Merge #382

382: Refactor attribute criterion r=Kerollmops a=ManyTheFish

### Re-implement set based algorithm for attribute criterion
#### Levels
Instead of doing level iteration and digging in the interesting level, we only iterate over the lowest level.

#### crossword iteration VS minimal position iteration
Instead of crossing word position in order to iterate strictly over the position that gives the best rank in good order; we iterate word by word starting with the word that increases the rank the little as possible.
This new method is a bit less precise but way simpler.

### Simplify word-level-position database
We don't use levels anymore in the attribute criterion, and so we removed the level complexity of the database making a word-position-docids database.

### Benchmarks on search on big datasets

#### songs main VS refactor-attribute-criterion
```diff
  group                                                   search_songsmain_31c18f09               search_songsrefactor-attribute-criterion_1bd15d84
  -----                                                   -------------------------               -------------------------------------------------
- smol-songs.csv: basic filter: <=/Notstandskomitee       1.00     84.8±0.58µs        ? ?/sec     1.09     92.2±8.98µs        ? ?/sec
+ smol-songs.csv: basic filter: TO/Notstandskomitee       1.18     98.0±6.30µs        ? ?/sec     1.00     83.2±0.97µs        ? ?/sec
+ smol-songs.csv: basic with quote/"david" "bowie"        114.68    76.0±0.20ms        ? ?/sec    1.00    662.5±5.03µs        ? ?/sec
- smol-songs.csv: basic with quote/"john"                 1.00    197.4±1.06µs        ? ?/sec     1.05    208.1±1.53µs        ? ?/sec
+ smol-songs.csv: basic with quote/"michael" "jackson"    2.75      2.0±0.01ms        ? ?/sec     1.00    738.9±3.91µs        ? ?/sec
+ smol-songs.csv: basic without quote/david bowie         297.42  1499.3±0.86ms        ? ?/sec    1.00      5.0±0.02ms        ? ?/sec
+ smol-songs.csv: basic without quote/michael jackson     2.55      8.9±0.02ms        ? ?/sec     1.00      3.5±0.01ms        ? ?/sec
+ smol-songs.csv: big filter/john                         1.08    473.6±2.25µs        ? ?/sec     1.00    438.1±2.59µs        ? ?/sec
- smol-songs.csv: prefix search/a                         1.00    446.9±1.81µs        ? ?/sec     1.79    800.5±4.45µs        ? ?/sec
- smol-songs.csv: prefix search/b                         1.00    398.5±2.74µs        ? ?/sec     1.81    723.1±5.46µs        ? ?/sec
- smol-songs.csv: prefix search/i                         1.00    486.3±1.99µs        ? ?/sec     1.69    823.6±9.42µs        ? ?/sec
- smol-songs.csv: prefix search/s                         1.00    229.6±3.29µs        ? ?/sec     2.59    594.4±2.22µs        ? ?/sec
- smol-songs.csv: prefix search/x                         1.00    150.2±0.76µs        ? ?/sec     1.11    166.0±0.87µs        ? ?/sec
```

On songs, the new algorithm gives a big improvement on slow queries, and is slower on one char prefix search (fast queries <1ms).

#### wiki main VS refactor-attribute-criterion
```diff
  group                                                           search_wikimain_31c18f09               search_wikirefactor-attribute-criterion_1bd15d84
  -----                                                           ------------------------               ------------------------------------------------
- smol-wiki-articles.csv: basic with quote/"rock" "and" "roll"    1.00      3.2±0.01ms        ? ?/sec    1.15      3.7±0.01ms        ? ?/sec
- smol-wiki-articles.csv: basic without quote/film                1.00    351.5±2.47µs        ? ?/sec    1.13    396.8±1.63µs        ? ?/sec
+ smol-wiki-articles.csv: basic without quote/rock and roll       1.10      9.4±0.02ms        ? ?/sec    1.00      8.6±0.04ms        ? ?/sec
- smol-wiki-articles.csv: basic without quote/spain               1.00    446.0±3.23µs        ? ?/sec    1.11    496.6±7.75µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/c                         1.00    115.6±0.61µs        ? ?/sec    2.22    256.7±1.24µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/g                         1.00    189.7±2.03µs        ? ?/sec    1.57    297.0±1.35µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/j                         1.00    209.2±1.11µs        ? ?/sec    1.40    293.0±2.09µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/q                         1.00     79.0±0.44µs        ? ?/sec    1.10     87.2±0.69µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/t                         1.00    270.1±1.15µs        ? ?/sec    1.55    419.9±5.16µs        ? ?/sec
- smol-wiki-articles.csv: prefix search/x                         1.00    244.9±1.33µs        ? ?/sec    1.07    260.9±1.95µs        ? ?/sec
- smol-wiki-articles.csv: words/Abraham machin                    1.00      8.1±0.03ms        ? ?/sec    1.17      9.4±0.02ms        ? ?/sec
- smol-wiki-articles.csv: words/Idaho Bellevue pizza              1.00     19.3±0.07ms        ? ?/sec    1.07     20.6±0.05ms        ? ?/sec
```
On wiki we have some regressions `+17%` and `+15%` on request `>1ms`.

Co-authored-by: many <maxime@meilisearch.com>

http-ui/src/main.rs

@@ -361,6 +361,7 @@ async fn main() -> anyhow::Result<()> {
                         // We must use the write transaction of the update here.
                         let mut wtxn = index_cloned.write_txn()?;
                         let mut builder = update_builder.index_documents(&mut wtxn, &index_cloned);
+                        builder.enable_autogenerate_docids();
 
                         match method.as_str() {
                             "replace" => builder

infos/src/main.rs

@@ -7,7 +7,7 @@ use byte_unit::Byte;
 use heed::EnvOpenOptions;
 use milli::facet::FacetType;
 use milli::index::db_name::*;
-use milli::{FieldId, Index, TreeLevel};
+use milli::{FieldId, Index};
 use structopt::StructOpt;
 use Command::*;
 
@@ -22,8 +22,8 @@ const ALL_DATABASE_NAMES: &[&str] = &[
     DOCID_WORD_POSITIONS,
     WORD_PAIR_PROXIMITY_DOCIDS,
     WORD_PREFIX_PAIR_PROXIMITY_DOCIDS,
-    WORD_LEVEL_POSITION_DOCIDS,
+    WORD_POSITION_DOCIDS,
-    WORD_PREFIX_LEVEL_POSITION_DOCIDS,
+    WORD_PREFIX_POSITION_DOCIDS,
     FIELD_ID_WORD_COUNT_DOCIDS,
     FACET_ID_F64_DOCIDS,
     FACET_ID_STRING_DOCIDS,
@@ -281,10 +281,10 @@ fn main() -> anyhow::Result<()> {
             facet_values_docids(&index, &rtxn, !full_display, FacetType::String, field_name)
         }
         WordsLevelPositionsDocids { full_display, words } => {
-            words_level_positions_docids(&index, &rtxn, !full_display, words)
+            words_positions_docids(&index, &rtxn, !full_display, words)
         }
         WordPrefixesLevelPositionsDocids { full_display, prefixes } => {
-            word_prefixes_level_positions_docids(&index, &rtxn, !full_display, prefixes)
+            word_prefixes_positions_docids(&index, &rtxn, !full_display, prefixes)
         }
         FieldIdWordCountDocids { full_display, field_name } => {
             field_id_word_count_docids(&index, &rtxn, !full_display, field_name)
@@ -379,8 +379,8 @@ fn biggest_value_sizes(index: &Index, rtxn: &heed::RoTxn, limit: usize) -> anyho
         docid_word_positions,
         word_pair_proximity_docids,
         word_prefix_pair_proximity_docids,
-        word_level_position_docids,
+        word_position_docids,
-        word_prefix_level_position_docids,
+        word_prefix_position_docids,
         field_id_word_count_docids,
         facet_id_f64_docids,
         facet_id_string_docids,
@@ -395,8 +395,8 @@ fn biggest_value_sizes(index: &Index, rtxn: &heed::RoTxn, limit: usize) -> anyho
     let docid_word_positions_name = "docid_word_positions";
     let word_prefix_pair_proximity_docids_name = "word_prefix_pair_proximity_docids";
     let word_pair_proximity_docids_name = "word_pair_proximity_docids";
-    let word_level_position_docids_name = "word_level_position_docids";
+    let word_position_docids_name = "word_position_docids";
-    let word_prefix_level_position_docids_name = "word_prefix_level_position_docids";
+    let word_prefix_position_docids_name = "word_prefix_position_docids";
     let field_id_word_count_docids_name = "field_id_word_count_docids";
     let facet_id_f64_docids_name = "facet_id_f64_docids";
     let facet_id_string_docids_name = "facet_id_string_docids";
@@ -471,19 +471,19 @@ fn biggest_value_sizes(index: &Index, rtxn: &heed::RoTxn, limit: usize) -> anyho
             }
         }
 
-        for result in word_level_position_docids.remap_data_type::<ByteSlice>().iter(rtxn)? {
+        for result in word_position_docids.remap_data_type::<ByteSlice>().iter(rtxn)? {
-            let ((word, level, left, right), value) = result?;
+            let ((word, pos), value) = result?;
-            let key = format!("{} {} {:?}", word, level, left..=right);
+            let key = format!("{} {}", word, pos);
-            heap.push(Reverse((value.len(), key, word_level_position_docids_name)));
+            heap.push(Reverse((value.len(), key, word_position_docids_name)));
             if heap.len() > limit {
                 heap.pop();
             }
         }
 
-        for result in word_prefix_level_position_docids.remap_data_type::<ByteSlice>().iter(rtxn)? {
+        for result in word_prefix_position_docids.remap_data_type::<ByteSlice>().iter(rtxn)? {
-            let ((word, level, left, right), value) = result?;
+            let ((word, pos), value) = result?;
-            let key = format!("{} {} {:?}", word, level, left..=right);
+            let key = format!("{} {}", word, pos);
-            heap.push(Reverse((value.len(), key, word_prefix_level_position_docids_name)));
+            heap.push(Reverse((value.len(), key, word_prefix_position_docids_name)));
             if heap.len() > limit {
                 heap.pop();
             }
@@ -663,7 +663,7 @@ fn facet_values_docids(
     Ok(wtr.flush()?)
 }
 
-fn words_level_positions_docids(
+fn words_positions_docids(
     index: &Index,
     rtxn: &heed::RoTxn,
     debug: bool,
@@ -671,16 +671,16 @@ fn words_level_positions_docids(
 ) -> anyhow::Result<()> {
     let stdout = io::stdout();
     let mut wtr = csv::Writer::from_writer(stdout.lock());
-    wtr.write_record(&["word", "level", "positions", "documents_count", "documents_ids"])?;
+    wtr.write_record(&["word", "position", "documents_count", "documents_ids"])?;
 
     for word in words.iter().map(AsRef::as_ref) {
         let range = {
-            let left = (word, TreeLevel::min_value(), u32::min_value(), u32::min_value());
+            let left = (word, u32::min_value());
-            let right = (word, TreeLevel::max_value(), u32::max_value(), u32::max_value());
+            let right = (word, u32::max_value());
             left..=right
         };
-        for result in index.word_level_position_docids.range(rtxn, &range)? {
+        for result in index.word_position_docids.range(rtxn, &range)? {
-            let ((w, level, left, right), docids) = result?;
+            let ((w, pos), docids) = result?;
 
             let count = docids.len().to_string();
             let docids = if debug {
@@ -688,20 +688,15 @@ fn words_level_positions_docids(
             } else {
                 format!("{:?}", docids.iter().collect::<Vec<_>>())
             };
-            let position_range = if level == TreeLevel::min_value() {
+            let position = format!("{:?}", pos);
-                format!("{:?}", left)
+            wtr.write_record(&[w, &position, &count, &docids])?;
-            } else {
-                format!("{:?}", left..=right)
-            };
-            let level = level.to_string();
-            wtr.write_record(&[w, &level, &position_range, &count, &docids])?;
         }
     }
 
