Merge branch 'master' into blockwand

go_to_first_doc was typically calling seek with a target smaller than
doc.

Since SegmentPostings typically do a linear search on the full block,
regardless of the current position, it could have our segment postings
go backward.

CHANGELOG.md

@@ -16,7 +16,7 @@ while doc != TERMINATED {
 ```
 The change made it possible to greatly simplify a lot of the docset's code.
 - Misc internal optimization and introduction of the `Scorer::for_each_pruning` function. (@fulmicoton)
-
+- Added an offset option to the Top(.*)Collectors. (@robyoung)
 
 Tantivy 0.12.0
 ======================

Cargo.toml

@@ -58,6 +58,7 @@ winapi = "0.3"
 rand = "0.7"
 maplit = "1"
 matches = "0.1.8"
+proptest = "0.10"
 
 [dev-dependencies.fail]
 version = "0.4"

examples/iterating_docs_and_positions.rs

@@ -117,11 +117,16 @@ fn main() -> tantivy::Result<()> {
         if let Some(mut block_segment_postings) =
             inverted_index.read_block_postings(&term_the, IndexRecordOption::Basic)
         {
-            while block_segment_postings.advance() {
+            loop {
+                let docs = block_segment_postings.docs();
+                if docs.is_empty() {
+                    break;
+                }
                 // Once again these docs MAY contains deleted documents as well.
                 let docs = block_segment_postings.docs();
                 // Prints `Docs [0, 2].`
                 println!("Docs {:?}", docs);
+                block_segment_postings.advance();
             }
         }
     }

src/collector/custom_score_top_collector.rs

@@ -11,13 +11,13 @@ impl<TCustomScorer, TScore> CustomScoreTopCollector<TCustomScorer, TScore>
 where
     TScore: Clone + PartialOrd,
 {
-    pub fn new(
+    pub(crate) fn new(
         custom_scorer: TCustomScorer,
-        limit: usize,
+        collector: TopCollector<TScore>,
     ) -> CustomScoreTopCollector<TCustomScorer, TScore> {
         CustomScoreTopCollector {
             custom_scorer,
-            collector: TopCollector::with_limit(limit),
+            collector,
         }
     }
 }

src/collector/mod.rs

@@ -109,7 +109,7 @@ pub use self::tweak_score_top_collector::{ScoreSegmentTweaker, ScoreTweaker};
 
 mod facet_collector;
 pub use self::facet_collector::FacetCollector;
-use crate::query::Scorer;
+use crate::query::Weight;
 
 /// `Fruit` is the type for the result of our collection.
 /// e.g. `usize` for the `Count` collector.
@@ -159,21 +159,22 @@ pub trait Collector: Sync {
     /// Created a segment collector and
     fn collect_segment(
         &self,
-        scorer: &mut dyn Scorer,
+        weight: &dyn Weight,
         segment_ord: u32,
-        segment_reader: &SegmentReader,
+        reader: &SegmentReader,
     ) -> crate::Result<<Self::Child as SegmentCollector>::Fruit> {
-        let mut segment_collector = self.for_segment(segment_ord as u32, segment_reader)?;
-        if let Some(delete_bitset) = segment_reader.delete_bitset() {
-            scorer.for_each(&mut |doc, score| {
+        let mut segment_collector = self.for_segment(segment_ord as u32, reader)?;
+
+        if let Some(delete_bitset) = reader.delete_bitset() {
+            weight.for_each(reader, &mut |doc, score| {
                 if delete_bitset.is_alive(doc) {
                     segment_collector.collect(doc, score);
                 }
-            });
+            })?;
         } else {
-            scorer.for_each(&mut |doc, score| {
+            weight.for_each(reader, &mut |doc, score| {
                 segment_collector.collect(doc, score);
-            })
+            })?;
         }
         Ok(segment_collector.harvest())
     }

src/collector/top_collector.rs

@@ -57,6 +57,7 @@ impl<T: PartialOrd, D: PartialOrd> Eq for ComparableDoc<T, D> {}
 
 pub(crate) struct TopCollector<T> {
     pub limit: usize,
+    pub offset: usize,
     _marker: PhantomData<T>,
 }
 
@@ -72,14 +73,20 @@ where
         if limit < 1 {
             panic!("Limit must be strictly greater than 0.");
         }
-        TopCollector {
+        Self {
             limit,
+            offset: 0,
             _marker: PhantomData,
         }
     }
 
-    pub fn limit(&self) -> usize {
-        self.limit
+    /// Skip the first "offset" documents when collecting.
+    ///
+    /// This is equivalent to `OFFSET` in MySQL or PostgreSQL and `start` in
+    /// Lucene's TopDocsCollector.
+    pub fn and_offset(mut self, offset: usize) -> TopCollector<T> {
+        self.offset = offset;
+        self
     }
 
     pub fn merge_fruits(
@@ -92,7 +99,7 @@ where
         let mut top_collector = BinaryHeap::new();
         for child_fruit in children {
             for (feature, doc) in child_fruit {
-                if top_collector.len() < self.limit {
+                if top_collector.len() < (self.limit + self.offset) {
                     top_collector.push(ComparableDoc { feature, doc });
                 } else if let Some(mut head) = top_collector.peek_mut() {
                     if head.feature < feature {
@@ -104,6 +111,7 @@ where
         Ok(top_collector
             .into_sorted_vec()
             .into_iter()
+            .skip(self.offset)
             .map(|cdoc| (cdoc.feature, cdoc.doc))
             .collect())
     }
@@ -113,7 +121,23 @@ where
         segment_id: SegmentLocalId,
         _: &SegmentReader,
     ) -> crate::Result<TopSegmentCollector<F>> {
-        Ok(TopSegmentCollector::new(segment_id, self.limit))
+        Ok(TopSegmentCollector::new(
+            segment_id,
+            self.limit + self.offset,
+        ))
+    }
+
+    /// Create a new TopCollector with the same limit and offset.
+    ///
+    /// Ideally we would use Into but the blanket implementation seems to cause the Scorer traits
+    /// to fail.
+    #[doc(hidden)]
+    pub(crate) fn into_tscore<TScore: PartialOrd + Clone>(self) -> TopCollector<TScore> {
+        TopCollector {
+            limit: self.limit,
+            offset: self.offset,
+            _marker: PhantomData,
+        }
     }
 }
 
@@ -187,7 +211,7 @@ impl<T: PartialOrd + Clone> TopSegmentCollector<T> {
 
 #[cfg(test)]
 mod tests {
-    use super::TopSegmentCollector;
+    use super::{TopCollector, TopSegmentCollector};
     use crate::DocAddress;
 
     #[test]
@@ -248,6 +272,48 @@ mod tests {
             top_collector_limit_3.harvest()[..2].to_vec(),
         );
     }
+
+    #[test]
+    fn test_top_collector_with_limit_and_offset() {
+        let collector = TopCollector::with_limit(2).and_offset(1);
+
+        let results = collector
+            .merge_fruits(vec![vec![
+                (0.9, DocAddress(0, 1)),
+                (0.8, DocAddress(0, 2)),
+                (0.7, DocAddress(0, 3)),
+                (0.6, DocAddress(0, 4)),
+                (0.5, DocAddress(0, 5)),
+            ]])
+            .unwrap();
+
+        assert_eq!(
+            results,
+            vec![(0.8, DocAddress(0, 2)), (0.7, DocAddress(0, 3)),]
+        );
+    }
+
+    #[test]
+    fn test_top_collector_with_limit_larger_than_set_and_offset() {
+        let collector = TopCollector::with_limit(2).and_offset(1);
+
+        let results = collector
+            .merge_fruits(vec![vec![(0.9, DocAddress(0, 1)), (0.8, DocAddress(0, 2))]])
+            .unwrap();
+
+        assert_eq!(results, vec![(0.8, DocAddress(0, 2)),]);
+    }
+
+    #[test]
+    fn test_top_collector_with_limit_and_offset_larger_than_set() {
+        let collector = TopCollector::with_limit(2).and_offset(20);
+
+        let results = collector
+            .merge_fruits(vec![vec![(0.9, DocAddress(0, 1)), (0.8, DocAddress(0, 2))]])
+            .unwrap();
+
+        assert_eq!(results, vec![]);
+    }
 }
 
 #[cfg(all(test, feature = "unstable"))]

src/collector/top_score_collector.rs

@@ -6,9 +6,8 @@ use crate::collector::tweak_score_top_collector::TweakedScoreTopCollector;
 use crate::collector::{
     CustomScorer, CustomSegmentScorer, ScoreSegmentTweaker, ScoreTweaker, SegmentCollector,
 };
-use crate::docset::TERMINATED;
 use crate::fastfield::FastFieldReader;
-use crate::query::Scorer;
+use crate::query::Weight;
 use crate::schema::Field;
 use crate::DocAddress;
 use crate::DocId;
@@ -60,7 +59,11 @@ pub struct TopDocs(TopCollector<Score>);
 
 impl fmt::Debug for TopDocs {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        write!(f, "TopDocs({})", self.0.limit())
+        write!(
+            f,
+            "TopDocs(limit={}, offset={})",
+            self.0.limit, self.0.offset
+        )
     }
 }
 
@@ -104,6 +107,45 @@ impl TopDocs {
         TopDocs(TopCollector::with_limit(limit))
     }
 
+    /// Skip the first "offset" documents when collecting.
+    ///
+    /// This is equivalent to `OFFSET` in MySQL or PostgreSQL and `start` in
+    /// Lucene's TopDocsCollector.
+    ///
+    /// ```rust
+    /// use tantivy::collector::TopDocs;
+    /// use tantivy::query::QueryParser;
+    /// use tantivy::schema::{Schema, TEXT};
+    /// use tantivy::{doc, DocAddress, Index};
+    ///
+    /// let mut schema_builder = Schema::builder();
+    /// let title = schema_builder.add_text_field("title", TEXT);
+    /// let schema = schema_builder.build();
+    /// let index = Index::create_in_ram(schema);
+    ///
+    /// let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+    /// index_writer.add_document(doc!(title => "The Name of the Wind"));
+    /// index_writer.add_document(doc!(title => "The Diary of Muadib"));
+    /// index_writer.add_document(doc!(title => "A Dairy Cow"));
+    /// index_writer.add_document(doc!(title => "The Diary of a Young Girl"));
+    /// index_writer.add_document(doc!(title => "The Diary of Lena Mukhina"));
+    /// assert!(index_writer.commit().is_ok());
+    ///
+    /// let reader = index.reader().unwrap();
+    /// let searcher = reader.searcher();
+    ///
+    /// let query_parser = QueryParser::for_index(&index, vec![title]);
+    /// let query = query_parser.parse_query("diary").unwrap();
+    /// let top_docs = searcher.search(&query, &TopDocs::with_limit(2).and_offset(1)).unwrap();
+    ///
+    /// assert_eq!(top_docs.len(), 2);
+    /// assert_eq!(&top_docs[0], &(0.5204813, DocAddress(0, 4)));
+    /// assert_eq!(&top_docs[1], &(0.4793185, DocAddress(0, 3)));
+    /// ```
+    pub fn and_offset(self, offset: usize) -> TopDocs {
+        TopDocs(self.0.and_offset(offset))
+    }
+
     /// Set top-K to rank documents by a given fast field.
     ///
     /// ```rust
@@ -284,7 +326,7 @@ impl TopDocs {
         TScoreSegmentTweaker: ScoreSegmentTweaker<TScore> + 'static,
         TScoreTweaker: ScoreTweaker<TScore, Child = TScoreSegmentTweaker>,
     {
-        TweakedScoreTopCollector::new(score_tweaker, self.0.limit())
+        TweakedScoreTopCollector::new(score_tweaker, self.0.into_tscore())
     }
 
     /// Ranks the documents using a custom score.
@@ -398,7 +440,7 @@ impl TopDocs {
         TCustomSegmentScorer: CustomSegmentScorer<TScore> + 'static,
         TCustomScorer: CustomScorer<TScore, Child = TCustomSegmentScorer>,
     {
-        CustomScoreTopCollector::new(custom_score, self.0.limit())
+        CustomScoreTopCollector::new(custom_score, self.0.into_tscore())
     }
 }
 
@@ -429,53 +471,59 @@ impl Collector for TopDocs {
 
     fn collect_segment(
         &self,
-        scorer: &mut dyn Scorer,
+        weight: &dyn Weight,
         segment_ord: u32,
-        segment_reader: &SegmentReader,
+        reader: &SegmentReader,
     ) -> crate::Result<<Self::Child as SegmentCollector>::Fruit> {
-        let mut heap: BinaryHeap<ComparableDoc<Score, DocId>> =
-            BinaryHeap::with_capacity(self.0.limit);
-        // first we fill the heap with the first `limit` elements.
-        let mut doc = scorer.doc();
-        while doc != TERMINATED && heap.len() < self.0.limit {
-            if !segment_reader.is_deleted(doc) {
-                let score = scorer.score();
-                heap.push(ComparableDoc {
-                    feature: score,
-                    doc,
-                });
-            }
-            doc = scorer.advance();
-        }
+        let heap_len = self.0.limit + self.0.offset;
+        let mut heap: BinaryHeap<ComparableDoc<Score, DocId>> = BinaryHeap::with_capacity(heap_len);
 
-        let threshold = heap.peek().map(|el| el.feature).unwrap_or(f32::MIN);
-
-        if let Some(delete_bitset) = segment_reader.delete_bitset() {
-            scorer.for_each_pruning(threshold, &mut |doc, score| {
-                if delete_bitset.is_alive(doc) {
-                    *heap.peek_mut().unwrap() = ComparableDoc {
-                        feature: score,
-                        doc,
-                    };
+        if let Some(delete_bitset) = reader.delete_bitset() {
+            let mut threshold = f32::MIN;
+            weight.for_each_pruning(threshold, reader, &mut |doc, score| {
+                if delete_bitset.is_deleted(doc) {
+                    return threshold;
                 }
-                heap.peek().map(|el| el.feature).unwrap_or(f32::MIN)
-            });
-        } else {
-            scorer.for_each_pruning(threshold, &mut |doc, score| {
-                *heap.peek_mut().unwrap() = ComparableDoc {
+                let heap_item = ComparableDoc {
                     feature: score,
                     doc,
                 };
-                heap.peek().map(|el| el.feature).unwrap_or(f32::MIN)
-            });
+                if heap.len() < heap_len {
+                    heap.push(heap_item);
+                    if heap.len() == heap_len {
+                        threshold = heap.peek().map(|el| el.feature).unwrap_or(f32::MIN);
+                    }
+                    return threshold;
+                }
+                *heap.peek_mut().unwrap() = heap_item;
+                threshold = heap.peek().map(|el| el.feature).unwrap_or(std::f32::MIN);
+                threshold
+            })?;
+        } else {
+            weight.for_each_pruning(f32::MIN, reader, &mut |doc, score| {
+                let heap_item = ComparableDoc {
+                    feature: score,
+                    doc,
+                };
+                if heap.len() < heap_len {
+                    heap.push(heap_item);
+                    // TODO the threshold is suboptimal for heap.len == heap_len
+                    if heap.len() == heap_len {
+                        return heap.peek().map(|el| el.feature).unwrap_or(f32::MIN);
+                    } else {
+                        return f32::MIN;
+                    }
+                }
+                *heap.peek_mut().unwrap() = heap_item;
+                heap.peek().map(|el| el.feature).unwrap_or(std::f32::MIN)
+            })?;
         }
 
-        let fruit = heap
+        let fruit: Vec<(Score, DocAddress)> = heap
             .into_sorted_vec()
             .into_iter()
             .map(|cid| (cid.feature, DocAddress(segment_ord, cid.doc)))
             .collect();
-
         Ok(fruit)
     }
 }
@@ -501,10 +549,10 @@ mod tests {
     use crate::collector::Collector;
     use crate::query::{AllQuery, Query, QueryParser};
     use crate::schema::{Field, Schema, FAST, STORED, TEXT};
-    use crate::DocAddress;
     use crate::Index;
     use crate::IndexWriter;
     use crate::Score;
+    use crate::{DocAddress, DocId, SegmentReader};
 
     fn make_index() -> Index {
         let mut schema_builder = Schema::builder();
@@ -544,6 +592,21 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_top_collector_not_at_capacity_with_offset() {
+        let index = make_index();
+        let field = index.schema().get_field("text").unwrap();
+        let query_parser = QueryParser::for_index(&index, vec![field]);
+        let text_query = query_parser.parse_query("droopy tax").unwrap();
+        let score_docs: Vec<(Score, DocAddress)> = index
+            .reader()
+            .unwrap()
+            .searcher()
+            .search(&text_query, &TopDocs::with_limit(4).and_offset(2))
+            .unwrap();
+        assert_eq!(score_docs, vec![(0.48527452, DocAddress(0, 0))]);
+    }
+
     #[test]
     fn test_top_collector_at_capacity() {
         let index = make_index();
@@ -565,6 +628,27 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_top_collector_at_capacity_with_offset() {
+        let index = make_index();
+        let field = index.schema().get_field("text").unwrap();
+        let query_parser = QueryParser::for_index(&index, vec![field]);
+        let text_query = query_parser.parse_query("droopy tax").unwrap();
+        let score_docs: Vec<(Score, DocAddress)> = index
+            .reader()
+            .unwrap()
+            .searcher()
+            .search(&text_query, &TopDocs::with_limit(2).and_offset(1))
+            .unwrap();
+        assert_eq!(
+            score_docs,
+            vec![
+                (0.5376842, DocAddress(0u32, 2)),
+                (0.48527452, DocAddress(0, 0))
+            ]
+        );
+    }
+
     #[test]
     fn test_top_collector_stable_sorting() {
         let index = make_index();
@@ -678,6 +762,50 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_tweak_score_top_collector_with_offset() {
+        let index = make_index();
+        let field = index.schema().get_field("text").unwrap();
+        let query_parser = QueryParser::for_index(&index, vec![field]);
+        let text_query = query_parser.parse_query("droopy tax").unwrap();
+        let collector = TopDocs::with_limit(2).and_offset(1).tweak_score(
+            move |_segment_reader: &SegmentReader| move |doc: DocId, _original_score: Score| doc,
+        );
+        let score_docs: Vec<(u32, DocAddress)> = index
+            .reader()
+            .unwrap()
+            .searcher()
+            .search(&text_query, &collector)
+            .unwrap();
+
+        assert_eq!(
+            score_docs,
+            vec![(1, DocAddress(0, 1)), (0, DocAddress(0, 0)),]
+        );
+    }
+
+    #[test]
+    fn test_custom_score_top_collector_with_offset() {
+        let index = make_index();
+        let field = index.schema().get_field("text").unwrap();
+        let query_parser = QueryParser::for_index(&index, vec![field]);
+        let text_query = query_parser.parse_query("droopy tax").unwrap();
+        let collector = TopDocs::with_limit(2)
+            .and_offset(1)
+            .custom_score(move |_segment_reader: &SegmentReader| move |doc: DocId| doc);
+        let score_docs: Vec<(u32, DocAddress)> = index
+            .reader()
+            .unwrap()
+            .searcher()
+            .search(&text_query, &collector)
+            .unwrap();
+
+        assert_eq!(
+            score_docs,
+            vec![(1, DocAddress(0, 1)), (0, DocAddress(0, 0)),]
+        );
+    }
+
     fn index(
         query: &str,
         query_field: Field,

src/collector/tweak_score_top_collector.rs

@@ -14,11 +14,11 @@ where
 {
     pub fn new(
         score_tweaker: TScoreTweaker,
-        limit: usize,
+        collector: TopCollector<TScore>,
     ) -> TweakedScoreTopCollector<TScoreTweaker, TScore> {
         TweakedScoreTopCollector {
             score_tweaker,
-            collector: TopCollector::with_limit(limit),
+            collector,
         }
     }
 }

src/common/serialize.rs

@@ -89,6 +89,19 @@ impl FixedSize for u64 {
     const SIZE_IN_BYTES: usize = 8;
 }
 
+impl BinarySerializable for f32 {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_f32::<Endianness>(*self)
+    }
+    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
+        reader.read_f32::<Endianness>()
+    }
+}
+
+impl FixedSize for f32 {
+    const SIZE_IN_BYTES: usize = 4;
+}
+
 impl BinarySerializable for i64 {
     fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_i64::<Endianness>(*self)

src/common/vint.rs

@@ -5,7 +5,7 @@ use std::io::Read;
 use std::io::Write;
 
 ///   Wrapper over a `u64` that serializes as a variable int.
-#[derive(Debug, Eq, PartialEq)]
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub struct VInt(pub u64);
 
 const STOP_BIT: u8 = 128;

src/core/index.rs

@@ -24,8 +24,10 @@ use crate::IndexWriter;
 use std::borrow::BorrowMut;
 use std::collections::HashSet;
 use std::fmt;
+
 #[cfg(feature = "mmap")]
-use std::path::{Path, PathBuf};
+use std::path::Path;
+use std::path::PathBuf;
 use std::sync::Arc;
 
 fn load_metas(
@@ -281,7 +283,7 @@ impl Index {
                 TantivyError::LockFailure(
                     err,
                     Some(
-                        "Failed to acquire index lock. If you are using\
+                        "Failed to acquire index lock. If you are using \
                          a regular directory, this means there is already an \
                          `IndexWriter` working on this `Directory`, in this process \
                          or in a different process."

src/core/index_meta.rs

@@ -213,7 +213,7 @@ pub struct IndexMeta {
     #[serde(skip_serializing_if = "Option::is_none")]
     /// Payload associated to the last commit.
     ///
-    /// Upon commit, clients can optionally add a small `Striing` payload to their commit
+    /// Upon commit, clients can optionally add a small `String` payload to their commit
     /// to help identify this commit.
     /// This payload is entirely unused by tantivy.
     pub payload: Option<String>,

src/core/inverted_index_reader.rs

@@ -7,7 +7,6 @@ use crate::schema::FieldType;
 use crate::schema::IndexRecordOption;
 use crate::schema::Term;
 use crate::termdict::TermDictionary;
-use owned_read::OwnedRead;
 
 /// The inverted index reader is in charge of accessing
 /// the inverted index associated to a specific field.
@@ -97,8 +96,7 @@ impl InvertedIndexReader {
         let offset = term_info.postings_offset as usize;
         let end_source = self.postings_source.len();
         let postings_slice = self.postings_source.slice(offset, end_source);
-        let postings_reader = OwnedRead::new(postings_slice);
-        block_postings.reset(term_info.doc_freq, postings_reader);
+        block_postings.reset(term_info.doc_freq, postings_slice);
     }
 
     /// Returns a block postings given a `Term`.
@@ -127,7 +125,7 @@ impl InvertedIndexReader {
         let postings_data = self.postings_source.slice_from(offset);
         BlockSegmentPostings::from_data(
             term_info.doc_freq,
-            OwnedRead::new(postings_data),
+            postings_data,
             self.record_option,
             requested_option,
         )

src/core/searcher.rs

@@ -140,8 +140,7 @@ impl Searcher {
         let segment_readers = self.segment_readers();
         let fruits = executor.map(
             |(segment_ord, segment_reader)| {
-                let mut scorer = weight.scorer(segment_reader, 1.0f32)?;
-                collector.collect_segment(scorer.as_mut(), segment_ord as u32, segment_reader)
+                collector.collect_segment(weight.as_ref(), segment_ord as u32, segment_reader)
             },
             segment_readers.iter().enumerate(),
         )?;

src/core/segment_reader.rs

@@ -8,7 +8,7 @@ use crate::directory::ReadOnlySource;
 use crate::fastfield::DeleteBitSet;
 use crate::fastfield::FacetReader;
 use crate::fastfield::FastFieldReaders;
-use crate::fieldnorm::FieldNormReader;
+use crate::fieldnorm::{FieldNormReader, FieldNormReaders};
 use crate::schema::Field;
 use crate::schema::FieldType;
 use crate::schema::Schema;
@@ -48,7 +48,7 @@ pub struct SegmentReader {
     positions_composite: CompositeFile,
     positions_idx_composite: CompositeFile,
     fast_fields_readers: Arc<FastFieldReaders>,
-    fieldnorms_composite: CompositeFile,
+    fieldnorm_readers: FieldNormReaders,
 
     store_source: ReadOnlySource,
     delete_bitset_opt: Option<DeleteBitSet>,
@@ -126,8 +126,8 @@ impl SegmentReader {
     /// They are simply stored as a fast field, serialized in
     /// the `.fieldnorm` file of the segment.
     pub fn get_fieldnorms_reader(&self, field: Field) -> FieldNormReader {
-        if let Some(fieldnorm_source) = self.fieldnorms_composite.open_read(field) {
-            FieldNormReader::open(fieldnorm_source)
+        if let Some(fieldnorm_source) = self.fieldnorm_readers.get_field(field) {
+            fieldnorm_source
         } else {
             let field_name = self.schema.get_field_name(field);
             let err_msg = format!(
@@ -178,8 +178,8 @@ impl SegmentReader {
         let fast_field_readers =
             Arc::new(FastFieldReaders::load_all(&schema, &fast_fields_composite)?);
 
-        let fieldnorms_data = segment.open_read(SegmentComponent::FIELDNORMS)?;
-        let fieldnorms_composite = CompositeFile::open(&fieldnorms_data)?;
+        let fieldnorm_data = segment.open_read(SegmentComponent::FIELDNORMS)?;
+        let fieldnorm_readers = FieldNormReaders::new(fieldnorm_data)?;
 
         let delete_bitset_opt = if segment.meta().has_deletes() {
             let delete_data = segment.open_read(SegmentComponent::DELETE)?;
@@ -195,7 +195,7 @@ impl SegmentReader {
             termdict_composite,
             postings_composite,
             fast_fields_readers: fast_field_readers,
-            fieldnorms_composite,
+            fieldnorm_readers,
             segment_id: segment.id(),
             store_source,
             delete_bitset_opt,
@@ -295,8 +295,8 @@ impl SegmentReader {
     }
 
     /// Returns an iterator that will iterate over the alive document ids
-    pub fn doc_ids_alive(&self) -> SegmentReaderAliveDocsIterator<'_> {
-        SegmentReaderAliveDocsIterator::new(&self)
+    pub fn doc_ids_alive<'a>(&'a self) -> impl Iterator<Item = DocId> + 'a {
+        (0u32..self.max_doc).filter(move |doc| !self.is_deleted(*doc))
     }
 
     /// Summarize total space usage of this segment.
@@ -308,7 +308,7 @@ impl SegmentReader {
             self.positions_composite.space_usage(),
             self.positions_idx_composite.space_usage(),
             self.fast_fields_readers.space_usage(),
-            self.fieldnorms_composite.space_usage(),
+            self.fieldnorm_readers.space_usage(),
             self.get_store_reader().space_usage(),
             self.delete_bitset_opt
                 .as_ref()
@@ -324,52 +324,6 @@ impl fmt::Debug for SegmentReader {
     }
 }
 
-/// Implements the iterator trait to allow easy iteration
-/// over non-deleted ("alive") DocIds in a SegmentReader
-pub struct SegmentReaderAliveDocsIterator<'a> {
-    reader: &'a SegmentReader,
-    max_doc: DocId,
-    current: DocId,
-}
-
-impl<'a> SegmentReaderAliveDocsIterator<'a> {
-    pub fn new(reader: &'a SegmentReader) -> SegmentReaderAliveDocsIterator<'a> {
-        SegmentReaderAliveDocsIterator {
-            reader,
-            max_doc: reader.max_doc(),
-            current: 0,
-        }
-    }
-}
-
-impl<'a> Iterator for SegmentReaderAliveDocsIterator<'a> {
-    type Item = DocId;
-
-    fn next(&mut self) -> Option<Self::Item> {
-        // TODO: Use TinySet (like in BitSetDocSet) to speed this process up
-        if self.current >= self.max_doc {
-            return None;
-        }
-
-        // find the next alive doc id
-        while self.reader.is_deleted(self.current) {
-            self.current += 1;
-
-            if self.current >= self.max_doc {
-                return None;
-            }
-        }
-
-        // capture the current alive DocId
-        let result = Some(self.current);
-
-        // move down the chain
-        self.current += 1;
-
-        result
-    }
-}
-
 #[cfg(test)]
 mod test {
     use crate::core::Index;

src/directory/footer.rs

@@ -94,12 +94,24 @@ impl Footer {
         match &self.versioned_footer {
             VersionedFooter::V1 {
                 crc32: _crc,
-                store_compression: compression,
+                store_compression,
             } => {
-                if &library_version.store_compression != compression {
+                if &library_version.store_compression != store_compression {
                     return Err(Incompatibility::CompressionMismatch {
                         library_compression_format: library_version.store_compression.to_string(),
-                        index_compression_format: compression.to_string(),
+                        index_compression_format: store_compression.to_string(),
+                    });
+                }
+                Ok(())
+            }
+            VersionedFooter::V2 {
+                crc32: _crc,
+                store_compression,
+            } => {
+                if &library_version.store_compression != store_compression {
+                    return Err(Incompatibility::CompressionMismatch {
+                        library_compression_format: library_version.store_compression.to_string(),
+                        index_compression_format: store_compression.to_string(),
                     });
                 }
                 Ok(())
@@ -120,24 +132,29 @@ pub enum VersionedFooter {
         crc32: CrcHashU32,
         store_compression: String,
     },
+    // Introduction of the Block WAND information.
