Added simple columnar CLI program

See README for more information.

.gitignore

@@ -13,3 +13,5 @@ benchmark
 .idea
 trace.dat
 cargo-timing*
+columnar/columnar-cli/*.json
+**/perf.data*

CHANGELOG.md

@@ -2,22 +2,21 @@ Tantivy 0.19
 ================================
 #### Bugfixes
 - Fix missing fieldnorms for u64, i64, f64, bool, bytes and date [#1620](https://github.com/quickwit-oss/tantivy/pull/1620) (@PSeitz)
-- Fix interpolation overflow in linear interpolation fastfield codec [#1480](https://github.com/quickwit-oss/tantivy/pull/1480 (@PSeitz @fulmicoton)
+- Fix interpolation overflow in linear interpolation fastfield codec [#1480](https://github.com/quickwit-oss/tantivy/pull/1480) (@PSeitz @fulmicoton)
 
 #### Features/Improvements
 - Add support for `IN` in queryparser , e.g. `field: IN [val1 val2 val3]` [#1683](https://github.com/quickwit-oss/tantivy/pull/1683) (@trinity-1686a)
 - Skip score calculation, when no scoring is required [#1646](https://github.com/quickwit-oss/tantivy/pull/1646) (@PSeitz)
 - Limit fast fields to u32 (`get_val(u32)`) [#1644](https://github.com/quickwit-oss/tantivy/pull/1644) (@PSeitz)
-- Updated [Date Field Type](https://github.com/quickwit-oss/tantivy/pull/1396)
+- The `DateTime` type has been updated to hold timestamps with microseconds precision.
-  The `DateTime` type has been updated to hold timestamps with microseconds precision.
+  `DateOptions` and `DatePrecision` have been added to configure Date fields. The precision is used to hint on fast values compression. Otherwise, seconds precision is used everywhere else (i.e terms, indexing) [#1396](https://github.com/quickwit-oss/tantivy/pull/1396) (@evanxg852000)
-  `DateOptions` and `DatePrecision` have been added to configure Date fields. The precision is used to hint on fast values compression. Otherwise, seconds precision is used everywhere else (i.e terms, indexing). (@evanxg852000)
 - Add IP address field type [#1553](https://github.com/quickwit-oss/tantivy/pull/1553) (@PSeitz)
 - Add boolean field type [#1382](https://github.com/quickwit-oss/tantivy/pull/1382) (@boraarslan)
 - Remove Searcher pool and make `Searcher` cloneable. (@PSeitz)
-- Validate settings on create [#1570](https://github.com/quickwit-oss/tantivy/pull/1570 (@PSeitz)
+- Validate settings on create [#1570](https://github.com/quickwit-oss/tantivy/pull/1570) (@PSeitz)
 - Detect and apply gcd on fastfield codecs [#1418](https://github.com/quickwit-oss/tantivy/pull/1418) (@PSeitz)
 - Doc store
-  - use separate thread to compress block store [#1389](https://github.com/quickwit-oss/tantivy/pull/1389) [#1510](https://github.com/quickwit-oss/tantivy/pull/1510 (@PSeitz @fulmicoton)
+  - use separate thread to compress block store [#1389](https://github.com/quickwit-oss/tantivy/pull/1389) [#1510](https://github.com/quickwit-oss/tantivy/pull/1510) (@PSeitz @fulmicoton)
   - Expose doc store cache size [#1403](https://github.com/quickwit-oss/tantivy/pull/1403) (@PSeitz)
   - Enable compression levels for doc store [#1378](https://github.com/quickwit-oss/tantivy/pull/1378) (@PSeitz)
   - Make block size configurable [#1374](https://github.com/quickwit-oss/tantivy/pull/1374) (@kryesh)

Cargo.toml

@@ -15,7 +15,7 @@ rust-version = "1.62"
 
 [dependencies]
 oneshot = "0.1.5"
-base64 = "0.13.0"
+base64 = "0.20.0"
 byteorder = "1.4.3"
 crc32fast = "1.3.2"
 once_cell = "1.10.0"
@@ -36,7 +36,6 @@ fs2 = { version = "0.4.3", optional = true }
 levenshtein_automata = "0.2.1"
 uuid = { version = "1.0.0", features = ["v4", "serde"] }
 crossbeam-channel = "0.5.4"
-stable_deref_trait = "1.2.0"
 rust-stemmers = "1.2.0"
 downcast-rs = "1.2.0"
 bitpacking = { version = "0.8.4", default-features = false, features = ["bitpacker4x"] }
@@ -53,15 +52,15 @@ lru = "0.7.5"
 fastdivide = "0.4.0"
 itertools = "0.10.3"
 measure_time = "0.8.2"
-ciborium = { version = "0.2", optional = true}
 async-trait = "0.1.53"
 arc-swap = "1.5.0"
 
+sstable = { version="0.1", path="./sstable", package ="tantivy-sstable", optional = true }
+stacker = { version="0.1", path="./stacker", package ="tantivy-stacker" }
 tantivy-query-grammar = { version= "0.19.0", path="./query-grammar" }
 tantivy-bitpacker = 		{ version= "0.3", path="./bitpacker" }
-common = 								{ version= "0.4", path = "./common/", package = "tantivy-common" }
+common = 								{ version= "0.5", path = "./common/", package = "tantivy-common" }
 fastfield_codecs = 			{ version= "0.3", path="./fastfield_codecs", default-features = false }
-ownedbytes = 						{ version= "0.4", path="./ownedbytes" }
 
 [target.'cfg(windows)'.dependencies]
 winapi = "0.3.9"
@@ -104,10 +103,10 @@ zstd-compression = ["zstd"]
 failpoints = ["fail/failpoints"]
 unstable = [] # useful for benches.
 
-quickwit = ["ciborium"]
+quickwit = ["sstable"]
 
 [workspace]
-members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes"]
+members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable", "columnar"]
 
 # Following the "fail" crate best practises, we isolate
 # tests that define specific behavior in fail check points

bitpacker/src/bitpacker.rs

@@ -25,15 +25,14 @@ impl BitPacker {
         num_bits: u8,
         output: &mut TWrite,
     ) -> io::Result<()> {
-        let val_u64 = val as u64;
         let num_bits = num_bits as usize;
         if self.mini_buffer_written + num_bits > 64 {
-            self.mini_buffer |= val_u64.wrapping_shl(self.mini_buffer_written as u32);
+            self.mini_buffer |= val.wrapping_shl(self.mini_buffer_written as u32);
             output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
-            self.mini_buffer = val_u64.wrapping_shr((64 - self.mini_buffer_written) as u32);
+            self.mini_buffer = val.wrapping_shr((64 - self.mini_buffer_written) as u32);
             self.mini_buffer_written = self.mini_buffer_written + num_bits - 64;
         } else {
-            self.mini_buffer |= val_u64 << self.mini_buffer_written;
+            self.mini_buffer |= val << self.mini_buffer_written;
             self.mini_buffer_written += num_bits;
             if self.mini_buffer_written == 64 {
                 output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
@@ -92,17 +91,15 @@ impl BitUnpacker {
             return 0u64;
         }
         let addr_in_bits = idx * self.num_bits as u32;
-        let addr = addr_in_bits >> 3;
+        let addr = (addr_in_bits >> 3) as usize;
         let bit_shift = addr_in_bits & 7;
         debug_assert!(
-            addr + 8 <= data.len() as u32,
+            addr + 8 <= data.len(),
             "The fast field field should have been padded with 7 bytes."
         );
-        let bytes: [u8; 8] = (&data[(addr as usize)..(addr as usize) + 8])
+        let bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
-            .try_into()
-            .unwrap();
         let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
-        let val_shifted = (val_unshifted_unmasked >> bit_shift) as u64;
+        let val_shifted = val_unshifted_unmasked >> bit_shift;
         val_shifted & self.mask
     }
 }

bitpacker/src/blocked_bitpacker.rs

@@ -84,7 +84,7 @@ impl BlockedBitpacker {
     #[inline]
     pub fn add(&mut self, val: u64) {
         self.buffer.push(val);
-        if self.buffer.len() == BLOCK_SIZE as usize {
+        if self.buffer.len() == BLOCK_SIZE {
             self.flush();
         }
     }
@@ -126,8 +126,8 @@ impl BlockedBitpacker {
     }
     #[inline]
     pub fn get(&self, idx: usize) -> u64 {
-        let metadata_pos = idx / BLOCK_SIZE as usize;
+        let metadata_pos = idx / BLOCK_SIZE;
-        let pos_in_block = idx % BLOCK_SIZE as usize;
+        let pos_in_block = idx % BLOCK_SIZE;
         if let Some(metadata) = self.offset_and_bits.get(metadata_pos) {
             let unpacked = BitUnpacker::new(metadata.num_bits()).get(
                 pos_in_block as u32,

bitpacker/src/lib.rs

@@ -1,6 +1,8 @@
 mod bitpacker;
 mod blocked_bitpacker;
 
+use std::cmp::Ordering;
+
 pub use crate::bitpacker::{BitPacker, BitUnpacker};
 pub use crate::blocked_bitpacker::BlockedBitpacker;
 
@@ -37,44 +39,104 @@ pub fn compute_num_bits(n: u64) -> u8 {
     }
 }
 
+/// Computes the (min, max) of an iterator of `PartialOrd` values.
+///
+/// For values implementing `Ord` (in a way consistent to their `PartialOrd` impl),
+/// this function behaves as expected.
+///
+/// For values with partial ordering, the behavior is non-trivial and may
+/// depends on the order of the values.
+/// For floats however, it simply returns the same results as if NaN were
+/// skipped.
 pub fn minmax<I, T>(mut vals: I) -> Option<(T, T)>
 where
     I: Iterator<Item = T>,
-    T: Copy + Ord,
+    T: Copy + PartialOrd,
 {
-    if let Some(first_el) = vals.next() {
+    let first_el = vals.find(|val| {
-        return Some(vals.fold((first_el, first_el), |(min_val, max_val), el| {
+        // We use this to make sure we skip all NaN values when
-            (min_val.min(el), max_val.max(el))
+        // working with a float type.
-        }));
+        val.partial_cmp(val) == Some(Ordering::Equal)
+    })?;
+    let mut min_so_far: T = first_el;
+    let mut max_so_far: T = first_el;
+    for val in vals {
+        if val.partial_cmp(&min_so_far) == Some(Ordering::Less) {
+            min_so_far = val;
+        }
+        if val.partial_cmp(&max_so_far) == Some(Ordering::Greater) {
+            max_so_far = val;
+        }
     }
-    None
+    Some((min_so_far, max_so_far))
 }
 
-#[test]
+#[cfg(test)]
-fn test_compute_num_bits() {
+mod tests {
-    assert_eq!(compute_num_bits(1), 1u8);
+    use super::*;
-    assert_eq!(compute_num_bits(0), 0u8);
-    assert_eq!(compute_num_bits(2), 2u8);
-    assert_eq!(compute_num_bits(3), 2u8);
-    assert_eq!(compute_num_bits(4), 3u8);
-    assert_eq!(compute_num_bits(255), 8u8);
-    assert_eq!(compute_num_bits(256), 9u8);
-    assert_eq!(compute_num_bits(5_000_000_000), 33u8);
-}
 
-#[test]
+    #[test]
-fn test_minmax_empty() {
+    fn test_compute_num_bits() {
-    let vals: Vec<u32> = vec![];
+        assert_eq!(compute_num_bits(1), 1u8);
-    assert_eq!(minmax(vals.into_iter()), None);
+        assert_eq!(compute_num_bits(0), 0u8);
-}
+        assert_eq!(compute_num_bits(2), 2u8);
+        assert_eq!(compute_num_bits(3), 2u8);
+        assert_eq!(compute_num_bits(4), 3u8);
+        assert_eq!(compute_num_bits(255), 8u8);
+        assert_eq!(compute_num_bits(256), 9u8);
+        assert_eq!(compute_num_bits(5_000_000_000), 33u8);
+    }
 
-#[test]
+    #[test]
-fn test_minmax_one() {
+    fn test_minmax_empty() {
-    assert_eq!(minmax(vec![1].into_iter()), Some((1, 1)));
+        let vals: Vec<u32> = vec![];
-}
+        assert_eq!(minmax(vals.into_iter()), None);
+    }
 
-#[test]
+    #[test]
-fn test_minmax_two() {
+    fn test_minmax_one() {
-    assert_eq!(minmax(vec![1, 2].into_iter()), Some((1, 2)));
+        assert_eq!(minmax(vec![1].into_iter()), Some((1, 1)));
-    assert_eq!(minmax(vec![2, 1].into_iter()), Some((1, 2)));
+    }
+
+    #[test]
+    fn test_minmax_two() {
+        assert_eq!(minmax(vec![1, 2].into_iter()), Some((1, 2)));
+        assert_eq!(minmax(vec![2, 1].into_iter()), Some((1, 2)));
+    }
+
+    #[test]
+    fn test_minmax_nan() {
+        assert_eq!(
+            minmax(vec![f64::NAN, 1f64, 2f64].into_iter()),
+            Some((1f64, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, f64::NAN, 1f64].into_iter()),
+            Some((1f64, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, 1f64, f64::NAN].into_iter()),
+            Some((1f64, 2f64))
+        );
+    }
+
+    #[test]
+    fn test_minmax_inf() {
+        assert_eq!(
+            minmax(vec![f64::INFINITY, 1f64, 2f64].into_iter()),
+            Some((1f64, f64::INFINITY))
+        );
+        assert_eq!(
+            minmax(vec![-f64::INFINITY, 1f64, 2f64].into_iter()),
+            Some((-f64::INFINITY, 2f64))
+        );
+        assert_eq!(
+            minmax(vec![2f64, f64::INFINITY, 1f64].into_iter()),
+            Some((1f64, f64::INFINITY))
+        );
+        assert_eq!(
+            minmax(vec![2f64, 1f64, -f64::INFINITY].into_iter()),
+            Some((-f64::INFINITY, 2f64))
+        );
+    }
 }

columnar/Cargo.toml

@@ -0,0 +1,19 @@
+[package]
+name = "tantivy-columnar"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+stacker = { path = "../stacker", package="tantivy-stacker"}
+serde_json = "1"
+thiserror = "1"
+fnv = "1"
+sstable = { path = "../sstable", package = "tantivy-sstable" }
+zstd = "0.12"
+common = { path = "../common", package = "tantivy-common" }
+fastfield_codecs = { path = "../fastfield_codecs"}
+itertools = "0.10"
+
+[dev-dependencies]
+proptest = "1"

columnar/README.md

@@ -0,0 +1,73 @@
+# Columnar format
+
+This crate describes columnar format used in tantivy.
+
+## Goals
+
+This format is special in the following way.
+- it needs to be compact
+- it does not required to be loaded in memory.
+- it is designed to fit well with quickwit's strange constraint:
+we need to be able to load columns rapidly.
+- columns of several types can be associated with the same column name.
+- it needs to support columns with different types `(str, u64, i64, f64)`
+and different cardinality `(required, optional, multivalued)`.
+- columns, once loaded, offer cheap random access.
+
+# Coercion rules
+
+Users can create a columnar by appending rows to a writer.
+Nothing prevents a user from recording values with different to a same `column_key`.
+
+In that case, `tantivy-columnar`'s behavior is as follows:
+- Values that corresponds to different JsonValue type are mapped to different columns. For instance, String values are treated independently from Number or boolean values. `tantivy-columnar` will simply emit several columns associated to a given column_name.
+- Only one column for a given json value type is emitted.  If number values with different number types are recorded (e.g. u64, i64, f64), `tantivy-columnar` will pick the first type that can represents the set of appended value, with the following prioriy order (`i64`, `u64`, `f64`). `i64` is picked over `u64` as it is likely to  yield less change of types. Most use cases strictly requiring `u64` show the restriction on 50% of the values (e.g. a 64-bit hash). On the other hand, a lot of use cases can show rare negative value.
+
+# Columnar format
+
+Because this columnar format tries to avoid some coercion.
+There can be several columns (with different type) associated to a single `column_name`.
+
+Each column is associated to `column_key`.
+The format of that key is:
+`[column_name][ZERO_BYTE][column_type_header: u8]`
+
+```
+COLUMNAR:=
+    [COLUMNAR_DATA]
+    [COLUMNAR_INDEX]
+    [COLUMNAR_FOOTER];
+
+
+# Columns are sorted by their column key.
+COLUMNAR_DATA:=
+    [COLUMN]+;
+
+COLUMN:=
+    COMPRESSED_COLUMN | NON_COMPRESSED_COLUMN;
+
+# COLUMN_DATA is compressed when it exceeds a threshold of 100KB.
+
+COMPRESSED_COLUMN := [b'1'][zstd(COLUMN_DATA)]
+NON_COMPRESSED_COLUMN:= [b'0'][COLUMN_DATA]
+
+COLUMNAR_INDEX := [RANGE_SSTABLE_BYTES]
+
+COLUMNAR_FOOTER := [RANGE_SSTABLE_BYTES_LEN: 8 bytes little endian]
+
+```
+
+The columnar file starts by the actual column data, concatenated one after the other,
+sorted by column key.
+
+A quickwit/tantivy style sstable associates
+`(column names, column_cardinality, column_type) to range of bytes.
+
+Column name may not contain the zero byte.
+
+Listing all columns associated to `column_name` can therefore
+be done by listing all keys prefixed by
+`[column_name][ZERO_BYTE]`
+
+The associated range of bytes refer to a range of bytes
+

columnar/columnar-cli/Cargo.toml

@@ -0,0 +1,17 @@
+[package]
+name = "tantivy-columnar-cli"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+columnar = {path="../", package="tantivy-columnar"}
+serde_json = "1"
+serde_json_borrow = {git="https://github.com/PSeitz/serde_json_borrow/"}
+serde = "1"
+
+[workspace]
+members = []
+
+[profile.release]
+debug = true

columnar/columnar-cli/src/main.rs

@@ -0,0 +1,126 @@
+use columnar::ColumnarWriter;
+use columnar::NumericalValue;
+use serde_json_borrow;
+use std::fs::File;
+use std::io;
+use std::io::BufRead;
+use std::io::BufReader;
+use std::time::Instant;
+
+#[derive(Default)]
+struct JsonStack {
+    path: String,
+    stack: Vec<usize>,
+}
+
+impl JsonStack {
+    fn push(&mut self, seg: &str) {
+        let len = self.path.len();
+        self.stack.push(len);
+        self.path.push('.');
+        self.path.push_str(seg);
+    }
+
+    fn pop(&mut self) {
+        if let Some(len) = self.stack.pop() {
+            self.path.truncate(len);
+        }
+    }
+
+    fn path(&self) -> &str {
+        &self.path[1..]
+    }
+}
+
+fn append_json_to_columnar(
+    doc: u32,
+    json_value: &serde_json_borrow::Value,
+    columnar: &mut ColumnarWriter,
+    stack: &mut JsonStack,
+) -> usize {
+    let mut count = 0;
+    match json_value {
+        serde_json_borrow::Value::Null => {}
+        serde_json_borrow::Value::Bool(val) => {
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                NumericalValue::from(if *val { 1u64 } else { 0u64 }),
+            );
+            count += 1;
+        }
+        serde_json_borrow::Value::Number(num) => {
+            let numerical_value: NumericalValue = if let Some(num_i64) = num.as_i64() {
+                num_i64.into()
+            } else if let Some(num_u64) = num.as_u64() {
+                num_u64.into()
+            } else if let Some(num_f64) = num.as_f64() {
+                num_f64.into()
+            } else {
+                panic!();
+            };
+            count += 1;
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                numerical_value,
+            );
+        }
+        serde_json_borrow::Value::Str(msg) => {
+            columnar.record_str(
+                doc,
+                stack.path(),
+                msg.as_bytes(),
+            );
+            count += 1;
+        },
+        serde_json_borrow::Value::Array(vals) => {
+            for val in vals {
+                count += append_json_to_columnar(doc, val, columnar, stack);
+            }
+        },
+        serde_json_borrow::Value::Object(json_map) => {
+            for (child_key, child_val) in json_map {
+                stack.push(child_key);
+                count += append_json_to_columnar(doc, child_val, columnar, stack);
+                stack.pop();
+            }
+        },
+    }
+    count
+}
+
+fn main() -> io::Result<()> {
+    let file = File::open("gh_small.json")?;
+    let mut reader = BufReader::new(file);
+    let mut line = String::with_capacity(100);
+    let mut columnar = columnar::ColumnarWriter::default();
+    let mut doc = 0;
+    let start = Instant::now();
+    let mut stack = JsonStack::default();
+    let mut total_count = 0;
+    loop {
+        line.clear();
+        let len = reader.read_line(&mut line)?;
+        if len == 0 {
+            break;
+        }
+        let Ok(json_value) = serde_json::from_str::<serde_json_borrow::Value>(&line) else { continue; };
+        total_count += append_json_to_columnar(doc, &json_value, &mut columnar, &mut stack);
+        doc += 1;
+    }
+    println!("value count {total_count}");
+    println!("record {:?}", start.elapsed());
+    let mut buffer = Vec::new();
+    columnar.serialize(doc, &mut buffer)?;
+    println!("num docs: {doc}, {:?}", start.elapsed());
+    println!("buffer len {} MB", buffer.len() / 1_000_000);
+    let columnar = columnar::ColumnarReader::open(buffer)?;
+    for (column_name, typ, offsets, num_bytes) in columnar.list_columns()? {
+        if num_bytes>1_000_000 {
+            println!("{column_name} {typ:?} {offsets:?} {}", num_bytes / 1_000_000);
+        }
+    }
+    println!("{} columns", columnar.num_columns());
+    Ok(())
+}

