term_freq in TermFrequencyRecorder untested

PR to demonstrate #2285

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tantivy"
-version = "0.21.0"
+version = "0.22.0-dev"
 authors = ["Paul Masurel <paul.masurel@gmail.com>"]
 license = "MIT"
 categories = ["database-implementations", "data-structures"]
@@ -22,7 +22,7 @@ crc32fast = "1.3.2"
 once_cell = "1.10.0"
 regex = { version = "1.5.5", default-features = false, features = ["std", "unicode"] }
 aho-corasick = "1.0"
-tantivy-fst = "0.4.0"
+tantivy-fst = "0.5"
 memmap2 = { version = "0.9.0", optional = true }
 lz4_flex = { version = "0.11", default-features = false, optional = true }
 zstd = { version = "0.13", optional = true, default-features = false }
@@ -37,21 +37,19 @@ uuid = { version = "1.0.0", features = ["v4", "serde"] }
 crossbeam-channel = "0.5.4"
 rust-stemmers = "1.2.0"
 downcast-rs = "1.2.0"
-bitpacking = { git = "https://github.com/quickwit-oss/bitpacking", rev = "f730b75", default-features = false, features = ["bitpacker4x"] }
+bitpacking = { version = "0.9.2", default-features = false, features = ["bitpacker4x"] }
 census = "0.4.0"
 rustc-hash = "1.1.0"
 thiserror = "1.0.30"
 htmlescape = "0.3.1"
 fail = { version = "0.5.0", optional = true }
-murmurhash32 = "0.3.0"
 time = { version = "0.3.10", features = ["serde-well-known"] }
 smallvec = "1.8.0"
 rayon = "1.5.2"
 lru = "0.12.0"
 fastdivide = "0.4.0"
-itertools = "0.11.0"
+itertools = "0.12.0"
 measure_time = "0.8.2"
-async-trait = "0.1.53"
 arc-swap = "1.5.0"
 
 columnar = { version= "0.2", path="./columnar", package ="tantivy-columnar" }
@@ -75,15 +73,13 @@ matches = "0.1.9"
 pretty_assertions = "1.2.1"
 proptest = "1.0.0"
 test-log = "0.2.10"
-env_logger = "0.10.0"
 futures = "0.3.21"
 paste = "1.0.11"
 more-asserts = "0.3.1"
 rand_distr = "0.4.3"
 
 [target.'cfg(not(windows))'.dev-dependencies]
-criterion = "0.5"
-pprof = { git = "https://github.com/PSeitz/pprof-rs/", rev = "53af24b", features = ["flamegraph", "criterion"] } # temp fork that works with criterion 0.5
+criterion = { version = "0.5", default-features = false }
 
 [dev-dependencies.fail]
 version = "0.5.0"

benches/index-bench.rs

@@ -1,14 +1,99 @@
-use criterion::{criterion_group, criterion_main, Criterion, Throughput};
-use pprof::criterion::{Output, PProfProfiler};
+use criterion::{criterion_group, criterion_main, BatchSize, Bencher, Criterion, Throughput};
 use tantivy::schema::{TantivyDocument, FAST, INDEXED, STORED, STRING, TEXT};
-use tantivy::{Index, IndexWriter};
+use tantivy::{tokenizer, Index, IndexWriter};
 
 const HDFS_LOGS: &str = include_str!("hdfs.json");
 const GH_LOGS: &str = include_str!("gh.json");
 const WIKI: &str = include_str!("wiki.json");
 
-fn get_lines(input: &str) -> Vec<&str> {
-    input.trim().split('\n').collect()
+fn benchmark(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    parse_json: bool,
+    is_dynamic: bool,
+) {
+    if is_dynamic {
+        benchmark_dynamic_json(b, input, schema, commit, parse_json)
+    } else {
+        _benchmark(b, input, schema, commit, parse_json, |schema, doc_json| {
+            TantivyDocument::parse_json(&schema, doc_json).unwrap()
+        })
+    }
+}
+
+fn get_index(schema: tantivy::schema::Schema) -> Index {
+    let mut index = Index::create_in_ram(schema.clone());
+    let ff_tokenizer_manager = tokenizer::TokenizerManager::default();
+    ff_tokenizer_manager.register(
+        "raw",
+        tokenizer::TextAnalyzer::builder(tokenizer::RawTokenizer::default())
+            .filter(tokenizer::RemoveLongFilter::limit(255))
+            .build(),
+    );
+    index.set_fast_field_tokenizers(ff_tokenizer_manager.clone());
+    index
+}
+
+fn _benchmark(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    include_json_parsing: bool,
+    create_doc: impl Fn(&tantivy::schema::Schema, &str) -> TantivyDocument,
+) {
+    if include_json_parsing {
+        let lines: Vec<&str> = input.trim().split('\n').collect();
+        b.iter(|| {
+            let index = get_index(schema.clone());
+            let mut index_writer: IndexWriter =
+                index.writer_with_num_threads(1, 100_000_000).unwrap();
+            for doc_json in &lines {
+                let doc = create_doc(&schema, doc_json);
+                index_writer.add_document(doc).unwrap();
+            }
+            if commit {
+                index_writer.commit().unwrap();
+            }
+        })
+    } else {
+        let docs: Vec<_> = input
+            .trim()
+            .split('\n')
+            .map(|doc_json| create_doc(&schema, doc_json))
+            .collect();
+        b.iter_batched(
+            || docs.clone(),
+            |docs| {
+                let index = get_index(schema.clone());
+                let mut index_writer: IndexWriter =
+                    index.writer_with_num_threads(1, 100_000_000).unwrap();
+                for doc in docs {
+                    index_writer.add_document(doc).unwrap();
+                }
+                if commit {
+                    index_writer.commit().unwrap();
+                }
+            },
+            BatchSize::SmallInput,
+        )
+    }
+}
+fn benchmark_dynamic_json(
+    b: &mut Bencher,
+    input: &str,
+    schema: tantivy::schema::Schema,
+    commit: bool,
+    parse_json: bool,
+) {
+    let json_field = schema.get_field("json").unwrap();
+    _benchmark(b, input, schema, commit, parse_json, |_schema, doc_json| {
+        let json_val: serde_json::Map<String, serde_json::Value> =
+            serde_json::from_str(doc_json).unwrap();
+        tantivy::doc!(json_field=>json_val)
+    })
 }
 
 pub fn hdfs_index_benchmark(c: &mut Criterion) {
@@ -19,7 +104,14 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
         schema_builder.add_text_field("severity", STRING);
         schema_builder.build()
     };
-    let schema_with_store = {
+    let schema_only_fast = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_u64_field("timestamp", FAST);
+        schema_builder.add_text_field("body", FAST);
+        schema_builder.add_text_field("severity", FAST);
+        schema_builder.build()
+    };
+    let _schema_with_store = {
         let mut schema_builder = tantivy::schema::SchemaBuilder::new();
         schema_builder.add_u64_field("timestamp", INDEXED | STORED);
         schema_builder.add_text_field("body", TEXT | STORED);
@@ -28,77 +120,39 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
     };
     let dynamic_schema = {
         let mut schema_builder = tantivy::schema::SchemaBuilder::new();
-        schema_builder.add_json_field("json", TEXT);
+        schema_builder.add_json_field("json", TEXT | FAST);
         schema_builder.build()
     };
 
     let mut group = c.benchmark_group("index-hdfs");
     group.throughput(Throughput::Bytes(HDFS_LOGS.len() as u64));
     group.sample_size(20);
-    group.bench_function("index-hdfs-no-commit", |b| {
-        let lines = get_lines(HDFS_LOGS);
-        b.iter(|| {
-            let index = Index::create_in_ram(schema.clone());
-            let index_writer: IndexWriter = index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let doc = TantivyDocument::parse_json(&schema, doc_json).unwrap();
-                index_writer.add_document(doc).unwrap();
+
+    let benches = [
+        ("only-indexed-".to_string(), schema, false),
+        //("stored-".to_string(), _schema_with_store, false),
+        ("only-fast-".to_string(), schema_only_fast, false),
+        ("dynamic-".to_string(), dynamic_schema, true),
+    ];
+
+    for (prefix, schema, is_dynamic) in benches {
+        for commit in [false, true] {
+            let suffix = if commit { "with-commit" } else { "no-commit" };
+            for parse_json in [false] {
+                // for parse_json in [false, true] {
+                let suffix = if parse_json {
+                    format!("{}-with-json-parsing", suffix)
+                } else {
+                    format!("{}", suffix)
+                };
+
+                let bench_name = format!("{}{}", prefix, suffix);
+                group.bench_function(bench_name, |b| {
+                    benchmark(b, HDFS_LOGS, schema.clone(), commit, parse_json, is_dynamic)
+                });
             }
-        })
-    });
-    group.bench_function("index-hdfs-with-commit", |b| {
-        let lines = get_lines(HDFS_LOGS);
-        b.iter(|| {
-            let index = Index::create_in_ram(schema.clone());
-            let mut index_writer: IndexWriter =
-                index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let doc = TantivyDocument::parse_json(&schema, doc_json).unwrap();
-                index_writer.add_document(doc).unwrap();
-            }
-            index_writer.commit().unwrap();
-        })
-    });
-    group.bench_function("index-hdfs-no-commit-with-docstore", |b| {
-        let lines = get_lines(HDFS_LOGS);
-        b.iter(|| {
-            let index = Index::create_in_ram(schema_with_store.clone());
-            let index_writer: IndexWriter = index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let doc = TantivyDocument::parse_json(&schema, doc_json).unwrap();
-                index_writer.add_document(doc).unwrap();
-            }
-        })
-    });
-    group.bench_function("index-hdfs-with-commit-with-docstore", |b| {
-        let lines = get_lines(HDFS_LOGS);
-        b.iter(|| {
-            let index = Index::create_in_ram(schema_with_store.clone());
-            let mut index_writer: IndexWriter =
-                index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let doc = TantivyDocument::parse_json(&schema, doc_json).unwrap();
-                index_writer.add_document(doc).unwrap();
-            }
-            index_writer.commit().unwrap();
-        })
-    });
-    group.bench_function("index-hdfs-no-commit-json-without-docstore", |b| {
-        let lines = get_lines(HDFS_LOGS);
-        b.iter(|| {
-            let index = Index::create_in_ram(dynamic_schema.clone());
-            let json_field = dynamic_schema.get_field("json").unwrap();
-            let mut index_writer: IndexWriter =
-                index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let json_val: serde_json::Map<String, serde_json::Value> =
-                    serde_json::from_str(doc_json).unwrap();
-                let doc = tantivy::doc!(json_field=>json_val);
-                index_writer.add_document(doc).unwrap();
-            }
-            index_writer.commit().unwrap();
-        })
-    });
+        }
+    }
 }
 
 pub fn gh_index_benchmark(c: &mut Criterion) {
@@ -107,39 +161,24 @@ pub fn gh_index_benchmark(c: &mut Criterion) {
         schema_builder.add_json_field("json", TEXT | FAST);
         schema_builder.build()
     };
+    let dynamic_schema_fast = {
+        let mut schema_builder = tantivy::schema::SchemaBuilder::new();
+        schema_builder.add_json_field("json", FAST);
+        schema_builder.build()
+    };
 
     let mut group = c.benchmark_group("index-gh");
     group.throughput(Throughput::Bytes(GH_LOGS.len() as u64));
 
     group.bench_function("index-gh-no-commit", |b| {
-        let lines = get_lines(GH_LOGS);
-        b.iter(|| {
-            let json_field = dynamic_schema.get_field("json").unwrap();
-            let index = Index::create_in_ram(dynamic_schema.clone());
-            let index_writer: IndexWriter = index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let json_val: serde_json::Map<String, serde_json::Value> =
-                    serde_json::from_str(doc_json).unwrap();
-                let doc = tantivy::doc!(json_field=>json_val);
-                index_writer.add_document(doc).unwrap();
-            }
-        })
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema.clone(), false, false)
     });
-    group.bench_function("index-gh-with-commit", |b| {
-        let lines = get_lines(GH_LOGS);
-        b.iter(|| {
-            let json_field = dynamic_schema.get_field("json").unwrap();
-            let index = Index::create_in_ram(dynamic_schema.clone());
-            let mut index_writer: IndexWriter =
-                index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let json_val: serde_json::Map<String, serde_json::Value> =
-                    serde_json::from_str(doc_json).unwrap();
-                let doc = tantivy::doc!(json_field=>json_val);
-                index_writer.add_document(doc).unwrap();
-            }
-            index_writer.commit().unwrap();
-        })
+    group.bench_function("index-gh-fast", |b| {
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema_fast.clone(), false, false)
+    });
+
+    group.bench_function("index-gh-fast-with-commit", |b| {
+        benchmark_dynamic_json(b, GH_LOGS, dynamic_schema_fast.clone(), true, false)
     });
 }
 
@@ -154,34 +193,10 @@ pub fn wiki_index_benchmark(c: &mut Criterion) {
     group.throughput(Throughput::Bytes(WIKI.len() as u64));
 
     group.bench_function("index-wiki-no-commit", |b| {
-        let lines = get_lines(WIKI);
-        b.iter(|| {
-            let json_field = dynamic_schema.get_field("json").unwrap();
-            let index = Index::create_in_ram(dynamic_schema.clone());
-            let index_writer: IndexWriter = index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let json_val: serde_json::Map<String, serde_json::Value> =
-                    serde_json::from_str(doc_json).unwrap();
-                let doc = tantivy::doc!(json_field=>json_val);
-                index_writer.add_document(doc).unwrap();
-            }
-        })
+        benchmark_dynamic_json(b, WIKI, dynamic_schema.clone(), false, false)
     });
     group.bench_function("index-wiki-with-commit", |b| {
-        let lines = get_lines(WIKI);
-        b.iter(|| {
-            let json_field = dynamic_schema.get_field("json").unwrap();
-            let index = Index::create_in_ram(dynamic_schema.clone());
-            let mut index_writer: IndexWriter =
-                index.writer_with_num_threads(1, 100_000_000).unwrap();
-            for doc_json in &lines {
-                let json_val: serde_json::Map<String, serde_json::Value> =
-                    serde_json::from_str(doc_json).unwrap();
-                let doc = tantivy::doc!(json_field=>json_val);
-                index_writer.add_document(doc).unwrap();
-            }
-            index_writer.commit().unwrap();
-        })
+        benchmark_dynamic_json(b, WIKI, dynamic_schema.clone(), true, false)
     });
 }
 
@@ -192,12 +207,12 @@ criterion_group! {
 }
 criterion_group! {
     name = gh_benches;
-    config = Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None)));
+    config = Criterion::default();
     targets = gh_index_benchmark
 }
 criterion_group! {
     name = wiki_benches;
-    config = Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None)));
+    config = Criterion::default();
     targets = wiki_index_benchmark
 }
 criterion_main!(benches, gh_benches, wiki_benches);

bitpacker/Cargo.toml

@@ -15,7 +15,7 @@ homepage = "https://github.com/quickwit-oss/tantivy"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-bitpacking = {version="0.8", default-features=false, features = ["bitpacker1x"]}
+bitpacking = { version = "0.9.2", default-features = false, features = ["bitpacker1x"] }
 
 [dev-dependencies]
 rand = "0.8"

columnar/Cargo.toml

@@ -9,8 +9,7 @@ description = "column oriented storage for tantivy"
 categories = ["database-implementations", "data-structures", "compression"]
 
 [dependencies]
-itertools = "0.11.0"
-fnv = "1.0.7"
+itertools = "0.12.0"
 fastdivide = "0.4.0"
 
 stacker = { version= "0.2", path = "../stacker", package="tantivy-stacker"}

columnar/columnar-cli/Cargo.toml

@@ -8,7 +8,6 @@ license = "MIT"
 columnar = {path="../", package="tantivy-columnar"}
 serde_json = "1"
 serde_json_borrow = {git="https://github.com/PSeitz/serde_json_borrow/"}
-serde = "1"
 
 [workspace]
 members = []

columnar/src/columnar/writer/column_writers.rs

@@ -269,7 +269,8 @@ impl StrOrBytesColumnWriter {
         dictionaries: &mut [DictionaryBuilder],
         arena: &mut MemoryArena,
     ) {
-        let unordered_id = dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes);
+        let unordered_id =
+            dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes, arena);
         self.column_writer.record(doc, unordered_id, arena);
     }
 

columnar/src/columnar/writer/mod.rs

@@ -437,6 +437,7 @@ impl ColumnarWriter {
                             &mut symbol_byte_buffer,
                         ),
                         buffers,
+                        &self.arena,
                         &mut column_serializer,
                     )?;
                     column_serializer.finalize()?;
@@ -490,6 +491,7 @@ impl ColumnarWriter {
 
 // Serialize [Dictionary, Column, dictionary num bytes U32::LE]
 // Column: [Column Index, Column Values, column index num bytes U32::LE]
+#[allow(clippy::too_many_arguments)]
 fn serialize_bytes_or_str_column(
     cardinality: Cardinality,
     num_docs: RowId,
@@ -497,6 +499,7 @@ fn serialize_bytes_or_str_column(
     dictionary_builder: &DictionaryBuilder,
     operation_it: impl Iterator<Item = ColumnOperation<UnorderedId>>,
     buffers: &mut SpareBuffers,
+    arena: &MemoryArena,
     wrt: impl io::Write,
 ) -> io::Result<()> {
     let SpareBuffers {
@@ -505,7 +508,8 @@ fn serialize_bytes_or_str_column(
         ..
     } = buffers;
     let mut counting_writer = CountingWriter::wrap(wrt);
-    let term_id_mapping: TermIdMapping = dictionary_builder.serialize(&mut counting_writer)?;
+    let term_id_mapping: TermIdMapping =
+        dictionary_builder.serialize(arena, &mut counting_writer)?;
     let dictionary_num_bytes: u32 = counting_writer.written_bytes() as u32;
     let mut wrt = counting_writer.finish();
     let operation_iterator = operation_it.map(|symbol: ColumnOperation<UnorderedId>| {

columnar/src/dictionary.rs

@@ -1,7 +1,7 @@
 use std::io;
 
-use fnv::FnvHashMap;
 use sstable::SSTable;
+use stacker::{MemoryArena, SharedArenaHashMap};
 
 pub(crate) struct TermIdMapping {
     unordered_to_ord: Vec<OrderedId>,
@@ -31,29 +31,38 @@ pub struct OrderedId(pub u32);
 /// mapping.
 #[derive(Default)]
 pub(crate) struct DictionaryBuilder {
-    dict: FnvHashMap<Vec<u8>, UnorderedId>,
-    memory_consumption: usize,
+    dict: SharedArenaHashMap,
 }
 
 impl DictionaryBuilder {
     /// Get or allocate an unordered id.
     /// (This ID is simply an auto-incremented id.)
-    pub fn get_or_allocate_id(&mut self, term: &[u8]) -> UnorderedId {
-        if let Some(term_id) = self.dict.get(term) {
-            return *term_id;
-        }
-        let new_id = UnorderedId(self.dict.len() as u32);
-        self.dict.insert(term.to_vec(), new_id);
-        self.memory_consumption += term.len();
-        self.memory_consumption += 40; // Term Metadata + HashMap overhead
-        new_id
+    pub fn get_or_allocate_id(&mut self, term: &[u8], arena: &mut MemoryArena) -> UnorderedId {
+        let next_id = self.dict.len() as u32;
+        let unordered_id = self
+            .dict
+            .mutate_or_create(term, arena, |unordered_id: Option<u32>| {
+                if let Some(unordered_id) = unordered_id {
+                    unordered_id
+                } else {
+                    next_id
+                }
+            });
+        UnorderedId(unordered_id)
     }
 
     /// Serialize the dictionary into an fst, and returns the
     /// `UnorderedId -> TermOrdinal` map.
-    pub fn serialize<'a, W: io::Write + 'a>(&self, wrt: &mut W) -> io::Result<TermIdMapping> {
-        let mut terms: Vec<(&[u8], UnorderedId)> =
-            self.dict.iter().map(|(k, v)| (k.as_slice(), *v)).collect();
+    pub fn serialize<'a, W: io::Write + 'a>(
+        &self,
+        arena: &MemoryArena,
+        wrt: &mut W,
+    ) -> io::Result<TermIdMapping> {
+        let mut terms: Vec<(&[u8], UnorderedId)> = self
+            .dict
+            .iter(arena)
+            .map(|(k, v)| (k, arena.read(v)))
+            .collect();
         terms.sort_unstable_by_key(|(key, _)| *key);
         // TODO Remove the allocation.
         let mut unordered_to_ord: Vec<OrderedId> = vec![OrderedId(0u32); terms.len()];
@@ -68,7 +77,7 @@ impl DictionaryBuilder {
     }
 
     pub(crate) fn mem_usage(&self) -> usize {
-        self.memory_consumption
+        self.dict.mem_usage()
     }
 }
 
@@ -78,12 +87,13 @@ mod tests {
 
     #[test]
     fn test_dictionary_builder() {
+        let mut arena = MemoryArena::default();
         let mut dictionary_builder = DictionaryBuilder::default();
-        let hello_uid = dictionary_builder.get_or_allocate_id(b"hello");
-        let happy_uid = dictionary_builder.get_or_allocate_id(b"happy");
-        let tax_uid = dictionary_builder.get_or_allocate_id(b"tax");
+        let hello_uid = dictionary_builder.get_or_allocate_id(b"hello", &mut arena);
+        let happy_uid = dictionary_builder.get_or_allocate_id(b"happy", &mut arena);
+        let tax_uid = dictionary_builder.get_or_allocate_id(b"tax", &mut arena);
         let mut buffer = Vec::new();
-        let id_mapping = dictionary_builder.serialize(&mut buffer).unwrap();
+        let id_mapping = dictionary_builder.serialize(&arena, &mut buffer).unwrap();
         assert_eq!(id_mapping.to_ord(hello_uid), OrderedId(1));
         assert_eq!(id_mapping.to_ord(happy_uid), OrderedId(0));
         assert_eq!(id_mapping.to_ord(tax_uid), OrderedId(2));

columnar/src/lib.rs

@@ -1,3 +1,22 @@
+//! # Tantivy-Columnar
+//!
+//! `tantivy-columnar`provides a columnar storage for tantivy.
+//! The crate allows for efficient read operations on specific columns rather than entire records.
+//!
+//! ## Overview
+//!
+//! - **columnar**: Reading, writing, and merging multiple columns:
+//!   - **[ColumnarWriter]**: Makes it possible to create a new columnar.
+//!   - **[ColumnarReader]**: The ColumnarReader makes it possible to access a set of columns
+//!     associated to field names.
+//!   - **[merge_columnar]**: Contains the functionalities to merge multiple ColumnarReader or
+//!     segments into a single one.
+//!
+//! - **column**: A single column, which contains
+//!     - [column_index]: Resolves the rows for a document id. Manages the cardinality of the
+//!       column.
+//!     - [column_values]: Stores the values of a column in a dense format.
+
 #![cfg_attr(all(feature = "unstable", test), feature(test))]
 