+    V2 {
+        crc32: CrcHashU32,
+        store_compression: String,
+    },
 }
 
 impl BinarySerializable for VersionedFooter {
     fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         let mut buf = Vec::new();
         match self {
-            VersionedFooter::V1 {
+            VersionedFooter::V2 {
                 crc32,
                 store_compression: compression,
             } => {
                 // Serializes a valid `VersionedFooter` or panics if the version is unknown
                 // [   version    |   crc_hash  | compression_mode ]
                 // [    0..4      |     4..8    |     variable     ]
-                BinarySerializable::serialize(&1u32, &mut buf)?;
+                BinarySerializable::serialize(&2u32, &mut buf)?;
                 BinarySerializable::serialize(crc32, &mut buf)?;
                 BinarySerializable::serialize(compression, &mut buf)?;
             }
-            VersionedFooter::UnknownVersion => {
+            VersionedFooter::V1 { .. } | VersionedFooter::UnknownVersion => {
                 return Err(io::Error::new(
                     io::ErrorKind::InvalidInput,
                     "Cannot serialize an unknown versioned footer ",
@@ -168,10 +185,17 @@ impl BinarySerializable for VersionedFooter {
         let version = u32::deserialize(&mut cursor)?;
         if version == 1 {
             let crc32 = u32::deserialize(&mut cursor)?;
-            let compression = String::deserialize(&mut cursor)?;
+            let store_compression = String::deserialize(&mut cursor)?;
             Ok(VersionedFooter::V1 {
                 crc32,
-                store_compression: compression,
+                store_compression,
+            })
+        } else if version == 2 {
+            let crc32 = u32::deserialize(&mut cursor)?;
+            let store_compression = String::deserialize(&mut cursor)?;
+            Ok(VersionedFooter::V2 {
+                crc32,
+                store_compression,
             })
         } else {
             Ok(VersionedFooter::UnknownVersion)
@@ -182,6 +206,7 @@ impl BinarySerializable for VersionedFooter {
 impl VersionedFooter {
     pub fn crc(&self) -> Option<CrcHashU32> {
         match self {
+            VersionedFooter::V2 { crc32, .. } => Some(*crc32),
             VersionedFooter::V1 { crc32, .. } => Some(*crc32),
             VersionedFooter::UnknownVersion { .. } => None,
         }
@@ -219,7 +244,7 @@ impl<W: TerminatingWrite> Write for FooterProxy<W> {
 impl<W: TerminatingWrite> TerminatingWrite for FooterProxy<W> {
     fn terminate_ref(&mut self, _: AntiCallToken) -> io::Result<()> {
         let crc32 = self.hasher.take().unwrap().finalize();
-        let footer = Footer::new(VersionedFooter::V1 {
+        let footer = Footer::new(VersionedFooter::V2 {
             crc32,
             store_compression: crate::store::COMPRESSION.to_string(),
         });
@@ -248,15 +273,11 @@ mod tests {
         assert!(footer_proxy.terminate().is_ok());
         assert_eq!(vec.len(), 167);
         let footer = Footer::deserialize(&mut &vec[..]).unwrap();
-        if let VersionedFooter::V1 {
-            crc32: _,
-            store_compression,
-        } = footer.versioned_footer
-        {
-            assert_eq!(store_compression, crate::store::COMPRESSION);
-        } else {
-            panic!("Versioned footer should be V1.");
-        }
+        assert!(matches!(
+           footer.versioned_footer,
+           VersionedFooter::V2 { store_compression, .. }
+           if store_compression == crate::store::COMPRESSION
+        ));
         assert_eq!(&footer.version, crate::version());
     }
 
@@ -264,7 +285,7 @@ mod tests {
     fn test_serialize_deserialize_footer() {
         let mut buffer = Vec::new();
         let crc32 = 123456u32;
-        let footer: Footer = Footer::new(VersionedFooter::V1 {
+        let footer: Footer = Footer::new(VersionedFooter::V2 {
             crc32,
             store_compression: "lz4".to_string(),
         });
@@ -276,7 +297,7 @@ mod tests {
     #[test]
     fn footer_length() {
         let crc32 = 1111111u32;
-        let versioned_footer = VersionedFooter::V1 {
+        let versioned_footer = VersionedFooter::V2 {
             crc32,
             store_compression: "lz4".to_string(),
         };
@@ -297,7 +318,7 @@ mod tests {
             // versionned footer length
             12 | 128,
             // index format version
-            1,
+            2,
             0,
             0,
             0,
@@ -316,7 +337,7 @@ mod tests {
         let versioned_footer = VersionedFooter::deserialize(&mut cursor).unwrap();
         assert!(cursor.is_empty());
         let expected_crc: u32 = LittleEndian::read_u32(&v_footer_bytes[5..9]) as CrcHashU32;
-        let expected_versioned_footer: VersionedFooter = VersionedFooter::V1 {
+        let expected_versioned_footer: VersionedFooter = VersionedFooter::V2 {
             crc32: expected_crc,
             store_compression: "lz4".to_string(),
         };

src/docset.rs

@@ -7,7 +7,7 @@ use std::borrow::BorrowMut;
 ///
 /// This is not u32::MAX as one would have expected, due to the lack of SSE2 instructions
 /// to compare [u32; 4].
-pub const TERMINATED: DocId = i32::MAX as u32;
+pub const TERMINATED: DocId = std::i32::MAX as u32;
 
 /// Represents an iterable set of sorted doc ids.
 pub trait DocSet {
@@ -38,6 +38,7 @@ pub trait DocSet {
     /// Calling `seek(TERMINATED)` is also legal and is the normal way to consume a DocSet.
     fn seek(&mut self, target: DocId) -> DocId {
         let mut doc = self.doc();
+        debug_assert!(doc <= target);
         while doc < target {
             doc = self.advance();
         }

src/fieldnorm/mod.rs

@@ -21,7 +21,7 @@ mod reader;
 mod serializer;
 mod writer;
 
-pub use self::reader::FieldNormReader;
+pub use self::reader::{FieldNormReader, FieldNormReaders};
 pub use self::serializer::FieldNormsSerializer;
 pub use self::writer::FieldNormsWriter;
 

src/fieldnorm/reader.rs

@@ -1,6 +1,41 @@
 use super::{fieldnorm_to_id, id_to_fieldnorm};
+use crate::common::CompositeFile;
 use crate::directory::ReadOnlySource;
+use crate::schema::Field;
+use crate::space_usage::PerFieldSpaceUsage;
 use crate::DocId;
+use std::sync::Arc;
+
+/// Reader for the fieldnorm (for each document, the number of tokens indexed in the
+/// field) of all indexed fields in the index.
+///
+/// Each fieldnorm is approximately compressed over one byte. We refer to this byte as
+/// `fieldnorm_id`.
+/// The mapping from `fieldnorm` to `fieldnorm_id` is given by monotonic.
+#[derive(Clone)]
+pub struct FieldNormReaders {
+    data: Arc<CompositeFile>,
+}
+
+impl FieldNormReaders {
+    /// Creates a field norm reader.
+    pub fn new(source: ReadOnlySource) -> crate::Result<FieldNormReaders> {
+        let data = CompositeFile::open(&source)?;
+        Ok(FieldNormReaders {
+            data: Arc::new(data),
+        })
+    }
+
+    /// Returns the FieldNormReader for a specific field.
+    pub fn get_field(&self, field: Field) -> Option<FieldNormReader> {
+        self.data.open_read(field).map(FieldNormReader::open)
+    }
+
+    /// Return a break down of the space usage per field.
+    pub fn space_usage(&self) -> PerFieldSpaceUsage {
+        self.data.space_usage()
+    }
+}
 
 /// Reads the fieldnorm associated to a document.
 /// The fieldnorm represents the length associated to
@@ -19,6 +54,7 @@ use crate::DocId;
 /// Apart from compression, this scale also makes it possible to
 /// precompute computationally expensive functions of the fieldnorm
 /// in a very short array.
+#[derive(Clone)]
 pub struct FieldNormReader {
     data: ReadOnlySource,
 }
@@ -29,6 +65,11 @@ impl FieldNormReader {
         FieldNormReader { data }
     }
 
+    /// Returns the number of documents in this segment.
+    pub fn num_docs(&self) -> u32 {
+        self.data.len() as u32
+    }
+
     /// Returns the `fieldnorm` associated to a doc id.
     /// The fieldnorm is a value approximating the number
     /// of tokens in a given field of the `doc_id`.
@@ -65,10 +106,11 @@ impl FieldNormReader {
 }
 
 #[cfg(test)]
-impl From<Vec<u32>> for FieldNormReader {
-    fn from(field_norms: Vec<u32>) -> FieldNormReader {
+impl From<&[u32]> for FieldNormReader {
+    fn from(field_norms: &[u32]) -> FieldNormReader {
         let field_norms_id = field_norms
-            .into_iter()
+            .iter()
+            .cloned()
             .map(FieldNormReader::fieldnorm_to_id)
             .collect::<Vec<u8>>();
         let field_norms_data = ReadOnlySource::from(field_norms_id);

src/fieldnorm/writer.rs

@@ -78,11 +78,12 @@ impl FieldNormsWriter {
     }
 
     /// Serialize the seen fieldnorm values to the serializer for all fields.
-    pub fn serialize(&self, fieldnorms_serializer: &mut FieldNormsSerializer) -> io::Result<()> {
+    pub fn serialize(&self, mut fieldnorms_serializer: FieldNormsSerializer) -> io::Result<()> {
         for &field in self.fields.iter() {
             let fieldnorm_values: &[u8] = &self.fieldnorms_buffer[field.field_id() as usize][..];
             fieldnorms_serializer.serialize_field(field, fieldnorm_values)?;
         }
+        fieldnorms_serializer.close()?;
         Ok(())
     }
 }

src/indexer/index_writer.rs

@@ -346,7 +346,7 @@ impl IndexWriter {
 
     fn drop_sender(&mut self) {
         let (sender, _receiver) = channel::bounded(1);
-        mem::replace(&mut self.operation_sender, sender);
+        self.operation_sender = sender;
     }
 
     /// If there are some merging threads, blocks until they all finish their work and

src/indexer/log_merge_policy.rs

@@ -54,10 +54,6 @@ impl LogMergePolicy {
 
 impl MergePolicy for LogMergePolicy {
     fn compute_merge_candidates(&self, segments: &[SegmentMeta]) -> Vec<MergeCandidate> {
-        if segments.is_empty() {
-            return Vec::new();
-        }
-
         let mut size_sorted_tuples = segments
             .iter()
             .map(SegmentMeta::num_docs)
@@ -67,27 +63,35 @@ impl MergePolicy for LogMergePolicy {
 
         size_sorted_tuples.sort_by(|x, y| y.1.cmp(&(x.1)));
 
+        if size_sorted_tuples.len() <= 1 {
+            return Vec::new();
+        }
+
         let size_sorted_log_tuples: Vec<_> = size_sorted_tuples
             .into_iter()
             .map(|(ind, num_docs)| (ind, f64::from(self.clip_min_size(num_docs)).log2()))
             .collect();
 
-        let (first_ind, first_score) = size_sorted_log_tuples[0];
-        let mut current_max_log_size = first_score;
-        let mut levels = vec![vec![first_ind]];
-        for &(ind, score) in (&size_sorted_log_tuples).iter().skip(1) {
-            if score < (current_max_log_size - self.level_log_size) {
-                current_max_log_size = score;
-                levels.push(Vec::new());
+        if let Some(&(first_ind, first_score)) = size_sorted_log_tuples.first() {
+            let mut current_max_log_size = first_score;
+            let mut levels = vec![vec![first_ind]];
+            for &(ind, score) in (&size_sorted_log_tuples).iter().skip(1) {
+                if score < (current_max_log_size - self.level_log_size) {
+                    current_max_log_size = score;
+                    levels.push(Vec::new());
+                }
+                levels.last_mut().unwrap().push(ind);
             }
-            levels.last_mut().unwrap().push(ind);
+            levels
+                .iter()
+                .filter(|level| level.len() >= self.min_merge_size)
+                .map(|ind_vec| {
+                    MergeCandidate(ind_vec.iter().map(|&ind| segments[ind].id()).collect())
+                })
+                .collect()
+        } else {
+            return vec![];
         }
-
-        levels
-            .iter()
-            .filter(|level| level.len() >= self.min_merge_size)
-            .map(|ind_vec| MergeCandidate(ind_vec.iter().map(|&ind| segments[ind].id()).collect()))
-            .collect()
     }
 }
 
@@ -179,6 +183,7 @@ mod tests {
         let result_list = test_merge_policy().compute_merge_candidates(&test_input);
         assert_eq!(result_list.len(), 2);
     }
+
     #[test]
     fn test_log_merge_policy_small_segments() {
         // segments under min_layer_size are merged together
@@ -194,6 +199,17 @@ mod tests {
         assert_eq!(result_list.len(), 1);
     }
 
+    #[test]
+    fn test_log_merge_policy_all_segments_too_large_to_merge() {
+        let eight_large_segments: Vec<SegmentMeta> =
+            std::iter::repeat_with(|| create_random_segment_meta(100_001))
+                .take(8)
+                .collect();
+        assert!(test_merge_policy()
+            .compute_merge_candidates(&eight_large_segments)
+            .is_empty());
+    }
+
     #[test]
     fn test_large_merge_segments() {
         let test_input = vec![

src/indexer/merger.rs

@@ -8,9 +8,9 @@ use crate::fastfield::DeleteBitSet;
 use crate::fastfield::FastFieldReader;
 use crate::fastfield::FastFieldSerializer;
 use crate::fastfield::MultiValueIntFastFieldReader;
-use crate::fieldnorm::FieldNormReader;
 use crate::fieldnorm::FieldNormsSerializer;
 use crate::fieldnorm::FieldNormsWriter;
+use crate::fieldnorm::{FieldNormReader, FieldNormReaders};
 use crate::indexer::SegmentSerializer;
 use crate::postings::InvertedIndexSerializer;
 use crate::postings::Postings;
@@ -20,7 +20,7 @@ use crate::schema::{Field, Schema};
 use crate::store::StoreWriter;
 use crate::termdict::TermMerger;
 use crate::termdict::TermOrdinal;
-use crate::DocId;
+use crate::{DocId, SegmentComponent};
 use std::cmp;
 use std::collections::HashMap;
 
@@ -167,7 +167,7 @@ impl IndexMerger {
 
     fn write_fieldnorms(
         &self,
-        fieldnorms_serializer: &mut FieldNormsSerializer,
+        mut fieldnorms_serializer: FieldNormsSerializer,
     ) -> crate::Result<()> {
         let fields = FieldNormsWriter::fields_with_fieldnorm(&self.schema);
         let mut fieldnorms_data = Vec::with_capacity(self.max_doc as usize);
@@ -182,6 +182,7 @@ impl IndexMerger {
             }
             fieldnorms_serializer.serialize_field(field, &fieldnorms_data[..])?;
         }
+        fieldnorms_serializer.close()?;
         Ok(())
     }
 
@@ -492,6 +493,7 @@ impl IndexMerger {
         indexed_field: Field,
         field_type: &FieldType,
         serializer: &mut InvertedIndexSerializer,
+        fieldnorm_reader: Option<FieldNormReader>,
     ) -> crate::Result<Option<TermOrdinalMapping>> {
         let mut positions_buffer: Vec<u32> = Vec::with_capacity(1_000);
         let mut delta_computer = DeltaComputer::new();
@@ -550,7 +552,8 @@ impl IndexMerger {
         // - Segment 2's doc ids become  [seg0.max_doc + seg1.max_doc,
         //                                seg0.max_doc + seg1.max_doc + seg2.max_doc]
         // ...
-        let mut field_serializer = serializer.new_field(indexed_field, total_num_tokens)?;
+        let mut field_serializer =
+            serializer.new_field(indexed_field, total_num_tokens, fieldnorm_reader)?;
 
         let field_entry = self.schema.get_field_entry(indexed_field);
 
@@ -589,48 +592,49 @@ impl IndexMerger {
             // of all of the segments containing the given term.
             //
             // These segments are non-empty and advance has already been called.
-            if !segment_postings.is_empty() {
-                // If not, the `term` will be entirely removed.
-
-                // We know that there is at least one document containing
-                // the term, so we add it.
-                let to_term_ord = field_serializer.new_term(term_bytes)?;
-
-                if let Some(ref mut term_ord_mapping) = term_ord_mapping_opt {
-                    for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
-                        term_ord_mapping.register_from_to(segment_ord, from_term_ord, to_term_ord);
-                    }
-                }
-
-                // We can now serialize this postings, by pushing each document to the
-                // postings serializer.
-                for (segment_ord, mut segment_postings) in segment_postings {
-                    let old_to_new_doc_id = &merged_doc_id_map[segment_ord];
-
-                    let mut doc = segment_postings.doc();
-                    while doc != TERMINATED {
-                        // deleted doc are skipped as they do not have a `remapped_doc_id`.
-                        if let Some(remapped_doc_id) = old_to_new_doc_id[doc as usize] {
-                            // we make sure to only write the term iff
-                            // there is at least one document.
-                            let term_freq = segment_postings.term_freq();
-                            segment_postings.positions(&mut positions_buffer);
-
-                            let delta_positions = delta_computer.compute_delta(&positions_buffer);
-                            field_serializer.write_doc(
-                                remapped_doc_id,
-                                term_freq,
-                                delta_positions,
-                            )?;
-                        }
-
-                        doc = segment_postings.advance();
-                    }
-                }
-
-                // closing the term.
-                field_serializer.close_term()?;
+            if segment_postings.is_empty() {
+                continue;
             }
+            // If not, the `term` will be entirely removed.
+
+            // We know that there is at least one document containing
+            // the term, so we add it.
+            let term_doc_freq = segment_postings
+                .iter()
+                .map(|(_, segment_posting)| segment_posting.doc_freq())
+                .sum();
+            let to_term_ord = field_serializer.new_term(term_bytes, term_doc_freq)?;
+
+            if let Some(ref mut term_ord_mapping) = term_ord_mapping_opt {
+                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
+                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, to_term_ord);
+                }
+            }
+
+            // We can now serialize this postings, by pushing each document to the
+            // postings serializer.
+            for (segment_ord, mut segment_postings) in segment_postings {
+                let old_to_new_doc_id = &merged_doc_id_map[segment_ord];
+
+                let mut doc = segment_postings.doc();
+                while doc != TERMINATED {
+                    // deleted doc are skipped as they do not have a `remapped_doc_id`.
+                    if let Some(remapped_doc_id) = old_to_new_doc_id[doc as usize] {
+                        // we make sure to only write the term iff
+                        // there is at least one document.
+                        let term_freq = segment_postings.term_freq();
+                        segment_postings.positions(&mut positions_buffer);
+
+                        let delta_positions = delta_computer.compute_delta(&positions_buffer);
+                        field_serializer.write_doc(remapped_doc_id, term_freq, delta_positions)?;
+                    }
+
+                    doc = segment_postings.advance();
+                }
+            }
+
+            // closing the term.
+            field_serializer.close_term()?;
         }
         field_serializer.close()?;
         Ok(term_ord_mapping_opt)
@@ -639,13 +643,18 @@ impl IndexMerger {
     fn write_postings(
         &self,
         serializer: &mut InvertedIndexSerializer,
+        fieldnorm_readers: FieldNormReaders,
     ) -> crate::Result<HashMap<Field, TermOrdinalMapping>> {
         let mut term_ordinal_mappings = HashMap::new();
         for (field, field_entry) in self.schema.fields() {
+            let fieldnorm_reader = fieldnorm_readers.get_field(field);
             if field_entry.is_indexed() {
-                if let Some(term_ordinal_mapping) =
-                    self.write_postings_for_field(field, field_entry.field_type(), serializer)?
-                {
+                if let Some(term_ordinal_mapping) = self.write_postings_for_field(
+                    field,
+                    field_entry.field_type(),
+                    serializer,
+                    fieldnorm_reader,
+                )? {
                     term_ordinal_mappings.insert(field, term_ordinal_mapping);
                 }
             }
@@ -671,8 +680,15 @@ impl IndexMerger {
 
 impl SerializableSegment for IndexMerger {
     fn write(&self, mut serializer: SegmentSerializer) -> crate::Result<u32> {
-        let term_ord_mappings = self.write_postings(serializer.get_postings_serializer())?;
-        self.write_fieldnorms(serializer.get_fieldnorms_serializer())?;
+        if let Some(fieldnorms_serializer) = serializer.extract_fieldnorms_serializer() {
+            self.write_fieldnorms(fieldnorms_serializer)?;
+        }
+        let fieldnorm_data = serializer
+            .segment()
+            .open_read(SegmentComponent::FIELDNORMS)?;
+        let fieldnorm_readers = FieldNormReaders::new(fieldnorm_data)?;
+        let term_ord_mappings =
+            self.write_postings(serializer.get_postings_serializer(), fieldnorm_readers)?;
         self.write_fast_fields(serializer.get_fast_field_serializer(), term_ord_mappings)?;
         self.write_storable_fields(serializer.get_store_writer())?;
         serializer.close()?;
@@ -682,15 +698,15 @@ impl SerializableSegment for IndexMerger {
 
 #[cfg(test)]
 mod tests {
+    use crate::assert_nearly_equals;
     use crate::collector::tests::TEST_COLLECTOR_WITH_SCORE;
     use crate::collector::tests::{BytesFastFieldTestCollector, FastFieldTestCollector};
     use crate::collector::{Count, FacetCollector};
     use crate::core::Index;
     use crate::query::AllQuery;
     use crate::query::BooleanQuery;
+    use crate::query::Scorer;
     use crate::query::TermQuery;
-    use crate::schema;
-    use crate::schema::Cardinality;
     use crate::schema::Document;
     use crate::schema::Facet;
     use crate::schema::IndexRecordOption;
@@ -698,9 +714,11 @@ mod tests {
     use crate::schema::Term;
     use crate::schema::TextFieldIndexing;
     use crate::schema::INDEXED;
+    use crate::schema::{Cardinality, TEXT};
     use crate::DocAddress;
     use crate::IndexWriter;
     use crate::Searcher;
+    use crate::{schema, DocSet, SegmentId};
     use byteorder::{BigEndian, ReadBytesExt, WriteBytesExt};
     use futures::executor::block_on;
     use std::io::Cursor;
@@ -1507,12 +1525,9 @@ mod tests {
         for i in 0..100 {
             let mut doc = Document::new();
             doc.add_f64(field, 42.0);
-
             doc.add_f64(multi_field, 0.24);
             doc.add_f64(multi_field, 0.27);
-
             writer.add_document(doc);
-
             if i % 5 == 0 {
                 writer.commit()?;
             }
@@ -1524,6 +1539,72 @@ mod tests {
         // If a merging thread fails, we should end up with more
         // than one segment here
         assert_eq!(1, index.searchable_segments()?.len());
+        Ok(())
+    }
+
+    #[test]
+    fn test_merged_index_has_blockwand() -> crate::Result<()> {
+        let mut builder = schema::SchemaBuilder::new();
+        let text = builder.add_text_field("text", TEXT);
+        let index = Index::create_in_ram(builder.build());
+        let mut writer = index.writer_with_num_threads(1, 3_000_000)?;
+        let happy_term = Term::from_field_text(text, "happy");
+        let term_query = TermQuery::new(happy_term, IndexRecordOption::WithFreqs);
+        for _ in 0..62 {
+            writer.add_document(doc!(text=>"hello happy tax payer"));
+        }
+        writer.commit()?;
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+        let mut term_scorer = term_query
+            .specialized_weight(&searcher, true)
+            .specialized_scorer(searcher.segment_reader(0u32), 1.0f32)?;
+        assert_eq!(term_scorer.doc(), 0);
+        assert_nearly_equals!(term_scorer.block_max_score(), 0.0079681855);
+        assert_nearly_equals!(term_scorer.score(), 0.0079681855);
+        for _ in 0..81 {
+            writer.add_document(doc!(text=>"hello happy tax payer"));
+        }
+        writer.commit()?;
+        reader.reload()?;
+        let searcher = reader.searcher();
+
+        assert_eq!(searcher.segment_readers().len(), 2);
+        for segment_reader in searcher.segment_readers() {
+            let mut term_scorer = term_query
+                .specialized_weight(&searcher, true)
+                .specialized_scorer(segment_reader, 1.0f32)?;
+            // the difference compared to before is instrinsic to the bm25 formula. no worries there.
+            for doc in segment_reader.doc_ids_alive() {
+                assert_eq!(term_scorer.doc(), doc);
+                assert_nearly_equals!(term_scorer.block_max_score(), 0.003478312);
+                assert_nearly_equals!(term_scorer.score(), 0.003478312);
+                term_scorer.advance();
+            }
+        }
+
+        let segment_ids: Vec<SegmentId> = searcher
+            .segment_readers()
+            .iter()
+            .map(|reader| reader.segment_id())
+            .collect();
+        block_on(writer.merge(&segment_ids[..]))?;
+
+        reader.reload()?;
+        let searcher = reader.searcher();
+        assert_eq!(searcher.segment_readers().len(), 1);
+
+        let segment_reader = searcher.segment_reader(0u32);
+        let mut term_scorer = term_query
+            .specialized_weight(&searcher, true)
+            .specialized_scorer(segment_reader, 1.0f32)?;
+        // the difference compared to before is instrinsic to the bm25 formula. no worries there.
+        for doc in segment_reader.doc_ids_alive() {
+            assert_eq!(term_scorer.doc(), doc);
+            assert_nearly_equals!(term_scorer.block_max_score(), 0.003478312);
+            assert_nearly_equals!(term_scorer.score(), 0.003478312);
+            term_scorer.advance();
+        }
 
         Ok(())
     }

src/indexer/segment_serializer.rs

@@ -8,15 +8,16 @@ use crate::store::StoreWriter;
 /// Segment serializer is in charge of laying out on disk
 /// the data accumulated and sorted by the `SegmentWriter`.
 pub struct SegmentSerializer {
+    segment: Segment,
     store_writer: StoreWriter,
     fast_field_serializer: FastFieldSerializer,
-    fieldnorms_serializer: FieldNormsSerializer,
+    fieldnorms_serializer: Option<FieldNormsSerializer>,
     postings_serializer: InvertedIndexSerializer,
 }
 
 impl SegmentSerializer {
     /// Creates a new `SegmentSerializer`.
-    pub fn for_segment(segment: &mut Segment) -> crate::Result<SegmentSerializer> {
+    pub fn for_segment(mut segment: Segment) -> crate::Result<SegmentSerializer> {
         let store_write = segment.open_write(SegmentComponent::STORE)?;
 
         let fast_field_write = segment.open_write(SegmentComponent::FASTFIELDS)?;
@@ -25,15 +26,20 @@ impl SegmentSerializer {
         let fieldnorms_write = segment.open_write(SegmentComponent::FIELDNORMS)?;
         let fieldnorms_serializer = FieldNormsSerializer::from_write(fieldnorms_write)?;
 
-        let postings_serializer = InvertedIndexSerializer::open(segment)?;
+        let postings_serializer = InvertedIndexSerializer::open(&mut segment)?;
         Ok(SegmentSerializer {
+            segment,
             store_writer: StoreWriter::new(store_write),
             fast_field_serializer,
-            fieldnorms_serializer,
+            fieldnorms_serializer: Some(fieldnorms_serializer),
             postings_serializer,
         })
     }
 
+    pub fn segment(&self) -> &Segment {
+        &self.segment
+    }
+
     /// Accessor to the `PostingsSerializer`.
     pub fn get_postings_serializer(&mut self) -> &mut InvertedIndexSerializer {
         &mut self.postings_serializer
@@ -44,9 +50,11 @@ impl SegmentSerializer {
         &mut self.fast_field_serializer
     }
 
-    /// Accessor to the field norm serializer.
