
@@ -181,6 +181,7 @@ impl BatchQueue {
                 stats: batch.stats,
                 started_at: batch.started_at,
                 finished_at: batch.finished_at,
+                enqueued_at: batch.enqueued_at,
             },
         )?;
 
@@ -234,34 +235,25 @@ impl BatchQueue {
         // What we know, though, is that the task date is from before the enqueued_at, and max two timestamps have been written
         // to the DB per batches.
         if let Some(ref old_batch) = old_batch {
-            let started_at = old_batch.started_at.unix_timestamp_nanos();
-
-            // We have either one or two enqueued at to remove
-            let mut exit = old_batch.stats.total_nb_tasks.clamp(0, 2);
-            let mut iterator = self.enqueued_at.rev_iter_mut(wtxn)?;
-            while let Some(entry) = iterator.next() {
-                let (key, mut value) = entry?;
-                if key > started_at {
-                    continue;
-                }
-                if value.remove(old_batch.uid) {
-                    exit = exit.saturating_sub(1);
-                    // Safe because the key and value are owned
-                    unsafe {
-                        iterator.put_current(&key, &value)?;
-                    }
-                    if exit == 0 {
-                        break;
-                    }
-                }
+            if let Some(enqueued_at) = old_batch.enqueued_at {
+                remove_task_datetime(wtxn, self.enqueued_at, enqueued_at.earliest, old_batch.uid)?;
+                remove_task_datetime(wtxn, self.enqueued_at, enqueued_at.oldest, old_batch.uid)?;
+            } else {
+                // If we don't have the enqueued at in the batch it means the database comes from the v1.12
+                // and we still need to find the date by scrolling the database
+                remove_n_tasks_datetime_earlier_than(
+                    wtxn,
+                    self.enqueued_at,
+                    old_batch.started_at,
+                    old_batch.stats.total_nb_tasks.clamp(1, 2) as usize,
+                    old_batch.uid,
+                )?;
             }
         }
-        if let Some(enqueued_at) = batch.oldest_enqueued_at {
-            insert_task_datetime(wtxn, self.enqueued_at, enqueued_at, batch.uid)?;
-        }
-        if let Some(enqueued_at) = batch.earliest_enqueued_at {
-            insert_task_datetime(wtxn, self.enqueued_at, enqueued_at, batch.uid)?;
-        }
+        // A finished batch MUST contains at least one task and have an enqueued_at
+        let enqueued_at = batch.enqueued_at.as_ref().unwrap();
+        insert_task_datetime(wtxn, self.enqueued_at, enqueued_at.earliest, batch.uid)?;
+        insert_task_datetime(wtxn, self.enqueued_at, enqueued_at.oldest, batch.uid)?;
 
         // Update the started at and finished at
         if let Some(ref old_batch) = old_batch {

crates/index-scheduler/src/queue/batches_test.rs

@@ -102,30 +102,33 @@ fn query_batches_simple() {
         .unwrap();
     assert_eq!(batches.len(), 1);
     batches[0].started_at = OffsetDateTime::UNIX_EPOCH;
+    assert!(batches[0].enqueued_at.is_some());
+    batches[0].enqueued_at = None;
     // Insta cannot snapshot our batches because the batch stats contains an enum as key: https://github.com/mitsuhiko/insta/issues/689
     let batch = serde_json::to_string_pretty(&batches[0]).unwrap();
     snapshot!(batch, @r#"
-        {
-          "uid": 0,
-          "details": {
-            "primaryKey": "mouse"
-          },
-          "stats": {
-            "totalNbTasks": 1,
-            "status": {
-              "processing": 1
-            },
-            "types": {
-              "indexCreation": 1
-            },
-            "indexUids": {
-              "catto": 1
-            }
-          },
-          "startedAt": "1970-01-01T00:00:00Z",
-          "finishedAt": null
+    {
+      "uid": 0,
+      "details": {
+        "primaryKey": "mouse"
+      },
+      "stats": {
+        "totalNbTasks": 1,
+        "status": {
+          "processing": 1
+        },
+        "types": {
+          "indexCreation": 1
+        },
+        "indexUids": {
+          "catto": 1
         }
-        "#);
+      },
+      "startedAt": "1970-01-01T00:00:00Z",
+      "finishedAt": null,
+      "enqueuedAt": null
+    }
+    "#);
 
     let query = Query { statuses: Some(vec![Status::Enqueued]), ..Default::default() };
     let (batches, _) = index_scheduler

crates/index-scheduler/src/scheduler/process_batch.rs

@@ -2,7 +2,7 @@ use std::collections::{BTreeSet, HashMap, HashSet};
 use std::panic::{catch_unwind, AssertUnwindSafe};
 use std::sync::atomic::Ordering;
 