 #[cfg(test)]

columnar/src/tests.rs

@@ -26,7 +26,7 @@ fn test_dataframe_writer_str() {
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
     assert_eq!(cols.len(), 1);
-    assert_eq!(cols[0].num_bytes(), 87);
+    assert_eq!(cols[0].num_bytes(), 73);
 }
 
 #[test]
@@ -40,7 +40,7 @@ fn test_dataframe_writer_bytes() {
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
     assert_eq!(cols.len(), 1);
-    assert_eq!(cols[0].num_bytes(), 87);
+    assert_eq!(cols[0].num_bytes(), 73);
 }
 
 #[test]

src/aggregation/agg_tests.rs

@@ -624,6 +624,65 @@ fn test_aggregation_on_json_object() {
     );
 }
 
+#[test]
+fn test_aggregation_on_nested_json_object() {
+    let mut schema_builder = Schema::builder();
+    let json = schema_builder.add_json_field("json.blub", FAST);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema);
+    let mut index_writer: IndexWriter = index.writer_for_tests().unwrap();
+    index_writer
+        .add_document(doc!(json => json!({"color.dot": "red", "color": {"nested":"red"} })))
+        .unwrap();
+    index_writer
+        .add_document(doc!(json => json!({"color.dot": "blue", "color": {"nested":"blue"} })))
+        .unwrap();
+    index_writer.commit().unwrap();
+    let reader = index.reader().unwrap();
+    let searcher = reader.searcher();
+
+    let agg: Aggregations = serde_json::from_value(json!({
+        "jsonagg1": {
+            "terms": {
+                "field": "json\\.blub.color\\.dot",
+            }
+        },
+        "jsonagg2": {
+            "terms": {
+                "field": "json\\.blub.color.nested",
+            }
+        }
+
+    }))
+    .unwrap();
+
+    let aggregation_collector = get_collector(agg);
+    let aggregation_results = searcher.search(&AllQuery, &aggregation_collector).unwrap();
+    let aggregation_res_json = serde_json::to_value(aggregation_results).unwrap();
+    assert_eq!(
+        &aggregation_res_json,
+        &serde_json::json!({
+            "jsonagg1": {
+                "buckets": [
+                    {"doc_count": 1, "key": "blue"},
+                    {"doc_count": 1, "key": "red"}
+                ],
+                "doc_count_error_upper_bound": 0,
+                "sum_other_doc_count": 0
+            },
+            "jsonagg2": {
+                "buckets": [
+                    {"doc_count": 1, "key": "blue"},
+                    {"doc_count": 1, "key": "red"}
+                ],
+                "doc_count_error_upper_bound": 0,
+                "sum_other_doc_count": 0
+            }
+
+        })
+    );
+}
+
 #[test]
 fn test_aggregation_on_json_object_empty_columns() {
     let mut schema_builder = Schema::builder();

src/core/index.rs

@@ -23,6 +23,7 @@ use crate::reader::{IndexReader, IndexReaderBuilder};
 use crate::schema::document::Document;
 use crate::schema::{Field, FieldType, Schema};
 use crate::tokenizer::{TextAnalyzer, TokenizerManager};
+use crate::{merge_field_meta_data, FieldMetadata, SegmentReader};
 
 fn load_metas(
     directory: &dyn Directory,
@@ -489,6 +490,28 @@ impl Index {
         self.inventory.all()
     }
 
+    /// Returns the list of fields that have been indexed in the Index.
+    /// The field list includes the field defined in the schema as well as the fields
+    /// that have been indexed as a part of a JSON field.
+    /// The returned field name is the full field name, including the name of the JSON field.
+    ///
+    /// The returned field names can be used in queries.
+    ///
+    /// Notice: If your data contains JSON fields this is **very expensive**, as it requires
+    /// browsing through the inverted index term dictionary and the columnar field dictionary.
+    ///
+    /// Disclaimer: Some fields may not be listed here. For instance, if the schema contains a json
+    /// field that is not indexed nor a fast field but is stored, it is possible for the field
+    /// to not be listed.
+    pub fn fields_metadata(&self) -> crate::Result<Vec<FieldMetadata>> {
+        let segments = self.searchable_segments()?;
+        let fields_metadata: Vec<Vec<FieldMetadata>> = segments
+            .into_iter()
+            .map(|segment| SegmentReader::open(&segment)?.fields_metadata())
+            .collect::<Result<_, _>>()?;
+        Ok(merge_field_meta_data(fields_metadata, &self.schema()))
+    }
+
     /// Creates a new segment_meta (Advanced user only).
     ///
     /// As long as the `SegmentMeta` lives, the files associated with the

src/core/inverted_index_reader.rs

@@ -75,7 +75,7 @@ impl InvertedIndexReader {
     ///
     /// Notice: This requires a full scan and therefore **very expensive**.
     /// TODO: Move to sstable to use the index.
-    pub fn list_fields(&self) -> io::Result<Vec<(String, Type)>> {
+    pub fn list_encoded_fields(&self) -> io::Result<Vec<(String, Type)>> {
         let mut stream = self.termdict.stream()?;
         let mut fields = Vec::new();
         let mut fields_set = FnvHashSet::default();

src/core/json_utils.rs

@@ -62,6 +62,14 @@ impl IndexingPositionsPerPath {
     }
 }
 
+/// Convert JSON_PATH_SEGMENT_SEP to a dot.
+pub fn json_path_sep_to_dot(path: &mut str) {
+    // This is safe since we are replacing a ASCII character by another ASCII character.
+    unsafe {
+        replace_in_place(JSON_PATH_SEGMENT_SEP, b'.', path.as_bytes_mut());
+    }
+}
+
 #[allow(clippy::too_many_arguments)]
 pub(crate) fn index_json_values<'a, V: Value<'a>>(
     doc: DocId,
@@ -320,7 +328,7 @@ pub struct JsonTermWriter<'a> {
 /// In other words,
 /// - `k8s.node` ends up as `["k8s", "node"]`.
 /// - `k8s\.node` ends up as `["k8s.node"]`.
-fn split_json_path(json_path: &str) -> Vec<String> {
+pub fn split_json_path(json_path: &str) -> Vec<String> {
     let mut escaped_state: bool = false;
     let mut json_path_segments = Vec::new();
     let mut buffer = String::new();

src/core/mod.rs

@@ -25,7 +25,7 @@ pub use self::searcher::{Searcher, SearcherGeneration};
 pub use self::segment::Segment;
 pub use self::segment_component::SegmentComponent;
 pub use self::segment_id::SegmentId;
-pub use self::segment_reader::SegmentReader;
+pub use self::segment_reader::{merge_field_meta_data, FieldMetadata, SegmentReader};
 pub use self::single_segment_index_writer::SingleSegmentIndexWriter;
 
 /// The meta file contains all the information about the list of segments and the schema

src/core/segment_reader.rs

@@ -1,12 +1,17 @@
 use std::collections::HashMap;
+use std::ops::BitOrAssign;
 use std::sync::{Arc, RwLock};
 use std::{fmt, io};
 
+use fnv::FnvHashMap;
+use itertools::Itertools;
+
 use crate::core::{InvertedIndexReader, Segment, SegmentComponent, SegmentId};
 use crate::directory::{CompositeFile, FileSlice};
 use crate::error::DataCorruption;
 use crate::fastfield::{intersect_alive_bitsets, AliveBitSet, FacetReader, FastFieldReaders};
 use crate::fieldnorm::{FieldNormReader, FieldNormReaders};
+use crate::json_utils::json_path_sep_to_dot;
 use crate::schema::{Field, IndexRecordOption, Schema, Type};
 use crate::space_usage::SegmentSpaceUsage;
 use crate::store::StoreReader;
@@ -280,6 +285,103 @@ impl SegmentReader {
         Ok(inv_idx_reader)
     }
 
+    /// Returns the list of fields that have been indexed in the segment.
+    /// The field list includes the field defined in the schema as well as the fields
+    /// that have been indexed as a part of a JSON field.
+    /// The returned field name is the full field name, including the name of the JSON field.
+    ///
+    /// The returned field names can be used in queries.
+    ///
+    /// Notice: If your data contains JSON fields this is **very expensive**, as it requires
+    /// browsing through the inverted index term dictionary and the columnar field dictionary.
+    ///
+    /// Disclaimer: Some fields may not be listed here. For instance, if the schema contains a json
+    /// field that is not indexed nor a fast field but is stored, it is possible for the field
+    /// to not be listed.
+    pub fn fields_metadata(&self) -> crate::Result<Vec<FieldMetadata>> {
+        let mut indexed_fields: Vec<FieldMetadata> = Vec::new();
+        let mut map_to_canonical = FnvHashMap::default();
+        for (field, field_entry) in self.schema().fields() {
+            let field_name = field_entry.name().to_string();
+            let is_indexed = field_entry.is_indexed();
+
+            if is_indexed {
+                let is_json = field_entry.field_type().value_type() == Type::Json;
+                if is_json {
+                    let inv_index = self.inverted_index(field)?;
+                    let encoded_fields_in_index = inv_index.list_encoded_fields()?;
+                    let mut build_path = |field_name: &str, mut json_path: String| {
+                        // In this case we need to map the potential fast field to the field name
+                        // accepted by the query parser.
+                        let create_canonical =
+                            !field_entry.is_expand_dots_enabled() && json_path.contains('.');
+                        if create_canonical {
+                            // Without expand dots enabled dots need to be escaped.
+                            let escaped_json_path = json_path.replace('.', "\\.");
+                            let full_path = format!("{}.{}", field_name, escaped_json_path);
+                            let full_path_unescaped = format!("{}.{}", field_name, &json_path);
+                            map_to_canonical.insert(full_path_unescaped, full_path.to_string());
+                            full_path
+                        } else {
+                            // With expand dots enabled, we can use '.' instead of '\u{1}'.
+                            json_path_sep_to_dot(&mut json_path);
+                            format!("{}.{}", field_name, json_path)
+                        }
+                    };
+                    indexed_fields.extend(
+                        encoded_fields_in_index
+                            .into_iter()
+                            .map(|(name, typ)| (build_path(&field_name, name), typ))
+                            .map(|(field_name, typ)| FieldMetadata {
+                                indexed: true,
+                                stored: false,
+                                field_name,
+                                fast: false,
+                                typ,
+                            }),
+                    );
+                } else {
+                    indexed_fields.push(FieldMetadata {
+                        indexed: true,
+                        stored: false,
+                        field_name: field_name.to_string(),
+                        fast: false,
+                        typ: field_entry.field_type().value_type(),
+                    });
+                }
+            }
+        }
+        let mut fast_fields: Vec<FieldMetadata> = self
+            .fast_fields()
+            .columnar()
+            .iter_columns()?
+            .map(|(mut field_name, handle)| {
+                json_path_sep_to_dot(&mut field_name);
+                // map to canonical path, to avoid similar but different entries.
+                // Eventually we should just accept '.' seperated for all cases.
+                let field_name = map_to_canonical
+                    .get(&field_name)
+                    .unwrap_or(&field_name)
+                    .to_string();
+                FieldMetadata {
+                    indexed: false,
+                    stored: false,
+                    field_name,
+                    fast: true,
+                    typ: Type::from(handle.column_type()),
+                }
+            })
+            .collect();
+        // Since the type is encoded differently in the fast field and in the inverted index,
+        // the order of the fields is not guaranteed to be the same. Therefore, we sort the fields.
+        // If we are sure that the order is the same, we can remove this sort.
+        indexed_fields.sort_unstable();
+        fast_fields.sort_unstable();
+        let merged = merge_field_meta_data(vec![indexed_fields, fast_fields], &self.schema);
+
+        Ok(merged)
+    }
+
     /// Returns the segment id
     pub fn segment_id(&self) -> SegmentId {
         self.segment_id
@@ -330,6 +432,65 @@ impl SegmentReader {
     }
 }
 
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
+/// FieldMetadata
+pub struct FieldMetadata {
+    /// The field name
+    // Notice: Don't reorder the declaration of 1.field_name 2.typ, as it is used for ordering by
+    // field_name then typ.
+    pub field_name: String,
+    /// The field type
+    // Notice: Don't reorder the declaration of 1.field_name 2.typ, as it is used for ordering by
+    // field_name then typ.
+    pub typ: Type,
+    /// Is the field indexed for search
+    pub indexed: bool,
+    /// Is the field stored in the doc store
+    pub stored: bool,
+    /// Is the field stored in the columnar storage
+    pub fast: bool,
+}
+impl BitOrAssign for FieldMetadata {
+    fn bitor_assign(&mut self, rhs: Self) {
+        assert!(self.field_name == rhs.field_name);
+        assert!(self.typ == rhs.typ);
+        self.indexed |= rhs.indexed;
+        self.stored |= rhs.stored;
+        self.fast |= rhs.fast;
+    }
+}
+
+// Maybe too slow for the high cardinality case
+fn is_field_stored(field_name: &str, schema: &Schema) -> bool {
+    schema
+        .find_field(field_name)
+        .map(|(field, _path)| schema.get_field_entry(field).is_stored())
+        .unwrap_or(false)
+}
+
+/// Helper to merge the field metadata from multiple segments.
+pub fn merge_field_meta_data(
+    field_metadatas: Vec<Vec<FieldMetadata>>,
+    schema: &Schema,
+) -> Vec<FieldMetadata> {
+    let mut merged_field_metadata = Vec::new();
+    for (_key, mut group) in &field_metadatas
+        .into_iter()
+        .kmerge_by(|left, right| left < right)
+        // TODO: Remove allocation
+        .group_by(|el| (el.field_name.to_string(), el.typ))
+    {
+        let mut merged: FieldMetadata = group.next().unwrap();
+        for el in group {
+            merged |= el;
+        }
+        // Currently is_field_stored is maybe too slow for the high cardinality case
+        merged.stored = is_field_stored(&merged.field_name, schema);
+        merged_field_metadata.push(merged);
+    }
+    merged_field_metadata
+}
+
 fn intersect_alive_bitset(
     left_opt: Option<AliveBitSet>,
     right_opt: Option<AliveBitSet>,
@@ -353,9 +514,127 @@ impl fmt::Debug for SegmentReader {
 
 #[cfg(test)]
 mod test {
+    use super::*;
     use crate::core::Index;
-    use crate::schema::{Schema, Term, STORED, TEXT};
-    use crate::{DocId, IndexWriter};
+    use crate::schema::{Schema, SchemaBuilder, Term, STORED, TEXT};
+    use crate::{DocId, FieldMetadata, IndexWriter};
+
+    #[test]
+    fn test_merge_field_meta_data_same() {
+        let schema = SchemaBuilder::new().build();
+        let field_metadata1 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: true,
+        };
+        let field_metadata2 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: true,
+        };
+        let res = merge_field_meta_data(
+            vec![vec![field_metadata1.clone()], vec![field_metadata2]],
+            &schema,
+        );
+        assert_eq!(res, vec![field_metadata1]);
+    }
+    #[test]
+    fn test_merge_field_meta_data_different() {
+        let schema = SchemaBuilder::new().build();
+        let field_metadata1 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: false,
+            stored: false,
+            fast: true,
+        };
+        let field_metadata2 = FieldMetadata {
+            field_name: "b".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: false,
+            stored: false,
+            fast: true,
+        };
+        let field_metadata3 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: false,
+        };
+        let res = merge_field_meta_data(
+            vec![
+                vec![field_metadata1.clone(), field_metadata2.clone()],
+                vec![field_metadata3],
+            ],
+            &schema,
+        );
+        let field_metadata_expected1 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: true,
+        };
+        assert_eq!(res, vec![field_metadata_expected1, field_metadata2.clone()]);
+    }
+    #[test]
+    fn test_merge_field_meta_data_merge() {
+        use pretty_assertions::assert_eq;
+        let get_meta_data = |name: &str, typ: Type| FieldMetadata {
+            field_name: name.to_string(),
+            typ,
+            indexed: false,
+            stored: false,
+            fast: true,
+        };
+        let schema = SchemaBuilder::new().build();
+        let mut metas = vec![get_meta_data("d", Type::Str), get_meta_data("e", Type::U64)];
+        metas.sort();
+        let res = merge_field_meta_data(vec![vec![get_meta_data("e", Type::Str)], metas], &schema);
+        assert_eq!(
+            res,
+            vec![
+                get_meta_data("d", Type::Str),
+                get_meta_data("e", Type::Str),
+                get_meta_data("e", Type::U64),
+            ]
+        );
+    }
+    #[test]
+    fn test_merge_field_meta_data_bitxor() {
+        let field_metadata1 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: false,
+            stored: false,
+            fast: true,
+        };
+        let field_metadata2 = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: false,
+        };
+        let field_metadata_expected = FieldMetadata {
+            field_name: "a".to_string(),
+            typ: crate::schema::Type::Str,
+            indexed: true,
+            stored: false,
+            fast: true,
+        };
+        let mut res1 = field_metadata1.clone();
+        res1 |= field_metadata2.clone();
+        let mut res2 = field_metadata2.clone();
+        res2 |= field_metadata1;
+        assert_eq!(res1, field_metadata_expected);
+        assert_eq!(res2, field_metadata_expected);
+    }
 
     #[test]
     fn test_num_alive() -> crate::Result<()> {

src/core/tests.rs

@@ -1,12 +1,13 @@
 use crate::collector::Count;
 use crate::directory::{RamDirectory, WatchCallback};
-use crate::indexer::NoMergePolicy;
+use crate::indexer::{LogMergePolicy, NoMergePolicy};
+use crate::json_utils::JsonTermWriter;
 use crate::query::TermQuery;
-use crate::schema::{Field, IndexRecordOption, Schema, INDEXED, STRING, TEXT};
+use crate::schema::{Field, IndexRecordOption, Schema, Type, INDEXED, STRING, TEXT};
 use crate::tokenizer::TokenizerManager;
 use crate::{
-    Directory, Index, IndexBuilder, IndexReader, IndexSettings, IndexWriter, ReloadPolicy,
-    SegmentId, TantivyDocument, Term,
+    Directory, DocSet, Index, IndexBuilder, IndexReader, IndexSettings, IndexWriter, Postings,
+    ReloadPolicy, SegmentId, TantivyDocument, Term,
 };
 
 #[test]
@@ -344,3 +345,132 @@ fn test_merging_segment_update_docfreq() {
     let term_info = inv_index.get_term_info(&term).unwrap().unwrap();
     assert_eq!(term_info.doc_freq, 12);
 }
+
+// motivated by https://github.com/quickwit-oss/quickwit/issues/4130
+#[test]
+fn test_positions_merge_bug_non_text_json_vint() {
+    let mut schema_builder = Schema::builder();
+    let field = schema_builder.add_json_field("dynamic", TEXT);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema.clone());
+    let mut writer: IndexWriter = index.writer_for_tests().unwrap();
+    let mut merge_policy = LogMergePolicy::default();
+    merge_policy.set_min_num_segments(2);
+    writer.set_merge_policy(Box::new(merge_policy));
+    // Here a string would work.
+    let doc_json = r#"{"tenant_id":75}"#;
+    let vals = serde_json::from_str(doc_json).unwrap();
+    let mut doc = TantivyDocument::default();
+    doc.add_object(field, vals);
+    writer.add_document(doc.clone()).unwrap();
+    writer.commit().unwrap();
+    writer.add_document(doc.clone()).unwrap();
+    writer.commit().unwrap();
+    writer.wait_merging_threads().unwrap();
+    let reader = index.reader().unwrap();
+    assert_eq!(reader.searcher().segment_readers().len(), 1);
+}
+
+// Same as above but with bitpacked blocks
+#[test]
+fn test_positions_merge_bug_non_text_json_bitpacked_block() {
+    let mut schema_builder = Schema::builder();
+    let field = schema_builder.add_json_field("dynamic", TEXT);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema.clone());
+    let mut writer: IndexWriter = index.writer_for_tests().unwrap();
+    let mut merge_policy = LogMergePolicy::default();
+    merge_policy.set_min_num_segments(2);
+    writer.set_merge_policy(Box::new(merge_policy));
+    // Here a string would work.
+    let doc_json = r#"{"tenant_id":75}"#;
+    let vals = serde_json::from_str(doc_json).unwrap();
+    let mut doc = TantivyDocument::default();
+    doc.add_object(field, vals);
+    for _ in 0..128 {
+        writer.add_document(doc.clone()).unwrap();
+    }
+    writer.commit().unwrap();
+    writer.add_document(doc.clone()).unwrap();
+    writer.commit().unwrap();
+    writer.wait_merging_threads().unwrap();
+    let reader = index.reader().unwrap();
+    assert_eq!(reader.searcher().segment_readers().len(), 1);
+}
+
+#[test]
+fn test_non_text_json_term_freq() {
+    let mut schema_builder = Schema::builder();
+    let field = schema_builder.add_json_field("dynamic", TEXT);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema.clone());
+    let mut writer: IndexWriter = index.writer_for_tests().unwrap();
+    // Here a string would work.
+    let doc_json = r#"{"tenant_id":75}"#;
+    let vals = serde_json::from_str(doc_json).unwrap();
+    let mut doc = TantivyDocument::default();
+    doc.add_object(field, vals);
+    writer.add_document(doc.clone()).unwrap();
+    writer.commit().unwrap();
+    let reader = index.reader().unwrap();
+    assert_eq!(reader.searcher().segment_readers().len(), 1);
+    let searcher = reader.searcher();
+    let segment_reader = searcher.segment_reader(0u32);
+    let inv_idx = segment_reader.inverted_index(field).unwrap();
+    let mut term = Term::with_type_and_field(Type::Json, field);
+    let mut json_term_writer = JsonTermWriter::wrap(&mut term, false);
+    json_term_writer.push_path_segment("tenant_id");
+    json_term_writer.close_path_and_set_type(Type::U64);
+    json_term_writer.set_fast_value(75u64);
+    let postings = inv_idx
+        .read_postings(
+            &json_term_writer.term(),
+            IndexRecordOption::WithFreqsAndPositions,
+        )
+        .unwrap()
+        .unwrap();
+    assert_eq!(postings.doc(), 0);
+    assert_eq!(postings.term_freq(), 1u32);
+}
+
+#[test]
+fn test_non_text_json_term_freq_bitpacked() {
+    let mut schema_builder = Schema::builder();
+    let field = schema_builder.add_json_field("dynamic", TEXT);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema.clone());
+    let mut writer: IndexWriter = index.writer_for_tests().unwrap();
+    // Here a string would work.
+    let doc_json = r#"{"tenant_id":75}"#;
+    let vals = serde_json::from_str(doc_json).unwrap();
+    let mut doc = TantivyDocument::default();
+    doc.add_object(field, vals);
+    let num_docs = 132;
+    for _ in 0..num_docs {
+        writer.add_document(doc.clone()).unwrap();
+    }
+    writer.commit().unwrap();
+    let reader = index.reader().unwrap();
+    assert_eq!(reader.searcher().segment_readers().len(), 1);
+    let searcher = reader.searcher();
+    let segment_reader = searcher.segment_reader(0u32);
+    let inv_idx = segment_reader.inverted_index(field).unwrap();
+    let mut term = Term::with_type_and_field(Type::Json, field);
+    let mut json_term_writer = JsonTermWriter::wrap(&mut term, false);
+    json_term_writer.push_path_segment("tenant_id");
+    json_term_writer.close_path_and_set_type(Type::U64);
+    json_term_writer.set_fast_value(75u64);
+    let mut postings = inv_idx
+        .read_postings(
+            &json_term_writer.term(),
+            IndexRecordOption::WithFreqsAndPositions,
+        )
+        .unwrap()
+        .unwrap();
+    assert_eq!(postings.doc(), 0);
+    assert_eq!(postings.term_freq(), 1u32);
+    for i in 1..num_docs {
+        assert_eq!(postings.advance(), i);
+        assert_eq!(postings.term_freq(), 1u32);
+    }
+}

src/fastfield/mod.rs

@@ -131,7 +131,7 @@ mod tests {
         }
         let file = directory.open_read(path).unwrap();
 