-    pub fn get_fieldnorms_serializer(&mut self) -> &mut FieldNormsSerializer {
-        &mut self.fieldnorms_serializer
+    /// Extract the field norm serializer.
+    ///
+    /// Note the fieldnorms serializer can only be extracted once.
+    pub fn extract_fieldnorms_serializer(&mut self) -> Option<FieldNormsSerializer> {
+        self.fieldnorms_serializer.take()
     }
 
     /// Accessor to the `StoreWriter`.
@@ -55,11 +63,13 @@ impl SegmentSerializer {
     }
 
     /// Finalize the segment serialization.
-    pub fn close(self) -> crate::Result<()> {
+    pub fn close(mut self) -> crate::Result<()> {
+        if let Some(fieldnorms_serializer) = self.extract_fieldnorms_serializer() {
+            fieldnorms_serializer.close()?;
+        }
         self.fast_field_serializer.close()?;
         self.postings_serializer.close()?;
         self.store_writer.close()?;
-        self.fieldnorms_serializer.close()?;
         Ok(())
     }
 }

src/indexer/segment_updater.rs

@@ -112,7 +112,7 @@ fn merge(
     target_opstamp: Opstamp,
 ) -> crate::Result<SegmentEntry> {
     // first we need to apply deletes to our segment.
-    let mut merged_segment = index.new_segment();
+    let merged_segment = index.new_segment();
 
     // First we apply all of the delet to the merged segment, up to the target opstamp.
     for segment_entry in &mut segment_entries {
@@ -130,12 +130,14 @@ fn merge(
     // An IndexMerger is like a "view" of our merged segments.
     let merger: IndexMerger = IndexMerger::open(index.schema(), &segments[..])?;
 
+    let merged_segment_id = merged_segment.id();
+
     // ... we just serialize this index merger in our new segment to merge the two segments.
-    let segment_serializer = SegmentSerializer::for_segment(&mut merged_segment)?;
+    let segment_serializer = SegmentSerializer::for_segment(merged_segment)?;
 
     let num_docs = merger.write(segment_serializer)?;
 
-    let segment_meta = index.new_segment_meta(merged_segment.id(), num_docs);
+    let segment_meta = index.new_segment_meta(merged_segment_id, num_docs);
 
     Ok(SegmentEntry::new(segment_meta, delete_cursor, None))
 }
@@ -521,7 +523,7 @@ impl SegmentUpdater {
     ///
     /// Upon termination of the current merging threads,
     /// merge opportunity may appear.
-    //
+    ///
     /// We keep waiting until the merge policy judges that
     /// no opportunity is available.
     ///

src/indexer/segment_writer.rs

@@ -2,7 +2,7 @@ use super::operation::AddOperation;
 use crate::core::Segment;
 use crate::core::SerializableSegment;
 use crate::fastfield::FastFieldsWriter;
-use crate::fieldnorm::FieldNormsWriter;
+use crate::fieldnorm::{FieldNormReaders, FieldNormsWriter};
 use crate::indexer::segment_serializer::SegmentSerializer;
 use crate::postings::compute_table_size;
 use crate::postings::MultiFieldPostingsWriter;
@@ -14,8 +14,8 @@ use crate::schema::{Field, FieldEntry};
 use crate::tokenizer::{BoxTokenStream, PreTokenizedStream};
 use crate::tokenizer::{FacetTokenizer, TextAnalyzer};
 use crate::tokenizer::{TokenStreamChain, Tokenizer};
-use crate::DocId;
 use crate::Opstamp;
+use crate::{DocId, SegmentComponent};
 use std::io;
 use std::str;
 
@@ -62,11 +62,12 @@ impl SegmentWriter {
     /// - schema
     pub fn for_segment(
         memory_budget: usize,
-        mut segment: Segment,
+        segment: Segment,
         schema: &Schema,
     ) -> crate::Result<SegmentWriter> {
+        let tokenizer_manager = segment.index().tokenizers().clone();
         let table_num_bits = initial_table_size(memory_budget)?;
-        let segment_serializer = SegmentSerializer::for_segment(&mut segment)?;
+        let segment_serializer = SegmentSerializer::for_segment(segment)?;
         let multifield_postings = MultiFieldPostingsWriter::new(schema, table_num_bits);
         let tokenizers = schema
             .fields()
@@ -76,7 +77,7 @@ impl SegmentWriter {
                         .get_indexing_options()
                         .and_then(|text_index_option| {
                             let tokenizer_name = &text_index_option.tokenizer();
-                            segment.index().tokenizers().get(tokenizer_name)
+                            tokenizer_manager.get(tokenizer_name)
                         }),
                     _ => None,
                 },
@@ -280,9 +281,16 @@ fn write(
     fieldnorms_writer: &FieldNormsWriter,
     mut serializer: SegmentSerializer,
 ) -> crate::Result<()> {
-    let term_ord_map = multifield_postings.serialize(serializer.get_postings_serializer())?;
+    if let Some(fieldnorms_serializer) = serializer.extract_fieldnorms_serializer() {
+        fieldnorms_writer.serialize(fieldnorms_serializer)?;
+    }
+    let fieldnorm_data = serializer
+        .segment()
+        .open_read(SegmentComponent::FIELDNORMS)?;
+    let fieldnorm_readers = FieldNormReaders::new(fieldnorm_data)?;
+    let term_ord_map =
+        multifield_postings.serialize(serializer.get_postings_serializer(), fieldnorm_readers)?;
     fast_field_writers.serialize(serializer.get_fast_field_serializer(), &term_ord_map)?;
-    fieldnorms_writer.serialize(serializer.get_fieldnorms_serializer())?;
     serializer.close()?;
     Ok(())
 }

src/lib.rs

@@ -173,7 +173,7 @@ use once_cell::sync::Lazy;
 use serde::{Deserialize, Serialize};
 
 /// Index format version.
-const INDEX_FORMAT_VERSION: u32 = 1;
+const INDEX_FORMAT_VERSION: u32 = 2;
 
 /// Structure version for the index.
 #[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -298,17 +298,26 @@ mod tests {
     use rand::rngs::StdRng;
     use rand::{Rng, SeedableRng};
 
-    pub fn assert_nearly_equals(expected: f32, val: f32) {
-        assert!(
-            nearly_equals(val, expected),
-            "Got {}, expected {}.",
-            val,
-            expected
-        );
-    }
-
-    pub fn nearly_equals(a: f32, b: f32) -> bool {
-        (a - b).abs() < 0.0005 * (a + b).abs()
+    /// Checks if left and right are close one to each other.
+    /// Panics if the two values are more than 0.5% apart.
+    #[macro_export]
+    macro_rules! assert_nearly_equals {
+        ($left:expr, $right:expr) => {{
+            match (&$left, &$right) {
+                (left_val, right_val) => {
+                    let diff = (left_val - right_val).abs();
+                    let add = left_val.abs() + right_val.abs();
+                    if diff > 0.0005 * add {
+                        panic!(
+                            r#"assertion failed: `(left ~= right)`
+  left: `{:?}`,
+ right: `{:?}`"#,
+                            &*left_val, &*right_val
+                        )
+                    }
+                }
+            }
+        }};
     }
 
     pub fn generate_nonunique_unsorted(max_value: u32, n_elems: usize) -> Vec<u32> {

src/positions/mod.rs

@@ -37,9 +37,9 @@ const LONG_SKIP_INTERVAL: u64 = (LONG_SKIP_IN_BLOCKS * COMPRESSION_BLOCK_SIZE) a
 #[cfg(test)]
 pub mod tests {
 
-    use super::{PositionReader, PositionSerializer};
+    use super::PositionSerializer;
     use crate::directory::ReadOnlySource;
-    use crate::positions::COMPRESSION_BLOCK_SIZE;
+    use crate::positions::reader::PositionReader;
     use std::iter;
 
     fn create_stream_buffer(vals: &[u32]) -> (ReadOnlySource, ReadOnlySource) {
@@ -68,7 +68,7 @@ pub mod tests {
         let mut position_reader = PositionReader::new(stream, skip, 0u64);
         for &n in &[1, 10, 127, 128, 130, 312] {
             let mut v = vec![0u32; n];
-            position_reader.read(&mut v[..n]);
+            position_reader.read(0, &mut v[..]);
             for i in 0..n {
                 assert_eq!(v[i], i as u32);
             }
@@ -76,19 +76,19 @@ pub mod tests {
     }
 
     #[test]
-    fn test_position_skip() {
-        let v: Vec<u32> = (0..1_000).collect();
+    fn test_position_read_with_offset() {
+        let v: Vec<u32> = (0..1000).collect();
         let (stream, skip) = create_stream_buffer(&v[..]);
         assert_eq!(skip.len(), 12);
         assert_eq!(stream.len(), 1168);
-
         let mut position_reader = PositionReader::new(stream, skip, 0u64);
-        position_reader.skip(10);
-        for &n in &[10, 127, COMPRESSION_BLOCK_SIZE, 130, 312] {
-            let mut v = vec![0u32; n];
-            position_reader.read(&mut v[..n]);
-            for i in 0..n {
-                assert_eq!(v[i], 10u32 + i as u32);
+        for &offset in &[1u64, 10u64, 127u64, 128u64, 130u64, 312u64] {
+            for &len in &[1, 10, 130, 500] {
+                let mut v = vec![0u32; len];
+                position_reader.read(offset, &mut v[..]);
+                for i in 0..len {
+                    assert_eq!(v[i], i as u32 + offset as u32);
+                }
             }
         }
     }
@@ -103,11 +103,12 @@ pub mod tests {
         let mut position_reader = PositionReader::new(stream, skip, 0u64);
         let mut buf = [0u32; 7];
         let mut c = 0;
+
+        let mut offset = 0;
         for _ in 0..100 {
-            position_reader.read(&mut buf);
-            position_reader.read(&mut buf);
-            position_reader.skip(4);
-            position_reader.skip(3);
+            position_reader.read(offset, &mut buf);
+            position_reader.read(offset, &mut buf);
+            offset += 7;
             for &el in &buf {
                 assert_eq!(c, el);
                 c += 1;
@@ -115,6 +116,58 @@ pub mod tests {
         }
     }
 
+    #[test]
+    fn test_position_reread_anchor_different_than_block() {
+        let v: Vec<u32> = (0..2_000_000).collect();
+        let (stream, skip) = create_stream_buffer(&v[..]);
+        assert_eq!(skip.len(), 15_749);
+        assert_eq!(stream.len(), 4_987_872);
+        let mut position_reader = PositionReader::new(stream.clone(), skip.clone(), 0);
+        let mut buf = [0u32; 256];
+        position_reader.read(128, &mut buf);
+        for i in 0..256 {
+            assert_eq!(buf[i], (128 + i) as u32);
+        }
+        position_reader.read(128, &mut buf);
+        for i in 0..256 {
+            assert_eq!(buf[i], (128 + i) as u32);
+        }
+    }
+
+    #[test]
+    #[should_panic(expected = "offset arguments should be increasing.")]
+    fn test_position_panic_if_called_previous_anchor() {
+        let v: Vec<u32> = (0..2_000_000).collect();
+        let (stream, skip) = create_stream_buffer(&v[..]);
+        assert_eq!(skip.len(), 15_749);
+        assert_eq!(stream.len(), 4_987_872);
+        let mut buf = [0u32; 1];
+        let mut position_reader = PositionReader::new(stream.clone(), skip.clone(), 200_000);
+        position_reader.read(230, &mut buf);
+        position_reader.read(9, &mut buf);
+    }
+
+    #[test]
+    fn test_positions_bug() {
+        let mut v: Vec<u32> = vec![];
+        for i in 1..200 {
+            for j in 0..i {
+                v.push(j);
+            }
+        }
+        let (stream, skip) = create_stream_buffer(&v[..]);
+        let mut buf = Vec::new();
+        let mut position_reader = PositionReader::new(stream.clone(), skip.clone(), 0);
+        let mut offset = 0;
+        for i in 1..24 {
+            buf.resize(i, 0);
+            position_reader.read(offset, &mut buf[..]);
+            offset += i as u64;
+            let r: Vec<u32> = (0..i).map(|el| el as u32).collect();
+            assert_eq!(buf, &r[..]);
+        }
+    }
+
     #[test]
     fn test_position_long_skip_const() {
         const CONST_VAL: u32 = 9u32;
@@ -124,7 +177,7 @@ pub mod tests {
         assert_eq!(stream.len(), 1_000_000);
         let mut position_reader = PositionReader::new(stream, skip, 128 * 1024);
         let mut buf = [0u32; 1];
-        position_reader.read(&mut buf);
+        position_reader.read(0, &mut buf);
         assert_eq!(buf[0], CONST_VAL);
     }
 
@@ -143,7 +196,7 @@ pub mod tests {
         ] {
             let mut position_reader = PositionReader::new(stream.clone(), skip.clone(), offset);
             let mut buf = [0u32; 1];
-            position_reader.read(&mut buf);
+            position_reader.read(0, &mut buf);
             assert_eq!(buf[0], offset as u32);
         }
     }

src/positions/reader.rs

@@ -3,7 +3,6 @@ use crate::directory::ReadOnlySource;
 use crate::positions::COMPRESSION_BLOCK_SIZE;
 use crate::positions::LONG_SKIP_INTERVAL;
 use crate::positions::LONG_SKIP_IN_BLOCKS;
-use crate::postings::compression::compressed_block_size;
 /// Positions works as a long sequence of compressed block.
 /// All terms are chained one after the other.
 ///
@@ -62,22 +61,20 @@ impl Positions {
 
     fn reader(&self, offset: u64) -> PositionReader {
         let long_skip_id = (offset / LONG_SKIP_INTERVAL) as usize;
-        let small_skip = (offset % LONG_SKIP_INTERVAL) as usize;
         let offset_num_bytes: u64 = self.long_skip(long_skip_id);
         let mut position_read = OwnedRead::new(self.position_source.clone());
         position_read.advance(offset_num_bytes as usize);
         let mut skip_read = OwnedRead::new(self.skip_source.clone());
         skip_read.advance(long_skip_id * LONG_SKIP_IN_BLOCKS);
-        let mut position_reader = PositionReader {
+        PositionReader {
             bit_packer: self.bit_packer,
             skip_read,
             position_read,
-            inner_offset: 0,
             buffer: Box::new([0u32; 128]),
-            ahead: None,
-        };
-        position_reader.skip(small_skip);
-        position_reader
+            block_offset: std::i64::MAX as u64,
+            anchor_offset: (long_skip_id as u64) * LONG_SKIP_INTERVAL,
+            abs_offset: offset,
+        }
     }
 }
 
@@ -85,51 +82,12 @@ pub struct PositionReader {
     skip_read: OwnedRead,
     position_read: OwnedRead,
     bit_packer: BitPacker4x,
-    inner_offset: usize,
-    buffer: Box<[u32; 128]>,
-    ahead: Option<usize>, // if None, no block is loaded.
-                          // if Some(num_blocks), the block currently loaded is num_blocks ahead
-                          // of the block of the next int to read.
-}
+    buffer: Box<[u32; COMPRESSION_BLOCK_SIZE]>,
 
-// `ahead` represents the offset of the block currently loaded
-// compared to the cursor of the actual stream.
-//
-// By contract, when this function is called, the current block has to be
-// decompressed.
-//
-// If the requested number of els ends exactly at a given block, the next
-// block is not decompressed.
-fn read_impl(
-    bit_packer: BitPacker4x,
-    mut position: &[u8],
-    buffer: &mut [u32; 128],
-    mut inner_offset: usize,
-    num_bits: &[u8],
-    output: &mut [u32],
-) -> usize {
-    let mut output_start = 0;
-    let mut output_len = output.len();
-    let mut ahead = 0;
-    loop {
-        let available_len = COMPRESSION_BLOCK_SIZE - inner_offset;
-        // We have enough elements in the current block.
-        // Let's copy the requested elements in the output buffer,
-        // and return.
-        if output_len <= available_len {
-            output[output_start..].copy_from_slice(&buffer[inner_offset..][..output_len]);
-            return ahead;
-        }
-        output[output_start..][..available_len].copy_from_slice(&buffer[inner_offset..]);
-        output_len -= available_len;
-        output_start += available_len;
-        inner_offset = 0;
-        let num_bits = num_bits[ahead];
-        bit_packer.decompress(position, &mut buffer[..], num_bits);
-        let block_len = compressed_block_size(num_bits);
-        position = &position[block_len..];
-        ahead += 1;
-    }
+    block_offset: u64,
+    anchor_offset: u64,
+
+    abs_offset: u64,
 }
 
 impl PositionReader {
@@ -141,57 +99,65 @@ impl PositionReader {
         Positions::new(position_source, skip_source).reader(offset)
     }
 
-    /// Fills a buffer with the next `output.len()` integers.
-    /// This does not consume / advance the stream.
-    pub fn read(&mut self, output: &mut [u32]) {
-        let skip_data = self.skip_read.as_ref();
-        let position_data = self.position_read.as_ref();
-        let num_bits = self.skip_read.get(0);
-        if self.ahead != Some(0) {
-            // the block currently available is not the block
-            // for the current position
+    fn advance_num_blocks(&mut self, num_blocks: usize) {
+        let num_bits: usize = self.skip_read.as_ref()[..num_blocks]
+            .iter()
+            .cloned()
+            .map(|num_bits| num_bits as usize)
+            .sum();
+        let num_bytes_to_skip = num_bits * COMPRESSION_BLOCK_SIZE / 8;
+        self.skip_read.advance(num_blocks as usize);
+        self.position_read.advance(num_bytes_to_skip);
+    }
+
+    /// Fills a buffer with the positions `[offset..offset+output.len())` integers.
+    ///
+    /// `offset` is required to have a value >= to the offsets given in previous calls
+    /// for the given `PositionReaderAbsolute` instance.
+    pub fn read(&mut self, mut offset: u64, mut output: &mut [u32]) {
+        offset += self.abs_offset;
+        assert!(
+            offset >= self.anchor_offset,
+            "offset arguments should be increasing."
+        );
+        let delta_to_block_offset = offset as i64 - self.block_offset as i64;
+        if delta_to_block_offset < 0 || delta_to_block_offset >= 128 {
+            // The first position is not within the first block.
+            // We need to decompress the first block.
+            let delta_to_anchor_offset = offset - self.anchor_offset;
+            let num_blocks_to_skip =
+                (delta_to_anchor_offset / (COMPRESSION_BLOCK_SIZE as u64)) as usize;
+            self.advance_num_blocks(num_blocks_to_skip);
+            self.anchor_offset = offset - (offset % COMPRESSION_BLOCK_SIZE as u64);
+            self.block_offset = self.anchor_offset;
+            let num_bits = self.skip_read.get(0);
+            self.bit_packer
+                .decompress(self.position_read.as_ref(), self.buffer.as_mut(), num_bits);
+        } else {
+            let num_blocks_to_skip =
+                ((self.block_offset - self.anchor_offset) / COMPRESSION_BLOCK_SIZE as u64) as usize;
+            self.advance_num_blocks(num_blocks_to_skip);
+            self.anchor_offset = self.block_offset;
+        }
+
+        let mut num_bits = self.skip_read.get(0);
+        let mut position_data = self.position_read.as_ref();
+
+        for i in 1.. {
+            let offset_in_block = (offset as usize) % COMPRESSION_BLOCK_SIZE;
+            let remaining_in_block = COMPRESSION_BLOCK_SIZE - offset_in_block;
+            if remaining_in_block >= output.len() {
+                output.copy_from_slice(&self.buffer[offset_in_block..][..output.len()]);
+                break;
+            }
+            output[..remaining_in_block].copy_from_slice(&self.buffer[offset_in_block..]);
+            output = &mut output[remaining_in_block..];
+            offset += remaining_in_block as u64;
+            position_data = &position_data[(num_bits as usize * COMPRESSION_BLOCK_SIZE / 8)..];
+            num_bits = self.skip_read.get(i);
             self.bit_packer
                 .decompress(position_data, self.buffer.as_mut(), num_bits);
-            self.ahead = Some(0);
+            self.block_offset += COMPRESSION_BLOCK_SIZE as u64;
         }
-        let block_len = compressed_block_size(num_bits);
-        self.ahead = Some(read_impl(
-            self.bit_packer,
-            &position_data[block_len..],
-            self.buffer.as_mut(),
-            self.inner_offset,
-            &skip_data[1..],
-            output,
-        ));
-    }
-
-    /// Skip the next `skip_len` integer.
-    ///
-    /// If a full block is skipped, calling
-    /// `.skip(...)` will avoid decompressing it.
-    ///
-    /// May panic if the end of the stream is reached.
-    pub fn skip(&mut self, skip_len: usize) {
-        let skip_len_plus_inner_offset = skip_len + self.inner_offset;
-
-        let num_blocks_to_advance = skip_len_plus_inner_offset / COMPRESSION_BLOCK_SIZE;
-        self.inner_offset = skip_len_plus_inner_offset % COMPRESSION_BLOCK_SIZE;
-
-        self.ahead = self.ahead.and_then(|num_blocks| {
-            if num_blocks >= num_blocks_to_advance {
-                Some(num_blocks - num_blocks_to_advance)
-            } else {
-                None
-            }
-        });
-
-        let skip_len_in_bits = self.skip_read.as_ref()[..num_blocks_to_advance]
-            .iter()
-            .map(|num_bits| *num_bits as usize)
-            .sum::<usize>()
-            * COMPRESSION_BLOCK_SIZE;
-        let skip_len_in_bytes = skip_len_in_bits / 8;
-        self.skip_read.advance(num_blocks_to_advance);
-        self.position_read.advance(skip_len_in_bytes);
     }
 }

src/postings/block_search.rs

@@ -87,6 +87,7 @@ fn exponential_search(arr: &[u32], target: u32) -> (usize, usize) {
     (begin, end)
 }
 
+#[inline(never)]
 fn galloping(block_docs: &[u32], target: u32) -> usize {
     let (start, end) = exponential_search(&block_docs, target);
     start + linear_search(&block_docs[start..end], target)
@@ -133,19 +134,14 @@ impl BlockSearcher {
     ///
     /// Indeed, if the block is not full, the remaining items are TERMINATED.
     /// It is surprisingly faster, most likely because of the lack of branch misprediction.
-    pub(crate) fn search_in_block(
-        self,
-        block_docs: &AlignedBuffer,
-        start: usize,
-        target: u32,
-    ) -> usize {
+    pub(crate) fn search_in_block(self, block_docs: &AlignedBuffer, target: u32) -> usize {
         #[cfg(target_arch = "x86_64")]
         {
             if self == BlockSearcher::SSE2 {
                 return sse2::linear_search_sse2_128(block_docs, target);
             }
         }
-        start + galloping(&block_docs.0[start..], target)
+        galloping(&block_docs.0[..], target)
     }
 }
 
@@ -199,12 +195,10 @@ mod tests {
         assert!(block.len() < COMPRESSION_BLOCK_SIZE);
         let mut output_buffer = [TERMINATED; COMPRESSION_BLOCK_SIZE];
         output_buffer[..block.len()].copy_from_slice(block);
-        for i in 0..cursor {
-            assert_eq!(
-                block_searcher.search_in_block(&AlignedBuffer(output_buffer), i, target),
-                cursor
-            );
-        }
+        assert_eq!(
+            block_searcher.search_in_block(&AlignedBuffer(output_buffer), target),
+            cursor
+        );
     }
 
     fn util_test_search_in_block_all(block_searcher: BlockSearcher, block: &[u32]) {

src/postings/block_segment_postings.rs

@@ -0,0 +1,515 @@
+use crate::common::{BinarySerializable, VInt};
+use crate::directory::ReadOnlySource;
+use crate::fieldnorm::FieldNormReader;
+use crate::postings::compression::{
+    AlignedBuffer, BlockDecoder, VIntDecoder, COMPRESSION_BLOCK_SIZE,
+};
+use crate::postings::{BlockInfo, FreqReadingOption, SkipReader};
+use crate::query::BM25Weight;
+use crate::schema::IndexRecordOption;
+use crate::{DocId, Score, TERMINATED};
+
+fn max_f32<I: Iterator<Item = f32>>(mut it: I) -> Option<f32> {
+    if let Some(first) = it.next() {
+        Some(it.fold(first, f32::max))
+    } else {
+        None
+    }
+}
+
+/// `BlockSegmentPostings` is a cursor iterating over blocks
+/// of documents.
+///
+/// # Warning
+///
+/// While it is useful for some very specific high-performance
+/// use cases, you should prefer using `SegmentPostings` for most usage.
+pub struct BlockSegmentPostings {
+    pub(crate) doc_decoder: BlockDecoder,
+    loaded_offset: usize,
+    freq_decoder: BlockDecoder,
+    freq_reading_option: FreqReadingOption,
+    block_max_score_cache: Option<Score>,
+
+    doc_freq: u32,
+
+    data: ReadOnlySource,
+    pub(crate) skip_reader: SkipReader,
+}
+
+fn decode_bitpacked_block(
+    doc_decoder: &mut BlockDecoder,
+    freq_decoder_opt: Option<&mut BlockDecoder>,
+    data: &[u8],
+    doc_offset: DocId,
+    doc_num_bits: u8,
+    tf_num_bits: u8,
+) {
+    let num_consumed_bytes = doc_decoder.uncompress_block_sorted(data, doc_offset, doc_num_bits);
+    if let Some(freq_decoder) = freq_decoder_opt {
+        freq_decoder.uncompress_block_unsorted(&data[num_consumed_bytes..], tf_num_bits);
+    }
+}
+
+fn decode_vint_block(
+    doc_decoder: &mut BlockDecoder,
+    freq_decoder_opt: Option<&mut BlockDecoder>,
+    data: &[u8],
+    doc_offset: DocId,
+    num_vint_docs: usize,
+) {
+    doc_decoder.fill(TERMINATED);
+    let num_consumed_bytes = doc_decoder.uncompress_vint_sorted(data, doc_offset, num_vint_docs);
+    if let Some(freq_decoder) = freq_decoder_opt {
+        freq_decoder.uncompress_vint_unsorted(&data[num_consumed_bytes..], num_vint_docs);
+    }
+}
+
+fn split_into_skips_and_postings(
+    doc_freq: u32,
+    data: ReadOnlySource,
+) -> (Option<ReadOnlySource>, ReadOnlySource) {
+    if doc_freq < COMPRESSION_BLOCK_SIZE as u32 {
+        return (None, data);
+    }
+    let mut data_byte_arr = data.as_slice();
+    let skip_len = VInt::deserialize(&mut data_byte_arr)
+        .expect("Data corrupted")
+        .0 as usize;
+    let vint_len = data.len() - data_byte_arr.len();
+    let (skip_data, postings_data) = data.slice_from(vint_len).split(skip_len);
+    (Some(skip_data), postings_data)
+}
+
+impl BlockSegmentPostings {
+    pub(crate) fn from_data(
+        doc_freq: u32,
+        data: ReadOnlySource,
+        record_option: IndexRecordOption,
+        requested_option: IndexRecordOption,
+    ) -> BlockSegmentPostings {
+        let freq_reading_option = match (record_option, requested_option) {
+            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
+            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
+            (_, _) => FreqReadingOption::ReadFreq,
+        };
+
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
+        let skip_reader = match skip_data_opt {
+            Some(skip_data) => SkipReader::new(skip_data, doc_freq, record_option),
+            None => SkipReader::new(ReadOnlySource::empty(), doc_freq, record_option),
+        };
+
+        let mut block_segment_postings = BlockSegmentPostings {
+            doc_decoder: BlockDecoder::with_val(TERMINATED),
+            loaded_offset: std::usize::MAX,
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option,
+            block_max_score_cache: None,
+            doc_freq,
+            data: postings_data,
+            skip_reader,
+        };
+        block_segment_postings.load_block();
+        block_segment_postings
+    }
+
+    /// Returns the block_max_score for the current block.
+    /// It does not require the block to be loaded. For instance, it is ok to call this method
+    /// after having called `.shallow_advance(..)`.
+    ///
+    /// See `TermScorer::block_max_score(..)` for more information.