-use meilisearch_types::batches::BatchId;
+use meilisearch_types::batches::{BatchEnqueuedAt, BatchId};
 use meilisearch_types::heed::{RoTxn, RwTxn};
 use meilisearch_types::milli::progress::{Progress, VariableNameStep};
 use meilisearch_types::milli::{self};
@@ -16,7 +16,10 @@ use crate::processing::{
     InnerSwappingTwoIndexes, SwappingTheIndexes, TaskCancelationProgress, TaskDeletionProgress,
     UpdateIndexProgress,
 };
-use crate::utils::{self, swap_index_uid_in_task, ProcessingBatch};
+use crate::utils::{
+    self, remove_n_tasks_datetime_earlier_than, remove_task_datetime, swap_index_uid_in_task,
+    ProcessingBatch,
+};
 use crate::{Error, IndexScheduler, Result, TaskId};
 
 impl IndexScheduler {
@@ -418,7 +421,6 @@ impl IndexScheduler {
         to_delete_tasks -= &enqueued_tasks;
 
         // 2. We now have a list of tasks to delete, delete them
-
         let mut affected_indexes = HashSet::new();
         let mut affected_statuses = HashSet::new();
         let mut affected_kinds = HashSet::new();
@@ -515,9 +517,51 @@ impl IndexScheduler {
                 tasks -= &to_delete_tasks;
                 // We must remove the batch entirely
                 if tasks.is_empty() {
-                    self.queue.batches.all_batches.delete(wtxn, &batch_id)?;
-                    self.queue.batch_to_tasks_mapping.delete(wtxn, &batch_id)?;
+                    if let Some(batch) = self.queue.batches.get_batch(wtxn, batch_id)? {
+                        if let Some(BatchEnqueuedAt { earliest, oldest }) = batch.enqueued_at {
+                            remove_task_datetime(
+                                wtxn,
+                                self.queue.batches.enqueued_at,
+                                earliest,
+                                batch_id,
+                            )?;
+                            remove_task_datetime(
+                                wtxn,
+                                self.queue.batches.enqueued_at,
+                                oldest,
+                                batch_id,
+                            )?;
+                        } else {
+                            // If we don't have the enqueued at in the batch it means the database comes from the v1.12
+                            // and we still need to find the date by scrolling the database
+                            remove_n_tasks_datetime_earlier_than(
+                                wtxn,
+                                self.queue.batches.enqueued_at,
+                                batch.started_at,
+                                batch.stats.total_nb_tasks.clamp(1, 2) as usize,
+                                batch_id,
+                            )?;
+                        }
+                        remove_task_datetime(
+                            wtxn,
+                            self.queue.batches.started_at,
+                            batch.started_at,
+                            batch_id,
+                        )?;
+                        if let Some(finished_at) = batch.finished_at {
+                            remove_task_datetime(
+                                wtxn,
+                                self.queue.batches.finished_at,
+                                finished_at,
+                                batch_id,
+                            )?;
+                        }
+
+                        self.queue.batches.all_batches.delete(wtxn, &batch_id)?;
+                        self.queue.batch_to_tasks_mapping.delete(wtxn, &batch_id)?;
+                    }
                 }
+
                 // Anyway, we must remove the batch from all its reverse indexes.
                 // The only way to do that is to check
 

crates/index-scheduler/src/scheduler/snapshots/test.rs/task_deletion_delete_same_task_twice/task_deletion_processed.snap

@@ -56,16 +56,13 @@ succeeded [1,]
 ### Batches Index Tasks:
 ----------------------------------------------------------------------
 ### Batches Enqueued At:
-[timestamp] [0,]
 [timestamp] [1,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### Batches Started At:
-[timestamp] [0,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### Batches Finished At:
-[timestamp] [0,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### File Store:

crates/index-scheduler/src/scheduler/snapshots/test.rs/task_deletion_deleteable/task_deletion_processed.snap