columnar/src/column_type_header.rs

@@ -0,0 +1,188 @@
+use crate::utils::{place_bits, select_bits};
+use crate::value::NumericalType;
+
+/// Enum describing the number of values that can exist per document
+/// (or per row if you will).
+#[derive(Clone, Copy, Hash, Default, Debug, PartialEq, Eq, PartialOrd, Ord)]
+#[repr(u8)]
+pub enum Cardinality {
+    /// All documents contain exactly one value.
+    #[default]
+    Required = 0,
+    /// All documents contain at most one value.
+    Optional = 1,
+    /// All documents may contain any number of values.
+    Multivalued = 2,
+}
+
+impl Cardinality {
+    pub(crate) fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn try_from_code(code: u8) -> Option<Cardinality> {
+        match code {
+            0 => Some(Cardinality::Required),
+            1 => Some(Cardinality::Optional),
+            2 => Some(Cardinality::Multivalued),
+            _ => None,
+        }
+    }
+}
+
+#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy)]
+pub enum ColumnType {
+    Bytes,
+    Numerical(NumericalType),
+    Bool,
+}
+
+impl ColumnType {
+    /// Encoded over 6 bits.
+    pub(crate) fn to_code(self) -> u8 {
+        let high_type;
+        let low_code: u8;
+        match self {
+            ColumnType::Bytes => {
+                high_type = GeneralType::Str;
+                low_code = 0u8;
+            }
+            ColumnType::Numerical(numerical_type) => {
+                high_type = GeneralType::Numerical;
+                low_code = numerical_type.to_code();
+            }
+            ColumnType::Bool => {
+                high_type = GeneralType::Bool;
+                low_code = 0u8;
+            }
+        }
+        place_bits::<3, 6>(high_type.to_code()) | place_bits::<0, 3>(low_code)
+    }
+
+    pub(crate) fn try_from_code(code: u8) -> Option<ColumnType> {
+        if select_bits::<6, 8>(code) != 0u8 {
+            return None;
+        }
+        let high_code = select_bits::<3, 6>(code);
+        let low_code = select_bits::<0, 3>(code);
+        let high_type = GeneralType::try_from_code(high_code)?;
+        match high_type {
+            GeneralType::Bool => {
+                if low_code != 0u8 {
+                    return None;
+                }
+                Some(ColumnType::Bool)
+            }
+            GeneralType::Str => {
+                if low_code != 0u8 {
+                    return None;
+                }
+                Some(ColumnType::Bytes)
+            }
+            GeneralType::Numerical => {
+                let numerical_type = NumericalType::try_from_code(low_code)?;
+                Some(ColumnType::Numerical(numerical_type))
+            }
+        }
+    }
+}
+
+/// This corresponds to the JsonType.
+#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq, Debug)]
+#[repr(u8)]
+pub(crate) enum GeneralType {
+    Bool = 0u8,
+    Str = 1u8,
+    Numerical = 2u8,
+}
+
+impl GeneralType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub fn try_from_code(code: u8) -> Option<Self> {
+        match code {
+            0u8 => Some(Self::Bool),
+            1u8 => Some(Self::Str),
+            2u8 => Some(Self::Numerical),
+            _ => None,
+        }
+    }
+}
+
+/// Represents the type and cardinality of a column.
+/// This is encoded over one-byte and added to a column key in the
+/// columnar sstable.
+///
+/// Cardinality is encoded as the first two highest two bits.
+/// The low 6 bits encode the column type.
+#[derive(Eq, Hash, PartialEq, Debug, Copy, Clone)]
+pub struct ColumnTypeAndCardinality {
+    pub cardinality: Cardinality,
+    pub typ: ColumnType,
+}
+
+impl ColumnTypeAndCardinality {
+    pub fn to_code(self) -> u8 {
+        place_bits::<6, 8>(self.cardinality.to_code()) | place_bits::<0, 6>(self.typ.to_code())
+    }
+
+    pub fn try_from_code(code: u8) -> Option<ColumnTypeAndCardinality> {
+        let typ_code = select_bits::<0, 6>(code);
+        let cardinality_code = select_bits::<6, 8>(code);
+        let cardinality = Cardinality::try_from_code(cardinality_code)?;
+        let typ = ColumnType::try_from_code(typ_code)?;
+        assert_eq!(typ.to_code(), typ_code);
+        Some(ColumnTypeAndCardinality { cardinality, typ })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashSet;
+
+    use super::ColumnTypeAndCardinality;
+    use crate::column_type_header::{Cardinality, ColumnType};
+
+    #[test]
+    fn test_column_type_header_to_code() {
+        let mut column_type_header_set: HashSet<ColumnTypeAndCardinality> = HashSet::new();
+        for code in u8::MIN..=u8::MAX {
+            if let Some(column_type_header) = ColumnTypeAndCardinality::try_from_code(code) {
+                assert_eq!(column_type_header.to_code(), code);
+                assert!(column_type_header_set.insert(column_type_header));
+            }
+        }
+        assert_eq!(
+            column_type_header_set.len(),
+            3 /* cardinality */ *
+            (1 + 1 + 3) // column_types (str, bool, numerical x 3)
+        );
+    }
+
+    #[test]
+    fn test_column_type_to_code() {
+        let mut column_type_set: HashSet<ColumnType> = HashSet::new();
+        for code in u8::MIN..=u8::MAX {
+            if let Some(column_type) = ColumnType::try_from_code(code) {
+                assert_eq!(column_type.to_code(), code);
+                assert!(column_type_set.insert(column_type));
+            }
+        }
+        assert_eq!(column_type_set.len(), 2 + 3);
+    }
+
+    #[test]
+    fn test_cardinality_to_code() {
+        let mut num_cardinality = 0;
+        for code in u8::MIN..=u8::MAX {
+            let cardinality_opt = Cardinality::try_from_code(code);
+            if let Some(cardinality) = cardinality_opt {
+                assert_eq!(cardinality.to_code(), code);
+                num_cardinality += 1;
+            }
+        }
+        assert_eq!(num_cardinality, 3);
+    }
+}

columnar/src/dictionary.rs

@@ -0,0 +1,84 @@
+use std::io;
+
+use fnv::FnvHashMap;
+use sstable::SSTable;
+
+pub(crate) struct IdMapping {
+    unordered_to_ord: Vec<OrderedId>,
+}
+
+impl IdMapping {
+    pub fn to_ord(&self, unordered: UnorderedId) -> OrderedId {
+        self.unordered_to_ord[unordered.0 as usize]
+    }
+}
+
+/// When we add values, we cannot know their ordered id yet.
+/// For this reason, we temporarily assign them a `UnorderedId`
+/// that will be mapped to an `OrderedId` upon serialization.
+#[derive(Clone, Copy, Debug, Hash, PartialEq, Eq)]
+pub struct UnorderedId(pub u32);
+
+#[derive(Clone, Copy, Hash, PartialEq, Eq, Debug)]
+pub struct OrderedId(pub u32);
+
+/// `DictionaryBuilder` for dictionary encoding.
+///
+/// It stores the different terms encounterred and assigns them a temporary value
+/// we call unordered id.
+///
+/// Upon serialization, we will sort the ids and hence build a `UnorderedId -> Term ordinal`
+/// mapping.
+#[derive(Default)]
+pub(crate) struct DictionaryBuilder {
+    dict: FnvHashMap<Vec<u8>, UnorderedId>,
+}
+
+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);
+        new_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<IdMapping> {
+        let mut terms: Vec<(&[u8], UnorderedId)> =
+            self.dict.iter().map(|(k, v)| (k.as_slice(), *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()];
+        let mut sstable_builder = sstable::VoidSSTable::writer(wrt);
+        for (ord, (key, unordered_id)) in terms.into_iter().enumerate() {
+            let ordered_id = OrderedId(ord as u32);
+            sstable_builder.insert(key, &())?;
+            unordered_to_ord[unordered_id.0 as usize] = ordered_id;
+        }
+        sstable_builder.finish()?;
+        Ok(IdMapping { unordered_to_ord })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_dictionary_builder() {
+        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 mut buffer = Vec::new();
+        let id_mapping = dictionary_builder.serialize(&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

@@ -0,0 +1,86 @@
+mod column_type_header;
+mod dictionary;
+mod reader;
+pub(crate) mod utils;
+mod value;
+mod writer;
+
+pub use column_type_header::Cardinality;
+pub use reader::ColumnarReader;
+pub use value::{NumericalType, NumericalValue};
+pub use writer::ColumnarWriter;
+
+pub type DocId = u32;
+
+#[cfg(test)]
+mod tests {
+    use std::ops::Range;
+
+    use common::file_slice::FileSlice;
+
+    use crate::column_type_header::{ColumnType, ColumnTypeAndCardinality};
+    use crate::reader::ColumnarReader;
+    use crate::value::NumericalValue;
+    use crate::{Cardinality, ColumnarWriter};
+
+    #[test]
+    fn test_dataframe_writer_bytes() {
+        let mut dataframe_writer = ColumnarWriter::default();
+        dataframe_writer.record_str(1u32, "my_string", b"hello");
+        dataframe_writer.record_str(3u32, "my_string", b"helloeee");
+        let mut buffer: Vec<u8> = Vec::new();
+        dataframe_writer.serialize(5, &mut buffer).unwrap();
+        let columnar_fileslice = FileSlice::from(buffer);
+        let columnar = ColumnarReader::open(columnar_fileslice).unwrap();
+        assert_eq!(columnar.num_columns(), 1);
+        let cols: Vec<(ColumnTypeAndCardinality, Range<u64>)> =
+            columnar.read_columns("my_string").unwrap();
+        assert_eq!(cols.len(), 1);
+        assert_eq!(cols[0].1, 0..159);
+    }
+
+    #[test]
+    fn test_dataframe_writer_bool() {
+        let mut dataframe_writer = ColumnarWriter::default();
+        dataframe_writer.record_bool(1u32, "bool.value", false);
+        let mut buffer: Vec<u8> = Vec::new();
+        dataframe_writer.serialize(5, &mut buffer).unwrap();
+        let columnar_fileslice = FileSlice::from(buffer);
+        let columnar = ColumnarReader::open(columnar_fileslice).unwrap();
+        assert_eq!(columnar.num_columns(), 1);
+        let cols: Vec<(ColumnTypeAndCardinality, Range<u64>)> =
+            columnar.read_columns("bool.value").unwrap();
+        assert_eq!(cols.len(), 1);
+        assert_eq!(
+            cols[0].0,
+            ColumnTypeAndCardinality {
+                cardinality: Cardinality::Optional,
+                typ: ColumnType::Bool
+            }
+        );
+        assert_eq!(cols[0].1, 0..22);
+    }
+
+    #[test]
+    fn test_dataframe_writer_numerical() {
+        let mut dataframe_writer = ColumnarWriter::default();
+        dataframe_writer.record_numerical(1u32, "srical.value", NumericalValue::U64(12u64));
+        dataframe_writer.record_numerical(2u32, "srical.value", NumericalValue::U64(13u64));
+        dataframe_writer.record_numerical(4u32, "srical.value", NumericalValue::U64(15u64));
+        let mut buffer: Vec<u8> = Vec::new();
+        dataframe_writer.serialize(5, &mut buffer).unwrap();
+        let columnar_fileslice = FileSlice::from(buffer);
+        let columnar = ColumnarReader::open(columnar_fileslice).unwrap();
+        assert_eq!(columnar.num_columns(), 1);
+        let cols: Vec<(ColumnTypeAndCardinality, Range<u64>)> =
+            columnar.read_columns("srical.value").unwrap();
+        assert_eq!(cols.len(), 1);
+        // Right now this 31 bytes are spent as follows
+        //
+        // - header 14 bytes
+        // - vals  8 //< due to padding? could have been 1byte?.
+        // - null footer 6 bytes
+        // - version footer 3 bytes // Should be file-wide
+        assert_eq!(cols[0].1, 0..32);
+    }
+}

columnar/src/reader/mod.rs

@@ -0,0 +1,102 @@
+use std::ops::Range;
+use std::{io, mem};
+
+use common::file_slice::FileSlice;
+use common::BinarySerializable;
+use sstable::{Dictionary, RangeSSTable};
+
+use crate::column_type_header::ColumnTypeAndCardinality;
+
+fn io_invalid_data(msg: String) -> io::Error {
+    io::Error::new(io::ErrorKind::InvalidData, msg) // format!("Invalid key found.
+                                                    // {key_bytes:?}")));
+}
+
+/// The ColumnarReader makes it possible to access a set of columns
+/// associated to field names.
+pub struct ColumnarReader {
+    column_dictionary: Dictionary<RangeSSTable>,
+    column_data: FileSlice,
+}
+
+impl ColumnarReader {
+    /// Opens a new Columnar file.
+    pub fn open<F>(file_slice: F) -> io::Result<ColumnarReader>
+    where FileSlice: From<F> {
+        Self::open_inner(file_slice.into())
+    }
+
+    fn open_inner(file_slice: FileSlice) -> io::Result<ColumnarReader> {
+        let (file_slice_without_sstable_len, sstable_len_bytes) =
+            file_slice.split_from_end(mem::size_of::<u64>());
+        let mut sstable_len_bytes = sstable_len_bytes.read_bytes()?;
+        let sstable_len = u64::deserialize(&mut sstable_len_bytes)?;
+        let (column_data, sstable) =
+            file_slice_without_sstable_len.split_from_end(sstable_len as usize);
+        let column_dictionary = Dictionary::open(sstable)?;
+        Ok(ColumnarReader {
+            column_dictionary,
+            column_data,
+        })
+    }
+
+    // TODO fix ugly API
+    pub fn list_columns(
+        &self,
+    ) -> io::Result<Vec<(String, ColumnTypeAndCardinality, Range<u64>, u64)>> {
+        let mut stream = self.column_dictionary.stream()?;
+        let mut results = Vec::new();
+        while stream.advance() {
+            let key_bytes: &[u8] = stream.key();
+            let column_code: u8 = key_bytes.last().cloned().unwrap();
+            let column_type_and_cardinality = ColumnTypeAndCardinality::try_from_code(column_code)
+                .ok_or_else(|| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
+            let range = stream.value().clone();
+            let column_name = String::from_utf8_lossy(&key_bytes[..key_bytes.len() - 1]);
+            let range_len = range.end - range.start;
+            results.push((
+                column_name.to_string(),
+                column_type_and_cardinality,
+                range,
+                range_len,
+            ));
+        }
+        Ok(results)
+    }
+
+    /// Get all columns for the given field_name.
+    // TODO fix ugly API
+    pub fn read_columns(
+        &self,
+        field_name: &str,
+    ) -> io::Result<Vec<(ColumnTypeAndCardinality, Range<u64>)>> {
+        let mut start_key = field_name.to_string();
+        start_key.push('\0');
+        let mut end_key = field_name.to_string();
+        end_key.push(1u8 as char);
+        let mut stream = self
+            .column_dictionary
+            .range()
+            .ge(start_key.as_bytes())
+            .lt(end_key.as_bytes())
+            .into_stream()?;
+        let mut results = Vec::new();
+        while stream.advance() {
+            let key_bytes: &[u8] = stream.key();
+            if !key_bytes.starts_with(start_key.as_bytes()) {
+                return Err(io_invalid_data(format!("Invalid key found. {key_bytes:?}")));
+            }
+            let column_code: u8 = key_bytes.last().cloned().unwrap();
+            let column_type_and_cardinality = ColumnTypeAndCardinality::try_from_code(column_code)
+                .ok_or_else(|| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
+            let range = stream.value().clone();
+            results.push((column_type_and_cardinality, range));
+        }
+        Ok(results)
+    }
+
+    /// Return the number of columns in the columnar.
+    pub fn num_columns(&self) -> usize {
+        self.column_dictionary.num_terms()
+    }
+}

columnar/src/utils.rs

@@ -0,0 +1,76 @@
+const fn compute_mask(num_bits: u8) -> u8 {
+    if num_bits == 8 {
+        u8::MAX
+    } else {
+        (1u8 << num_bits) - 1
+    }
+}
+
+#[inline(always)]
+#[must_use]
+pub(crate) fn select_bits<const START: u8, const END: u8>(code: u8) -> u8 {
+    assert!(START <= END);
+    assert!(END <= 8);
+    let num_bits: u8 = END - START;
+    let mask: u8 = compute_mask(num_bits);
+    (code >> START) & mask
+}
+
+#[inline(always)]
+#[must_use]
+pub(crate) fn place_bits<const START: u8, const END: u8>(code: u8) -> u8 {
+    assert!(START <= END);
+    assert!(END <= 8);
+    let num_bits: u8 = END - START;
+    let mask: u8 = compute_mask(num_bits);
+    assert!(code <= mask);
+    code << START
+}
+
+/// Pop-front one bytes from a slice of bytes.
+#[inline(always)]
+pub fn pop_first_byte(bytes: &mut &[u8]) -> Option<u8> {
+    if bytes.is_empty() {
+        return None;
+    }
+    let first_byte = bytes[0];
+    *bytes = &bytes[1..];
+    Some(first_byte)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_select_bits() {
+        assert_eq!(255u8, select_bits::<0, 8>(255u8));
+        assert_eq!(0u8, select_bits::<0, 0>(255u8));
+        assert_eq!(8u8, select_bits::<0, 4>(8u8));
+        assert_eq!(4u8, select_bits::<1, 4>(8u8));
+        assert_eq!(0u8, select_bits::<1, 3>(8u8));
+    }
+
+    #[test]
+    fn test_place_bits() {
+        assert_eq!(255u8, place_bits::<0, 8>(255u8));
+        assert_eq!(4u8, place_bits::<2, 3>(1u8));
+        assert_eq!(0u8, place_bits::<2, 2>(0u8));
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_place_bits_overflows() {
+        let _ = place_bits::<1, 4>(8u8);
+    }
+
+    #[test]
+    fn test_pop_first_byte() {
+        let mut cursor: &[u8] = &b"abcd"[..];
+        assert_eq!(pop_first_byte(&mut cursor), Some(b'a'));
+        assert_eq!(pop_first_byte(&mut cursor), Some(b'b'));
+        assert_eq!(pop_first_byte(&mut cursor), Some(b'c'));
+        assert_eq!(pop_first_byte(&mut cursor), Some(b'd'));
+        assert_eq!(pop_first_byte(&mut cursor), None);
+    }
+}

