move the leader to a new design

Cargo.lock

@@ -523,6 +523,17 @@ version = "3.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "572f695136211188308f16ad2ca5c851a712c464060ae6974944458eb83880ba"
 
+[[package]]
+name = "bus"
+version = "2.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "80cb4625f5b60155ff1018c9d4ce2e38bf5ae3e5780dfab9fa68bb44a6b751e2"
+dependencies = [
+ "crossbeam-channel",
+ "num_cpus",
+ "parking_lot_core",
+]
+
 [[package]]
 name = "byte-unit"
 version = "4.0.18"
@@ -795,6 +806,8 @@ dependencies = [
 name = "cluster"
 version = "1.1.0"
 dependencies = [
+ "bus",
+ "crossbeam",
  "ductile",
  "log",
  "meilisearch-types",

cluster/Cargo.toml

@@ -18,3 +18,5 @@ thiserror = "1.0.39"
 meilisearch-types = { path = "../meilisearch-types" }
 roaring = "0.10.1"
 log = "0.4.17"
+crossbeam = "0.8.2"
+bus = "2.3.0"

cluster/src/leader.rs

@@ -1,111 +1,151 @@
 use std::net::ToSocketAddrs;
-use std::time::Duration;
+use std::sync::atomic::AtomicUsize;
+use std::sync::{atomic, Arc, Mutex};
 
+use bus::{Bus, BusReader};
+use crossbeam::channel::{unbounded, Receiver, Sender};
 use ductile::{ChannelReceiver, ChannelSender, ChannelServer};
-use serde::de::DeserializeOwned;
+use log::info;
-use serde::Serialize;
 
-use crate::{Consistency, Error, FollowerMsg, LeaderMsg};
+use crate::{Consistency, FollowerMsg, LeaderMsg};
 
 pub struct Leader {
-    listener: ChannelServer<LeaderMsg, FollowerMsg>,
+    task_ready_to_commit: Receiver<u32>,
-    active_followers: Vec<Follower>,
+    broadcast_to_follower: Sender<LeaderMsg>,
-    new_followers: Vec<Follower>,
+
-    dead_followers: Vec<Follower>,
+    cluster_size: Arc<AtomicUsize>,
 
     batch_id: u32,
-    tick: Duration,
-}
-
-struct Follower {
-    sender: ChannelSender<LeaderMsg>,
-    receiver: ChannelReceiver<FollowerMsg>,
 }
 
 impl Leader {
-    pub fn new(listen_on: impl ToSocketAddrs) -> Leader {
+    pub fn new(listen_on: impl ToSocketAddrs + Send + 'static) -> Leader {
-        let listener = ChannelServer::bind(listen_on).unwrap();
+        let cluster_size = Arc::new(AtomicUsize::new(1));
+        let (process_batch_sender, process_batch_receiver) = unbounded();
+        let (task_finished_sender, task_finished_receiver) = unbounded();
+
+        let cs = cluster_size.clone();
+        std::thread::spawn(move || {
+            Self::listener(listen_on, cs, process_batch_receiver, task_finished_sender)
+        });
 
         Leader {
-            listener,
+            task_ready_to_commit: task_finished_receiver,
-            active_followers: Vec::new(),
+            broadcast_to_follower: process_batch_sender,
-            new_followers: Vec::new(),
+            cluster_size,
-            dead_followers: Vec::new(),
             batch_id: 0,
-            tick: Duration::new(1, 0),
         }
     }
 
-    pub fn starts_batch(&mut self, batch: Vec<u32>) -> Result<(), Error> {
+    /// Takes all the necessary channels to chat with the scheduler and give them
-        let mut dead_nodes = Vec::new();
+    /// to each new followers
+    fn listener(
+        listen_on: impl ToSocketAddrs,
+        cluster_size: Arc<AtomicUsize>,
+        broadcast_to_follower: Receiver<LeaderMsg>,
+        task_finished: Sender<u32>,
+    ) {
+        let listener: ChannelServer<LeaderMsg, FollowerMsg> =
+            ChannelServer::bind(listen_on).unwrap();
 