@@ -54,15 +54,12 @@ succeeded [1,]
 ### Batches Index Tasks:
 ----------------------------------------------------------------------
 ### Batches Enqueued At:
-[timestamp] [0,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### Batches Started At:
-[timestamp] [0,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### Batches Finished At:
-[timestamp] [0,]
 [timestamp] [1,]
 ----------------------------------------------------------------------
 ### File Store:

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/after_removing_the_upgrade_tasks.snap

@@ -87,7 +87,6 @@ doggo [2,3,]
 girafo [4,]
 ----------------------------------------------------------------------
 ### Batches Enqueued At:
-[timestamp] [0,]
 [timestamp] [1,]
 [timestamp] [2,]
 [timestamp] [3,]
@@ -95,7 +94,6 @@ girafo [4,]
 [timestamp] [5,]
 ----------------------------------------------------------------------
 ### Batches Started At:
-[timestamp] [0,]
 [timestamp] [1,]
 [timestamp] [2,]
 [timestamp] [3,]
@@ -103,7 +101,6 @@ girafo [4,]
 [timestamp] [5,]
 ----------------------------------------------------------------------
 ### Batches Finished At:
-[timestamp] [0,]
 [timestamp] [1,]
 [timestamp] [2,]
 [timestamp] [3,]

crates/index-scheduler/src/utils.rs

@@ -3,7 +3,7 @@
 use std::collections::{BTreeSet, HashSet};
 use std::ops::Bound;
 
-use meilisearch_types::batches::{Batch, BatchId, BatchStats};
+use meilisearch_types::batches::{Batch, BatchEnqueuedAt, BatchId, BatchStats};
 use meilisearch_types::heed::{Database, RoTxn, RwTxn};
 use meilisearch_types::milli::CboRoaringBitmapCodec;
 use meilisearch_types::task_view::DetailsView;
@@ -30,8 +30,7 @@ pub struct ProcessingBatch {
     pub kinds: HashSet<Kind>,
     pub indexes: HashSet<String>,
     pub canceled_by: HashSet<TaskId>,
-    pub oldest_enqueued_at: Option<OffsetDateTime>,
-    pub earliest_enqueued_at: Option<OffsetDateTime>,
+    pub enqueued_at: Option<BatchEnqueuedAt>,
     pub started_at: OffsetDateTime,
     pub finished_at: Option<OffsetDateTime>,
 }
@@ -51,8 +50,7 @@ impl ProcessingBatch {
             kinds: HashSet::default(),
             indexes: HashSet::default(),
             canceled_by: HashSet::default(),
-            oldest_enqueued_at: None,
-            earliest_enqueued_at: None,
+            enqueued_at: None,
             started_at: OffsetDateTime::now_utc(),
             finished_at: None,
         }
@@ -80,14 +78,18 @@ impl ProcessingBatch {
             if let Some(canceled_by) = task.canceled_by {
                 self.canceled_by.insert(canceled_by);
             }
-            self.oldest_enqueued_at =
-                Some(self.oldest_enqueued_at.map_or(task.enqueued_at, |oldest_enqueued_at| {
-                    task.enqueued_at.min(oldest_enqueued_at)
-                }));
-            self.earliest_enqueued_at =
-                Some(self.earliest_enqueued_at.map_or(task.enqueued_at, |earliest_enqueued_at| {
-                    task.enqueued_at.max(earliest_enqueued_at)
-                }));
+            match self.enqueued_at.as_mut() {
+                Some(BatchEnqueuedAt { earliest, oldest }) => {
+                    *oldest = task.enqueued_at.min(*oldest);
+                    *earliest = task.enqueued_at.max(*earliest);
+                }
+                None => {
+                    self.enqueued_at = Some(BatchEnqueuedAt {
+                        earliest: task.enqueued_at,
+                        oldest: task.enqueued_at,
+                    });
+                }
+            }
         }
     }
 
@@ -138,6 +140,7 @@ impl ProcessingBatch {
             stats: self.stats.clone(),
             started_at: self.started_at,
             finished_at: self.finished_at,
+            enqueued_at: self.enqueued_at,
         }
     }
 }
@@ -174,6 +177,33 @@ pub(crate) fn remove_task_datetime(
     Ok(())
 }
 