columnar/src/value.rs

@@ -0,0 +1,121 @@
+#[derive(Copy, Clone, Debug, PartialEq)]
+pub enum NumericalValue {
+    I64(i64),
+    U64(u64),
+    F64(f64),
+}
+
+impl From<u64> for NumericalValue {
+    fn from(val: u64) -> NumericalValue {
+        NumericalValue::U64(val)
+    }
+}
+
+impl From<i64> for NumericalValue {
+    fn from(val: i64) -> Self {
+        NumericalValue::I64(val)
+    }
+}
+
+impl From<f64> for NumericalValue {
+    fn from(val: f64) -> Self {
+        NumericalValue::F64(val)
+    }
+}
+
+impl NumericalValue {
+    pub fn numerical_type(&self) -> NumericalType {
+        match self {
+            NumericalValue::F64(_) => NumericalType::F64,
+            NumericalValue::I64(_) => NumericalType::I64,
+            NumericalValue::U64(_) => NumericalType::U64,
+        }
+    }
+}
+
+impl Eq for NumericalValue {}
+
+#[derive(Clone, Copy, Debug, Default, Hash, Eq, PartialEq)]
+#[repr(u8)]
+pub enum NumericalType {
+    #[default]
+    I64 = 0,
+    U64 = 1,
+    F64 = 2,
+}
+
+impl NumericalType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub fn try_from_code(code: u8) -> Option<NumericalType> {
+        match code {
+            0 => Some(NumericalType::I64),
+            1 => Some(NumericalType::U64),
+            2 => Some(NumericalType::F64),
+            _ => None,
+        }
+    }
+}
+
+/// We voluntarily avoid using `Into` here to keep this
+/// implementation quirk as private as possible.
+///
+/// This coercion trait actually panics if it is used
+/// to convert a loose types to a stricter type.
+///
+/// The level is strictness is somewhat arbitrary.
+/// - i64
+/// - u64
+/// - f64.
+pub(crate) trait Coerce {
+    fn coerce(numerical_value: NumericalValue) -> Self;
+}
+
+impl Coerce for i64 {
+    fn coerce(value: NumericalValue) -> Self {
+        match value {
+            NumericalValue::I64(val) => val,
+            NumericalValue::U64(val) => val as i64,
+            NumericalValue::F64(_) => unreachable!(),
+        }
+    }
+}
+
+impl Coerce for u64 {
+    fn coerce(value: NumericalValue) -> Self {
+        match value {
+            NumericalValue::I64(val) => val as u64,
+            NumericalValue::U64(val) => val,
+            NumericalValue::F64(_) => unreachable!(),
+        }
+    }
+}
+
+impl Coerce for f64 {
+    fn coerce(value: NumericalValue) -> Self {
+        match value {
+            NumericalValue::I64(val) => val as f64,
+            NumericalValue::U64(val) => val as f64,
+            NumericalValue::F64(val) => val,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::NumericalType;
+
+    #[test]
+    fn test_numerical_type_code() {
+        let mut num_numerical_type = 0;
+        for code in u8::MIN..=u8::MAX {
+            if let Some(numerical_type) = NumericalType::try_from_code(code) {
+                assert_eq!(numerical_type.to_code(), code);
+                num_numerical_type += 1;
+            }
+        }
+        assert_eq!(num_numerical_type, 3);
+    }
+}

columnar/src/writer/column_operation.rs

@@ -0,0 +1,311 @@
+use crate::dictionary::UnorderedId;
+use crate::utils::{place_bits, pop_first_byte, select_bits};
+use crate::value::NumericalValue;
+use crate::{DocId, NumericalType};
+
+/// When we build a columnar dataframe, we first just group
+/// all mutations per column, and append them in append-only object.
+///
+/// We represents all of these operations as `ColumnOperation`.
+#[derive(Eq, PartialEq, Debug, Clone, Copy)]
+pub(crate) enum ColumnOperation<T> {
+    NewDoc(DocId),
+    Value(T),
+}
+
+#[derive(Copy, Clone, Debug, Eq, PartialEq)]
+struct ColumnOperationHeader {
+    typ_code: u8,
+    len: u8,
+}
+
+impl ColumnOperationHeader {
+    fn to_code(self) -> u8 {
+        place_bits::<0, 4>(self.len) | place_bits::<4, 8>(self.typ_code)
+    }
+
+    fn from_code(code: u8) -> Self {
+        let len = select_bits::<0, 4>(code);
+        let typ_code = select_bits::<4, 8>(code);
+        ColumnOperationHeader { typ_code, len }
+    }
+}
+
+const NEW_DOC_CODE: u8 = 0u8;
+const NEW_VALUE_CODE: u8 = 1u8;
+
+impl<V: SymbolValue> ColumnOperation<V> {
+    pub fn serialize(self) -> impl AsRef<[u8]> {
+        let mut minibuf = MiniBuffer::default();
+        let header = match self {
+            ColumnOperation::NewDoc(new_doc) => {
+                let symbol_len = new_doc.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationHeader {
+                    typ_code: NEW_DOC_CODE,
+                    len: symbol_len,
+                }
+            }
+            ColumnOperation::Value(val) => {
+                let symbol_len = val.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationHeader {
+                    typ_code: NEW_VALUE_CODE,
+                    len: symbol_len,
+                }
+            }
+        };
+        minibuf.bytes[0] = header.to_code();
+        minibuf.len = 1 + header.len;
+        minibuf
+    }
+
+    /// Deserialize a colummn operation.
+    /// Returns None if the buffer is empty.
+    ///
+    /// Panics if the payload is invalid.
+    pub fn deserialize(bytes: &mut &[u8]) -> Option<Self> {
+        let header_byte = pop_first_byte(bytes)?;
+        let column_op_header = ColumnOperationHeader::from_code(header_byte);
+        let symbol_bytes: &[u8];
+        (symbol_bytes, *bytes) = bytes.split_at(column_op_header.len as usize);
+        match column_op_header.typ_code {
+            NEW_DOC_CODE => {
+                let new_doc = u32::deserialize(symbol_bytes);
+                Some(ColumnOperation::NewDoc(new_doc))
+            }
+            NEW_VALUE_CODE => {
+                let value = V::deserialize(symbol_bytes);
+                Some(ColumnOperation::Value(value))
+            }
+            _ => {
+                panic!("Unknown code {}", column_op_header.typ_code);
+            }
+        }
+    }
+}
+
+impl<T> From<T> for ColumnOperation<T> {
+    fn from(value: T) -> Self {
+        ColumnOperation::Value(value)
+    }
+}
+
+#[allow(clippy::from_over_into)]
+pub(crate) trait SymbolValue: Clone + Copy {
+    fn serialize(self, buffer: &mut [u8]) -> u8;
+
+    // Reads the header type and the given bytes.
+    //
+    // `bytes` does not contain the header byte.
+    // This method should advance bytes by the number of bytes that were consumed.
+    fn deserialize(bytes: &[u8]) -> Self;
+}
+
+impl SymbolValue for bool {
+    fn serialize(self, buffer: &mut [u8]) -> u8 {
+        buffer[0] = if self { 1u8 } else { 0u8 };
+        1u8
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        bytes[0] == 1u8
+    }
+}
+
+#[derive(Default)]
+struct MiniBuffer {
+    pub bytes: [u8; 10],
+    pub len: u8,
+}
+
+impl AsRef<[u8]> for MiniBuffer {
+    fn as_ref(&self) -> &[u8] {
+        &self.bytes[..self.len as usize]
+    }
+}
+
+impl SymbolValue for NumericalValue {
+    fn deserialize(mut bytes: &[u8]) -> Self {
+        let type_code = pop_first_byte(&mut bytes).unwrap();
+        let symbol_type = NumericalType::try_from_code(type_code).unwrap();
+        let mut octet: [u8; 8] = [0u8; 8];
+        octet[..bytes.len()].copy_from_slice(bytes);
+        match symbol_type {
+            NumericalType::U64 => {
+                let val: u64 = u64::from_le_bytes(octet);
+                NumericalValue::U64(val)
+            }
+            NumericalType::I64 => {
+                let encoded: u64 = u64::from_le_bytes(octet);
+                let val: i64 = decode_zig_zag(encoded);
+                NumericalValue::I64(val)
+            }
+            NumericalType::F64 => {
+                debug_assert_eq!(bytes.len(), 8);
+                let val: f64 = f64::from_le_bytes(octet);
+                NumericalValue::F64(val)
+            }
+        }
+    }
+
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        match self {
+            NumericalValue::F64(val) => {
+                output[0] = NumericalType::F64 as u8;
+                output[1..9].copy_from_slice(&val.to_le_bytes());
+                9u8
+            }
+            NumericalValue::U64(val) => {
+                let len = compute_num_bytes_for_u64(val) as u8;
+                output[0] = NumericalType::U64 as u8;
+                output[1..9].copy_from_slice(&val.to_le_bytes());
+                len + 1u8
+            }
+            NumericalValue::I64(val) => {
+                let zig_zag_encoded = encode_zig_zag(val);
+                let len = compute_num_bytes_for_u64(zig_zag_encoded) as u8;
+                output[0] = NumericalType::I64 as u8;
+                output[1..9].copy_from_slice(&zig_zag_encoded.to_le_bytes());
+                len + 1u8
+            }
+        }
+    }
+}
+
+impl SymbolValue for u32 {
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        let len = compute_num_bytes_for_u64(self as u64);
+        output[0..4].copy_from_slice(&self.to_le_bytes());
+        len as u8
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        let mut quartet: [u8; 4] = [0u8; 4];
+        quartet[..bytes.len()].copy_from_slice(bytes);
+        u32::from_le_bytes(quartet)
+    }
+}
+
+impl SymbolValue for UnorderedId {
+    fn serialize(self, output: &mut [u8]) -> u8 {
+        self.0.serialize(output)
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        UnorderedId(u32::deserialize(bytes))
+    }
+}
+
+fn compute_num_bytes_for_u64(val: u64) -> usize {
+    let msb = (64u32 - val.leading_zeros()) as usize;
+    (msb + 7) / 8
+}
+
+fn encode_zig_zag(n: i64) -> u64 {
+    ((n << 1) ^ (n >> 63)) as u64
+}
+
+fn decode_zig_zag(n: u64) -> i64 {
+    ((n >> 1) as i64) ^ (-((n & 1) as i64))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[track_caller]
+    fn test_zig_zag_aux(val: i64) {
+        let encoded = super::encode_zig_zag(val);
+        assert_eq!(decode_zig_zag(encoded), val);
+        if let Some(abs_val) = val.checked_abs() {
+            let abs_val = abs_val as u64;
+            assert!(encoded <= abs_val * 2);
+        }
+    }
+
+    #[test]
+    fn test_zig_zag() {
+        assert_eq!(encode_zig_zag(0i64), 0u64);
+        assert_eq!(encode_zig_zag(-1i64), 1u64);
+        assert_eq!(encode_zig_zag(1i64), 2u64);
+        test_zig_zag_aux(0i64);
+        test_zig_zag_aux(i64::MIN);
+        test_zig_zag_aux(i64::MAX);
+    }
+
+    use proptest::prelude::any;
+    use proptest::proptest;
+
+    proptest! {
+        #[test]
+        fn test_proptest_zig_zag(val in any::<i64>()) {
+            test_zig_zag_aux(val);
+        }
+    }
+
+    #[test]
+    fn test_header_byte_serialization() {
+        for len in 0..=15 {
+            for typ_code in 0..=15 {
+                let header = ColumnOperationHeader { typ_code, len };
+                let header_code = header.to_code();
+                let serdeser_header = ColumnOperationHeader::from_code(header_code);
+                assert_eq!(header, serdeser_header);
+            }
+        }
+    }
+
+    #[track_caller]
+    fn ser_deser_symbol(column_op: ColumnOperation<NumericalValue>) {
+        let buf = column_op.serialize();
+        let mut buffer = buf.as_ref().to_vec();
+        buffer.extend_from_slice(b"234234");
+        let mut bytes = &buffer[..];
+        let serdeser_symbol = ColumnOperation::deserialize(&mut bytes).unwrap();
+        assert_eq!(bytes.len() + buf.as_ref().len() as usize, buffer.len());
+        assert_eq!(column_op, serdeser_symbol);
+    }
+
+    #[test]
+    fn test_compute_num_bytes_for_u64() {
+        assert_eq!(compute_num_bytes_for_u64(0), 0);
+        assert_eq!(compute_num_bytes_for_u64(1), 1);
+        assert_eq!(compute_num_bytes_for_u64(255), 1);
+        assert_eq!(compute_num_bytes_for_u64(256), 2);
+        assert_eq!(compute_num_bytes_for_u64((1 << 16) - 1), 2);
+        assert_eq!(compute_num_bytes_for_u64(1 << 16), 3);
+    }
+
+    #[test]
+    fn test_symbol_serialization() {
+        ser_deser_symbol(ColumnOperation::NewDoc(0));
+        ser_deser_symbol(ColumnOperation::NewDoc(3));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(0i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(1i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(257u64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(-257i64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(i64::MIN)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(0u64)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MIN)));
+        ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MAX)));
+    }
+
+    fn test_column_operation_unordered_aux(val: u32, expected_len: usize) {
+        let column_op = ColumnOperation::Value(UnorderedId(val));
+        let minibuf = column_op.serialize();
+        assert_eq!(minibuf.as_ref().len() as usize, expected_len);
+        let mut buf = minibuf.as_ref().to_vec();
+        buf.extend_from_slice(&[2, 2, 2, 2, 2, 2]);
+        let mut cursor = &buf[..];
+        let column_op_serdeser: ColumnOperation<UnorderedId> =
+            ColumnOperation::deserialize(&mut cursor).unwrap();
+        assert_eq!(column_op_serdeser, ColumnOperation::Value(UnorderedId(val)));
+        assert_eq!(cursor.len() + expected_len, buf.len());
+    }
+
+    #[test]
+    fn test_column_operation_unordered() {
+        test_column_operation_unordered_aux(300u32, 3);
+        test_column_operation_unordered_aux(1u32, 2);
+        test_column_operation_unordered_aux(0u32, 1);
+    }
+}

columnar/src/writer/column_writers.rs

@@ -0,0 +1,270 @@
+use std::cmp::Ordering;
+
+use stacker::{ExpUnrolledLinkedList, MemoryArena};
+
+use crate::dictionary::{DictionaryBuilder, UnorderedId};
+use crate::writer::column_operation::{ColumnOperation, SymbolValue};
+use crate::{Cardinality, DocId, NumericalType, NumericalValue};
+
+#[derive(Copy, Clone, Debug, Eq, PartialEq)]
+#[repr(u8)]
+enum DocumentStep {
+    SameDoc = 0,
+    NextDoc = 1,
+    SkippedDoc = 2,
+}
+
+#[inline(always)]
+fn delta_with_last_doc(last_doc_opt: Option<u32>, doc: u32) -> DocumentStep {
+    let expected_next_doc = last_doc_opt.map(|last_doc| last_doc + 1).unwrap_or(0u32);
+    match doc.cmp(&expected_next_doc) {
+        Ordering::Less => DocumentStep::SameDoc,
+        Ordering::Equal => DocumentStep::NextDoc,
+        Ordering::Greater => DocumentStep::SkippedDoc,
+    }
+}
+
+#[derive(Copy, Clone, Default)]
+pub struct ColumnWriter {
+    // Detected cardinality of the column so far.
+    cardinality: Cardinality,
+    // Last document inserted.
+    // None if no doc has been added yet.
+    last_doc_opt: Option<u32>,
+    // Buffer containing the serialized values.
+    values: ExpUnrolledLinkedList,
+}
+
+impl ColumnWriter {
+    /// Returns an iterator over the Symbol that have been recorded
+    /// for the given column.
+    pub(crate) fn operation_iterator<'a, V: SymbolValue>(
+        &self,
+        arena: &MemoryArena,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<V>> + 'a {
+        buffer.clear();
+        self.values.read_to_end(arena, buffer);
+        let mut cursor: &[u8] = &buffer[..];
+        std::iter::from_fn(move || ColumnOperation::deserialize(&mut cursor))
+    }
+
+    /// Records a change of the document being recorded.
+    ///
+    /// This function will also update the cardinality of the column
+    /// if necessary.
+    pub(crate) fn record<S: SymbolValue>(&mut self, doc: DocId, value: S, arena: &mut MemoryArena) {
+        // Difference between `doc` and the last doc.
+        match delta_with_last_doc(self.last_doc_opt, doc) {
+            DocumentStep::SameDoc => {
+                // This is the last encounterred document.
+                self.cardinality = Cardinality::Multivalued;
+            }
+            DocumentStep::NextDoc => {
+                self.last_doc_opt = Some(doc);
+                self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
+            }
+            DocumentStep::SkippedDoc => {
+                self.cardinality = self.cardinality.max(Cardinality::Optional);
+                self.last_doc_opt = Some(doc);
+                self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
+            }
+        }
+        self.write_symbol(ColumnOperation::Value(value), arena);
+    }
+
+    // Get the cardinality.
+    // The overall number of docs in the column is necessary to
+    // deal with the case where the all docs contain 1 value, except some documents
+    // at the end of the column.
+    pub fn get_cardinality(&self, num_docs: DocId) -> Cardinality {
+        match delta_with_last_doc(self.last_doc_opt, num_docs) {
+            DocumentStep::SameDoc | DocumentStep::NextDoc => self.cardinality,
+            DocumentStep::SkippedDoc => self.cardinality.max(Cardinality::Optional),
+        }
+    }
+
+    /// Appends a new symbol to the `ColumnWriter`.
+    fn write_symbol<V: SymbolValue>(
+        &mut self,
+        column_operation: ColumnOperation<V>,
+        arena: &mut MemoryArena,
+    ) {
+        self.values
+            .writer(arena)
+            .extend_from_slice(column_operation.serialize().as_ref());
+    }
+}
+
+#[derive(Clone, Copy, Default)]
+pub(crate) struct NumericalColumnWriter {
+    compatible_numerical_types: CompatibleNumericalTypes,
+    column_writer: ColumnWriter,
+}
+
+/// State used to store what types are still acceptable
+/// after having seen a set of numerical values.
+#[derive(Clone, Copy)]
+pub(crate) struct CompatibleNumericalTypes {
+    all_values_within_i64_range: bool,
+    all_values_within_u64_range: bool,
+    // f64 is always acceptable.
+}
+
+impl Default for CompatibleNumericalTypes {
+    fn default() -> CompatibleNumericalTypes {
+        CompatibleNumericalTypes {
+            all_values_within_i64_range: true,
+            all_values_within_u64_range: true,
+        }
+    }
+}
+
+impl CompatibleNumericalTypes {
+    fn accept_value(&mut self, numerical_value: NumericalValue) {
+        match numerical_value {
+            NumericalValue::I64(val_i64) => {
+                let value_within_u64_range = val_i64 >= 0i64;
+                self.all_values_within_u64_range &= value_within_u64_range;
+            }
+            NumericalValue::U64(val_u64) => {
+                let value_within_i64_range = val_u64 < i64::MAX as u64;
+                self.all_values_within_i64_range &= value_within_i64_range;
+            }
+            NumericalValue::F64(_) => {
+                self.all_values_within_i64_range = false;
+                self.all_values_within_u64_range = false;
+            }
+        }
+    }
+
+    pub fn to_numerical_type(self) -> NumericalType {
+        if self.all_values_within_i64_range {
+            NumericalType::I64
+        } else if self.all_values_within_u64_range {
+            NumericalType::U64
+        } else {
+            NumericalType::F64
+        }
+    }
+}
+
+impl NumericalColumnWriter {
+    pub fn column_type_and_cardinality(&self, num_docs: DocId) -> (NumericalType, Cardinality) {
+        let numerical_type = self.compatible_numerical_types.to_numerical_type();
+        let cardinality = self.column_writer.get_cardinality(num_docs);
+        (numerical_type, cardinality)
+    }
+    pub fn record_numerical_value(
+        &mut self,
+        doc: DocId,
+        value: NumericalValue,
+        arena: &mut MemoryArena,
+    ) {
+        self.compatible_numerical_types.accept_value(value);
+        self.column_writer.record(doc, value, arena);
+    }
+
+    pub fn operation_iterator<'a>(
+        self,
+        arena: &MemoryArena,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<NumericalValue>> + 'a {
+        self.column_writer.operation_iterator(arena, buffer)
+    }
+}
+
+#[derive(Copy, Clone, Default)]
+pub struct StrColumnWriter {
+    pub(crate) dictionary_id: u32,
+    pub(crate) column_writer: ColumnWriter,
+}
+
+impl StrColumnWriter {
+    pub fn with_dictionary_id(dictionary_id: u32) -> StrColumnWriter {
+        StrColumnWriter {
+            dictionary_id,
+            column_writer: Default::default(),
+        }
+    }
+
+    pub(crate) fn record_bytes(
+        &mut self,
+        doc: DocId,
+        bytes: &[u8],
+        dictionaries: &mut [DictionaryBuilder],
+        arena: &mut MemoryArena,
+    ) {
+        let unordered_id = dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes);
+        self.column_writer.record(doc, unordered_id, arena);
+    }
+
+    pub(crate) fn operation_iterator<'a>(
+        &self,
+        arena: &MemoryArena,
+        byte_buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<UnorderedId>> + 'a {
+        self.column_writer.operation_iterator(arena, byte_buffer)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_delta_with_last_doc() {
+        assert_eq!(delta_with_last_doc(None, 0u32), DocumentStep::NextDoc);
+        assert_eq!(delta_with_last_doc(None, 1u32), DocumentStep::SkippedDoc);
+        assert_eq!(delta_with_last_doc(None, 2u32), DocumentStep::SkippedDoc);
+        assert_eq!(delta_with_last_doc(Some(0u32), 0u32), DocumentStep::SameDoc);
+        assert_eq!(delta_with_last_doc(Some(1u32), 1u32), DocumentStep::SameDoc);
+        assert_eq!(delta_with_last_doc(Some(1u32), 2u32), DocumentStep::NextDoc);
+        assert_eq!(
+            delta_with_last_doc(Some(1u32), 3u32),
+            DocumentStep::SkippedDoc
+        );
+        assert_eq!(
+            delta_with_last_doc(Some(1u32), 4u32),
+            DocumentStep::SkippedDoc
+        );
+    }
+
+    #[track_caller]
+    fn test_column_writer_coercion_iter_aux(
+        values: impl Iterator<Item = NumericalValue>,
+        expected_numerical_type: NumericalType,
+    ) {
+        let mut compatible_numerical_types = CompatibleNumericalTypes::default();
+        for value in values {
+            compatible_numerical_types.accept_value(value);
+        }
+        assert_eq!(
+            compatible_numerical_types.to_numerical_type(),
+            expected_numerical_type
+        );
+    }
+
+    #[track_caller]
+    fn test_column_writer_coercion_aux(
+        values: &[NumericalValue],
+        expected_numerical_type: NumericalType,
+    ) {
+        test_column_writer_coercion_iter_aux(values.iter().copied(), expected_numerical_type);
+        test_column_writer_coercion_iter_aux(values.iter().rev().copied(), expected_numerical_type);
+    }
+
+    #[test]
+    fn test_column_writer_coercion() {
+        test_column_writer_coercion_aux(&[], NumericalType::I64);
+        test_column_writer_coercion_aux(&[1i64.into()], NumericalType::I64);
+        test_column_writer_coercion_aux(&[1u64.into()], NumericalType::I64);
+        // We don't detect exact integer at the moment. We could!
+        test_column_writer_coercion_aux(&[1f64.into()], NumericalType::F64);
+        test_column_writer_coercion_aux(&[u64::MAX.into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[(i64::MAX as u64).into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[(1u64 << 63).into()], NumericalType::U64);
+        test_column_writer_coercion_aux(&[1i64.into(), 1u64.into()], NumericalType::I64);
+        test_column_writer_coercion_aux(&[u64::MAX.into(), (-1i64).into()], NumericalType::F64);
+    }
+}