+    pub fn block_max_score(
+        &mut self,
+        fieldnorm_reader: &FieldNormReader,
+        bm25_weight: &BM25Weight,
+    ) -> Score {
+        let (block_max_score_cache, skip_reader, doc_decoder, freq_decoder) = (
+            &mut self.block_max_score_cache,
+            &self.skip_reader,
+            &self.doc_decoder,
+            &self.freq_decoder,
+        );
+        *block_max_score_cache.get_or_insert_with(|| {
+            skip_reader
+                .block_max_score(bm25_weight)
+                .or_else(|| {
+                    let docs = doc_decoder.output_array();
+                    let freqs = freq_decoder.output_array();
+                    max_f32(docs.iter().cloned().zip(freqs.iter().cloned()).map(
+                        |(doc, term_freq)| {
+                            let fieldnorm_id = fieldnorm_reader.fieldnorm_id(doc);
+                            bm25_weight.score(fieldnorm_id, term_freq)
+                        },
+                    ))
+                })
+                .unwrap_or(0f32)
+        })
+    }
+
+    pub(crate) fn freq_reading_option(&self) -> FreqReadingOption {
+        self.freq_reading_option
+    }
+
+    // Resets the block segment postings on another position
+    // in the postings file.
+    //
+    // This is useful for enumerating through a list of terms,
+    // and consuming the associated posting lists while avoiding
+    // reallocating a `BlockSegmentPostings`.
+    //
+    // # Warning
+    //
+    // This does not reset the positions list.
+    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: ReadOnlySource) {
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
+        self.data = ReadOnlySource::new(postings_data);
+        self.loaded_offset = std::usize::MAX;
+        if let Some(skip_data) = skip_data_opt {
+            self.skip_reader.reset(skip_data, doc_freq);
+        } else {
+            self.skip_reader.reset(ReadOnlySource::empty(), doc_freq);
+        }
+        self.doc_freq = doc_freq;
+        self.load_block();
+    }
+
+    /// Returns the overall number of documents in the block postings.
+    /// It does not take in account whether documents are deleted or not.
+    ///
+    /// This `doc_freq` is simply the sum of the length of all of the blocks
+    /// length, and it does not take in account deleted documents.
+    pub fn doc_freq(&self) -> u32 {
+        self.doc_freq
+    }
+
+    /// Returns the array of docs in the current block.
+    ///
+    /// Before the first call to `.advance()`, the block
+    /// returned by `.docs()` is empty.
+    #[inline]
+    pub fn docs(&self) -> &[DocId] {
+        debug_assert!(self.block_is_loaded());
+        self.doc_decoder.output_array()
+    }
+
+    /// Returns a full block, regardless of whetehr the block is complete or incomplete (
+    /// as it happens for the last block of the posting list).
+    ///
+    /// In the latter case, the block is guaranteed to be padded with the sentinel value:
+    /// `TERMINATED`. The array is also guaranteed to be aligned on 16 bytes = 128 bits.
+    ///
+    /// This method is useful to run SSE2 linear search.
+    #[inline(always)]
+    pub(crate) fn docs_aligned(&self) -> &AlignedBuffer {
+        debug_assert!(self.block_is_loaded());
+        self.doc_decoder.output_aligned()
+    }
+
+    /// Return the document at index `idx` of the block.
+    #[inline(always)]
+    pub fn doc(&self, idx: usize) -> u32 {
+        self.doc_decoder.output(idx)
+    }
+
+    /// Return the array of `term freq` in the block.
+    #[inline]
+    pub fn freqs(&self) -> &[u32] {
+        debug_assert!(self.block_is_loaded());
+        self.freq_decoder.output_array()
+    }
+
+    /// Return the frequency at index `idx` of the block.
+    #[inline]
+    pub fn freq(&self, idx: usize) -> u32 {
+        debug_assert!(self.block_is_loaded());
+        self.freq_decoder.output(idx)
+    }
+
+    /// Returns the length of the current block.
+    ///
+    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
+    /// except the last block that may have a length
+    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
+    #[inline]
+    pub fn block_len(&self) -> usize {
+        debug_assert!(self.block_is_loaded());
+        self.doc_decoder.output_len
+    }
+
+    /// Position on a block that may contains `target_doc`.
+    ///
+    /// If all docs are smaller than target, the block loaded may be empty,
+    /// or be the last an incomplete VInt block.
+    pub fn seek(&mut self, target_doc: DocId) {
+        self.skip_reader.seek(target_doc);
+        self.load_block();
+    }
+
+    pub(crate) fn position_offset(&self) -> u64 {
+        self.skip_reader.position_offset()
+    }
+
+    /// Dangerous API! This calls seek on the skip list,
+    /// but does not `.load_block()` afterwards.
+    ///
+    /// `.load_block()` needs to be called manually afterwards.
+    /// If all docs are smaller than target, the block loaded may be empty,
+    /// or be the last an incomplete VInt block.
+    pub(crate) fn shallow_seek(&mut self, target_doc: DocId) {
+        self.skip_reader.seek(target_doc);
+    }
+
+    pub(crate) fn block_is_loaded(&self) -> bool {
+        self.loaded_offset == self.skip_reader.byte_offset()
+    }
+
+    pub(crate) fn load_block(&mut self) {
+        self.block_max_score_cache = None;
+        let offset = self.skip_reader.byte_offset();
+        if self.loaded_offset == offset {
+            return;
+        }
+        self.loaded_offset = offset;
+        match self.skip_reader.block_info() {
+            BlockInfo::BitPacked {
+                doc_num_bits,
+                tf_num_bits,
+                ..
+            } => {
+                decode_bitpacked_block(
+                    &mut self.doc_decoder,
+                    if let FreqReadingOption::ReadFreq = self.freq_reading_option {
+                        Some(&mut self.freq_decoder)
+                    } else {
+                        None
+                    },
+                    &self.data.as_slice()[offset..],
+                    self.skip_reader.last_doc_in_previous_block,
+                    doc_num_bits,
+                    tf_num_bits,
+                );
+            }
+            BlockInfo::VInt { num_docs } => {
+                let data = {
+                    if num_docs == 0 {
+                        &[]
+                    } else {
+                        &self.data.as_slice()[offset..]
+                    }
+                };
+                decode_vint_block(
+                    &mut self.doc_decoder,
+                    if let FreqReadingOption::ReadFreq = self.freq_reading_option {
+                        Some(&mut self.freq_decoder)
+                    } else {
+                        None
+                    },
+                    data,
+                    self.skip_reader.last_doc_in_previous_block,
+                    num_docs as usize,
+                );
+            }
+        }
+    }
+
+    /// Advance to the next block.
+    ///
+    /// Returns false iff there was no remaining blocks.
+    pub fn advance(&mut self) {
+        self.skip_reader.advance();
+        self.load_block();
+    }
+
+    /// Returns an empty segment postings object
+    pub fn empty() -> BlockSegmentPostings {
+        BlockSegmentPostings {
+            doc_decoder: BlockDecoder::with_val(TERMINATED),
+            loaded_offset: 0,
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option: FreqReadingOption::NoFreq,
+            block_max_score_cache: None,
+            doc_freq: 0,
+            data: ReadOnlySource::new(vec![]),
+            skip_reader: SkipReader::new(ReadOnlySource::new(vec![]), 0, IndexRecordOption::Basic),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::BlockSegmentPostings;
+    use crate::common::HasLen;
+    use crate::core::Index;
+    use crate::docset::{DocSet, TERMINATED};
+    use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+    use crate::postings::postings::Postings;
+    use crate::postings::SegmentPostings;
+    use crate::schema::IndexRecordOption;
+    use crate::schema::Schema;
+    use crate::schema::Term;
+    use crate::schema::INDEXED;
+    use crate::DocId;
+
+    #[test]
+    fn test_empty_segment_postings() {
+        let mut postings = SegmentPostings::empty();
+        assert_eq!(postings.doc(), TERMINATED);
+        assert_eq!(postings.advance(), TERMINATED);
+        assert_eq!(postings.advance(), TERMINATED);
+        assert_eq!(postings.doc_freq(), 0);
+        assert_eq!(postings.len(), 0);
+    }
+
+    #[test]
+    fn test_empty_postings_doc_returns_terminated() {
+        let mut postings = SegmentPostings::empty();
+        assert_eq!(postings.doc(), TERMINATED);
+        assert_eq!(postings.advance(), TERMINATED);
+    }
+
+    #[test]
+    fn test_empty_postings_doc_term_freq_returns_0() {
+        let postings = SegmentPostings::empty();
+        assert_eq!(postings.term_freq(), 1);
+    }
+
+    #[test]
+    fn test_empty_block_segment_postings() {
+        let mut postings = BlockSegmentPostings::empty();
+        assert!(postings.docs().is_empty());
+        assert_eq!(postings.doc_freq(), 0);
+        postings.advance();
+        assert!(postings.docs().is_empty());
+        assert_eq!(postings.doc_freq(), 0);
+    }
+
+    #[test]
+    fn test_block_segment_postings() {
+        let mut block_segments = build_block_postings(&(0..100_000).collect::<Vec<u32>>());
+        let mut offset: u32 = 0u32;
+        // checking that the `doc_freq` is correct
+        assert_eq!(block_segments.doc_freq(), 100_000);
+        loop {
+            let block = block_segments.docs();
+            if block.is_empty() {
+                break;
+            }
+            for (i, doc) in block.iter().cloned().enumerate() {
+                assert_eq!(offset + (i as u32), doc);
+            }
+            offset += block.len() as u32;
+            block_segments.advance();
+        }
+    }
+
+    #[test]
+    fn test_skip_right_at_new_block() {
+        let mut doc_ids = (0..128).collect::<Vec<u32>>();
+        // 128 is missing
+        doc_ids.push(129);
+        doc_ids.push(130);
+        {
+            let block_segments = build_block_postings(&doc_ids);
+            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
+            assert_eq!(docset.seek(128), 129);
+            assert_eq!(docset.doc(), 129);
+            assert_eq!(docset.advance(), 130);
+            assert_eq!(docset.doc(), 130);
+            assert_eq!(docset.advance(), TERMINATED);
+        }
+        {
+            let block_segments = build_block_postings(&doc_ids);
+            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
+            assert_eq!(docset.seek(129), 129);
+            assert_eq!(docset.doc(), 129);
+            assert_eq!(docset.advance(), 130);
+            assert_eq!(docset.doc(), 130);
+            assert_eq!(docset.advance(), TERMINATED);
+        }
+        {
+            let block_segments = build_block_postings(&doc_ids);
+            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
+            assert_eq!(docset.doc(), 0);
+            assert_eq!(docset.seek(131), TERMINATED);
+            assert_eq!(docset.doc(), TERMINATED);
+        }
+    }
+
+    fn build_block_postings(docs: &[DocId]) -> BlockSegmentPostings {
+        let mut schema_builder = Schema::builder();
+        let int_field = schema_builder.add_u64_field("id", INDEXED);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        let mut last_doc = 0u32;
+        for &doc in docs {
+            for _ in last_doc..doc {
+                index_writer.add_document(doc!(int_field=>1u64));
+            }
+            index_writer.add_document(doc!(int_field=>0u64));
+            last_doc = doc + 1;
+        }
+        index_writer.commit().unwrap();
+        let searcher = index.reader().unwrap().searcher();
+        let segment_reader = searcher.segment_reader(0);
+        let inverted_index = segment_reader.inverted_index(int_field);
+        let term = Term::from_field_u64(int_field, 0u64);
+        let term_info = inverted_index.get_term_info(&term).unwrap();
+        inverted_index.read_block_postings_from_terminfo(&term_info, IndexRecordOption::Basic)
+    }
+
+    #[test]
+    fn test_block_segment_postings_seek2() {
+        let mut docs = vec![0];
+        for i in 0..1300 {
+            docs.push((i * i / 100) + i);
+        }
+        let mut block_postings = build_block_postings(&docs[..]);
+        for i in vec![0, 424, 10000] {
+            block_postings.shallow_seek(i);
+            block_postings.load_block();
+            let docs = block_postings.docs();
+            assert!(docs[0] <= i);
+            assert!(docs.last().cloned().unwrap_or(0u32) >= i);
+        }
+        block_postings.shallow_seek(100_000);
+        block_postings.load_block();
+        assert_eq!(block_postings.doc(COMPRESSION_BLOCK_SIZE - 1), TERMINATED);
+    }
+
+    #[test]
+    fn test_reset_block_segment_postings() {
+        let mut schema_builder = Schema::builder();
+        let int_field = schema_builder.add_u64_field("id", INDEXED);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        // create two postings list, one containg even number,
+        // the other containing odd numbers.
+        for i in 0..6 {
+            let doc = doc!(int_field=> (i % 2) as u64);
+            index_writer.add_document(doc);
+        }
+        index_writer.commit().unwrap();
+        let searcher = index.reader().unwrap().searcher();
+        let segment_reader = searcher.segment_reader(0);
+
+        let mut block_segments;
+        {
+            let term = Term::from_field_u64(int_field, 0u64);
+            let inverted_index = segment_reader.inverted_index(int_field);
+            let term_info = inverted_index.get_term_info(&term).unwrap();
+            block_segments = inverted_index
+                .read_block_postings_from_terminfo(&term_info, IndexRecordOption::Basic);
+        }
+        assert_eq!(block_segments.docs(), &[0, 2, 4]);
+        {
+            let term = Term::from_field_u64(int_field, 1u64);
+            let inverted_index = segment_reader.inverted_index(int_field);
+            let term_info = inverted_index.get_term_info(&term).unwrap();
+            inverted_index.reset_block_postings_from_terminfo(&term_info, &mut block_segments);
+        }
+        assert_eq!(block_segments.docs(), &[1, 3, 5]);
+    }
+}

src/postings/compression/mod.rs

@@ -18,6 +18,12 @@ pub struct BlockEncoder {
     pub output_len: usize,
 }
 
+impl Default for BlockEncoder {
+    fn default() -> Self {
+        BlockEncoder::new()
+    }
+}
+
 impl BlockEncoder {
     pub fn new() -> BlockEncoder {
         BlockEncoder {
@@ -55,11 +61,13 @@ pub struct BlockDecoder {
     pub output_len: usize,
 }
 
-impl BlockDecoder {
-    pub fn new() -> BlockDecoder {
+impl Default for BlockDecoder {
+    fn default() -> Self {
         BlockDecoder::with_val(0u32)
     }
+}
 
+impl BlockDecoder {
     pub fn with_val(val: u32) -> BlockDecoder {
         BlockDecoder {
             bitpacker: BitPacker4x::new(),
@@ -100,8 +108,8 @@ impl BlockDecoder {
         self.output.0[idx]
     }
 
-    pub fn clear(&mut self) {
-        self.output.0.iter_mut().for_each(|el| *el = TERMINATED);
+    pub fn fill(&mut self, val: u32) {
+        self.output.0.iter_mut().for_each(|el| *el = val);
     }
 }
 
@@ -175,7 +183,7 @@ impl VIntDecoder for BlockDecoder {
         vint::uncompress_sorted(compressed_data, &mut self.output.0[..num_els], offset)
     }
 
-    fn uncompress_vint_unsorted<'a>(&mut self, compressed_data: &'a [u8], num_els: usize) -> usize {
+    fn uncompress_vint_unsorted(&mut self, compressed_data: &[u8], num_els: usize) -> usize {
         self.output_len = num_els;
         vint::uncompress_unsorted(compressed_data, &mut self.output.0[..num_els])
     }
@@ -191,7 +199,7 @@ pub mod tests {
         let vals: Vec<u32> = (0u32..128u32).map(|i| i * 7).collect();
         let mut encoder = BlockEncoder::new();
         let (num_bits, compressed_data) = encoder.compress_block_sorted(&vals, 0);
-        let mut decoder = BlockDecoder::new();
+        let mut decoder = BlockDecoder::default();
         {
             let consumed_num_bytes = decoder.uncompress_block_sorted(compressed_data, 0, num_bits);
             assert_eq!(consumed_num_bytes, compressed_data.len());
@@ -204,9 +212,9 @@ pub mod tests {
     #[test]
     fn test_encode_sorted_block_with_offset() {
         let vals: Vec<u32> = (0u32..128u32).map(|i| 11 + i * 7).collect();
-        let mut encoder = BlockEncoder::new();
+        let mut encoder = BlockEncoder::default();
         let (num_bits, compressed_data) = encoder.compress_block_sorted(&vals, 10);
-        let mut decoder = BlockDecoder::new();
+        let mut decoder = BlockDecoder::default();
         {
             let consumed_num_bytes = decoder.uncompress_block_sorted(compressed_data, 10, num_bits);
             assert_eq!(consumed_num_bytes, compressed_data.len());
@@ -221,11 +229,11 @@ pub mod tests {
         let mut compressed: Vec<u8> = Vec::new();
         let n = 128;
         let vals: Vec<u32> = (0..n).map(|i| 11u32 + (i as u32) * 7u32).collect();
-        let mut encoder = BlockEncoder::new();
+        let mut encoder = BlockEncoder::default();
         let (num_bits, compressed_data) = encoder.compress_block_sorted(&vals, 10);
         compressed.extend_from_slice(compressed_data);
         compressed.push(173u8);
-        let mut decoder = BlockDecoder::new();
+        let mut decoder = BlockDecoder::default();
         {
             let consumed_num_bytes = decoder.uncompress_block_sorted(&compressed, 10, num_bits);
             assert_eq!(consumed_num_bytes, compressed.len() - 1);
@@ -241,11 +249,11 @@ pub mod tests {
         let mut compressed: Vec<u8> = Vec::new();
         let n = 128;
         let vals: Vec<u32> = (0..n).map(|i| 11u32 + (i as u32) * 7u32 % 12).collect();
-        let mut encoder = BlockEncoder::new();
+        let mut encoder = BlockEncoder::default();
         let (num_bits, compressed_data) = encoder.compress_block_unsorted(&vals);
         compressed.extend_from_slice(compressed_data);
         compressed.push(173u8);
-        let mut decoder = BlockDecoder::new();
+        let mut decoder = BlockDecoder::default();
         {
             let consumed_num_bytes = decoder.uncompress_block_unsorted(&compressed, num_bits);
             assert_eq!(consumed_num_bytes + 1, compressed.len());
@@ -256,6 +264,11 @@ pub mod tests {
         }
     }
 
+    #[test]
+    fn test_block_decoder_initialization() {
+        let block = BlockDecoder::with_val(TERMINATED);
+        assert_eq!(block.output(0), TERMINATED);
+    }
     #[test]
     fn test_encode_vint() {
         {
@@ -265,7 +278,7 @@ pub mod tests {
             for offset in &[0u32, 1u32, 2u32] {
                 let encoded_data = encoder.compress_vint_sorted(&input, *offset);
                 assert!(encoded_data.len() <= expected_length);
-                let mut decoder = BlockDecoder::new();
+                let mut decoder = BlockDecoder::default();
                 let consumed_num_bytes =
                     decoder.uncompress_vint_sorted(&encoded_data, *offset, input.len());
                 assert_eq!(consumed_num_bytes, encoded_data.len());

src/postings/mod.rs

@@ -3,11 +3,8 @@ Postings module (also called inverted index)
 */
 
 mod block_search;
+mod block_segment_postings;
 pub(crate) mod compression;
-/// Postings module
-///
-/// Postings, also called inverted lists, is the key datastructure
-/// to full-text search.
 mod postings;
 mod postings_writer;
 mod recorder;
@@ -22,18 +19,17 @@ pub(crate) use self::block_search::BlockSearcher;
 pub(crate) use self::postings_writer::MultiFieldPostingsWriter;
 pub use self::serializer::{FieldSerializer, InvertedIndexSerializer};
 
-use self::compression::COMPRESSION_BLOCK_SIZE;
 pub use self::postings::Postings;
-pub(crate) use self::skip::SkipReader;
+pub(crate) use self::skip::{BlockInfo, SkipReader};
 pub use self::term_info::TermInfo;
 
-pub use self::segment_postings::{BlockSegmentPostings, SegmentPostings};
+pub use self::block_segment_postings::BlockSegmentPostings;
+pub use self::segment_postings::SegmentPostings;
 
 pub(crate) use self::stacker::compute_table_size;
 
 pub use crate::common::HasLen;
 
-pub(crate) const USE_SKIP_INFO_LIMIT: u32 = COMPRESSION_BLOCK_SIZE as u32;
 pub(crate) type UnorderedTermId = u64;
 
 #[cfg_attr(feature = "cargo-clippy", allow(clippy::enum_variant_names))]
@@ -77,8 +73,10 @@ pub mod tests {
         let mut segment = index.new_segment();
         let mut posting_serializer = InvertedIndexSerializer::open(&mut segment).unwrap();
         {
-            let mut field_serializer = posting_serializer.new_field(text_field, 120 * 4).unwrap();
-            field_serializer.new_term("abc".as_bytes()).unwrap();
+            let mut field_serializer = posting_serializer
+                .new_field(text_field, 120 * 4, None)
+                .unwrap();
+            field_serializer.new_term("abc".as_bytes(), 12u32).unwrap();
             for doc_id in 0u32..120u32 {
                 let delta_positions = vec![1, 2, 3, 2];
                 field_serializer
@@ -586,6 +584,9 @@ pub mod tests {
     ) {
         for target in targets {
             let mut postings_opt = postings_factory();
+            if target < postings_opt.doc() {
+                continue;
+            }
             let mut postings_unopt = UnoptimizedDocSet::wrap(postings_factory());
             let skip_result_opt = postings_opt.seek(target);
             let skip_result_unopt = postings_unopt.seek(target);

src/postings/postings_writer.rs

@@ -1,5 +1,6 @@
 use super::stacker::{Addr, MemoryArena, TermHashMap};
 
+use crate::fieldnorm::FieldNormReaders;
 use crate::postings::recorder::{
     BufferLender, NothingRecorder, Recorder, TFAndPositionRecorder, TermFrequencyRecorder,
 };
@@ -128,6 +129,7 @@ impl MultiFieldPostingsWriter {
     pub fn serialize(
         &self,
         serializer: &mut InvertedIndexSerializer,
+        fieldnorm_readers: FieldNormReaders,
     ) -> crate::Result<HashMap<Field, FnvHashMap<UnorderedTermId, TermOrdinal>>> {
         let mut term_offsets: Vec<(&[u8], Addr, UnorderedTermId)> =
             self.term_index.iter().collect();
@@ -161,8 +163,12 @@ impl MultiFieldPostingsWriter {
             }
 
             let postings_writer = &self.per_field_postings_writers[field.field_id() as usize];
-            let mut field_serializer =
-                serializer.new_field(field, postings_writer.total_num_tokens())?;
+            let fieldnorm_reader = fieldnorm_readers.get_field(field);
+            let mut field_serializer = serializer.new_field(
+                field,
+                postings_writer.total_num_tokens(),
+                fieldnorm_reader,
+            )?;
             postings_writer.serialize(
                 &term_offsets[start..stop],
                 &mut field_serializer,
@@ -297,7 +303,8 @@ impl<Rec: Recorder + 'static> PostingsWriter for SpecializedPostingsWriter<Rec>
         let mut buffer_lender = BufferLender::default();
         for &(term_bytes, addr, _) in term_addrs {
             let recorder: Rec = termdict_heap.read(addr);
-            serializer.new_term(&term_bytes[4..])?;
+            let term_doc_freq = recorder.term_doc_freq().unwrap_or(0u32);
+            serializer.new_term(&term_bytes[4..], term_doc_freq)?;
             recorder.serialize(&mut buffer_lender, serializer, heap)?;
             serializer.close_term()?;
         }

src/postings/recorder.rs

@@ -75,6 +75,8 @@ pub(crate) trait Recorder: Copy + 'static {
         serializer: &mut FieldSerializer<'_>,
         heap: &MemoryArena,
     ) -> io::Result<()>;
+    /// Returns the number of document containg this term.
+    fn term_doc_freq(&self) -> Option<u32>;
 }
 
 /// Only records the doc ids
@@ -113,11 +115,16 @@ impl Recorder for NothingRecorder {
     ) -> io::Result<()> {
         let buffer = buffer_lender.lend_u8();
         self.stack.read_to_end(heap, buffer);
+        // TODO avoid reading twice.
         for doc in VInt32Reader::new(&buffer[..]) {
             serializer.write_doc(doc as u32, 0u32, &[][..])?;
         }
         Ok(())
     }
+
+    fn term_doc_freq(&self) -> Option<u32> {
+        None
+    }
 }
 
 /// Recorder encoding document ids, and term frequencies
@@ -126,6 +133,7 @@ pub struct TermFrequencyRecorder {
     stack: ExpUnrolledLinkedList,
     current_doc: DocId,
     current_tf: u32,
+    term_doc_freq: u32,
 }
 
 impl Recorder for TermFrequencyRecorder {
@@ -134,6 +142,7 @@ impl Recorder for TermFrequencyRecorder {
             stack: ExpUnrolledLinkedList::new(),
             current_doc: u32::max_value(),
             current_tf: 0u32,
+            term_doc_freq: 0u32,
         }
     }
 
@@ -142,6 +151,7 @@ impl Recorder for TermFrequencyRecorder {
     }
 
     fn new_doc(&mut self, doc: DocId, heap: &mut MemoryArena) {
+        self.term_doc_freq += 1;
         self.current_doc = doc;
         let _ = write_u32_vint(doc, &mut self.stack.writer(heap));
     }
@@ -172,6 +182,10 @@ impl Recorder for TermFrequencyRecorder {
 
         Ok(())
     }
+
+    fn term_doc_freq(&self) -> Option<u32> {
+        Some(self.term_doc_freq)
+    }
 }
 
 /// Recorder encoding term frequencies as well as positions.
@@ -179,12 +193,14 @@ impl Recorder for TermFrequencyRecorder {
 pub struct TFAndPositionRecorder {
     stack: ExpUnrolledLinkedList,
     current_doc: DocId,
+    term_doc_freq: u32,
 }
 impl Recorder for TFAndPositionRecorder {
     fn new() -> Self {
         TFAndPositionRecorder {
             stack: ExpUnrolledLinkedList::new(),
             current_doc: u32::max_value(),
+            term_doc_freq: 0u32,
         }
     }
 
@@ -194,6 +210,7 @@ impl Recorder for TFAndPositionRecorder {
 
     fn new_doc(&mut self, doc: DocId, heap: &mut MemoryArena) {
         self.current_doc = doc;
+        self.term_doc_freq += 1u32;
         let _ = write_u32_vint(doc, &mut self.stack.writer(heap));
     }
 
@@ -233,6 +250,10 @@ impl Recorder for TFAndPositionRecorder {
         }
         Ok(())
     }
+
+    fn term_doc_freq(&self) -> Option<u32> {
+        Some(self.term_doc_freq)
+    }
 }
 
 #[cfg(test)]

src/postings/segment_postings.rs

@@ -1,54 +1,20 @@
 use crate::common::HasLen;
-use crate::common::{BinarySerializable, VInt};
-use crate::docset::{DocSet, TERMINATED};
+
+use crate::docset::DocSet;
 use crate::positions::PositionReader;
-use crate::postings::compression::{compressed_block_size, AlignedBuffer};
-use crate::postings::compression::{BlockDecoder, VIntDecoder, COMPRESSION_BLOCK_SIZE};
+
+use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
 use crate::postings::serializer::PostingsSerializer;
 use crate::postings::BlockSearcher;
-use crate::postings::FreqReadingOption;
+
 use crate::postings::Postings;
-use crate::postings::SkipReader;
-use crate::postings::USE_SKIP_INFO_LIMIT;
+
 use crate::schema::IndexRecordOption;
 use crate::DocId;
-use owned_read::OwnedRead;
-use tantivy_fst::Streamer;
 
-struct PositionComputer {
-    // store the amount of position int
-    // before reading positions.
-    //
-    // if none, position are already loaded in
-    // the positions vec.
-    position_to_skip: usize,
-    position_reader: PositionReader,
-}
-
-impl PositionComputer {
-    pub fn new(position_reader: PositionReader) -> PositionComputer {
-        PositionComputer {
-            position_to_skip: 0,
-            position_reader,
-        }
-    }
-
-    pub fn add_skip(&mut self, num_skip: usize) {
-        self.position_to_skip += num_skip;
-    }
-
-    // Positions can only be read once.
-    pub fn positions_with_offset(&mut self, offset: u32, output: &mut [u32]) {
-        self.position_reader.skip(self.position_to_skip);
-        self.position_to_skip = 0;
-        self.position_reader.read(output);
-        let mut cum = offset;
-        for output_mut in output.iter_mut() {
-            cum += *output_mut;
-            *output_mut = cum;
-        }
-    }
-}
+use crate::directory::ReadOnlySource;
+use crate::fieldnorm::FieldNormReader;
+use crate::postings::BlockSegmentPostings;
 
 /// `SegmentPostings` represents the inverted list or postings associated to
 /// a term in a `Segment`.