     Ok(wtr.flush()?)
 }
 
-fn word_prefixes_level_positions_docids(
+fn word_prefixes_positions_docids(
     index: &Index,
     rtxn: &heed::RoTxn,
     debug: bool,
@@ -709,16 +704,16 @@ fn word_prefixes_level_positions_docids(
 ) -> anyhow::Result<()> {
     let stdout = io::stdout();
     let mut wtr = csv::Writer::from_writer(stdout.lock());
-    wtr.write_record(&["prefix", "level", "positions", "documents_count", "documents_ids"])?;
+    wtr.write_record(&["prefix", "position", "documents_count", "documents_ids"])?;
 
     for word in prefixes.iter().map(AsRef::as_ref) {
         let range = {
-            let left = (word, TreeLevel::min_value(), u32::min_value(), u32::min_value());
+            let left = (word, u32::min_value());
-            let right = (word, TreeLevel::max_value(), u32::max_value(), u32::max_value());
+            let right = (word, u32::max_value());
             left..=right
         };
-        for result in index.word_prefix_level_position_docids.range(rtxn, &range)? {
+        for result in index.word_prefix_position_docids.range(rtxn, &range)? {
-            let ((w, level, left, right), docids) = result?;
+            let ((w, pos), docids) = result?;
 
             let count = docids.len().to_string();
             let docids = if debug {
@@ -726,13 +721,8 @@ fn word_prefixes_level_positions_docids(
             } else {
                 format!("{:?}", docids.iter().collect::<Vec<_>>())
             };
-            let position_range = if level == TreeLevel::min_value() {
+            let position = format!("{:?}", pos);
-                format!("{:?}", left)
+            wtr.write_record(&[w, &position, &count, &docids])?;
-            } else {
-                format!("{:?}", left..=right)
-            };
-            let level = level.to_string();
-            wtr.write_record(&[w, &level, &position_range, &count, &docids])?;
         }
     }
 
@@ -970,8 +960,8 @@ fn size_of_databases(index: &Index, rtxn: &heed::RoTxn, names: Vec<String>) -> a
         docid_word_positions,
         word_pair_proximity_docids,
         word_prefix_pair_proximity_docids,
-        word_level_position_docids,
+        word_position_docids,
-        word_prefix_level_position_docids,
+        word_prefix_position_docids,
         field_id_word_count_docids,
         facet_id_f64_docids,
         facet_id_string_docids,
@@ -994,8 +984,8 @@ fn size_of_databases(index: &Index, rtxn: &heed::RoTxn, names: Vec<String>) -> a
             DOCID_WORD_POSITIONS => docid_word_positions.as_polymorph(),
             WORD_PAIR_PROXIMITY_DOCIDS => word_pair_proximity_docids.as_polymorph(),
             WORD_PREFIX_PAIR_PROXIMITY_DOCIDS => word_prefix_pair_proximity_docids.as_polymorph(),
-            WORD_LEVEL_POSITION_DOCIDS => word_level_position_docids.as_polymorph(),
+            WORD_POSITION_DOCIDS => word_position_docids.as_polymorph(),
-            WORD_PREFIX_LEVEL_POSITION_DOCIDS => word_prefix_level_position_docids.as_polymorph(),
+            WORD_PREFIX_POSITION_DOCIDS => word_prefix_position_docids.as_polymorph(),
             FIELD_ID_WORD_COUNT_DOCIDS => field_id_word_count_docids.as_polymorph(),
             FACET_ID_F64_DOCIDS => facet_id_f64_docids.as_polymorph(),
             FACET_ID_STRING_DOCIDS => facet_id_string_docids.as_polymorph(),

milli/src/heed_codec/mod.rs

@@ -4,7 +4,7 @@ mod field_id_word_count_codec;
 mod obkv_codec;
 mod roaring_bitmap;
 mod roaring_bitmap_length;
-mod str_level_position_codec;
+mod str_beu32_codec;
 mod str_str_u8_codec;
 
 pub use self::beu32_str_codec::BEU32StrCodec;
@@ -14,5 +14,5 @@ pub use self::roaring_bitmap::{BoRoaringBitmapCodec, CboRoaringBitmapCodec, Roar
 pub use self::roaring_bitmap_length::{
     BoRoaringBitmapLenCodec, CboRoaringBitmapLenCodec, RoaringBitmapLenCodec,
 };
-pub use self::str_level_position_codec::StrLevelPositionCodec;
+pub use self::str_beu32_codec::StrBEU32Codec;
 pub use self::str_str_u8_codec::StrStrU8Codec;

milli/src/heed_codec/str_beu32_codec.rs

@@ -0,0 +1,38 @@
+use std::borrow::Cow;
+use std::convert::TryInto;
+use std::mem::size_of;
+use std::str;
+
+pub struct StrBEU32Codec;
+
+impl<'a> heed::BytesDecode<'a> for StrBEU32Codec {
+    type DItem = (&'a str, u32);
+
+    fn bytes_decode(bytes: &'a [u8]) -> Option<Self::DItem> {
+        let footer_len = size_of::<u32>();
+
+        if bytes.len() < footer_len {
+            return None;
+        }
+
+        let (word, bytes) = bytes.split_at(bytes.len() - footer_len);
+        let word = str::from_utf8(word).ok()?;
+        let pos = bytes.try_into().map(u32::from_be_bytes).ok()?;
+
+        Some((word, pos))
+    }
+}
+
+impl<'a> heed::BytesEncode<'a> for StrBEU32Codec {
+    type EItem = (&'a str, u32);
+
+    fn bytes_encode((word, pos): &Self::EItem) -> Option<Cow<[u8]>> {
+        let pos = pos.to_be_bytes();
+
+        let mut bytes = Vec::with_capacity(word.len() + pos.len());
+        bytes.extend_from_slice(word.as_bytes());
+        bytes.extend_from_slice(&pos[..]);
+
+        Some(Cow::Owned(bytes))
+    }
+}

milli/src/heed_codec/str_level_position_codec.rs

@@ -1,47 +0,0 @@
-use std::borrow::Cow;
-use std::convert::{TryFrom, TryInto};
-use std::mem::size_of;
-use std::str;
-
-use crate::TreeLevel;
-
-pub struct StrLevelPositionCodec;
-
-impl<'a> heed::BytesDecode<'a> for StrLevelPositionCodec {
-    type DItem = (&'a str, TreeLevel, u32, u32);
-
-    fn bytes_decode(bytes: &'a [u8]) -> Option<Self::DItem> {
-        let footer_len = size_of::<u8>() + size_of::<u32>() * 2;
-
-        if bytes.len() < footer_len {
-            return None;
-        }
-
-        let (word, bytes) = bytes.split_at(bytes.len() - footer_len);
-        let word = str::from_utf8(word).ok()?;
-
-        let (level, bytes) = bytes.split_first()?;
-        let left = bytes[..4].try_into().map(u32::from_be_bytes).ok()?;
-        let right = bytes[4..].try_into().map(u32::from_be_bytes).ok()?;
-        let level = TreeLevel::try_from(*level).ok()?;
-
-        Some((word, level, left, right))
-    }
-}
-
-impl<'a> heed::BytesEncode<'a> for StrLevelPositionCodec {
-    type EItem = (&'a str, TreeLevel, u32, u32);
-
-    fn bytes_encode((word, level, left, right): &Self::EItem) -> Option<Cow<[u8]>> {
-        let left = left.to_be_bytes();
-        let right = right.to_be_bytes();
-
-        let mut bytes = Vec::with_capacity(word.len() + 1 + left.len() + right.len());
-        bytes.extend_from_slice(word.as_bytes());
-        bytes.push((*level).into());
-        bytes.extend_from_slice(&left[..]);
-        bytes.extend_from_slice(&right[..]);
-
-        Some(Cow::Owned(bytes))
-    }
-}

milli/src/index.rs

@@ -20,7 +20,7 @@ use crate::{
     default_criteria, BEU32StrCodec, BoRoaringBitmapCodec, CboRoaringBitmapCodec, Criterion,
     DocumentId, ExternalDocumentsIds, FacetDistribution, FieldDistribution, FieldId,
     FieldIdWordCountCodec, GeoPoint, ObkvCodec, Result, RoaringBitmapCodec, RoaringBitmapLenCodec,
-    Search, StrLevelPositionCodec, StrStrU8Codec, BEU32,
+    Search, StrBEU32Codec, StrStrU8Codec, BEU32,
 };
 
 pub mod main_key {
@@ -55,8 +55,8 @@ pub mod db_name {
     pub const DOCID_WORD_POSITIONS: &str = "docid-word-positions";
     pub const WORD_PAIR_PROXIMITY_DOCIDS: &str = "word-pair-proximity-docids";
     pub const WORD_PREFIX_PAIR_PROXIMITY_DOCIDS: &str = "word-prefix-pair-proximity-docids";
-    pub const WORD_LEVEL_POSITION_DOCIDS: &str = "word-level-position-docids";
+    pub const WORD_POSITION_DOCIDS: &str = "word-position-docids";
-    pub const WORD_PREFIX_LEVEL_POSITION_DOCIDS: &str = "word-prefix-level-position-docids";
+    pub const WORD_PREFIX_POSITION_DOCIDS: &str = "word-prefix-position-docids";
     pub const FIELD_ID_WORD_COUNT_DOCIDS: &str = "field-id-word-count-docids";
     pub const FACET_ID_F64_DOCIDS: &str = "facet-id-f64-docids";
     pub const FACET_ID_STRING_DOCIDS: &str = "facet-id-string-docids";
@@ -86,12 +86,12 @@ pub struct Index {
     /// Maps the proximity between a pair of word and prefix with all the docids where this relation appears.
     pub word_prefix_pair_proximity_docids: Database<StrStrU8Codec, CboRoaringBitmapCodec>,
 
-    /// Maps the word, level and position range with the docids that corresponds to it.
+    /// Maps the word and the position with the docids that corresponds to it.
-    pub word_level_position_docids: Database<StrLevelPositionCodec, CboRoaringBitmapCodec>,
+    pub word_position_docids: Database<StrBEU32Codec, CboRoaringBitmapCodec>,
     /// Maps the field id and the word count with the docids that corresponds to it.
     pub field_id_word_count_docids: Database<FieldIdWordCountCodec, CboRoaringBitmapCodec>,
-    /// Maps the level positions of a word prefix with all the docids where this prefix appears.
+    /// Maps the position of a word prefix with all the docids where this prefix appears.
-    pub word_prefix_level_position_docids: Database<StrLevelPositionCodec, CboRoaringBitmapCodec>,
+    pub word_prefix_position_docids: Database<StrBEU32Codec, CboRoaringBitmapCodec>,
 