columnar/src/writer/mod.rs

@@ -0,0 +1,526 @@
+mod column_operation;
+mod column_writers;
+mod serializer;
+mod value_index;
+
+use std::io::{self, Write};
+
+use column_operation::ColumnOperation;
+use fastfield_codecs::serialize::ValueIndexInfo;
+use fastfield_codecs::{Column, MonotonicallyMappableToU64, VecColumn};
+use serializer::ColumnarSerializer;
+use stacker::{Addr, ArenaHashMap, MemoryArena};
+
+use crate::column_type_header::{ColumnType, ColumnTypeAndCardinality, GeneralType};
+use crate::dictionary::{DictionaryBuilder, IdMapping, UnorderedId};
+use crate::value::{Coerce, NumericalType, NumericalValue};
+use crate::writer::column_writers::{ColumnWriter, NumericalColumnWriter, StrColumnWriter};
+use crate::writer::value_index::{IndexBuilder, SpareIndexBuilders};
+use crate::{Cardinality, DocId};
+
+/// Threshold above which a column data will be compressed
+/// using ZSTD.
+const COLUMN_COMPRESSION_THRESHOLD: usize = 100_000;
+
+/// This is a set of buffers that are only here
+/// to limit the amount of allocation.
+#[derive(Default)]
+struct SpareBuffers {
+    value_index_builders: SpareIndexBuilders,
+    i64_values: Vec<i64>,
+    u64_values: Vec<u64>,
+    f64_values: Vec<f64>,
+    bool_values: Vec<bool>,
+    column_buffer: Vec<u8>,
+}
+
+pub struct ColumnarWriter {
+    numerical_field_hash_map: ArenaHashMap,
+    bool_field_hash_map: ArenaHashMap,
+    bytes_field_hash_map: ArenaHashMap,
+    arena: MemoryArena,
+    // Dictionaries used to store dictionary-encoded values.
+    dictionaries: Vec<DictionaryBuilder>,
+    buffers: SpareBuffers,
+}
+
+impl Default for ColumnarWriter {
+    fn default() -> Self {
+        ColumnarWriter {
+            numerical_field_hash_map: ArenaHashMap::new(10_000),
+            bool_field_hash_map: ArenaHashMap::new(10_000),
+            bytes_field_hash_map: ArenaHashMap::new(10_000),
+            dictionaries: Vec::new(),
+            arena: MemoryArena::default(),
+            buffers: SpareBuffers::default(),
+        }
+    }
+}
+
+impl ColumnarWriter {
+    pub fn record_numerical(
+        &mut self,
+        doc: DocId,
+        column_name: &str,
+        numerical_value: NumericalValue,
+    ) {
+        assert!(
+            !column_name.as_bytes().contains(&0u8),
+            "key may not contain the 0 byte"
+        );
+        let (hash_map, arena) = (&mut self.numerical_field_hash_map, &mut self.arena);
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<NumericalColumnWriter>| {
+                let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
+                column.record_numerical_value(doc, numerical_value, arena);
+                column
+            },
+        );
+    }
+
+    pub fn record_bool(&mut self, doc: DocId, column_name: &str, val: bool) {
+        assert!(
+            !column_name.as_bytes().contains(&0u8),
+            "key may not contain the 0 byte"
+        );
+        let (hash_map, arena) = (&mut self.bool_field_hash_map, &mut self.arena);
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<ColumnWriter>| {
+                let mut column: ColumnWriter = column_opt.unwrap_or_default();
+                column.record(doc, val, arena);
+                column
+            },
+        );
+    }
+
+    pub fn record_str(&mut self, doc: DocId, column_name: &str, value: &[u8]) {
+        assert!(
+            !column_name.as_bytes().contains(&0u8),
+            "key may not contain the 0 byte"
+        );
+        let (hash_map, arena, dictionaries) = (
+            &mut self.bytes_field_hash_map,
+            &mut self.arena,
+            &mut self.dictionaries,
+        );
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<StrColumnWriter>| {
+                let mut column: StrColumnWriter = column_opt.unwrap_or_else(|| {
+                    let dictionary_id = dictionaries.len() as u32;
+                    dictionaries.push(DictionaryBuilder::default());
+                    StrColumnWriter::with_dictionary_id(dictionary_id)
+                });
+                column.record_bytes(doc, value, dictionaries, arena);
+                column
+            },
+        );
+    }
+
+    pub fn serialize(&mut self, num_docs: DocId, wrt: &mut dyn io::Write) -> io::Result<()> {
+        let mut serializer = ColumnarSerializer::new(wrt);
+        let mut field_columns: Vec<(&[u8], GeneralType, Addr)> = self
+            .numerical_field_hash_map
+            .iter()
+            .map(|(term, addr, _)| (term, GeneralType::Numerical, addr))
+            .collect();
+        field_columns.extend(
+            self.bytes_field_hash_map
+                .iter()
+                .map(|(term, addr, _)| (term, GeneralType::Str, addr)),
+        );
+        field_columns.extend(
+            self.bool_field_hash_map
+                .iter()
+                .map(|(term, addr, _)| (term, GeneralType::Bool, addr)),
+        );
+        field_columns.sort_unstable_by_key(|(column_name, col_type, _)| (*column_name, *col_type));
+        let (arena, buffers, dictionaries) = (&self.arena, &mut self.buffers, &self.dictionaries);
+        let mut symbol_byte_buffer: Vec<u8> = Vec::new();
+        for (column_name, bytes_or_numerical, addr) in field_columns {
+            match bytes_or_numerical {
+                GeneralType::Bool => {
+                    let column_writer: ColumnWriter = self.bool_field_hash_map.read(addr);
+                    let cardinality = column_writer.get_cardinality(num_docs);
+                    let column_type_and_cardinality = ColumnTypeAndCardinality {
+                        cardinality,
+                        typ: ColumnType::Bool,
+                    };
+                    let column_serializer =
+                        serializer.serialize_column(column_name, column_type_and_cardinality);
+                    serialize_bool_column(
+                        cardinality,
+                        num_docs,
+                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        column_serializer,
+                    )?;
+                }
+                GeneralType::Str => {
+                    let str_column_writer: StrColumnWriter = self.bytes_field_hash_map.read(addr);
+                    let dictionary_builder =
+                        &dictionaries[str_column_writer.dictionary_id as usize];
+                    let cardinality = str_column_writer.column_writer.get_cardinality(num_docs);
+                    let column_type_and_cardinality = ColumnTypeAndCardinality {
+                        cardinality,
+                        typ: ColumnType::Bytes,
+                    };
+                    let column_serializer =
+                        serializer.serialize_column(column_name, column_type_and_cardinality);
+                    serialize_bytes_column(
+                        cardinality,
+                        num_docs,
+                        dictionary_builder,
+                        str_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        column_serializer,
+                    )?;
+                }
+                GeneralType::Numerical => {
+                    let numerical_column_writer: NumericalColumnWriter =
+                        self.numerical_field_hash_map.read(addr);
+                    let (numerical_type, cardinality) =
+                        numerical_column_writer.column_type_and_cardinality(num_docs);
+                    let column_type_and_cardinality = ColumnTypeAndCardinality {
+                        cardinality,
+                        typ: ColumnType::Numerical(numerical_type),
+                    };
+                    let column_serializer =
+                        serializer.serialize_column(column_name, column_type_and_cardinality);
+                    serialize_numerical_column(
+                        cardinality,
+                        num_docs,
+                        numerical_type,
+                        numerical_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        column_serializer,
+                    )?;
+                }
+            };
+        }
+        serializer.finalize()?;
+        Ok(())
+    }
+}
+
+fn compress_and_write_column<W: io::Write>(column_bytes: &[u8], wrt: &mut W) -> io::Result<()> {
+    if column_bytes.len() >= COLUMN_COMPRESSION_THRESHOLD {
+        wrt.write_all(&[1])?;
+        let mut encoder = zstd::Encoder::new(wrt, 3)?;
+        encoder.write_all(column_bytes)?;
+        encoder.finish()?;
+    } else {
+        wrt.write_all(&[0])?;
+        wrt.write_all(column_bytes)?;
+    }
+    Ok(())
+}
+
+fn serialize_bytes_column<W: io::Write>(
+    cardinality: Cardinality,
+    num_docs: DocId,
+    dictionary_builder: &DictionaryBuilder,
+    operation_it: impl Iterator<Item = ColumnOperation<UnorderedId>>,
+    buffers: &mut SpareBuffers,
+    mut wrt: W,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        u64_values,
+        column_buffer,
+        ..
+    } = buffers;
+    column_buffer.clear();
+    let id_mapping: IdMapping = dictionary_builder.serialize(column_buffer)?;
+    let dictionary_num_bytes: u32 = column_buffer.len() as u32;
+    let operation_iterator = operation_it.map(|symbol: ColumnOperation<UnorderedId>| {
+        // We map unordered ids to ordered ids.
+        match symbol {
+            ColumnOperation::Value(unordered_id) => {
+                let ordered_id = id_mapping.to_ord(unordered_id);
+                ColumnOperation::Value(ordered_id.0 as u64)
+            }
+            ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
+        }
+    });
+    serialize_column(
+        operation_iterator,
+        cardinality,
+        num_docs,
+        value_index_builders,
+        u64_values,
+        column_buffer,
+    )?;
+    column_buffer.write_all(&dictionary_num_bytes.to_le_bytes()[..])?;
+    compress_and_write_column(column_buffer, &mut wrt)?;
+    Ok(())
+}
+
+fn serialize_numerical_column<W: io::Write>(
+    cardinality: Cardinality,
+    num_docs: DocId,
+    numerical_type: NumericalType,
+    op_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
+    buffers: &mut SpareBuffers,
+    mut wrt: W,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        u64_values,
+        i64_values,
+        f64_values,
+        column_buffer,
+        ..
+    } = buffers;
+    column_buffer.clear();
+    match numerical_type {
+        NumericalType::I64 => {
+            serialize_column(
+                coerce_numerical_symbol::<i64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                i64_values,
+                column_buffer,
+            )?;
+        }
+        NumericalType::U64 => {
+            serialize_column(
+                coerce_numerical_symbol::<u64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                u64_values,
+                column_buffer,
+            )?;
+        }
+        NumericalType::F64 => {
+            serialize_column(
+                coerce_numerical_symbol::<f64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                f64_values,
+                column_buffer,
+            )?;
+        }
+    };
+    compress_and_write_column(column_buffer, &mut wrt)?;
+    Ok(())
+}
+
+fn serialize_bool_column<W: io::Write>(
+    cardinality: Cardinality,
+    num_docs: DocId,
+    column_operations_it: impl Iterator<Item = ColumnOperation<bool>>,
+    buffers: &mut SpareBuffers,
+    mut wrt: W,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        bool_values,
+        column_buffer,
+        ..
+    } = buffers;
+    column_buffer.clear();
+    serialize_column(
+        column_operations_it,
+        cardinality,
+        num_docs,
+        value_index_builders,
+        bool_values,
+        column_buffer,
+    )?;
+    compress_and_write_column(column_buffer, &mut wrt)?;
+    Ok(())
+}
+
+fn serialize_column<
+    T: Copy + Default + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU64 + PartialOrd,
+>(
+    op_iterator: impl Iterator<Item = ColumnOperation<T>>,
+    cardinality: Cardinality,
+    num_docs: DocId,
+    value_index_builders: &mut SpareIndexBuilders,
+    values: &mut Vec<T>,
+    wrt: &mut Vec<u8>,
+) -> io::Result<()>
+where
+    for<'a> VecColumn<'a, T>: Column<T>,
+{
+    values.clear();
+    match cardinality {
+        Cardinality::Required => {
+            consume_operation_iterator(
+                op_iterator,
+                value_index_builders.borrow_required_index_builder(),
+                values,
+            );
+            fastfield_codecs::serialize(
+                VecColumn::from(&values[..]),
+                wrt,
+                &fastfield_codecs::ALL_CODEC_TYPES[..],
+            )?;
+        }
+        Cardinality::Optional => {
+            let optional_index_builder = value_index_builders.borrow_optional_index_builder();
+            consume_operation_iterator(op_iterator, optional_index_builder, values);
+            let optional_index = optional_index_builder.finish(num_docs);
+            fastfield_codecs::serialize::serialize_new(
+                ValueIndexInfo::SingleValue(Box::new(optional_index)),
+                VecColumn::from(&values[..]),
+                wrt,
+                &fastfield_codecs::ALL_CODEC_TYPES[..],
+            )?;
+        }
+        Cardinality::Multivalued => {
+            let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
+            consume_operation_iterator(op_iterator, multivalued_index_builder, values);
+            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            fastfield_codecs::serialize::serialize_new(
+                ValueIndexInfo::MultiValue(Box::new(multivalued_index)),
+                VecColumn::from(&values[..]),
+                wrt,
+                &fastfield_codecs::ALL_CODEC_TYPES[..],
+            )?;
+        }
+    }
+    Ok(())
+}
+
+fn coerce_numerical_symbol<T>(
+    operation_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
+) -> impl Iterator<Item = ColumnOperation<T>>
+where T: Coerce {
+    operation_iterator.map(|symbol| match symbol {
+        ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
+        ColumnOperation::Value(numerical_value) => {
+            ColumnOperation::Value(Coerce::coerce(numerical_value))
+        }
+    })
+}
+
+fn consume_operation_iterator<T: std::fmt::Debug, TIndexBuilder: IndexBuilder>(
+    operation_iterator: impl Iterator<Item = ColumnOperation<T>>,
+    index_builder: &mut TIndexBuilder,
+    values: &mut Vec<T>,
+) {
+    for symbol in operation_iterator {
+        match symbol {
+            ColumnOperation::NewDoc(doc) => {
+                index_builder.record_doc(doc);
+            }
+            ColumnOperation::Value(value) => {
+                index_builder.record_value();
+                values.push(value);
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use column_operation::ColumnOperation;
+    use stacker::MemoryArena;
+
+    use super::*;
+    use crate::value::NumericalValue;
+    use crate::Cardinality;
+
+    #[test]
+    fn test_column_writer_required_simple() {
+        let mut arena = MemoryArena::default();
+        let mut column_writer = super::ColumnWriter::default();
+        column_writer.record(0u32, NumericalValue::from(14i64), &mut arena);
+        column_writer.record(1u32, NumericalValue::from(15i64), &mut arena);
+        column_writer.record(2u32, NumericalValue::from(-16i64), &mut arena);
+        assert_eq!(column_writer.get_cardinality(3), Cardinality::Required);
+        let mut buffer = Vec::new();
+        let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
+            .operation_iterator(&mut arena, &mut buffer)
+            .collect();
+        assert_eq!(symbols.len(), 6);
+        assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
+        assert!(matches!(
+            symbols[1],
+            ColumnOperation::Value(NumericalValue::I64(14i64))
+        ));
+        assert!(matches!(symbols[2], ColumnOperation::NewDoc(1u32)));
+        assert!(matches!(
+            symbols[3],
+            ColumnOperation::Value(NumericalValue::I64(15i64))
+        ));
+        assert!(matches!(symbols[4], ColumnOperation::NewDoc(2u32)));
+        assert!(matches!(
+            symbols[5],
+            ColumnOperation::Value(NumericalValue::I64(-16i64))
+        ));
+    }
+
+    #[test]
+    fn test_column_writer_optional_cardinality_missing_first() {
+        let mut arena = MemoryArena::default();
+        let mut column_writer = super::ColumnWriter::default();
+        column_writer.record(1u32, NumericalValue::from(15i64), &mut arena);
+        column_writer.record(2u32, NumericalValue::from(-16i64), &mut arena);
+        assert_eq!(column_writer.get_cardinality(3), Cardinality::Optional);
+        let mut buffer = Vec::new();
+        let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
+            .operation_iterator(&mut arena, &mut buffer)
+            .collect();
+        assert_eq!(symbols.len(), 4);
+        assert!(matches!(symbols[0], ColumnOperation::NewDoc(1u32)));
+        assert!(matches!(
+            symbols[1],
+            ColumnOperation::Value(NumericalValue::I64(15i64))
+        ));
+        assert!(matches!(symbols[2], ColumnOperation::NewDoc(2u32)));
+        assert!(matches!(
+            symbols[3],
+            ColumnOperation::Value(NumericalValue::I64(-16i64))
+        ));
+    }
+
+    #[test]
+    fn test_column_writer_optional_cardinality_missing_last() {
+        let mut arena = MemoryArena::default();
+        let mut column_writer = super::ColumnWriter::default();
+        column_writer.record(0u32, NumericalValue::from(15i64), &mut arena);
+        assert_eq!(column_writer.get_cardinality(2), Cardinality::Optional);
+        let mut buffer = Vec::new();
+        let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
+            .operation_iterator(&mut arena, &mut buffer)
+            .collect();
+        assert_eq!(symbols.len(), 2);
+        assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
+        assert!(matches!(
+            symbols[1],
+            ColumnOperation::Value(NumericalValue::I64(15i64))
+        ));
+    }
+
+    #[test]
+    fn test_column_writer_multivalued() {
+        let mut arena = MemoryArena::default();
+        let mut column_writer = super::ColumnWriter::default();
+        column_writer.record(0u32, NumericalValue::from(16i64), &mut arena);
+        column_writer.record(0u32, NumericalValue::from(17i64), &mut arena);
+        assert_eq!(column_writer.get_cardinality(1), Cardinality::Multivalued);
+        let mut buffer = Vec::new();
+        let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
+            .operation_iterator(&mut arena, &mut buffer)
+            .collect();
+        assert_eq!(symbols.len(), 3);
+        assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
+        assert!(matches!(
+            symbols[1],
+            ColumnOperation::Value(NumericalValue::I64(16i64))
+        ));
+        assert!(matches!(
+            symbols[2],
+            ColumnOperation::Value(NumericalValue::I64(17i64))
+        ));
+    }
+}

