Remove useless heed path

Bump actions/setup-python from 5 to 6

.github/dependabot.yml

@@ -7,6 +7,5 @@ updates:
     schedule:
       interval: "monthly"
     labels:
-      - 'skip changelog'
       - 'dependencies'
     rebase-strategy: disabled

.github/release-draft-template.yml

@@ -18,6 +18,7 @@ categories:
     label: 'security'
   - title: '⚙️ Maintenance/misc'
     label:
+      - 'dependencies'
       - 'maintenance'
       - 'documentation'
 template: |
@@ -26,8 +27,3 @@ template: |
   ❤️  Huge thanks to our contributors: $CONTRIBUTORS.
 no-changes-template: 'Changes are coming soon 😎'
 sort-direction: 'ascending'
-replacers:
-  - search: '/(?:and )?@dependabot-preview(?:\[bot\])?,?/g'
-    replace: ''
-  - search: '/(?:and )?@dependabot(?:\[bot\])?,?/g'
-    replace: ''

.github/workflows/publish-docker-images.yml

@@ -65,7 +65,7 @@ jobs:
         uses: docker/setup-buildx-action@v3
 
       - name: Install cosign
-        uses: sigstore/cosign-installer@d58896d6a1865668819e1d91763c7751a165e159 # tag=v3.9.2
+        uses: sigstore/cosign-installer@d7543c93d881b35a8faa02e8e3605f69b7a1ce62 # tag=v3.10.0
 
       - name: Login to Docker Hub
         uses: docker/login-action@v3

.github/workflows/publish-release-assets.yml

@@ -11,7 +11,7 @@ jobs:
   check-version:
     name: Check the version validity
     runs-on: ubuntu-latest
-    # No need to check the version for dry run (cron)
+    # No need to check the version for dry run (cron or workflow_dispatch)
     steps:
       - uses: actions/checkout@v5
       # Check if the tag has the v<nmumber>.<number>.<number> format.
@@ -48,7 +48,7 @@ jobs:
       - uses: dtolnay/rust-toolchain@1.89
       - name: Build
         run: cargo build --release --locked
-      # No need to upload binaries for dry run (cron)
+      # No need to upload binaries for dry run (cron or workflow_dispatch)
       - name: Upload binaries to release
         if: github.event_name == 'release'
         uses: svenstaro/upload-release-action@2.11.2
@@ -78,7 +78,7 @@ jobs:
       - uses: dtolnay/rust-toolchain@1.89
       - name: Build
         run: cargo build --release --locked
-      # No need to upload binaries for dry run (cron)
+      # No need to upload binaries for dry run (cron or workflow_dispatch)
       - name: Upload binaries to release
         if: github.event_name == 'release'
         uses: svenstaro/upload-release-action@2.11.2
@@ -111,7 +111,7 @@ jobs:
           command: build
           args: --release --target ${{ matrix.target }}
       - name: Upload the binary to release
-        # No need to upload binaries for dry run (cron)
+        # No need to upload binaries for dry run (cron or workflow_dispatch)
         if: github.event_name == 'release'
         uses: svenstaro/upload-release-action@2.11.2
         with:
@@ -176,7 +176,7 @@ jobs:
       - name: List target output files
         run: ls -lR ./target
       - name: Upload the binary to release
-        # No need to upload binaries for dry run (cron)
+        # No need to upload binaries for dry run (cron or workflow_dispatch)
         if: github.event_name == 'release'
         uses: svenstaro/upload-release-action@2.11.2
         with:
@@ -187,6 +187,7 @@ jobs:
 
   publish-openapi-file:
     name: Publish OpenAPI file
+    needs: check-version
     runs-on: ubuntu-latest
     steps:
       - name: Checkout code
@@ -201,7 +202,7 @@ jobs:
           cd crates/openapi-generator
           cargo run --release -- --pretty --output ../../meilisearch.json
       - name: Upload OpenAPI to Release
-        # No need to upload for dry run (cron)
+        # No need to upload for dry run (cron or workflow_dispatch)
         if: github.event_name == 'release'
         uses: svenstaro/upload-release-action@2.11.2
         with:

.github/workflows/sdks-tests.yml

@@ -50,7 +50,7 @@ jobs:
         with:
           repository: meilisearch/meilisearch-dotnet
       - name: Setup .NET Core
-        uses: actions/setup-dotnet@v4
+        uses: actions/setup-dotnet@v5
         with:
           dotnet-version: "8.0.x"
       - name: Install dependencies
@@ -100,7 +100,7 @@ jobs:
           - '7700:7700'
     steps:
       - name: Set up Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
         with:
           go-version: stable
       - uses: actions/checkout@v5
@@ -135,13 +135,13 @@ jobs:
       - name: Set up Java
         uses: actions/setup-java@v5
         with:
-          java-version: 8
-          distribution: 'zulu'
+          java-version: 17
+          distribution: 'temurin'
           cache: gradle
       - name: Grant execute permission for gradlew
         run: chmod +x gradlew
       - name: Build and run unit and integration tests
-        run: ./gradlew build integrationTest
+        run: ./gradlew build integrationTest --info
 
   meilisearch-js-tests:
     needs: define-docker-image
@@ -160,7 +160,7 @@ jobs:
         with:
           repository: meilisearch/meilisearch-js
       - name: Setup node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v5
         with:
           cache: 'yarn'
       - name: Install dependencies
@@ -224,7 +224,7 @@ jobs:
         with:
           repository: meilisearch/meilisearch-python
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
       - name: Install pipenv
         uses: dschep/install-pipenv-action@v1
       - name: Install dependencies
@@ -318,7 +318,7 @@ jobs:
         with:
           repository: meilisearch/meilisearch-js-plugins
       - name: Setup node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v5
         with:
           cache: yarn
       - name: Install dependencies

Cargo.lock

@@ -453,9 +453,8 @@ checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
 
 [[package]]
 name = "arroy"
-version = "0.6.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8578a72223dfa13dfd9fc144d15260d134361789ebdea9b16e85a511edc73c7d"
+version = "0.6.4-nested-rtxns"
+source = "git+https://github.com/meilisearch/arroy?branch=use-heed-nested-rtxns#61c8f4f0addeff968e80438018d0aee2c1eb8d67"
 dependencies = [
  "bytemuck",
  "byteorder",
@@ -1075,9 +1074,8 @@ dependencies = [
 
 [[package]]
 name = "cellulite"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "71a41aa2cd021bc3f23d97cc1e645848ca8c279fc757d1570ba7fe7ddc021290"
+version = "0.3.1-nested-rtxns"
+source = "git+https://github.com/meilisearch/cellulite?branch=use-heed-nested-rtxns#9fb1866cc49277d26f606769112fa704944ccc61"
 dependencies = [
  "crossbeam",
  "geo",
@@ -2758,9 +2756,8 @@ dependencies = [
 
 [[package]]
 name = "hannoy"
-version = "0.0.8"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0dba13a271c49a119a97862ebf0a74131d879832868400d9fcd937b790058fdd"
+version = "0.0.9-nested-rtxns"
+source = "git+https://github.com/nnethercott/hannoy?branch=use-heed-nested-rtxns#d4ca5454eff6539e9fc2119f07113abebbda0a39"
 dependencies = [
  "bytemuck",
  "byteorder",
@@ -2838,9 +2835,9 @@ checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
 
 [[package]]
 name = "heed"
-version = "0.22.0"
+version = "0.22.1-nested-rtxns"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a56c94661ddfb51aa9cdfbf102cfcc340aa69267f95ebccc4af08d7c530d393"
+checksum = "0ff115ba5712b1f1fc7617b195f5c2f139e29c397ff79da040cd19db75ccc240"
 dependencies = [
  "bitflags 2.9.4",
  "byteorder",
@@ -2850,7 +2847,6 @@ dependencies = [
  "lmdb-master-sys",
  "once_cell",
  "page_size",
- "serde",
  "synchronoise",
  "url",
 ]
@@ -3889,9 +3885,9 @@ checksum = "241eaef5fd12c88705a01fc1066c48c4b36e0dd4377dcdc7ec3942cea7a69956"
 
 [[package]]
 name = "lmdb-master-sys"
-version = "0.2.5"
+version = "0.2.6-nested-rtxns"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "864808e0b19fb6dd3b70ba94ee671b82fce17554cf80aeb0a155c65bb08027df"
+checksum = "f4ff85130e3c994b36877045fbbb138d521dea7197bfc19dc3d5d95101a8e20a"
 dependencies = [
  "cc",
  "doxygen-rs",