     /// Maps the facet field id, level and the number with the docids that corresponds to it.
     pub facet_id_f64_docids: Database<FacetLevelValueF64Codec, CboRoaringBitmapCodec>,
@@ -122,10 +122,9 @@ impl Index {
         let word_pair_proximity_docids = env.create_database(Some(WORD_PAIR_PROXIMITY_DOCIDS))?;
         let word_prefix_pair_proximity_docids =
             env.create_database(Some(WORD_PREFIX_PAIR_PROXIMITY_DOCIDS))?;
-        let word_level_position_docids = env.create_database(Some(WORD_LEVEL_POSITION_DOCIDS))?;
+        let word_position_docids = env.create_database(Some(WORD_POSITION_DOCIDS))?;
         let field_id_word_count_docids = env.create_database(Some(FIELD_ID_WORD_COUNT_DOCIDS))?;
-        let word_prefix_level_position_docids =
+        let word_prefix_position_docids = env.create_database(Some(WORD_PREFIX_POSITION_DOCIDS))?;
-            env.create_database(Some(WORD_PREFIX_LEVEL_POSITION_DOCIDS))?;
         let facet_id_f64_docids = env.create_database(Some(FACET_ID_F64_DOCIDS))?;
         let facet_id_string_docids = env.create_database(Some(FACET_ID_STRING_DOCIDS))?;
         let field_id_docid_facet_f64s = env.create_database(Some(FIELD_ID_DOCID_FACET_F64S))?;
@@ -143,8 +142,8 @@ impl Index {
             docid_word_positions,
             word_pair_proximity_docids,
             word_prefix_pair_proximity_docids,
-            word_level_position_docids,
+            word_position_docids,
-            word_prefix_level_position_docids,
+            word_prefix_position_docids,
             field_id_word_count_docids,
             facet_id_f64_docids,
             facet_id_string_docids,

milli/src/lib.rs

@@ -14,7 +14,6 @@ pub mod heed_codec;
 pub mod index;
 pub mod proximity;
 mod search;
-pub mod tree_level;
 pub mod update;
 
 use std::collections::{BTreeMap, HashMap};
@@ -35,11 +34,10 @@ pub use self::fields_ids_map::FieldsIdsMap;
 pub use self::heed_codec::{
     BEU32StrCodec, BoRoaringBitmapCodec, BoRoaringBitmapLenCodec, CboRoaringBitmapCodec,
     CboRoaringBitmapLenCodec, FieldIdWordCountCodec, ObkvCodec, RoaringBitmapCodec,
-    RoaringBitmapLenCodec, StrLevelPositionCodec, StrStrU8Codec,
+    RoaringBitmapLenCodec, StrBEU32Codec, StrStrU8Codec,
 };
 pub use self::index::Index;
 pub use self::search::{FacetDistribution, FilterCondition, MatchingWords, Search, SearchResult};
-pub use self::tree_level::TreeLevel;
 
 pub type Result<T> = std::result::Result<T, error::Error>;
 

milli/src/search/criteria/attribute.rs

@@ -1,7 +1,7 @@
-use std::borrow::Cow;
 use std::cmp::{self, Ordering};
 use std::collections::binary_heap::PeekMut;
 use std::collections::{btree_map, BTreeMap, BinaryHeap, HashMap};
+use std::iter::Peekable;
 use std::mem::take;
 
 use roaring::RoaringBitmap;
@@ -10,20 +10,16 @@ use super::{resolve_query_tree, Context, Criterion, CriterionParameters, Criteri
 use crate::search::criteria::Query;
 use crate::search::query_tree::{Operation, QueryKind};
 use crate::search::{build_dfa, word_derivations, WordDerivationsCache};
-use crate::{Result, TreeLevel};
+use crate::Result;
 
 /// To be able to divide integers by the number of words in the query
 /// we want to find a multiplier that allow us to divide by any number between 1 and 10.
 /// We chose the LCM of all numbers between 1 and 10 as the multiplier (https://en.wikipedia.org/wiki/Least_common_multiple).
 const LCM_10_FIRST_NUMBERS: u32 = 2520;
 
-/// To compute the interval size of a level,
-/// we use 4 as the exponentiation base and the level as the exponent.
-const LEVEL_EXPONENTIATION_BASE: u32 = 4;
-
 /// Threshold on the number of candidates that will make
 /// the system to choose between one algorithm or another.
-const CANDIDATES_THRESHOLD: u64 = 1000;
+const CANDIDATES_THRESHOLD: u64 = 500;
 
 type FlattenedQueryTree = Vec<Vec<Vec<Query>>>;
 
@@ -32,7 +28,8 @@ pub struct Attribute<'t> {
     state: Option<(Operation, FlattenedQueryTree, RoaringBitmap)>,
     bucket_candidates: RoaringBitmap,
     parent: Box<dyn Criterion + 't>,
-    current_buckets: Option<btree_map::IntoIter<u64, RoaringBitmap>>,
+    linear_buckets: Option<btree_map::IntoIter<u64, RoaringBitmap>>,
+    set_buckets: Option<BinaryHeap<Branch<'t>>>,
 }
 
 impl<'t> Attribute<'t> {
@@ -42,7 +39,8 @@ impl<'t> Attribute<'t> {
             state: None,
             bucket_candidates: RoaringBitmap::new(),
             parent,
-            current_buckets: None,
+            linear_buckets: None,
+            set_buckets: None,
         }
     }
 }
@@ -67,19 +65,19 @@ impl<'t> Criterion for Attribute<'t> {
                 }
                 Some((query_tree, flattened_query_tree, mut allowed_candidates)) => {
                     let found_candidates = if allowed_candidates.len() < CANDIDATES_THRESHOLD {
-                        let current_buckets = match self.current_buckets.as_mut() {
+                        let linear_buckets = match self.linear_buckets.as_mut() {
-                            Some(current_buckets) => current_buckets,
+                            Some(linear_buckets) => linear_buckets,
                             None => {
-                                let new_buckets = linear_compute_candidates(
+                                let new_buckets = initialize_linear_buckets(
                                     self.ctx,
                                     &flattened_query_tree,
                                     &allowed_candidates,
                                 )?;
-                                self.current_buckets.get_or_insert(new_buckets.into_iter())
+                                self.linear_buckets.get_or_insert(new_buckets.into_iter())
                             }
                         };
 
-                        match current_buckets.next() {
+                        match linear_buckets.next() {
                             Some((_score, candidates)) => candidates,
                             None => {
                                 return Ok(Some(CriterionResult {
@@ -91,13 +89,21 @@ impl<'t> Criterion for Attribute<'t> {
                             }
                         }
                     } else {
-                        match set_compute_candidates(
+                        let mut set_buckets = match self.set_buckets.as_mut() {
-                            self.ctx,
+                            Some(set_buckets) => set_buckets,
-                            &flattened_query_tree,
+                            None => {
-                            &allowed_candidates,
+                                let new_buckets = initialize_set_buckets(
-                            params.wdcache,
+                                    self.ctx,
-                        )? {
+                                    &flattened_query_tree,
-                            Some(candidates) => candidates,
+                                    &allowed_candidates,
+                                    params.wdcache,
+                                )?;
+                                self.set_buckets.get_or_insert(new_buckets)
+                            }
+                        };
+
+                        match set_compute_candidates(&mut set_buckets, &allowed_candidates)? {
+                            Some((_score, candidates)) => candidates,
                             None => {
                                 return Ok(Some(CriterionResult {
                                     query_tree: Some(query_tree),
@@ -148,7 +154,7 @@ impl<'t> Criterion for Attribute<'t> {
                         }
 
                         self.state = Some((query_tree, flattened_query_tree, candidates));
-                        self.current_buckets = None;
+                        self.linear_buckets = None;
                     }
                     Some(CriterionResult {
                         query_tree: None,
@@ -170,142 +176,33 @@ impl<'t> Criterion for Attribute<'t> {
     }
 }
 
-/// WordLevelIterator is an pseudo-Iterator over intervals of word-position for one word,
+/// QueryPositionIterator is an Iterator over positions of a Query,
-/// it will begin at the first non-empty interval and will return every interval without
+/// It contains iterators over words positions.
-/// jumping over empty intervals.
+struct QueryPositionIterator<'t> {
-struct WordLevelIterator<'t, 'q> {
+    inner:
-    inner: Box<
+        Vec<Peekable<Box<dyn Iterator<Item = heed::Result<((&'t str, u32), RoaringBitmap)>> + 't>>>,
-        dyn Iterator<Item = heed::Result<((&'t str, TreeLevel, u32, u32), RoaringBitmap)>> + 't,
-    >,
-    level: TreeLevel,
-    interval_size: u32,
-    word: Cow<'q, str>,
-    in_prefix_cache: bool,
-    inner_next: Option<(u32, u32, RoaringBitmap)>,
-    current_interval: Option<(u32, u32)>,
 }
 