columnar/src/writer/serializer.rs

@@ -0,0 +1,116 @@
+use std::io;
+use std::io::Write;
+
+use common::CountingWriter;
+use sstable::value::RangeValueWriter;
+use sstable::RangeSSTable;
+
+use crate::column_type_header::ColumnTypeAndCardinality;
+
+pub struct ColumnarSerializer<W: io::Write> {
+    wrt: CountingWriter<W>,
+    sstable_range: sstable::Writer<Vec<u8>, RangeValueWriter>,
+    prepare_key_buffer: Vec<u8>,
+}
+
+/// Returns a key consisting of the concatenation of the key and the column_type_and_cardinality
+/// code.
+fn prepare_key<'a>(
+    key: &[u8],
+    column_type_cardinality: ColumnTypeAndCardinality,
+    buffer: &'a mut Vec<u8>,
+) {
+    buffer.clear();
+    buffer.extend_from_slice(key);
+    buffer.push(0u8);
+    buffer.push(column_type_cardinality.to_code());
+}
+
+impl<W: io::Write> ColumnarSerializer<W> {
+    pub(crate) fn new(wrt: W) -> ColumnarSerializer<W> {
+        let sstable_range: sstable::Writer<Vec<u8>, RangeValueWriter> =
+            sstable::Dictionary::<RangeSSTable>::builder(Vec::with_capacity(100_000)).unwrap();
+        ColumnarSerializer {
+            wrt: CountingWriter::wrap(wrt),
+            sstable_range,
+            prepare_key_buffer: Vec::new(),
+        }
+    }
+
+    pub fn serialize_column<'a>(
+        &'a mut self,
+        column_name: &[u8],
+        column_type_cardinality: ColumnTypeAndCardinality,
+    ) -> impl io::Write + 'a {
+        let start_offset = self.wrt.written_bytes();
+        prepare_key(
+            column_name,
+            column_type_cardinality,
+            &mut self.prepare_key_buffer,
+        );
+        ColumnSerializer {
+            columnar_serializer: self,
+            start_offset,
+        }
+    }
+
+    pub(crate) fn finalize(mut self) -> io::Result<()> {
+        let sstable_bytes: Vec<u8> = self.sstable_range.finish()?;
+        let sstable_num_bytes: u64 = sstable_bytes.len() as u64;
+        self.wrt.write_all(&sstable_bytes)?;
+        self.wrt.write_all(&sstable_num_bytes.to_le_bytes()[..])?;
+        Ok(())
+    }
+}
+
+struct ColumnSerializer<'a, W: io::Write> {
+    columnar_serializer: &'a mut ColumnarSerializer<W>,
+    start_offset: u64,
+}
+
+impl<'a, W: io::Write> Drop for ColumnSerializer<'a, W> {
+    fn drop(&mut self) {
+        let end_offset: u64 = self.columnar_serializer.wrt.written_bytes();
+        let byte_range = self.start_offset..end_offset;
+        self.columnar_serializer.sstable_range.insert_cannot_fail(
+            &self.columnar_serializer.prepare_key_buffer[..],
+            &byte_range,
+        );
+        self.columnar_serializer.prepare_key_buffer.clear();
+    }
+}
+
+impl<'a, W: io::Write> io::Write for ColumnSerializer<'a, W> {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        self.columnar_serializer.wrt.write(buf)
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        self.columnar_serializer.wrt.flush()
+    }
+
+    fn write_all(&mut self, buf: &[u8]) -> io::Result<()> {
+        self.columnar_serializer.wrt.write_all(buf)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_type_header::ColumnType;
+    use crate::Cardinality;
+
+    #[test]
+    fn test_prepare_key_bytes() {
+        let mut buffer: Vec<u8> = b"somegarbage".to_vec();
+        let column_type_and_cardinality = ColumnTypeAndCardinality {
+            typ: ColumnType::Bytes,
+            cardinality: Cardinality::Optional,
+        };
+        prepare_key(b"root\0child", column_type_and_cardinality, &mut buffer);
+        assert_eq!(buffer.len(), 12);
+        assert_eq!(&buffer[..10], b"root\0child");
+        assert_eq!(buffer[10], 0u8);
+        assert_eq!(buffer[11], column_type_and_cardinality.to_code());
+    }
+}

columnar/src/writer/value_index.rs

@@ -0,0 +1,218 @@
+use fastfield_codecs::serialize::{MultiValueIndexInfo, SingleValueIndexInfo};
+
+use crate::DocId;
+
+/// The `IndexBuilder` interprets a sequence of
+/// calls of the form:
+/// (record_doc,record_value+)*
+/// and can then serialize the results into an index.
+///
+/// It has different implementation depending on whether the
+/// cardinality is required, optional, or multivalued.
+pub(crate) trait IndexBuilder {
+    fn record_doc(&mut self, doc: DocId);
+    #[inline]
+    fn record_value(&mut self) {}
+}
+
+/// The RequiredIndexBuilder does nothing.
+#[derive(Default)]
+pub struct RequiredIndexBuilder;
+
+impl IndexBuilder for RequiredIndexBuilder {
+    #[inline(always)]
+    fn record_doc(&mut self, _doc: DocId) {}
+}
+
+#[derive(Default)]
+pub struct OptionalIndexBuilder {
+    docs: Vec<DocId>,
+}
+
+struct SingleValueArrayIndex<'a> {
+    docs: &'a [DocId],
+    num_docs: DocId,
+}
+
+impl<'a> SingleValueIndexInfo for SingleValueArrayIndex<'a> {
+    fn num_vals(&self) -> u32 {
+        self.num_docs as u32
+    }
+
+    fn num_non_nulls(&self) -> u32 {
+        self.docs.len() as u32
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new(self.docs.iter().copied())
+    }
+}
+
+impl OptionalIndexBuilder {
+    pub fn finish(&mut self, num_docs: DocId) -> impl SingleValueIndexInfo + '_ {
+        debug_assert!(self
+            .docs
+            .last()
+            .copied()
+            .map(|last_doc| last_doc < num_docs)
+            .unwrap_or(true));
+        SingleValueArrayIndex {
+            docs: &self.docs[..],
+            num_docs,
+        }
+    }
+
+    fn reset(&mut self) {
+        self.docs.clear();
+    }
+}
+
+impl IndexBuilder for OptionalIndexBuilder {
+    #[inline(always)]
+    fn record_doc(&mut self, doc: DocId) {
+        debug_assert!(self
+            .docs
+            .last()
+            .copied()
+            .map(|prev_doc| doc > prev_doc)
+            .unwrap_or(true));
+        self.docs.push(doc);
+    }
+}
+
+#[derive(Default)]
+pub struct MultivaluedIndexBuilder {
+    // TODO should we switch to `start_offset`?
+    end_values: Vec<DocId>,
+    total_num_vals_seen: u32,
+}
+
+pub struct MultivaluedValueArrayIndex<'a> {
+    end_offsets: &'a [DocId],
+}
+
+impl<'a> MultiValueIndexInfo for MultivaluedValueArrayIndex<'a> {
+    fn num_docs(&self) -> u32 {
+        self.end_offsets.len() as u32
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.end_offsets.last().copied().unwrap_or(0u32)
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        if self.end_offsets.is_empty() {
+            return Box::new(std::iter::empty());
+        }
+        let n = self.end_offsets.len();
+        Box::new(std::iter::once(0u32).chain(self.end_offsets[..n - 1].iter().copied()))
+    }
+}
+
+impl MultivaluedIndexBuilder {
+    pub fn finish(&mut self, num_docs: DocId) -> impl MultiValueIndexInfo + '_ {
+        self.end_values
+            .resize(num_docs as usize, self.total_num_vals_seen);
+        MultivaluedValueArrayIndex {
+            end_offsets: &self.end_values[..],
+        }
+    }
+
+    fn reset(&mut self) {
+        self.end_values.clear();
+        self.total_num_vals_seen = 0;
+    }
+}
+
+impl IndexBuilder for MultivaluedIndexBuilder {
+    fn record_doc(&mut self, doc: DocId) {
+        self.end_values
+            .resize(doc as usize, self.total_num_vals_seen);
+    }
+
+    fn record_value(&mut self) {
+        self.total_num_vals_seen += 1;
+    }
+}
+
+/// The `SpareIndexBuilders` is there to avoid allocating a
+/// new index builder for every single column.
+#[derive(Default)]
+pub struct SpareIndexBuilders {
+    required_index_builder: RequiredIndexBuilder,
+    optional_index_builder: OptionalIndexBuilder,
+    multivalued_index_builder: MultivaluedIndexBuilder,
+}
+
+impl SpareIndexBuilders {
+    pub fn borrow_required_index_builder(&mut self) -> &mut RequiredIndexBuilder {
+        &mut self.required_index_builder
+    }
+
+    pub fn borrow_optional_index_builder(&mut self) -> &mut OptionalIndexBuilder {
+        self.optional_index_builder.reset();
+        &mut self.optional_index_builder
+    }
+
+    pub fn borrow_multivalued_index_builder(&mut self) -> &mut MultivaluedIndexBuilder {
+        self.multivalued_index_builder.reset();
+        &mut self.multivalued_index_builder
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_optional_value_index_builder() {
+        let mut opt_value_index_builder = OptionalIndexBuilder::default();
+        opt_value_index_builder.record_doc(0u32);
+        opt_value_index_builder.record_value();
+        assert_eq!(
+            &opt_value_index_builder
+                .finish(1u32)
+                .iter()
+                .collect::<Vec<u32>>(),
+            &[0]
+        );
+        opt_value_index_builder.reset();
+        opt_value_index_builder.record_doc(1u32);
+        opt_value_index_builder.record_value();
+        assert_eq!(
+            &opt_value_index_builder
+                .finish(2u32)
+                .iter()
+                .collect::<Vec<u32>>(),
+            &[1]
+        );
+    }
+
+    #[test]
+    fn test_multivalued_value_index_builder() {
+        let mut multivalued_value_index_builder = MultivaluedIndexBuilder::default();
+        multivalued_value_index_builder.record_doc(1u32);
+        multivalued_value_index_builder.record_value();
+        multivalued_value_index_builder.record_value();
+        multivalued_value_index_builder.record_doc(2u32);
+        multivalued_value_index_builder.record_value();
+        assert_eq!(
+            multivalued_value_index_builder
+                .finish(4u32)
+                .iter()
+                .collect::<Vec<u32>>(),
+            vec![0, 0, 2, 3]
+        );
+        multivalued_value_index_builder.reset();
+        multivalued_value_index_builder.record_doc(2u32);
+        multivalued_value_index_builder.record_value();
+        multivalued_value_index_builder.record_value();
+        assert_eq!(
+            multivalued_value_index_builder
+                .finish(4u32)
+                .iter()
+                .collect::<Vec<u32>>(),
+            vec![0, 0, 0, 2]
+        );
+    }
+}

common/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-common"
-version = "0.4.0"
+version = "0.5.0"
 authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
 license = "MIT"
 edition = "2021"
@@ -14,7 +14,8 @@ repository = "https://github.com/quickwit-oss/tantivy"
 
 [dependencies]
 byteorder = "1.4.3"
-ownedbytes = { version= "0.4", path="../ownedbytes" }
+ownedbytes = { version= "0.5", path="../ownedbytes" }
+async-trait = "0.1"
 
 [dev-dependencies]
 proptest = "1.0.0"

common/src/bitset.rs

@@ -151,7 +151,7 @@ impl TinySet {
         if self.is_empty() {
             None
         } else {
-            let lowest = self.0.trailing_zeros() as u32;
+            let lowest = self.0.trailing_zeros();
             self.0 ^= TinySet::singleton(lowest).0;
             Some(lowest)
         }
@@ -421,7 +421,7 @@ mod tests {
             bitset.serialize(&mut out).unwrap();
 
             let bitset = ReadOnlyBitSet::open(OwnedBytes::new(out));
-            assert_eq!(bitset.len() as usize, i as usize);
+            assert_eq!(bitset.len(), i as usize);
         }
     }
 
@@ -432,7 +432,7 @@ mod tests {
         bitset.serialize(&mut out).unwrap();
 
         let bitset = ReadOnlyBitSet::open(OwnedBytes::new(out));
-        assert_eq!(bitset.len() as usize, 64);
+        assert_eq!(bitset.len(), 64);
     }
 
     #[test]

common/src/file_slice.rs

@@ -1,19 +1,18 @@
-use std::ops::{Deref, Range};
+use std::ops::{Deref, Range, RangeBounds};
 use std::sync::Arc;
 use std::{fmt, io};
 
 use async_trait::async_trait;
-use common::HasLen;
+use ownedbytes::{OwnedBytes, StableDeref};
-use stable_deref_trait::StableDeref;
 
-use crate::directory::OwnedBytes;
+use crate::HasLen;
 
 /// Objects that represents files sections in tantivy.
 ///
 /// By contract, whatever happens to the directory file, as long as a FileHandle
 /// is alive, the data associated with it cannot be altered or destroyed.
 ///
-/// The underlying behavior is therefore specific to the [`Directory`](crate::Directory) that
+/// The underlying behavior is therefore specific to the `Directory` that
 /// created it. Despite its name, a [`FileSlice`] may or may not directly map to an actual file
 /// on the filesystem.
 
@@ -24,13 +23,12 @@ pub trait FileHandle: 'static + Send + Sync + HasLen + fmt::Debug {
     /// This method may panic if the range requested is invalid.
     fn read_bytes(&self, range: Range<usize>) -> io::Result<OwnedBytes>;
 
-    #[cfg(feature = "quickwit")]
     #[doc(hidden)]
-    async fn read_bytes_async(
+    async fn read_bytes_async(&self, _byte_range: Range<usize>) -> io::Result<OwnedBytes> {
-        &self,
+        Err(io::Error::new(
-        _byte_range: Range<usize>,
+            io::ErrorKind::Unsupported,
-    ) -> crate::AsyncIoResult<OwnedBytes> {
+            "Async read is not supported.",
-        Err(crate::error::AsyncIoError::AsyncUnsupported)
+        ))
     }
 }
 
@@ -41,8 +39,7 @@ impl FileHandle for &'static [u8] {
         Ok(OwnedBytes::new(bytes))
     }
 
-    #[cfg(feature = "quickwit")]
+    async fn read_bytes_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
-    async fn read_bytes_async(&self, byte_range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
         Ok(self.read_bytes(byte_range)?)
     }
 }
@@ -70,6 +67,34 @@ impl fmt::Debug for FileSlice {
     }
 }
 
+/// Takes a range, a `RangeBounds` object, and returns
+/// a `Range` that corresponds to the relative application of the
+/// `RangeBounds` object to the original `Range`.
+///
+/// For instance, combine_ranges(`[2..11)`, `[5..7]`) returns `[7..10]`
+/// as it reads, what is the sub-range that starts at the 5 element of
+/// `[2..11)` and ends at the 9th element included.
+///
+/// This function panics, if the result would suggest something outside
+/// of the bounds of the original range.
+fn combine_ranges<R: RangeBounds<usize>>(orig_range: Range<usize>, rel_range: R) -> Range<usize> {
+    let start: usize = orig_range.start
+        + match rel_range.start_bound().cloned() {
+            std::ops::Bound::Included(rel_start) => rel_start,
+            std::ops::Bound::Excluded(rel_start) => rel_start + 1,
+            std::ops::Bound::Unbounded => 0,
+        };
+    assert!(start <= orig_range.end);
+    let end: usize = match rel_range.end_bound().cloned() {
+        std::ops::Bound::Included(rel_end) => orig_range.start + rel_end + 1,
+        std::ops::Bound::Excluded(rel_end) => orig_range.start + rel_end,
+        std::ops::Bound::Unbounded => orig_range.end,
+    };
+    assert!(end >= start);
+    assert!(end <= orig_range.end);
+    start..end
+}
+
 impl FileSlice {
     /// Wraps a FileHandle.
     pub fn new(file_handle: Arc<dyn FileHandle>) -> Self {
@@ -93,11 +118,11 @@ impl FileSlice {
     ///
     /// Panics if `byte_range.end` exceeds the filesize.
     #[must_use]
-    pub fn slice(&self, byte_range: Range<usize>) -> FileSlice {
+    #[inline]
-        assert!(byte_range.end <= self.len());
+    pub fn slice<R: RangeBounds<usize>>(&self, byte_range: R) -> FileSlice {
         FileSlice {
             data: self.data.clone(),
-            range: self.range.start + byte_range.start..self.range.start + byte_range.end,
+            range: combine_ranges(self.range.clone(), byte_range),
         }
     }
 
@@ -117,9 +142,8 @@ impl FileSlice {
         self.data.read_bytes(self.range.clone())
     }
 
-    #[cfg(feature = "quickwit")]
     #[doc(hidden)]
-    pub async fn read_bytes_async(&self) -> crate::AsyncIoResult<OwnedBytes> {
+    pub async fn read_bytes_async(&self) -> io::Result<OwnedBytes> {
         self.data.read_bytes_async(self.range.clone()).await
     }
 
@@ -137,12 +161,8 @@ impl FileSlice {
             .read_bytes(self.range.start + range.start..self.range.start + range.end)
     }
 
-    #[cfg(feature = "quickwit")]
     #[doc(hidden)]
-    pub async fn read_bytes_slice_async(
+    pub async fn read_bytes_slice_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
-        &self,
-        byte_range: Range<usize>,
-    ) -> crate::AsyncIoResult<OwnedBytes> {
         assert!(
             self.range.start + byte_range.end <= self.range.end,
             "`to` exceeds the fileslice length"
@@ -204,8 +224,7 @@ impl FileHandle for FileSlice {
         self.read_bytes_slice(range)
     }
 
-    #[cfg(feature = "quickwit")]
+    async fn read_bytes_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
-    async fn read_bytes_async(&self, byte_range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
         self.read_bytes_slice_async(byte_range).await
     }
 }
@@ -222,21 +241,20 @@ impl FileHandle for OwnedBytes {
         Ok(self.slice(range))
     }
 
-    #[cfg(feature = "quickwit")]
+    async fn read_bytes_async(&self, range: Range<usize>) -> io::Result<OwnedBytes> {
-    async fn read_bytes_async(&self, range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
+        self.read_bytes(range)
-        let bytes = self.read_bytes(range)?;
-        Ok(bytes)
     }
 }
 
 #[cfg(test)]
 mod tests {
     use std::io;
+    use std::ops::Bound;
     use std::sync::Arc;
 