@@ -56,24 +22,27 @@ impl PositionComputer {
 /// As we iterate through the `SegmentPostings`, the frequencies are optionally decoded.
 /// Positions on the other hand, are optionally entirely decoded upfront.
 pub struct SegmentPostings {
-    block_cursor: BlockSegmentPostings,
+    pub(crate) block_cursor: BlockSegmentPostings,
     cur: usize,
-    position_computer: Option<PositionComputer>,
+    position_reader: Option<PositionReader>,
     block_searcher: BlockSearcher,
 }
 
 impl SegmentPostings {
     /// Returns an empty segment postings object
     pub fn empty() -> Self {
-        let empty_block_cursor = BlockSegmentPostings::empty();
-        let mut segment_postings = SegmentPostings {
-            block_cursor: empty_block_cursor,
-            cur: COMPRESSION_BLOCK_SIZE,
-            position_computer: None,
+        SegmentPostings {
+            block_cursor: BlockSegmentPostings::empty(),
+            cur: 0,
+            position_reader: None,
             block_searcher: BlockSearcher::default(),
-        };
-        segment_postings.advance();
-        segment_postings
+        }
+    }
+
+    /// Returns the overall number of documents in the block postings.
+    /// It does not take in account whether documents are deleted or not.
+    pub fn doc_freq(&self) -> u32 {
+        self.block_cursor.doc_freq()
     }
 
     /// Creates a segment postings object with the given documents
@@ -87,7 +56,8 @@ impl SegmentPostings {
     pub fn create_from_docs(docs: &[u32]) -> SegmentPostings {
         let mut buffer = Vec::new();
         {
-            let mut postings_serializer = PostingsSerializer::new(&mut buffer, false, false);
+            let mut postings_serializer = PostingsSerializer::new(&mut buffer, false, false, None);
+            postings_serializer.new_term(docs.len() as u32);
             for &doc in docs {
                 postings_serializer.write_doc(doc, 1u32);
             }
@@ -97,15 +67,36 @@ impl SegmentPostings {
         }
         let block_segment_postings = BlockSegmentPostings::from_data(
             docs.len() as u32,
-            OwnedRead::new(buffer),
+            ReadOnlySource::from(buffer),
             IndexRecordOption::Basic,
             IndexRecordOption::Basic,
         );
         SegmentPostings::from_block_postings(block_segment_postings, None)
     }
-}
 
-impl SegmentPostings {
+    /// Helper functions to create `SegmentPostings` for tests.
+    pub fn create_from_docs_and_tfs(
+        doc_and_tfs: &[(u32, u32)],
+        fieldnorm_reader: Option<FieldNormReader>,
+    ) -> crate::Result<SegmentPostings> {
+        let mut buffer = Vec::new();
+        let mut postings_serializer =
+            PostingsSerializer::new(&mut buffer, true, false, fieldnorm_reader);
+        postings_serializer.new_term(doc_and_tfs.len() as u32);
+        for &(doc, tf) in doc_and_tfs {
+            postings_serializer.write_doc(doc, tf);
+        }
+        postings_serializer
+            .close_term(doc_and_tfs.len() as u32)?;
+        let block_segment_postings = BlockSegmentPostings::from_data(
+            doc_and_tfs.len() as u32,
+            ReadOnlySource::from(buffer),
+            IndexRecordOption::WithFreqs,
+            IndexRecordOption::WithFreqs,
+        );
+        Ok(SegmentPostings::from_block_postings(block_segment_postings, None))
+    }
+
     /// Reads a Segment postings from an &[u8]
     ///
     /// * `len` - number of document in the posting lists.
@@ -114,16 +105,14 @@ impl SegmentPostings {
     ///   frequencies and/or positions
     pub(crate) fn from_block_postings(
         segment_block_postings: BlockSegmentPostings,
-        positions_stream_opt: Option<PositionReader>,
+        position_reader: Option<PositionReader>,
     ) -> SegmentPostings {
-        let mut postings = SegmentPostings {
+        SegmentPostings {
             block_cursor: segment_block_postings,
-            cur: COMPRESSION_BLOCK_SIZE, // cursor within the block
-            position_computer: positions_stream_opt.map(PositionComputer::new),
+            cur: 0, // cursor within the block
+            position_reader,
             block_searcher: BlockSearcher::default(),
-        };
-        postings.advance();
-        postings
+        }
     }
 }
 
@@ -132,91 +121,46 @@ impl DocSet for SegmentPostings {
     // next needs to be called a first time to point to the correct element.
     #[inline]
     fn advance(&mut self) -> DocId {
-        if self.position_computer.is_some() && self.cur < COMPRESSION_BLOCK_SIZE {
-            let term_freq = self.term_freq() as usize;
-            if let Some(position_computer) = self.position_computer.as_mut() {
-                position_computer.add_skip(term_freq);
-            }
-        }
-        self.cur += 1;
-        if self.cur >= self.block_cursor.block_len() {
-            if self.block_cursor.advance() {
-                self.cur = 0;
-            } else {
-                self.cur = COMPRESSION_BLOCK_SIZE - 1;
-                return TERMINATED;
-            }
+        assert!(self.block_cursor.block_is_loaded());
+        if self.cur == COMPRESSION_BLOCK_SIZE - 1 {
+            self.cur = 0;
+            self.block_cursor.advance();
+        } else {
+            self.cur += 1;
         }
         self.doc()
     }
 
     fn seek(&mut self, target: DocId) -> DocId {
+        debug_assert!(self.doc() <= target);
         if self.doc() >= target {
             return self.doc();
         }
 
-        // In the following, thanks to the call to advance above,
-        // we know that the position is not loaded and we need
-        // to skip every doc_freq we cross.
-
-        // skip blocks until one that might contain the target
-        // check if we need to go to the next block
-        let mut sum_freqs_skipped: u32 = 0;
-        if self
-            .block_cursor
-            .docs()
-            .last()
-            .map(|&doc| doc < target)
-            .unwrap_or(true)
-        {
-            // We are not in the right block.
-            if self.position_computer.is_some() {
-                // First compute all of the freqs skipped from the current block.
-                sum_freqs_skipped = self.block_cursor.freqs()[self.cur..].iter().sum::<u32>();
-                match self.block_cursor.skip_to(target) {
-                    BlockSegmentPostingsSkipResult::Success(block_skip_freqs) => {
-                        sum_freqs_skipped += block_skip_freqs;
-                    }
-                    BlockSegmentPostingsSkipResult::Terminated => {
-                        self.block_cursor.doc_decoder.clear();
-                        self.cur = 0;
-                        return TERMINATED;
-                    }
-                }
-            } else if self.block_cursor.skip_to(target)
-                == BlockSegmentPostingsSkipResult::Terminated
-            {
-                // no positions needed. no need to sum freqs.
-                self.block_cursor.doc_decoder.clear();
-                self.cur = 0;
-                return TERMINATED;
-            }
-            self.cur = 0;
-        }
+        self.block_cursor.seek(target);
 
         // At this point we are on the block, that might contain our document.
-
-        let cur = self.cur;
-
         let output = self.block_cursor.docs_aligned();
-        let new_cur = self.block_searcher.search_in_block(&output, cur, target);
-        if let Some(position_computer) = self.position_computer.as_mut() {
-            sum_freqs_skipped += self.block_cursor.freqs()[cur..new_cur].iter().sum::<u32>();
-            position_computer.add_skip(sum_freqs_skipped as usize);
-        }
-        self.cur = new_cur;
+        self.cur = self.block_searcher.search_in_block(&output, target);
+
+        // The last block is not full and padded with the value TERMINATED,
+        // so that we are guaranteed to have at least doc in the block (a real one or the padding)
+        // that is greater or equal to the target.
+        debug_assert!(self.cur < COMPRESSION_BLOCK_SIZE);
 
         // `doc` is now the first element >= `target`
-        let doc = output.0[new_cur];
+
+        // If all docs are smaller than target the current block should be incomplemented and padded
+        // with the value `TERMINATED`.
+        //
+        // After the search, the cursor should point to the first value of TERMINATED.
+        let doc = output.0[self.cur];
         debug_assert!(doc >= target);
+        debug_assert_eq!(doc, self.doc());
         doc
     }
 
     /// Return the current document's `DocId`.
-    ///
-    /// # Panics
-    ///
-    /// Will panics if called without having called advance before.
     #[inline]
     fn doc(&self) -> DocId {
         self.block_cursor.doc(self.cur)
@@ -225,28 +169,11 @@ impl DocSet for SegmentPostings {
     fn size_hint(&self) -> u32 {
         self.len() as u32
     }
-
-    /*
-    fn append_to_bitset(&mut self, bitset: &mut BitSet) {
-        // finish the current block
-        if self.advance() {
-            for &doc in &self.block_cursor.docs()[self.cur..] {
-                bitset.insert(doc);
-            }
-            // ... iterate through the remaining blocks.
-            while self.block_cursor.advance() {
-                for &doc in self.block_cursor.docs() {
-                    bitset.insert(doc);
-                }
-            }
-        }
-    }
-     */
 }
 
 impl HasLen for SegmentPostings {
     fn len(&self) -> usize {
-        self.block_cursor.doc_freq()
+        self.block_cursor.doc_freq() as usize
     }
 }
 
@@ -275,338 +202,33 @@ impl Postings for SegmentPostings {
 
     fn positions_with_offset(&mut self, offset: u32, output: &mut Vec<u32>) {
         let term_freq = self.term_freq() as usize;
-        if let Some(position_comp) = self.position_computer.as_mut() {
+        if let Some(position_reader) = self.position_reader.as_mut() {
+            let read_offset = self.block_cursor.position_offset()
+                + (self.block_cursor.freqs()[..self.cur]
+                    .iter()
+                    .cloned()
+                    .sum::<u32>() as u64);
             output.resize(term_freq, 0u32);
-            position_comp.positions_with_offset(offset, &mut output[..]);
+            position_reader.read(read_offset, &mut output[..]);
+            let mut cum = offset;
+            for output_mut in output.iter_mut() {
+                cum += *output_mut;
+                *output_mut = cum;
+            }
         } else {
             output.clear();
         }
     }
 }
 
-/// `BlockSegmentPostings` is a cursor iterating over blocks
-/// of documents.
-///
-/// # Warning
-///
-/// While it is useful for some very specific high-performance
-/// use cases, you should prefer using `SegmentPostings` for most usage.
-pub struct BlockSegmentPostings {
-    doc_decoder: BlockDecoder,
-    freq_decoder: BlockDecoder,
-    freq_reading_option: FreqReadingOption,
-
-    doc_freq: usize,
-    doc_offset: DocId,
-
-    num_vint_docs: usize,
-
-    remaining_data: OwnedRead,
-    skip_reader: SkipReader,
-}
-
-fn split_into_skips_and_postings(
-    doc_freq: u32,
-    mut data: OwnedRead,
-) -> (Option<OwnedRead>, OwnedRead) {
-    if doc_freq < USE_SKIP_INFO_LIMIT {
-        return (None, data);
-    }
-    let skip_len = VInt::deserialize(&mut data).expect("Data corrupted").0 as usize;
-    let mut postings_data = data.clone();
-    postings_data.advance(skip_len);
-    data.clip(skip_len);
-    (Some(data), postings_data)
-}
-
-#[derive(Debug, Eq, PartialEq)]
-pub enum BlockSegmentPostingsSkipResult {
-    Terminated,
-    Success(u32), //< number of term freqs to skip
-}
-
-impl BlockSegmentPostings {
-    pub(crate) fn from_data(
-        doc_freq: u32,
-        data: OwnedRead,
-        record_option: IndexRecordOption,
-        requested_option: IndexRecordOption,
-    ) -> BlockSegmentPostings {
-        let freq_reading_option = match (record_option, requested_option) {
-            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
-            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
-            (_, _) => FreqReadingOption::ReadFreq,
-        };
-
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
-        let skip_reader = match skip_data_opt {
-            Some(skip_data) => SkipReader::new(skip_data, record_option),
-            None => SkipReader::new(OwnedRead::new(&[][..]), record_option),
-        };
-        let doc_freq = doc_freq as usize;
-        let num_vint_docs = doc_freq % COMPRESSION_BLOCK_SIZE;
-        BlockSegmentPostings {
-            num_vint_docs,
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option,
-            doc_offset: 0,
-            doc_freq,
-            remaining_data: postings_data,
-            skip_reader,
-        }
-    }
-
-    // Resets the block segment postings on another position
-    // in the postings file.
-    //
-    // This is useful for enumerating through a list of terms,
-    // and consuming the associated posting lists while avoiding
-    // reallocating a `BlockSegmentPostings`.
-    //
-    // # Warning
-    //
-    // This does not reset the positions list.
-    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: OwnedRead) {
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
-        let num_vint_docs = (doc_freq as usize) & (COMPRESSION_BLOCK_SIZE - 1);
-        self.num_vint_docs = num_vint_docs;
-        self.remaining_data = postings_data;
-        if let Some(skip_data) = skip_data_opt {
-            self.skip_reader.reset(skip_data);
-        } else {
-            self.skip_reader.reset(OwnedRead::new(&[][..]))
-        }
-        self.doc_offset = 0;
-        self.doc_freq = doc_freq as usize;
-    }
-
-    /// Returns the document frequency associated to this block postings.
-    ///
-    /// This `doc_freq` is simply the sum of the length of all of the blocks
-    /// length, and it does not take in account deleted documents.
-    pub fn doc_freq(&self) -> usize {
-        self.doc_freq
-    }
-
-    /// Returns the array of docs in the current block.
-    ///
-    /// Before the first call to `.advance()`, the block
-    /// returned by `.docs()` is empty.
-    #[inline]
-    pub fn docs(&self) -> &[DocId] {
-        self.doc_decoder.output_array()
-    }
-
-    pub(crate) fn docs_aligned(&self) -> &AlignedBuffer {
-        self.doc_decoder.output_aligned()
-    }
-
-    /// Return the document at index `idx` of the block.
-    #[inline]
-    pub fn doc(&self, idx: usize) -> u32 {
-        self.doc_decoder.output(idx)
-    }
-
-    /// Return the array of `term freq` in the block.
-    #[inline]
-    pub fn freqs(&self) -> &[u32] {
-        self.freq_decoder.output_array()
-    }
-
-    /// Return the frequency at index `idx` of the block.
-    #[inline]
-    pub fn freq(&self, idx: usize) -> u32 {
-        self.freq_decoder.output(idx)
-    }
-
-    /// Returns the length of the current block.
-    ///
-    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
-    /// except the last block that may have a length
-    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
-    #[inline]
-    fn block_len(&self) -> usize {
-        self.doc_decoder.output_len
-    }
-
-    /// position on a block that may contains `doc_id`.
-    /// Always advance the current block.
-    ///
-    /// Returns true if a block that has an element greater or equal to the target is found.
-    /// Returning true does not guarantee that the smallest element of the block is smaller
-    /// than the target. It only guarantees that the last element is greater or equal.
-    ///
-    /// Returns false iff all of the document remaining are smaller than
-    /// `doc_id`. In that case, all of these document are consumed.
-    ///
-    pub fn skip_to(&mut self, target_doc: DocId) -> BlockSegmentPostingsSkipResult {
-        let mut skip_freqs = 0u32;
-        while self.skip_reader.advance() {
-            if self.skip_reader.doc() >= target_doc {
-                // the last document of the current block is larger
-                // than the target.
-                //
-                // We found our block!
-                let num_bits = self.skip_reader.doc_num_bits();
-                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                    self.remaining_data.as_ref(),
-                    self.doc_offset,
-                    num_bits,
-                );
-                self.remaining_data.advance(num_consumed_bytes);
-                let tf_num_bits = self.skip_reader.tf_num_bits();
-                match self.freq_reading_option {
-                    FreqReadingOption::NoFreq => {}
-                    FreqReadingOption::SkipFreq => {
-                        let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                        self.remaining_data.advance(num_bytes_to_skip);
-                    }
-                    FreqReadingOption::ReadFreq => {
-                        let num_consumed_bytes = self
-                            .freq_decoder
-                            .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                        self.remaining_data.advance(num_consumed_bytes);
-                    }
-                }
-                self.doc_offset = self.skip_reader.doc();
-                return BlockSegmentPostingsSkipResult::Success(skip_freqs);
-            } else {
-                skip_freqs += self.skip_reader.tf_sum();
-                let advance_len = self.skip_reader.total_block_len();
-                self.doc_offset = self.skip_reader.doc();
-                self.remaining_data.advance(advance_len);
-            }
-        }
-
-        self.doc_decoder.clear();
-
-        if self.num_vint_docs == 0 {
-            return BlockSegmentPostingsSkipResult::Terminated;
-        }
-        // we are now on the last, incomplete, variable encoded block.
-        let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-            self.remaining_data.as_ref(),
-            self.doc_offset,
-            self.num_vint_docs,
-        );
-        self.remaining_data.advance(num_compressed_bytes);
-        match self.freq_reading_option {
-            FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-            FreqReadingOption::ReadFreq => {
-                self.freq_decoder
-                    .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-            }
-        }
-        self.num_vint_docs = 0;
-        self.docs()
-            .last()
-            .map(|last_doc| {
-                if *last_doc >= target_doc {
-                    BlockSegmentPostingsSkipResult::Success(skip_freqs)
-                } else {
-                    BlockSegmentPostingsSkipResult::Terminated
-                }
-            })
-            .unwrap_or(BlockSegmentPostingsSkipResult::Terminated)
-    }
-
-    /// Advance to the next block.
-    ///
-    /// Returns false iff there was no remaining blocks.
-    pub fn advance(&mut self) -> bool {
-        if self.skip_reader.advance() {
-            let num_bits = self.skip_reader.doc_num_bits();
-            let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                num_bits,
-            );
-            self.remaining_data.advance(num_consumed_bytes);
-            let tf_num_bits = self.skip_reader.tf_num_bits();
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq => {}
-                FreqReadingOption::SkipFreq => {
-                    let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                    self.remaining_data.advance(num_bytes_to_skip);
-                }
-                FreqReadingOption::ReadFreq => {
-                    let num_consumed_bytes = self
-                        .freq_decoder
-                        .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                    self.remaining_data.advance(num_consumed_bytes);
-                }
-            }
-            // it will be used as the next offset.
-            self.doc_offset = self.doc_decoder.output(COMPRESSION_BLOCK_SIZE - 1);
-            return true;
-        }
-        self.doc_decoder.clear();
-        if self.num_vint_docs == 0 {
-            return false;
-        }
-        let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-            self.remaining_data.as_ref(),
-            self.doc_offset,
-            self.num_vint_docs,
-        );
-        self.remaining_data.advance(num_compressed_bytes);
-        match self.freq_reading_option {
-            FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-            FreqReadingOption::ReadFreq => {
-                self.freq_decoder
-                    .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-            }
-        }
-        self.num_vint_docs = 0;
-        true
-    }
-
-    /// Returns an empty segment postings object
-    pub fn empty() -> BlockSegmentPostings {
-        BlockSegmentPostings {
-            num_vint_docs: 0,
-
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option: FreqReadingOption::NoFreq,
-
-            doc_offset: 0,
-            doc_freq: 0,
-
-            remaining_data: OwnedRead::new(vec![]),
-            skip_reader: SkipReader::new(OwnedRead::new(vec![]), IndexRecordOption::Basic),
-        }
-    }
-}
-
-impl<'b> Streamer<'b> for BlockSegmentPostings {
-    type Item = &'b [DocId];
-
-    fn next(&'b mut self) -> Option<&'b [DocId]> {
-        if self.advance() {
-            Some(self.docs())
-        } else {
-            None
-        }
-    }
-}
-
 #[cfg(test)]
 mod tests {
-    use super::BlockSegmentPostings;
-    use super::BlockSegmentPostingsSkipResult;
+
     use super::SegmentPostings;
     use crate::common::HasLen;
-    use crate::core::Index;
+
     use crate::docset::{DocSet, TERMINATED};
     use crate::postings::postings::Postings;
-    use crate::schema::IndexRecordOption;
-    use crate::schema::Schema;
-    use crate::schema::Term;
-    use crate::schema::INDEXED;
-    use crate::DocId;
-    use tantivy_fst::Streamer;
 
     #[test]
     fn test_empty_segment_postings() {
@@ -628,165 +250,4 @@ mod tests {
         let postings = SegmentPostings::empty();
         assert_eq!(postings.term_freq(), 1);
     }
-
-    #[test]
-    fn test_empty_block_segment_postings() {
-        let mut postings = BlockSegmentPostings::empty();
-        assert!(!postings.advance());
-        assert_eq!(postings.doc_freq(), 0);
-    }
-
-    #[test]
-    fn test_block_segment_postings() {
-        let mut block_segments = build_block_postings(&(0..100_000).collect::<Vec<u32>>());
-        let mut offset: u32 = 0u32;
-        // checking that the block before calling advance is empty
-        assert!(block_segments.docs().is_empty());
-        // checking that the `doc_freq` is correct
-        assert_eq!(block_segments.doc_freq(), 100_000);
-        while let Some(block) = block_segments.next() {
-            for (i, doc) in block.iter().cloned().enumerate() {
-                assert_eq!(offset + (i as u32), doc);
-            }
-            offset += block.len() as u32;
-        }
-    }
-
-    #[test]
-    fn test_skip_right_at_new_block() {
-        let mut doc_ids = (0..128).collect::<Vec<u32>>();
-        doc_ids.push(129);
-        doc_ids.push(130);
-        {
-            let block_segments = build_block_postings(&doc_ids);
-            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
-            assert_eq!(docset.seek(128), 129);
-            assert_eq!(docset.doc(), 129);
-            assert_eq!(docset.advance(), 130);
-            assert_eq!(docset.doc(), 130);
-            assert_eq!(docset.advance(), TERMINATED);
-        }
-        {
-            let block_segments = build_block_postings(&doc_ids);
-            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
-            assert_eq!(docset.seek(129), 129);
-            assert_eq!(docset.doc(), 129);
-            assert_eq!(docset.advance(), 130);
-            assert_eq!(docset.doc(), 130);
-            assert_eq!(docset.advance(), TERMINATED);
-        }
-        {
-            let block_segments = build_block_postings(&doc_ids);
-            let mut docset = SegmentPostings::from_block_postings(block_segments, None);
-            assert_eq!(docset.doc(), 0);
-            assert_eq!(docset.seek(131), TERMINATED);
-            assert_eq!(docset.doc(), TERMINATED);
-        }
-    }
-
-    fn build_block_postings(docs: &[DocId]) -> BlockSegmentPostings {
-        let mut schema_builder = Schema::builder();
-        let int_field = schema_builder.add_u64_field("id", INDEXED);
-        let schema = schema_builder.build();
-        let index = Index::create_in_ram(schema);
-        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
-        let mut last_doc = 0u32;
-        for &doc in docs {
-            for _ in last_doc..doc {
-                index_writer.add_document(doc!(int_field=>1u64));
-            }
-            index_writer.add_document(doc!(int_field=>0u64));
-            last_doc = doc + 1;
-        }
-        index_writer.commit().unwrap();
-        let searcher = index.reader().unwrap().searcher();
-        let segment_reader = searcher.segment_reader(0);
-        let inverted_index = segment_reader.inverted_index(int_field);
-        let term = Term::from_field_u64(int_field, 0u64);
-        let term_info = inverted_index.get_term_info(&term).unwrap();
-        inverted_index.read_block_postings_from_terminfo(&term_info, IndexRecordOption::Basic)
-    }
-
-    #[test]
-    fn test_block_segment_postings_skip() {
-        for i in 0..4 {
-            let mut block_postings = build_block_postings(&[3]);
-            assert_eq!(
-                block_postings.skip_to(i),
-                BlockSegmentPostingsSkipResult::Success(0u32)
-            );
-            assert_eq!(
-                block_postings.skip_to(i),
-                BlockSegmentPostingsSkipResult::Terminated
-            );
-        }
-        let mut block_postings = build_block_postings(&[3]);
-        assert_eq!(
-            block_postings.skip_to(4u32),
-            BlockSegmentPostingsSkipResult::Terminated
-        );
-    }
-
-    #[test]
-    fn test_block_segment_postings_skip2() {
-        let mut docs = vec![0];
-        for i in 0..1300 {
-            docs.push((i * i / 100) + i);
-        }
-        let mut block_postings = build_block_postings(&docs[..]);
-        for i in vec![0, 424, 10000] {
-            assert_eq!(
-                block_postings.skip_to(i),
-                BlockSegmentPostingsSkipResult::Success(0u32)
-            );
-            let docs = block_postings.docs();
-            assert!(docs[0] <= i);
-            assert!(docs.last().cloned().unwrap_or(0u32) >= i);
-        }
-        assert_eq!(
-            block_postings.skip_to(100_000),
-            BlockSegmentPostingsSkipResult::Terminated
-        );
-        assert_eq!(
-            block_postings.skip_to(101_000),
-            BlockSegmentPostingsSkipResult::Terminated
-        );
-    }
-
-    #[test]
-    fn test_reset_block_segment_postings() {
-        let mut schema_builder = Schema::builder();
-        let int_field = schema_builder.add_u64_field("id", INDEXED);
-        let schema = schema_builder.build();
-        let index = Index::create_in_ram(schema);
-        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
-        // create two postings list, one containg even number,
-        // the other containing odd numbers.
-        for i in 0..6 {
-            let doc = doc!(int_field=> (i % 2) as u64);
-            index_writer.add_document(doc);
-        }
-        index_writer.commit().unwrap();
-        let searcher = index.reader().unwrap().searcher();
-        let segment_reader = searcher.segment_reader(0);
-
-        let mut block_segments;
-        {
-            let term = Term::from_field_u64(int_field, 0u64);
-            let inverted_index = segment_reader.inverted_index(int_field);
-            let term_info = inverted_index.get_term_info(&term).unwrap();
-            block_segments = inverted_index
-                .read_block_postings_from_terminfo(&term_info, IndexRecordOption::Basic);
-        }
-        assert!(block_segments.advance());
-        assert_eq!(block_segments.docs(), &[0, 2, 4]);
-        {
-            let term = Term::from_field_u64(int_field, 1u64);
-            let inverted_index = segment_reader.inverted_index(int_field);
-            let term_info = inverted_index.get_term_info(&term).unwrap();
-            inverted_index.reset_block_postings_from_terminfo(&term_info, &mut block_segments);
-        }
-        assert!(block_segments.advance());
-        assert_eq!(block_segments.docs(), &[1, 3, 5]);
-    }
 }

src/postings/serializer.rs

@@ -3,14 +3,16 @@ use crate::common::{BinarySerializable, VInt};
 use crate::common::{CompositeWrite, CountingWriter};
 use crate::core::Segment;
 use crate::directory::WritePtr;
+use crate::fieldnorm::FieldNormReader;
 use crate::positions::PositionSerializer;
 use crate::postings::compression::{BlockEncoder, VIntEncoder, COMPRESSION_BLOCK_SIZE};
 use crate::postings::skip::SkipSerializer;
-use crate::postings::USE_SKIP_INFO_LIMIT;
+use crate::query::BM25Weight;
 use crate::schema::Schema;
 use crate::schema::{Field, FieldEntry, FieldType};
 use crate::termdict::{TermDictionaryBuilder, TermOrdinal};
 use crate::DocId;
+use std::cmp::Ordering;
 use std::io::{self, Write};
 
 /// `InvertedIndexSerializer` is in charge of serializing
@@ -90,6 +92,7 @@ impl InvertedIndexSerializer {
         &mut self,
         field: Field,
         total_num_tokens: u64,
+        fieldnorm_reader: Option<FieldNormReader>,
     ) -> io::Result<FieldSerializer<'_>> {
         let field_entry: &FieldEntry = self.schema.get_field_entry(field);
         let term_dictionary_write = self.terms_write.for_field(field);
@@ -104,6 +107,7 @@ impl InvertedIndexSerializer {
             postings_write,
             positions_write,
             positionsidx_write,
+            fieldnorm_reader,
         )
     }
 
@@ -135,6 +139,7 @@ impl<'a> FieldSerializer<'a> {
         postings_write: &'a mut CountingWriter<WritePtr>,
         positions_write: &'a mut CountingWriter<WritePtr>,
         positionsidx_write: &'a mut CountingWriter<WritePtr>,
+        fieldnorm_reader: Option<FieldNormReader>,
     ) -> io::Result<FieldSerializer<'a>> {
         let (term_freq_enabled, position_enabled): (bool, bool) = match field_type {
             FieldType::Str(ref text_options) => {
@@ -148,8 +153,12 @@ impl<'a> FieldSerializer<'a> {
             _ => (false, false),
         };
         let term_dictionary_builder = TermDictionaryBuilder::create(term_dictionary_write)?;
-        let postings_serializer =
-            PostingsSerializer::new(postings_write, term_freq_enabled, position_enabled);
+        let postings_serializer = PostingsSerializer::new(
+            postings_write,
+            term_freq_enabled,
+            position_enabled,
+            fieldnorm_reader,
+        );
         let positions_serializer_opt = if position_enabled {
             Some(PositionSerializer::new(positions_write, positionsidx_write))
         } else {
@@ -182,8 +191,8 @@ impl<'a> FieldSerializer<'a> {
     /// Starts the postings for a new term.