-        assert_eq!(file.len(), 93);
+        assert_eq!(file.len(), 80);
         let fast_field_readers = FastFieldReaders::open(file, SCHEMA.clone()).unwrap();
         let column = fast_field_readers
             .u64("field")
@@ -181,7 +181,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 121);
+        assert_eq!(file.len(), 108);
         let fast_field_readers = FastFieldReaders::open(file, SCHEMA.clone()).unwrap();
         let col = fast_field_readers
             .u64("field")
@@ -214,7 +214,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 94);
+        assert_eq!(file.len(), 81);
         let fast_field_readers = FastFieldReaders::open(file, SCHEMA.clone()).unwrap();
         let fast_field_reader = fast_field_readers
             .u64("field")
@@ -246,7 +246,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 4489);
+        assert_eq!(file.len(), 4476);
         {
             let fast_field_readers = FastFieldReaders::open(file, SCHEMA.clone()).unwrap();
             let col = fast_field_readers
@@ -279,7 +279,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 265);
+        assert_eq!(file.len(), 252);
 
         {
             let fast_field_readers = FastFieldReaders::open(file, schema).unwrap();
@@ -773,7 +773,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 102);
+        assert_eq!(file.len(), 84);
         let fast_field_readers = FastFieldReaders::open(file, schema).unwrap();
         let bool_col = fast_field_readers.bool("field_bool").unwrap();
         assert_eq!(bool_col.first(0), Some(true));
@@ -805,7 +805,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 114);
+        assert_eq!(file.len(), 96);
         let readers = FastFieldReaders::open(file, schema).unwrap();
         let bool_col = readers.bool("field_bool").unwrap();
         for i in 0..25 {
@@ -830,7 +830,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 104);
+        assert_eq!(file.len(), 86);
         let fastfield_readers = FastFieldReaders::open(file, schema).unwrap();
         let col = fastfield_readers.bool("field_bool").unwrap();
         assert_eq!(col.first(0), None);
@@ -1288,11 +1288,18 @@ mod tests {
         index_writer.commit().unwrap();
         let searcher = index.reader().unwrap().searcher();
         let fast_field_reader = searcher.segment_reader(0u32).fast_fields();
+        // Supported for now, maybe dropped in the future.
         let column = fast_field_reader
             .column_opt::<i64>("jsonfield.attr.age")
             .unwrap()
             .unwrap();
         let vals: Vec<i64> = column.values_for_doc(0u32).collect();
         assert_eq!(&vals, &[33]);
+        let column = fast_field_reader
+            .column_opt::<i64>("jsonfield\\.attr.age")
+            .unwrap()
+            .unwrap();
+        let vals: Vec<i64> = column.values_for_doc(0u32).collect();
+        assert_eq!(&vals, &[33]);
     }
 }

src/indexer/merger.rs

@@ -552,7 +552,41 @@ impl IndexMerger {
                 continue;
             }
 
-            field_serializer.new_term(term_bytes, total_doc_freq)?;
+            // This should never happen as we early exited for total_doc_freq == 0.
+            assert!(!segment_postings_containing_the_term.is_empty());
+
+            let has_term_freq = {
+                let has_term_freq = !segment_postings_containing_the_term[0]
+                    .1
+                    .block_cursor
+                    .freqs()
+                    .is_empty();
+                for (_, postings) in &segment_postings_containing_the_term[1..] {
+                    // This may look at a strange way to test whether we have term freq or not.
+                    // With JSON object, the schema is not sufficient to know whether a term
+                    // has its term frequency encoded or not:
+                    // strings may have term frequencies, while number terms never have one.
+                    //
+                    // Ideally, we should have burnt one bit of two in the `TermInfo`.
+                    // However, we preferred not changing the codec too much and detect this
+                    // instead by
+                    // - looking at the size of the skip data for bitpacked blocks
+                    // - observing the absence of remaining data after reading the docs for vint
+                    // blocks.
+                    //
+                    // Overall the reliable way to know if we have actual frequencies loaded or not
+                    // is to check whether the actual decoded array is empty or not.
+                    if has_term_freq != !postings.block_cursor.freqs().is_empty() {
+                        return Err(DataCorruption::comment_only(
+                            "Term freqs are inconsistent across segments",
+                        )
+                        .into());
+                    }
+                }
+                has_term_freq
+            };
+
+            field_serializer.new_term(term_bytes, total_doc_freq, has_term_freq)?;
 
             // We can now serialize this postings, by pushing each document to the
             // postings serializer.
@@ -567,8 +601,13 @@ impl IndexMerger {
                     if let Some(remapped_doc_id) = old_to_new_doc_id[doc as usize] {
                         // we make sure to only write the term if
                         // there is at least one document.
-                        let term_freq = segment_postings.term_freq();
-                        segment_postings.positions(&mut positions_buffer);
+                        let term_freq = if has_term_freq {
+                            segment_postings.positions(&mut positions_buffer);
+                            segment_postings.term_freq()
+                        } else {
+                            0u32
+                        };
+
                         // if doc_id_mapping exists, the doc_ids are reordered, they are
                         // not just stacked. The field serializer expects monotonically increasing
                         // doc_ids, so we collect and sort them first, before writing.

src/indexer/mod.rs

@@ -59,10 +59,13 @@ type AddBatchReceiver<D> = channel::Receiver<AddBatch<D>>;
 #[cfg(test)]
 mod tests_mmap {
 
-    use crate::collector::Count;
-    use crate::query::QueryParser;
-    use crate::schema::{JsonObjectOptions, Schema, Type, TEXT};
-    use crate::{Index, IndexWriter, Term};
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::agg_result::AggregationResults;
+    use crate::aggregation::AggregationCollector;
+    use crate::collector::{Count, TopDocs};
+    use crate::query::{AllQuery, QueryParser};
+    use crate::schema::{JsonObjectOptions, Schema, Type, FAST, INDEXED, STORED, TEXT};
+    use crate::{FieldMetadata, Index, IndexWriter, Term};
 
     #[test]
     fn test_advance_delete_bug() -> crate::Result<()> {
@@ -173,8 +176,7 @@ mod tests_mmap {
     #[test]
     fn test_json_field_list_fields() {
         let mut schema_builder = Schema::builder();
-        let json_options: JsonObjectOptions =
-            JsonObjectOptions::from(TEXT).set_expand_dots_enabled();
+        let json_options: JsonObjectOptions = JsonObjectOptions::from(TEXT);
         let json_field = schema_builder.add_json_field("json", json_options);
         let index = Index::create_in_ram(schema_builder.build());
         let mut index_writer = index.writer_for_tests().unwrap();
@@ -193,9 +195,9 @@ mod tests_mmap {
         let reader = &searcher.segment_readers()[0];
         let inverted_index = reader.inverted_index(json_field).unwrap();
         assert_eq!(
-            inverted_index.list_fields().unwrap(),
+            inverted_index.list_encoded_fields().unwrap(),
             [
-                ("k8s\u{1}container\u{1}name".to_string(), Type::Str),
+                ("k8s.container.name".to_string(), Type::Str),
                 ("sub\u{1}a".to_string(), Type::I64),
                 ("sub\u{1}b".to_string(), Type::I64),
                 ("suber\u{1}a".to_string(), Type::I64),
@@ -205,4 +207,240 @@ mod tests_mmap {
             ]
         );
     }
+
+    #[test]
+    fn test_json_fields_metadata_expanded_dots_one_segment() {
+        test_json_fields_metadata(true, true);
+    }
+    #[test]
+    fn test_json_fields_metadata_expanded_dots_multi_segment() {
+        test_json_fields_metadata(true, false);
+    }
+    #[test]
+    fn test_json_fields_metadata_no_expanded_dots_one_segment() {
+        test_json_fields_metadata(false, true);
+    }
+    #[test]
+    fn test_json_fields_metadata_no_expanded_dots_multi_segment() {
+        test_json_fields_metadata(false, false);
+    }
+
+    fn test_json_fields_metadata(expanded_dots: bool, one_segment: bool) {
+        use pretty_assertions::assert_eq;
+        let mut schema_builder = Schema::builder();
+        let json_options: JsonObjectOptions =
+            JsonObjectOptions::from(TEXT).set_fast(None).set_stored();
+        let json_options = if expanded_dots {
+            json_options.set_expand_dots_enabled()
+        } else {
+            json_options
+        };
+        schema_builder.add_json_field("json.confusing", json_options.clone());
+        let json_field = schema_builder.add_json_field("json.shadow", json_options.clone());
+        let json_field2 = schema_builder.add_json_field("json", json_options.clone());
+        schema_builder.add_json_field("empty_json", json_options);
+        let number_field = schema_builder.add_u64_field("numbers", FAST);
+        schema_builder.add_u64_field("empty", FAST | INDEXED | STORED);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer = index.writer_for_tests().unwrap();
+        let json =
+            serde_json::json!({"k8s.container.name": "a", "val": "a", "sub": {"a": 1, "b": 1}});
+        index_writer.add_document(doc!(json_field=>json)).unwrap();
+        let json =
+            serde_json::json!({"k8s.container.name": "a", "val": "a", "suber": {"a": 1, "b": 1}});
+        if !one_segment {
+            index_writer.commit().unwrap();
+        }
+        index_writer.add_document(doc!(json_field=>json)).unwrap();
+        let json = serde_json::json!({"k8s.container.name": "a", "k8s.container.name": "a", "val": "a", "suber": {"a": "a", "b": 1}});
+        index_writer
+            .add_document(doc!(number_field => 50u64, json_field=>json, json_field2=>json!({"shadow": {"val": "a"}})))
+            .unwrap();
+        index_writer.commit().unwrap();
+        let reader = index.reader().unwrap();
+
+        let searcher = reader.searcher();
+        assert_eq!(searcher.num_docs(), 3);
+
+        let fields_metadata = index.fields_metadata().unwrap();
+        assert_eq!(
+            fields_metadata,
+            [
+                FieldMetadata {
+                    field_name: "empty".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::U64
+                },
+                FieldMetadata {
+                    field_name: if expanded_dots {
+                        "json.shadow.k8s.container.name".to_string()
+                    } else {
+                        "json.shadow.k8s\\.container\\.name".to_string()
+                    },
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::Str
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.sub.a".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::I64
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.sub.b".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::I64
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.suber.a".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::I64
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.suber.a".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::Str
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.suber.b".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::I64
+                },
+                FieldMetadata {
+                    field_name: "json.shadow.val".to_string(),
+                    indexed: true,
+                    stored: true,
+                    fast: true,
+                    typ: Type::Str
+                },
+                FieldMetadata {
+                    field_name: "numbers".to_string(),
+                    indexed: false,
+                    stored: false,
+                    fast: true,
+                    typ: Type::U64
+                }
+            ]
+        );
+        let query_parser = QueryParser::for_index(&index, vec![]);
+        // Test if returned field name can be queried
+        for indexed_field in fields_metadata.iter().filter(|meta| meta.indexed) {
+            let val = if indexed_field.typ == Type::Str {
+                "a"
+            } else {
+                "1"
+            };
+            let query_str = &format!("{}:{}", indexed_field.field_name, val);
+            let query = query_parser.parse_query(query_str).unwrap();
+            let count_docs = searcher.search(&*query, &TopDocs::with_limit(2)).unwrap();
+            if indexed_field.field_name.contains("empty") || indexed_field.typ == Type::Json {
+                assert_eq!(count_docs.len(), 0);
+            } else {
+                assert!(!count_docs.is_empty(), "{}", indexed_field.field_name);
+            }
+        }
+        // Test if returned field name can be used for aggregation
+        for fast_field in fields_metadata.iter().filter(|meta| meta.fast) {
+            let agg_req_str = json!(
+            {
+              "termagg": {
+                "terms": {
+                  "field": fast_field.field_name,
+                }
+              }
+            });
+
+            let agg_req: Aggregations = serde_json::from_value(agg_req_str).unwrap();
+            let collector = AggregationCollector::from_aggs(agg_req, Default::default());
+            let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
+            let res = serde_json::to_value(agg_res).unwrap();
+            if !fast_field.field_name.contains("empty") && fast_field.typ != Type::Json {
+                assert!(
+                    !res["termagg"]["buckets"].as_array().unwrap().is_empty(),
+                    "{}",
+                    fast_field.field_name
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_json_field_shadowing_field_name_bug() {
+        /// This test is only there to display a bug on addressing a field if it gets shadowed
+        /// The issues only occurs if the field name that shadows contains a dot.
+        ///
+        /// Happens independently of the `expand_dots` option. Since that option does not
+        /// affect the field name itself.
+        use pretty_assertions::assert_eq;
+        let mut schema_builder = Schema::builder();
+        let json_options: JsonObjectOptions =
+            JsonObjectOptions::from(TEXT).set_fast(None).set_stored();
+        // let json_options = json_options.set_expand_dots_enabled();
+        let json_field_shadow = schema_builder.add_json_field("json.shadow", json_options.clone());
+        let json_field = schema_builder.add_json_field("json", json_options.clone());
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer = index.writer_for_tests().unwrap();
+        index_writer
+            .add_document(
+                doc!(json_field_shadow=>json!({"val": "b"}), json_field=>json!({"shadow": {"val": "a"}})),
+            )
+            .unwrap();
+        index_writer.commit().unwrap();
+        let reader = index.reader().unwrap();
+
+        let searcher = reader.searcher();
+
+        let fields_and_vals = vec![
+            // Only way to address or it gets shadowed by `json.shadow` field
+            ("json.shadow\u{1}val".to_string(), "a"), // Succeeds
+            //("json.shadow.val".to_string(), "a"),   // Fails
+            ("json.shadow.val".to_string(), "b"), // Succeeds
+        ];
+
+        let query_parser = QueryParser::for_index(&index, vec![]);
+        // Test if field name can be queried
+        for (indexed_field, val) in fields_and_vals.iter() {
+            let query_str = &format!("{}:{}", indexed_field, val);
+            let query = query_parser.parse_query(query_str).unwrap();
+            let count_docs = searcher.search(&*query, &TopDocs::with_limit(2)).unwrap();
+            assert!(!count_docs.is_empty(), "{}:{}", indexed_field, val);
+        }
+        // Test if field name can be used for aggregation
+        for (field_name, val) in fields_and_vals.iter() {
+            let agg_req_str = json!(
+            {
+              "termagg": {
+                "terms": {
+                  "field": field_name,
+                }
+              }
+            });
+
+            let agg_req: Aggregations = serde_json::from_value(agg_req_str).unwrap();
+            let collector = AggregationCollector::from_aggs(agg_req, Default::default());
+            let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
+            let res = serde_json::to_value(agg_res).unwrap();
+            assert_eq!(
+                res["termagg"]["buckets"].as_array().unwrap()[0]["key"]
+                    .as_str()
+                    .unwrap(),
+                *val,
+                "{}",
+                field_name
+            );
+        }
+    }
 }

src/lib.rs

@@ -221,9 +221,9 @@ pub use self::snippet::{Snippet, SnippetGenerator};
 #[doc(hidden)]
 pub use crate::core::json_utils;
 pub use crate::core::{
-    Executor, Index, IndexBuilder, IndexMeta, IndexSettings, IndexSortByField, InvertedIndexReader,
-    Order, Searcher, SearcherGeneration, Segment, SegmentComponent, SegmentId, SegmentMeta,
-    SegmentReader, SingleSegmentIndexWriter,
+    merge_field_meta_data, Executor, FieldMetadata, Index, IndexBuilder, IndexMeta, IndexSettings,
+    IndexSortByField, InvertedIndexReader, Order, Searcher, SearcherGeneration, Segment,
+    SegmentComponent, SegmentId, SegmentMeta, SegmentReader, SingleSegmentIndexWriter,
 };
 pub use crate::directory::Directory;
 pub use crate::indexer::IndexWriter;

src/postings/json_postings_writer.rs

@@ -11,6 +11,10 @@ use crate::schema::{Field, Type, JSON_END_OF_PATH};
 use crate::tokenizer::TokenStream;
 use crate::{DocId, Term};
 
+/// The `JsonPostingsWriter` is odd in that it relies on a hidden contract:
+///
+/// `subscribe` is called directly to index non-text tokens, while
+/// `index_text` is used to index text.
 #[derive(Default)]
 pub(crate) struct JsonPostingsWriter<Rec: Recorder> {
     str_posting_writer: SpecializedPostingsWriter<Rec>,

src/postings/mod.rs

@@ -63,7 +63,7 @@ pub mod tests {
         let mut segment = index.new_segment();
         let mut posting_serializer = InvertedIndexSerializer::open(&mut segment)?;
         let mut field_serializer = posting_serializer.new_field(text_field, 120 * 4, None)?;
-        field_serializer.new_term("abc".as_bytes(), 12u32)?;
+        field_serializer.new_term("abc".as_bytes(), 12u32, true)?;
         for doc_id in 0u32..120u32 {
             let delta_positions = vec![1, 2, 3, 2];
             field_serializer.write_doc(doc_id, 4, &delta_positions);

src/postings/postings_writer.rs

@@ -194,7 +194,7 @@ impl<Rec: Recorder> SpecializedPostingsWriter<Rec> {
     ) -> io::Result<()> {
         let recorder: Rec = ctx.term_index.read(addr);
         let term_doc_freq = recorder.term_doc_freq().unwrap_or(0u32);
-        serializer.new_term(term, term_doc_freq)?;
+        serializer.new_term(term, term_doc_freq, recorder.has_term_freq())?;
         recorder.serialize(&ctx.arena, doc_id_map, serializer, buffer_lender);
         serializer.close_term()?;
         Ok(())

src/postings/recorder.rs

@@ -79,24 +79,20 @@ pub(crate) trait Recorder: Copy + Default + Send + Sync + 'static {
     ///
     /// Returns `None` if not available.
     fn term_doc_freq(&self) -> Option<u32>;
+
+    #[inline]
+    fn has_term_freq(&self) -> bool {
+        true
+    }
 }
 
 /// Only records the doc ids
-#[derive(Clone, Copy)]
+#[derive(Clone, Copy, Default)]
 pub struct DocIdRecorder {
     stack: ExpUnrolledLinkedList,
     current_doc: DocId,
 }
 
-impl Default for DocIdRecorder {
-    fn default() -> Self {
-        DocIdRecorder {
-            stack: ExpUnrolledLinkedList::default(),
-            current_doc: u32::MAX,
-        }
-    }
-}
-
 impl Recorder for DocIdRecorder {
     #[inline]
     fn current_doc(&self) -> DocId {
@@ -105,8 +101,9 @@ impl Recorder for DocIdRecorder {
 
     #[inline]
     fn new_doc(&mut self, doc: DocId, arena: &mut MemoryArena) {
+        let delta = doc - self.current_doc;
         self.current_doc = doc;
-        self.stack.writer(arena).write_u32_vint(doc);
+        self.stack.writer(arena).write_u32_vint(delta);
     }
 