-    use common::HasLen;
-
     use super::{FileHandle, FileSlice};
+    use crate::file_slice::combine_ranges;
+    use crate::HasLen;
 
     #[test]
     fn test_file_slice() -> io::Result<()> {
@@ -307,4 +325,23 @@ mod tests {
             b"bcd"
         );
     }
+
+    #[test]
+    fn test_combine_range() {
+        assert_eq!(combine_ranges(1..3, 0..1), 1..2);
+        assert_eq!(combine_ranges(1..3, 1..), 2..3);
+        assert_eq!(combine_ranges(1..4, ..2), 1..3);
+        assert_eq!(combine_ranges(3..10, 2..5), 5..8);
+        assert_eq!(combine_ranges(2..11, 5..=7), 7..10);
+        assert_eq!(
+            combine_ranges(2..11, (Bound::Excluded(5), Bound::Unbounded)),
+            8..11
+        );
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_combine_range_panics() {
+        let _ = combine_ranges(3..5, 1..4);
+    }
 }

common/src/lib.rs

@@ -5,11 +5,12 @@ use std::ops::Deref;
 pub use byteorder::LittleEndian as Endianness;
 
 mod bitset;
+pub mod file_slice;
 mod serialize;
 mod vint;
 mod writer;
-
 pub use bitset::*;
+pub use ownedbytes::{OwnedBytes, StableDeref};
 pub use serialize::{BinarySerializable, DeserializeFrom, FixedSize};
 pub use vint::{
     deserialize_vint_u128, read_u32_vint, read_u32_vint_no_advance, serialize_vint_u128,

fastfield_codecs/Cargo.toml

@@ -12,9 +12,8 @@ repository = "https://github.com/quickwit-oss/tantivy"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-common = { version = "0.4", path = "../common/", package = "tantivy-common" }
+common = { version = "0.5", path = "../common/", package = "tantivy-common" }
 tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-ownedbytes = { version = "0.4.0", path = "../ownedbytes" }
 prettytable-rs = {version="0.9.0", optional= true}
 rand = {version="0.8.3", optional= true}
 fastdivide = "0.4"

fastfield_codecs/benches/bench.rs

@@ -7,8 +7,8 @@ mod tests {
     use std::iter;
     use std::sync::Arc;
 
+    use common::OwnedBytes;
     use fastfield_codecs::*;
-    use ownedbytes::OwnedBytes;
     use rand::prelude::*;
     use test::Bencher;
 

fastfield_codecs/src/bitpacked.rs

@@ -1,6 +1,6 @@
 use std::io::{self, Write};
 
-use ownedbytes::OwnedBytes;
+use common::OwnedBytes;
 use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
 
 use crate::serialize::NormalizedHeader;

fastfield_codecs/src/blockwise_linear.rs

@@ -1,8 +1,7 @@
 use std::sync::Arc;
 use std::{io, iter};
 
-use common::{BinarySerializable, CountingWriter, DeserializeFrom};
+use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
-use ownedbytes::OwnedBytes;
 use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
 
 use crate::line::Line;
@@ -47,7 +46,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
     type Reader = BlockwiseLinearReader;
 
     fn open_from_bytes(
-        bytes: ownedbytes::OwnedBytes,
+        bytes: common::OwnedBytes,
         normalized_header: NormalizedHeader,
     ) -> io::Result<Self::Reader> {
         let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
@@ -75,7 +74,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
         if column.num_vals() < 10 * CHUNK_SIZE as u32 {
             return None;
         }
-        let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE as usize).collect();
+        let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE).collect();
         let line = Line::train(&VecColumn::from(&first_chunk));
         for (i, buffer_val) in first_chunk.iter_mut().enumerate() {
             let interpolated_val = line.eval(i as u32);
@@ -171,15 +170,18 @@ impl Column for BlockwiseLinearReader {
         interpoled_val.wrapping_add(bitpacked_diff)
     }
 
+    #[inline(always)]
     fn min_value(&self) -> u64 {
         // The BlockwiseLinearReader assumes a normalized vector.
         0u64
     }
 
+    #[inline(always)]
     fn max_value(&self) -> u64 {
         self.normalized_header.max_value
     }
 
+    #[inline(always)]
     fn num_vals(&self) -> u32 {
         self.normalized_header.num_vals
     }

fastfield_codecs/src/column.rs

@@ -135,7 +135,7 @@ impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
     }
 }
 
-impl<'a, T: Copy + Ord + Default, V> From<&'a V> for VecColumn<'a, T>
+impl<'a, T: Copy + PartialOrd + Default, V> From<&'a V> for VecColumn<'a, T>
 where V: AsRef<[T]> + ?Sized
 {
     fn from(values: &'a V) -> Self {

fastfield_codecs/src/compact_space/build_compact_space.rs

@@ -208,7 +208,7 @@ impl CompactSpaceBuilder {
             };
             let covered_range_len = range_mapping.range_length();
             ranges_mapping.push(range_mapping);
-            compact_start += covered_range_len as u64;
+            compact_start += covered_range_len;
         }
         // println!("num ranges {}", ranges_mapping.len());
         CompactSpace { ranges_mapping }

fastfield_codecs/src/compact_space/mod.rs

@@ -17,8 +17,7 @@ use std::{
     ops::{Range, RangeInclusive},
 };
 
-use common::{BinarySerializable, CountingWriter, VInt, VIntU128};
+use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
-use ownedbytes::OwnedBytes;
 use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
 
 use crate::compact_space::build_compact_space::get_compact_space;
@@ -97,7 +96,7 @@ impl BinarySerializable for CompactSpace {
             };
             let range_length = range_mapping.range_length();
             ranges_mapping.push(range_mapping);
-            compact_start += range_length as u64;
+            compact_start += range_length;
         }
 
         Ok(Self { ranges_mapping })
@@ -407,10 +406,10 @@ impl CompactSpaceDecompressor {
             let idx2 = idx + 1;
             let idx3 = idx + 2;
             let idx4 = idx + 3;
-            let val1 = get_val(idx1 as u32);
+            let val1 = get_val(idx1);
-            let val2 = get_val(idx2 as u32);
+            let val2 = get_val(idx2);
-            let val3 = get_val(idx3 as u32);
+            let val3 = get_val(idx3);
-            let val4 = get_val(idx4 as u32);
+            let val4 = get_val(idx4);
             push_if_in_range(idx1, val1);
             push_if_in_range(idx2, val2);
             push_if_in_range(idx3, val3);
@@ -419,14 +418,13 @@ impl CompactSpaceDecompressor {
 
         // handle rest
         for idx in cutoff..position_range.end {
-            push_if_in_range(idx, get_val(idx as u32));
+            push_if_in_range(idx, get_val(idx));
         }
     }
 
     #[inline]
     fn iter_compact(&self) -> impl Iterator<Item = u64> + '_ {
-        (0..self.params.num_vals)
+        (0..self.params.num_vals).map(move |idx| self.params.bit_unpacker.get(idx, &self.data))
-            .map(move |idx| self.params.bit_unpacker.get(idx, &self.data) as u64)
     }
 
     #[inline]
@@ -569,7 +567,7 @@ mod tests {
         let decomp = CompactSpaceDecompressor::open(data).unwrap();
         let complete_range = 0..vals.len() as u32;
         for (pos, val) in vals.iter().enumerate() {
-            let val = *val as u128;
+            let val = *val;
             let pos = pos as u32;
             let mut positions = Vec::new();
             decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
@@ -666,7 +664,7 @@ mod tests {
             get_positions_for_value_range_helper(
                 &decomp,
                 4_000_211_221u128..=5_000_000_000u128,
-                complete_range.clone()
+                complete_range
             ),
             vec![6, 7]
         );
@@ -703,7 +701,7 @@ mod tests {
             vec![0]
         );
         assert_eq!(
-            get_positions_for_value_range_helper(&decomp, 0..=105, complete_range.clone()),
+            get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
             vec![0]
         );
     }
@@ -756,11 +754,7 @@ mod tests {
         );
 
         assert_eq!(
-            get_positions_for_value_range_helper(
+            get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
-                &*decomp,
-                1_000_000..=1_000_000,
-                complete_range.clone()
-            ),
             vec![11]
         );
     }

fastfield_codecs/src/format_version.rs

@@ -1,7 +1,6 @@
 use std::io;
 
-use common::BinarySerializable;
+use common::{BinarySerializable, OwnedBytes};
-use ownedbytes::OwnedBytes;
 
 const MAGIC_NUMBER: u16 = 4335u16;
 const FASTFIELD_FORMAT_VERSION: u8 = 1;

fastfield_codecs/src/gcd.rs

@@ -45,7 +45,7 @@ mod tests {
     use std::io;
     use std::num::NonZeroU64;
 
-    use ownedbytes::OwnedBytes;
+    use common::OwnedBytes;
 
     use crate::gcd::{compute_gcd, find_gcd};
     use crate::{FastFieldCodecType, VecColumn};

fastfield_codecs/src/lib.rs

@@ -18,7 +18,7 @@ use std::io;
 use std::io::Write;
 use std::sync::Arc;
 
-use common::BinarySerializable;
+use common::{BinarySerializable, OwnedBytes};
 use compact_space::CompactSpaceDecompressor;
 use format_version::read_format_version;
 use monotonic_mapping::{
@@ -26,7 +26,6 @@ use monotonic_mapping::{
     StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
 };
 use null_index_footer::read_null_index_footer;
-use ownedbytes::OwnedBytes;
 use serialize::{Header, U128Header};
 
 mod bitpacked;
@@ -37,11 +36,13 @@ mod line;
 mod linear;
 mod monotonic_mapping;
 mod monotonic_mapping_u128;
+#[allow(dead_code)]
+mod null_index;
 mod null_index_footer;
 
 mod column;
 mod gcd;
-mod serialize;
+pub mod serialize;
 
 use self::bitpacked::BitpackedCodec;
 use self::blockwise_linear::BlockwiseLinearCodec;
@@ -434,7 +435,7 @@ mod tests {
 mod bench {
     use std::sync::Arc;
 
-    use ownedbytes::OwnedBytes;
+    use common::OwnedBytes;
     use rand::rngs::StdRng;
     use rand::{Rng, SeedableRng};
     use test::{self, Bencher};

fastfield_codecs/src/linear.rs

@@ -1,7 +1,6 @@
 use std::io::{self, Write};
 
-use common::BinarySerializable;
+use common::{BinarySerializable, OwnedBytes};
-use ownedbytes::OwnedBytes;
 use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
 
 use crate::line::Line;
@@ -25,13 +24,13 @@ impl Column for LinearReader {
         interpoled_val.wrapping_add(bitpacked_diff)
     }
 
-    #[inline]
+    #[inline(always)]
     fn min_value(&self) -> u64 {
         // The LinearReader assumes a normalized vector.
         0u64
     }
 
-    #[inline]
+    #[inline(always)]
     fn max_value(&self) -> u64 {
         self.header.max_value
     }

fastfield_codecs/src/main.rs

@@ -6,10 +6,10 @@ use std::io::BufRead;
 use std::net::{IpAddr, Ipv6Addr};
 use std::str::FromStr;
 
+use common::OwnedBytes;
 use fastfield_codecs::{open_u128, serialize_u128, Column, FastFieldCodecType, VecColumn};
 use itertools::Itertools;
 use measure_time::print_time;
-use ownedbytes::OwnedBytes;
 use prettytable::{Cell, Row, Table};
 
 fn print_set_stats(ip_addrs: &[u128]) {

fastfield_codecs/src/monotonic_mapping.rs

@@ -56,10 +56,12 @@ impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
 impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
 where T: StrictlyMonotonicFn<From, To>
 {
+    #[inline(always)]
     fn mapping(&self, val: To) -> From {
         self.orig_mapping.inverse(val)
     }
 
+    #[inline(always)]
     fn inverse(&self, val: From) -> To {
         self.orig_mapping.mapping(val)
     }
@@ -82,10 +84,12 @@ impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
     StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
 where T: MonotonicallyMappableToU128
 {
+    #[inline(always)]
     fn mapping(&self, inp: External) -> u128 {
         External::to_u128(inp)
     }
 
+    #[inline(always)]
     fn inverse(&self, out: u128) -> External {
         External::from_u128(out)
     }
@@ -95,10 +99,12 @@ impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
     StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
 where T: MonotonicallyMappableToU64
 {
+    #[inline(always)]
     fn mapping(&self, inp: External) -> u64 {
         External::to_u64(inp)
     }
 
+    #[inline(always)]
     fn inverse(&self, out: u64) -> External {
         External::from_u64(out)
     }
@@ -126,11 +132,13 @@ impl StrictlyMonotonicMappingToInternalGCDBaseval {
 impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
     for StrictlyMonotonicMappingToInternalGCDBaseval
 {
+    #[inline(always)]
     fn mapping(&self, inp: External) -> u64 {
         self.gcd_divider
             .divide(External::to_u64(inp) - self.min_value)
     }
 
+    #[inline(always)]
     fn inverse(&self, out: u64) -> External {
         External::from_u64(self.min_value + out * self.gcd)
     }
@@ -141,6 +149,7 @@ pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
     min_value: u64,
 }
 impl StrictlyMonotonicMappingToInternalBaseval {
+    #[inline(always)]
     pub(crate) fn new(min_value: u64) -> Self {
         Self { min_value }
     }
@@ -149,20 +158,24 @@ impl StrictlyMonotonicMappingToInternalBaseval {
 impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
     for StrictlyMonotonicMappingToInternalBaseval
 {
+    #[inline(always)]
     fn mapping(&self, val: External) -> u64 {
         External::to_u64(val) - self.min_value
     }
 
+    #[inline(always)]
     fn inverse(&self, val: u64) -> External {
         External::from_u64(self.min_value + val)
     }
 }
 
 impl MonotonicallyMappableToU64 for u64 {
+    #[inline(always)]
     fn to_u64(self) -> u64 {
         self
     }
 
+    #[inline(always)]
     fn from_u64(val: u64) -> Self {
         val
     }
@@ -192,11 +205,15 @@ impl MonotonicallyMappableToU64 for bool {
     }
 }
 
+// TODO remove me.
+// Tantivy should refuse NaN values and work with NotNaN internally.
 impl MonotonicallyMappableToU64 for f64 {
+    #[inline(always)]
     fn to_u64(self) -> u64 {
         common::f64_to_u64(self)
     }
 
+    #[inline(always)]
     fn from_u64(val: u64) -> Self {
         common::u64_to_f64(val)
     }