-impl<'t, 'q> WordLevelIterator<'t, 'q> {
+impl<'t> QueryPositionIterator<'t> {
     fn new(
         ctx: &'t dyn Context<'t>,
-        word: Cow<'q, str>,
+        queries: &[Query],
-        in_prefix_cache: bool,
-    ) -> heed::Result<Option<Self>> {
-        match ctx.word_position_last_level(&word, in_prefix_cache)? {
-            Some(_) => {
-                // HOTFIX Meilisearch#1707: it is better to only iterate over level 0 for performances reasons.
-                let level = TreeLevel::min_value();
-                let interval_size = LEVEL_EXPONENTIATION_BASE.pow(Into::<u8>::into(level) as u32);
-                let inner =
-                    ctx.word_position_iterator(&word, level, in_prefix_cache, None, None)?;
-                Ok(Some(Self {
-                    inner,
-                    level,
-                    interval_size,
-                    word,
-                    in_prefix_cache,
-                    inner_next: None,
-                    current_interval: None,
-                }))
-            }
-            None => Ok(None),
-        }
-    }
-
-    fn dig(
-        &self,
-        ctx: &'t dyn Context<'t>,
-        level: &TreeLevel,
-        left_interval: Option<u32>,
-    ) -> heed::Result<Self> {
-        let level = *level.min(&self.level);
-        let interval_size = LEVEL_EXPONENTIATION_BASE.pow(Into::<u8>::into(level) as u32);
-        let word = self.word.clone();
-        let in_prefix_cache = self.in_prefix_cache;
-        let inner =
-            ctx.word_position_iterator(&word, level, in_prefix_cache, left_interval, None)?;
-
-        Ok(Self {
-            inner,
-            level,
-            interval_size,
-            word,
-            in_prefix_cache,
-            inner_next: None,
-            current_interval: None,
-        })
-    }
-
-    fn next(&mut self) -> heed::Result<Option<(u32, u32, RoaringBitmap)>> {
-        fn is_next_interval(last_right: u32, next_left: u32) -> bool {
-            last_right + 1 == next_left
-        }
-
-        let inner_next = match self.inner_next.take() {
-            Some(inner_next) => Some(inner_next),
-            None => self
-                .inner
-                .next()
-                .transpose()?
-                .map(|((_, _, left, right), docids)| (left, right, docids)),
-        };
-
-        match inner_next {
-            Some((left, right, docids)) => match self.current_interval {
-                Some((last_left, last_right)) if !is_next_interval(last_right, left) => {
-                    let blank_left = last_left + self.interval_size;
-                    let blank_right = last_right + self.interval_size;
-                    self.current_interval = Some((blank_left, blank_right));
-                    self.inner_next = Some((left, right, docids));
-                    Ok(Some((blank_left, blank_right, RoaringBitmap::new())))
-                }
-                _ => {
-                    self.current_interval = Some((left, right));
-                    Ok(Some((left, right, docids)))
-                }
-            },
-            None => Ok(None),
-        }
-    }
-}
-
-/// QueryLevelIterator is an pseudo-Iterator for a Query,
-/// It contains WordLevelIterators and is chainned with other QueryLevelIterator.
-struct QueryLevelIterator<'t, 'q> {
-    parent: Option<Box<QueryLevelIterator<'t, 'q>>>,
-    inner: Vec<WordLevelIterator<'t, 'q>>,
-    level: TreeLevel,
-    accumulator: Vec<Option<(u32, u32, RoaringBitmap)>>,
-    parent_accumulator: Vec<Option<(u32, u32, RoaringBitmap)>>,
-    interval_to_skip: usize,
-}
-
-impl<'t, 'q> QueryLevelIterator<'t, 'q> {
-    fn new(
-        ctx: &'t dyn Context<'t>,
-        queries: &'q [Query],
         wdcache: &mut WordDerivationsCache,
-    ) -> Result<Option<Self>> {
+    ) -> Result<Self> {
         let mut inner = Vec::with_capacity(queries.len());
         for query in queries {
+            let in_prefix_cache = query.prefix && ctx.in_prefix_cache(query.kind.word());
             match &query.kind {
                 QueryKind::Exact { word, .. } => {
-                    if !query.prefix || ctx.in_prefix_cache(&word) {
+                    if !query.prefix || in_prefix_cache {
-                        let word = Cow::Borrowed(query.kind.word());
+                        let word = query.kind.word();
-                        if let Some(word_level_iterator) =
+                        let iter = ctx.word_position_iterator(word, in_prefix_cache)?;
-                            WordLevelIterator::new(ctx, word, query.prefix)?
+                        inner.push(iter.peekable());
-                        {
-                            inner.push(word_level_iterator);
-                        }
                     } else {
                         for (word, _) in word_derivations(&word, true, 0, ctx.words_fst(), wdcache)?
                         {
-                            let word = Cow::Owned(word.to_owned());
+                            let iter = ctx.word_position_iterator(&word, in_prefix_cache)?;
-                            if let Some(word_level_iterator) =
+                            inner.push(iter.peekable());
-                                WordLevelIterator::new(ctx, word, false)?
-                            {
-                                inner.push(word_level_iterator);
-                            }
                         }
                     }
                 }
@@ -313,360 +210,247 @@ impl<'t, 'q> QueryLevelIterator<'t, 'q> {
                     for (word, _) in
                         word_derivations(&word, query.prefix, *typo, ctx.words_fst(), wdcache)?
                     {
-                        let word = Cow::Owned(word.to_owned());
+                        let iter = ctx.word_position_iterator(&word, in_prefix_cache)?;
-                        if let Some(word_level_iterator) = WordLevelIterator::new(ctx, word, false)?
+                        inner.push(iter.peekable());
-                        {
-                            inner.push(word_level_iterator);
-                        }
                     }
                 }
-            }
+            };
         }
 
-        let highest = inner.iter().max_by_key(|wli| wli.level).map(|wli| wli.level);
+        Ok(Self { inner })
-        match highest {
-            Some(level) => Ok(Some(Self {
-                parent: None,
-                inner,
-                level,
-                accumulator: vec![],
-                parent_accumulator: vec![],
-                interval_to_skip: 0,
-            })),
-            None => Ok(None),
-        }
-    }
-
-    fn parent(&mut self, parent: QueryLevelIterator<'t, 'q>) -> &Self {
-        self.parent = Some(Box::new(parent));
-        self
-    }
-
-    /// create a new QueryLevelIterator with a lower level than the current one.
-    fn dig(&self, ctx: &'t dyn Context<'t>) -> heed::Result<Self> {
-        let (level, parent) = match &self.parent {
-            Some(parent) => {
-                let parent = parent.dig(ctx)?;
-                (parent.level.min(self.level), Some(Box::new(parent)))
-            }
-            None => (self.level.saturating_sub(1), None),
-        };
-
-        let left_interval = self
-            .accumulator
-            .get(self.interval_to_skip)
-            .map(|opt| opt.as_ref().map(|(left, _, _)| *left))
-            .flatten();
-        let mut inner = Vec::with_capacity(self.inner.len());
-        for word_level_iterator in self.inner.iter() {
-            inner.push(word_level_iterator.dig(ctx, &level, left_interval)?);
-        }
-
-        Ok(Self {
-            parent,
-            inner,
-            level,
-            accumulator: vec![],
-            parent_accumulator: vec![],
-            interval_to_skip: 0,
-        })
-    }
-
-    fn inner_next(&mut self, level: TreeLevel) -> heed::Result<Option<(u32, u32, RoaringBitmap)>> {
-        let mut accumulated: Option<(u32, u32, RoaringBitmap)> = None;
-        let u8_level = Into::<u8>::into(level);
-        let interval_size = LEVEL_EXPONENTIATION_BASE.pow(u8_level as u32);
-        for wli in self.inner.iter_mut() {
-            let wli_u8_level = Into::<u8>::into(wli.level);
-            let accumulated_count = LEVEL_EXPONENTIATION_BASE.pow((u8_level - wli_u8_level) as u32);
-            for _ in 0..accumulated_count {
-                if let Some((next_left, _, next_docids)) = wli.next()? {
-                    accumulated = match accumulated.take() {
-                        Some((acc_left, acc_right, mut acc_docids)) => {
-                            acc_docids |= next_docids;
-                            Some((acc_left, acc_right, acc_docids))
-                        }
-                        None => Some((next_left, next_left + interval_size, next_docids)),
-                    };
-                }
-            }
-        }
-
-        Ok(accumulated)
-    }
-
-    /// return the next meta-interval created from inner WordLevelIterators,
-    /// and from eventual chainned QueryLevelIterator.
-    fn next(
-        &mut self,
-        allowed_candidates: &RoaringBitmap,
-        tree_level: TreeLevel,
-    ) -> heed::Result<Option<(u32, u32, RoaringBitmap)>> {
-        let parent_result = match self.parent.as_mut() {
-            Some(parent) => Some(parent.next(allowed_candidates, tree_level)?),
-            None => None,
-        };
-
-        match parent_result {
-            Some(parent_next) => {
-                let inner_next = self.inner_next(tree_level)?;
-                self.interval_to_skip += interval_to_skip(
-                    &self.parent_accumulator,
-                    &self.accumulator,
-                    self.interval_to_skip,
-                    allowed_candidates,
-                );
-                self.accumulator.push(inner_next);
-                self.parent_accumulator.push(parent_next);
-                let mut merged_interval: Option<(u32, u32, RoaringBitmap)> = None;
-
-                for current in self
-                    .accumulator
-                    .iter()
-                    .rev()
-                    .zip(self.parent_accumulator.iter())
-                    .skip(self.interval_to_skip)
-                {
-                    if let (Some((left_a, right_a, a)), Some((left_b, right_b, b))) = current {
-                        match merged_interval.as_mut() {
-                            Some((_, _, merged_docids)) => *merged_docids |= a & b,
-                            None => {
-                                merged_interval = Some((left_a + left_b, right_a + right_b, a & b))
-                            }
-                        }
-                    }
-                }
-                Ok(merged_interval)
-            }
-            None => {
-                let level = self.level;
-                match self.inner_next(level)? {
-                    Some((left, right, mut candidates)) => {
-                        self.accumulator = vec![Some((left, right, RoaringBitmap::new()))];
-                        candidates &= allowed_candidates;
-                        Ok(Some((left, right, candidates)))
-                    }
-                    None => {
-                        self.accumulator = vec![None];
-                        Ok(None)
-                    }
-                }
-            }
-        }
     }
 }
 
-/// Count the number of interval that can be skiped when we make the cross-intersections
+impl<'t> Iterator for QueryPositionIterator<'t> {
-/// in order to compute the next meta-interval.
+    type Item = heed::Result<(u32, RoaringBitmap)>;
-/// A pair of intervals is skiped when both intervals doesn't contain any allowed docids.
+
-fn interval_to_skip(
+    fn next(&mut self) -> Option<Self::Item> {
-    parent_accumulator: &[Option<(u32, u32, RoaringBitmap)>],
+        // sort inner words from the closest next position to the farthest next position.
-    current_accumulator: &[Option<(u32, u32, RoaringBitmap)>],
+        let expected_pos = self
-    already_skiped: usize,
+            .inner
-    allowed_candidates: &RoaringBitmap,
+            .iter_mut()
-) -> usize {
+            .filter_map(|wli| match wli.peek() {
-    parent_accumulator
+                Some(Ok(((_, pos), _))) => Some(*pos),
-        .iter()
+                _ => None,
-        .zip(current_accumulator.iter())
+            })
-        .skip(already_skiped)
+            .min()?;
-        .take_while(|(parent, current)| {
+
-            let skip_parent = parent.as_ref().map_or(true, |(_, _, docids)| docids.is_empty());
+        let mut candidates = None;
-            let skip_current = current
+        for wli in self.inner.iter_mut() {
-                .as_ref()
+            if let Some(Ok(((_, pos), _))) = wli.peek() {
-                .map_or(true, |(_, _, docids)| docids.is_disjoint(allowed_candidates));
+                if *pos > expected_pos {
-            skip_parent && skip_current
+                    continue;
-        })
+                }
-        .count()
+            }
+
+            match wli.next() {
+                Some(Ok((_, docids))) => {
+                    candidates = match candidates.take() {
+                        Some(candidates) => Some(candidates | docids),
+                        None => Some(docids),
+                    }
+                }
+                Some(Err(e)) => return Some(Err(e)),
+                None => continue,
+            }
+        }
+
+        candidates.map(|candidates| Ok((expected_pos, candidates)))
+    }
 }
 
 /// A Branch is represent a possible alternative of the original query and is build with the Query Tree,
-/// This branch allows us to iterate over meta-interval of position and to dig in it if it contains interesting candidates.
+/// This branch allows us to iterate over meta-interval of positions.
-struct Branch<'t, 'q> {
+struct Branch<'t> {
-    query_level_iterator: QueryLevelIterator<'t, 'q>,
+    query_level_iterator: Vec<(u32, RoaringBitmap, Peekable<QueryPositionIterator<'t>>)>,
-    last_result: (u32, u32, RoaringBitmap),
+    last_result: (u32, RoaringBitmap),
-    tree_level: TreeLevel,
     branch_size: u32,
 }
 