README.md

@@ -121,7 +121,7 @@ If you want to know more about the kind of data we collect and what we use it fo
 
 Meilisearch is a search engine created by [Meili](https://www.meilisearch.com/careers), a software development company headquartered in France and with team members all over the world. Want to know more about us? [Check out our blog!](https://blog.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=contact)
 
-🗞 [Subscribe to our newsletter](https://meilisearch.us2.list-manage.com/subscribe?u=27870f7b71c908a8b359599fb&id=79582d828e) if you don't want to miss any updates! We promise we won't clutter your mailbox: we only send one edition every two months.
+🗞 [Subscribe to our newsletter](https://share-eu1.hsforms.com/1LN5N0x_GQgq7ss7tXmSykwfg3aq) if you don't want to miss any updates! We promise we won't clutter your mailbox: we only send one edition every two months.
 
 💌 Want to make a suggestion or give feedback? Here are some of the channels where you can reach us:
 

crates/milli/Cargo.toml

@@ -19,7 +19,7 @@ bstr = "1.12.0"
 bytemuck = { version = "1.23.1", features = ["extern_crate_alloc"] }
 byteorder = "1.5.0"
 charabia = { version = "0.9.7", default-features = false }
-cellulite = "0.3.0"
+cellulite = { version = "0.3.1-nested-rtxns", git = "https://github.com/meilisearch/cellulite", "branch" = "use-heed-nested-rtxns" }
 concat-arrays = "0.1.2"
 convert_case = "0.8.0"
 crossbeam-channel = "0.5.15"
@@ -34,7 +34,7 @@ grenad = { version = "0.5.0", default-features = false, features = [
     "rayon",
     "tempfile",
 ] }
-heed = { version = "0.22.0", default-features = false, features = [
+heed = { version = "0.22.1-nested-rtxns", default-features = false, features = [
     "serde-json",
     "serde-bincode",
 ] }
@@ -89,8 +89,8 @@ rhai = { version = "1.22.2", features = [
     "no_time",
     "sync",
 ] }
-arroy = "0.6.3"
-hannoy = { version = "0.0.8", features = ["arroy"] }
+arroy = { version = "0.6.4-nested-rtxns", git = "https://github.com/meilisearch/arroy", "branch" = "use-heed-nested-rtxns" }
+hannoy = { version = "0.0.9-nested-rtxns", git = "https://github.com/nnethercott/hannoy", "branch" = "use-heed-nested-rtxns", features = ["arroy"] }
 rand = "0.8.5"
 tracing = "0.1.41"
 ureq = { version = "2.12.1", features = ["json"] }

crates/milli/src/update/new/indexer/mod.rs

@@ -180,12 +180,15 @@ where
         })
         .unwrap()?;
 
-        post_processing::post_process(
-            indexing_context,
-            wtxn,
-            global_fields_ids_map,
-            facet_field_ids_delta,
-        )?;
+        pool.install(|| {
+            post_processing::post_process(
+                indexing_context,
+                wtxn,
+                global_fields_ids_map,
+                facet_field_ids_delta,
+            )
+        })
+        .unwrap()?;
 
         indexing_context.progress.update_progress(IndexingStep::BuildingGeoJson);
         index.cellulite.build(

