Merge #5223

5223: Limit batched tasks total size r=curquiza a=Kerollmops

Introduce a new engine parameter (env and config, too) to limit the maximum payload size processed by the engine in batches. You can [review the Discussion and usage on GitHub](https://github.com/orgs/meilisearch/discussions/801).

Co-authored-by: Clément Renault <clement@meilisearch.com>

crates/index-scheduler/src/lib.rs

@@ -115,6 +115,9 @@ pub struct IndexSchedulerOptions {
     /// If the autobatcher is allowed to automatically batch tasks
     /// it will only batch this defined number of tasks at once.
     pub max_number_of_batched_tasks: usize,
+    /// If the autobatcher is allowed to automatically batch tasks
+    /// it will only batch this defined maximum size (in bytes) of tasks at once.
+    pub batched_tasks_size_limit: u64,
     /// The experimental features enabled for this instance.
     pub instance_features: InstanceTogglableFeatures,
 }

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

@@ -497,17 +497,26 @@ impl IndexScheduler {
             1
         };
 
-        let enqueued = index_tasks
-            .into_iter()
-            .take(tasks_limit)
-            .map(|task_id| {
-                self.queue
-                    .tasks
-                    .get_task(rtxn, task_id)
-                    .and_then(|task| task.ok_or(Error::CorruptedTaskQueue))
-                    .map(|task| (task.uid, task.kind))
-            })
-            .collect::<Result<Vec<_>>>()?;
+        let mut enqueued = Vec::new();
+        let mut total_size: u64 = 0;
+        for task_id in index_tasks.into_iter().take(tasks_limit) {
+            let task = self
+                .queue
+                .tasks
+                .get_task(rtxn, task_id)
+                .and_then(|task| task.ok_or(Error::CorruptedTaskQueue))?;
+
+            if let Some(uuid) = task.content_uuid() {
+                let content_size = self.queue.file_store.compute_size(uuid)?;
+                total_size = total_size.saturating_add(content_size);
+            }
+
+            if total_size > self.scheduler.batched_tasks_size_limit && !enqueued.is_empty() {
+                break;
+            }
+
+            enqueued.push((task.uid, task.kind));
+        }
 
         if let Some((batchkind, create_index)) =
             autobatcher::autobatch(enqueued, index_already_exists, primary_key.as_deref())

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

@@ -60,6 +60,9 @@ pub struct Scheduler {
     /// The maximum number of tasks that will be batched together.
     pub(crate) max_number_of_batched_tasks: usize,
 
+    /// The maximum size, in bytes, of tasks in a batch.
+    pub(crate) batched_tasks_size_limit: u64,
+
     /// The path used to create the dumps.
     pub(crate) dumps_path: PathBuf,
 
@@ -80,6 +83,7 @@ impl Scheduler {
             wake_up: self.wake_up.clone(),
             autobatching_enabled: self.autobatching_enabled,
             max_number_of_batched_tasks: self.max_number_of_batched_tasks,
+            batched_tasks_size_limit: self.batched_tasks_size_limit,
             dumps_path: self.dumps_path.clone(),
             snapshots_path: self.snapshots_path.clone(),
             auth_path: self.auth_path.clone(),
@@ -94,6 +98,7 @@ impl Scheduler {
             wake_up: Arc::new(SignalEvent::auto(true)),
             autobatching_enabled: options.autobatching_enabled,
             max_number_of_batched_tasks: options.max_number_of_batched_tasks,
+            batched_tasks_size_limit: options.batched_tasks_size_limit,
             dumps_path: options.dumps_path.clone(),
             snapshots_path: options.snapshots_path.clone(),
             auth_path: options.auth_path.clone(),

crates/index-scheduler/src/test_utils.rs

@@ -107,6 +107,7 @@ impl IndexScheduler {
             cleanup_enabled: true,
             max_number_of_tasks: 1_000_000,
             max_number_of_batched_tasks: usize::MAX,
+            batched_tasks_size_limit: u64::MAX,
             instance_features: Default::default(),
         };
         configuration(&mut options);