fastfield_codecs/src/null_index/dense.rs

@@ -0,0 +1,453 @@
+use std::convert::TryInto;
+use std::io::{self, Write};
+
+use common::{BinarySerializable, OwnedBytes};
+use itertools::Itertools;
+
+use super::{get_bit_at, set_bit_at};
+
+/// For the `DenseCodec`, `data` which contains the encoded blocks.
+/// Each block consists of [u8; 12]. The first 8 bytes is a bitvec for 64 elements.
+/// The last 4 bytes are the offset, the number of set bits so far.
+///
+/// When translating the original index to a dense index, the correct block can be computed
+/// directly `orig_idx/64`. Inside the block the position is `orig_idx%64`.
+///
+/// When translating a dense index to the original index, we can use the offset to find the correct
+/// block. Direct computation is not possible, but we can employ a linear or binary search.
+#[derive(Clone)]
+pub struct DenseCodec {
+    // data consists of blocks of 64 bits.
+    //
+    // The format is &[(u64, u32)]
+    // u64 is the bitvec
+    // u32 is the offset of the block, the number of set bits so far.
+    //
+    // At the end one block is appended, to store the number of values in the index in offset.
+    data: OwnedBytes,
+}
+const ELEMENTS_PER_BLOCK: u32 = 64;
+const BLOCK_BITVEC_SIZE: usize = 8;
+const BLOCK_OFFSET_SIZE: usize = 4;
+const SERIALIZED_BLOCK_SIZE: usize = BLOCK_BITVEC_SIZE + BLOCK_OFFSET_SIZE;
+
+#[inline]
+fn count_ones(bitvec: u64, pos_in_bitvec: u32) -> u32 {
+    if pos_in_bitvec == 63 {
+        bitvec.count_ones()
+    } else {
+        let mask = (1u64 << (pos_in_bitvec + 1)) - 1;
+        let masked_bitvec = bitvec & mask;
+        masked_bitvec.count_ones()
+    }
+}
+
+#[derive(Clone, Copy)]
+struct DenseIndexBlock {
+    bitvec: u64,
+    offset: u32,
+}
+
+impl From<[u8; SERIALIZED_BLOCK_SIZE]> for DenseIndexBlock {
+    fn from(data: [u8; SERIALIZED_BLOCK_SIZE]) -> Self {
+        let bitvec = u64::from_le_bytes(data[..BLOCK_BITVEC_SIZE].try_into().unwrap());
+        let offset = u32::from_le_bytes(data[BLOCK_BITVEC_SIZE..].try_into().unwrap());
+        Self { bitvec, offset }
+    }
+}
+
+impl DenseCodec {
+    /// Open the DenseCodec from OwnedBytes
+    pub fn open(data: OwnedBytes) -> Self {
+        Self { data }
+    }
+    #[inline]
+    /// Check if value at position is not null.
+    pub fn exists(&self, idx: u32) -> bool {
+        let block_pos = idx / ELEMENTS_PER_BLOCK;
+        let bitvec = self.dense_index_block(block_pos).bitvec;
+
+        let pos_in_bitvec = idx % ELEMENTS_PER_BLOCK;
+
+        get_bit_at(bitvec, pos_in_bitvec)
+    }
+    #[inline]
+    fn dense_index_block(&self, block_pos: u32) -> DenseIndexBlock {
+        dense_index_block(&self.data, block_pos)
+    }
+
+    /// Return the number of non-null values in an index
+    pub fn num_non_nulls(&self) -> u32 {
+        let last_block = (self.data.len() / SERIALIZED_BLOCK_SIZE) - 1;
+        self.dense_index_block(last_block as u32).offset
+    }
+
+    #[inline]
+    /// Translate from the original index to the codec index.
+    pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
+        let block_pos = idx / ELEMENTS_PER_BLOCK;
+        let index_block = self.dense_index_block(block_pos);
+        let pos_in_block_bit_vec = idx % ELEMENTS_PER_BLOCK;
+        let ones_in_block = count_ones(index_block.bitvec, pos_in_block_bit_vec);
+        if get_bit_at(index_block.bitvec, pos_in_block_bit_vec) {
+            // -1 is ok, since idx does exist, so there's at least one
+            Some(index_block.offset + ones_in_block - 1)
+        } else {
+            None
+        }
+    }
+
+    /// Translate positions from the codec index to the original index.
+    ///
+    /// # Panics
+    ///
+    /// May panic if any `idx` is greater than the max codec index.
+    pub fn translate_codec_idx_to_original_idx<'a>(
+        &'a self,
+        iter: impl Iterator<Item = u32> + 'a,
+    ) -> impl Iterator<Item = u32> + 'a {
+        let mut block_pos = 0u32;
+        iter.map(move |dense_idx| {
+            // update block_pos to limit search scope
+            block_pos = find_block(dense_idx, block_pos, &self.data);
+            let index_block = self.dense_index_block(block_pos);
+
+            // The next offset is higher than dense_idx and therefore:
+            // dense_idx <= offset + num_set_bits in block
+            let mut num_set_bits = 0;
+            for idx_in_bitvec in 0..ELEMENTS_PER_BLOCK {
+                if get_bit_at(index_block.bitvec, idx_in_bitvec) {
+                    num_set_bits += 1;
+                }
+                if num_set_bits == (dense_idx - index_block.offset + 1) {
+                    let orig_idx = block_pos * ELEMENTS_PER_BLOCK + idx_in_bitvec;
+                    return orig_idx;
+                }
+            }
+            panic!("Internal Error: Offset calculation in dense idx seems to be wrong.");
+        })
+    }
+}
+
+#[inline]
+fn dense_index_block(data: &[u8], block_pos: u32) -> DenseIndexBlock {
+    let data_start_pos = block_pos as usize * SERIALIZED_BLOCK_SIZE;
+    let block_data: [u8; SERIALIZED_BLOCK_SIZE] = data[data_start_pos..][..SERIALIZED_BLOCK_SIZE]
+        .try_into()
+        .unwrap();
+    block_data.into()
+}
+
+#[inline]
+/// Finds the block position containing the dense_idx.
+///
+/// # Correctness
+/// dense_idx needs to be smaller than the number of values in the index
+///
+/// The last offset number is equal to the number of values in the index.
+fn find_block(dense_idx: u32, mut block_pos: u32, data: &[u8]) -> u32 {
+    loop {
+        let offset = dense_index_block(data, block_pos).offset;
+        if offset > dense_idx {
+            return block_pos - 1;
+        }
+        block_pos += 1;
+    }
+}
+
+/// Iterator over all values, true if set, otherwise false
+pub fn serialize_dense_codec(
+    iter: impl Iterator<Item = bool>,
+    mut out: impl Write,
+) -> io::Result<()> {
+    let mut offset: u32 = 0;
+
+    for chunk in &iter.chunks(ELEMENTS_PER_BLOCK as usize) {
+        let mut block: u64 = 0;
+        for (pos, is_bit_set) in chunk.enumerate() {
+            if is_bit_set {
+                set_bit_at(&mut block, pos as u64);
+            }
+        }
+
+        block.serialize(&mut out)?;
+        offset.serialize(&mut out)?;
+
+        offset += block.count_ones();
+    }
+    // Add sentinal block for the offset
+    let block: u64 = 0;
+    block.serialize(&mut out)?;
+    offset.serialize(&mut out)?;
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use proptest::prelude::{any, prop, *};
+    use proptest::strategy::Strategy;
+    use proptest::{prop_oneof, proptest};
+
+    use super::*;
+
+    fn random_bitvec() -> BoxedStrategy<Vec<bool>> {
+        prop_oneof![
+            1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..100),
+            1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..64),
+            1 => prop::collection::vec(proptest::bool::weighted(0.0), 0..100),
+            1 => prop::collection::vec(proptest::bool::weighted(0.0), 0..64),
+            8 => vec![any::<bool>()],
+            2 => prop::collection::vec(any::<bool>(), 0..50),
+        ]
+        .boxed()
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(500))]
+        #[test]
+        fn test_with_random_bitvecs(bitvec1 in random_bitvec(), bitvec2 in random_bitvec(), bitvec3 in random_bitvec()) {
+            let mut bitvec = Vec::new();
+            bitvec.extend_from_slice(&bitvec1);
+            bitvec.extend_from_slice(&bitvec2);
+            bitvec.extend_from_slice(&bitvec3);
+            test_null_index(bitvec);
+        }
+    }
+
+    #[test]
+    fn dense_codec_test_one_block_false() {
+        let mut iter = vec![false; 64];
+        iter.push(true);
+        test_null_index(iter);
+    }
+
+    fn test_null_index(data: Vec<bool>) {
+        let mut out = vec![];
+
+        serialize_dense_codec(data.iter().cloned(), &mut out).unwrap();
+        let null_index = DenseCodec::open(OwnedBytes::new(out));
+
+        let orig_idx_with_value: Vec<u32> = data
+            .iter()
+            .enumerate()
+            .filter(|(_pos, val)| **val)
+            .map(|(pos, _val)| pos as u32)
+            .collect();
+
+        assert_eq!(
+            null_index
+                .translate_codec_idx_to_original_idx(0..orig_idx_with_value.len() as u32)
+                .collect_vec(),
+            orig_idx_with_value
+        );
+
+        for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate() {
+            assert_eq!(
+                null_index.translate_to_codec_idx(*orig_idx),
+                Some(dense_idx as u32)
+            );
+        }
+
+        for (pos, value) in data.iter().enumerate() {
+            assert_eq!(null_index.exists(pos as u32), *value);
+        }
+    }
+
+    #[test]
+    fn dense_codec_test_translation() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_dense_codec(iter, &mut out).unwrap();
+        let null_index = DenseCodec::open(OwnedBytes::new(out));
+
+        assert_eq!(
+            null_index
+                .translate_codec_idx_to_original_idx(0..2)
+                .collect_vec(),
+            vec![0, 2]
+        );
+    }
+
+    #[test]
+    fn dense_codec_translate() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_dense_codec(iter, &mut out).unwrap();
+        let null_index = DenseCodec::open(OwnedBytes::new(out));
+        assert_eq!(null_index.translate_to_codec_idx(0), Some(0));
+        assert_eq!(null_index.translate_to_codec_idx(2), Some(1));
+    }
+
+    #[test]
+    fn dense_codec_test_small() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_dense_codec(iter, &mut out).unwrap();
+        let null_index = DenseCodec::open(OwnedBytes::new(out));
+        assert!(null_index.exists(0));
+        assert!(!null_index.exists(1));
+        assert!(null_index.exists(2));
+        assert!(!null_index.exists(3));
+    }
+
+    #[test]
+    fn dense_codec_test_large() {
+        let mut docs = vec![];
+        docs.extend((0..1000).map(|_idx| false));
+        docs.extend((0..=1000).map(|_idx| true));
+
+        let iter = docs.iter().cloned();
+        let mut out = vec![];
+        serialize_dense_codec(iter, &mut out).unwrap();
+        let null_index = DenseCodec::open(OwnedBytes::new(out));
+        assert!(!null_index.exists(0));
+        assert!(!null_index.exists(100));
+        assert!(!null_index.exists(999));
+        assert!(null_index.exists(1000));
+        assert!(null_index.exists(1999));
+        assert!(null_index.exists(2000));
+        assert!(!null_index.exists(2001));
+    }
+
+    #[test]
+    fn test_count_ones() {
+        let mut block = 0;
+        set_bit_at(&mut block, 0);
+        set_bit_at(&mut block, 2);
+
+        assert_eq!(count_ones(block, 0), 1);
+        assert_eq!(count_ones(block, 1), 1);
+        assert_eq!(count_ones(block, 2), 2);
+    }
+}
+
+#[cfg(all(test, feature = "unstable"))]
+mod bench {
+
+    use rand::rngs::StdRng;
+    use rand::{Rng, SeedableRng};
+    use test::Bencher;
+
+    use super::*;
+
+    const TOTAL_NUM_VALUES: u32 = 1_000_000;
+    fn gen_bools(fill_ratio: f64) -> DenseCodec {
+        let mut out = Vec::new();
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let bools: Vec<_> = (0..TOTAL_NUM_VALUES)
+            .map(|_| rng.gen_bool(fill_ratio))
+            .collect();
+        serialize_dense_codec(bools.into_iter(), &mut out).unwrap();
+
+        let codec = DenseCodec::open(OwnedBytes::new(out));
+        codec
+    }
+
+    fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let mut current = start;
+        std::iter::from_fn(move || {
+            current += rng.gen_range(1..step_size + 1);
+            if current >= end {
+                None
+            } else {
+                Some(current)
+            }
+        })
+    }
+
+    fn walk_over_data(codec: &DenseCodec, max_step_size: u32) -> Option<u32> {
+        walk_over_data_from_positions(
+            codec,
+            random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
+        )
+    }
+
+    fn walk_over_data_from_positions(
+        codec: &DenseCodec,
+        positions: impl Iterator<Item = u32>,
+    ) -> Option<u32> {
+        let mut dense_idx: Option<u32> = None;
+        for idx in positions {
+            dense_idx = dense_idx.or(codec.translate_to_codec_idx(idx));
+        }
+        dense_idx
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_orig_to_codec_90percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_orig_to_codec_50percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.5f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_orig_to_codec_full_scan_10percent(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_orig_to_codec_full_scan_90percent(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_orig_to_codec_10percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_codec_to_orig_90percent_filled_random_stride_big_step(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.9f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_codec_to_orig_90percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.9f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_dense_codec_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(0..num_vals)
+                .last()
+        });
+    }
+}

fastfield_codecs/src/null_index/mod.rs

@@ -0,0 +1,14 @@
+pub use dense::{serialize_dense_codec, DenseCodec};
+
+mod dense;
+mod sparse;
+
+#[inline]
+fn get_bit_at(input: u64, n: u32) -> bool {
+    input & (1 << n) != 0
+}
+
+#[inline]
+fn set_bit_at(input: &mut u64, n: u64) {
+    *input |= 1 << n;
+}

fastfield_codecs/src/null_index/sparse.rs

@@ -0,0 +1,752 @@
+use std::io::{self, Write};
+
+use common::{BitSet, OwnedBytes};
+
+use super::{serialize_dense_codec, DenseCodec};
+
+/// `SparseCodec` is the codec for data, when only few documents have values.
+/// In contrast to `DenseCodec` opening a `SparseCodec` causes runtime data to be produced, for
+/// faster access.
+///
+/// The lower 16 bits of doc ids are stored as u16 while the upper 16 bits are given by the block
+/// id. Each block contains 1<<16 docids.
+///
+/// # Serialized Data Layout
+/// The data starts with the block data. Each block is either dense or sparse encoded, depending on
+/// the number of values in the block. A block is sparse when it contains less than
+/// DENSE_BLOCK_THRESHOLD (6144) values.
+/// [Sparse data block | dense data block, .. #repeat*; Desc: Either a sparse or dense encoded
+/// block]
+/// ### Sparse block data
+/// [u16 LE, .. #repeat*; Desc: Positions with values in a block]
+/// ### Dense block data
+/// [Dense codec for the whole block; Desc: Similar to a bitvec(0..ELEMENTS_PER_BLOCK) + Metadata
+/// for faster lookups. See dense.rs]
+///
+/// The data is followed by block metadata, to know which area of the raw block data belongs to
+/// which block. Only metadata for blocks with elements is recorded to
+/// keep the overhead low for scenarios with many very sparse columns. The block metadata consists
+/// of the block index and the number of values in the block. Since we don't store empty blocks
+/// num_vals is incremented by 1, e.g. 0 means 1 value.
+///
+/// The last u16 is storing the number of metadata blocks.
+/// [u16 LE, .. #repeat*; Desc: Positions with values in a block][(u16 LE, u16 LE), .. #repeat*;
+/// Desc: (Block Id u16, Num Elements u16)][u16 LE; Desc: num blocks with values u16]
+///
+/// # Opening
+/// When opening the data layout, the data is expanded to `Vec<SparseCodecBlockVariant>`, where the
+/// index is the block index. For each block `byte_start` and `offset` is computed.
+pub struct SparseCodec {
+    data: OwnedBytes,
+    blocks: Vec<SparseCodecBlockVariant>,
+}
+
+/// The threshold for for number of elements after which we switch to dense block encoding
+const DENSE_BLOCK_THRESHOLD: u32 = 6144;
+
+const ELEMENTS_PER_BLOCK: u32 = u16::MAX as u32 + 1;
+
+/// 1.5 bit per Element + 12 bytes for the sentinal block
+const NUM_BYTES_DENSE_BLOCK: u32 = (ELEMENTS_PER_BLOCK + ELEMENTS_PER_BLOCK / 2 + 64 + 32) / 8;
+
+#[derive(Clone)]
+enum SparseCodecBlockVariant {
+    Empty { offset: u32 },
+    Dense(DenseBlock),
+    Sparse(SparseBlock),
+}
+
+impl SparseCodecBlockVariant {
+    /// The number of non-null values that preceeded that block.
+    #[inline]
+    fn offset(&self) -> u32 {
+        match self {
+            SparseCodecBlockVariant::Empty { offset } => *offset,
+            SparseCodecBlockVariant::Dense(dense) => dense.offset,
+            SparseCodecBlockVariant::Sparse(sparse) => sparse.offset,
+        }
+    }
+}
+
+/// A block consists of max u16 values
+#[derive(Clone)]
+struct DenseBlock {
+    /// The number of values set before the block
+    offset: u32,
+    /// The data for the dense encoding
+    codec: DenseCodec,
+}
+
+impl DenseBlock {
+    pub fn exists(&self, idx: u32) -> bool {
+        self.codec.exists(idx)
+    }
+    pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
+        self.codec.translate_to_codec_idx(idx)
+    }
+    pub fn translate_codec_idx_to_original_idx(&self, idx: u32) -> u32 {
+        self.codec
+            .translate_codec_idx_to_original_idx(idx..=idx)
+            .next()
+            .unwrap()
+    }
+}
+
+/// A block consists of max u16 values
+#[derive(Debug, Copy, Clone)]
+struct SparseBlock {
+    /// The number of values in the block
+    num_vals: u32,
+    /// The number of values set before the block
+    offset: u32,
+    /// The start position of the data for the block
+    byte_start: u32,
+}
+
+impl SparseBlock {
+    fn empty_block(offset: u32) -> Self {
+        Self {
+            num_vals: 0,
+            byte_start: 0,
+            offset,
+        }
+    }
+
+    #[inline]
+    fn value_at_idx(&self, data: &[u8], idx: u16) -> u16 {
+        let start_offset: usize = self.byte_start as usize + (idx as u32 as usize * 2);
+        get_u16(data, start_offset)
+    }
+
+    #[inline]
+    #[allow(clippy::comparison_chain)]
+    // Looks for the element in the block. Returns the positions if found.
+    fn binary_search(&self, data: &[u8], target: u16) -> Option<u16> {
+        let mut size = self.num_vals as u16;
+        let mut left = 0;
+        let mut right = size;
+        // TODO try different implem.
+        //  e.g. exponential search into binary search
+        while left < right {
+            let mid = left + size / 2;
+
+            // TODO do boundary check only once, and then use an
+            // unsafe `value_at_idx`
+            let mid_val = self.value_at_idx(data, mid);
+
+            if target > mid_val {
+                left = mid + 1;
+            } else if target < mid_val {
+                right = mid;
+            } else {
+                return Some(mid);
+            }
+
+            size = right - left;
+        }
+        None
+    }
+}
+
+#[inline]
+fn get_u16(data: &[u8], byte_position: usize) -> u16 {
+    let bytes: [u8; 2] = data[byte_position..byte_position + 2].try_into().unwrap();
+    u16::from_le_bytes(bytes)
+}
+
+const SERIALIZED_BLOCK_METADATA_SIZE: usize = 4;
+
+fn deserialize_sparse_codec_block(data: &OwnedBytes) -> Vec<SparseCodecBlockVariant> {
+    // The number of vals so far
+    let mut offset = 0;
+    let mut sparse_codec_blocks = Vec::new();
+    let num_blocks = get_u16(data, data.len() - 2);
+    let block_data_index_start =
+        data.len() - 2 - num_blocks as usize * SERIALIZED_BLOCK_METADATA_SIZE;
+    let mut byte_start = 0;
+    for block_num in 0..num_blocks as usize {
+        let block_data_index = block_data_index_start + SERIALIZED_BLOCK_METADATA_SIZE * block_num;
+        let block_idx = get_u16(data, block_data_index);
+        let num_vals = get_u16(data, block_data_index + 2) as u32 + 1;
+        sparse_codec_blocks.resize(
+            block_idx as usize,
+            SparseCodecBlockVariant::Empty { offset },
+        );
+
+        if is_sparse(num_vals) {
+            let block = SparseBlock {
+                num_vals,
+                offset,
+                byte_start,
+            };
+            sparse_codec_blocks.push(SparseCodecBlockVariant::Sparse(block));
+            byte_start += 2 * num_vals;
+        } else {
+            let block = DenseBlock {
+                offset,
+                codec: DenseCodec::open(data.slice(byte_start as usize..data.len()).clone()),
+            };
+            sparse_codec_blocks.push(SparseCodecBlockVariant::Dense(block));
+            // Dense blocks have a fixed size spanning ELEMENTS_PER_BLOCK.
+            byte_start += NUM_BYTES_DENSE_BLOCK;
+        }
+
+        offset += num_vals;
+    }
+    sparse_codec_blocks.push(SparseCodecBlockVariant::Empty { offset });
+    sparse_codec_blocks
+}
+
+/// Splits a value address into lower and upper 16bits.
+/// The lower 16 bits are the value in the block
+/// The upper 16 bits are the block index
+#[derive(Debug, Clone, Copy)]
+struct ValueAddr {
+    block_idx: u16,
+    value_in_block: u16,
+}
+
+/// Splits a idx into block index and value in the block
+fn value_addr(idx: u32) -> ValueAddr {
+    /// Static assert number elements per block this method expects
+    #[allow(clippy::assertions_on_constants)]
+    const _: () = assert!(ELEMENTS_PER_BLOCK == (1 << 16));
+
+    let value_in_block = idx as u16;
+    let block_idx = (idx >> 16) as u16;
+    ValueAddr {
+        block_idx,
+        value_in_block,
+    }
+}
+
+impl SparseCodec {
+    /// Open the SparseCodec from OwnedBytes
+    pub fn open(data: OwnedBytes) -> Self {
+        let blocks = deserialize_sparse_codec_block(&data);
+        Self { data, blocks }
+    }
+
+    #[inline]
+    /// Check if value at position is not null.
+    pub fn exists(&self, idx: u32) -> bool {
+        let value_addr = value_addr(idx);
+        // There may be trailing nulls without data, those are not stored as blocks. It would be
+        // possible to create empty blocks, but for that we would need to serialize the number of
+        // values or pass them when opening
+
+        if let Some(block) = self.blocks.get(value_addr.block_idx as usize) {
+            match block {
+                SparseCodecBlockVariant::Empty { offset: _ } => false,
+                SparseCodecBlockVariant::Dense(block) => {
+                    block.exists(value_addr.value_in_block as u32)
+                }
+                SparseCodecBlockVariant::Sparse(block) => block
+                    .binary_search(&self.data, value_addr.value_in_block)
+                    .is_some(),
+            }
+        } else {
+            false
+        }
+    }
+
+    /// Return the number of non-null values in an index
+    pub fn num_non_nulls(&self) -> u32 {
+        self.blocks.last().map(|block| block.offset()).unwrap_or(0)
+    }
+
+    #[inline]
+    /// Translate from the original index to the codec index.
+    pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
+        let value_addr = value_addr(idx);
+        let block = self.blocks.get(value_addr.block_idx as usize)?;
+
+        match block {
+            SparseCodecBlockVariant::Empty { offset: _ } => None,
+            SparseCodecBlockVariant::Dense(block) => block
+                .translate_to_codec_idx(value_addr.value_in_block as u32)
+                .map(|pos_in_block| pos_in_block + block.offset),
+            SparseCodecBlockVariant::Sparse(block) => {
+                let pos_in_block = block.binary_search(&self.data, value_addr.value_in_block);
+                pos_in_block.map(|pos_in_block: u16| block.offset + pos_in_block as u32)
+            }
+        }
+    }
+
+    fn find_block(&self, dense_idx: u32, mut block_pos: u32) -> u32 {
+        loop {
+            let offset = self.blocks[block_pos as usize].offset();
+            if offset > dense_idx {
+                return block_pos - 1;
+            }
+            block_pos += 1;
+        }
+    }
+
+    /// Translate positions from the codec index to the original index.
+    ///
+    /// # Panics
+    ///
+    /// May panic if any `idx` is greater than the max codec index.
+    pub fn translate_codec_idx_to_original_idx<'a>(
+        &'a self,
+        iter: impl Iterator<Item = u32> + 'a,
+    ) -> impl Iterator<Item = u32> + 'a {
+        // TODO: There's a big potential performance gain, by using iterators per block instead of
+        // random access for each element in a block
+        // group_by itertools won't help though, since it requires a temporary local variable
+        let mut block_pos = 0u32;
+        iter.map(move |codec_idx| {
+            // update block_pos to limit search scope
+            block_pos = self.find_block(codec_idx, block_pos);
+            let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
+            let block = &self.blocks[block_pos as usize];
+            let idx_in_block = codec_idx - block.offset();
+            match block {
+                SparseCodecBlockVariant::Empty { offset: _ } => {
+                    panic!(
+                        "invalid input, cannot translate to original index. associated empty \
+                         block with dense idx. block_pos {}, idx_in_block {}",
+                        block_pos, idx_in_block
+                    )
+                }
+                SparseCodecBlockVariant::Dense(dense) => {
+                    dense.translate_codec_idx_to_original_idx(idx_in_block) + block_doc_idx_start
+                }
+                SparseCodecBlockVariant::Sparse(block) => {
+                    block.value_at_idx(&self.data, idx_in_block as u16) as u32 + block_doc_idx_start
+                }
+            }
+        })
+    }
+}
+
+fn is_sparse(num_elem_in_block: u32) -> bool {
+    num_elem_in_block < DENSE_BLOCK_THRESHOLD
+}
+
+#[derive(Default)]
+struct BlockDataSerialized {
+    block_idx: u16,
+    num_vals: u32,
+}
+
+/// Iterator over positions of set values.
+pub fn serialize_sparse_codec<W: Write>(
+    mut iter: impl Iterator<Item = u32>,
+    mut out: W,
+) -> io::Result<()> {
+    let mut block_metadata: Vec<BlockDataSerialized> = Vec::new();
+    let mut current_block = Vec::new();
+    // This if-statement for the first element ensures that
+    // `block_metadata` is not empty in the loop below.
+    if let Some(idx) = iter.next() {
+        let value_addr = value_addr(idx);
+        block_metadata.push(BlockDataSerialized {
+            block_idx: value_addr.block_idx,
+            num_vals: 1,
+        });
+        current_block.push(value_addr.value_in_block);
+    }
+    let flush_block = |current_block: &mut Vec<u16>, out: &mut W| -> io::Result<()> {
+        let is_sparse = is_sparse(current_block.len() as u32);
+        if is_sparse {
+            for val_in_block in current_block.iter() {
+                out.write_all(val_in_block.to_le_bytes().as_ref())?;
+            }
+        } else {
+            let mut bitset = BitSet::with_max_value(ELEMENTS_PER_BLOCK + 1);
+            for val_in_block in current_block.iter() {
+                bitset.insert(*val_in_block as u32);
+            }
+
+            let iter = (0..ELEMENTS_PER_BLOCK).map(|idx| bitset.contains(idx));
+            serialize_dense_codec(iter, out)?;
+        }
+        current_block.clear();
+        Ok(())
+    };
+    for idx in iter {
+        let value_addr = value_addr(idx);
+        if block_metadata[block_metadata.len() - 1].block_idx == value_addr.block_idx {
+            let last_idx_metadata = block_metadata.len() - 1;
+            block_metadata[last_idx_metadata].num_vals += 1;
+        } else {
+            // flush prev block
+            flush_block(&mut current_block, &mut out)?;
+
+            block_metadata.push(BlockDataSerialized {
+                block_idx: value_addr.block_idx,
+                num_vals: 1,
+            });
+        }
+        current_block.push(value_addr.value_in_block);
+    }
+    // handle last block
+    flush_block(&mut current_block, &mut out)?;
+
+    for block in &block_metadata {
+        out.write_all(block.block_idx.to_le_bytes().as_ref())?;
+        // We don't store empty blocks, therefore we can subtract 1.
+        // This way we will be able to use u16 when the number of elements is 1 << 16 or u16::MAX+1
+        out.write_all(((block.num_vals - 1) as u16).to_le_bytes().as_ref())?;
+    }
+    out.write_all((block_metadata.len() as u16).to_le_bytes().as_ref())?;
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use itertools::Itertools;
+    use proptest::prelude::{any, prop, *};
+    use proptest::strategy::Strategy;
+    use proptest::{prop_oneof, proptest};
+
+    use super::*;
+
+    fn random_bitvec() -> BoxedStrategy<Vec<bool>> {
+        prop_oneof![
+            1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..100),
+            1 => prop::collection::vec(proptest::bool::weighted(0.00), 0..(ELEMENTS_PER_BLOCK as usize * 3)), // empty blocks
+            1 => prop::collection::vec(proptest::bool::weighted(1.00), 0..(ELEMENTS_PER_BLOCK as usize + 10)), // full block
+            1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..100),
+            1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..u16::MAX as usize),
+            8 => vec![any::<bool>()],
+        ]
+        .boxed()
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(50))]
+        #[test]
+        fn test_with_random_bitvecs(bitvec1 in random_bitvec(), bitvec2 in random_bitvec(), bitvec3 in random_bitvec()) {
+            let mut bitvec = Vec::new();
+            bitvec.extend_from_slice(&bitvec1);
+            bitvec.extend_from_slice(&bitvec2);
+            bitvec.extend_from_slice(&bitvec3);
+            test_null_index(bitvec);
+        }
+    }
+
+    #[test]
+    fn sparse_codec_test_one_block_false() {
+        let mut iter = vec![false; ELEMENTS_PER_BLOCK as usize];
+        iter.push(true);
+        test_null_index(iter);
+    }
+
+    #[test]
+    fn sparse_codec_test_one_block_true() {
+        let mut iter = vec![true; ELEMENTS_PER_BLOCK as usize];
+        iter.push(true);
+        test_null_index(iter);
+    }
+
+    fn test_null_index(data: Vec<bool>) {
+        let mut out = vec![];
+
+        serialize_sparse_codec(
+            data.iter()
+                .cloned()
+                .enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+        let null_index = SparseCodec::open(OwnedBytes::new(out));
+
+        let orig_idx_with_value: Vec<u32> = data
+            .iter()
+            .enumerate()
+            .filter(|(_pos, val)| **val)
+            .map(|(pos, _val)| pos as u32)
+            .collect();
+
+        assert_eq!(
+            null_index
+                .translate_codec_idx_to_original_idx(0..orig_idx_with_value.len() as u32)
+                .collect_vec(),
+            orig_idx_with_value
+        );
+
+        let step_size = (orig_idx_with_value.len() / 100).max(1);
+        for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate().step_by(step_size) {
+            assert_eq!(
+                null_index.translate_to_codec_idx(*orig_idx),
+                Some(dense_idx as u32)
+            );
+        }
+
+        // 100 samples
+        let step_size = (data.len() / 100).max(1);
+        for (pos, value) in data.iter().enumerate().step_by(step_size) {
+            assert_eq!(null_index.exists(pos as u32), *value);
+        }
+    }
+
+    #[test]
+    fn sparse_codec_test_translation() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_sparse_codec(
+            iter.enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+        let null_index = SparseCodec::open(OwnedBytes::new(out));
+
+        assert_eq!(
+            null_index
+                .translate_codec_idx_to_original_idx(0..2)
+                .collect_vec(),
+            vec![0, 2]
+        );
+    }
+
+    #[test]
+    fn sparse_codec_translate() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_sparse_codec(
+            iter.enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+        let null_index = SparseCodec::open(OwnedBytes::new(out));
+        assert_eq!(null_index.translate_to_codec_idx(0), Some(0));
+        assert_eq!(null_index.translate_to_codec_idx(2), Some(1));
+    }
+
+    #[test]
+    fn sparse_codec_test_small() {
+        let mut out = vec![];
+
+        let iter = ([true, false, true, false]).iter().cloned();
+        serialize_sparse_codec(
+            iter.enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+        let null_index = SparseCodec::open(OwnedBytes::new(out));
+        assert!(null_index.exists(0));
+        assert!(!null_index.exists(1));
+        assert!(null_index.exists(2));
+        assert!(!null_index.exists(3));
+    }
+
+    #[test]
+    fn sparse_codec_test_large() {
+        let mut docs = vec![];
+        docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
+        docs.extend((0..=1).map(|_idx| true));
+
+        let iter = docs.iter().cloned();
+        let mut out = vec![];
+        serialize_sparse_codec(
+            iter.enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+        let null_index = SparseCodec::open(OwnedBytes::new(out));
+        assert!(!null_index.exists(0));
+        assert!(!null_index.exists(100));
+        assert!(!null_index.exists(ELEMENTS_PER_BLOCK - 1));
+        assert!(null_index.exists(ELEMENTS_PER_BLOCK));
+        assert!(null_index.exists(ELEMENTS_PER_BLOCK + 1));
+    }
+}
+
+#[cfg(all(test, feature = "unstable"))]
+mod bench {
+
+    use rand::rngs::StdRng;
+    use rand::{Rng, SeedableRng};
+    use test::Bencher;
+
+    use super::*;
+
+    const TOTAL_NUM_VALUES: u32 = 1_000_000;
+    fn gen_bools(fill_ratio: f64) -> SparseCodec {
+        let mut out = Vec::new();
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        serialize_sparse_codec(
+            (0..TOTAL_NUM_VALUES)
+                .map(|_| rng.gen_bool(fill_ratio))
+                .enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+            &mut out,
+        )
+        .unwrap();
+
+        let codec = SparseCodec::open(OwnedBytes::new(out));
+        codec
+    }
+
+    fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let mut current = start;
+        std::iter::from_fn(move || {
+            current += rng.gen_range(1..step_size + 1);
+            if current >= end {
+                None
+            } else {
+                Some(current)
+            }
+        })
+    }
+
+    fn walk_over_data(codec: &SparseCodec, max_step_size: u32) -> Option<u32> {
+        walk_over_data_from_positions(
+            codec,
+            random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
+        )
+    }
+
+    fn walk_over_data_from_positions(
+        codec: &SparseCodec,
+        positions: impl Iterator<Item = u32>,
+    ) -> Option<u32> {
+        let mut dense_idx: Option<u32> = None;
+        for idx in positions {
+            dense_idx = dense_idx.or(codec.translate_to_codec_idx(idx));
+        }
+        dense_idx
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_1percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_5percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_full_scan_10percent(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_full_scan_90percent(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_full_scan_1percent(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_10percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_orig_to_codec_90percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_random_stride_big_step(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.01f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.01f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(0..num_vals)
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_random_stride_big_step(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.90f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_random_stride(
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(0.9f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        let num_vals = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(0..num_vals)
+                .last()
+        });
+    }
+}

