From cf8f369e407bd0b10ed63f6bea25a077dc63f739 Mon Sep 17 00:00:00 2001
From: Pascal Seitz <pascal.seitz@gmail.com>
Date: Wed, 22 Apr 2026 15:10:02 +0200
Subject: [PATCH] sort batch docs by partition to reduce mem usage lookups

Sort the incoming ProcessedDocBatch by partition and group with
chunk_by so get_or_create_indexed_split and the index_writer
mem_usage() probes are done once per partition group instead of
once per doc.
---
 .../quickwit-indexing/src/actors/indexer.rs   | 55 +++++++++++--------
 1 file changed, 31 insertions(+), 24 deletions(-)

diff --git a/quickwit/quickwit-indexing/src/actors/indexer.rs b/quickwit/quickwit-indexing/src/actors/indexer.rs
index 3d272caf11a..60c5b1fd081 100644
--- a/quickwit/quickwit-indexing/src/actors/indexer.rs
+++ b/quickwit/quickwit-indexing/src/actors/indexer.rs
@@ -145,7 +145,7 @@ impl IndexerState {
         last_delete_opstamp: u64,
         splits: &'a mut FnvHashMap<u64, IndexedSplitBuilder>,
         other_split_opt: &'a mut Option<IndexedSplitBuilder>,
-        counter: &'a mut IndexerCounters,
+        num_docs_in_workbench: u64,
         ctx: &ActorContext<Indexer>,
     ) -> anyhow::Result<(&'a mut IndexedSplitBuilder, bool)> {
         let num_splits = splits.len();
@@ -157,7 +157,7 @@ impl IndexerState {
                     // `OTHER` special partition.
                     if other_split_opt.is_none() {
                         warn!(
-                            num_docs_in_workbench = counter.num_docs_in_workbench,
+                            num_docs_in_workbench = num_docs_in_workbench,
                             max_num_partition = self.max_num_partitions.get(),
                             "Exceeding max_num_partition"
                         );
@@ -272,7 +272,7 @@ impl IndexerState {
 
     async fn index_batch(
         &self,
-        batch: ProcessedDocBatch,
+        mut batch: ProcessedDocBatch,
         indexing_workbench_opt: &mut Option<IndexingWorkbench>,
         counters: &mut IndexerCounters,
         ctx: &ActorContext<Indexer>,
@@ -297,22 +297,19 @@ impl IndexerState {
             .source_delta
             .extend(batch.checkpoint_delta)
             .context("batch delta does not follow indexer checkpoint")?;
-        let mut memory_usage_delta: i64 = 0;
         counters.num_doc_batches_in_workbench += 1;
-        for doc in batch.docs {
-            let ProcessedDoc {
-                doc,
-                timestamp_opt,
-                partition,
-                num_bytes,
-            } = doc;
-            counters.num_docs_in_workbench += 1;
+        batch.docs.sort_unstable_by_key(|doc| doc.partition);
+        // sort by partition to minimize the number of splits we need to write to for this batch.
+
+        let docs_grouped_by_partition = batch.docs.into_iter().chunk_by(|doc| doc.partition);
+        for (partition, docs) in &docs_grouped_by_partition {
+            let mut memory_usage_delta: i64 = 0;
             let (indexed_split, split_created) = self.get_or_create_indexed_split(
                 partition,
                 *last_delete_opstamp,
                 indexed_splits,
                 other_indexed_split_opt,
-                counters,
+                counters.num_docs_in_workbench,
                 ctx,
             )?;
             let mem_usage_before = indexed_split.index_writer.mem_usage() as u64;
@@ -321,21 +318,31 @@ impl IndexerState {
                 // memory usage.
                 memory_usage_delta += mem_usage_before as i64;
             }
-            indexed_split.split_attrs.uncompressed_docs_size_in_bytes += num_bytes as u64;
-            indexed_split.split_attrs.num_docs += 1;
-            if let Some(timestamp) = timestamp_opt {
-                record_timestamp(timestamp, &mut indexed_split.split_attrs.time_range);
+
+            for doc in docs {
+                let ProcessedDoc {
+                    doc,
+                    timestamp_opt,
+                    num_bytes,
+                    partition: _,
+                } = doc;
+                counters.num_docs_in_workbench += 1;
+                indexed_split.split_attrs.uncompressed_docs_size_in_bytes += num_bytes as u64;
+                indexed_split.split_attrs.num_docs += 1;
+                if let Some(timestamp) = timestamp_opt {
+                    record_timestamp(timestamp, &mut indexed_split.split_attrs.time_range);
+                }
+                let _protect_guard = ctx.protect_zone();
+                indexed_split
+                    .index_writer
+                    .add_document(doc)
+                    .context("failed to add document")?;
+                ctx.record_progress();
             }
-            let _protect_guard = ctx.protect_zone();
-            indexed_split
-                .index_writer
-                .add_document(doc)
-                .context("failed to add document")?;
             let mem_usage_after = indexed_split.index_writer.mem_usage() as u64;
             memory_usage_delta += mem_usage_after as i64 - mem_usage_before as i64;
-            ctx.record_progress();
+            memory_usage.add(memory_usage_delta);
         }
-        memory_usage.add(memory_usage_delta);
         Ok(())
     }
 }