-impl<'t, 'q> Branch<'t, 'q> {
+impl<'t> Branch<'t> {
+    fn new(
+        ctx: &'t dyn Context<'t>,
+        flatten_branch: &[Vec<Query>],
+        wdcache: &mut WordDerivationsCache,
+        allowed_candidates: &RoaringBitmap,
+    ) -> Result<Self> {
+        let mut query_level_iterator = Vec::new();
+        for queries in flatten_branch {
+            let mut qli = QueryPositionIterator::new(ctx, queries, wdcache)?.peekable();
+            let (pos, docids) = qli.next().transpose()?.unwrap_or((0, RoaringBitmap::new()));
+            query_level_iterator.push((pos, docids & allowed_candidates, qli));
+        }
+
+        let mut branch = Self {
+            query_level_iterator,
+            last_result: (0, RoaringBitmap::new()),
+            branch_size: flatten_branch.len() as u32,
+        };
+
+        branch.update_last_result();
+
+        Ok(branch)
+    }
+
     /// return the next meta-interval of the branch,
     /// and update inner interval in order to be ranked by the BinaryHeap.
     fn next(&mut self, allowed_candidates: &RoaringBitmap) -> heed::Result<bool> {
-        let tree_level = self.query_level_iterator.level;
+        // update the first query.
-        match self.query_level_iterator.next(allowed_candidates, tree_level)? {
+        let index = self.lowest_iterator_index();
-            Some(last_result) => {
+        match self.query_level_iterator.get_mut(index) {
-                self.last_result = last_result;
+            Some((cur_pos, cur_docids, qli)) => match qli.next().transpose()? {
-                self.tree_level = tree_level;
+                Some((next_pos, next_docids)) => {
-                Ok(true)
+                    *cur_pos = next_pos;
-            }
+                    *cur_docids |= next_docids & allowed_candidates;
+                    self.update_last_result();
+                    Ok(true)
+                }
+                None => Ok(false),
+            },
             None => Ok(false),
         }
     }
 
-    /// make the current Branch iterate over smaller intervals.
+    fn lowest_iterator_index(&mut self) -> usize {
-    fn dig(&mut self, ctx: &'t dyn Context<'t>) -> heed::Result<()> {
+        let (index, _) = self
-        self.query_level_iterator = self.query_level_iterator.dig(ctx)?;
+            .query_level_iterator
-        Ok(())
+            .iter_mut()
+            .map(|(pos, docids, qli)| {
+                if docids.is_empty() {
+                    0
+                } else {
+                    match qli.peek() {
+                        Some(result) => {
+                            result.as_ref().map(|(next_pos, _)| *next_pos - *pos).unwrap_or(0)
+                        }
+                        None => u32::MAX,
+                    }
+                }
+            })
+            .enumerate()
+            .min_by_key(|(_, diff)| *diff)
+            .unwrap_or((0, 0));
+
+        index
     }
 
-    /// because next() method could be time consuming,
+    fn update_last_result(&mut self) {
-    /// update inner interval in order to be ranked by the binary_heap without computing it,
+        let mut result_pos = 0;
-    /// the next() method should be called when the real interval is needed.
+        let mut result_docids = None;
-    fn lazy_next(&mut self) {
-        let u8_level = Into::<u8>::into(self.tree_level);
-        let interval_size = LEVEL_EXPONENTIATION_BASE.pow(u8_level as u32);
-        let (left, right, _) = self.last_result;
 
-        self.last_result = (left + interval_size, right + interval_size, RoaringBitmap::new());
+        for (pos, docids, _qli) in self.query_level_iterator.iter() {
+            result_pos += pos;
+            result_docids = result_docids
+                .take()
+                .map_or_else(|| Some(docids.clone()), |candidates| Some(candidates & docids));
+        }
+
+        // remove last result docids from inner iterators
+        if let Some(docids) = result_docids.as_ref() {
+            for (_, query_docids, _) in self.query_level_iterator.iter_mut() {
+                *query_docids -= docids;
+            }
+        }
+
+        self.last_result = (result_pos, result_docids.unwrap_or_default());
     }
 
     /// return the score of the current inner interval.
     fn compute_rank(&self) -> u32 {
-        // we compute a rank from the left interval.
+        // we compute a rank from the position.
-        let (left, _, _) = self.last_result;
+        let (pos, _) = self.last_result;
-        left.saturating_sub((0..self.branch_size).sum()) * LCM_10_FIRST_NUMBERS / self.branch_size
+        pos.saturating_sub((0..self.branch_size).sum()) * LCM_10_FIRST_NUMBERS / self.branch_size
     }
 
     fn cmp(&self, other: &Self) -> Ordering {
         let self_rank = self.compute_rank();
         let other_rank = other.compute_rank();
-        let left_cmp = self_rank.cmp(&other_rank);
-        // on level: lower is better,
-        // we want to dig faster into levels on interesting branches.
-        let level_cmp = self.tree_level.cmp(&other.tree_level);
 
-        left_cmp.then(level_cmp).then(self.last_result.2.len().cmp(&other.last_result.2.len()))
+        // lower rank is better, and because BinaryHeap give the higher ranked branch, we reverse it.
+        self_rank.cmp(&other_rank).reverse()
     }
 }
 
-impl<'t, 'q> Ord for Branch<'t, 'q> {
+impl<'t> Ord for Branch<'t> {
     fn cmp(&self, other: &Self) -> Ordering {
         self.cmp(other)
     }
 }
 
-impl<'t, 'q> PartialOrd for Branch<'t, 'q> {
+impl<'t> PartialOrd for Branch<'t> {
     fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
         Some(self.cmp(other))
     }
 }
 
-impl<'t, 'q> PartialEq for Branch<'t, 'q> {
+impl<'t> PartialEq for Branch<'t> {
     fn eq(&self, other: &Self) -> bool {
         self.cmp(other) == Ordering::Equal
     }
 }
 
-impl<'t, 'q> Eq for Branch<'t, 'q> {}
+impl<'t> Eq for Branch<'t> {}
 
-fn initialize_query_level_iterators<'t, 'q>(
+fn initialize_set_buckets<'t>(
-    ctx: &'t dyn Context<'t>,
-    branches: &'q FlattenedQueryTree,
-    allowed_candidates: &RoaringBitmap,
-    wdcache: &mut WordDerivationsCache,
-) -> Result<BinaryHeap<Branch<'t, 'q>>> {
-    let mut positions = BinaryHeap::with_capacity(branches.len());
-    for branch in branches {
-        let mut branch_positions = Vec::with_capacity(branch.len());
-        for queries in branch {
-            match QueryLevelIterator::new(ctx, queries, wdcache)? {
-                Some(qli) => branch_positions.push(qli),
-                None => {
-                    // the branch seems to be invalid, so we skip it.
-                    branch_positions.clear();
-                    break;
-                }
-            }
-        }
-        // QueryLevelIterator need to be sorted by level and folded in descending order.
-        branch_positions.sort_unstable_by_key(|qli| qli.level);
-        let folded_query_level_iterators =
-            branch_positions.into_iter().fold(None, |fold: Option<QueryLevelIterator>, mut qli| {
-                match fold {
-                    Some(fold) => {
-                        qli.parent(fold);
-                        Some(qli)
-                    }
-                    None => Some(qli),
-                }
-            });
-
-        if let Some(mut folded_query_level_iterators) = folded_query_level_iterators {
-            let tree_level = folded_query_level_iterators.level;
-            let last_result = folded_query_level_iterators.next(allowed_candidates, tree_level)?;
-            if let Some(last_result) = last_result {
-                let branch = Branch {
-                    last_result,
-                    tree_level,
-                    query_level_iterator: folded_query_level_iterators,
-                    branch_size: branch.len() as u32,
-                };
-                positions.push(branch);
-            }
-        }
-    }
-
-    Ok(positions)
-}
-
-fn set_compute_candidates<'t>(
     ctx: &'t dyn Context<'t>,
     branches: &FlattenedQueryTree,
     allowed_candidates: &RoaringBitmap,
     wdcache: &mut WordDerivationsCache,
-) -> Result<Option<RoaringBitmap>> {
+) -> Result<BinaryHeap<Branch<'t>>> {
-    let mut branches_heap =
+    let mut heap = BinaryHeap::new();
-        initialize_query_level_iterators(ctx, branches, allowed_candidates, wdcache)?;
+    for flatten_branch in branches {
-    let lowest_level = TreeLevel::min_value();
+        let branch = Branch::new(ctx, flatten_branch, wdcache, allowed_candidates)?;
+        heap.push(branch);
+    }
+
+    Ok(heap)
+}
+
+fn set_compute_candidates(
+    branches_heap: &mut BinaryHeap<Branch>,
+    allowed_candidates: &RoaringBitmap,
+) -> Result<Option<(u32, RoaringBitmap)>> {
     let mut final_candidates: Option<(u32, RoaringBitmap)> = None;
     let mut allowed_candidates = allowed_candidates.clone();
 
     while let Some(mut branch) = branches_heap.peek_mut() {
-        let is_lowest_level = branch.tree_level == lowest_level;
-        let branch_rank = branch.compute_rank();
         // if current is worst than best we break to return
         // candidates that correspond to the best rank
+        let branch_rank = branch.compute_rank();
         if let Some((best_rank, _)) = final_candidates {
             if branch_rank > best_rank {
                 break;
             }
         }
-        let _left = branch.last_result.0;
+
-        let candidates = take(&mut branch.last_result.2);
+        let candidates = take(&mut branch.last_result.1);
         if candidates.is_empty() {
             // we don't have candidates, get next interval.
             if !branch.next(&allowed_candidates)? {
                 PeekMut::pop(branch);
             }
-        } else if is_lowest_level {
+        } else {
-            // we have candidates, but we can't dig deeper.
             allowed_candidates -= &candidates;
             final_candidates = match final_candidates.take() {
                 // we add current candidates to best candidates
                 Some((best_rank, mut best_candidates)) => {
                     best_candidates |= candidates;
-                    branch.lazy_next();
+                    branch.next(&allowed_candidates)?;
                     Some((best_rank, best_candidates))
                 }
                 // we take current candidates as best candidates
                 None => {
-                    branch.lazy_next();
+                    branch.next(&allowed_candidates)?;
                     Some((branch_rank, candidates))
                 }
             };
-        } else {
-            // we have candidates, lets dig deeper in levels.
-            branch.dig(ctx)?;
-            if !branch.next(&allowed_candidates)? {
-                PeekMut::pop(branch);
-            }
         }
     }
 
-    Ok(final_candidates.map(|(_rank, candidates)| candidates))
+    Ok(final_candidates)
 }
 