     #[inline]
@@ -123,21 +120,20 @@ impl Recorder for DocIdRecorder {
         buffer_lender: &mut BufferLender,
     ) {
         let (buffer, doc_ids) = buffer_lender.lend_all();
-        self.stack.read_to_end(arena, buffer);
         // TODO avoid reading twice.
+        self.stack.read_to_end(arena, buffer);
         if let Some(doc_id_map) = doc_id_map {
-            doc_ids.extend(
-                VInt32Reader::new(&buffer[..])
-                    .map(|old_doc_id| doc_id_map.get_new_doc_id(old_doc_id)),
-            );
+            let iter = get_sum_reader(VInt32Reader::new(&buffer[..]));
+            doc_ids.extend(iter.map(|old_doc_id| doc_id_map.get_new_doc_id(old_doc_id)));
             doc_ids.sort_unstable();
 
             for doc in doc_ids {
                 serializer.write_doc(*doc, 0u32, &[][..]);
             }
         } else {
-            for doc in VInt32Reader::new(&buffer[..]) {
-                serializer.write_doc(doc, 0u32, &[][..]);
+            let iter = get_sum_reader(VInt32Reader::new(&buffer[..]));
+            for doc_id in iter {
+                serializer.write_doc(doc_id, 0u32, &[][..]);
             }
         }
     }
@@ -145,6 +141,19 @@ impl Recorder for DocIdRecorder {
     fn term_doc_freq(&self) -> Option<u32> {
         None
     }
+
+    fn has_term_freq(&self) -> bool {
+        false
+    }
+}
+
+/// Takes an Iterator of delta encoded elements and returns an iterator
+/// that yields the sum of the elements.
+fn get_sum_reader(iter: impl Iterator<Item = u32>) -> impl Iterator<Item = u32> {
+    iter.scan(0, |state, delta| {
+        *state += delta;
+        Some(*state)
+    })
 }
 
 /// Recorder encoding document ids, and term frequencies
@@ -164,9 +173,10 @@ impl Recorder for TermFrequencyRecorder {
 
     #[inline]
     fn new_doc(&mut self, doc: DocId, arena: &mut MemoryArena) {
+        let delta = doc - self.current_doc;
         self.term_doc_freq += 1;
         self.current_doc = doc;
-        self.stack.writer(arena).write_u32_vint(doc);
+        self.stack.writer(arena).write_u32_vint(delta);
     }
 
     #[inline]
@@ -193,19 +203,25 @@ impl Recorder for TermFrequencyRecorder {
         let mut u32_it = VInt32Reader::new(&buffer[..]);
         if let Some(doc_id_map) = doc_id_map {
             let mut doc_id_and_tf = vec![];
-            while let Some(old_doc_id) = u32_it.next() {
+            let mut prev_doc = 0;
+            while let Some(delta_doc_id) = u32_it.next() {
+                let doc_id = prev_doc + delta_doc_id;
+                prev_doc = doc_id;
                 let term_freq = u32_it.next().unwrap_or(self.current_tf);
-                doc_id_and_tf.push((doc_id_map.get_new_doc_id(old_doc_id), term_freq));
+                doc_id_and_tf.push((doc_id_map.get_new_doc_id(doc_id), term_freq));
             }
             doc_id_and_tf.sort_unstable_by_key(|&(doc_id, _)| doc_id);
 
             for (doc_id, tf) in doc_id_and_tf {
-                serializer.write_doc(doc_id, tf, &[][..]);
+                serializer.write_doc(doc_id, 0, &[][..]);
             }
         } else {
-            while let Some(doc) = u32_it.next() {
+            let mut prev_doc = 0;
+            while let Some(delta_doc_id) = u32_it.next() {
+                let doc_id = prev_doc + delta_doc_id;
+                prev_doc = doc_id;
                 let term_freq = u32_it.next().unwrap_or(self.current_tf);
-                serializer.write_doc(doc, term_freq, &[][..]);
+                serializer.write_doc(doc_id, 0, &[][..]);
             }
         }
     }
@@ -216,23 +232,13 @@ impl Recorder for TermFrequencyRecorder {
 }
 
 /// Recorder encoding term frequencies as well as positions.
-#[derive(Clone, Copy)]
+#[derive(Clone, Copy, Default)]
 pub struct TfAndPositionRecorder {
     stack: ExpUnrolledLinkedList,
     current_doc: DocId,
     term_doc_freq: u32,
 }
 
-impl Default for TfAndPositionRecorder {
-    fn default() -> Self {
-        TfAndPositionRecorder {
-            stack: ExpUnrolledLinkedList::default(),
-            current_doc: u32::MAX,
-            term_doc_freq: 0u32,
-        }
-    }
-}
-
 impl Recorder for TfAndPositionRecorder {
     #[inline]
     fn current_doc(&self) -> DocId {
@@ -241,9 +247,10 @@ impl Recorder for TfAndPositionRecorder {
 
     #[inline]
     fn new_doc(&mut self, doc: DocId, arena: &mut MemoryArena) {
+        let delta = doc - self.current_doc;
         self.current_doc = doc;
         self.term_doc_freq += 1u32;
-        self.stack.writer(arena).write_u32_vint(doc);
+        self.stack.writer(arena).write_u32_vint(delta);
     }
 
     #[inline]
@@ -269,7 +276,10 @@ impl Recorder for TfAndPositionRecorder {
         self.stack.read_to_end(arena, buffer_u8);
         let mut u32_it = VInt32Reader::new(&buffer_u8[..]);
         let mut doc_id_and_positions = vec![];
-        while let Some(doc) = u32_it.next() {
+        let mut prev_doc = 0;
+        while let Some(delta_doc_id) = u32_it.next() {
+            let doc_id = prev_doc + delta_doc_id;
+            prev_doc = doc_id;
             let mut prev_position_plus_one = 1u32;
             buffer_positions.clear();
             loop {
@@ -287,9 +297,9 @@ impl Recorder for TfAndPositionRecorder {
             if let Some(doc_id_map) = doc_id_map {
                 // this simple variant to remap may consume to much memory
                 doc_id_and_positions
-                    .push((doc_id_map.get_new_doc_id(doc), buffer_positions.to_vec()));
+                    .push((doc_id_map.get_new_doc_id(doc_id), buffer_positions.to_vec()));
             } else {
-                serializer.write_doc(doc, buffer_positions.len() as u32, buffer_positions);
+                serializer.write_doc(doc_id, buffer_positions.len() as u32, buffer_positions);
             }
         }
         if doc_id_map.is_some() {

src/postings/segment_postings.rs

@@ -71,7 +71,7 @@ impl SegmentPostings {
         {
             let mut postings_serializer =
                 PostingsSerializer::new(&mut buffer, 0.0, IndexRecordOption::Basic, None);
-            postings_serializer.new_term(docs.len() as u32);
+            postings_serializer.new_term(docs.len() as u32, false);
             for &doc in docs {
                 postings_serializer.write_doc(doc, 1u32);
             }
@@ -120,7 +120,7 @@ impl SegmentPostings {
             IndexRecordOption::WithFreqs,
             fieldnorm_reader,
         );
-        postings_serializer.new_term(doc_and_tfs.len() as u32);
+        postings_serializer.new_term(doc_and_tfs.len() as u32, true);
         for &(doc, tf) in doc_and_tfs {
             postings_serializer.write_doc(doc, tf);
         }
@@ -238,14 +238,18 @@ impl Postings for SegmentPostings {
     }
 
     fn positions_with_offset(&mut self, offset: u32, output: &mut Vec<u32>) {
-        let term_freq = self.term_freq() as usize;
+        let term_freq = self.term_freq();
         if let Some(position_reader) = self.position_reader.as_mut() {
+            debug_assert!(
+                !self.block_cursor.freqs().is_empty(),
+                "No positions available"
+            );
             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);
+            output.resize(term_freq as usize, 0u32);
             position_reader.read(read_offset, &mut output[..]);
             let mut cum = offset;
             for output_mut in output.iter_mut() {

src/postings/serializer.rs

@@ -168,7 +168,12 @@ impl<'a> FieldSerializer<'a> {
     /// * term - the term. It needs to come after the previous term according to the lexicographical
     ///   order.
     /// * 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<()> {
+    pub fn new_term(
+        &mut self,
+        term: &[u8],
+        term_doc_freq: u32,
+        record_term_freq: bool,
+    ) -> io::Result<()> {
         assert!(
             !self.term_open,
             "Called new_term, while the previous term was not closed."
@@ -177,7 +182,8 @@ impl<'a> FieldSerializer<'a> {
         self.postings_serializer.clear();
         self.current_term_info = self.current_term_info();
         self.term_dictionary_builder.insert_key(term)?;
-        self.postings_serializer.new_term(term_doc_freq);
+        self.postings_serializer
+            .new_term(term_doc_freq, record_term_freq);
         Ok(())
     }
 
@@ -330,10 +336,10 @@ impl<W: Write> PostingsSerializer<W> {
         }
     }
 
-    pub fn new_term(&mut self, term_doc_freq: u32) {
+    pub fn new_term(&mut self, term_doc_freq: u32, record_term_freq: bool) {
         self.bm25_weight = None;
 
-        self.term_has_freq = self.mode.has_freq() && term_doc_freq != 0;
+        self.term_has_freq = self.mode.has_freq() && record_term_freq;
         if !self.term_has_freq {
             return;
         }
@@ -349,7 +355,7 @@ impl<W: Write> PostingsSerializer<W> {
             return;
         }
 
-        self.bm25_weight = Some(Bm25Weight::for_one_term(
+        self.bm25_weight = Some(Bm25Weight::for_one_term_without_explain(
             term_doc_freq as u64,
             num_docs_in_segment,
             self.avg_fieldnorm,

src/query/bm25.rs

@@ -77,7 +77,7 @@ pub struct Bm25Params {
 /// A struct used for computing BM25 scores.
 #[derive(Clone)]
 pub struct Bm25Weight {
-    idf_explain: Explanation,
+    idf_explain: Option<Explanation>,
     weight: Score,
     cache: [Score; 256],
     average_fieldnorm: Score,
@@ -147,11 +147,30 @@ impl Bm25Weight {
         idf_explain.add_const("N, total number of docs", total_num_docs as Score);
         Bm25Weight::new(idf_explain, avg_fieldnorm)
     }
+    /// Construct a [Bm25Weight] for a single term.
+    /// This method does not carry the [Explanation] for the idf.
+    pub fn for_one_term_without_explain(
+        term_doc_freq: u64,
+        total_num_docs: u64,
+        avg_fieldnorm: Score,
+    ) -> Bm25Weight {
+        let idf = idf(term_doc_freq, total_num_docs);
+        Bm25Weight::new_without_explain(idf, avg_fieldnorm)
+    }
 
     pub(crate) fn new(idf_explain: Explanation, average_fieldnorm: Score) -> Bm25Weight {
         let weight = idf_explain.value() * (1.0 + K1);
         Bm25Weight {
-            idf_explain,
+            idf_explain: Some(idf_explain),
+            weight,
+            cache: compute_tf_cache(average_fieldnorm),
+            average_fieldnorm,
+        }
+    }
+    pub(crate) fn new_without_explain(idf: f32, average_fieldnorm: Score) -> Bm25Weight {
+        let weight = idf * (1.0 + K1);
+        Bm25Weight {
+            idf_explain: None,
             weight,
             cache: compute_tf_cache(average_fieldnorm),
             average_fieldnorm,
@@ -202,7 +221,9 @@ impl Bm25Weight {
 
         let mut explanation = Explanation::new("TermQuery, product of...", score);
         explanation.add_detail(Explanation::new("(K1+1)", K1 + 1.0));
-        explanation.add_detail(self.idf_explain.clone());
+        if let Some(idf_explain) = &self.idf_explain {
+            explanation.add_detail(idf_explain.clone());
+        }
         explanation.add_detail(tf_explanation);
         explanation
     }

src/query/boost_query.rs

@@ -74,7 +74,8 @@ impl Weight for BoostWeight {
     fn explain(&self, reader: &SegmentReader, doc: u32) -> crate::Result<Explanation> {
         let underlying_explanation = self.weight.explain(reader, doc)?;
         let score = underlying_explanation.value() * self.boost;
-        let mut explanation = Explanation::new(format!("Boost x{} of ...", self.boost), score);
+        let mut explanation =
+            Explanation::new_with_string(format!("Boost x{} of ...", self.boost), score);
         explanation.add_detail(underlying_explanation);
         Ok(explanation)
     }
@@ -151,7 +152,7 @@ mod tests {
         let explanation = query.explain(&searcher, DocAddress::new(0, 0u32)).unwrap();
         assert_eq!(
             explanation.to_pretty_json(),
-            "{\n  \"value\": 0.2,\n  \"description\": \"Boost x0.2 of ...\",\n  \"details\": [\n    {\n      \"value\": 1.0,\n      \"description\": \"AllQuery\",\n      \"context\": []\n    }\n  ],\n  \"context\": []\n}"
+            "{\n  \"value\": 0.2,\n  \"description\": \"Boost x0.2 of ...\",\n  \"details\": [\n    {\n      \"value\": 1.0,\n      \"description\": \"AllQuery\"\n    }\n  ]\n}"
         );
         Ok(())
     }

src/query/const_score_query.rs

@@ -164,11 +164,9 @@ mod tests {
   "details": [
     {
       "value": 1.0,
-      "description": "AllQuery",
-      "context": []
+      "description": "AllQuery"
     }
-  ],
-  "context": []
+  ]
 }"#
         );
         Ok(())

src/query/explanation.rs

@@ -1,3 +1,4 @@
+use std::borrow::Cow;
 use std::fmt;
 
 use serde::Serialize;
@@ -16,12 +17,12 @@ pub(crate) fn does_not_match(doc: DocId) -> TantivyError {
 #[derive(Clone, Serialize)]
 pub struct Explanation {
     value: Score,
-    description: String,
-    #[serde(skip_serializing_if = "Vec::is_empty")]
-    details: Vec<Explanation>,
-    context: Vec<String>,
+    description: Cow<'static, str>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    details: Option<Vec<Explanation>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    context: Option<Vec<String>>,
 }
-
 impl fmt::Debug for Explanation {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         write!(f, "Explanation({})", self.to_pretty_json())
@@ -30,12 +31,21 @@ impl fmt::Debug for Explanation {
 
 impl Explanation {
     /// Creates a new explanation object.
-    pub fn new<T: ToString>(description: T, value: Score) -> Explanation {
+    pub fn new_with_string(description: String, value: Score) -> Explanation {
         Explanation {
             value,
-            description: description.to_string(),
-            details: vec![],
-            context: vec![],
+            description: Cow::Owned(description),
+            details: None,
+            context: None,
+        }
+    }
+    /// Creates a new explanation object.
+    pub fn new(description: &'static str, value: Score) -> Explanation {
+        Explanation {
+            value,
+            description: Cow::Borrowed(description),
+            details: None,
+            context: None,
         }
     }
 
@@ -48,17 +58,21 @@ impl Explanation {
     ///
     /// Details are treated as child of the current node.
     pub fn add_detail(&mut self, child_explanation: Explanation) {
-        self.details.push(child_explanation);
+        self.details
+            .get_or_insert_with(Vec::new)
+            .push(child_explanation);
     }
 
     /// Adds some extra context to the explanation.
     pub fn add_context(&mut self, context: String) {
-        self.context.push(context);
+        self.context.get_or_insert_with(Vec::new).push(context);
     }
 
     /// Shortcut for `self.details.push(Explanation::new(name, value));`
-    pub fn add_const<T: ToString>(&mut self, name: T, value: Score) {
-        self.details.push(Explanation::new(name, value));
+    pub fn add_const(&mut self, name: &'static str, value: Score) {
+        self.details
+            .get_or_insert_with(Vec::new)
+            .push(Explanation::new(name, value));
     }
 
     /// Returns an indented json representation of the explanation tree for debug usage.

src/query/term_query/term_query.rs

@@ -101,7 +101,7 @@ impl TermQuery {
                 ..
             } => Bm25Weight::for_terms(statistics_provider, &[self.term.clone()])?,
             EnableScoring::Disabled { .. } => {
-                Bm25Weight::new(Explanation::new("<no score>".to_string(), 1.0f32), 1.0f32)
+                Bm25Weight::new(Explanation::new("<no score>", 1.0f32), 1.0f32)
             }
         };
         let scoring_enabled = enable_scoring.is_scoring_enabled();

src/schema/bytes_options.rs

@@ -40,21 +40,25 @@ impl From<BytesOptionsDeser> for BytesOptions {
 
 impl BytesOptions {
     /// Returns true if the value is indexed.
+    #[inline]
     pub fn is_indexed(&self) -> bool {
         self.indexed
     }
 
     /// Returns true if and only if the value is normed.
+    #[inline]
     pub fn fieldnorms(&self) -> bool {
         self.fieldnorms
     }
 
     /// Returns true if the value is a fast field.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         self.fast
     }
 
     /// Returns true if the value is stored.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }

src/schema/date_time_options.rs

@@ -27,21 +27,25 @@ pub struct DateOptions {
 
 impl DateOptions {
     /// Returns true iff the value is stored.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }
 
     /// Returns true iff the value is indexed and therefore searchable.
+    #[inline]
     pub fn is_indexed(&self) -> bool {
         self.indexed
     }
 
     /// Returns true iff the field has fieldnorm.
+    #[inline]
     pub fn fieldnorms(&self) -> bool {
         self.fieldnorms && self.indexed
     }
 
     /// Returns true iff the value is a fast field.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         self.fast
     }

src/schema/facet.rs

@@ -131,16 +131,16 @@ impl Facet {
     pub fn from_path<Path>(path: Path) -> Facet
     where
         Path: IntoIterator,
-        Path::Item: ToString,
+        Path::Item: AsRef<str>,
     {
         let mut facet_string: String = String::with_capacity(100);
         let mut step_it = path.into_iter();
         if let Some(step) = step_it.next() {
-            facet_string.push_str(&step.to_string());
+            facet_string.push_str(step.as_ref());
         }
         for step in step_it {
             facet_string.push(FACET_SEP_CHAR);
-            facet_string.push_str(&step.to_string());
+            facet_string.push_str(step.as_ref());
         }
         Facet(facet_string)
     }

src/schema/facet_options.rs

@@ -14,6 +14,7 @@ pub struct FacetOptions {
 
 impl FacetOptions {
     /// Returns true if the value is stored.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }

src/schema/field_entry.rs

@@ -108,7 +108,16 @@ impl FieldEntry {
         self.field_type.is_fast()
     }
 
+    /// Returns true if the field has the expand dots option set (for json fields)
+    pub fn is_expand_dots_enabled(&self) -> bool {
+        match self.field_type {
+            FieldType::JsonObject(ref options) => options.is_expand_dots_enabled(),
+            _ => false,
+        }
+    }
+
     /// Returns true if the field is stored
+    #[inline]
     pub fn is_stored(&self) -> bool {
         match self.field_type {
             FieldType::U64(ref options)

src/schema/field_type.rs

@@ -3,6 +3,7 @@ use std::str::FromStr;
 
 use base64::engine::general_purpose::STANDARD as BASE64;
 use base64::Engine;
+use columnar::ColumnType;
 use serde::{Deserialize, Serialize};
 use serde_json::Value as JsonValue;
 use thiserror::Error;
@@ -47,7 +48,7 @@ pub enum ValueParsingError {
 ///
 /// Contrary to FieldType, this does
 /// not include the way the field must be indexed.
-#[derive(Copy, Clone, Debug, Eq, PartialEq)]
+#[derive(Copy, Clone, Debug, Eq, PartialEq, Ord, PartialOrd, Hash)]
 #[repr(u8)]
 pub enum Type {
     /// `&str`
@@ -72,6 +73,21 @@ pub enum Type {
     IpAddr = b'p',
 }
 
+impl From<ColumnType> for Type {
+    fn from(value: ColumnType) -> Self {
+        match value {
+            ColumnType::Str => Type::Str,
+            ColumnType::U64 => Type::U64,
+            ColumnType::I64 => Type::I64,
+            ColumnType::F64 => Type::F64,
+            ColumnType::Bool => Type::Bool,
+            ColumnType::DateTime => Type::Date,
+            ColumnType::Bytes => Type::Bytes,
+            ColumnType::IpAddr => Type::IpAddr,
+        }
+    }
+}
+
 const ALL_TYPES: [Type; 10] = [
     Type::Str,
     Type::U64,

src/schema/ip_options.rs

@@ -31,21 +31,25 @@ pub struct IpAddrOptions {
 
 impl IpAddrOptions {
     /// Returns true iff the value is a fast field.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         self.fast
     }
 
     /// Returns `true` if the ip address should be stored in the doc store.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }
 
     /// Returns true iff the value is indexed and therefore searchable.
+    #[inline]
     pub fn is_indexed(&self) -> bool {
         self.indexed
     }
 
     /// Returns true if and only if the value is normed.
+    #[inline]
     pub fn fieldnorms(&self) -> bool {
         self.fieldnorms
     }

src/schema/json_object_options.rs

@@ -46,17 +46,20 @@ pub struct JsonObjectOptions {
 
 impl JsonObjectOptions {
     /// Returns `true` if the json object should be stored.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }
 
     /// Returns `true` iff the json object should be indexed.
+    #[inline]
     pub fn is_indexed(&self) -> bool {
         self.indexing.is_some()
     }
 
     /// Returns true if and only if the json object fields are
     /// to be treated as fast fields.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         matches!(self.fast, FastFieldTextOptions::IsEnabled(true))
             || matches!(
@@ -66,6 +69,7 @@ impl JsonObjectOptions {
     }
 
     /// Returns true if and only if the value is a fast field.
+    #[inline]
     pub fn get_fast_field_tokenizer_name(&self) -> Option<&str> {
         match &self.fast {
             FastFieldTextOptions::IsEnabled(true) | FastFieldTextOptions::IsEnabled(false) => None,
@@ -87,6 +91,7 @@ impl JsonObjectOptions {
     ///
     /// If disabled, the "." needs to be escaped:
     /// `k8s\.node\.id:5`.
+    #[inline]
     pub fn is_expand_dots_enabled(&self) -> bool {
         self.expand_dots_enabled
     }
@@ -103,6 +108,7 @@ impl JsonObjectOptions {
     /// If set to `Some` then both int and str values will be indexed.
     /// The inner `TextFieldIndexing` will however, only apply to the str values
     /// in the json object.
+    #[inline]
     pub fn get_text_indexing_options(&self) -> Option<&TextFieldIndexing> {
         self.indexing.as_ref()
     }

src/schema/numeric_options.rs

@@ -57,26 +57,31 @@ impl From<NumericOptionsDeser> for NumericOptions {
 
 impl NumericOptions {
     /// Returns true iff the value is stored in the doc store.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }
 