crates/milli/src/update/new/indexer/post_processing/facet_bulk.rs

@@ -0,0 +1,164 @@
+use std::fs::File;
+use std::io::BufReader;
+use std::{iter, mem};
+
+use grenad::CompressionType;
+use heed::types::{Bytes, LazyDecode};
+use heed::{Database, RwTxn};
+use rayon::prelude::*;
+use roaring::MultiOps;
+use tempfile::tempfile;
+
+use crate::facet::FacetType;
+use crate::heed_codec::facet::{FacetGroupKey, FacetGroupKeyCodec, FacetGroupValueCodec};
+use crate::heed_codec::BytesRefCodec;
+use crate::update::facet::{FACET_GROUP_SIZE, FACET_MIN_LEVEL_SIZE};
+use crate::update::{create_writer, writer_into_reader};
+use crate::{CboRoaringBitmapCodec, FieldId, Index};
+
+/// Generate the facet level based on the level 0.
+///
+/// The function will generate all the group levels from
+/// the group 1 to the level n until the number of group
+/// is smaller than the minimum required size.
+pub fn generate_facet_levels(
+    index: &Index,
+    wtxn: &mut RwTxn,
+    field_id: FieldId,
+    facet_type: FacetType,
+) -> crate::Result<()> {
+    let db = match facet_type {
+        FacetType::String => index
+            .facet_id_string_docids
+            .remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>()
+            .lazily_decode_data(),
+        FacetType::Number => index
+            .facet_id_f64_docids
+            .remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>()
+            .lazily_decode_data(),
+    };
+
+    clear_levels(db, wtxn, field_id)?;
+
+    let mut base_level = 0;
+    // That's a do-while loop
+    while {
+        let mut level_size = 0;
+        for reader in compute_level(index, wtxn, db, field_id, base_level)? {
+            let mut cursor = reader.into_cursor()?;
+            while let Some((left_bound, facet_group_value)) = cursor.move_on_next()? {
+                level_size += 1;
+                let level = base_level.checked_add(1).unwrap();
+                let key = FacetGroupKey { field_id, level, left_bound };
+                debug_assert!(
+                    db.get(wtxn, &key).transpose().is_none(),
+                    "entry must not be there and must have already been deleted: {key:?}"
+                );
+                db.remap_data_type::<Bytes>().put(wtxn, &key, facet_group_value)?;
+            }
+        }
+
+        base_level += 1;
+
+        // If the next level will have the minimum required groups, continue.
+        (level_size / FACET_GROUP_SIZE as usize) >= FACET_MIN_LEVEL_SIZE as usize
+    } {}
+
+    Ok(())
+}
+
+/// Compute the groups of facets from the provided base level
+/// and write the content into different grenad files.
+fn compute_level(
+    index: &Index,
+    wtxn: &heed::RwTxn,
+    db: Database<FacetGroupKeyCodec<BytesRefCodec>, LazyDecode<FacetGroupValueCodec>>,
+    field_id: FieldId,
+    base_level: u8,
+) -> Result<Vec<grenad::Reader<BufReader<File>>>, crate::Error> {
+    let thread_count = rayon::current_num_threads();
+    let rtxns = iter::repeat_with(|| index.env.nested_read_txn(wtxn))
+        .take(thread_count)
+        .collect::<heed::Result<Vec<_>>>()?;
+
+    let range = {
+        // Based on the first possible value for the base level up to
+        // the first possible value for the next level *excluded*.
+        let left = FacetGroupKey::<&[u8]> { field_id, level: base_level, left_bound: &[] };
+        let right = FacetGroupKey::<&[u8]> {
+            field_id,
+            level: base_level.checked_add(1).unwrap(),
+            left_bound: &[],
+        };
+        left..right
+    };
+
+    rtxns
+        .into_par_iter()
+        .enumerate()
+        .map(|(thread_id, rtxn)| {
+            let mut writer = tempfile().map(|f| create_writer(CompressionType::None, None, f))?;
+
+            let mut left_bound = None;
+            let mut group_docids = Vec::new();
+            let mut ser_buffer = Vec::new();
+            for (i, result) in db.range(&rtxn, &range)?.enumerate() {
+                let (key, lazy_value) = result?;
+
+                let start_of_group = i % FACET_GROUP_SIZE as usize == 0;
+                let group_index = i / FACET_GROUP_SIZE as usize;
+                let group_for_thread = group_index % thread_count == thread_id;
+
+                if group_for_thread {
+                    if start_of_group {
+                        if let Some(left_bound) = left_bound.take() {
+                            // We store the bitmaps in a Vec this way we can use
+                            // the MultiOps operations that tends to be more efficient
+                            // for unions. The Vec is empty after the operation.
+                            //
+                            // We also don't forget to store the group size corresponding
+                            // to the number of entries merged in this group.
+                            ser_buffer.clear();
+                            let group_len: u8 = group_docids.len().try_into().unwrap();
+                            ser_buffer.push(group_len);
+                            let group_docids = mem::take(&mut group_docids);
+                            let docids = group_docids.into_iter().union();
+                            CboRoaringBitmapCodec::serialize_into_vec(&docids, &mut ser_buffer);
+                            writer.insert(left_bound, &ser_buffer)?;
+                        }
+                        left_bound = Some(key.left_bound);
+                    }
+
+                    // Lazily decode the bitmaps we are interested in.
+                    let value = lazy_value.decode().map_err(heed::Error::Decoding)?;
+                    group_docids.push(value.bitmap);
+                }
+            }
+
+            if let Some(left_bound) = left_bound.take() {
+                ser_buffer.clear();
+                // We don't forget to store the group size corresponding
+                // to the number of entries merged in this group.
+                let group_len: u8 = group_docids.len().try_into().unwrap();
+                ser_buffer.push(group_len);
+                let group_docids = group_docids.into_iter().union();
+                CboRoaringBitmapCodec::serialize_into_vec(&group_docids, &mut ser_buffer);
+                writer.insert(left_bound, &ser_buffer)?;
+            }
+
+            writer_into_reader(writer)
+        })
+        .collect()
+}
+
+/// Clears all the levels and only keeps the level 0 of the specified field id.
+fn clear_levels(
+    db: Database<FacetGroupKeyCodec<BytesRefCodec>, LazyDecode<FacetGroupValueCodec>>,
+    wtxn: &mut RwTxn<'_>,
+    field_id: FieldId,
+) -> heed::Result<()> {
+    let left = FacetGroupKey::<&[u8]> { field_id, level: 1, left_bound: &[] };
+    let right = FacetGroupKey::<&[u8]> { field_id, level: u8::MAX, left_bound: &[] };
+    let range = left..=right;
+    db.delete_range(wtxn, &range).map(drop)
+}