-fn linear_compute_candidates(
+fn initialize_linear_buckets(
     ctx: &dyn Context,
     branches: &FlattenedQueryTree,
     allowed_candidates: &RoaringBitmap,

milli/src/search/criteria/exactness.rs

@@ -10,7 +10,7 @@ use crate::search::criteria::{
     resolve_query_tree, Context, Criterion, CriterionParameters, CriterionResult,
 };
 use crate::search::query_tree::{Operation, PrimitiveQueryPart};
-use crate::{Result, TreeLevel};
+use crate::Result;
 
 pub struct Exactness<'t> {
     ctx: &'t dyn Context<'t>,
@@ -293,7 +293,6 @@ fn attribute_start_with_docids(
     attribute_id: u32,
     query: &[ExactQueryPart],
 ) -> heed::Result<Vec<RoaringBitmap>> {
-    let lowest_level = TreeLevel::min_value();
     let mut attribute_candidates_array = Vec::new();
     // start from attribute first position
     let mut pos = attribute_id * 1000;
@@ -303,7 +302,7 @@ fn attribute_start_with_docids(
             Synonyms(synonyms) => {
                 let mut synonyms_candidates = RoaringBitmap::new();
                 for word in synonyms {
-                    let wc = ctx.word_level_position_docids(word, lowest_level, pos, pos)?;
+                    let wc = ctx.word_position_docids(word, pos)?;
                     if let Some(word_candidates) = wc {
                         synonyms_candidates |= word_candidates;
                     }
@@ -313,7 +312,7 @@ fn attribute_start_with_docids(
             }
             Phrase(phrase) => {
                 for word in phrase {
-                    let wc = ctx.word_level_position_docids(word, lowest_level, pos, pos)?;
+                    let wc = ctx.word_position_docids(word, pos)?;
                     if let Some(word_candidates) = wc {
                         attribute_candidates_array.push(word_candidates);
                     }

milli/src/search/criteria/mod.rs

@@ -14,7 +14,7 @@ use self::words::Words;
 use super::query_tree::{Operation, PrimitiveQueryPart, Query, QueryKind};
 use crate::search::criteria::geo::Geo;
 use crate::search::{word_derivations, WordDerivationsCache};
-use crate::{AscDesc as AscDescName, DocumentId, FieldId, Index, Member, Result, TreeLevel};
+use crate::{AscDesc as AscDescName, DocumentId, FieldId, Index, Member, Result};
 
 mod asc_desc;
 mod attribute;
@@ -90,20 +90,8 @@ pub trait Context<'c> {
     fn word_position_iterator(
         &self,
         word: &str,
-        level: TreeLevel,
         in_prefix_cache: bool,
-        left: Option<u32>,
+    ) -> heed::Result<Box<dyn Iterator<Item = heed::Result<((&'c str, u32), RoaringBitmap)>> + 'c>>;
-        right: Option<u32>,
-    ) -> heed::Result<
-        Box<
-            dyn Iterator<Item = heed::Result<((&'c str, TreeLevel, u32, u32), RoaringBitmap)>> + 'c,
-        >,
-    >;
-    fn word_position_last_level(
-        &self,
-        word: &str,
-        in_prefix_cache: bool,
-    ) -> heed::Result<Option<TreeLevel>>;
     fn synonyms(&self, word: &str) -> heed::Result<Option<Vec<Vec<String>>>>;
     fn searchable_fields_ids(&self) -> Result<Vec<FieldId>>;
     fn field_id_word_count_docids(
@@ -111,13 +99,7 @@ pub trait Context<'c> {
         field_id: FieldId,
         word_count: u8,
     ) -> heed::Result<Option<RoaringBitmap>>;
-    fn word_level_position_docids(
+    fn word_position_docids(&self, word: &str, pos: u32) -> heed::Result<Option<RoaringBitmap>>;
-        &self,
-        word: &str,
-        level: TreeLevel,
-        left: u32,
-        right: u32,
-    ) -> heed::Result<Option<RoaringBitmap>>;
 }
 
 pub struct CriteriaBuilder<'t> {
@@ -183,54 +165,24 @@ impl<'c> Context<'c> for CriteriaBuilder<'c> {
     fn word_position_iterator(
         &self,
         word: &str,
-        level: TreeLevel,
         in_prefix_cache: bool,
-        left: Option<u32>,
+    ) -> heed::Result<Box<dyn Iterator<Item = heed::Result<((&'c str, u32), RoaringBitmap)>> + 'c>>
-        right: Option<u32>,
+    {
-    ) -> heed::Result<
-        Box<
-            dyn Iterator<Item = heed::Result<((&'c str, TreeLevel, u32, u32), RoaringBitmap)>> + 'c,
-        >,
-    > {
         let range = {
-            let left = left.unwrap_or(u32::min_value());
+            let left = u32::min_value();
-            let right = right.unwrap_or(u32::max_value());
+            let right = u32::max_value();
-            let left = (word, level, left, left);
+            let left = (word, left);
-            let right = (word, level, right, right);
+            let right = (word, right);
             left..=right
         };
         let db = match in_prefix_cache {
-            true => self.index.word_prefix_level_position_docids,
+            true => self.index.word_prefix_position_docids,
-            false => self.index.word_level_position_docids,
+            false => self.index.word_position_docids,
         };
 
         Ok(Box::new(db.range(self.rtxn, &range)?))
     }
 
-    fn word_position_last_level(
-        &self,
-        word: &str,
-        in_prefix_cache: bool,
-    ) -> heed::Result<Option<TreeLevel>> {
-        let range = {
-            let left = (word, TreeLevel::min_value(), u32::min_value(), u32::min_value());
-            let right = (word, TreeLevel::max_value(), u32::max_value(), u32::max_value());
-            left..=right
-        };
-        let db = match in_prefix_cache {
-            true => self.index.word_prefix_level_position_docids,
-            false => self.index.word_level_position_docids,
-        };
-        let last_level = db
-            .remap_data_type::<heed::types::DecodeIgnore>()
-            .range(self.rtxn, &range)?
-            .last()
-            .transpose()?
-            .map(|((_, level, _, _), _)| level);
-
-        Ok(last_level)
-    }
-
     fn synonyms(&self, word: &str) -> heed::Result<Option<Vec<Vec<String>>>> {
         self.index.words_synonyms(self.rtxn, &[word])
     }
@@ -251,15 +203,9 @@ impl<'c> Context<'c> for CriteriaBuilder<'c> {
         self.index.field_id_word_count_docids.get(self.rtxn, &key)
     }
 
-    fn word_level_position_docids(
+    fn word_position_docids(&self, word: &str, pos: u32) -> heed::Result<Option<RoaringBitmap>> {
-        &self,
+        let key = (word, pos);
-        word: &str,
+        self.index.word_position_docids.get(self.rtxn, &key)
-        level: TreeLevel,
-        left: u32,
-        right: u32,
-    ) -> heed::Result<Option<RoaringBitmap>> {
-        let key = (word, level, left, right);
-        self.index.word_level_position_docids.get(self.rtxn, &key)
     }
 }
 
@@ -616,27 +562,13 @@ pub mod test {
         fn word_position_iterator(
             &self,
             _word: &str,
-            _level: TreeLevel,
             _in_prefix_cache: bool,
-            _left: Option<u32>,
-            _right: Option<u32>,
         ) -> heed::Result<
-            Box<
+            Box<dyn Iterator<Item = heed::Result<((&'c str, u32), RoaringBitmap)>> + 'c>,
-                dyn Iterator<Item = heed::Result<((&'c str, TreeLevel, u32, u32), RoaringBitmap)>>
-                    + 'c,
-            >,
         > {
             todo!()
         }
 
-        fn word_position_last_level(
-            &self,
-            _word: &str,
-            _in_prefix_cache: bool,
-        ) -> heed::Result<Option<TreeLevel>> {
-            todo!()
-        }
-
         fn synonyms(&self, _word: &str) -> heed::Result<Option<Vec<Vec<String>>>> {
             todo!()
         }
@@ -645,12 +577,10 @@ pub mod test {
             todo!()
         }
 
-        fn word_level_position_docids(
+        fn word_position_docids(
             &self,
             _word: &str,
-            _level: TreeLevel,
+            _pos: u32,
-            _left: u32,
-            _right: u32,
         ) -> heed::Result<Option<RoaringBitmap>> {
             todo!()
         }

milli/src/tree_level.rs

@@ -1,51 +0,0 @@
-use std::convert::TryFrom;
-use std::fmt;
-
-/// This is just before the lowest printable character (space, sp, 32)
-const MAX_VALUE: u8 = 31;
-
-#[derive(Debug, Copy, Clone)]
-pub enum Error {
-    LevelTooHigh(u8),
-}
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
-#[repr(transparent)]
-pub struct TreeLevel(u8);
-
-impl TreeLevel {
-    pub const fn max_value() -> TreeLevel {
-        TreeLevel(MAX_VALUE)
-    }
-
-    pub const fn min_value() -> TreeLevel {
-        TreeLevel(0)
-    }
-
-    pub fn saturating_sub(&self, lhs: u8) -> TreeLevel {
-        TreeLevel(self.0.saturating_sub(lhs))
-    }
-}
-
-impl Into<u8> for TreeLevel {
-    fn into(self) -> u8 {
-        self.0
-    }
-}
-
-impl TryFrom<u8> for TreeLevel {
-    type Error = Error;
-
-    fn try_from(value: u8) -> Result<TreeLevel, Error> {
-        match value {
-            0..=MAX_VALUE => Ok(TreeLevel(value)),
-            _ => Err(Error::LevelTooHigh(value)),
-        }
-    }
-}
-
-impl fmt::Display for TreeLevel {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        write!(f, "{}", self.0)
-    }
-}

milli/src/update/clear_documents.rs

@@ -28,9 +28,9 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
             docid_word_positions,
             word_pair_proximity_docids,
             word_prefix_pair_proximity_docids,
-            word_level_position_docids,
+            word_position_docids,
             field_id_word_count_docids,
-            word_prefix_level_position_docids,
+            word_prefix_position_docids,
             facet_id_f64_docids,
             facet_id_string_docids,
             field_id_docid_facet_f64s,
@@ -64,9 +64,9 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
         docid_word_positions.clear(self.wtxn)?;
         word_pair_proximity_docids.clear(self.wtxn)?;
         word_prefix_pair_proximity_docids.clear(self.wtxn)?;
-        word_level_position_docids.clear(self.wtxn)?;
+        word_position_docids.clear(self.wtxn)?;
         field_id_word_count_docids.clear(self.wtxn)?;
-        word_prefix_level_position_docids.clear(self.wtxn)?;
+        word_prefix_position_docids.clear(self.wtxn)?;
         facet_id_f64_docids.clear(self.wtxn)?;
         facet_id_string_docids.clear(self.wtxn)?;
         field_id_docid_facet_f64s.clear(self.wtxn)?;

milli/src/update/delete_documents.rs

@@ -102,8 +102,8 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
             word_pair_proximity_docids,
             field_id_word_count_docids,
             word_prefix_pair_proximity_docids,
-            word_level_position_docids,
+            word_position_docids,
-            word_prefix_level_position_docids,
+            word_prefix_position_docids,
             facet_id_f64_docids,
             facet_id_string_docids,
             field_id_docid_facet_f64s,
@@ -326,8 +326,7 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
         drop(iter);
 
         // We delete the documents ids that are under the word level position docids.
-        let mut iter =
+        let mut iter = word_position_docids.iter_mut(self.wtxn)?.remap_key_type::<ByteSlice>();
-            word_level_position_docids.iter_mut(self.wtxn)?.remap_key_type::<ByteSlice>();
         while let Some(result) = iter.next() {
             let (bytes, mut docids) = result?;
             let previous_len = docids.len();
@@ -346,7 +345,7 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
 