     /// Returns true iff the value is indexed and therefore searchable.
+    #[inline]
     pub fn is_indexed(&self) -> bool {
         self.indexed
     }
 
     /// Returns true iff the field has fieldnorm.
+    #[inline]
     pub fn fieldnorms(&self) -> bool {
         self.fieldnorms && self.indexed
     }
 
     /// Returns true iff the value is a fast field.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         self.fast
     }
 
     /// Returns true if values should be coerced to numbers.
+    #[inline]
     pub fn should_coerce(&self) -> bool {
         self.coerce
     }

src/schema/schema.rs

@@ -8,6 +8,7 @@ use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
 use super::ip_options::IpAddrOptions;
 use super::*;
+use crate::json_utils::split_json_path;
 use crate::schema::bytes_options::BytesOptions;
 use crate::TantivyError;
 
@@ -30,7 +31,7 @@ use crate::TantivyError;
 /// let body_field = schema_builder.add_text_field("body", TEXT);
 /// let schema = schema_builder.build();
 /// ```
-#[derive(Default)]
+#[derive(Debug, Default)]
 pub struct SchemaBuilder {
     fields: Vec<FieldEntry>,
     fields_map: HashMap<String, Field>,
@@ -328,12 +329,19 @@ impl Schema {
         if let Some(field) = self.0.fields_map.get(full_path) {
             return Some((*field, ""));
         }
+
         let mut splitting_period_pos: Vec<usize> = locate_splitting_dots(full_path);
         while let Some(pos) = splitting_period_pos.pop() {
             let (prefix, suffix) = full_path.split_at(pos);
+
             if let Some(field) = self.0.fields_map.get(prefix) {
                 return Some((*field, &suffix[1..]));
             }
+            // JSON path may contain a dot, for now we try both variants to find the field.
+            let prefix = split_json_path(prefix).join(".");
+            if let Some(field) = self.0.fields_map.get(&prefix) {
+                return Some((*field, &suffix[1..]));
+            }
         }
         None
     }
@@ -349,6 +357,7 @@ impl Schema {
     pub fn find_field_with_default<'a>(
         &self,
         full_path: &'a str,
+
         default_field_opt: Option<Field>,
     ) -> Option<(Field, &'a str)> {
         let (field, json_path) = self

src/schema/text_options.rs

@@ -72,16 +72,19 @@ fn is_false(val: &bool) -> bool {
 
 impl TextOptions {
     /// Returns the indexing options.
+    #[inline]
     pub fn get_indexing_options(&self) -> Option<&TextFieldIndexing> {
         self.indexing.as_ref()
     }
 
     /// Returns true if the text is to be stored.
+    #[inline]
     pub fn is_stored(&self) -> bool {
         self.stored
     }
 
     /// Returns true if and only if the value is a fast field.
+    #[inline]
     pub fn is_fast(&self) -> bool {
         matches!(self.fast, FastFieldTextOptions::IsEnabled(true))
             || matches!(
@@ -91,6 +94,7 @@ impl TextOptions {
     }
 
     /// Returns true if and only if the value is a fast field.
+    #[inline]
     pub fn get_fast_field_tokenizer_name(&self) -> Option<&str> {
         match &self.fast {
             FastFieldTextOptions::IsEnabled(true) | FastFieldTextOptions::IsEnabled(false) => None,
@@ -101,6 +105,7 @@ impl TextOptions {
     }
 
     /// Returns true if values should be coerced to strings (numbers, null).
+    #[inline]
     pub fn should_coerce(&self) -> bool {
         self.coerce
     }

sstable/Cargo.toml

@@ -11,16 +11,22 @@ description = "sstables for tantivy"
 
 [dependencies]
 common = {version= "0.6", path="../common", package="tantivy-common"}
-tantivy-fst = "0.4"
+tantivy-bitpacker = { version= "0.5", path="../bitpacker" }
+tantivy-fst = "0.5"
 # experimental gives us access to Decompressor::upper_bound
 zstd = { version = "0.13", features = ["experimental"] }
 
 [dev-dependencies]
 proptest = "1"
-criterion = "0.5"
+criterion = { version = "0.5", default-features = false }
 names = "0.14"
 rand = "0.8"
 
 [[bench]]
 name = "stream_bench"
 harness = false
+
+[[bench]]
+name = "ord_to_term"
+harness = false
+

sstable/README.md

@@ -89,33 +89,71 @@ Note: as the SSTable does not support redundant keys, there is no ambiguity betw
 
 ### SSTFooter
 ```
-+-------+-------+-----+-------------+---------+---------+
-| Block | Block | ... | IndexOffset | NumTerm | Version |
-+-------+-------+-----+-------------+---------+---------+
-|----( # of blocks)---|
++-----+----------------+-------------+-------------+---------+---------+
+| Fst | BlockAddrStore | StoreOffset | IndexOffset | NumTerm | Version |
++-----+----------------+-------------+-------------+---------+---------+
 ```
-- Block(SSTBlock): uses IndexValue for its Values format
+- Fst(Fst): finite state transducer mapping keys to a block number
+- BlockAddrStore(BlockAddrStore): store mapping a block number to its BlockAddr
+- StoreOffset(u64): Offset to start of the BlockAddrStore. If zero, see the SingleBlockSStable section
 - IndexOffset(u64): Offset to the start of the SSTFooter
 - NumTerm(u64): number of terms in the sstable
-- Version(u32): Currently equal to 2
+- Version(u32): Currently equal to 3
 
-### IndexValue
-```
-+------------+----------+-------+-------+-----+
-| EntryCount | StartPos | Entry | Entry | ... |
-+------------+----------+-------+-------+-----+
-                        |---( # of entries)---|
-```
+### Fst
 
-- EntryCount(VInt): number of entries
-- StartPos(VInt): the start pos of the first (data) block referenced by this (index) block
-- Entry (IndexEntry)
+Fst is in the format of tantivy\_fst
 
-### Entry
-```
-+----------+--------------+
-| BlockLen | FirstOrdinal |
-+----------+--------------+
-```
-- BlockLen(VInt): length of the block
-- FirstOrdinal(VInt): ordinal of the first element in the given block
+### BlockAddrStore
+
++---------+-----------+-----------+-----+-----------+-----------+-----+
+| MetaLen | BlockMeta | BlockMeta | ... | BlockData | BlockData | ... |
++---------+-----------+-----------+-----+-----------+-----------+-----+
+          |---------(N blocks)----------|---------(N blocks)----------|
+
+- MetaLen(u64): length of the BlockMeta section
+- BlockMeta(BlockAddrBlockMetadata): metadata to seek through BlockData
+- BlockData(CompactedBlockAddr): bitpacked per block metadata
+
+### BlockAddrBlockMetadata
+
++--------+------------+--------------+------------+--------------+-------------------+-----------------+----------+
+| Offset | RangeStart | FirstOrdinal | RangeSlope | OrdinalSlope | FirstOrdinalNBits | RangeStartNBits | BlockLen |
++--------+------------+--------------+------------+--------------+-------------------+-----------------+----------+
+
+- Offset(u64): offset of the corresponding BlockData in the datastream
+- RangeStart(u64): the start position of the first block
+- FirstOrdinal(u64): the first ordinal of the first block
+- RangeSlope(u32): slope predicted for start range evolution (see computation in BlockData)
+- OrdinalSlope(u64): slope predicted for first ordinal evolution (see computation in BlockData)
+- FirstOrdinalNBits(u8): number of bits per ordinal in datastream (see computation in BlockData)
+- RangeStartNBits(u8): number of bits per range start in datastream (see computation in BlockData)
+
+### BlockData
+
++-----------------+-------------------+---------------+
+| RangeStartDelta | FirstOrdinalDelta | FinalRangeEnd |
++-----------------+-------------------+---------------+
+|------(BlockLen repetitions)---------|
+
+- RangeStartDelta(var): RangeStartNBits *bits* of little endian number. See below for decoding
+- FirstOrdinalDelta(var): FirstOrdinalNBits *bits* of little endian number. See below for decoding
+- FinalRangeEnd(var): RangeStartNBits *bits* of integer. See below for decoding
+
+converting a BlockData of index Index and a BlockAddrBlockMetadata to an actual block address is done as follow:
+range\_prediction := RangeStart + Index * RangeSlop;
+range\_derivation := RangeStartDelta - (1 << (RangeStartNBits-1));
+range\_start := range\_prediction + range\_derivation
+
+The same computation can be done for ordinal.
+
+Note that `range_derivation` can take negative value. `RangeStartDelta` is just its translation to a positive range.
+
+
+## SingleBlockSStable
+
+The format used for the index is meant to be compact, however it has a constant cost of around 70
+bytes, which isn't negligible for a table containing very few keys.
+To limit the impact of that constant cost, single block sstable omit the Fst and BlockAddrStore from
+their index. Instead a block with first ordinal of 0, range start of 0 and range end of IndexOffset
+is implicitly used for every operations.

sstable/benches/ord_to_term.rs

@@ -0,0 +1,110 @@
+use std::sync::Arc;
+
+use common::file_slice::FileSlice;
+use common::OwnedBytes;
+use criterion::{criterion_group, criterion_main, Criterion};
+use tantivy_sstable::{self, Dictionary, MonotonicU64SSTable};
+
+fn make_test_sstable(suffix: &str) -> FileSlice {
+    let mut builder = Dictionary::<MonotonicU64SSTable>::builder(Vec::new()).unwrap();
+
+    // 125 mio elements
+    for elem in 0..125_000_000 {
+        let key = format!("prefix.{elem:07X}{suffix}").into_bytes();
+        builder.insert(&key, &elem).unwrap();
+    }
+
+    let table = builder.finish().unwrap();
+    let table = Arc::new(OwnedBytes::new(table));
+    let slice = common::file_slice::FileSlice::new(table.clone());
+
+    slice
+}
+
+pub fn criterion_benchmark(c: &mut Criterion) {
+    {
+        let slice = make_test_sstable(".suffix");
+        let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+        c.bench_function("ord_to_term_suffix", |b| {
+            let mut res = Vec::new();
+            b.iter(|| {
+                assert!(dict.ord_to_term(100_000, &mut res).unwrap());
+                assert!(dict.ord_to_term(19_000_000, &mut res).unwrap());
+            })
+        });
+        c.bench_function("open_and_ord_to_term_suffix", |b| {
+            let mut res = Vec::new();
+            b.iter(|| {
+                let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+                assert!(dict.ord_to_term(100_000, &mut res).unwrap());
+                assert!(dict.ord_to_term(19_000_000, &mut res).unwrap());
+            })
+        });
+        c.bench_function("term_ord_suffix", |b| {
+            b.iter(|| {
+                assert_eq!(
+                    dict.term_ord(b"prefix.00186A0.suffix").unwrap().unwrap(),
+                    100_000
+                );
+                assert_eq!(
+                    dict.term_ord(b"prefix.121EAC0.suffix").unwrap().unwrap(),
+                    19_000_000
+                );
+            })
+        });
+        c.bench_function("open_and_term_ord_suffix", |b| {
+            b.iter(|| {
+                let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+                assert_eq!(
+                    dict.term_ord(b"prefix.00186A0.suffix").unwrap().unwrap(),
+                    100_000
+                );
+                assert_eq!(
+                    dict.term_ord(b"prefix.121EAC0.suffix").unwrap().unwrap(),
+                    19_000_000
+                );
+            })
+        });
+    }
+    {
+        let slice = make_test_sstable("");
+        let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+        c.bench_function("ord_to_term", |b| {
+            let mut res = Vec::new();
+            b.iter(|| {
+                assert!(dict.ord_to_term(100_000, &mut res).unwrap());
+                assert!(dict.ord_to_term(19_000_000, &mut res).unwrap());
+            })
+        });
+        c.bench_function("open_and_ord_to_term", |b| {
+            let mut res = Vec::new();
+            b.iter(|| {
+                let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+                assert!(dict.ord_to_term(100_000, &mut res).unwrap());
+                assert!(dict.ord_to_term(19_000_000, &mut res).unwrap());
+            })
+        });
+        c.bench_function("term_ord", |b| {
+            b.iter(|| {
+                assert_eq!(dict.term_ord(b"prefix.00186A0").unwrap().unwrap(), 100_000);
+                assert_eq!(
+                    dict.term_ord(b"prefix.121EAC0").unwrap().unwrap(),
+                    19_000_000
+                );
+            })
+        });
+        c.bench_function("open_and_term_ord", |b| {
+            b.iter(|| {
+                let dict = Dictionary::<MonotonicU64SSTable>::open(slice.clone()).unwrap();
+                assert_eq!(dict.term_ord(b"prefix.00186A0").unwrap().unwrap(), 100_000);
+                assert_eq!(
+                    dict.term_ord(b"prefix.121EAC0").unwrap().unwrap(),
+                    19_000_000
+                );
+            })
+        });
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);

sstable/src/dictionary.rs

@@ -9,8 +9,11 @@ use common::{BinarySerializable, OwnedBytes};
 use tantivy_fst::automaton::AlwaysMatch;
 use tantivy_fst::Automaton;
 
+use crate::sstable_index_v3::SSTableIndexV3Empty;
 use crate::streamer::{Streamer, StreamerBuilder};
-use crate::{BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, TermOrdinal, VoidSSTable};
+use crate::{
+    BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, SSTableIndexV3, TermOrdinal, VoidSSTable,
+};
 
 /// An SSTable is a sorted map that associates sorted `&[u8]` keys
 /// to any kind of typed values.
@@ -180,24 +183,41 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
     pub fn open(term_dictionary_file: FileSlice) -> io::Result<Self> {
         let (main_slice, footer_len_slice) = term_dictionary_file.split_from_end(20);
         let mut footer_len_bytes: OwnedBytes = footer_len_slice.read_bytes()?;
-
         let index_offset = u64::deserialize(&mut footer_len_bytes)?;
         let num_terms = u64::deserialize(&mut footer_len_bytes)?;
         let version = u32::deserialize(&mut footer_len_bytes)?;
-        if version != crate::SSTABLE_VERSION {
-            return Err(io::Error::new(
-                io::ErrorKind::Other,
-                format!(
-                    "Unsuported sstable version, expected {version}, found {}",
-                    crate::SSTABLE_VERSION,
-                ),
-            ));
-        }
-
         let (sstable_slice, index_slice) = main_slice.split(index_offset as usize);
         let sstable_index_bytes = index_slice.read_bytes()?;
-        let sstable_index = SSTableIndex::load(sstable_index_bytes)
-            .map_err(|_| io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption"))?;
+
+        let sstable_index = match version {
+            2 => SSTableIndex::V2(
+                crate::sstable_index_v2::SSTableIndex::load(sstable_index_bytes).map_err(|_| {
+                    io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption")
+                })?,
+            ),
+            3 => {
+                let (sstable_index_bytes, mut footerv3_len_bytes) = sstable_index_bytes.rsplit(8);
+                let store_offset = u64::deserialize(&mut footerv3_len_bytes)?;
+                if store_offset != 0 {
+                    SSTableIndex::V3(
+                        SSTableIndexV3::load(sstable_index_bytes, store_offset).map_err(|_| {
+                            io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption")
+                        })?,
+                    )
+                } else {
+                    // if store_offset is zero, there is no index, so we build a pseudo-index
+                    // assuming a single block of sstable covering everything.
+                    SSTableIndex::V3Empty(SSTableIndexV3Empty::load(index_offset as usize))
+                }
+            }
+            _ => {
+                return Err(io::Error::new(
+                    io::ErrorKind::Other,
+                    format!("Unsuported sstable version, expected one of [2, 3], found {version}"),
+                ))
+            }
+        };
+
         Ok(Dictionary {
             sstable_slice,
             sstable_index,

sstable/src/lib.rs

@@ -10,8 +10,9 @@ pub mod merge;
 mod streamer;
 pub mod value;
 
-mod sstable_index;
-pub use sstable_index::{BlockAddr, SSTableIndex, SSTableIndexBuilder};
+mod sstable_index_v3;
+pub use sstable_index_v3::{BlockAddr, SSTableIndex, SSTableIndexBuilder, SSTableIndexV3};
+mod sstable_index_v2;
 pub(crate) mod vint;
 pub use dictionary::Dictionary;
 pub use streamer::{Streamer, StreamerBuilder};
@@ -28,7 +29,7 @@ use crate::value::{RangeValueReader, RangeValueWriter};
 pub type TermOrdinal = u64;
 
 const DEFAULT_KEY_CAPACITY: usize = 50;
-const SSTABLE_VERSION: u32 = 2;
+const SSTABLE_VERSION: u32 = 3;
 
 /// Given two byte string returns the length of
 /// the longest common prefix.
@@ -304,7 +305,8 @@ where
 
         let offset = wrt.written_bytes();
 
-        self.index_builder.serialize(&mut wrt)?;
+        let fst_len: u64 = self.index_builder.serialize(&mut wrt)?;
+        wrt.write_all(&fst_len.to_le_bytes())?;
         wrt.write_all(&offset.to_le_bytes())?;
         wrt.write_all(&self.num_terms.to_le_bytes())?;
 
@@ -385,13 +387,10 @@ mod test {
                 16, 17, 33, 18, 19, 17, 20, // data block
                 0, 0, 0, 0, // no more block
                 // index
-                8, 0, 0, 0, // size of index block
-                0, // compression
-                1, 0, 12, 0, 32, 17, 20, // index block
-                0, 0, 0, 0, // no more index block
+                0, 0, 0, 0, 0, 0, 0, 0, // fst lenght
                 16, 0, 0, 0, 0, 0, 0, 0, // index start offset
                 3, 0, 0, 0, 0, 0, 0, 0, // num term
-                2, 0, 0, 0, // version
+                3, 0, 0, 0, // version
             ]
         );
         let buffer = OwnedBytes::new(buffer);