crates/milli/src/update/new/indexer/post_processing/mod.rs

@@ -1,5 +1,6 @@
 use std::cmp::Ordering;
 
+use facet_bulk::generate_facet_levels;
 use heed::types::{Bytes, DecodeIgnore, Str};
 use heed::RwTxn;
 use itertools::{merge_join_by, EitherOrBoth};
@@ -23,6 +24,8 @@ use crate::update::new::FacetFieldIdsDelta;
 use crate::update::{FacetsUpdateBulk, GrenadParameters};
 use crate::{GlobalFieldsIdsMap, Index, Result};
 
+mod facet_bulk;
+
 pub(super) fn post_process<MSP>(
     indexing_context: IndexingContext<MSP>,
     wtxn: &mut RwTxn<'_>,
@@ -239,9 +242,8 @@ fn compute_facet_level_database(
         match delta {
             FacetFieldIdDelta::Bulk => {
                 progress.update_progress(PostProcessingFacets::StringsBulk);
-                tracing::debug!(%fid, "bulk string facet processing");
-                FacetsUpdateBulk::new_not_updating_level_0(index, vec![fid], FacetType::String)
-                    .execute(wtxn)?
+                tracing::debug!(%fid, "bulk string facet processing in parallel");
+                generate_facet_levels(index, wtxn, fid, FacetType::String)?
             }
             FacetFieldIdDelta::Incremental(delta_data) => {
                 progress.update_progress(PostProcessingFacets::StringsIncremental);

crates/milli/src/update/new/words_prefix_docids.rs

@@ -1,11 +1,12 @@
 use std::cell::RefCell;
 use std::collections::BTreeSet;
 use std::io::{BufReader, BufWriter, Read, Seek, Write};
+use std::iter;
 
 use hashbrown::HashMap;
 use heed::types::Bytes;
 use heed::{BytesDecode, Database, Error, RoTxn, RwTxn};
-use rayon::iter::{IntoParallelIterator, ParallelIterator as _};
+use rayon::iter::{IndexedParallelIterator as _, IntoParallelIterator, ParallelIterator as _};
 use roaring::MultiOps;
 use tempfile::spooled_tempfile;
 use thread_local::ThreadLocal;
@@ -151,25 +152,35 @@ impl<'a, 'rtxn> FrozenPrefixBitmaps<'a, 'rtxn> {
 
 unsafe impl Sync for FrozenPrefixBitmaps<'_, '_> {}
 
-struct WordPrefixIntegerDocids {
+struct WordPrefixIntegerDocids<'i> {
+    index: &'i Index,
     database: Database<Bytes, CboRoaringBitmapCodec>,
     prefix_database: Database<Bytes, CboRoaringBitmapCodec>,
     max_memory_by_thread: Option<usize>,
+    read_uncommitted_in_parallel: bool,
 }
 