+pub(crate) fn remove_n_tasks_datetime_earlier_than(
+    wtxn: &mut RwTxn,
+    database: Database<BEI128, CboRoaringBitmapCodec>,
+    earlier_than: OffsetDateTime,
+    mut count: usize,
+    task_id: TaskId,
+) -> Result<()> {
+    let earlier_than = earlier_than.unix_timestamp_nanos();
+    let mut iter = database.rev_range_mut(wtxn, &(..earlier_than))?;
+    while let Some((current, mut existing)) = iter.next().transpose()? {
+        count -= existing.remove(task_id) as usize;
+
+        if existing.is_empty() {
+            // safety: We don't keep references to the database
+            unsafe { iter.del_current()? };
+        } else {
+            // safety: We don't keep references to the database
+            unsafe { iter.put_current(&current, &existing)? };
+        }
+        if count == 0 {
+            break;
+        }
+    }
+
+    Ok(())
+}
+
 pub(crate) fn keep_ids_within_datetimes(
     rtxn: &RoTxn,
     ids: &mut RoaringBitmap,
@@ -329,14 +359,27 @@ impl crate::IndexScheduler {
                 kind,
             } = task;
             assert_eq!(uid, task.uid);
-            if let Some(ref batch) = batch_uid {
+            if task.status != Status::Enqueued {
+                let batch_uid = batch_uid.expect("All non enqueued tasks must be part of a batch");
                 assert!(self
                     .queue
                     .batch_to_tasks_mapping
-                    .get(&rtxn, batch)
+                    .get(&rtxn, &batch_uid)
                     .unwrap()
                     .unwrap()
                     .contains(uid));
+                let batch = self.queue.batches.get_batch(&rtxn, batch_uid).unwrap().unwrap();
+                assert_eq!(batch.uid, batch_uid);
+                if task.status == Status::Processing {
+                    assert!(batch.progress.is_some());
+                } else {
+                    assert!(batch.progress.is_none());
+                }
+                assert_eq!(batch.started_at, task.started_at.unwrap());
+                assert_eq!(batch.finished_at, task.finished_at);
+                let enqueued_at = batch.enqueued_at.unwrap();
+                assert!(task.enqueued_at >= enqueued_at.oldest);
+                assert!(task.enqueued_at <= enqueued_at.earliest);
             }
             if let Some(task_index_uid) = &task_index_uid {
                 assert!(self

crates/meilisearch-types/src/batches.rs

@@ -24,6 +24,18 @@ pub struct Batch {
     pub started_at: OffsetDateTime,
     #[serde(with = "time::serde::rfc3339::option")]
     pub finished_at: Option<OffsetDateTime>,
+
+    // Enqueued at is never displayed and is only required when removing a batch.
+    // It's always some except when upgrading from a database pre v1.12
+    pub enqueued_at: Option<BatchEnqueuedAt>,
+}
+
+#[derive(Clone, Copy, Debug, Serialize, Deserialize)]
+pub struct BatchEnqueuedAt {
+    #[serde(with = "time::serde::rfc3339")]
+    pub earliest: OffsetDateTime,
+    #[serde(with = "time::serde::rfc3339")]
+    pub oldest: OffsetDateTime,
 }
 
 #[derive(Default, Debug, Clone, Serialize, Deserialize, ToSchema)]

crates/meilisearch/tests/batches/mod.rs

@@ -41,9 +41,8 @@ async fn list_batches() {
     let index = server.index("test");
     let (task, _status_code) = index.create(None).await;
     index.wait_task(task.uid()).await.succeeded();
-    index
-        .add_documents(serde_json::from_str(include_str!("../assets/test_set.json")).unwrap(), None)
-        .await;
+    let (task, _status_code) = index.create(None).await;
+    index.wait_task(task.uid()).await.failed();
     let (response, code) = index.list_batches().await;
     assert_eq!(code, 200);
     assert_eq!(
@@ -96,11 +95,12 @@ async fn list_batches_pagination_and_reverse() {
 async fn list_batches_with_star_filters() {
     let server = Server::new().await;
     let index = server.index("test");
-    let (batch, _code) = index.create(None).await;
-    index.wait_task(batch.uid()).await.succeeded();
-    index
-        .add_documents(serde_json::from_str(include_str!("../assets/test_set.json")).unwrap(), None)
-        .await;
+    let (task, _code) = index.create(None).await;
+    index.wait_task(task.uid()).await.succeeded();
+    let index = server.index("test");
+    let (task, _code) = index.create(None).await;
+    index.wait_task(task.uid()).await.failed();
+
     let (response, code) = index.service.get("/batches?indexUids=test").await;
     assert_eq!(code, 200);
     assert_eq!(response["results"].as_array().unwrap().len(), 2);
@@ -187,9 +187,6 @@ async fn list_batches_invalid_canceled_by_filter() {
     let index = server.index("test");
     let (task, _status_code) = index.create(None).await;
     index.wait_task(task.uid()).await.succeeded();
-    index
-        .add_documents(serde_json::from_str(include_str!("../assets/test_set.json")).unwrap(), None)
-        .await;
 