     /// * term - the term. It needs to come after the previous term according
     ///   to the lexicographical order.
-    /// * doc_freq - return the number of document containing the term.
-    pub fn new_term(&mut self, term: &[u8]) -> io::Result<TermOrdinal> {
+    /// * term_doc_freq - return the number of document containing the term.
+    pub fn new_term(&mut self, term: &[u8], term_doc_freq: u32) -> io::Result<TermOrdinal> {
         assert!(
             !self.term_open,
             "Called new_term, while the previous term was not closed."
@@ -194,6 +203,7 @@ impl<'a> FieldSerializer<'a> {
         self.term_dictionary_builder.insert_key(term)?;
         let term_ordinal = self.num_terms;
         self.num_terms += 1;
+        self.postings_serializer.new_term(term_doc_freq);
         Ok(term_ordinal)
     }
 
@@ -307,6 +317,21 @@ pub struct PostingsSerializer<W: Write> {
 
     termfreq_enabled: bool,
     termfreq_sum_enabled: bool,
+    fieldnorm_reader: Option<FieldNormReader>,
+
+    bm25_weight: Option<BM25Weight>,
+
+    num_docs: u32, // Number of docs in the segment
+    avg_fieldnorm: f32, // Average number of term in the field for that segment.
+                   // this value is used to compute the block wand information.
+}
+
+fn get_avg_fieldnorm(fieldnorm_reader: &FieldNormReader) -> f32 {
+    let num_docs = fieldnorm_reader.num_docs();
+    let sum_fieldnorm: f32 = (0u32..num_docs)
+        .map(|doc| fieldnorm_reader.fieldnorm(doc) as f32)
+        .sum();
+    sum_fieldnorm / (num_docs as f32)
 }
 
 impl<W: Write> PostingsSerializer<W> {
@@ -314,7 +339,16 @@ impl<W: Write> PostingsSerializer<W> {
         write: W,
         termfreq_enabled: bool,
         termfreq_sum_enabled: bool,
+        fieldnorm_reader: Option<FieldNormReader>,
     ) -> PostingsSerializer<W> {
+        let avg_fieldnorm: f32 = fieldnorm_reader
+            .as_ref()
+            .map(get_avg_fieldnorm)
+            .unwrap_or(0f32);
+        let num_docs = fieldnorm_reader
+            .as_ref()
+            .map(|fieldnorm_reader| fieldnorm_reader.num_docs())
+            .unwrap_or(0u32);
         PostingsSerializer {
             output_write: CountingWriter::wrap(write),
 
@@ -327,6 +361,23 @@ impl<W: Write> PostingsSerializer<W> {
             last_doc_id_encoded: 0u32,
             termfreq_enabled,
             termfreq_sum_enabled,
+
+            fieldnorm_reader,
+            bm25_weight: None,
+
+            num_docs,
+            avg_fieldnorm,
+        }
+    }
+
+    pub fn new_term(&mut self, term_doc_freq: u32) {
+        if self.termfreq_enabled && self.num_docs > 0 {
+            let bm25_weight = BM25Weight::for_one_term(
+                term_doc_freq as u64,
+                self.num_docs as u64,
+                self.avg_fieldnorm,
+            );
+            self.bm25_weight = Some(bm25_weight);
         }
     }
 
@@ -343,7 +394,6 @@ impl<W: Write> PostingsSerializer<W> {
             self.postings_write.extend(block_encoded);
         }
         if self.termfreq_enabled {
-            // encode the term_freqs
             let (num_bits, block_encoded): (u8, &[u8]) = self
                 .block_encoder
                 .compress_block_unsorted(&self.block.term_freqs());
@@ -353,6 +403,32 @@ impl<W: Write> PostingsSerializer<W> {
                 let sum_freq = self.block.term_freqs().iter().cloned().sum();
                 self.skip_write.write_total_term_freq(sum_freq);
             }
+            let mut blockwand_params_opt = None;
+            if let Some(bm25_weight) = self.bm25_weight.as_ref() {
+                if let Some(fieldnorm_reader) = self.fieldnorm_reader.as_ref() {
+                    let docs = self.block.doc_ids();
+                    let term_freqs = self.block.term_freqs();
+                    blockwand_params_opt = docs
+                        .iter()
+                        .cloned()
+                        .map(|doc| fieldnorm_reader.fieldnorm_id(doc))
+                        .zip(term_freqs.iter().cloned())
+                        .max_by(
+                            |(left_fieldnorm_id, left_term_freq),
+                             (right_fieldnorm_id, right_term_freq)| {
+                                let left_score =
+                                    bm25_weight.tf_factor(*left_fieldnorm_id, *left_term_freq);
+                                let right_score =
+                                    bm25_weight.tf_factor(*right_fieldnorm_id, *right_term_freq);
+                                left_score
+                                    .partial_cmp(&right_score)
+                                    .unwrap_or(Ordering::Equal)
+                            },
+                        );
+                }
+            }
+            let (fieldnorm_id, term_freq) = blockwand_params_opt.unwrap_or((0u8, 0u32));
+            self.skip_write.write_blockwand_max(fieldnorm_id, term_freq);
         }
         self.block.clear();
     }
@@ -391,7 +467,7 @@ impl<W: Write> PostingsSerializer<W> {
             }
             self.block.clear();
         }
-        if doc_freq >= USE_SKIP_INFO_LIMIT {
+        if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
             let skip_data = self.skip_write.data();
             VInt(skip_data.len() as u64).serialize(&mut self.output_write)?;
             self.output_write.write_all(skip_data)?;
@@ -401,6 +477,7 @@ impl<W: Write> PostingsSerializer<W> {
         }
         self.skip_write.clear();
         self.postings_write.clear();
+        self.bm25_weight = None;
         Ok(())
     }
 

src/postings/skip.rs

@@ -1,7 +1,9 @@
-use crate::common::BinarySerializable;
-use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+use crate::common::{BinarySerializable, VInt};
+use crate::directory::ReadOnlySource;
+use crate::postings::compression::{compressed_block_size, COMPRESSION_BLOCK_SIZE};
+use crate::query::BM25Weight;
 use crate::schema::IndexRecordOption;
-use crate::DocId;
+use crate::{DocId, Score, TERMINATED};
 use owned_read::OwnedRead;
 
 pub struct SkipSerializer {
@@ -39,6 +41,11 @@ impl SkipSerializer {
             .expect("Should never fail");
     }
 
+    pub fn write_blockwand_max(&mut self, fieldnorm_id: u8, term_freq: u32) {
+        self.buffer.push(fieldnorm_id);
+        VInt(term_freq as u64).serialize_into_vec(&mut self.buffer);
+    }
+
     pub fn data(&self) -> &[u8] {
         &self.buffer[..]
     }
@@ -50,80 +57,186 @@ impl SkipSerializer {
 }
 
 pub(crate) struct SkipReader {
-    doc: DocId,
+    last_doc_in_block: DocId,
+    pub(crate) last_doc_in_previous_block: DocId,
     owned_read: OwnedRead,
-    doc_num_bits: u8,
-    tf_num_bits: u8,
-    tf_sum: u32,
     skip_info: IndexRecordOption,
+    byte_offset: usize,
+    remaining_docs: u32, // number of docs remaining, including the
+    // documents in the current block.
+    block_info: BlockInfo,
+
+    position_offset: u64,
+}
+
+#[derive(Clone, Copy, Debug)]
+pub(crate) enum BlockInfo {
+    BitPacked {
+        doc_num_bits: u8,
+        tf_num_bits: u8,
+        tf_sum: u32,
+        block_wand_fieldnorm_id: u8,
+        block_wand_term_freq: u32,
+    },
+    VInt {
+        num_docs: u32,
+    },
+}
+
+impl Default for BlockInfo {
+    fn default() -> Self {
+        BlockInfo::VInt { num_docs: 0u32 }
+    }
 }
 
 impl SkipReader {
-    pub fn new(data: OwnedRead, skip_info: IndexRecordOption) -> SkipReader {
-        SkipReader {
-            doc: 0u32,
-            owned_read: data,
+    pub fn new(data: ReadOnlySource, doc_freq: u32, skip_info: IndexRecordOption) -> SkipReader {
+        let mut skip_reader = SkipReader {
+            last_doc_in_block: if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
+                0
+            } else {
+                TERMINATED
+            },
+            last_doc_in_previous_block: 0u32,
+            owned_read: OwnedRead::new(data),
             skip_info,
-            doc_num_bits: 0u8,
-            tf_num_bits: 0u8,
-            tf_sum: 0u32,
+            block_info: BlockInfo::VInt { num_docs: doc_freq },
+            byte_offset: 0,
+            remaining_docs: doc_freq,
+            position_offset: 0u64,
+        };
+        if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
+            skip_reader.read_block_info();
+        }
+        skip_reader
+    }
+
+    pub fn reset(&mut self, data: ReadOnlySource, doc_freq: u32) {
+        self.last_doc_in_block = if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
+            0
+        } else {
+            TERMINATED
+        };
+        self.last_doc_in_previous_block = 0u32;
+        self.owned_read = OwnedRead::new(data);
+        self.block_info = BlockInfo::VInt { num_docs: doc_freq };
+        self.byte_offset = 0;
+        self.remaining_docs = doc_freq;
+        self.position_offset = 0u64;
+        if doc_freq >= COMPRESSION_BLOCK_SIZE as u32 {
+            self.read_block_info();
         }
     }
 
-    pub fn reset(&mut self, data: OwnedRead) {
-        self.doc = 0u32;
-        self.owned_read = data;
-        self.doc_num_bits = 0u8;
-        self.tf_num_bits = 0u8;
-        self.tf_sum = 0u32;
+    pub fn block_max_score(&self, bm25_weight: &BM25Weight) -> Option<Score> {
+        match self.block_info {
+            BlockInfo::BitPacked {
+                block_wand_fieldnorm_id,
+                block_wand_term_freq,
+                ..
+            } => Some(bm25_weight.score(block_wand_fieldnorm_id, block_wand_term_freq)),
+            BlockInfo::VInt { .. } => None,
+        }
     }
 
-    pub fn total_block_len(&self) -> usize {
-        (self.doc_num_bits + self.tf_num_bits) as usize * COMPRESSION_BLOCK_SIZE / 8
+    pub(crate) fn last_doc_in_block(&self) -> DocId {
+        self.last_doc_in_block
     }
 
-    pub fn doc(&self) -> DocId {
-        self.doc
+    pub fn position_offset(&self) -> u64 {
+        self.position_offset
     }
 
-    pub fn doc_num_bits(&self) -> u8 {
-        self.doc_num_bits
+    pub fn byte_offset(&self) -> usize {
+        self.byte_offset
     }
 
-    /// Number of bits used to encode term frequencies
-    ///
-    /// 0 if term frequencies are not enabled.
-    pub fn tf_num_bits(&self) -> u8 {
-        self.tf_num_bits
-    }
-
-    pub fn tf_sum(&self) -> u32 {
-        self.tf_sum
-    }
-
-    pub fn advance(&mut self) -> bool {
-        if self.owned_read.as_ref().is_empty() {
-            false
-        } else {
-            let doc_delta = u32::deserialize(&mut self.owned_read).expect("Skip data corrupted");
-            self.doc += doc_delta as DocId;
-            self.doc_num_bits = self.owned_read.get(0);
-            match self.skip_info {
-                IndexRecordOption::Basic => {
-                    self.owned_read.advance(1);
-                }
-                IndexRecordOption::WithFreqs => {
-                    self.tf_num_bits = self.owned_read.get(1);
-                    self.owned_read.advance(2);
-                }
-                IndexRecordOption::WithFreqsAndPositions => {
-                    self.tf_num_bits = self.owned_read.get(1);
-                    self.owned_read.advance(2);
-                    self.tf_sum =
-                        u32::deserialize(&mut self.owned_read).expect("Failed reading tf_sum");
-                }
+    fn read_block_info(&mut self) {
+        let doc_delta = u32::deserialize(&mut self.owned_read).expect("Skip data corrupted");
+        self.last_doc_in_block += doc_delta as DocId;
+        let doc_num_bits = self.owned_read.get(0);
+        match self.skip_info {
+            IndexRecordOption::Basic => {
+                self.owned_read.advance(1);
+                self.block_info = BlockInfo::BitPacked {
+                    doc_num_bits,
+                    tf_num_bits: 0,
+                    tf_sum: 0,
+                    block_wand_fieldnorm_id: 0,
+                    block_wand_term_freq: 0,
+                };
             }
-            true
+            IndexRecordOption::WithFreqs => {
+                let tf_num_bits = self.owned_read.get(1);
+                let block_wand_fieldnorm_id = self.owned_read.get(2);
+                self.owned_read.advance(3);
+                let block_wand_term_freq =
+                    VInt::deserialize_u64(&mut self.owned_read).unwrap() as u32;
+                self.block_info = BlockInfo::BitPacked {
+                    doc_num_bits,
+                    tf_num_bits,
+                    tf_sum: 0,
+                    block_wand_fieldnorm_id,
+                    block_wand_term_freq,
+                };
+            }
+            IndexRecordOption::WithFreqsAndPositions => {
+                let tf_num_bits = self.owned_read.get(1);
+                self.owned_read.advance(2);
+                let tf_sum = u32::deserialize(&mut self.owned_read).expect("Failed reading tf_sum");
+                let block_wand_fieldnorm_id = self.owned_read.get(0);
+                self.owned_read.advance(1);
+                let block_wand_term_freq =
+                    VInt::deserialize_u64(&mut self.owned_read).unwrap() as u32;
+                self.block_info = BlockInfo::BitPacked {
+                    doc_num_bits,
+                    tf_num_bits,
+                    tf_sum,
+                    block_wand_fieldnorm_id,
+                    block_wand_term_freq,
+                };
+            }
+        }
+    }
+
+    pub fn block_info(&self) -> BlockInfo {
+        self.block_info
+    }
+
+    /// Advance the skip reader to the block that may contain the target.
+    ///
+    /// If the target is larger than all documents, the skip_reader
+    /// then advance to the last Variable In block.
+    pub fn seek(&mut self, target: DocId) {
+        while self.last_doc_in_block() < target {
+            self.advance();
+        }
+    }
+
+    pub fn advance(&mut self) {
+        match self.block_info {
+            BlockInfo::BitPacked {
+                doc_num_bits,
+                tf_num_bits,
+                tf_sum,
+                ..
+            } => {
+                self.remaining_docs -= COMPRESSION_BLOCK_SIZE as u32;
+                self.byte_offset += compressed_block_size(doc_num_bits + tf_num_bits);
+                self.position_offset += tf_sum as u64;
+            }
+            BlockInfo::VInt { num_docs} => {
+                debug_assert_eq!(num_docs, self.remaining_docs);
+                self.remaining_docs = 0;
+                self.byte_offset = std::usize::MAX;
+            }
+        }
+        self.last_doc_in_previous_block = self.last_doc_in_block;
+        if self.remaining_docs >= COMPRESSION_BLOCK_SIZE as u32 {
+            self.read_block_info();
+        } else {
+            self.last_doc_in_block = TERMINATED;
+            self.block_info = BlockInfo::VInt { num_docs: self.remaining_docs };
         }
     }
 }
@@ -131,9 +244,11 @@ impl SkipReader {
 #[cfg(test)]
 mod tests {
 
+    use super::BlockInfo;
     use super::IndexRecordOption;
     use super::{SkipReader, SkipSerializer};
-    use owned_read::OwnedRead;
+    use crate::directory::ReadOnlySource;
+    use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
 
     #[test]
     fn test_skip_with_freq() {
@@ -141,20 +256,47 @@ mod tests {
             let mut skip_serializer = SkipSerializer::new();
             skip_serializer.write_doc(1u32, 2u8);
             skip_serializer.write_term_freq(3u8);
+            skip_serializer.write_blockwand_max(13u8, 3u32);
             skip_serializer.write_doc(5u32, 5u8);
             skip_serializer.write_term_freq(2u8);
+            skip_serializer.write_blockwand_max(8u8, 2u32);
             skip_serializer.data().to_owned()
         };
-        let mut skip_reader = SkipReader::new(OwnedRead::new(buf), IndexRecordOption::WithFreqs);
-        assert!(skip_reader.advance());
-        assert_eq!(skip_reader.doc(), 1u32);
-        assert_eq!(skip_reader.doc_num_bits(), 2u8);
-        assert_eq!(skip_reader.tf_num_bits(), 3u8);
-        assert!(skip_reader.advance());
-        assert_eq!(skip_reader.doc(), 5u32);
-        assert_eq!(skip_reader.doc_num_bits(), 5u8);
-        assert_eq!(skip_reader.tf_num_bits(), 2u8);
-        assert!(!skip_reader.advance());
+        let doc_freq = 3u32 + (COMPRESSION_BLOCK_SIZE * 2) as u32;
+        let mut skip_reader = SkipReader::new(
+            ReadOnlySource::new(buf),
+            doc_freq,
+            IndexRecordOption::WithFreqs,
+        );
+        assert_eq!(skip_reader.last_doc_in_block(), 1u32);
+        assert!(matches!(
+            skip_reader.block_info,
+            BlockInfo::BitPacked {
+                doc_num_bits: 2u8,
+                tf_num_bits: 3u8,
+                tf_sum: 0,
+                block_wand_fieldnorm_id: 13,
+                block_wand_term_freq: 3
+            }
+       ));
+        skip_reader.advance();
+        assert_eq!(skip_reader.last_doc_in_block(), 5u32);
+        assert!(matches!(
+            skip_reader.block_info(),
+            BlockInfo::BitPacked {
+                doc_num_bits: 5u8,
+                tf_num_bits: 2u8,
+                tf_sum: 0,
+                block_wand_fieldnorm_id: 8,
+                block_wand_term_freq: 2
+            }
+       ));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 3u32 }));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 0u32 }));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 0u32 }));
     }
 
     #[test]
@@ -165,13 +307,68 @@ mod tests {
             skip_serializer.write_doc(5u32, 5u8);
             skip_serializer.data().to_owned()
         };
-        let mut skip_reader = SkipReader::new(OwnedRead::new(buf), IndexRecordOption::Basic);
-        assert!(skip_reader.advance());
-        assert_eq!(skip_reader.doc(), 1u32);
-        assert_eq!(skip_reader.doc_num_bits(), 2u8);
-        assert!(skip_reader.advance());
-        assert_eq!(skip_reader.doc(), 5u32);
-        assert_eq!(skip_reader.doc_num_bits(), 5u8);
-        assert!(!skip_reader.advance());
+        let doc_freq = 3u32 + (COMPRESSION_BLOCK_SIZE * 2) as u32;
+        let mut skip_reader = SkipReader::new(
+            ReadOnlySource::from(buf),
+            doc_freq,
+            IndexRecordOption::Basic,
+        );
+        assert_eq!(skip_reader.last_doc_in_block(), 1u32);
+        assert!(matches!(
+            skip_reader.block_info(),
+            BlockInfo::BitPacked {
+                doc_num_bits: 2u8,
+                tf_num_bits: 0,
+                tf_sum: 0u32,
+                block_wand_fieldnorm_id: 0,
+                block_wand_term_freq: 0
+            }
+        ));
+        skip_reader.advance();
+        assert_eq!(skip_reader.last_doc_in_block(), 5u32);
+        assert!(matches!(
+            skip_reader.block_info(),
+            BlockInfo::BitPacked {
+                doc_num_bits: 5u8,
+                tf_num_bits: 0,
+                tf_sum: 0u32,
+                block_wand_fieldnorm_id: 0,
+                block_wand_term_freq: 0
+            }
+        ));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 3u32 }));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 0u32 }));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 0u32 }));
+    }
+
+    #[test]
+    fn test_skip_multiple_of_block_size() {
+        let buf = {
+            let mut skip_serializer = SkipSerializer::new();
+            skip_serializer.write_doc(1u32, 2u8);
+            skip_serializer.data().to_owned()
+        };
+        let doc_freq = COMPRESSION_BLOCK_SIZE as u32;
+        let mut skip_reader = SkipReader::new(
+            ReadOnlySource::from(buf),
+            doc_freq,
+            IndexRecordOption::Basic,
+        );
+        assert_eq!(skip_reader.last_doc_in_block(), 1u32);
+        assert!(matches!(
+            skip_reader.block_info(),
+            BlockInfo::BitPacked {
+                doc_num_bits: 2u8,
+                tf_num_bits: 0,
+                tf_sum: 0u32,
+                block_wand_fieldnorm_id: 0,
+                block_wand_term_freq: 0
+            }
+        ));
+        skip_reader.advance();
+        assert!(matches!(skip_reader.block_info(), BlockInfo::VInt { num_docs: 0u32 }));
     }
 }

src/query/automaton_weight.rs

@@ -43,7 +43,6 @@ where
     fn scorer(&self, reader: &SegmentReader, boost: f32) -> Result<Box<dyn Scorer>> {
         let max_doc = reader.max_doc();
         let mut doc_bitset = BitSet::with_max_value(max_doc);
-
         let inverted_index = reader.inverted_index(self.field);
         let term_dict = inverted_index.terms();
         let mut term_stream = self.automaton_stream(term_dict);
@@ -51,10 +50,15 @@ where
             let term_info = term_stream.value();
             let mut block_segment_postings = inverted_index
                 .read_block_postings_from_terminfo(term_info, IndexRecordOption::Basic);
-            while block_segment_postings.advance() {
-                for &doc in block_segment_postings.docs() {
+            loop {
+                let docs = block_segment_postings.docs();
+                if docs.is_empty() {
+                    break;
+                }
+                for &doc in docs {
                     doc_bitset.insert(doc);
                 }
+                block_segment_postings.advance();
             }
         }
         let doc_bitset = BitSetDocSet::from(doc_bitset);

src/query/bm25.rs

@@ -3,11 +3,14 @@ use crate::query::Explanation;
 use crate::Score;
 use crate::Searcher;
 use crate::Term;
+use serde::Deserialize;
+use serde::Serialize;
 
 const K1: f32 = 1.2;
 const B: f32 = 0.75;
 
 fn idf(doc_freq: u64, doc_count: u64) -> f32 {
+    assert!(doc_count >= doc_freq, "{} >= {}", doc_count, doc_freq);
     let x = ((doc_count - doc_freq) as f32 + 0.5) / (doc_freq as f32 + 0.5);
     (1f32 + x).ln()
 }
@@ -25,6 +28,12 @@ fn compute_tf_cache(average_fieldnorm: f32) -> [f32; 256] {
     cache
 }
 
+#[derive(Clone, PartialEq, Debug, Serialize, Deserialize)]
+pub struct BM25Params {
+    pub idf: f32,
+    pub avg_fieldnorm: f32,
+}
+
 pub struct BM25Weight {
     idf_explain: Explanation,
     weight: f32,
@@ -62,17 +71,9 @@ impl BM25Weight {
         }
         let average_fieldnorm = total_num_tokens as f32 / total_num_docs as f32;
 
-        let mut idf_explain: Explanation;
         if terms.len() == 1 {
             let term_doc_freq = searcher.doc_freq(&terms[0]);
-            let idf = idf(term_doc_freq, total_num_docs);
-            idf_explain =
-                Explanation::new("idf, computed as log(1 + (N - n + 0.5) / (n + 0.5))", idf);
-            idf_explain.add_const(
-                "n, number of docs containing this term",
-                term_doc_freq as f32,
-            );
-            idf_explain.add_const("N, total number of docs", total_num_docs as f32);
+            BM25Weight::for_one_term(term_doc_freq, total_num_docs, average_fieldnorm)
         } else {
             let idf = terms
                 .iter()
@@ -81,9 +82,21 @@ impl BM25Weight {
                     idf(term_doc_freq, total_num_docs)
                 })
                 .sum::<f32>();
-            idf_explain = Explanation::new("idf", idf);
+            let idf_explain = Explanation::new("idf", idf);
+            BM25Weight::new(idf_explain, average_fieldnorm)
         }
-        BM25Weight::new(idf_explain, average_fieldnorm)
+    }
+
+    pub fn for_one_term(term_doc_freq: u64, total_num_docs: u64, avg_fieldnorm: f32) -> BM25Weight {
+        let idf = idf(term_doc_freq, total_num_docs);
+        let mut idf_explain =
+            Explanation::new("idf, computed as log(1 + (N - n + 0.5) / (n + 0.5))", idf);
+        idf_explain.add_const(
+            "n, number of docs containing this term",
+            term_doc_freq as f32,
+        );
+        idf_explain.add_const("N, total number of docs", total_num_docs as f32);
+        BM25Weight::new(idf_explain, avg_fieldnorm)
     }
 
     fn new(idf_explain: Explanation, average_fieldnorm: f32) -> BM25Weight {
@@ -98,15 +111,23 @@ impl BM25Weight {
 
     #[inline(always)]
     pub fn score(&self, fieldnorm_id: u8, term_freq: u32) -> Score {
-        let norm = self.cache[fieldnorm_id as usize];
+        self.weight * self.tf_factor(fieldnorm_id, term_freq)
+    }
+
+    pub fn max_score(&self) -> Score {
+        self.score(255u8, 2_013_265_944)
+    }
+
+    #[inline(always)]
+    pub(crate) fn tf_factor(&self, fieldnorm_id: u8, term_freq: u32) -> f32 {
         let term_freq = term_freq as f32;
-        self.weight * term_freq / (term_freq + norm)
+        let norm = self.cache[fieldnorm_id as usize];
+        term_freq / (term_freq + norm)
     }
 
     pub fn explain(&self, fieldnorm_id: u8, term_freq: u32) -> Explanation {
         // The explain format is directly copied from Lucene's.
         // (So, Kudos to Lucene)
-
         let score = self.score(fieldnorm_id, term_freq);
 
         let norm = self.cache[fieldnorm_id as usize];
@@ -139,10 +160,10 @@ impl BM25Weight {
 mod tests {
 
     use super::idf;
-    use crate::tests::assert_nearly_equals;
+    use crate::assert_nearly_equals;
 
     #[test]
     fn test_idf() {
-        assert_nearly_equals(idf(1, 2), 0.6931472);
+        assert_nearly_equals!(idf(1, 2), std::f32::consts::LN_2);
     }
 }

src/query/boolean_query/block_wand.rs

@@ -0,0 +1,206 @@
+use crate::query::term_query::TermScorer;
+use crate::query::Scorer;
+use crate::{DocId, DocSet, Score, TERMINATED};
+use std::ops::DerefMut;
+use std::ops::Deref;
+
+/// Takes a term_scorers sorted by their current doc() and a threshold and returns
+/// Returns (pivot_len, pivot_ord) defined as follows:
+/// - `pivot_doc` lowest document that has a chance of exceeding (>) the threshold score.
+/// - `before_pivot_len` number of term_scorers such that term_scorer.doc() < pivot.
+/// - `pivot_len` number of term_scorers such that term_scorer.doc() <= pivot.
+///
+/// We always have `before_pivot_len` < `pivot_len`.
+///
+/// None is returned if we establish that no document can exceed the threshold.
+fn find_pivot_doc(term_scorers: &[TermScorerWithMaxScore], threshold: f32) -> Option<(usize, usize, DocId)> {
+    let mut max_score = 0.0f32;
+    let mut before_pivot_len = 0;
+    let mut pivot_doc = TERMINATED;
+    while before_pivot_len < term_scorers.len() {
+        let term_scorer = &term_scorers[before_pivot_len];
+        max_score += term_scorer.max_score;
+        if max_score > threshold {
+            pivot_doc = term_scorer.doc();
+            break;
+        }
+        before_pivot_len += 1;
+    }
+    if pivot_doc == TERMINATED {
+        return None;
+    }
+    // Right now i is an ordinal, we want a len.
+    let mut pivot_len = before_pivot_len + 1;
+    // Some other term_scorer may be positioned on the same document.