-impl WordPrefixIntegerDocids {
+impl<'i> WordPrefixIntegerDocids<'i> {
     fn new(
+        index: &'i Index,
         database: Database<Bytes, CboRoaringBitmapCodec>,
         prefix_database: Database<Bytes, CboRoaringBitmapCodec>,
-        grenad_parameters: &GrenadParameters,
-    ) -> WordPrefixIntegerDocids {
+        grenad_parameters: &'_ GrenadParameters,
+    ) -> WordPrefixIntegerDocids<'i> {
         WordPrefixIntegerDocids {
+            index,
             database,
             prefix_database,
             max_memory_by_thread: grenad_parameters.max_memory_by_thread(),
+            read_uncommitted_in_parallel: false,
         }
     }
 
+    /// Use an experimental LMDB feature to read uncommitted data in parallel.
+    fn read_uncommitted_in_parallel(&mut self, value: bool) {
+        self.read_uncommitted_in_parallel = value;
+    }
+
     fn execute(
         self,
         wtxn: &mut heed::RwTxn,
@@ -177,7 +188,144 @@ impl WordPrefixIntegerDocids {
         prefix_to_delete: &BTreeSet<Prefix>,
     ) -> Result<()> {
         delete_prefixes(wtxn, &self.prefix_database, prefix_to_delete)?;
-        self.recompute_modified_prefixes(wtxn, prefix_to_compute)
+        if self.read_uncommitted_in_parallel {
+            self.recompute_modified_prefixes_no_frozen(wtxn, prefix_to_compute)
+        } else {
+            self.recompute_modified_prefixes(wtxn, prefix_to_compute)
+        }
+    }
+
+    /// Computes the same as `recompute_modified_prefixes`.
+    ///
+    /// ...but without aggregating the prefixes mmap pointers into a static HashMap
+    /// beforehand and rather use an experimental LMDB feature to read the subset
+    /// of prefixes in parallel from the uncommitted transaction.
+    #[tracing::instrument(level = "trace", skip_all, target = "indexing::prefix")]
+    fn recompute_modified_prefixes_no_frozen(
+        &self,
+        wtxn: &mut RwTxn,
+        prefixes: &BTreeSet<Prefix>,
+    ) -> Result<()> {
+        let thread_count = rayon::current_num_threads();
+        let rtxns = iter::repeat_with(|| self.index.env.nested_read_txn(wtxn))
+            .take(thread_count)
+            .collect::<heed::Result<Vec<_>>>()?;
+
+        let outputs = rtxns
+            .into_par_iter()
+            .enumerate()
+            .map(|(thread_id, rtxn)| {
+                // `indexes` represent offsets at which prefixes computations were stored in the `file`.
+                let mut indexes = Vec::new();
+                let mut file = BufWriter::new(spooled_tempfile(
+                    self.max_memory_by_thread.unwrap_or(usize::MAX),
+                ));
+
+                let mut buffer = Vec::new();
+                for (prefix_index, prefix) in prefixes.iter().enumerate() {
+                    // Is prefix for another thread?
+                    if prefix_index % thread_count != thread_id {
+                        continue;
+                    }
+
+                    let mut bitmaps_bytes = Vec::<&[u8]>::new();
+                    let mut prev_pos = None;
+                    for result in self
+                        .database
+                        .remap_data_type::<Bytes>()
+                        .prefix_iter(&rtxn, prefix.as_bytes())?
+                    {
+                        let (key, current_bitmap_bytes) = result?;
+                        let (_word, pos) =
+                            StrBEU16Codec::bytes_decode(key).map_err(Error::Decoding)?;
+
+                        if prev_pos.is_some_and(|p| p != pos) {
+                            if bitmaps_bytes.is_empty() {
+                                indexes.push(PrefixIntegerEntry {
+                                    prefix,
+                                    pos,
+                                    serialized_length: None,
+                                });
+                            } else {
+                                let output = bitmaps_bytes
+                                    .iter()
+                                    .map(|bytes| CboRoaringBitmapCodec::deserialize_from(bytes))
+                                    .union()?;
+                                buffer.clear();
+                                CboRoaringBitmapCodec::serialize_into_vec(&output, &mut buffer);
+                                indexes.push(PrefixIntegerEntry {
+                                    prefix,
+                                    pos,
+                                    serialized_length: Some(buffer.len()),
+                                });
+                                file.write_all(&buffer)?;
+                                bitmaps_bytes.clear();
+                            }
+                        }
+
+                        bitmaps_bytes.push(current_bitmap_bytes);
+                        prev_pos = Some(pos);
+                    }
+
+                    if let Some(pos) = prev_pos {
+                        if bitmaps_bytes.is_empty() {
+                            indexes.push(PrefixIntegerEntry {
+                                prefix,
+                                pos,
+                                serialized_length: None,
+                            });
+                        } else {
+                            let output = bitmaps_bytes
+                                .iter()
+                                .map(|bytes| CboRoaringBitmapCodec::deserialize_from(bytes))
+                                .union()?;
+                            buffer.clear();
+                            CboRoaringBitmapCodec::serialize_into_vec(&output, &mut buffer);
+                            indexes.push(PrefixIntegerEntry {
+                                prefix,
+                                pos,
+                                serialized_length: Some(buffer.len()),
+                            });
+                            file.write_all(&buffer)?;
+                        }
+                    }
+                }
+
+                Ok((indexes, file))
+            })
+            .collect::<Result<Vec<_>>>()?;
+
+        // We iterate over all the collected and serialized bitmaps through
+        // the files and entries to eventually put them in the final database.
+        let mut key_buffer = Vec::new();
+        let mut buffer = Vec::new();
+        for (index, file) in outputs {
+            let mut file = file.into_inner().map_err(|e| e.into_error())?;
+            file.rewind()?;
+            let mut file = BufReader::new(file);
+            for PrefixIntegerEntry { prefix, pos, serialized_length } in index {
+                key_buffer.clear();
+                key_buffer.extend_from_slice(prefix.as_bytes());
+                key_buffer.push(0);
+                key_buffer.extend_from_slice(&pos.to_be_bytes());
+                match serialized_length {
+                    Some(serialized_length) => {
+                        buffer.resize(serialized_length, 0);
+                        file.read_exact(&mut buffer)?;
+                        self.prefix_database.remap_data_type::<Bytes>().put(
+                            wtxn,
+                            &key_buffer,
+                            &buffer,
+                        )?;
+                    }
+                    None => {
+                        self.prefix_database.delete(wtxn, &key_buffer)?;
+                    }
+                }
+            }
+        }
+
+        Ok(())
     }
 