     let (response, code) = index.filtered_batches(&[], &[], &["0"]).await;
     assert_eq!(code, 200, "{}", response);
@@ -202,9 +199,8 @@ async fn list_batches_status_and_type_filtered() {
     let index = server.index("test");
     let (task, _status_code) = index.create(None).await;
     index.wait_task(task.uid()).await.succeeded();
-    index
-        .add_documents(serde_json::from_str(include_str!("../assets/test_set.json")).unwrap(), None)
-        .await;
+    let (task, _status_code) = index.update(Some("id")).await;
+    index.wait_task(task.uid()).await.succeeded();
 
     let (response, code) = index.filtered_batches(&["indexCreation"], &["failed"], &[]).await;
     assert_eq!(code, 200, "{}", response);
@@ -212,7 +208,7 @@ async fn list_batches_status_and_type_filtered() {
 
     let (response, code) = index
         .filtered_batches(
-            &["indexCreation", "documentAdditionOrUpdate"],
+            &["indexCreation", "IndexUpdate"],
             &["succeeded", "processing", "enqueued"],
             &[],
         )

crates/meilitool/src/main.rs

@@ -1,6 +1,7 @@
 use std::fs::{read_dir, read_to_string, remove_file, File};
 use std::io::BufWriter;
 use std::path::PathBuf;
+use std::time::Instant;
 
 use anyhow::Context;
 use clap::{Parser, Subcommand};
@@ -8,7 +9,9 @@ use dump::{DumpWriter, IndexMetadata};
 use file_store::FileStore;
 use meilisearch_auth::AuthController;
 use meilisearch_types::heed::types::{SerdeJson, Str};
-use meilisearch_types::heed::{Database, Env, EnvOpenOptions, RoTxn, RwTxn, Unspecified};
+use meilisearch_types::heed::{
+    CompactionOption, Database, Env, EnvOpenOptions, RoTxn, RwTxn, Unspecified,
+};
 use meilisearch_types::milli::documents::{obkv_to_object, DocumentsBatchReader};
 use meilisearch_types::milli::{obkv_to_json, BEU32};
 use meilisearch_types::tasks::{Status, Task};
@@ -78,6 +81,27 @@ enum Command {
         #[arg(long)]
         target_version: String,
     },
+
+    /// Compact the index by using LMDB.
+    ///
+    /// You must run this command while Meilisearch is off. The reason is that Meilisearch keep the
+    /// indexes opened and this compaction operation writes into another file. Meilisearch will not
+    /// switch to the new file.
+    ///
+    /// **Another possibility** is to keep Meilisearch running to serve search requests, run the
+    /// compaction and once done, close and immediately reopen Meilisearch. This way Meilisearch
+    /// will reopened the data.mdb file when rebooting and see the newly compacted file, ignoring
+    /// the previous non-compacted data.
+    ///
+    /// Note that the compaction will open the index, copy and compact the index into another file
+    /// **on the same disk as the index** and replace the previous index with the newly compacted
+    /// one. Which means that the disk must have enough room for at most two time the index size.
+    ///
+    /// To make sure not to loose any data, this tool takes a mutable transaction on the index
+    /// before running the copy and compaction. This way the current indexation must finish before
+    /// the compaction operation can start. Once the compaction is done, the big index is replaced
+    /// by the compacted one and the mutable transaction is released.
+    CompactIndex { index_name: String },
 }
 
 fn main() -> anyhow::Result<()> {
@@ -94,6 +118,7 @@ fn main() -> anyhow::Result<()> {
             let target_version = parse_version(&target_version).context("While parsing `--target-version`. Make sure `--target-version` is in the format MAJOR.MINOR.PATCH")?;
             OfflineUpgrade { db_path, current_version: detected_version, target_version }.upgrade()
         }
+        Command::CompactIndex { index_name } => compact_index(db_path, &index_name),
     }
 }
 