+    pivot_len += term_scorers[pivot_len..].iter()
+        .take_while(|term_scorer| term_scorer.doc() == pivot_doc)
+        .count();
+    Some((before_pivot_len, pivot_len, pivot_doc))
+}
+
+struct TermScorerWithMaxScore<'a> {
+    scorer: &'a mut TermScorer,
+    max_score: f32,
+}
+
+impl<'a> From<&'a mut TermScorer> for TermScorerWithMaxScore<'a> {
+    fn from(scorer: &'a mut TermScorer) -> Self {
+        let max_score = scorer.max_score();
+        TermScorerWithMaxScore {
+            scorer,
+            max_score
+        }
+    }
+}
+
+impl<'a> Deref for TermScorerWithMaxScore<'a> {
+    type Target = TermScorer;
+
+    fn deref(&self) -> &Self::Target {
+        self.scorer
+    }
+}
+
+
+impl<'a> DerefMut for TermScorerWithMaxScore<'a> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.scorer
+    }
+}
+
+// Before and after calling this method, scorers need to be sorted by their `.doc()`.
+fn block_max_was_too_low_advance_one_scorer(scorers: &mut Vec<TermScorerWithMaxScore>, pivot_len: usize) {
+    let mut scorer_to_seek = pivot_len - 1;
+    let mut doc_to_seek_after = scorers[scorer_to_seek].doc();
+    for scorer_ord in (0..pivot_len - 1).rev() {
+        let scorer = &scorers[scorer_ord];
+        if scorer.last_doc_in_block() <= doc_to_seek_after {
+            doc_to_seek_after = scorer.last_doc_in_block();
+            scorer_to_seek = scorer_ord;
+        }
+    }
+    for scorer in &scorers[pivot_len..] {
+        if scorer.doc() <= doc_to_seek_after {
+            doc_to_seek_after = scorer.doc();
+        }
+    }
+    scorers[scorer_to_seek].seek(doc_to_seek_after + 1);
+    restore_ordering(scorers, scorer_to_seek);
+}
+
+
+// Given a list of term_scorers and a `ord` and assuming that `term_scorers[ord]` is sorted
+// except term_scorers[ord] that might be in advance compared to its ranks,
+// bubble up term_scorers[ord] in order to restore the ordering.
+fn restore_ordering(term_scorers: &mut Vec<TermScorerWithMaxScore>, ord: usize) {
+    let doc = term_scorers[ord].doc();
+    for i in ord + 1..term_scorers.len() {
+        if term_scorers[i].doc() >= doc {
+            break;
+        }
+        term_scorers.swap(i, i - 1);
+    }
+}
+
+
+// Attempts to advance all term_scorers between `&term_scorers[0..before_len]` to the pivot.
+// If this works, return true.
+// If this fails (ie: one of the term_scorer does not contain `pivot_doc` and seek goes past the
+// pivot), reorder the term_scorers to ensure the list is still sorted and returns `false`.
+// If a term_scorer reach TERMINATED in the process return false remove the term_scorer and return.
+fn align_scorers(term_scorers: &mut Vec<TermScorerWithMaxScore>, pivot_doc: DocId, before_pivot_len: usize) -> bool {
+    debug_assert_ne!(pivot_doc, TERMINATED);
+    for i in (0..before_pivot_len).rev() {
+        let new_doc = term_scorers[i].seek(pivot_doc);
+        if new_doc != pivot_doc {
+            if new_doc == TERMINATED {
+                term_scorers.swap_remove(i);
+            }
+            // We went past the pivot.
+            // We just go through the outer loop mechanic (Note that pivot is
+            // still a possible candidate).
+            //
+            // Termination is still guaranteed since we can only consider the same
+            // pivot at most term_scorers.len() - 1 times.
+            restore_ordering(term_scorers, i);
+            return false;
+        }
+    }
+    return true;
+}
+
+// Assumes terms_scorers[..pivot_len] are positioned on the same doc (pivot_doc).
+// Advance term_scorers[..pivot_len] and out of these removes the terminated scores.
+// Restores the ordering of term_scorers.
+fn advance_all_scorers_on_pivot(term_scorers: &mut Vec<TermScorerWithMaxScore>, pivot_len: usize) {
+    let mut i = 0;
+    for _ in 0..pivot_len {
+        if term_scorers[i].advance() == TERMINATED {
+            term_scorers.swap_remove(i);
+        } else {
+            i += 1;
+        }
+    }
+    term_scorers.sort_by_key(|scorer| scorer.doc());
+}
+
+pub fn block_wand(
+    mut scorers: Vec<TermScorer>,
+    mut threshold: f32,
+    callback: &mut dyn FnMut(u32, Score) -> Score,
+) {
+    let mut scorers: Vec<TermScorerWithMaxScore> = scorers.iter_mut().map(TermScorerWithMaxScore::from).collect();
+    scorers.sort_by_key(|scorer| scorer.doc());
+    loop {
+        // At this point we need to ensure that the scorers are sorted!
+        if let Some((before_pivot_len, pivot_len, pivot_doc)) = find_pivot_doc(&scorers[..], threshold) {
+            debug_assert_ne!(pivot_doc, TERMINATED);
+            debug_assert!(before_pivot_len < pivot_len);
+            let block_max_score_upperbound: Score =  scorers[..pivot_len].iter_mut()
+                .map(|scorer| {
+                    scorer.shallow_seek(pivot_doc);
+                    scorer.block_max_score()
+                })
+                .sum();
+
+            // Beware after shallow advance, skip readers can be in advance compared to
+            // the segment posting lists.
+            //
+            // `block_segment_postings.load_block()` need to be called separately.
+            if block_max_score_upperbound <= threshold {
+                // Block max condition was not reached.
+                // We could get away by simply advancing the scorers to DocId + 1 but it would
+                // be inefficient. The optimization requires proper explanation and was
+                // isolated in a different function.
+                block_max_was_too_low_advance_one_scorer(&mut scorers, pivot_len);
+                continue;
+            }
+
+            // Block max condition is observed.
+            //
+            // Let's try and advance all scorers before the pivot to the pivot.
+            if !align_scorers(&mut scorers, pivot_doc, before_pivot_len) {
+                // At least of the scorer does not contain the pivot.
+                //
+                // Let's stop scoring this pivot and go through the pivot selection again.
+                // Note that the current pivot is not necessarily a bad candidate and it
+                // may be picked again.
+                continue;
+            }
+
+            // At this point, all scorers are positioned on the doc.
+            let score = scorers[..pivot_len]
+                .iter_mut()
+                .map(|scorer| scorer.score())
+                .sum();
+            if score > threshold {
+                threshold = callback(pivot_doc, score);
+            }
+            // let's advance all of the scorers that are currently positioned on the pivot.
+            advance_all_scorers_on_pivot(&mut scorers, pivot_len);
+
+        } else {
+            return;
+        }
+    }
+}

src/query/boolean_query/boolean_weight.rs

@@ -1,7 +1,9 @@
 use crate::core::SegmentReader;
+use crate::postings::FreqReadingOption;
 use crate::query::explanation::does_not_match;
 use crate::query::score_combiner::{DoNothingCombiner, ScoreCombiner, SumWithCoordsCombiner};
 use crate::query::term_query::TermScorer;
+use crate::query::weight::{for_each_pruning_scorer, for_each_scorer};
 use crate::query::EmptyScorer;
 use crate::query::Exclude;
 use crate::query::Occur;
@@ -10,16 +12,21 @@ use crate::query::Scorer;
 use crate::query::Union;
 use crate::query::Weight;
 use crate::query::{intersect_scorers, Explanation};
-use crate::DocId;
+use crate::{DocId, Score};
 use std::collections::HashMap;
 
-fn scorer_union<TScoreCombiner>(scorers: Vec<Box<dyn Scorer>>) -> Box<dyn Scorer>
+enum SpecializedScorer {
+    TermUnion(Vec<TermScorer>),
+    Other(Box<dyn Scorer>),
+}
+
+fn scorer_union<TScoreCombiner>(scorers: Vec<Box<dyn Scorer>>) -> SpecializedScorer
 where
     TScoreCombiner: ScoreCombiner,
 {
     assert!(!scorers.is_empty());
     if scorers.len() == 1 {
-        return scorers.into_iter().next().unwrap(); //< we checked the size beforehands
+        return SpecializedScorer::Other(scorers.into_iter().next().unwrap()); //< we checked the size beforehands
     }
 
     {
@@ -29,14 +36,30 @@ where
                 .into_iter()
                 .map(|scorer| *(scorer.downcast::<TermScorer>().map_err(|_| ()).unwrap()))
                 .collect();
-            let scorer: Box<dyn Scorer> =
-                Box::new(Union::<TermScorer, TScoreCombiner>::from(scorers));
-            return scorer;
+            if scorers
+                .iter()
+                .all(|scorer| scorer.freq_reading_option() == FreqReadingOption::ReadFreq)
+            {
+                // Block wand is only available iff we read frequencies.
+                return SpecializedScorer::TermUnion(scorers);
+            } else {
+                return SpecializedScorer::Other(Box::new(Union::<_, TScoreCombiner>::from(
+                    scorers,
+                )));
+            }
         }
     }
+    SpecializedScorer::Other(Box::new(Union::<_, TScoreCombiner>::from(scorers)))
+}
 
-    let scorer: Box<dyn Scorer> = Box::new(Union::<_, TScoreCombiner>::from(scorers));
-    scorer
+fn into_box_scorer<TScoreCombiner: ScoreCombiner>(scorer: SpecializedScorer) -> Box<dyn Scorer> {
+    match scorer {
+        SpecializedScorer::TermUnion(term_scorers) => {
+            let union_scorer = Union::<TermScorer, TScoreCombiner>::from(term_scorers);
+            Box::new(union_scorer)
+        }
+        SpecializedScorer::Other(scorer) => scorer,
+    }
 }
 
 pub struct BooleanWeight {
@@ -72,41 +95,51 @@ impl BooleanWeight {
         &self,
         reader: &SegmentReader,
         boost: f32,
-    ) -> crate::Result<Box<dyn Scorer>> {
+    ) -> crate::Result<SpecializedScorer> {
         let mut per_occur_scorers = self.per_occur_scorers(reader, boost)?;
 
-        let should_scorer_opt: Option<Box<dyn Scorer>> = per_occur_scorers
+        let should_scorer_opt: Option<SpecializedScorer> = per_occur_scorers
             .remove(&Occur::Should)
             .map(scorer_union::<TScoreCombiner>);
 
         let exclude_scorer_opt: Option<Box<dyn Scorer>> = per_occur_scorers
             .remove(&Occur::MustNot)
-            .map(scorer_union::<TScoreCombiner>);
+            .map(scorer_union::<DoNothingCombiner>)
+            .map(into_box_scorer::<DoNothingCombiner>);
 
         let must_scorer_opt: Option<Box<dyn Scorer>> = per_occur_scorers
             .remove(&Occur::Must)
             .map(intersect_scorers);
 
-        let positive_scorer: Box<dyn Scorer> = match (should_scorer_opt, must_scorer_opt) {
+        let positive_scorer: SpecializedScorer = match (should_scorer_opt, must_scorer_opt) {
             (Some(should_scorer), Some(must_scorer)) => {
                 if self.scoring_enabled {
-                    Box::new(RequiredOptionalScorer::<_, _, TScoreCombiner>::new(
+                    SpecializedScorer::Other(Box::new(RequiredOptionalScorer::<
+                        Box<dyn Scorer>,
+                        Box<dyn Scorer>,
+                        TScoreCombiner,
+                    >::new(
                         must_scorer,
-                        should_scorer,
-                    ))
+                        into_box_scorer::<TScoreCombiner>(should_scorer),
+                    )))
                 } else {
-                    must_scorer
+                    SpecializedScorer::Other(must_scorer)
                 }
             }
-            (None, Some(must_scorer)) => must_scorer,
+            (None, Some(must_scorer)) => SpecializedScorer::Other(must_scorer),
             (Some(should_scorer), None) => should_scorer,
             (None, None) => {
-                return Ok(Box::new(EmptyScorer));
+                return Ok(SpecializedScorer::Other(Box::new(EmptyScorer)));
             }
         };
 
         if let Some(exclude_scorer) = exclude_scorer_opt {
-            Ok(Box::new(Exclude::new(positive_scorer, exclude_scorer)))
+            let positive_scorer_boxed: Box<dyn Scorer> =
+                into_box_scorer::<TScoreCombiner>(positive_scorer);
+            Ok(SpecializedScorer::Other(Box::new(Exclude::new(
+                positive_scorer_boxed,
+                exclude_scorer,
+            ))))
         } else {
             Ok(positive_scorer)
         }
@@ -126,8 +159,12 @@ impl Weight for BooleanWeight {
             }
         } else if self.scoring_enabled {
             self.complex_scorer::<SumWithCoordsCombiner>(reader, boost)
+                .map(|specialized_scorer| {
+                    into_box_scorer::<SumWithCoordsCombiner>(specialized_scorer)
+                })
         } else {
             self.complex_scorer::<DoNothingCombiner>(reader, boost)
+                .map(into_box_scorer::<DoNothingCombiner>)
         }
     }
 
@@ -150,6 +187,53 @@ impl Weight for BooleanWeight {
         }
         Ok(explanation)
     }
+
+    fn for_each(
+        &self,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score),
+    ) -> crate::Result<()> {
+        let scorer = self.complex_scorer::<SumWithCoordsCombiner>(reader, 1.0f32)?;
+        match scorer {
+            SpecializedScorer::TermUnion(term_scorers) => {
+                let mut union_scorer =
+                    Union::<TermScorer, SumWithCoordsCombiner>::from(term_scorers);
+                for_each_scorer(&mut union_scorer, callback);
+            }
+            SpecializedScorer::Other(mut scorer) => {
+                for_each_scorer(scorer.as_mut(), callback);
+            }
+        }
+        Ok(())
+    }
+
+    /// Calls `callback` with all of the `(doc, score)` for which score
+    /// is exceeding a given threshold.
+    ///
+    /// This method is useful for the TopDocs collector.
+    /// For all docsets, the blanket implementation has the benefit
+    /// of prefiltering (doc, score) pairs, avoiding the
+    /// virtual dispatch cost.
+    ///
+    /// More importantly, it makes it possible for scorers to implement
+    /// important optimization (e.g. BlockWAND for union).
+    fn for_each_pruning(
+        &self,
+        threshold: f32,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score) -> Score,
+    ) -> crate::Result<()> {
+        let scorer = self.complex_scorer::<SumWithCoordsCombiner>(reader, 1.0f32)?;
+        match scorer {
+            SpecializedScorer::TermUnion(term_scorers) => {
+                super::block_wand(term_scorers, threshold, callback);
+            }
+            SpecializedScorer::Other(mut scorer) => {
+                for_each_pruning_scorer(scorer.as_mut(), threshold, callback);
+            }
+        }
+        Ok(())
+    }
 }
 
 fn is_positive_occur(occur: Occur) -> bool {

src/query/boolean_query/mod.rs

@@ -1,13 +1,17 @@
+mod block_wand;
 mod boolean_query;
 mod boolean_weight;
 
+pub(crate) use self::block_wand::block_wand;
 pub use self::boolean_query::BooleanQuery;
 
 #[cfg(test)]
 mod tests {
 
     use super::*;
+    use crate::assert_nearly_equals;
     use crate::collector::tests::TEST_COLLECTOR_WITH_SCORE;
+    use crate::collector::TopDocs;
     use crate::query::score_combiner::SumWithCoordsCombiner;
     use crate::query::term_query::TermScorer;
     use crate::query::Intersection;
@@ -18,9 +22,8 @@ mod tests {
     use crate::query::Scorer;
     use crate::query::TermQuery;
     use crate::schema::*;
-    use crate::tests::assert_nearly_equals;
     use crate::Index;
-    use crate::{DocAddress, DocId};
+    use crate::{DocAddress, DocId, Score};
 
     fn aux_test_helper() -> (Index, Field) {
         let mut schema_builder = Schema::builder();
@@ -140,7 +143,6 @@ mod tests {
                 .map(|doc| doc.1)
                 .collect::<Vec<DocId>>()
         };
-
         {
             let boolean_query = BooleanQuery::from(vec![(Occur::Must, make_term_query("a"))]);
             assert_eq!(matching_docs(&boolean_query), vec![0, 1, 3]);
@@ -177,6 +179,54 @@ mod tests {
         }
     }
 
+    #[test]
+    pub fn test_boolean_query_two_excluded() {
+        let (index, text_field) = aux_test_helper();
+
+        let make_term_query = |text: &str| {
+            let term_query = TermQuery::new(
+                Term::from_field_text(text_field, text),
+                IndexRecordOption::Basic,
+            );
+            let query: Box<dyn Query> = Box::new(term_query);
+            query
+        };
+
+        let reader = index.reader().unwrap();
+
+        let matching_topdocs = |query: &dyn Query| {
+            reader
+                .searcher()
+                .search(query, &TopDocs::with_limit(3))
+                .unwrap()
+        };
+
+        let score_doc_4: Score; // score of doc 4 should not be influenced by exclusion
+        {
+            let boolean_query_no_excluded =
+                BooleanQuery::from(vec![(Occur::Must, make_term_query("d"))]);
+            let topdocs_no_excluded = matching_topdocs(&boolean_query_no_excluded);
+            assert_eq!(topdocs_no_excluded.len(), 2);
+            let (top_score, top_doc) = topdocs_no_excluded[0];
+            assert_eq!(top_doc, DocAddress(0, 4));
+            assert_eq!(topdocs_no_excluded[1].1, DocAddress(0, 3)); // ignore score of doc 3.
+            score_doc_4 = top_score;
+        }
+
+        {
+            let boolean_query_two_excluded = BooleanQuery::from(vec![
+                (Occur::Must, make_term_query("d")),
+                (Occur::MustNot, make_term_query("a")),
+                (Occur::MustNot, make_term_query("b")),
+            ]);
+            let topdocs_excluded = matching_topdocs(&boolean_query_two_excluded);
+            assert_eq!(topdocs_excluded.len(), 1);
+            let (top_score, top_doc) = topdocs_excluded[0];
+            assert_eq!(top_doc, DocAddress(0, 4));
+            assert_eq!(top_score, score_doc_4);
+        }
+    }
+
     #[test]
     pub fn test_boolean_query_with_weight() {
         let mut schema_builder = Schema::builder();
@@ -208,14 +258,14 @@ mod tests {
                 .scorer(searcher.segment_reader(0u32), 1.0f32)
                 .unwrap();
             assert_eq!(boolean_scorer.doc(), 0u32);
-            assert_nearly_equals(boolean_scorer.score(), 0.84163445f32);
+            assert_nearly_equals!(boolean_scorer.score(), 0.84163445f32);
         }
         {
             let mut boolean_scorer = boolean_weight
                 .scorer(searcher.segment_reader(0u32), 2.0f32)
                 .unwrap();
             assert_eq!(boolean_scorer.doc(), 0u32);
-            assert_nearly_equals(boolean_scorer.score(), 1.6832689f32);
+            assert_nearly_equals!(boolean_scorer.score(), 1.6832689f32);
         }
     }
 
@@ -274,7 +324,7 @@ mod tests {
         index_writer.add_document(doc!(
             // tf = 1 1
             title =>  "PDF Мастер Класс \"Морячок\" (Оксана Лифенко)",
-            // tf = 0 0 
+            // tf = 0 0
             text => "https://i.ibb.co/pzvHrDN/I3d U T6 Gg TM.jpg\nhttps://i.ibb.co/NFrb6v6/N0ls Z9nwjb U.jpg\nВ описание входит штаны, кофта, берет, матросский воротник. Описание продается в формате PDF, состоит из 12 страниц формата А4 и может быть напечатано на любом принтере.\nОписание предназначено для кукол BJD RealPuki от FairyLand, но может подойти и другим подобным куклам. Также вы можете вязать этот наряд из обычной пряжи, и он подойдет для куколок побольше.\nhttps://vk.com/market 95724412?w=product 95724412_2212"
         ));
         for _ in 0..1_000 {

src/query/exclude.rs

@@ -3,6 +3,11 @@ use crate::query::Scorer;
 use crate::DocId;
 use crate::Score;
 
+#[inline(always)]
+fn is_within<TDocSetExclude: DocSet>(docset: &mut TDocSetExclude, doc: DocId) -> bool {
+    docset.doc() <= doc && docset.seek(doc) == doc
+}
+
 /// Filters a given `DocSet` by removing the docs from a given `DocSet`.
 ///
 /// The excluding docset has no impact on scoring.
@@ -23,8 +28,7 @@ where
     ) -> Exclude<TDocSet, TDocSetExclude> {
         while underlying_docset.doc() != TERMINATED {
             let target = underlying_docset.doc();
-            if excluding_docset.seek(target) != target {
-                // this document is not excluded.
+            if !is_within(&mut excluding_docset, target) {
                 break;
             }
             underlying_docset.advance();
@@ -36,42 +40,30 @@ where
     }
 }
 
-impl<TDocSet, TDocSetExclude> Exclude<TDocSet, TDocSetExclude>
-where
-    TDocSet: DocSet,
-    TDocSetExclude: DocSet,
-{
-    /// Returns true iff the doc is not removed.
-    ///
-    /// The method has to be called with non strictly
-    /// increasing `doc`.
-    fn accept(&mut self) -> bool {
-        let doc = self.underlying_docset.doc();
-        self.excluding_docset.seek(doc) != doc
-    }
-}
-
 impl<TDocSet, TDocSetExclude> DocSet for Exclude<TDocSet, TDocSetExclude>
 where
     TDocSet: DocSet,
     TDocSetExclude: DocSet,
 {
     fn advance(&mut self) -> DocId {
-        while self.underlying_docset.advance() != TERMINATED {
-            if self.accept() {
-                return self.doc();
+        loop {
+            let candidate = self.underlying_docset.advance();
+            if candidate == TERMINATED {
+                return TERMINATED;
+            }
+            if !is_within(&mut self.excluding_docset, candidate) {
+                return candidate;
             }
         }
-        TERMINATED
     }
 
     fn seek(&mut self, target: DocId) -> DocId {
-        let underlying_seek_result = self.underlying_docset.seek(target);
-        if underlying_seek_result == TERMINATED {
+        let candidate = self.underlying_docset.seek(target);
+        if candidate == TERMINATED {
             return TERMINATED;
         }
-        if self.accept() {
-            return underlying_seek_result;
+        if !is_within(&mut self.excluding_docset, candidate) {
+            return candidate;
         }
         self.advance()
     }
@@ -129,7 +121,7 @@ mod tests {
                     VecDocSet::from(vec![1, 2, 3, 10, 16, 24]),
                 ))
             },
-            vec![1, 2, 5, 8, 10, 15, 24],
+            vec![5, 8, 10, 15, 24],
         );
     }
 

src/query/explanation.rs

@@ -1,5 +1,6 @@
 use crate::{DocId, TantivyError};
 use serde::Serialize;
+use std::fmt;
 
 pub(crate) fn does_not_match(doc: DocId) -> TantivyError {
     TantivyError::InvalidArgument(format!("Document #({}) does not match", doc))
@@ -18,6 +19,12 @@ pub struct Explanation {
     details: Vec<Explanation>,
 }
 
+impl fmt::Debug for Explanation {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "Explanation({})", self.to_pretty_json())
+    }
+}
+
 impl Explanation {
     /// Creates a new explanation object.
     pub fn new<T: ToString>(description: T, value: f32) -> Explanation {

src/query/fuzzy_query.rs

@@ -163,10 +163,10 @@ impl Query for FuzzyTermQuery {
 #[cfg(test)]
 mod test {
     use super::FuzzyTermQuery;
+    use crate::assert_nearly_equals;
     use crate::collector::TopDocs;
     use crate::schema::Schema;
     use crate::schema::TEXT;
-    use crate::tests::assert_nearly_equals;
     use crate::Index;
     use crate::Term;
 
@@ -199,7 +199,7 @@ mod test {
                 .unwrap();
             assert_eq!(top_docs.len(), 1, "Expected only 1 document");
             let (score, _) = top_docs[0];
-            assert_nearly_equals(1f32, score);
+            assert_nearly_equals!(1f32, score);
         }
 
         // fails because non-prefix Levenshtein distance is more than 1 (add 'a' and 'n')
@@ -223,7 +223,7 @@ mod test {
                 .unwrap();
             assert_eq!(top_docs.len(), 1, "Expected only 1 document");
             let (score, _) = top_docs[0];
-            assert_nearly_equals(1f32, score);
+            assert_nearly_equals!(1f32, score);
         }
     }
 }

src/query/intersection.rs

@@ -53,7 +53,8 @@ pub struct Intersection<TDocSet: DocSet, TOtherDocSet: DocSet = Box<dyn Scorer>>
 }
 
 fn go_to_first_doc<TDocSet: DocSet>(docsets: &mut [TDocSet]) -> DocId {
-    let mut candidate = 0;
+    assert!(!docsets.is_empty());
+    let mut candidate = docsets.iter().map(TDocSet::doc).max().unwrap();
     'outer: loop {
         for docset in docsets.iter_mut() {
             let seek_doc = docset.seek(candidate);
@@ -112,18 +113,15 @@ impl<TDocSet: DocSet, TOtherDocSet: DocSet> DocSet for Intersection<TDocSet, TOt
             debug_assert_eq!(left.doc(), right.doc());
             // test the remaining scorers;
             for docset in self.others.iter_mut() {
-                // `candidate_ord` is already at the
-                // right position.
-                //
-                // Calling `skip_next` would advance this docset
-                // and miss it.
                 let seek_doc = docset.seek(candidate);
                 if seek_doc > candidate {
                     candidate = left.seek(seek_doc);
                     continue 'outer;
                 }
             }
-
+            debug_assert_eq!(candidate, self.left.doc());
+            debug_assert_eq!(candidate, self.right.doc());
+            debug_assert!(self.others.iter().all(|docset| docset.doc() == candidate));
             return candidate;
         }
     }
@@ -134,7 +132,10 @@ impl<TDocSet: DocSet, TOtherDocSet: DocSet> DocSet for Intersection<TDocSet, TOt
         for docset in &mut self.others {
             docsets.push(docset);
         }
-        go_to_first_doc(&mut docsets[..])
+        let doc = go_to_first_doc(&mut docsets[..]);
+        debug_assert!(docsets.iter().all(|docset| docset.doc() == doc));
+        debug_assert!(doc >= target);
+        doc
     }
 
     fn doc(&self) -> DocId {

src/query/mod.rs

@@ -26,6 +26,7 @@ mod weight;
 mod vec_docset;
 
 pub(crate) mod score_combiner;
+pub(crate) use self::bm25::BM25Weight;
 pub use self::intersection::Intersection;
 pub use self::union::Union;
 

src/query/phrase_query/mod.rs

@@ -10,12 +10,13 @@ pub use self::phrase_weight::PhraseWeight;
 pub mod tests {
 
     use super::*;
+    use crate::assert_nearly_equals;
     use crate::collector::tests::{TEST_COLLECTOR_WITHOUT_SCORE, TEST_COLLECTOR_WITH_SCORE};
     use crate::core::Index;
+    use crate::query::Weight;
     use crate::schema::{Schema, Term, TEXT};
-    use crate::tests::assert_nearly_equals;
-    use crate::DocAddress;
     use crate::DocId;
+    use crate::{DocAddress, TERMINATED};
 
     pub fn create_index(texts: &[&'static str]) -> Index {
         let mut schema_builder = Schema::builder();
@@ -67,6 +68,23 @@ pub mod tests {
         assert!(test_query(vec!["g", "a"]).is_empty());
     }
 
+    #[test]
+    pub fn test_phrase_query_simple() -> crate::Result<()> {
+        let index = create_index(&["a b b d c g c", "a b a b c"]);
+        let text_field = index.schema().get_field("text").unwrap();
+        let searcher = index.reader()?.searcher();
+        let terms: Vec<Term> = vec!["a", "b", "c"]
+            .iter()
+            .map(|text| Term::from_field_text(text_field, text))
+            .collect();
+        let phrase_query = PhraseQuery::new(terms);
+        let phrase_weight = phrase_query.phrase_weight(&searcher, false)?;
+        let mut phrase_scorer = phrase_weight.scorer(searcher.segment_reader(0), 1.0f32)?;
+        assert_eq!(phrase_scorer.doc(), 1);
+        assert_eq!(phrase_scorer.advance(), TERMINATED);
+        Ok(())
+    }
+
     #[test]
     pub fn test_phrase_query_no_score() {
         let index = create_index(&[
@@ -157,8 +175,8 @@ pub mod tests {
                 .to_vec()
         };
         let scores = test_query(vec!["a", "b"]);
-        assert_nearly_equals(scores[0], 0.40618482);
-        assert_nearly_equals(scores[1], 0.46844664);
+        assert_nearly_equals!(scores[0], 0.40618482);
+        assert_nearly_equals!(scores[1], 0.46844664);
     }
 
     #[test] // motivated by #234

src/query/phrase_query/phrase_scorer.rs

@@ -239,6 +239,7 @@ impl<TPostings: Postings> DocSet for PhraseScorer<TPostings> {
     }
 
     fn seek(&mut self, target: DocId) -> DocId {
+        debug_assert!(target >= self.doc());
         let doc = self.intersection_docset.seek(target);
         if doc == TERMINATED || self.phrase_match() {
             return doc;
@@ -266,7 +267,6 @@ impl<TPostings: Postings> Scorer for PhraseScorer<TPostings> {
 
 #[cfg(test)]
 mod tests {
-
     use super::{intersection, intersection_count};
 
     fn test_intersection_sym(left: &[u32], right: &[u32], expected: &[u32]) {

src/query/query_parser/query_parser.rs

@@ -113,7 +113,7 @@ fn trim_ast(logical_ast: LogicalAST) -> Option<LogicalAST> {
 /// The language covered by the current parser is extremely simple.