-        for (idx, follower) in self.active_followers.iter_mut().enumerate() {
+        // We're going to broadcast all the batches to all our follower
-            match follower
+        let bus: Bus<LeaderMsg> = Bus::new(10);
-                .sender
+        let bus = Arc::new(Mutex::new(bus));
-                .send(LeaderMsg::StartBatch { id: self.batch_id, batch: batch.clone() })
+        let b = bus.clone();
+
+        std::thread::spawn(move || loop {
+            let msg = broadcast_to_follower.recv().expect("Main thread is dead");
+            b.lock().unwrap().broadcast(msg);
+        });
+
+        for (sender, receiver, _addr) in listener {
+            let task_finished = task_finished.clone();
+            let cs = cluster_size.clone();
+
+            let process_batch = bus.lock().unwrap().add_rx();
+
+            std::thread::spawn(move || {
+                Self::follower(sender, receiver, cs, process_batch, task_finished)
+            });
+        }
+    }
+
+    /// Allow a follower to chat with the scheduler
+    fn follower(
+        sender: ChannelSender<LeaderMsg>,
+        receiver: ChannelReceiver<FollowerMsg>,
+        cluster_size: Arc<AtomicUsize>,
+        mut broadcast_to_follower: BusReader<LeaderMsg>,
+        task_finished: Sender<u32>,
+    ) {
+        let size = cluster_size.fetch_add(1, atomic::Ordering::Relaxed) + 1;
+        info!("A new follower joined the cluster. {} members.", size);
+
+        // send messages to the follower
+        std::thread::spawn(move || loop {
+            let msg = broadcast_to_follower.recv().expect("Main thread died");
+            if sender.send(msg).is_err() {
+                // the follower died, the logging and cluster size update should be done
+                // in the other thread
+                break;
+            }
+        });
+
+        // receive messages from the follower
+        loop {
+            match receiver.recv() {
+                Err(_) => break,
+                Ok(msg) => match msg {
+                    FollowerMsg::ReadyToCommit(id) => {
+                        task_finished.send(id).expect("Can't reach the main thread")
+                    }
+                    FollowerMsg::RegisterNewTask(_) => todo!(),
+                },
+            }
+        }
+
+        // if we exited from the previous loop it means the follower is down and should
+        // be removed from the cluster
+        let size = cluster_size.fetch_sub(1, atomic::Ordering::Relaxed) - 1;
+        info!("A follower left the cluster. {} members.", size);
+    }
+
+    pub fn starts_batch(&mut self, batch: Vec<u32>) {
+        assert!(
+            self.batch_id % 2 == 0,
+            "Tried to start processing a batch before commiting the previous one"
+        );
+        self.batch_id += 1;
+
+        self.broadcast_to_follower
+            .send(LeaderMsg::StartBatch { id: self.batch_id, batch })
+            .expect("Can't reach the cluster");
+    }
+
+    pub fn commit(&mut self, consistency_level: Consistency) {
+        // if zero nodes needs to be sync we can commit right away and early exit
+        if consistency_level != Consistency::Zero {
+            // else, we wait till enough nodes are ready to commit
+            for (ready_to_commit, _id) in self
+                .task_ready_to_commit
+                .iter()
+                // we need to filter out the messages from the old batches
+                .filter(|id| *id == self.batch_id)
+                .enumerate()
             {
-                Ok(_) => (),
+                let cluster_size = self.cluster_size.load(atomic::Ordering::Relaxed);
-                // if a node can't be joined we consider it as dead
-                Err(_) => dead_nodes.push(idx),
-            }
-        }
 
-        // we do it from the end so the indices stays correct while removing elements
+                match consistency_level {
-        for dead_node in dead_nodes.into_iter().rev() {
+                    Consistency::One if ready_to_commit >= 1 => break,
-            let dead = self.active_followers.swap_remove(dead_node);
+                    Consistency::Two if ready_to_commit >= 2 => break,
-            self.dead_followers.push(dead);
+                    Consistency::Quorum if ready_to_commit >= (cluster_size / 2) => break,
-        }
+                    _ => (),
-
-        Ok(())
-    }
-
-    pub fn commit(&mut self, consistency_level: Consistency) -> Result<(), Error> {
-        let mut dead_nodes = Vec::new();
-        let mut ready_to_commit = 0;
-        // get the size of the cluster to compute what a quorum means
-        // it's mutable because if followers die we must remove them
-        // from the quorum
-        let mut cluster_size = self.active_followers.len();
-
-        // wait till enough nodes are ready to commit
-        for (idx, follower) in self.active_followers.iter_mut().enumerate() {
-            match consistency_level {
-                Consistency::Zero => break,
-                Consistency::One if ready_to_commit >= 1 => break,
-                Consistency::Two if ready_to_commit >= 2 => break,
-                Consistency::Quorum if ready_to_commit >= (cluster_size / 2) => break,
-                _ => (),
-            }
-            match follower.receiver.recv() {
-                Ok(FollowerMsg::ReadyToCommit(id)) if id == self.batch_id => ready_to_commit += 1,
-                Ok(FollowerMsg::RegisterNewTask(_)) => log::warn!("Missed a task"),
-                Ok(_) => (),
-                // if a node can't be joined we consider it as dead
-                Err(_) => {
-                    dead_nodes.push(idx);
-                    cluster_size -= 1
                 }
             }
         }
 
-        let dn = dead_nodes.clone();
+        self.broadcast_to_follower.send(LeaderMsg::Commit(self.batch_id)).unwrap();
-        for (idx, follower) in
-            self.active_followers.iter_mut().enumerate().filter(|(i, _)| !dn.contains(i))
-        {
-            match follower.sender.send(LeaderMsg::Commit(self.batch_id)) {
-                Ok(_) => (),
-                Err(_) => dead_nodes.push(idx),
-            }
-        }
-
-        // we do it from the end so the indices stays correct while removing elements
-        for dead_node in dead_nodes.into_iter().rev() {
-            let dead = self.active_followers.swap_remove(dead_node);
-            self.dead_followers.push(dead);
-        }
 
         self.batch_id += 1;
-
-        Ok(())
     }
 }

cluster/src/lib.rs

@@ -16,7 +16,7 @@ pub enum Error {
     SerdeJson(#[from] serde_json::Error),
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 pub enum LeaderMsg {
     // Starts a new batch
     StartBatch { id: u32, batch: Vec<u32> },