     #[tracing::instrument(level = "trace", skip_all, target = "indexing::prefix")]
@@ -262,7 +410,7 @@ impl WordPrefixIntegerDocids {
     }
 }
 
-/// Represents a prefix and the lenght the bitmap takes on disk.
+/// Represents a prefix and the length the bitmap takes on disk.
 struct PrefixIntegerEntry<'a> {
     prefix: &'a str,
     pos: u16,
@@ -362,12 +510,14 @@ pub fn compute_word_prefix_fid_docids(
     prefix_to_delete: &BTreeSet<Prefix>,
     grenad_parameters: &GrenadParameters,
 ) -> Result<()> {
-    WordPrefixIntegerDocids::new(
+    let mut builder = WordPrefixIntegerDocids::new(
+        index,
         index.word_fid_docids.remap_key_type(),
         index.word_prefix_fid_docids.remap_key_type(),
         grenad_parameters,
-    )
-    .execute(wtxn, prefix_to_compute, prefix_to_delete)
+    );
+    builder.read_uncommitted_in_parallel(true);
+    builder.execute(wtxn, prefix_to_compute, prefix_to_delete)
 }
 
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::prefix")]
@@ -378,10 +528,12 @@ pub fn compute_word_prefix_position_docids(
     prefix_to_delete: &BTreeSet<Prefix>,
     grenad_parameters: &GrenadParameters,
 ) -> Result<()> {
-    WordPrefixIntegerDocids::new(
+    let mut builder = WordPrefixIntegerDocids::new(
+        index,
         index.word_position_docids.remap_key_type(),
         index.word_prefix_position_docids.remap_key_type(),
         grenad_parameters,
-    )
-    .execute(wtxn, prefix_to_compute, prefix_to_delete)
+    );
+    builder.read_uncommitted_in_parallel(true);
+    builder.execute(wtxn, prefix_to_compute, prefix_to_delete)
 }