sstable/src/sstable_index.rs

@@ -1,266 +0,0 @@
-use std::io::{self, Write};
-use std::ops::Range;
-
-use common::OwnedBytes;
-
-use crate::{common_prefix_len, SSTable, SSTableDataCorruption, TermOrdinal};
-
-#[derive(Default, Debug, Clone)]
-pub struct SSTableIndex {
-    blocks: Vec<BlockMeta>,
-}
-
-impl SSTableIndex {
-    /// Load an index from its binary representation
-    pub fn load(data: OwnedBytes) -> Result<SSTableIndex, SSTableDataCorruption> {
-        let mut reader = IndexSSTable::reader(data);
-        let mut blocks = Vec::new();
-
-        while reader.advance().map_err(|_| SSTableDataCorruption)? {
-            blocks.push(BlockMeta {
-                last_key_or_greater: reader.key().to_vec(),
-                block_addr: reader.value().clone(),
-            });
-        }
-
-        Ok(SSTableIndex { blocks })
-    }
-
-    /// Get the [`BlockAddr`] of the requested block.
-    pub(crate) fn get_block(&self, block_id: usize) -> Option<BlockAddr> {
-        self.blocks
-            .get(block_id)
-            .map(|block_meta| block_meta.block_addr.clone())
-    }
-
-    /// Get the block id of the block that would contain `key`.
-    ///
-    /// Returns None if `key` is lexicographically after the last key recorded.
-    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<usize> {
-        let pos = self
-            .blocks
-            .binary_search_by_key(&key, |block| &block.last_key_or_greater);
-        match pos {
-            Ok(pos) => Some(pos),
-            Err(pos) => {
-                if pos < self.blocks.len() {
-                    Some(pos)
-                } else {
-                    // after end of last block: no block matches
-                    None
-                }
-            }
-        }
-    }
-
-    /// Get the [`BlockAddr`] of the block that would contain `key`.
-    ///
-    /// Returns None if `key` is lexicographically after the last key recorded.
-    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
-        self.locate_with_key(key).and_then(|id| self.get_block(id))
-    }
-
-    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> usize {
-        let pos = self
-            .blocks
-            .binary_search_by_key(&ord, |block| block.block_addr.first_ordinal);
-
-        match pos {
-            Ok(pos) => pos,
-            // Err(0) can't happen as the sstable starts with ordinal zero
-            Err(pos) => pos - 1,
-        }
-    }
-
-    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
-    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
-        // locate_with_ord always returns an index within range
-        self.get_block(self.locate_with_ord(ord)).unwrap()
-    }
-}
-
-#[derive(Clone, Eq, PartialEq, Debug)]
-pub struct BlockAddr {
-    pub byte_range: Range<usize>,
-    pub first_ordinal: u64,
-}
-
-#[derive(Debug, Clone)]
-pub(crate) struct BlockMeta {
-    /// Any byte string that is lexicographically greater or equal to
-    /// the last key in the block,
-    /// and yet strictly smaller than the first key in the next block.
-    pub last_key_or_greater: Vec<u8>,
-    pub block_addr: BlockAddr,
-}
-
-#[derive(Default)]
-pub struct SSTableIndexBuilder {
-    index: SSTableIndex,
-}
-
-/// Given that left < right,
-/// mutates `left into a shorter byte string left'` that
-/// matches `left <= left' < right`.
-fn find_shorter_str_in_between(left: &mut Vec<u8>, right: &[u8]) {
-    assert!(&left[..] < right);
-    let common_len = common_prefix_len(left, right);
-    if left.len() == common_len {
-        return;
-    }
-    // It is possible to do one character shorter in some case,
-    // but it is not worth the extra complexity
-    for pos in (common_len + 1)..left.len() {
-        if left[pos] != u8::MAX {
-            left[pos] += 1;
-            left.truncate(pos + 1);
-            return;
-        }
-    }
-}
-
-impl SSTableIndexBuilder {
-    /// In order to make the index as light as possible, we
-    /// try to find a shorter alternative to the last key of the last block
-    /// that is still smaller than the next key.
-    pub(crate) fn shorten_last_block_key_given_next_key(&mut self, next_key: &[u8]) {
-        if let Some(last_block) = self.index.blocks.last_mut() {
-            find_shorter_str_in_between(&mut last_block.last_key_or_greater, next_key);
-        }
-    }
-
-    pub fn add_block(&mut self, last_key: &[u8], byte_range: Range<usize>, first_ordinal: u64) {
-        self.index.blocks.push(BlockMeta {
-            last_key_or_greater: last_key.to_vec(),
-            block_addr: BlockAddr {
-                byte_range,
-                first_ordinal,
-            },
-        })
-    }
-
-    pub fn serialize<W: std::io::Write>(&self, wrt: W) -> io::Result<()> {
-        // we can't use a plain writer as it would generate an index
-        let mut sstable_writer = IndexSSTable::delta_writer(wrt);
-
-        // in tests, set a smaller block size to stress-test
-        #[cfg(test)]
-        sstable_writer.set_block_len(16);
-
-        let mut previous_key = Vec::with_capacity(crate::DEFAULT_KEY_CAPACITY);
-        for block in self.index.blocks.iter() {
-            let keep_len = common_prefix_len(&previous_key, &block.last_key_or_greater);
-
-            sstable_writer.write_suffix(keep_len, &block.last_key_or_greater[keep_len..]);
-            sstable_writer.write_value(&block.block_addr);
-            sstable_writer.flush_block_if_required()?;
-
-            previous_key.clear();
-            previous_key.extend_from_slice(&block.last_key_or_greater);
-        }
-        sstable_writer.flush_block()?;
-        sstable_writer.finish().write_all(&0u32.to_le_bytes())?;
-        Ok(())
-    }
-}
-
-/// SSTable representing an index
-///
-/// `last_key_or_greater` is used as the key, the value contains the
-/// length and first ordinal of each block. The start offset is implicitly
-/// obtained from lengths.
-struct IndexSSTable;
-
-impl SSTable for IndexSSTable {
-    type Value = BlockAddr;
-
-    type ValueReader = crate::value::index::IndexValueReader;
-
-    type ValueWriter = crate::value::index::IndexValueWriter;
-}
-
-#[cfg(test)]
-mod tests {
-    use common::OwnedBytes;
-
-    use super::{BlockAddr, SSTableIndex, SSTableIndexBuilder};
-    use crate::SSTableDataCorruption;
-
-    #[test]
-    fn test_sstable_index() {
-        let mut sstable_builder = SSTableIndexBuilder::default();
-        sstable_builder.add_block(b"aaa", 10..20, 0u64);
-        sstable_builder.add_block(b"bbbbbbb", 20..30, 5u64);
-        sstable_builder.add_block(b"ccc", 30..40, 10u64);
-        sstable_builder.add_block(b"dddd", 40..50, 15u64);
-        let mut buffer: Vec<u8> = Vec::new();
-        sstable_builder.serialize(&mut buffer).unwrap();
-        let buffer = OwnedBytes::new(buffer);
-        let sstable_index = SSTableIndex::load(buffer).unwrap();
-        assert_eq!(
-            sstable_index.get_block_with_key(b"bbbde"),
-            Some(BlockAddr {
-                first_ordinal: 10u64,
-                byte_range: 30..40
-            })
-        );
-
-        assert_eq!(sstable_index.locate_with_key(b"aa").unwrap(), 0);
-        assert_eq!(sstable_index.locate_with_key(b"aaa").unwrap(), 0);
-        assert_eq!(sstable_index.locate_with_key(b"aab").unwrap(), 1);
-        assert_eq!(sstable_index.locate_with_key(b"ccc").unwrap(), 2);
-        assert!(sstable_index.locate_with_key(b"e").is_none());
-
-        assert_eq!(sstable_index.locate_with_ord(0), 0);
-        assert_eq!(sstable_index.locate_with_ord(1), 0);
-        assert_eq!(sstable_index.locate_with_ord(4), 0);
-        assert_eq!(sstable_index.locate_with_ord(5), 1);
-        assert_eq!(sstable_index.locate_with_ord(100), 3);
-    }
-
-    #[test]
-    fn test_sstable_with_corrupted_data() {
-        let mut sstable_builder = SSTableIndexBuilder::default();
-        sstable_builder.add_block(b"aaa", 10..20, 0u64);
-        sstable_builder.add_block(b"bbbbbbb", 20..30, 5u64);
-        sstable_builder.add_block(b"ccc", 30..40, 10u64);
-        sstable_builder.add_block(b"dddd", 40..50, 15u64);
-        let mut buffer: Vec<u8> = Vec::new();
-        sstable_builder.serialize(&mut buffer).unwrap();
-        buffer[2] = 9u8;
-        let buffer = OwnedBytes::new(buffer);
-        let data_corruption_err = SSTableIndex::load(buffer).err().unwrap();
-        assert!(matches!(data_corruption_err, SSTableDataCorruption));
-    }
-
-    #[track_caller]
-    fn test_find_shorter_str_in_between_aux(left: &[u8], right: &[u8]) {
-        let mut left_buf = left.to_vec();
-        super::find_shorter_str_in_between(&mut left_buf, right);
-        assert!(left_buf.len() <= left.len());
-        assert!(left <= &left_buf);
-        assert!(&left_buf[..] < right);
-    }
-
-    #[test]
-    fn test_find_shorter_str_in_between() {
-        test_find_shorter_str_in_between_aux(b"", b"hello");
-        test_find_shorter_str_in_between_aux(b"abc", b"abcd");
-        test_find_shorter_str_in_between_aux(b"abcd", b"abd");
-        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[1]);
-        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[0, 0, 1]);
-        test_find_shorter_str_in_between_aux(&[0, 0, 255, 255, 255, 0u8], &[0, 1]);
-    }
-
-    use proptest::prelude::*;
-
-    proptest! {
-        #![proptest_config(ProptestConfig::with_cases(100))]
-        #[test]
-        fn test_proptest_find_shorter_str(left in any::<Vec<u8>>(), right in any::<Vec<u8>>()) {
-            if left < right {
-                test_find_shorter_str_in_between_aux(&left, &right);
-            }
-        }
-    }
-}

sstable/src/sstable_index_v2.rs

@@ -0,0 +1,101 @@
+use common::OwnedBytes;
+
+use crate::{BlockAddr, SSTable, SSTableDataCorruption, TermOrdinal};
+
+#[derive(Default, Debug, Clone)]
+pub struct SSTableIndex {
+    blocks: Vec<BlockMeta>,
+}
+
+impl SSTableIndex {
+    /// Load an index from its binary representation
+    pub fn load(data: OwnedBytes) -> Result<SSTableIndex, SSTableDataCorruption> {
+        let mut reader = IndexSSTable::reader(data);
+        let mut blocks = Vec::new();
+
+        while reader.advance().map_err(|_| SSTableDataCorruption)? {
+            blocks.push(BlockMeta {
+                last_key_or_greater: reader.key().to_vec(),
+                block_addr: reader.value().clone(),
+            });
+        }
+
+        Ok(SSTableIndex { blocks })
+    }
+
+    /// Get the [`BlockAddr`] of the requested block.
+    pub(crate) fn get_block(&self, block_id: usize) -> Option<BlockAddr> {
+        self.blocks
+            .get(block_id)
+            .map(|block_meta| block_meta.block_addr.clone())
+    }
+
+    /// Get the block id of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<usize> {
+        let pos = self
+            .blocks
+            .binary_search_by_key(&key, |block| &block.last_key_or_greater);
+        match pos {
+            Ok(pos) => Some(pos),
+            Err(pos) => {
+                if pos < self.blocks.len() {
+                    Some(pos)
+                } else {
+                    // after end of last block: no block matches
+                    None
+                }
+            }
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
+        self.locate_with_key(key).and_then(|id| self.get_block(id))
+    }
+
+    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> usize {
+        let pos = self
+            .blocks
+            .binary_search_by_key(&ord, |block| block.block_addr.first_ordinal);
+
+        match pos {
+            Ok(pos) => pos,
+            // Err(0) can't happen as the sstable starts with ordinal zero
+            Err(pos) => pos - 1,
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
+    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
+        // locate_with_ord always returns an index within range
+        self.get_block(self.locate_with_ord(ord)).unwrap()
+    }
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct BlockMeta {
+    /// Any byte string that is lexicographically greater or equal to
+    /// the last key in the block,
+    /// and yet strictly smaller than the first key in the next block.
+    pub last_key_or_greater: Vec<u8>,
+    pub block_addr: BlockAddr,
+}
+
+/// SSTable representing an index
+///
+/// `last_key_or_greater` is used as the key, the value contains the
+/// length and first ordinal of each block. The start offset is implicitly
+/// obtained from lengths.
+struct IndexSSTable;
+
+impl SSTable for IndexSSTable {
+    type Value = BlockAddr;
+
+    type ValueReader = crate::value::index::IndexValueReader;
+
+    type ValueWriter = crate::value::index::IndexValueWriter;
+}