         // We delete the documents ids that are under the word prefix level position docids.
         let mut iter =
-            word_prefix_level_position_docids.iter_mut(self.wtxn)?.remap_key_type::<ByteSlice>();
+            word_prefix_position_docids.iter_mut(self.wtxn)?.remap_key_type::<ByteSlice>();
         while let Some(result) = iter.next() {
             let (bytes, mut docids) = result?;
             let previous_len = docids.len();

milli/src/update/index_documents/extract/extract_word_position_docids.rs

@@ -14,13 +14,13 @@ use crate::{DocumentId, Result};
 /// Returns a grenad reader with the list of extracted words at positions and
 /// documents ids from the given chunk of docid word positions.
 #[logging_timer::time]
-pub fn extract_word_level_position_docids<R: io::Read>(
+pub fn extract_word_position_docids<R: io::Read>(
     mut docid_word_positions: grenad::Reader<R>,
     indexer: GrenadParameters,
 ) -> Result<grenad::Reader<File>> {
     let max_memory = indexer.max_memory_by_thread();
 
-    let mut word_level_position_docids_sorter = create_sorter(
+    let mut word_position_docids_sorter = create_sorter(
         merge_cbo_roaring_bitmaps,
         indexer.chunk_compression_type,
         indexer.chunk_compression_level,
@@ -37,15 +37,11 @@ pub fn extract_word_level_position_docids<R: io::Read>(
         for position in read_u32_ne_bytes(value) {
             key_buffer.clear();
             key_buffer.extend_from_slice(word_bytes);
-            key_buffer.push(0); // tree level
-
-            // Levels are composed of left and right bounds.
-            key_buffer.extend_from_slice(&position.to_be_bytes());
             key_buffer.extend_from_slice(&position.to_be_bytes());
 
-            word_level_position_docids_sorter.insert(&key_buffer, &document_id.to_ne_bytes())?;
+            word_position_docids_sorter.insert(&key_buffer, &document_id.to_ne_bytes())?;
         }
     }
 
-    sorter_into_reader(word_level_position_docids_sorter, indexer)
+    sorter_into_reader(word_position_docids_sorter, indexer)
 }

milli/src/update/index_documents/extract/mod.rs

@@ -5,8 +5,8 @@ mod extract_fid_docid_facet_values;
 mod extract_fid_word_count_docids;
 mod extract_geo_points;
 mod extract_word_docids;
-mod extract_word_level_position_docids;
 mod extract_word_pair_proximity_docids;
+mod extract_word_position_docids;
 
 use std::collections::HashSet;
 use std::fs::File;
@@ -22,8 +22,8 @@ use self::extract_fid_docid_facet_values::extract_fid_docid_facet_values;
 use self::extract_fid_word_count_docids::extract_fid_word_count_docids;
 use self::extract_geo_points::extract_geo_points;
 use self::extract_word_docids::extract_word_docids;
-use self::extract_word_level_position_docids::extract_word_level_position_docids;
 use self::extract_word_pair_proximity_docids::extract_word_pair_proximity_docids;
+use self::extract_word_position_docids::extract_word_position_docids;
 use super::helpers::{
     into_clonable_grenad, keep_first_prefix_value_merge_roaring_bitmaps, merge_cbo_roaring_bitmaps,
     merge_readers, merge_roaring_bitmaps, CursorClonableMmap, GrenadParameters, MergeFn,
@@ -98,10 +98,10 @@ pub(crate) fn data_from_obkv_documents(
         docid_word_positions_chunks.clone(),
         indexer.clone(),
         lmdb_writer_sx.clone(),
-        extract_word_level_position_docids,
+        extract_word_position_docids,
         merge_cbo_roaring_bitmaps,
-        TypedChunk::WordLevelPositionDocids,
+        TypedChunk::WordPositionDocids,
-        "word-level-position-docids",
+        "word-position-docids",
     );
 
     spawn_extraction_task(

milli/src/update/index_documents/mod.rs

@@ -27,7 +27,7 @@ pub use self::transform::{Transform, TransformOutput};
 use crate::documents::DocumentBatchReader;
 use crate::update::{
     Facets, UpdateBuilder, UpdateIndexingStep, WordPrefixDocids, WordPrefixPairProximityDocids,
-    WordsLevelPositions, WordsPrefixesFst,
+    WordPrefixPositionDocids, WordsPrefixesFst,
 };
 use crate::{Index, Result};
 
@@ -412,8 +412,8 @@ impl<'t, 'u, 'i, 'a> IndexDocuments<'t, 'u, 'i, 'a> {
             total_databases: TOTAL_POSTING_DATABASE_COUNT,
         });
 
-        // Run the words level positions update operation.
+        // Run the words prefix position docids update operation.
-        let mut builder = WordsLevelPositions::new(self.wtxn, self.index);
+        let mut builder = WordPrefixPositionDocids::new(self.wtxn, self.index);
         builder.chunk_compression_type = self.chunk_compression_type;
         builder.chunk_compression_level = self.chunk_compression_level;
         builder.max_nb_chunks = self.max_nb_chunks;

milli/src/update/index_documents/typed_chunk.rs

@@ -22,7 +22,7 @@ pub(crate) enum TypedChunk {
     FieldIdWordcountDocids(grenad::Reader<File>),
     NewDocumentsIds(RoaringBitmap),
     WordDocids(grenad::Reader<File>),
-    WordLevelPositionDocids(grenad::Reader<File>),
+    WordPositionDocids(grenad::Reader<File>),
     WordPairProximityDocids(grenad::Reader<File>),
     FieldIdFacetStringDocids(grenad::Reader<File>),
     FieldIdFacetNumberDocids(grenad::Reader<File>),
@@ -110,10 +110,10 @@ pub(crate) fn write_typed_chunk_into_index(
             index.put_words_fst(wtxn, &fst)?;
             is_merged_database = true;
         }
-        TypedChunk::WordLevelPositionDocids(word_level_position_docids_iter) => {
+        TypedChunk::WordPositionDocids(word_position_docids_iter) => {
             append_entries_into_database(
-                word_level_position_docids_iter,
+                word_position_docids_iter,
-                &index.word_level_position_docids,
+                &index.word_position_docids,
                 wtxn,
                 index_is_empty,
                 |value, _buffer| Ok(value),

milli/src/update/mod.rs

@@ -8,7 +8,7 @@ pub use self::update_builder::UpdateBuilder;
 pub use self::update_step::UpdateIndexingStep;
 pub use self::word_prefix_docids::WordPrefixDocids;
 pub use self::word_prefix_pair_proximity_docids::WordPrefixPairProximityDocids;
-pub use self::words_level_positions::WordsLevelPositions;
+pub use self::words_prefix_position_docids::WordPrefixPositionDocids;
 pub use self::words_prefixes_fst::WordsPrefixesFst;
 
 mod available_documents_ids;
@@ -21,5 +21,5 @@ mod update_builder;
 mod update_step;
 mod word_prefix_docids;
 mod word_prefix_pair_proximity_docids;
-mod words_level_positions;
+mod words_prefix_position_docids;
 mod words_prefixes_fst;