 ///
 /// * simple terms: "e.g.: `Barack Obama` are simply tokenized using
-///   tantivy's [`SimpleTokenizer`](tantivy::tokenizer::SimpleTokenizer), hence
+///   tantivy's [`SimpleTokenizer`](../tokenizer/struct.SimpleTokenizer.html), hence
 ///   becoming `["barack", "obama"]`. The terms are then searched within
 ///   the default terms of the query parser.
 ///

src/query/range_query.rs

@@ -300,10 +300,15 @@ impl Weight for RangeWeight {
             let term_info = term_range.value();
             let mut block_segment_postings = inverted_index
                 .read_block_postings_from_terminfo(term_info, IndexRecordOption::Basic);
-            while block_segment_postings.advance() {
+            loop {
+                let docs = block_segment_postings.docs();
+                if docs.is_empty() {
+                    break;
+                }
                 for &doc in block_segment_postings.docs() {
                     doc_bitset.insert(doc);
                 }
+                block_segment_postings.advance();
             }
         }
         let doc_bitset = BitSetDocSet::from(doc_bitset);

src/query/regex_query.rs

@@ -89,10 +89,10 @@ impl Query for RegexQuery {
 #[cfg(test)]
 mod test {
     use super::RegexQuery;
+    use crate::assert_nearly_equals;
     use crate::collector::TopDocs;
     use crate::schema::TEXT;
     use crate::schema::{Field, Schema};
-    use crate::tests::assert_nearly_equals;
     use crate::{Index, IndexReader};
     use std::sync::Arc;
     use tantivy_fst::Regex;
@@ -129,7 +129,7 @@ mod test {
                 .unwrap();
             assert_eq!(scored_docs.len(), 1, "Expected only 1 document");
             let (score, _) = scored_docs[0];
-            assert_nearly_equals(1f32, score);
+            assert_nearly_equals!(1f32, score);
         }
         let top_docs = searcher
             .search(&query_matching_zero, &TopDocs::with_limit(2))

src/query/reqopt_scorer.rs

@@ -72,7 +72,7 @@ where
         let doc = self.doc();
         let mut score_combiner = TScoreCombiner::default();
         score_combiner.update(&mut self.req_scorer);
-        if self.opt_scorer.seek(doc) == doc {
+        if self.opt_scorer.doc() <= doc && self.opt_scorer.seek(doc) == doc {
             score_combiner.update(&mut self.opt_scorer);
         }
         let score = score_combiner.score();
@@ -101,7 +101,10 @@ mod tests {
                 ConstScorer::from(VecDocSet::from(vec![])),
             );
         let mut docs = vec![];
-        reqoptscorer.for_each(&mut |doc, _| docs.push(doc));
+        while reqoptscorer.doc() != TERMINATED {
+            docs.push(reqoptscorer.doc());
+            reqoptscorer.advance();
+        }
         assert_eq!(docs, req);
     }
 

src/query/scorer.rs

@@ -1,4 +1,4 @@
-use crate::docset::{DocSet, TERMINATED};
+use crate::docset::DocSet;
 use crate::DocId;
 use crate::Score;
 use downcast_rs::impl_downcast;
@@ -12,41 +12,6 @@ pub trait Scorer: downcast_rs::Downcast + DocSet + 'static {
     ///
     /// This method will perform a bit of computation and is not cached.
     fn score(&mut self) -> Score;
-
-    /// Iterates through all of the document matched by the DocSet
-    /// `DocSet` and push the scored documents to the collector.
-    fn for_each(&mut self, callback: &mut dyn FnMut(DocId, Score)) {
-        let mut doc = self.doc();
-        while doc != TERMINATED {
-            callback(doc, self.score());
-            doc = self.advance();
-        }
-    }
-
-    /// Calls `callback` with all of the `(doc, score)` for which score
-    /// is exceeding a given threshold.
-    ///
-    /// This method is useful for the TopDocs collector.
-    /// For all docsets, the blanket implementation has the benefit
-    /// of prefiltering (doc, score) pairs, avoiding the
-    /// virtual dispatch cost.
-    ///
-    /// More importantly, it makes it possible for scorers to implement
-    /// important optimization (e.g. BlockWAND for union).
-    fn for_each_pruning(
-        &mut self,
-        mut threshold: f32,
-        callback: &mut dyn FnMut(DocId, Score) -> Score,
-    ) {
-        let mut doc = self.doc();
-        while doc != TERMINATED {
-            let score = self.score();
-            if score > threshold {
-                threshold = callback(doc, score);
-            }
-            doc = self.advance();
-        }
-    }
 }
 
 impl_downcast!(Scorer);
@@ -55,11 +20,6 @@ impl Scorer for Box<dyn Scorer> {
     fn score(&mut self) -> Score {
         self.deref_mut().score()
     }
-
-    fn for_each(&mut self, callback: &mut dyn FnMut(DocId, Score)) {
-        let scorer = self.deref_mut();
-        scorer.for_each(callback);
-    }
 }
 
 /// Wraps a `DocSet` and simply returns a constant `Scorer`.

src/query/term_query/mod.rs

@@ -9,13 +9,13 @@ pub use self::term_weight::TermWeight;
 #[cfg(test)]
 mod tests {
 
+    use crate::assert_nearly_equals;
     use crate::collector::TopDocs;
     use crate::docset::DocSet;
+    use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
     use crate::query::{Query, QueryParser, Scorer, TermQuery};
     use crate::schema::{Field, IndexRecordOption, Schema, STRING, TEXT};
-    use crate::tests::assert_nearly_equals;
-    use crate::Index;
-    use crate::Term;
+    use crate::{Term, Index, TERMINATED};
 
     #[test]
     pub fn test_term_query_no_freq() {
@@ -42,6 +42,41 @@ mod tests {
         assert_eq!(term_scorer.score(), 0.28768212);
     }
 
+    #[test]
+    pub fn test_term_query_multiple_of_block_len() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let text_field = schema_builder.add_text_field("text", STRING);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        {
+            // writing the segment
+            let mut index_writer = index.writer_with_num_threads(1, 3_000_000)?;
+            for _ in 0..COMPRESSION_BLOCK_SIZE {
+                let doc = doc!(text_field => "a");
+                index_writer.add_document(doc);
+            }
+            index_writer.commit()?;
+        }
+        let searcher = index.reader()?.searcher();
+        let term_query = TermQuery::new(
+            Term::from_field_text(text_field, "a"),
+            IndexRecordOption::Basic,
+        );
+        let term_weight = term_query.weight(&searcher, true)?;
+        let segment_reader = searcher.segment_reader(0);
+        let mut term_scorer = term_weight.scorer(segment_reader, 1.0f32)?;
+        for i in 0u32..COMPRESSION_BLOCK_SIZE as u32 {
+            assert_eq!(term_scorer.doc(), i);
+            if i == COMPRESSION_BLOCK_SIZE as u32 - 1u32 {
+                assert_eq!(term_scorer.advance(), TERMINATED);
+            } else {
+                assert_eq!(term_scorer.advance(), i + 1);
+            }
+        }
+        assert_eq!(term_scorer.doc(), TERMINATED);
+        Ok(())
+    }
+
     #[test]
     pub fn test_term_weight() {
         let mut schema_builder = Schema::builder();
@@ -69,7 +104,7 @@ mod tests {
                 .unwrap();
             assert_eq!(topdocs.len(), 1);
             let (score, _) = topdocs[0];
-            assert_nearly_equals(0.77802235, score);
+            assert_nearly_equals!(0.77802235, score);
         }
         {
             let term = Term::from_field_text(left_field, "left1");
@@ -79,9 +114,9 @@ mod tests {
                 .unwrap();
             assert_eq!(top_docs.len(), 2);
             let (score1, _) = top_docs[0];
-            assert_nearly_equals(0.27101856, score1);
+            assert_nearly_equals!(0.27101856, score1);
             let (score2, _) = top_docs[1];
-            assert_nearly_equals(0.13736556, score2);
+            assert_nearly_equals!(0.13736556, score2);
         }
         {
             let query_parser = QueryParser::for_index(&index, vec![]);
@@ -89,9 +124,9 @@ mod tests {
             let top_docs = searcher.search(&query, &TopDocs::with_limit(2)).unwrap();
             assert_eq!(top_docs.len(), 2);
             let (score1, _) = top_docs[0];
-            assert_nearly_equals(0.9153879, score1);
+            assert_nearly_equals!(0.9153879, score1);
             let (score2, _) = top_docs[1];
-            assert_nearly_equals(0.27101856, score2);
+            assert_nearly_equals!(0.27101856, score2);
         }
     }
 
@@ -112,6 +147,27 @@ mod tests {
         assert_eq!(term_query.count(&*reader.searcher()).unwrap(), 1);
     }
 
+    #[test]
+    fn test_term_query_simple_seek() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let text_field = schema_builder.add_text_field("text", TEXT);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        index_writer.add_document(doc!(text_field=>"a"));
+        index_writer.add_document(doc!(text_field=>"a"));
+        index_writer.commit()?;
+        let term_a = Term::from_field_text(text_field, "a");
+        let term_query = TermQuery::new(term_a, IndexRecordOption::Basic);
+        let searcher = index.reader()?.searcher();
+        let term_weight = term_query.weight(&searcher, false)?;
+        let mut term_scorer = term_weight.scorer(searcher.segment_reader(0u32), 1.0f32)?;
+        assert_eq!(term_scorer.doc(), 0u32);
+        term_scorer.seek(1u32);
+        assert_eq!(term_scorer.doc(), 1u32);
+        Ok(())
+    }
+
     #[test]
     fn test_term_query_debug() {
         let term_query = TermQuery::new(

src/query/term_query/term_scorer.rs

@@ -4,8 +4,8 @@ use crate::DocId;
 use crate::Score;
 
 use crate::fieldnorm::FieldNormReader;
-use crate::postings::Postings;
 use crate::postings::SegmentPostings;
+use crate::postings::{FreqReadingOption, Postings};
 use crate::query::bm25::BM25Weight;
 
 pub struct TermScorer {
@@ -26,13 +26,70 @@ impl TermScorer {
             similarity_weight,
         }
     }
-}
 
-impl TermScorer {
+    pub(crate) fn shallow_seek(&mut self, target_doc: DocId) {
+        self.postings.block_cursor.shallow_seek(target_doc)
+    }
+
+
+    #[cfg(test)]
+    pub fn create_for_test(
+        doc_and_tfs: &[(DocId, u32)],
+        fieldnorm_vals: &[u32],
+        similarity_weight: BM25Weight,
+    ) -> crate::Result<TermScorer> {
+        assert!(!doc_and_tfs.is_empty());
+        assert!(doc_and_tfs.len() <= fieldnorm_vals.len());
+
+        let doc_freq = doc_and_tfs.len();
+        let max_doc = doc_and_tfs.last().unwrap().0 + 1;
+        let mut fieldnorms: Vec<u32> = std::iter::repeat(1).take(max_doc as usize).collect();
+
+        for i in 0..doc_freq {
+            let doc = doc_and_tfs[i].0;
+            let fieldnorm = fieldnorm_vals[i];
+            fieldnorms[doc as usize] = fieldnorm;
+        }
+        let fieldnorm_reader = FieldNormReader::from(&fieldnorms[..]);
+
+        let segment_postings =
+            SegmentPostings::create_from_docs_and_tfs(doc_and_tfs, Some(fieldnorm_reader.clone()))?;
+
+        Ok(TermScorer::new(segment_postings, fieldnorm_reader, similarity_weight))
+    }
+
+    /// See `FreqReadingOption`.
+    pub(crate) fn freq_reading_option(&self) -> FreqReadingOption {
+        self.postings.block_cursor.freq_reading_option()
+    }
+
+    /// Returns the maximum score for the current block.
+    ///
+    /// In some rare case, the result may not be exact. In this case a lower value is returned,
+    /// (and may lead us to return a lesser document).
+    ///
+    /// At index time, we store the (fieldnorm_id, term frequency) pair that maximizes the
+    /// score assuming the average fieldnorm computed on this segment.
+    ///
+    /// Though extremely rare, it is theoretically possible that the actual average fieldnorm
+    /// is different enough from the current segment average fieldnorm that the maximum over a
+    /// specific is achieved on a different document.
+    ///
+    /// (The result is on the other hand guaranteed to be correct if there is only one segment).
+    pub fn block_max_score(&mut self) -> Score {
+        self.postings
+            .block_cursor
+            .block_max_score(&self.fieldnorm_reader, &self.similarity_weight)
+    }
+
     pub fn term_freq(&self) -> u32 {
         self.postings.term_freq()
     }
 
+    pub fn doc_freq(&self) -> usize {
+        self.postings.doc_freq() as usize
+    }
+
     pub fn fieldnorm_id(&self) -> u8 {
         self.fieldnorm_reader.fieldnorm_id(self.doc())
     }
@@ -42,6 +99,14 @@ impl TermScorer {
         let term_freq = self.term_freq();
         self.similarity_weight.explain(fieldnorm_id, term_freq)
     }
+
+    pub fn max_score(&self) -> f32 {
+        self.similarity_weight.max_score()
+    }
+
+    pub fn last_doc_in_block(&self) -> DocId {
+        self.postings.block_cursor.skip_reader.last_doc_in_block()
+    }
 }
 
 impl DocSet for TermScorer {
@@ -69,3 +134,99 @@ impl Scorer for TermScorer {
         self.similarity_weight.score(fieldnorm_id, term_freq)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::assert_nearly_equals;
+    use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+    use crate::query::term_query::TermScorer;
+    use crate::query::{BM25Weight, Scorer};
+    use crate::{DocId, DocSet, TERMINATED};
+    use proptest::prelude::*;
+
+    #[test]
+    fn test_term_scorer_max_score() -> crate::Result<()> {
+        let bm25_weight = BM25Weight::for_one_term(3, 6, 10f32);
+        let mut term_scorer =
+            TermScorer::create_for_test(&[(2, 3), (3, 12), (7, 8)], &[10, 12, 100], bm25_weight)?;
+        let max_scorer = term_scorer.max_score();
+        assert_eq!(max_scorer, 1.3990127f32);
+        assert_eq!(term_scorer.doc(), 2);
+        assert_eq!(term_scorer.term_freq(), 3);
+        assert_nearly_equals!(term_scorer.block_max_score(), 1.3676447f32);
+        assert_nearly_equals!(term_scorer.score(), 1.0892314f32);
+        assert_eq!(term_scorer.advance(), 3);
+        assert_eq!(term_scorer.doc(), 3);
+        assert_eq!(term_scorer.term_freq(), 12);
+        assert_nearly_equals!(term_scorer.score(), 1.3676447f32);
+        assert_eq!(term_scorer.advance(), 7);
+        assert_eq!(term_scorer.doc(), 7);
+        assert_eq!(term_scorer.term_freq(), 8);
+        assert_nearly_equals!(term_scorer.score(), 0.72015285f32);
+        assert_eq!(term_scorer.advance(), TERMINATED);
+        Ok(())
+    }
+
+
+    #[test]
+    fn test_term_scorer_shallow_advance() -> crate::Result<()> {
+        let bm25_weight = BM25Weight::for_one_term(300, 1024, 10f32);
+        let mut doc_and_tfs = vec![];
+        for i in 0u32..300u32 {
+            let doc = i * 10;
+            doc_and_tfs.push((doc, 1u32 + doc % 3u32));
+        }
+        let fieldnorms: Vec<u32> = std::iter::repeat(10u32).take(1024).collect();
+        let mut term_scorer =
+            TermScorer::create_for_test(&doc_and_tfs, &fieldnorms, bm25_weight)?;
+        assert_eq!(term_scorer.doc(), 0u32);
+        term_scorer.shallow_seek(1289);
+        assert_eq!(term_scorer.doc(), 0u32);
+        term_scorer.seek(1289);
+        assert_eq!(term_scorer.doc(), 1290);
+        Ok(())
+    }
+
+    proptest! {
+        #[test]
+        fn test_term_scorer_block_max_score(term_freqs_fieldnorms in proptest::collection::vec((1u32..10u32, 0u32..100u32), 80..300)) {
+        let term_doc_freq = term_freqs_fieldnorms.len();
+         let doc_tfs: Vec<(u32, u32)> = term_freqs_fieldnorms.iter()
+                   .cloned()
+                  .enumerate()
+                  .map(|(doc, (tf, _))| (doc as u32, tf))
+                  .collect();
+
+         let mut fieldnorms: Vec<u32> = vec![];
+         for i in 0..term_doc_freq {
+             let (tf, num_extra_terms) = term_freqs_fieldnorms[i];
+             fieldnorms.push(tf + num_extra_terms);
+         }
+         let average_fieldnorm = fieldnorms
+             .iter()
+             .cloned()
+             .sum::<u32>() as f32 / term_doc_freq as f32;
+             // Average fieldnorm is over the entire index,
+             // not necessarily the docs that are in the posting list.
+             // For this reason we multiply by 1.1 to make a realistic value.
+         let bm25_weight = BM25Weight::for_one_term(term_doc_freq as u64,
+            term_doc_freq as u64 * 10u64,
+            average_fieldnorm);
+
+         let mut term_scorer =
+              TermScorer::create_for_test(&doc_tfs[..], &fieldnorms[..], bm25_weight).unwrap();
+
+         let docs: Vec<DocId> = (0..term_doc_freq).map(|doc| doc as DocId).collect();
+         for block in docs.chunks(COMPRESSION_BLOCK_SIZE) {
+             let block_max_score = term_scorer.block_max_score();
+             let mut block_max_score_computed = 0.0f32;
+             for &doc in block {
+                assert_eq!(term_scorer.doc(), doc);
+                block_max_score_computed = block_max_score_computed.max(term_scorer.score());
+                term_scorer.advance();
+             }
+             assert_nearly_equals!(block_max_score_computed, block_max_score);
+         }
+        }
+    }
+}

src/query/term_query/term_weight.rs

@@ -4,12 +4,13 @@ use crate::docset::DocSet;
 use crate::postings::SegmentPostings;
 use crate::query::bm25::BM25Weight;
 use crate::query::explanation::does_not_match;
+use crate::query::weight::{for_each_pruning_scorer, for_each_scorer};
 use crate::query::Weight;
 use crate::query::{Explanation, Scorer};
 use crate::schema::IndexRecordOption;
-use crate::DocId;
 use crate::Result;
 use crate::Term;
+use crate::{DocId, Score};
 
 pub struct TermWeight {
     term: Term,
@@ -19,12 +20,12 @@ pub struct TermWeight {
 
 impl Weight for TermWeight {
     fn scorer(&self, reader: &SegmentReader, boost: f32) -> Result<Box<dyn Scorer>> {
-        let term_scorer = self.scorer_specialized(reader, boost)?;
+        let term_scorer = self.specialized_scorer(reader, boost)?;
         Ok(Box::new(term_scorer))
     }
 
     fn explain(&self, reader: &SegmentReader, doc: DocId) -> Result<Explanation> {
-        let mut scorer = self.scorer_specialized(reader, 1.0f32)?;
+        let mut scorer = self.specialized_scorer(reader, 1.0f32)?;
         if scorer.seek(doc) != doc {
             return Err(does_not_match(doc));
         }
@@ -43,6 +44,39 @@ impl Weight for TermWeight {
                 .unwrap_or(0))
         }
     }
+
+    /// Iterates through all of the document matched by the DocSet
+    /// `DocSet` and push the scored documents to the collector.
+    fn for_each(
+        &self,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score),
+    ) -> crate::Result<()> {
+        let mut scorer = self.specialized_scorer(reader, 1.0f32)?;
+        for_each_scorer(&mut scorer, callback);
+        Ok(())
+    }
+
+    /// Calls `callback` with all of the `(doc, score)` for which score
+    /// is exceeding a given threshold.
+    ///
+    /// This method is useful for the TopDocs collector.
+    /// For all docsets, the blanket implementation has the benefit
+    /// of prefiltering (doc, score) pairs, avoiding the
+    /// virtual dispatch cost.
+    ///
+    /// More importantly, it makes it possible for scorers to implement
+    /// important optimization (e.g. BlockWAND for union).
+    fn for_each_pruning(
+        &self,
+        threshold: f32,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score) -> Score,
+    ) -> crate::Result<()> {
+        let mut scorer = self.scorer(reader, 1.0f32)?;
+        for_each_pruning_scorer(&mut scorer, threshold, callback);
+        Ok(())
+    }
 }
 
 impl TermWeight {
@@ -58,7 +92,7 @@ impl TermWeight {
         }
     }
 
-    fn scorer_specialized(&self, reader: &SegmentReader, boost: f32) -> Result<TermScorer> {
+    pub fn specialized_scorer(&self, reader: &SegmentReader, boost: f32) -> Result<TermScorer> {
         let field = self.term.field();
         let inverted_index = reader.inverted_index(field);
         let fieldnorm_reader = reader.get_fieldnorms_reader(field);

src/query/union.rs

@@ -183,7 +183,10 @@ where
             // advance all docsets to a doc >= to the target.
             #[cfg_attr(feature = "cargo-clippy", allow(clippy::clippy::collapsible_if))]
             unordered_drain_filter(&mut self.docsets, |docset| {
-                docset.seek(target) == TERMINATED
+                if docset.doc() < target {
+                    docset.seek(target);
+                }
+                docset.doc() == TERMINATED
             });
 
             // at this point all of the docsets
@@ -198,6 +201,18 @@ where
 
     // TODO Also implement `count` with deletes efficiently.
 
+    fn doc(&self) -> DocId {
+        self.doc
+    }
+
+    fn size_hint(&self) -> u32 {
+        self.docsets
+            .iter()
+            .map(|docset| docset.size_hint())
+            .max()
+            .unwrap_or(0u32)
+    }
+
     fn count_including_deleted(&mut self) -> u32 {
         if self.doc == TERMINATED {
             return 0;
@@ -219,14 +234,6 @@ where
         self.cursor = HORIZON_NUM_TINYBITSETS;
         count
     }
-
-    fn doc(&self) -> DocId {
-        self.doc
-    }
-
-    fn size_hint(&self) -> u32 {
-        self.docsets.iter().map(|docset| docset.size_hint()).max().unwrap_or(0u32)
-    }
 }
 
 impl<TScorer, TScoreCombiner> Scorer for Union<TScorer, TScoreCombiner>

src/query/weight.rs

@@ -1,7 +1,45 @@
 use super::Scorer;
 use crate::core::SegmentReader;
 use crate::query::Explanation;
-use crate::DocId;
+use crate::{DocId, Score, TERMINATED};
+
+/// Iterates through all of the document matched by the DocSet
+/// `DocSet` and push the scored documents to the collector.
+pub(crate) fn for_each_scorer<TScorer: Scorer + ?Sized>(
+    scorer: &mut TScorer,
+    callback: &mut dyn FnMut(DocId, Score),
+) {
+    let mut doc = scorer.doc();
+    while doc != TERMINATED {
+        callback(doc, scorer.score());
+        doc = scorer.advance();
+    }
+}
+
+/// Calls `callback` with all of the `(doc, score)` for which score
+/// is exceeding a given threshold.
+///
+/// This method is useful for the TopDocs collector.
+/// For all docsets, the blanket implementation has the benefit
+/// of prefiltering (doc, score) pairs, avoiding the
+/// virtual dispatch cost.
+///
+/// More importantly, it makes it possible for scorers to implement
+/// important optimization (e.g. BlockWAND for union).
+pub(crate) fn for_each_pruning_scorer<TScorer: Scorer + ?Sized>(
+    scorer: &mut TScorer,
+    mut threshold: f32,
+    callback: &mut dyn FnMut(DocId, Score) -> Score,
+) {
+    let mut doc = scorer.doc();
+    while doc != TERMINATED {
+        let score = scorer.score();
+        if score > threshold {
+            threshold = callback(doc, score);
+        }
+        doc = scorer.advance();
+    }
+}
 
 /// A Weight is the specialization of a Query
 /// for a given set of segments.
@@ -27,4 +65,37 @@ pub trait Weight: Send + Sync + 'static {
             Ok(scorer.count_including_deleted())
         }
     }
+
+    /// Iterates through all of the document matched by the DocSet
+    /// `DocSet` and push the scored documents to the collector.
+    fn for_each(
+        &self,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score),
+    ) -> crate::Result<()> {
+        let mut scorer = self.scorer(reader, 1.0f32)?;
+        for_each_scorer(scorer.as_mut(), callback);
+        Ok(())
+    }
+
+    /// Calls `callback` with all of the `(doc, score)` for which score
+    /// is exceeding a given threshold.
+    ///
+    /// This method is useful for the TopDocs collector.
+    /// For all docsets, the blanket implementation has the benefit
+    /// of prefiltering (doc, score) pairs, avoiding the
+    /// virtual dispatch cost.
+    ///
+    /// More importantly, it makes it possible for scorers to implement
+    /// important optimization (e.g. BlockWAND for union).
+    fn for_each_pruning(
+        &self,
+        threshold: f32,
+        reader: &SegmentReader,
+        callback: &mut dyn FnMut(DocId, Score) -> Score,
+    ) -> crate::Result<()> {
+        let mut scorer = self.scorer(reader, 1.0f32)?;
+        for_each_pruning_scorer(scorer.as_mut(), threshold, callback);
+        Ok(())
+    }
 }

src/reader/mod.rs

@@ -22,7 +22,7 @@ pub enum ReloadPolicy {
     /// The index is entirely reloaded manually.
     /// All updates of the index should be manual.
     ///
-    /// No change is reflected automatically. You are required to call `.load_seacher()` manually.
+    /// No change is reflected automatically. You are required to call `IndexReader::reload()` manually.
     Manual,
     /// The index is reloaded within milliseconds after a new commit is available.
     /// This is made possible by watching changes in the `meta.json` file.

src/schema/field_entry.rs

@@ -14,7 +14,7 @@ use std::fmt;
 /// - a field name
 /// - a field type, itself wrapping up options describing
 /// how the field should be indexed.
-#[derive(Clone, Debug, Eq, PartialEq)]
+#[derive(Clone, Debug, PartialEq)]
 pub struct FieldEntry {
     name: String,
     field_type: FieldType,

src/schema/field_type.rs

@@ -48,7 +48,7 @@ pub enum Type {
 
 /// A `FieldType` describes the type (text, u64) of a field as well as
 /// how it should be handled by tantivy.
-#[derive(Clone, Debug, Eq, PartialEq)]
+#[derive(Clone, Debug, PartialEq)]
 pub enum FieldType {
     /// String field type configuration
     Str(TextOptions),

src/schema/text_options.rs

@@ -6,7 +6,7 @@ use std::borrow::Cow;
 use std::ops::BitOr;
 
 /// Define how a text field should be handled by tantivy.
-#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct TextOptions {
     indexing: Option<TextFieldIndexing>,
     stored: bool,
@@ -51,7 +51,7 @@ impl Default for TextOptions {
 /// - the amount of information that should be stored about the presence of a term in a document.
 /// Essentially, should we store the term frequency and/or the positions (See [`IndexRecordOption`](./enum.IndexRecordOption.html)).
 /// - the name of the `Tokenizer` that should be used to process the field.
-#[derive(Clone, PartialEq, Eq, Debug, Serialize, Deserialize)]
+#[derive(Clone, PartialEq, Debug, Serialize, Deserialize)]
 pub struct TextFieldIndexing {
     record: IndexRecordOption,
     tokenizer: Cow<'static, str>,