Merge #5303

5303: Bring back changes from v1.12.8 into v1.13.0 r=Kerollmops a=Kerollmops

Fixes #5087 and other problems that you can find in the original PR #5294.

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

crates/index-scheduler/src/index_mapper/index_map.rs

@@ -1,5 +1,7 @@
 use std::collections::BTreeMap;
+use std::env::VarError;
 use std::path::Path;
+use std::str::FromStr;
 use std::time::Duration;
 
 use meilisearch_types::heed::{EnvClosingEvent, EnvFlags, EnvOpenOptions};
@@ -304,7 +306,18 @@ fn create_or_open_index(
 ) -> Result<Index> {
     let mut options = EnvOpenOptions::new();
     options.map_size(clamp_to_page_size(map_size));
-    options.max_readers(1024);
+
+    // You can find more details about this experimental
+    // environment variable on the following GitHub discussion:
+    // <https://github.com/orgs/meilisearch/discussions/806>
+    let max_readers = match std::env::var("MEILI_EXPERIMENTAL_INDEX_MAX_READERS") {
+        Ok(value) => u32::from_str(&value).unwrap(),
+        Err(VarError::NotPresent) => 1024,
+        Err(VarError::NotUnicode(value)) => panic!(
+            "Invalid unicode for the `MEILI_EXPERIMENTAL_INDEX_MAX_READERS` env var: {value:?}"
+        ),
+    };
+    options.max_readers(max_readers);
     if enable_mdb_writemap {
         unsafe { options.flags(EnvFlags::WRITE_MAP) };
     }

crates/milli/src/thread_pool_no_abort.rs

@@ -1,4 +1,4 @@
-use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use std::sync::Arc;
 
 use rayon::{ThreadPool, ThreadPoolBuilder};
@@ -9,6 +9,8 @@ use thiserror::Error;
 #[derive(Debug)]
 pub struct ThreadPoolNoAbort {
     thread_pool: ThreadPool,
+    /// The number of active operations.
+    active_operations: AtomicUsize,
     /// Set to true if the thread pool catched a panic.
     pool_catched_panic: Arc<AtomicBool>,
 }
@@ -19,7 +21,9 @@ impl ThreadPoolNoAbort {
         OP: FnOnce() -> R + Send,
         R: Send,
     {
+        self.active_operations.fetch_add(1, Ordering::Relaxed);
         let output = self.thread_pool.install(op);
+        self.active_operations.fetch_sub(1, Ordering::Relaxed);
         // While reseting the pool panic catcher we return an error if we catched one.
         if self.pool_catched_panic.swap(false, Ordering::SeqCst) {
             Err(PanicCatched)
@@ -31,6 +35,11 @@ impl ThreadPoolNoAbort {
     pub fn current_num_threads(&self) -> usize {
         self.thread_pool.current_num_threads()
     }
+
+    /// The number of active operations.
+    pub fn active_operations(&self) -> usize {
+        self.active_operations.load(Ordering::Relaxed)
+    }
 }
 
 #[derive(Error, Debug)]
@@ -64,6 +73,10 @@ impl ThreadPoolNoAbortBuilder {
             let catched_panic = pool_catched_panic.clone();
             move |_result| catched_panic.store(true, Ordering::SeqCst)
         });
-        Ok(ThreadPoolNoAbort { thread_pool: self.0.build()?, pool_catched_panic })
+        Ok(ThreadPoolNoAbort {
+            thread_pool: self.0.build()?,
+            active_operations: AtomicUsize::new(0),
+            pool_catched_panic,
+        })
     }
 }

crates/milli/src/vector/ollama.rs

@@ -5,7 +5,7 @@ use rayon::slice::ParallelSlice as _;
 