@@ -347,3 +372,73 @@ fn export_a_dump(
 
     Ok(())
 }
+
+fn compact_index(db_path: PathBuf, index_name: &str) -> anyhow::Result<()> {
+    let index_scheduler_path = db_path.join("tasks");
+    let env = unsafe { EnvOpenOptions::new().max_dbs(100).open(&index_scheduler_path) }
+        .with_context(|| format!("While trying to open {:?}", index_scheduler_path.display()))?;
+
+    let rtxn = env.read_txn()?;
+    let index_mapping: Database<Str, UuidCodec> =
+        try_opening_database(&env, &rtxn, "index-mapping")?;
+
+    for result in index_mapping.iter(&rtxn)? {
+        let (uid, uuid) = result?;
+
+        if uid != index_name {
+            eprintln!("Found index {uid} and skipping it");
+            continue;
+        } else {
+            eprintln!("Found index {uid} 🎉");
+        }
+
+        let index_path = db_path.join("indexes").join(uuid.to_string());
+        let index = Index::new(EnvOpenOptions::new(), &index_path).with_context(|| {
+            format!("While trying to open the index at path {:?}", index_path.display())
+        })?;
+
+        eprintln!("Awaiting for a mutable transaction...");
+        let _wtxn = index.write_txn().context("While awaiting for a write transaction")?;
+
+        // We create and immediately drop the file because the
+        let non_compacted_index_file_path = index_path.join("data.mdb");
+        let compacted_index_file_path = index_path.join("data.mdb.cpy");
+
+        eprintln!("Compacting the index...");
+        let before_compaction = Instant::now();
+        let new_file = index
+            .copy_to_file(&compacted_index_file_path, CompactionOption::Enabled)
+            .with_context(|| format!("While compacting {}", compacted_index_file_path.display()))?;
+
+        let after_size = new_file.metadata()?.len();
+        let before_size = std::fs::metadata(&non_compacted_index_file_path)
+            .with_context(|| {
+                format!(
+                    "While retrieving the metadata of {}",
+                    non_compacted_index_file_path.display(),
+                )
+            })?
+            .len();
+
+        let reduction = before_size as f64 / after_size as f64;
+        println!("Compaction successful. Took around {:.2?}", before_compaction.elapsed());
+        eprintln!("The index went from {before_size} bytes to {after_size} bytes ({reduction:.2}x reduction)");
+
+        eprintln!("Replacing the non-compacted index by the compacted one...");
+        std::fs::rename(&compacted_index_file_path, &non_compacted_index_file_path).with_context(
+            || {
+                format!(
+                    "While renaming {} into {}",
+                    compacted_index_file_path.display(),
+                    non_compacted_index_file_path.display(),
+                )
+            },
+        )?;
+
+        drop(new_file);
+
+        println!("Everything's done 🎉");
+    }
+
+    Ok(())
+}

workloads/embeddings-movies-subset-hf.json

@@ -12,16 +12,6 @@
     }
   },
   "precommands": [
-    {
-      "route": "experimental-features",
-      "method": "PATCH",
-      "body": {
-        "inline": {
-          "vectorStore": true
-        }
-      },
-      "synchronous": "DontWait"
-    },
     {
       "route": "indexes/movies/settings",
       "method": "PATCH",

workloads/embeddings-settings-add.json

@@ -12,16 +12,6 @@
     }
   },
   "precommands": [
-    {
-      "route": "experimental-features",
-      "method": "PATCH",
-      "body": {
-        "inline": {
-          "vectorStore": true
-        }
-      },
-      "synchronous": "DontWait"
-    },
     {
       "route": "indexes/movies/settings",
       "method": "PATCH",

workloads/search/embeddings-movies-subset-hf.json

@@ -13,16 +13,6 @@
     }
   },
   "precommands": [
-    {
-      "route": "experimental-features",
-      "method": "PATCH",
-      "body": {
-        "inline": {
-          "vectorStore": true
-        }
-      },
-      "synchronous": "DontWait"
-    },
     {
       "route": "indexes/movies/settings",
       "method": "PATCH",