milli/src/update/words_level_positions.rs

@@ -1,268 +0,0 @@
-use std::convert::TryFrom;
-use std::fs::File;
-use std::num::NonZeroU32;
-use std::{cmp, str};
-
-use fst::Streamer;
-use grenad::{CompressionType, Reader, Writer};
-use heed::types::{ByteSlice, DecodeIgnore, Str};
-use heed::{BytesEncode, Error};
-use log::debug;
-use roaring::RoaringBitmap;
-
-use crate::error::{InternalError, SerializationError};
-use crate::heed_codec::{CboRoaringBitmapCodec, StrLevelPositionCodec};
-use crate::index::main_key::WORDS_PREFIXES_FST_KEY;
-use crate::update::index_documents::{
-    create_sorter, create_writer, merge_cbo_roaring_bitmaps, sorter_into_lmdb_database,
-    write_into_lmdb_database, writer_into_reader, WriteMethod,
-};
-use crate::{Index, Result, TreeLevel};
-
-pub struct WordsLevelPositions<'t, 'u, 'i> {
-    wtxn: &'t mut heed::RwTxn<'i, 'u>,
-    index: &'i Index,
-    pub(crate) chunk_compression_type: CompressionType,
-    pub(crate) chunk_compression_level: Option<u32>,
-    pub(crate) max_nb_chunks: Option<usize>,
-    pub(crate) max_memory: Option<usize>,
-    level_group_size: NonZeroU32,
-    min_level_size: NonZeroU32,
-}
-
-impl<'t, 'u, 'i> WordsLevelPositions<'t, 'u, 'i> {
-    pub fn new(
-        wtxn: &'t mut heed::RwTxn<'i, 'u>,
-        index: &'i Index,
-    ) -> WordsLevelPositions<'t, 'u, 'i> {
-        WordsLevelPositions {
-            wtxn,
-            index,
-            chunk_compression_type: CompressionType::None,
-            chunk_compression_level: None,
-            max_nb_chunks: None,
-            max_memory: None,
-            level_group_size: NonZeroU32::new(4).unwrap(),
-            min_level_size: NonZeroU32::new(5).unwrap(),
-        }
-    }
-
-    pub fn level_group_size(&mut self, value: NonZeroU32) -> &mut Self {
-        self.level_group_size = NonZeroU32::new(cmp::max(value.get(), 2)).unwrap();
-        self
-    }
-
-    pub fn min_level_size(&mut self, value: NonZeroU32) -> &mut Self {
-        self.min_level_size = value;
-        self
-    }
-
-    #[logging_timer::time("WordsLevelPositions::{}")]
-    pub fn execute(self) -> Result<()> {
-        debug!("Computing and writing the word levels positions docids into LMDB on disk...");
-
-        let entries = compute_positions_levels(
-            self.wtxn,
-            self.index.word_docids.remap_data_type::<DecodeIgnore>(),
-            self.index.word_level_position_docids,
-            self.chunk_compression_type,
-            self.chunk_compression_level,
-            self.level_group_size,
-            self.min_level_size,
-        )?;
-
-        // The previously computed entries also defines the level 0 entries
-        // so we can clear the database and append all of these entries.
-        self.index.word_level_position_docids.clear(self.wtxn)?;
-
-        write_into_lmdb_database(
-            self.wtxn,
-            *self.index.word_level_position_docids.as_polymorph(),
-            entries,
-            |_, _| Err(InternalError::IndexingMergingKeys { process: "word level position" })?,
-            WriteMethod::Append,
-        )?;
-
-        // We compute the word prefix level positions database.
-        self.index.word_prefix_level_position_docids.clear(self.wtxn)?;
-
-        let mut word_prefix_level_positions_docids_sorter = create_sorter(
-            merge_cbo_roaring_bitmaps,
-            self.chunk_compression_type,
-            self.chunk_compression_level,
-            self.max_nb_chunks,
-            self.max_memory,
-        );
-
-        // We insert the word prefix level positions where the level is equal to 0 and
-        // corresponds to the word-prefix level positions where the prefixes appears
-        // in the prefix FST previously constructed.
-        let prefix_fst = self.index.words_prefixes_fst(self.wtxn)?;
-        let db = self.index.word_level_position_docids.remap_data_type::<ByteSlice>();
-        // iter over all prefixes in the prefix fst.
-        let mut word_stream = prefix_fst.stream();
-        while let Some(prefix_bytes) = word_stream.next() {
-            let prefix = str::from_utf8(prefix_bytes).map_err(|_| {
-                SerializationError::Decoding { db_name: Some(WORDS_PREFIXES_FST_KEY) }
-            })?;
-
-            // iter over all lines of the DB where the key is prefixed by the current prefix.
-            let mut iter = db
-                .remap_key_type::<ByteSlice>()
-                .prefix_iter(self.wtxn, &prefix_bytes)?
-                .remap_key_type::<StrLevelPositionCodec>();
-            while let Some(((_word, level, left, right), data)) = iter.next().transpose()? {
-                // if level is 0, we push the line in the sorter
-                // replacing the complete word by the prefix.
-                if level == TreeLevel::min_value() {
-                    let key = (prefix, level, left, right);
-                    let bytes = StrLevelPositionCodec::bytes_encode(&key).unwrap();
-                    word_prefix_level_positions_docids_sorter.insert(bytes, data)?;
-                }
-            }
-        }
-
-        // We finally write all the word prefix level positions docids with
-        // a level equal to 0 into the LMDB database.
-        sorter_into_lmdb_database(
-            self.wtxn,
-            *self.index.word_prefix_level_position_docids.as_polymorph(),
-            word_prefix_level_positions_docids_sorter,
-            merge_cbo_roaring_bitmaps,
-            WriteMethod::Append,
-        )?;
-
-        let entries = compute_positions_levels(
-            self.wtxn,
-            self.index.word_prefix_docids.remap_data_type::<DecodeIgnore>(),
-            self.index.word_prefix_level_position_docids,
-            self.chunk_compression_type,
-            self.chunk_compression_level,
-            self.level_group_size,
-            self.min_level_size,
-        )?;
-
-        // The previously computed entries also defines the level 0 entries
-        // so we can clear the database and append all of these entries.
-        self.index.word_prefix_level_position_docids.clear(self.wtxn)?;
-
-        write_into_lmdb_database(
-            self.wtxn,
-            *self.index.word_prefix_level_position_docids.as_polymorph(),
-            entries,
-            |_, _| {
-                Err(InternalError::IndexingMergingKeys { process: "word prefix level position" })?
-            },
-            WriteMethod::Append,
-        )?;
-
-        Ok(())
-    }
-}
-
-/// Returns the next number after or equal to `x` that is divisible by `d`.
-fn next_divisible(x: u32, d: u32) -> u32 {
-    (x.saturating_sub(1) | (d - 1)) + 1
-}
-
-/// Returns the previous number after or equal to `x` that is divisible by `d`,
-/// saturates on zero.
-fn previous_divisible(x: u32, d: u32) -> u32 {
-    match x.checked_sub(d - 1) {
-        Some(0) | None => 0,
-        Some(x) => next_divisible(x, d),
-    }
-}
-
-/// Generates all the words positions levels based on the levels zero (including the level zero).
-fn compute_positions_levels(
-    rtxn: &heed::RoTxn,
-    words_db: heed::Database<Str, DecodeIgnore>,
-    words_positions_db: heed::Database<StrLevelPositionCodec, CboRoaringBitmapCodec>,
-    compression_type: CompressionType,
-    compression_level: Option<u32>,
-    level_group_size: NonZeroU32,
-    min_level_size: NonZeroU32,
-) -> Result<Reader<File>> {
-    // It is forbidden to keep a cursor and write in a database at the same time with LMDB
-    // therefore we write the facet levels entries into a grenad file before transfering them.
-    let mut writer = tempfile::tempfile()
-        .and_then(|file| create_writer(compression_type, compression_level, file))?;
-
-    for result in words_db.iter(rtxn)? {
-        let (word, ()) = result?;
-
-        let level_0_range = {
-            let left = (word, TreeLevel::min_value(), u32::min_value(), u32::min_value());
-            let right = (word, TreeLevel::min_value(), u32::max_value(), u32::max_value());
-            left..=right
-        };
-
-        let first_level_size = words_positions_db
-            .remap_data_type::<DecodeIgnore>()
-            .range(rtxn, &level_0_range)?
-            .fold(Ok(0u32), |count, result| result.and(count).map(|c| c + 1))?;
-
-        // Groups sizes are always a power of the original level_group_size and therefore a group
-        // always maps groups of the previous level and never splits previous levels groups in half.
-        let group_size_iter = (1u8..)
-            .map(|l| (TreeLevel::try_from(l).unwrap(), level_group_size.get().pow(l as u32)))
-            .take_while(|(_, s)| first_level_size / *s >= min_level_size.get());
-
-        // As specified in the documentation, we also write the level 0 entries.
-        for result in words_positions_db.range(rtxn, &level_0_range)? {
-            let ((word, level, left, right), docids) = result?;
-            write_level_entry(&mut writer, word, level, left, right, &docids)?;
-        }
-
-        for (level, group_size) in group_size_iter {
-            let mut left = 0;
-            let mut right = 0;
-            let mut group_docids = RoaringBitmap::new();
-
-            for (i, result) in words_positions_db.range(rtxn, &level_0_range)?.enumerate() {
-                let ((_word, _level, value, _right), docids) = result?;
-
-                if i == 0 {
-                    left = previous_divisible(value, group_size);
-                    right = left + (group_size - 1);
-                }
-
-                if value > right {
-                    // we found the first bound of the next group, we must store the left
-                    // and right bounds associated with the docids.
-                    write_level_entry(&mut writer, word, level, left, right, &group_docids)?;
-
-                    // We save the left bound for the new group and also reset the docids.
-                    group_docids = RoaringBitmap::new();
-                    left = previous_divisible(value, group_size);
-                    right = left + (group_size - 1);
-                }
-
-                // The right bound is always the bound we run through.
-                group_docids |= docids;
-            }
-
-            if !group_docids.is_empty() {
-                write_level_entry(&mut writer, word, level, left, right, &group_docids)?;
-            }
-        }
-    }
-
-    writer_into_reader(writer)
-}
-
-fn write_level_entry(
-    writer: &mut Writer<File>,
-    word: &str,
-    level: TreeLevel,
-    left: u32,
-    right: u32,
-    ids: &RoaringBitmap,
-) -> Result<()> {
-    let key = (word, level, left, right);
-    let key = StrLevelPositionCodec::bytes_encode(&key).ok_or(Error::Encoding)?;
-    let data = CboRoaringBitmapCodec::bytes_encode(&ids).ok_or(Error::Encoding)?;
-    writer.insert(&key, &data)?;
-    Ok(())
-}

milli/src/update/words_prefix_position_docids.rs

@@ -0,0 +1,105 @@
+use std::num::NonZeroU32;
+use std::{cmp, str};
+
+use fst::Streamer;
+use grenad::CompressionType;
+use heed::types::ByteSlice;
+use heed::BytesEncode;
+use log::debug;
+
+use crate::error::SerializationError;
+use crate::heed_codec::StrBEU32Codec;
+use crate::index::main_key::WORDS_PREFIXES_FST_KEY;
+use crate::update::index_documents::{
+    create_sorter, merge_cbo_roaring_bitmaps, sorter_into_lmdb_database, WriteMethod,
+};
+use crate::{Index, Result};
+
+pub struct WordPrefixPositionDocids<'t, 'u, 'i> {
+    wtxn: &'t mut heed::RwTxn<'i, 'u>,
+    index: &'i Index,
+    pub(crate) chunk_compression_type: CompressionType,
+    pub(crate) chunk_compression_level: Option<u32>,
+    pub(crate) max_nb_chunks: Option<usize>,
+    pub(crate) max_memory: Option<usize>,
+    level_group_size: NonZeroU32,
+    min_level_size: NonZeroU32,
+}
+
+impl<'t, 'u, 'i> WordPrefixPositionDocids<'t, 'u, 'i> {
+    pub fn new(
+        wtxn: &'t mut heed::RwTxn<'i, 'u>,
+        index: &'i Index,
+    ) -> WordPrefixPositionDocids<'t, 'u, 'i> {
+        WordPrefixPositionDocids {
+            wtxn,
+            index,
+            chunk_compression_type: CompressionType::None,
+            chunk_compression_level: None,
+            max_nb_chunks: None,
+            max_memory: None,
+            level_group_size: NonZeroU32::new(4).unwrap(),
+            min_level_size: NonZeroU32::new(5).unwrap(),
+        }
+    }
+
+    pub fn level_group_size(&mut self, value: NonZeroU32) -> &mut Self {
+        self.level_group_size = NonZeroU32::new(cmp::max(value.get(), 2)).unwrap();
+        self
+    }
+
+    pub fn min_level_size(&mut self, value: NonZeroU32) -> &mut Self {
+        self.min_level_size = value;
+        self
+    }
+
+    #[logging_timer::time("WordPrefixPositionDocids::{}")]
+    pub fn execute(self) -> Result<()> {
+        debug!("Computing and writing the word levels positions docids into LMDB on disk...");
+
+        self.index.word_prefix_position_docids.clear(self.wtxn)?;
+
+        let mut word_prefix_positions_docids_sorter = create_sorter(
+            merge_cbo_roaring_bitmaps,
+            self.chunk_compression_type,
+            self.chunk_compression_level,
+            self.max_nb_chunks,
+            self.max_memory,
+        );
+
+        // We insert the word prefix position and
+        // corresponds to the word-prefix position where the prefixes appears
+        // in the prefix FST previously constructed.
+        let prefix_fst = self.index.words_prefixes_fst(self.wtxn)?;
+        let db = self.index.word_position_docids.remap_data_type::<ByteSlice>();
+        // iter over all prefixes in the prefix fst.
+        let mut word_stream = prefix_fst.stream();
+        while let Some(prefix_bytes) = word_stream.next() {
+            let prefix = str::from_utf8(prefix_bytes).map_err(|_| {
+                SerializationError::Decoding { db_name: Some(WORDS_PREFIXES_FST_KEY) }
+            })?;
+
+            // iter over all lines of the DB where the key is prefixed by the current prefix.
+            let mut iter = db
+                .remap_key_type::<ByteSlice>()
+                .prefix_iter(self.wtxn, &prefix_bytes)?
+                .remap_key_type::<StrBEU32Codec>();
+            while let Some(((_word, pos), data)) = iter.next().transpose()? {
+                let key = (prefix, pos);
+                let bytes = StrBEU32Codec::bytes_encode(&key).unwrap();
+                word_prefix_positions_docids_sorter.insert(bytes, data)?;
+            }
+        }
+
+        // We finally write all the word prefix position docids into the LMDB database.
+        sorter_into_lmdb_database(
+            self.wtxn,
+            *self.index.word_prefix_position_docids.as_polymorph(),
+            word_prefix_positions_docids_sorter,
+            merge_cbo_roaring_bitmaps,
+            WriteMethod::Append,
+        )?;
+
+        Ok(())
+    }
+}