sstable/src/sstable_index_v3.rs

@@ -0,0 +1,826 @@
+use std::io::{self, Read, Write};
+use std::ops::Range;
+use std::sync::Arc;
+
+use common::{BinarySerializable, FixedSize, OwnedBytes};
+use tantivy_bitpacker::{compute_num_bits, BitPacker};
+use tantivy_fst::raw::Fst;
+use tantivy_fst::{IntoStreamer, Map, MapBuilder, Streamer};
+
+use crate::{common_prefix_len, SSTableDataCorruption, TermOrdinal};
+
+#[derive(Debug, Clone)]
+pub enum SSTableIndex {
+    V2(crate::sstable_index_v2::SSTableIndex),
+    V3(SSTableIndexV3),
+    V3Empty(SSTableIndexV3Empty),
+}
+
+impl SSTableIndex {
+    /// Get the [`BlockAddr`] of the requested block.
+    pub(crate) fn get_block(&self, block_id: u64) -> Option<BlockAddr> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block(block_id as usize),
+            SSTableIndex::V3(v3_index) => v3_index.get_block(block_id),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block(block_id),
+        }
+    }
+
+    /// Get the block id of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<u64> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.locate_with_key(key).map(|i| i as u64),
+            SSTableIndex::V3(v3_index) => v3_index.locate_with_key(key),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_key(key),
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block_with_key(key),
+            SSTableIndex::V3(v3_index) => v3_index.get_block_with_key(key),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_key(key),
+        }
+    }
+
+    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> u64 {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.locate_with_ord(ord) as u64,
+            SSTableIndex::V3(v3_index) => v3_index.locate_with_ord(ord),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.locate_with_ord(ord),
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
+    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
+        match self {
+            SSTableIndex::V2(v2_index) => v2_index.get_block_with_ord(ord),
+            SSTableIndex::V3(v3_index) => v3_index.get_block_with_ord(ord),
+            SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_ord(ord),
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct SSTableIndexV3 {
+    fst_index: Arc<Map<OwnedBytes>>,
+    block_addr_store: BlockAddrStore,
+}
+
+impl SSTableIndexV3 {
+    /// Load an index from its binary representation
+    pub fn load(
+        data: OwnedBytes,
+        fst_length: u64,
+    ) -> Result<SSTableIndexV3, SSTableDataCorruption> {
+        let (fst_slice, block_addr_store_slice) = data.split(fst_length as usize);
+        let fst_index = Fst::new(fst_slice)
+            .map_err(|_| SSTableDataCorruption)?
+            .into();
+        let block_addr_store =
+            BlockAddrStore::open(block_addr_store_slice).map_err(|_| SSTableDataCorruption)?;
+
+        Ok(SSTableIndexV3 {
+            fst_index: Arc::new(fst_index),
+            block_addr_store,
+        })
+    }
+
+    /// Get the [`BlockAddr`] of the requested block.
+    pub(crate) fn get_block(&self, block_id: u64) -> Option<BlockAddr> {
+        self.block_addr_store.get(block_id)
+    }
+
+    /// Get the block id of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub(crate) fn locate_with_key(&self, key: &[u8]) -> Option<u64> {
+        self.fst_index
+            .range()
+            .ge(key)
+            .into_stream()
+            .next()
+            .map(|(_key, id)| id)
+    }
+
+    /// Get the [`BlockAddr`] of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub fn get_block_with_key(&self, key: &[u8]) -> Option<BlockAddr> {
+        self.locate_with_key(key).and_then(|id| self.get_block(id))
+    }
+
+    pub(crate) fn locate_with_ord(&self, ord: TermOrdinal) -> u64 {
+        self.block_addr_store.binary_search_ord(ord).0
+    }
+
+    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
+    pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
+        self.block_addr_store.binary_search_ord(ord).1
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct SSTableIndexV3Empty {
+    block_addr: BlockAddr,
+}
+
+impl SSTableIndexV3Empty {
+    pub fn load(index_start_pos: usize) -> SSTableIndexV3Empty {
+        SSTableIndexV3Empty {
+            block_addr: BlockAddr {
+                first_ordinal: 0,
+                byte_range: 0..index_start_pos,
+            },
+        }
+    }
+
+    /// Get the [`BlockAddr`] of the requested block.
+    pub(crate) fn get_block(&self, _block_id: u64) -> Option<BlockAddr> {
+        Some(self.block_addr.clone())
+    }
+
+    /// Get the block id of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub(crate) fn locate_with_key(&self, _key: &[u8]) -> Option<u64> {
+        Some(0)
+    }
+
+    /// Get the [`BlockAddr`] of the block that would contain `key`.
+    ///
+    /// Returns None if `key` is lexicographically after the last key recorded.
+    pub fn get_block_with_key(&self, _key: &[u8]) -> Option<BlockAddr> {
+        Some(self.block_addr.clone())
+    }
+
+    pub(crate) fn locate_with_ord(&self, _ord: TermOrdinal) -> u64 {
+        0
+    }
+
+    /// Get the [`BlockAddr`] of the block containing the `ord`-th term.
+    pub(crate) fn get_block_with_ord(&self, _ord: TermOrdinal) -> BlockAddr {
+        self.block_addr.clone()
+    }
+}
+#[derive(Clone, Eq, PartialEq, Debug)]
+pub struct BlockAddr {
+    pub first_ordinal: u64,
+    pub byte_range: Range<usize>,
+}
+
+impl BlockAddr {
+    fn to_block_start(&self) -> BlockStartAddr {
+        BlockStartAddr {
+            first_ordinal: self.first_ordinal,
+            byte_range_start: self.byte_range.start,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct BlockStartAddr {
+    first_ordinal: u64,
+    byte_range_start: usize,
+}
+
+impl BlockStartAddr {
+    fn to_block_addr(&self, byte_range_end: usize) -> BlockAddr {
+        BlockAddr {
+            first_ordinal: self.first_ordinal,
+            byte_range: self.byte_range_start..byte_range_end,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct BlockMeta {
+    /// Any byte string that is lexicographically greater or equal to
+    /// the last key in the block,
+    /// and yet strictly smaller than the first key in the next block.
+    pub last_key_or_greater: Vec<u8>,
+    pub block_addr: BlockAddr,
+}
+
+impl BinarySerializable for BlockStartAddr {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        let start = self.byte_range_start as u64;
+        start.serialize(writer)?;
+        self.first_ordinal.serialize(writer)
+    }
+
+    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
+        let byte_range_start = u64::deserialize(reader)? as usize;
+        let first_ordinal = u64::deserialize(reader)?;
+        Ok(BlockStartAddr {
+            first_ordinal,
+            byte_range_start,
+        })
+    }
+
+    // Provided method
+    fn num_bytes(&self) -> u64 {
+        BlockStartAddr::SIZE_IN_BYTES as u64
+    }
+}
+
+impl FixedSize for BlockStartAddr {
+    const SIZE_IN_BYTES: usize = 2 * u64::SIZE_IN_BYTES;
+}
+
+/// Given that left < right,
+/// mutates `left into a shorter byte string left'` that
+/// matches `left <= left' < right`.
+fn find_shorter_str_in_between(left: &mut Vec<u8>, right: &[u8]) {
+    assert!(&left[..] < right);
+    let common_len = common_prefix_len(left, right);
+    if left.len() == common_len {
+        return;
+    }
+    // It is possible to do one character shorter in some case,
+    // but it is not worth the extra complexity
+    for pos in (common_len + 1)..left.len() {
+        if left[pos] != u8::MAX {
+            left[pos] += 1;
+            left.truncate(pos + 1);
+            return;
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct SSTableIndexBuilder {
+    blocks: Vec<BlockMeta>,
+}
+
+impl SSTableIndexBuilder {
+    /// In order to make the index as light as possible, we
+    /// try to find a shorter alternative to the last key of the last block
+    /// that is still smaller than the next key.
+    pub(crate) fn shorten_last_block_key_given_next_key(&mut self, next_key: &[u8]) {
+        if let Some(last_block) = self.blocks.last_mut() {
+            find_shorter_str_in_between(&mut last_block.last_key_or_greater, next_key);
+        }
+    }
+
+    pub fn add_block(&mut self, last_key: &[u8], byte_range: Range<usize>, first_ordinal: u64) {
+        self.blocks.push(BlockMeta {
+            last_key_or_greater: last_key.to_vec(),
+            block_addr: BlockAddr {
+                byte_range,
+                first_ordinal,
+            },
+        })
+    }
+
+    pub fn serialize<W: std::io::Write>(&self, wrt: W) -> io::Result<u64> {
+        if self.blocks.len() <= 1 {
+            return Ok(0);
+        }
+        let counting_writer = common::CountingWriter::wrap(wrt);
+        let mut map_builder = MapBuilder::new(counting_writer).map_err(fst_error_to_io_error)?;
+        for (i, block) in self.blocks.iter().enumerate() {
+            map_builder
+                .insert(&block.last_key_or_greater, i as u64)
+                .map_err(fst_error_to_io_error)?;
+        }
+        let counting_writer = map_builder.into_inner().map_err(fst_error_to_io_error)?;
+        let written_bytes = counting_writer.written_bytes();
+        let mut wrt = counting_writer.finish();
+
+        let mut block_store_writer = BlockAddrStoreWriter::new();
+        for block in &self.blocks {
+            block_store_writer.write_block_meta(block.block_addr.clone())?;
+        }
+        block_store_writer.serialize(&mut wrt)?;
+
+        Ok(written_bytes)
+    }
+}
+
+fn fst_error_to_io_error(error: tantivy_fst::Error) -> io::Error {
+    match error {
+        tantivy_fst::Error::Fst(fst_error) => io::Error::new(io::ErrorKind::Other, fst_error),
+        tantivy_fst::Error::Io(ioerror) => ioerror,
+    }
+}
+
+const STORE_BLOCK_LEN: usize = 128;
+
+#[derive(Debug)]
+struct BlockAddrBlockMetadata {
+    offset: u64,
+    ref_block_addr: BlockStartAddr,
+    range_start_slope: u32,
+    first_ordinal_slope: u32,
+    range_start_nbits: u8,
+    first_ordinal_nbits: u8,
+    block_len: u16,
+    // these fields are computed on deserialization, and not stored
+    range_shift: i64,
+    ordinal_shift: i64,
+}
+
+impl BlockAddrBlockMetadata {
+    fn num_bits(&self) -> u8 {
+        self.first_ordinal_nbits + self.range_start_nbits
+    }
+
+    fn deserialize_block_addr(&self, data: &[u8], inner_offset: usize) -> Option<BlockAddr> {
+        if inner_offset == 0 {
+            let range_end = self.ref_block_addr.byte_range_start
+                + extract_bits(data, 0, self.range_start_nbits) as usize
+                + self.range_start_slope as usize
+                - self.range_shift as usize;
+            return Some(self.ref_block_addr.to_block_addr(range_end));
+        }
+        let inner_offset = inner_offset - 1;
+        if inner_offset >= self.block_len as usize {
+            return None;
+        }
+        let num_bits = self.num_bits() as usize;
+
+        let range_start_addr = num_bits * inner_offset;
+        let ordinal_addr = range_start_addr + self.range_start_nbits as usize;
+        let range_end_addr = range_start_addr + num_bits;
+
+        if (range_end_addr + self.range_start_nbits as usize + 7) / 8 > data.len() {
+            return None;
+        }
+
+        let range_start = self.ref_block_addr.byte_range_start
+            + extract_bits(data, range_start_addr, self.range_start_nbits) as usize
+            + self.range_start_slope as usize * (inner_offset + 1)
+            - self.range_shift as usize;
+        let first_ordinal = self.ref_block_addr.first_ordinal
+            + extract_bits(data, ordinal_addr, self.first_ordinal_nbits)
+            + self.first_ordinal_slope as u64 * (inner_offset + 1) as u64
+            - self.ordinal_shift as u64;
+        let range_end = self.ref_block_addr.byte_range_start
+            + extract_bits(data, range_end_addr, self.range_start_nbits) as usize
+            + self.range_start_slope as usize * (inner_offset + 2)
+            - self.range_shift as usize;
+
+        Some(BlockAddr {
+            first_ordinal,
+            byte_range: range_start..range_end,
+        })
+    }
+
+    fn bisect_for_ord(&self, data: &[u8], target_ord: TermOrdinal) -> (u64, BlockAddr) {
+        let inner_target_ord = target_ord - self.ref_block_addr.first_ordinal;
+        let num_bits = self.num_bits() as usize;
+        let range_start_nbits = self.range_start_nbits as usize;
+        let get_ord = |index| {
+            extract_bits(
+                data,
+                num_bits * index as usize + range_start_nbits,
+                self.first_ordinal_nbits,
+            ) + self.first_ordinal_slope as u64 * (index + 1)
+                - self.ordinal_shift as u64
+        };
+
+        let inner_offset = match binary_search(self.block_len as u64, |index| {
+            get_ord(index).cmp(&inner_target_ord)
+        }) {
+            Ok(inner_offset) => inner_offset + 1,
+            Err(inner_offset) => inner_offset,
+        };
+        // we can unwrap because inner_offset <= self.block_len
+        (
+            inner_offset,
+            self.deserialize_block_addr(data, inner_offset as usize)
+                .unwrap(),
+        )
+    }
+}
+
+// TODO move this function to tantivy_common?
+#[inline(always)]
+fn extract_bits(data: &[u8], addr_bits: usize, num_bits: u8) -> u64 {
+    assert!(num_bits <= 56);
+    let addr_byte = addr_bits / 8;
+    let bit_shift = (addr_bits % 8) as u64;
+    let val_unshifted_unmasked: u64 = if data.len() >= addr_byte + 8 {
+        let b = data[addr_byte..addr_byte + 8].try_into().unwrap();
+        u64::from_le_bytes(b)
+    } else {
+        // the buffer is not large enough.
+        // Let's copy the few remaining bytes to a 8 byte buffer
+        // padded with 0s.
+        let mut buf = [0u8; 8];
+        let data_to_copy = &data[addr_byte..];
+        let nbytes = data_to_copy.len();
+        buf[..nbytes].copy_from_slice(data_to_copy);
+        u64::from_le_bytes(buf)
+    };
+    let val_shifted_unmasked = val_unshifted_unmasked >> bit_shift;
+    let mask = (1u64 << u64::from(num_bits)) - 1;
+    val_shifted_unmasked & mask
+}
+
+impl BinarySerializable for BlockAddrBlockMetadata {
+    fn serialize<W: Write + ?Sized>(&self, write: &mut W) -> io::Result<()> {
+        self.offset.serialize(write)?;
+        self.ref_block_addr.serialize(write)?;
+        self.range_start_slope.serialize(write)?;
+        self.first_ordinal_slope.serialize(write)?;
+        write.write_all(&[self.first_ordinal_nbits, self.range_start_nbits])?;
+        self.block_len.serialize(write)?;
+        self.num_bits();
+        Ok(())
+    }
+
+    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
+        let offset = u64::deserialize(reader)?;
+        let ref_block_addr = BlockStartAddr::deserialize(reader)?;
+        let range_start_slope = u32::deserialize(reader)?;
+        let first_ordinal_slope = u32::deserialize(reader)?;
+        let mut buffer = [0u8; 2];
+        reader.read_exact(&mut buffer)?;
+        let first_ordinal_nbits = buffer[0];
+        let range_start_nbits = buffer[1];
+        let block_len = u16::deserialize(reader)?;
+        Ok(BlockAddrBlockMetadata {
+            offset,
+            ref_block_addr,
+            range_start_slope,
+            first_ordinal_slope,
+            range_start_nbits,
+            first_ordinal_nbits,
+            block_len,
+            range_shift: 1 << (range_start_nbits - 1),
+            ordinal_shift: 1 << (first_ordinal_nbits - 1),
+        })
+    }
+}
+
+impl FixedSize for BlockAddrBlockMetadata {
+    const SIZE_IN_BYTES: usize = u64::SIZE_IN_BYTES
+        + BlockStartAddr::SIZE_IN_BYTES
+        + 2 * u32::SIZE_IN_BYTES
+        + 2 * u8::SIZE_IN_BYTES
+        + u16::SIZE_IN_BYTES;
+}
+
+#[derive(Debug, Clone)]
+struct BlockAddrStore {
+    block_meta_bytes: OwnedBytes,
+    addr_bytes: OwnedBytes,
+}
+
+impl BlockAddrStore {
+    fn open(term_info_store_file: OwnedBytes) -> io::Result<BlockAddrStore> {
+        let (mut len_slice, main_slice) = term_info_store_file.split(8);
+        let len = u64::deserialize(&mut len_slice)? as usize;
+        let (block_meta_bytes, addr_bytes) = main_slice.split(len);
+        Ok(BlockAddrStore {
+            block_meta_bytes,
+            addr_bytes,
+        })
+    }
+
+    fn get_block_meta(&self, store_block_id: usize) -> Option<BlockAddrBlockMetadata> {
+        let mut block_data: &[u8] = self
+            .block_meta_bytes
+            .get(store_block_id * BlockAddrBlockMetadata::SIZE_IN_BYTES..)?;
+        BlockAddrBlockMetadata::deserialize(&mut block_data).ok()
+    }
+
+    fn get(&self, block_id: u64) -> Option<BlockAddr> {
+        let store_block_id = (block_id as usize) / STORE_BLOCK_LEN;
+        let inner_offset = (block_id as usize) % STORE_BLOCK_LEN;
+        let block_addr_block_data = self.get_block_meta(store_block_id)?;
+        block_addr_block_data.deserialize_block_addr(
+            &self.addr_bytes[block_addr_block_data.offset as usize..],
+            inner_offset,
+        )
+    }
+
+    fn binary_search_ord(&self, ord: TermOrdinal) -> (u64, BlockAddr) {
+        let max_block =
+            (self.block_meta_bytes.len() / BlockAddrBlockMetadata::SIZE_IN_BYTES) as u64;
+        let get_first_ordinal = |block_id| {
+            // we can unwrap because block_id < max_block
+            self.get(block_id * STORE_BLOCK_LEN as u64)
+                .unwrap()
+                .first_ordinal
+        };
+        let store_block_id =
+            binary_search(max_block, |block_id| get_first_ordinal(block_id).cmp(&ord));
+        let store_block_id = match store_block_id {
+            Ok(store_block_id) => {
+                let block_id = store_block_id * STORE_BLOCK_LEN as u64;
+                // we can unwrap because store_block_id < max_block
+                return (block_id, self.get(block_id).unwrap());
+            }
+            Err(store_block_id) => store_block_id - 1,
+        };
+
+        // we can unwrap because store_block_id < max_block
+        let block_addr_block_data = self.get_block_meta(store_block_id as usize).unwrap();
+        let (inner_offset, block_addr) = block_addr_block_data.bisect_for_ord(
+            &self.addr_bytes[block_addr_block_data.offset as usize..],
+            ord,
+        );
+        (
+            store_block_id * STORE_BLOCK_LEN as u64 + inner_offset,
+            block_addr,
+        )
+    }
+}
+
+fn binary_search(max: u64, cmp_fn: impl Fn(u64) -> std::cmp::Ordering) -> Result<u64, u64> {
+    use std::cmp::Ordering::*;
+    let mut size = max;
+    let mut left = 0;
+    let mut right = size;
+    while left < right {
+        let mid = left + size / 2;
+
+        let cmp = cmp_fn(mid);
+
+        if cmp == Less {
+            left = mid + 1;
+        } else if cmp == Greater {
+            right = mid;
+        } else {
+            return Ok(mid);
+        }
+
+        size = right - left;
+    }
+    Err(left)
+}
+
+struct BlockAddrStoreWriter {
+    buffer_block_metas: Vec<u8>,
+    buffer_addrs: Vec<u8>,
+    block_addrs: Vec<BlockAddr>,
+}
+
+impl BlockAddrStoreWriter {
+    fn new() -> Self {
+        BlockAddrStoreWriter {
+            buffer_block_metas: Vec::new(),
+            buffer_addrs: Vec::new(),
+            block_addrs: Vec::with_capacity(STORE_BLOCK_LEN),
+        }
+    }
+
+    fn flush_block(&mut self) -> io::Result<()> {
+        if self.block_addrs.is_empty() {
+            return Ok(());
+        }
+        let ref_block_addr = self.block_addrs[0].clone();
+
+        for block_addr in &mut self.block_addrs {
+            block_addr.byte_range.start -= ref_block_addr.byte_range.start;
+            block_addr.first_ordinal -= ref_block_addr.first_ordinal;
+        }
+
+        // we are only called if block_addrs is not empty
+        let mut last_block_addr = self.block_addrs.last().unwrap().clone();
+        last_block_addr.byte_range.end -= ref_block_addr.byte_range.start;
+
+        // we skip(1), so we never give an index of 0 to find_best_slope
+        let (range_start_slope, range_start_nbits) = find_best_slope(
+            self.block_addrs
+                .iter()
+                .map(|block| block.byte_range.start as u64)
+                .chain(std::iter::once(last_block_addr.byte_range.end as u64))
+                .enumerate()
+                .skip(1),
+        );
+
+        // we skip(1), so we never give an index of 0 to find_best_slope
+        let (first_ordinal_slope, first_ordinal_nbits) = find_best_slope(
+            self.block_addrs
+                .iter()
+                .map(|block| block.first_ordinal)
+                .enumerate()
+                .skip(1),
+        );
+
+        let range_shift = 1 << (range_start_nbits - 1);
+        let ordinal_shift = 1 << (first_ordinal_nbits - 1);
+
+        let block_addr_block_meta = BlockAddrBlockMetadata {
+            offset: self.buffer_addrs.len() as u64,
+            ref_block_addr: ref_block_addr.to_block_start(),
+            range_start_slope,
+            first_ordinal_slope,
+            range_start_nbits,
+            first_ordinal_nbits,
+            block_len: self.block_addrs.len() as u16 - 1,
+            range_shift,
+            ordinal_shift,
+        };
+        block_addr_block_meta.serialize(&mut self.buffer_block_metas)?;
+
+        let mut bit_packer = BitPacker::new();
+
+        for (i, block_addr) in self.block_addrs.iter().enumerate().skip(1) {
+            let range_pred = (range_start_slope as usize * i) as i64;
+            bit_packer.write(
+                (block_addr.byte_range.start as i64 - range_pred + range_shift) as u64,
+                range_start_nbits,
+                &mut self.buffer_addrs,
+            )?;
+            let first_ordinal_pred = (first_ordinal_slope as u64 * i as u64) as i64;
+            bit_packer.write(
+                (block_addr.first_ordinal as i64 - first_ordinal_pred + ordinal_shift) as u64,
+                first_ordinal_nbits,
+                &mut self.buffer_addrs,
+            )?;
+        }
+
+        let range_pred = (range_start_slope as usize * self.block_addrs.len()) as i64;
+        bit_packer.write(
+            (last_block_addr.byte_range.end as i64 - range_pred + range_shift) as u64,
+            range_start_nbits,
+            &mut self.buffer_addrs,
+        )?;
+        bit_packer.flush(&mut self.buffer_addrs)?;
+
+        self.block_addrs.clear();
+        Ok(())
+    }
+
+    fn write_block_meta(&mut self, block_addr: BlockAddr) -> io::Result<()> {
+        self.block_addrs.push(block_addr);
+        if self.block_addrs.len() >= STORE_BLOCK_LEN {
+            self.flush_block()?;
+        }
+        Ok(())
+    }
+
+    fn serialize<W: std::io::Write>(&mut self, wrt: &mut W) -> io::Result<()> {
+        self.flush_block()?;
+        let len = self.buffer_block_metas.len() as u64;
+        len.serialize(wrt)?;
+        wrt.write_all(&self.buffer_block_metas)?;
+        wrt.write_all(&self.buffer_addrs)?;
+        Ok(())
+    }
+}
+
+/// Given an iterator over (index, value), returns the slope, and number of bits needed to
+/// represente the error to a prediction made by this slope.
+///
+/// The iterator may be empty, but all indexes in it must be non-zero.
+fn find_best_slope(elements: impl Iterator<Item = (usize, u64)> + Clone) -> (u32, u8) {
+    let slope_iterator = elements.clone();
+    let derivation_iterator = elements;
+
+    let mut min_slope_idx = 1;
+    let mut min_slope_val = 0;
+    let mut min_slope = u32::MAX;
+    let mut max_slope_idx = 1;
+    let mut max_slope_val = 0;
+    let mut max_slope = 0;
+    for (index, value) in slope_iterator {
+        let slope = (value / index as u64) as u32;
+        if slope <= min_slope {
+            min_slope = slope;
+            min_slope_idx = index;
+            min_slope_val = value;
+        }
+        if slope >= max_slope {
+            max_slope = slope;
+            max_slope_idx = index;
+            max_slope_val = value;
+        }
+    }
+
+    // above is an heuristic giving the "highest" and "lowest" point. It's imperfect in that in that
+    // a point that appear earlier might have a high slope derivation, but a smaller absolute
+    // derivation than a latter point.
+    // The actual best values can be obtained by using the symplex method, but the improvement is
+    // likely minimal, and computation is way more complexe.
+    //
+    // Assuming these point are the furthest up and down, we find the slope that would cause the
+    // same positive derivation for the highest as negative derivation for the lowest.
+    // A is the optimal slope. B is the derivation to the guess
+    //
+    // 0 = min_slope_val - min_slope_idx * A - B
+    // 0 = max_slope_val - max_slope_idx * A + B
+    //
+    // 0 = min_slope_val + max_slope_val - (min_slope_idx + max_slope_idx) * A
+    // (min_slope_val + max_slope_val) / (min_slope_idx + max_slope_idx) = A
+    //
+    // we actually add some correcting factor to have proper rounding, not truncation.
+
+    let denominator = (min_slope_idx + max_slope_idx) as u64;
+    let final_slope = ((min_slope_val + max_slope_val + denominator / 2) / denominator) as u32;
+
+    // we don't solve for B because our choice of point is suboptimal, so it's actually a lower
+    // bound and we need to iterate to find the actual worst value.
+
+    let max_derivation: u64 = derivation_iterator
+        .map(|(index, value)| (value as i64 - final_slope as i64 * index as i64).unsigned_abs())
+        .max()
+        .unwrap_or(0);
+
+    (final_slope, compute_num_bits(max_derivation) + 1)
+}
+
+#[cfg(test)]
+mod tests {
+    use common::OwnedBytes;
+
+    use super::{BlockAddr, SSTableIndexBuilder, SSTableIndexV3};
+    use crate::SSTableDataCorruption;
+
+    #[test]
+    fn test_sstable_index() {
+        let mut sstable_builder = SSTableIndexBuilder::default();
+        sstable_builder.add_block(b"aaa", 10..20, 0u64);
+        sstable_builder.add_block(b"bbbbbbb", 20..30, 5u64);
+        sstable_builder.add_block(b"ccc", 30..40, 10u64);
+        sstable_builder.add_block(b"dddd", 40..50, 15u64);
+        let mut buffer: Vec<u8> = Vec::new();
+        let fst_len = sstable_builder.serialize(&mut buffer).unwrap();
+        let buffer = OwnedBytes::new(buffer);
+        let sstable_index = SSTableIndexV3::load(buffer, fst_len).unwrap();
+        assert_eq!(
+            sstable_index.get_block_with_key(b"bbbde"),
+            Some(BlockAddr {
+                first_ordinal: 10u64,
+                byte_range: 30..40
+            })
+        );
+
+        assert_eq!(sstable_index.locate_with_key(b"aa").unwrap(), 0);
+        assert_eq!(sstable_index.locate_with_key(b"aaa").unwrap(), 0);
+        assert_eq!(sstable_index.locate_with_key(b"aab").unwrap(), 1);
+        assert_eq!(sstable_index.locate_with_key(b"ccc").unwrap(), 2);
+        assert!(sstable_index.locate_with_key(b"e").is_none());
+
+        assert_eq!(sstable_index.locate_with_ord(0), 0);
+        assert_eq!(sstable_index.locate_with_ord(1), 0);
+        assert_eq!(sstable_index.locate_with_ord(4), 0);
+        assert_eq!(sstable_index.locate_with_ord(5), 1);
+        assert_eq!(sstable_index.locate_with_ord(100), 3);
+    }
+
+    #[test]
+    fn test_sstable_with_corrupted_data() {
+        let mut sstable_builder = SSTableIndexBuilder::default();
+        sstable_builder.add_block(b"aaa", 10..20, 0u64);
+        sstable_builder.add_block(b"bbbbbbb", 20..30, 5u64);
+        sstable_builder.add_block(b"ccc", 30..40, 10u64);
+        sstable_builder.add_block(b"dddd", 40..50, 15u64);
+        let mut buffer: Vec<u8> = Vec::new();
+        let fst_len = sstable_builder.serialize(&mut buffer).unwrap();
+        buffer[2] = 9u8;
+        let buffer = OwnedBytes::new(buffer);
+        let data_corruption_err = SSTableIndexV3::load(buffer, fst_len).err().unwrap();
+        assert!(matches!(data_corruption_err, SSTableDataCorruption));
+    }
+
+    #[track_caller]
+    fn test_find_shorter_str_in_between_aux(left: &[u8], right: &[u8]) {
+        let mut left_buf = left.to_vec();
+        super::find_shorter_str_in_between(&mut left_buf, right);
+        assert!(left_buf.len() <= left.len());
+        assert!(left <= &left_buf);
+        assert!(&left_buf[..] < right);
+    }
+
+    #[test]
+    fn test_find_shorter_str_in_between() {
+        test_find_shorter_str_in_between_aux(b"", b"hello");
+        test_find_shorter_str_in_between_aux(b"abc", b"abcd");
+        test_find_shorter_str_in_between_aux(b"abcd", b"abd");
+        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[1]);
+        test_find_shorter_str_in_between_aux(&[0, 0, 0], &[0, 0, 1]);
+        test_find_shorter_str_in_between_aux(&[0, 0, 255, 255, 255, 0u8], &[0, 1]);
+    }
+
+    use proptest::prelude::*;
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(100))]
+        #[test]
+        fn test_proptest_find_shorter_str(left in any::<Vec<u8>>(), right in any::<Vec<u8>>()) {
+            if left < right {
+                test_find_shorter_str_in_between_aux(&left, &right);
+            }
+        }
+    }
+
+    #[test]
+    fn test_find_best_slop() {
+        assert_eq!(super::find_best_slope(std::iter::empty()), (0, 1));
+        assert_eq!(
+            super::find_best_slope(std::iter::once((1, 12345))),
+            (12345, 1)
+        );
+    }
+}

stacker/src/arena_hashmap.rs

@@ -1,51 +1,5 @@
-use std::iter::{Cloned, Filter};
-use std::mem;
-
 use super::{Addr, MemoryArena};
-use crate::fastcpy::fast_short_slice_copy;
-use crate::memory_arena::store;
-
-/// Returns the actual memory size in bytes
-/// required to create a table with a given capacity.
-/// required to create a table of size
-pub fn compute_table_memory_size(capacity: usize) -> usize {
-    capacity * mem::size_of::<KeyValue>()
-}
-
-#[cfg(not(feature = "compare_hash_only"))]
-type HashType = u32;
-
-#[cfg(feature = "compare_hash_only")]
-type HashType = u64;
-
-/// `KeyValue` is the item stored in the hash table.
-/// The key is actually a `BytesRef` object stored in an external memory arena.
-/// The `value_addr` also points to an address in the memory arena.
-#[derive(Copy, Clone)]
-struct KeyValue {
-    pub(crate) key_value_addr: Addr,
-    hash: HashType,
-}
-
-impl Default for KeyValue {
-    fn default() -> Self {
-        KeyValue {
-            key_value_addr: Addr::null_pointer(),
-            hash: 0,
-        }
-    }
-}
-
-impl KeyValue {
-    #[inline]
-    fn is_empty(&self) -> bool {
-        self.key_value_addr.is_null()
-    }
-    #[inline]
-    fn is_not_empty_ref(&self) -> bool {
-        !self.key_value_addr.is_null()
-    }
-}
+use crate::shared_arena_hashmap::SharedArenaHashMap;
 
 /// Customized `HashMap` with `&[u8]` keys
 ///
@@ -56,61 +10,13 @@ impl KeyValue {
 /// The quirky API has the benefit of avoiding
 /// the computation of the hash of the key twice,
 /// or copying the key as long as there is no insert.
-pub struct ArenaHashMap {
-    table: Vec<KeyValue>,
-    pub memory_arena: MemoryArena,
-    mask: usize,
-    len: usize,
-}
-
-struct LinearProbing {
-    pos: usize,
-    mask: usize,
-}
-
-impl LinearProbing {
-    #[inline]
-    fn compute(hash: HashType, mask: usize) -> LinearProbing {
-        LinearProbing {
-            pos: hash as usize,
-            mask,
-        }
-    }
-
-    #[inline]
-    fn next_probe(&mut self) -> usize {
-        // Not saving the masked version removes a dependency.
-        self.pos = self.pos.wrapping_add(1);
-        self.pos & self.mask
-    }
-}
-
-type IterNonEmpty<'a> = Filter<Cloned<std::slice::Iter<'a, KeyValue>>, fn(&KeyValue) -> bool>;
-
-pub struct Iter<'a> {
-    hashmap: &'a ArenaHashMap,
-    inner: IterNonEmpty<'a>,
-}
-
-impl<'a> Iterator for Iter<'a> {
-    type Item = (&'a [u8], Addr);
-
-    fn next(&mut self) -> Option<Self::Item> {
-        self.inner.next().map(move |kv| {
-            let (key, offset): (&'a [u8], Addr) = self.hashmap.get_key_value(kv.key_value_addr);
-            (key, offset)
-        })
-    }
-}
-
-/// Returns the greatest power of two lower or equal to `n`.
-/// Except if n == 0, in that case, return 1.
 ///
-/// # Panics if n == 0
-fn compute_previous_power_of_two(n: usize) -> usize {
-    assert!(n > 0);
-    let msb = (63u32 - (n as u64).leading_zeros()) as u8;
-    1 << msb
+/// ArenaHashMap is like SharedArenaHashMap but takes ownership
+/// of the memory arena. The memory arena stores the serialized
+/// keys and values.
+pub struct ArenaHashMap {
+    shared_arena_hashmap: SharedArenaHashMap,
+    pub memory_arena: MemoryArena,
 }
 
 impl Default for ArenaHashMap {
@@ -121,156 +27,44 @@ impl Default for ArenaHashMap {
 
 impl ArenaHashMap {
     pub fn with_capacity(table_size: usize) -> ArenaHashMap {
-        let table_size_power_of_2 = compute_previous_power_of_two(table_size);
         let memory_arena = MemoryArena::default();
-        let table = vec![KeyValue::default(); table_size_power_of_2];
 
         ArenaHashMap {
-            table,
+            shared_arena_hashmap: SharedArenaHashMap::with_capacity(table_size),
             memory_arena,
-            mask: table_size_power_of_2 - 1,
-            len: 0,
         }
     }
 
-    #[inline]
-    #[cfg(not(feature = "compare_hash_only"))]
-    fn get_hash(&self, key: &[u8]) -> HashType {
-        murmurhash32::murmurhash2(key)
-    }
-
-    #[inline]
-    #[cfg(feature = "compare_hash_only")]
-    fn get_hash(&self, key: &[u8]) -> HashType {
-        /// Since we compare only the hash we need a high quality hash.
-        use std::hash::Hasher;
-        let mut hasher = ahash::AHasher::default();
-        hasher.write(key);
-        hasher.finish() as HashType
-    }
-
     #[inline]
     pub fn read<Item: Copy + 'static>(&self, addr: Addr) -> Item {
         self.memory_arena.read(addr)
     }
 