 use super::error::{EmbedError, EmbedErrorKind, NewEmbedderError, NewEmbedderErrorKind};
 use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
-use super::DistributionShift;
+use super::{DistributionShift, REQUEST_PARALLELISM};
 use crate::error::FaultSource;
 use crate::vector::Embedding;
 use crate::ThreadPoolNoAbort;
@@ -118,6 +118,11 @@ impl Embedder {
         text_chunks: Vec<Vec<String>>,
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Vec<Embedding>>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            text_chunks.into_iter().map(move |chunk| self.embed(&chunk, None)).collect()
+        } else {
             threads
                 .install(move || {
                     text_chunks.into_par_iter().map(move |chunk| self.embed(&chunk, None)).collect()
@@ -127,12 +132,24 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub(crate) fn embed_chunks_ref(
         &self,
         texts: &[&str],
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Vec<f32>>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
+                .chunks(self.prompt_count_in_chunk_hint())
+                .map(move |chunk| self.embed(chunk, None))
+                .collect();
+
+            let embeddings = embeddings?;
+            Ok(embeddings.into_iter().flatten().collect())
+        } else {
             threads
                 .install(move || {
                     let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
@@ -148,6 +165,7 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub fn chunk_count_hint(&self) -> usize {
         self.rest_embedder.chunk_count_hint()

crates/milli/src/vector/openai.rs

@@ -7,7 +7,7 @@ use rayon::slice::ParallelSlice as _;
 
 use super::error::{EmbedError, NewEmbedderError};
 use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
-use super::DistributionShift;
+use super::{DistributionShift, REQUEST_PARALLELISM};
 use crate::error::FaultSource;
 use crate::vector::error::EmbedErrorKind;
 use crate::vector::Embedding;
@@ -255,6 +255,11 @@ impl Embedder {
         text_chunks: Vec<Vec<String>>,
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Vec<Embedding>>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            text_chunks.into_iter().map(move |chunk| self.embed(&chunk, None)).collect()
+        } else {
             threads
                 .install(move || {
                     text_chunks.into_par_iter().map(move |chunk| self.embed(&chunk, None)).collect()
@@ -264,12 +269,23 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub(crate) fn embed_chunks_ref(
         &self,
         texts: &[&str],
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Vec<f32>>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
+                .chunks(self.prompt_count_in_chunk_hint())
+                .map(move |chunk| self.embed(chunk, None))
+                .collect();
+            let embeddings = embeddings?;
+            Ok(embeddings.into_iter().flatten().collect())
+        } else {
             threads
                 .install(move || {
                     let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
@@ -285,6 +301,7 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub fn chunk_count_hint(&self) -> usize {
         self.rest_embedder.chunk_count_hint()

crates/milli/src/vector/rest.rs

@@ -188,6 +188,11 @@ impl Embedder {
         text_chunks: Vec<Vec<String>>,
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Vec<Embedding>>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            text_chunks.into_iter().map(move |chunk| self.embed(chunk, None)).collect()
+        } else {
             threads
                 .install(move || {
                     text_chunks.into_par_iter().map(move |chunk| self.embed(chunk, None)).collect()
@@ -197,12 +202,24 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub(crate) fn embed_chunks_ref(
         &self,
         texts: &[&str],
         threads: &ThreadPoolNoAbort,
     ) -> Result<Vec<Embedding>, EmbedError> {
+        // This condition helps reduce the number of active rayon jobs
+        // so that we avoid consuming all the LMDB rtxns and avoid stack overflows.
+        if threads.active_operations() >= REQUEST_PARALLELISM {
+            let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
+                .chunks(self.prompt_count_in_chunk_hint())
+                .map(move |chunk| self.embed_ref(chunk, None))
+                .collect();
+
+            let embeddings = embeddings?;
+            Ok(embeddings.into_iter().flatten().collect())
+        } else {
             threads
                 .install(move || {
                     let embeddings: Result<Vec<Vec<Embedding>>, _> = texts
@@ -218,6 +235,7 @@ impl Embedder {
                     fault: FaultSource::Bug,
                 })?
         }
+    }
 
     pub fn chunk_count_hint(&self) -> usize {
         super::REQUEST_PARALLELISM