-    #[inline]
-    fn probe(&self, hash: HashType) -> LinearProbing {
-        LinearProbing::compute(hash, self.mask)
-    }
-
     #[inline]
     pub fn mem_usage(&self) -> usize {
-        self.table.len() * mem::size_of::<KeyValue>() + self.memory_arena.mem_usage()
-    }
-
-    #[inline]
-    fn is_saturated(&self) -> bool {
-        self.table.len() <= self.len * 2
-    }
-
-    #[inline]
-    fn get_key_value(&self, addr: Addr) -> (&[u8], Addr) {
-        let data = self.memory_arena.slice_from(addr);
-        let key_bytes_len_bytes = unsafe { data.get_unchecked(..2) };
-        let key_bytes_len = u16::from_le_bytes(key_bytes_len_bytes.try_into().unwrap());
-        let key_bytes: &[u8] = unsafe { data.get_unchecked(2..2 + key_bytes_len as usize) };
-        (key_bytes, addr.offset(2 + key_bytes_len as u32))
-    }
-
-    #[inline]
-    #[cfg(not(feature = "compare_hash_only"))]
-    fn get_value_addr_if_key_match(&self, target_key: &[u8], addr: Addr) -> Option<Addr> {
-        use crate::fastcmp::fast_short_slice_compare;
-
-        let (stored_key, value_addr) = self.get_key_value(addr);
-        if fast_short_slice_compare(stored_key, target_key) {
-            Some(value_addr)
-        } else {
-            None
-        }
-    }
-    #[inline]
-    #[cfg(feature = "compare_hash_only")]
-    fn get_value_addr_if_key_match(&self, _target_key: &[u8], addr: Addr) -> Option<Addr> {
-        // For the compare_hash_only feature, it would make sense to store the keys at a different
-        // memory location. Here they will just pollute the cache.
-        let data = self.memory_arena.slice_from(addr);
-        let key_bytes_len_bytes = &data[..2];
-        let key_bytes_len = u16::from_le_bytes(key_bytes_len_bytes.try_into().unwrap());
-        let value_addr = addr.offset(2 + key_bytes_len as u32);
-
-        Some(value_addr)
-    }
-
-    #[inline]
-    fn set_bucket(&mut self, hash: HashType, key_value_addr: Addr, bucket: usize) {
-        self.len += 1;
-
-        self.table[bucket] = KeyValue {
-            key_value_addr,
-            hash,
-        };
+        self.shared_arena_hashmap.mem_usage() + self.memory_arena.mem_usage()
     }
 
     #[inline]
     pub fn is_empty(&self) -> bool {
-        self.len() == 0
+        self.shared_arena_hashmap.is_empty()
     }
 
     #[inline]
     pub fn len(&self) -> usize {
-        self.len
+        self.shared_arena_hashmap.len()
     }
 
     #[inline]
-    pub fn iter(&self) -> Iter<'_> {
-        Iter {
-            inner: self
-                .table
-                .iter()
-                .cloned()
-                .filter(KeyValue::is_not_empty_ref),
-            hashmap: self,
-        }
-    }
-
-    fn resize(&mut self) {
-        let new_len = (self.table.len() * 2).max(1 << 13);
-        let mask = new_len - 1;
-        self.mask = mask;
-        let new_table = vec![KeyValue::default(); new_len];
-        let old_table = mem::replace(&mut self.table, new_table);
-        for key_value in old_table.into_iter().filter(KeyValue::is_not_empty_ref) {
-            let mut probe = LinearProbing::compute(key_value.hash, mask);
-            loop {
-                let bucket = probe.next_probe();
-                if self.table[bucket].is_empty() {
-                    self.table[bucket] = key_value;
-                    break;
-                }
-            }
-        }
+    pub fn iter(&self) -> impl Iterator<Item = (&[u8], Addr)> {
+        self.shared_arena_hashmap.iter(&self.memory_arena)
     }
 
     /// Get a value associated to a key.
     #[inline]
     pub fn get<V>(&self, key: &[u8]) -> Option<V>
     where V: Copy + 'static {
-        let hash = self.get_hash(key);
-        let mut probe = self.probe(hash);
-        loop {
-            let bucket = probe.next_probe();
-            let kv: KeyValue = self.table[bucket];
-            if kv.is_empty() {
-                return None;
-            } else if kv.hash == hash {
-                if let Some(val_addr) = self.get_value_addr_if_key_match(key, kv.key_value_addr) {
-                    let v = self.memory_arena.read(val_addr);
-                    return Some(v);
-                }
-            }
-        }
+        self.shared_arena_hashmap.get(key, &self.memory_arena)
     }
 
     /// `update` create a new entry for a given key if it does not exist
@@ -284,45 +78,10 @@ impl ArenaHashMap {
     /// If the key already as an associated value, then it will be passed
     /// `Some(previous_value)`.
     #[inline]
-    pub fn mutate_or_create<V>(&mut self, key: &[u8], mut updater: impl FnMut(Option<V>) -> V)
+    pub fn mutate_or_create<V>(&mut self, key: &[u8], updater: impl FnMut(Option<V>) -> V)
     where V: Copy + 'static {
-        if self.is_saturated() {
-            self.resize();
-        }
-        let hash = self.get_hash(key);
-        let mut probe = self.probe(hash);
-        let mut bucket = probe.next_probe();
-        let mut kv: KeyValue = self.table[bucket];
-        loop {
-            if kv.is_empty() {
-                // The key does not exist yet.
-                let val = updater(None);
-                let num_bytes = std::mem::size_of::<u16>() + key.len() + std::mem::size_of::<V>();
-                let key_addr = self.memory_arena.allocate_space(num_bytes);
-                {
-                    let data = self.memory_arena.slice_mut(key_addr, num_bytes);
-                    let key_len_bytes: [u8; 2] = (key.len() as u16).to_le_bytes();
-                    data[..2].copy_from_slice(&key_len_bytes);
-                    let stop = 2 + key.len();
-                    fast_short_slice_copy(key, &mut data[2..stop]);
-                    store(&mut data[stop..], val);
-                }
-
-                self.set_bucket(hash, key_addr, bucket);
-                return;
-            }
-            if kv.hash == hash {
-                if let Some(val_addr) = self.get_value_addr_if_key_match(key, kv.key_value_addr) {
-                    let v = self.memory_arena.read(val_addr);
-                    let new_v = updater(Some(v));
-                    self.memory_arena.write_at(val_addr, new_v);
-                    return;
-                }
-            }
-            // This allows fetching the next bucket before the loop jmp
-            bucket = probe.next_probe();
-            kv = self.table[bucket];
-        }
+        self.shared_arena_hashmap
+            .mutate_or_create(key, &mut self.memory_arena, updater);
     }
 }
 
@@ -331,7 +90,7 @@ mod tests {
 
     use std::collections::HashMap;
 
-    use super::{compute_previous_power_of_two, ArenaHashMap};
+    use super::ArenaHashMap;
 
     #[test]
     fn test_hash_map() {
@@ -362,14 +121,6 @@ mod tests {
         assert_eq!(hash_map.get::<u32>(b"abc"), None);
     }
 
-    #[test]
-    fn test_compute_previous_power_of_two() {
-        assert_eq!(compute_previous_power_of_two(8), 8);
-        assert_eq!(compute_previous_power_of_two(9), 8);
-        assert_eq!(compute_previous_power_of_two(7), 4);
-        assert_eq!(compute_previous_power_of_two(u64::MAX as usize), 1 << 63);
-    }
-
     #[test]
     fn test_many_terms() {
         let mut terms: Vec<String> = (0..20_000).map(|val| val.to_string()).collect();

stacker/src/lib.rs

@@ -9,10 +9,12 @@ mod expull;
 mod fastcmp;
 mod fastcpy;
 mod memory_arena;
+mod shared_arena_hashmap;
 
-pub use self::arena_hashmap::{compute_table_memory_size, ArenaHashMap};
+pub use self::arena_hashmap::ArenaHashMap;
 pub use self::expull::ExpUnrolledLinkedList;
 pub use self::memory_arena::{Addr, MemoryArena};
+pub use self::shared_arena_hashmap::{compute_table_memory_size, SharedArenaHashMap};
 
 /// When adding an element in a `ArenaHashMap`, we get a unique id associated to the given key.
 pub type UnorderedId = u32;

stacker/src/shared_arena_hashmap.rs

@@ -0,0 +1,420 @@
+use std::iter::{Cloned, Filter};
+use std::mem;
+
+use super::{Addr, MemoryArena};
+use crate::fastcpy::fast_short_slice_copy;
+use crate::memory_arena::store;
+
+/// Returns the actual memory size in bytes
+/// required to create a table with a given capacity.
+/// required to create a table of size
+pub fn compute_table_memory_size(capacity: usize) -> usize {
+    capacity * mem::size_of::<KeyValue>()
+}
+
+#[cfg(not(feature = "compare_hash_only"))]
+type HashType = u32;
+
+#[cfg(feature = "compare_hash_only")]
+type HashType = u64;
+
+/// `KeyValue` is the item stored in the hash table.
+/// The key is actually a `BytesRef` object stored in an external memory arena.
+/// The `value_addr` also points to an address in the memory arena.
+#[derive(Copy, Clone)]
+struct KeyValue {
+    key_value_addr: Addr,
+    hash: HashType,
+}
+
+impl Default for KeyValue {
+    fn default() -> Self {
+        KeyValue {
+            key_value_addr: Addr::null_pointer(),
+            hash: 0,
+        }
+    }
+}
+
+impl KeyValue {
+    #[inline]
+    fn is_empty(&self) -> bool {
+        self.key_value_addr.is_null()
+    }
+    #[inline]
+    fn is_not_empty_ref(&self) -> bool {
+        !self.key_value_addr.is_null()
+    }
+}
+
+/// Customized `HashMap` with `&[u8]` keys
+///
+/// Its main particularity is that rather than storing its
+/// keys in the heap, keys are stored in a memory arena
+/// inline with the values.
+///
+/// The quirky API has the benefit of avoiding
+/// the computation of the hash of the key twice,
+/// or copying the key as long as there is no insert.
+///
+/// SharedArenaHashMap is like ArenaHashMap but gets the memory arena
+/// passed as an argument to the methods.
+/// So one MemoryArena can be shared with multiple SharedArenaHashMap.
+pub struct SharedArenaHashMap {
+    table: Vec<KeyValue>,
+    mask: usize,
+    len: usize,
+}
+
+struct LinearProbing {
+    pos: usize,
+    mask: usize,
+}
+
+impl LinearProbing {
+    #[inline]
+    fn compute(hash: HashType, mask: usize) -> LinearProbing {
+        LinearProbing {
+            pos: hash as usize,
+            mask,
+        }
+    }
+
+    #[inline]
+    fn next_probe(&mut self) -> usize {
+        // Not saving the masked version removes a dependency.
+        self.pos = self.pos.wrapping_add(1);
+        self.pos & self.mask
+    }
+}
+
+type IterNonEmpty<'a> = Filter<Cloned<std::slice::Iter<'a, KeyValue>>, fn(&KeyValue) -> bool>;
+
+pub struct Iter<'a> {
+    hashmap: &'a SharedArenaHashMap,
+    memory_arena: &'a MemoryArena,
+    inner: IterNonEmpty<'a>,
+}
+
+impl<'a> Iterator for Iter<'a> {
+    type Item = (&'a [u8], Addr);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        self.inner.next().map(move |kv| {
+            let (key, offset): (&'a [u8], Addr) = self
+                .hashmap
+                .get_key_value(kv.key_value_addr, self.memory_arena);
+            (key, offset)
+        })
+    }
+}
+
+/// Returns the greatest power of two lower or equal to `n`.
+/// Except if n == 0, in that case, return 1.
+///
+/// # Panics if n == 0
+fn compute_previous_power_of_two(n: usize) -> usize {
+    assert!(n > 0);
+    let msb = (63u32 - (n as u64).leading_zeros()) as u8;
+    1 << msb
+}
+
+impl Default for SharedArenaHashMap {
+    fn default() -> Self {
+        SharedArenaHashMap::with_capacity(4)
+    }
+}
+
+impl SharedArenaHashMap {
+    pub fn with_capacity(table_size: usize) -> SharedArenaHashMap {
+        let table_size_power_of_2 = compute_previous_power_of_two(table_size);
+        let table = vec![KeyValue::default(); table_size_power_of_2];
+
+        SharedArenaHashMap {
+            table,
+            mask: table_size_power_of_2 - 1,
+            len: 0,
+        }
+    }
+
+    #[inline]
+    #[cfg(not(feature = "compare_hash_only"))]
+    fn get_hash(&self, key: &[u8]) -> HashType {
+        murmurhash32::murmurhash2(key)
+    }
+
+    #[inline]
+    #[cfg(feature = "compare_hash_only")]
+    fn get_hash(&self, key: &[u8]) -> HashType {
+        /// Since we compare only the hash we need a high quality hash.
+        use std::hash::Hasher;
+        let mut hasher = ahash::AHasher::default();
+        hasher.write(key);
+        hasher.finish() as HashType
+    }
+
+    #[inline]
+    fn probe(&self, hash: HashType) -> LinearProbing {
+        LinearProbing::compute(hash, self.mask)
+    }
+
+    #[inline]
+    pub fn mem_usage(&self) -> usize {
+        self.table.len() * mem::size_of::<KeyValue>()
+    }
+
+    #[inline]
+    fn is_saturated(&self) -> bool {
+        self.table.len() <= self.len * 2
+    }
+
+    #[inline]
+    fn get_key_value<'a>(&'a self, addr: Addr, memory_arena: &'a MemoryArena) -> (&[u8], Addr) {
+        let data = memory_arena.slice_from(addr);
+        let key_bytes_len_bytes = unsafe { data.get_unchecked(..2) };
+        let key_bytes_len = u16::from_le_bytes(key_bytes_len_bytes.try_into().unwrap());
+        let key_bytes: &[u8] = unsafe { data.get_unchecked(2..2 + key_bytes_len as usize) };
+        (key_bytes, addr.offset(2 + key_bytes_len as u32))
+    }
+
+    #[inline]
+    #[cfg(not(feature = "compare_hash_only"))]
+    fn get_value_addr_if_key_match(
+        &self,
+        target_key: &[u8],
+        addr: Addr,
+        memory_arena: &MemoryArena,
+    ) -> Option<Addr> {
+        use crate::fastcmp::fast_short_slice_compare;
+
+        let (stored_key, value_addr) = self.get_key_value(addr, memory_arena);
+        if fast_short_slice_compare(stored_key, target_key) {
+            Some(value_addr)
+        } else {
+            None
+        }
+    }
+    #[inline]
+    #[cfg(feature = "compare_hash_only")]
+    fn get_value_addr_if_key_match(
+        &self,
+        _target_key: &[u8],
+        addr: Addr,
+        memory_arena: &MemoryArena,
+    ) -> Option<Addr> {
+        // For the compare_hash_only feature, it would make sense to store the keys at a different
+        // memory location. Here they will just pollute the cache.
+        let data = memory_arena.slice_from(addr);
+        let key_bytes_len_bytes = &data[..2];
+        let key_bytes_len = u16::from_le_bytes(key_bytes_len_bytes.try_into().unwrap());
+        let value_addr = addr.offset(2 + key_bytes_len as u32);
+
+        Some(value_addr)
+    }
+
+    #[inline]
+    fn set_bucket(&mut self, hash: HashType, key_value_addr: Addr, bucket: usize) {
+        self.len += 1;
+
+        self.table[bucket] = KeyValue {
+            key_value_addr,
+            hash,
+        };
+    }
+
+    #[inline]
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.len
+    }
+
+    #[inline]
+    pub fn iter<'a>(&'a self, memory_arena: &'a MemoryArena) -> Iter<'_> {
+        Iter {
+            inner: self
+                .table
+                .iter()
+                .cloned()
+                .filter(KeyValue::is_not_empty_ref),
+            hashmap: self,
+            memory_arena,
+        }
+    }
+
+    fn resize(&mut self) {
+        let new_len = (self.table.len() * 2).max(1 << 3);
+        let mask = new_len - 1;
+        self.mask = mask;
+        let new_table = vec![KeyValue::default(); new_len];
+        let old_table = mem::replace(&mut self.table, new_table);
+        for key_value in old_table.into_iter().filter(KeyValue::is_not_empty_ref) {
+            let mut probe = LinearProbing::compute(key_value.hash, mask);
+            loop {
+                let bucket = probe.next_probe();
+                if self.table[bucket].is_empty() {
+                    self.table[bucket] = key_value;
+                    break;
+                }
+            }
+        }
+    }
+
+    /// Get a value associated to a key.
+    #[inline]
+    pub fn get<V>(&self, key: &[u8], memory_arena: &MemoryArena) -> Option<V>
+    where V: Copy + 'static {
+        let hash = self.get_hash(key);
+        let mut probe = self.probe(hash);
+        loop {
+            let bucket = probe.next_probe();
+            let kv: KeyValue = self.table[bucket];
+            if kv.is_empty() {
+                return None;
+            } else if kv.hash == hash {
+                if let Some(val_addr) =
+                    self.get_value_addr_if_key_match(key, kv.key_value_addr, memory_arena)
+                {
+                    let v = memory_arena.read(val_addr);
+                    return Some(v);
+                }
+            }
+        }
+    }
+
+    /// `update` create a new entry for a given key if it does not exist
+    /// or updates the existing entry.
+    ///
+    /// The actual logic for this update is define in the `updater`
+    /// argument.
+    ///
+    /// If the key is not present, `updater` will receive `None` and
+    /// will be in charge of returning a default value.
+    /// If the key already as an associated value, then it will be passed
+    /// `Some(previous_value)`.
+    #[inline]
+    pub fn mutate_or_create<V>(
+        &mut self,
+        key: &[u8],
+        memory_arena: &mut MemoryArena,
+        mut updater: impl FnMut(Option<V>) -> V,
+    ) -> V
+    where
+        V: Copy + 'static,
+    {
+        if self.is_saturated() {
+            self.resize();
+        }
+        let hash = self.get_hash(key);
+        let mut probe = self.probe(hash);
+        let mut bucket = probe.next_probe();
+        let mut kv: KeyValue = self.table[bucket];
+        loop {
+            if kv.is_empty() {
+                // The key does not exist yet.
+                let val = updater(None);
+                let num_bytes = std::mem::size_of::<u16>() + key.len() + std::mem::size_of::<V>();
+                let key_addr = memory_arena.allocate_space(num_bytes);
+                {
+                    let data = memory_arena.slice_mut(key_addr, num_bytes);
+                    let key_len_bytes: [u8; 2] = (key.len() as u16).to_le_bytes();
+                    data[..2].copy_from_slice(&key_len_bytes);
+                    let stop = 2 + key.len();
+                    fast_short_slice_copy(key, &mut data[2..stop]);
+                    store(&mut data[stop..], val);
+                }
+
+                self.set_bucket(hash, key_addr, bucket);
+                return val;
+            }
+            if kv.hash == hash {
+                if let Some(val_addr) =
+                    self.get_value_addr_if_key_match(key, kv.key_value_addr, memory_arena)
+                {
+                    let v = memory_arena.read(val_addr);
+                    let new_v = updater(Some(v));
+                    memory_arena.write_at(val_addr, new_v);
+                    return new_v;
+                }
+            }
+            // This allows fetching the next bucket before the loop jmp
+            bucket = probe.next_probe();
+            kv = self.table[bucket];
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use std::collections::HashMap;
+
+    use super::{compute_previous_power_of_two, SharedArenaHashMap};
+    use crate::MemoryArena;
+
+    #[test]
+    fn test_hash_map() {
+        let mut memory_arena = MemoryArena::default();
+        let mut hash_map: SharedArenaHashMap = SharedArenaHashMap::default();
+        hash_map.mutate_or_create(b"abc", &mut memory_arena, |opt_val: Option<u32>| {
+            assert_eq!(opt_val, None);
+            3u32
+        });
+        hash_map.mutate_or_create(b"abcd", &mut memory_arena, |opt_val: Option<u32>| {
+            assert_eq!(opt_val, None);
+            4u32
+        });
+        hash_map.mutate_or_create(b"abc", &mut memory_arena, |opt_val: Option<u32>| {
+            assert_eq!(opt_val, Some(3u32));
+            5u32
+        });
+        let mut vanilla_hash_map = HashMap::new();
+        let iter_values = hash_map.iter(&memory_arena);
+        for (key, addr) in iter_values {
+            let val: u32 = memory_arena.read(addr);
+            vanilla_hash_map.insert(key.to_owned(), val);
+        }
+        assert_eq!(vanilla_hash_map.len(), 2);
+    }
+    #[test]
+    fn test_empty_hashmap() {
+        let memory_arena = MemoryArena::default();
+        let hash_map: SharedArenaHashMap = SharedArenaHashMap::default();
+        assert_eq!(hash_map.get::<u32>(b"abc", &memory_arena), None);
+    }
+
+    #[test]
+    fn test_compute_previous_power_of_two() {
+        assert_eq!(compute_previous_power_of_two(8), 8);
+        assert_eq!(compute_previous_power_of_two(9), 8);
+        assert_eq!(compute_previous_power_of_two(7), 4);
+        assert_eq!(compute_previous_power_of_two(u64::MAX as usize), 1 << 63);
+    }
+
+    #[test]
+    fn test_many_terms() {
+        let mut memory_arena = MemoryArena::default();
+        let mut terms: Vec<String> = (0..20_000).map(|val| val.to_string()).collect();
+        let mut hash_map: SharedArenaHashMap = SharedArenaHashMap::default();
+        for term in terms.iter() {
+            hash_map.mutate_or_create(
+                term.as_bytes(),
+                &mut memory_arena,
+                |_opt_val: Option<u32>| 5u32,
+            );
+        }
+        let mut terms_back: Vec<String> = hash_map
+            .iter(&memory_arena)
+            .map(|(bytes, _)| String::from_utf8(bytes.to_vec()).unwrap())
+            .collect();
+        terms_back.sort();
+        terms.sort();
+
+        for pos in 0..terms.len() {
+            assert_eq!(terms[pos], terms_back[pos]);
+        }
+    }
+}