fastfield_codecs/src/null_index_footer.rs

@@ -1,12 +1,12 @@
 use std::io::{self, Write};
 use std::ops::Range;
 
-use common::{BinarySerializable, CountingWriter, VInt};
+use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt};
-use ownedbytes::OwnedBytes;
 
 #[derive(Debug, Clone, Copy, Eq, PartialEq)]
 pub(crate) enum FastFieldCardinality {
     Single = 1,
+    Multi = 2,
 }
 
 impl BinarySerializable for FastFieldCardinality {
@@ -30,6 +30,7 @@ impl FastFieldCardinality {
     pub(crate) fn from_code(code: u8) -> Option<Self> {
         match code {
             1 => Some(Self::Single),
+            2 => Some(Self::Multi),
             _ => None,
         }
     }

fastfield_codecs/src/serialize.rs

@@ -21,9 +21,8 @@ use std::io;
 use std::num::NonZeroU64;
 use std::sync::Arc;
 
-use common::{BinarySerializable, VInt};
+use common::{BinarySerializable, OwnedBytes, VInt};
 use log::warn;
-use ownedbytes::OwnedBytes;
 
 use crate::bitpacked::BitpackedCodec;
 use crate::blockwise_linear::BlockwiseLinearCodec;
@@ -193,6 +192,69 @@ pub fn serialize_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
     iter_gen: F,
     num_vals: u32,
     output: &mut impl io::Write,
+) -> io::Result<()> {
+    serialize_u128_new(ValueIndexInfo::default(), iter_gen, num_vals, output)
+}
+
+#[allow(dead_code)]
+pub enum ValueIndexInfo<'a> {
+    MultiValue(Box<dyn MultiValueIndexInfo + 'a>),
+    SingleValue(Box<dyn SingleValueIndexInfo + 'a>),
+}
+
+// TODO Remove me
+impl Default for ValueIndexInfo<'static> {
+    fn default() -> Self {
+        struct Dummy {}
+        impl SingleValueIndexInfo for Dummy {
+            fn num_vals(&self) -> u32 {
+                todo!()
+            }
+            fn num_non_nulls(&self) -> u32 {
+                todo!()
+            }
+            fn iter(&self) -> Box<dyn Iterator<Item = u32>> {
+                todo!()
+            }
+        }
+
+        Self::SingleValue(Box::new(Dummy {}))
+    }
+}
+
+impl<'a> ValueIndexInfo<'a> {
+    fn get_cardinality(&self) -> FastFieldCardinality {
+        match self {
+            ValueIndexInfo::MultiValue(_) => FastFieldCardinality::Multi,
+            ValueIndexInfo::SingleValue(_) => FastFieldCardinality::Single,
+        }
+    }
+}
+
+pub trait MultiValueIndexInfo {
+    /// The number of docs in the column.
+    fn num_docs(&self) -> u32;
+    /// The number of values in the column.
+    fn num_vals(&self) -> u32;
+    /// Return the start index of the values for each doc
+    fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_>;
+}
+
+pub trait SingleValueIndexInfo {
+    /// The number of values including nulls in the column.
+    fn num_vals(&self) -> u32;
+    /// The number of non-null values in the column.
+    fn num_non_nulls(&self) -> u32;
+    /// Return a iterator of the positions of docs with a value
+    fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_>;
+}
+
+/// Serializes u128 values with the compact space codec.
+pub fn serialize_u128_new<F: Fn() -> I, I: Iterator<Item = u128>>(
+    value_index: ValueIndexInfo,
+    iter_gen: F,
+    num_vals: u32,
+    output: &mut impl io::Write,
 ) -> io::Result<()> {
     let header = U128Header {
         num_vals,
@@ -203,7 +265,7 @@ pub fn serialize_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
     compressor.compress_into(iter_gen(), output).unwrap();
 
     let null_index_footer = NullIndexFooter {
-        cardinality: FastFieldCardinality::Single,
+        cardinality: value_index.get_cardinality(),
         null_index_codec: NullIndexCodec::Full,
         null_index_byte_range: 0..0,
     };
@@ -218,6 +280,16 @@ pub fn serialize<T: MonotonicallyMappableToU64>(
     typed_column: impl Column<T>,
     output: &mut impl io::Write,
     codecs: &[FastFieldCodecType],
+) -> io::Result<()> {
+    serialize_new(ValueIndexInfo::default(), typed_column, output, codecs)
+}
+
+/// Serializes the column with the codec with the best estimate on the data.
+pub fn serialize_new<T: MonotonicallyMappableToU64>(
+    value_index: ValueIndexInfo,
+    typed_column: impl Column<T>,
+    output: &mut impl io::Write,
+    codecs: &[FastFieldCodecType],
 ) -> io::Result<()> {
     let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
     let header = Header::compute_header(&column, codecs).ok_or_else(|| {
@@ -235,7 +307,7 @@ pub fn serialize<T: MonotonicallyMappableToU64>(
     serialize_given_codec(normalized_column, header.codec_type, output)?;
 
     let null_index_footer = NullIndexFooter {
-        cardinality: FastFieldCardinality::Single,
+        cardinality: value_index.get_cardinality(),
         null_index_codec: NullIndexCodec::Full,
         null_index_byte_range: 0..0,
     };

ownedbytes/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
 name = "ownedbytes"
-version = "0.4.0"
+version = "0.5.0"
 edition = "2021"
 description = "Expose data as static slice"
 license = "MIT"

ownedbytes/src/lib.rs

@@ -3,7 +3,7 @@ use std::ops::{Deref, Range};
 use std::sync::Arc;
 use std::{fmt, io, mem};
 
-use stable_deref_trait::StableDeref;
+pub use stable_deref_trait::StableDeref;
 
 /// An OwnedBytes simply wraps an object that owns a slice of data and exposes
 /// this data as a slice.

src/aggregation/bucket/histogram/histogram.rs

@@ -206,6 +206,7 @@ pub struct SegmentHistogramCollector {
     field_type: Type,
     interval: f64,
     offset: f64,
+    min_doc_count: u64,
     first_bucket_num: i64,
     bounds: HistogramBounds,
 }
@@ -215,6 +216,30 @@ impl SegmentHistogramCollector {
         self,
         agg_with_accessor: &BucketAggregationWithAccessor,
     ) -> crate::Result<IntermediateBucketResult> {
+        // Compute the number of buckets to validate against max num buckets
+        // Note: We use min_doc_count here, but it's only an lowerbound here, since were are on the
+        // intermediate level and after merging the number of documents of a bucket could exceed
+        // `min_doc_count`.
+        {
+            let cut_off_buckets_front = self
+                .buckets
+                .iter()
+                .take_while(|bucket| bucket.doc_count <= self.min_doc_count)
+                .count();
+            let cut_off_buckets_back = self.buckets[cut_off_buckets_front..]
+                .iter()
+                .rev()
+                .take_while(|bucket| bucket.doc_count <= self.min_doc_count)
+                .count();
+            let estimate_num_buckets =
+                self.buckets.len() - cut_off_buckets_front - cut_off_buckets_back;
+
+            agg_with_accessor
+                .bucket_count
+                .add_count(estimate_num_buckets as u32);
+            agg_with_accessor.bucket_count.validate_bucket_count()?;
+        }
+
         let mut buckets = Vec::with_capacity(
             self.buckets
                 .iter()
@@ -251,11 +276,6 @@ impl SegmentHistogramCollector {
             );
         };
 
-        agg_with_accessor
-            .bucket_count
-            .add_count(buckets.len() as u32);
-        agg_with_accessor.bucket_count.validate_bucket_count()?;
-
         Ok(IntermediateBucketResult::Histogram { buckets })
     }
 
@@ -308,6 +328,7 @@ impl SegmentHistogramCollector {
             first_bucket_num,
             bounds,
             sub_aggregations,
+            min_doc_count: req.min_doc_count(),
         })
     }
 
@@ -380,7 +401,7 @@ impl SegmentHistogramCollector {
 
             debug_assert_eq!(
                 self.buckets[bucket_pos].key,
-                get_bucket_val(val, self.interval, self.offset) as f64
+                get_bucket_val(val, self.interval, self.offset)
             );
             self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
         }
@@ -407,7 +428,7 @@ impl SegmentHistogramCollector {
         if bounds.contains(val) {
             debug_assert_eq!(
                 self.buckets[bucket_pos].key,
-                get_bucket_val(val, self.interval, self.offset) as f64
+                get_bucket_val(val, self.interval, self.offset)
             );
 
             self.increment_bucket(bucket_pos, doc, bucket_with_accessor)?;
@@ -1521,4 +1542,36 @@ mod tests {
 
         Ok(())
     }
+
+    #[test]
+    fn histogram_test_max_buckets_segments() -> crate::Result<()> {
+        let values = vec![0.0, 70000.0];
+
+        let index = get_test_index_from_values(true, &values)?;
+
+        let agg_req: Aggregations = vec![(
+            "my_interval".to_string(),
+            Aggregation::Bucket(BucketAggregation {
+                bucket_agg: BucketAggregationType::Histogram(HistogramAggregation {
+                    field: "score_f64".to_string(),
+                    interval: 1.0,
+                    ..Default::default()
+                }),
+                sub_aggregation: Default::default(),
+            }),
+        )]
+        .into_iter()
+        .collect();
+
+        let res = exec_request(agg_req, &index);
+
+        assert_eq!(
+            res.unwrap_err().to_string(),
+            "An invalid argument was passed: 'Aborting aggregation because too many buckets were \
+             created'"
+                .to_string()
+        );
+
+        Ok(())
+    }
 }

src/aggregation/intermediate_agg_result.rs

@@ -282,8 +282,8 @@ impl IntermediateBucketResult {
             IntermediateBucketResult::Range(range_res) => {
                 let mut buckets: Vec<RangeBucketEntry> = range_res
                     .buckets
-                    .into_iter()
+                    .into_values()
-                    .map(|(_, bucket)| {
+                    .map(|bucket| {
                         bucket.into_final_bucket_entry(
                             &req.sub_aggregation,
                             schema,

src/aggregation/mod.rs

@@ -451,9 +451,9 @@ mod tests {
                         text_field_id => term.to_string(),
                         string_field_id => term.to_string(),
                         score_field => i as u64,
-                        score_field_f64 => i as f64,
+                        score_field_f64 => i,
                         score_field_i64 => i as i64,
-                        fraction_field => i as f64/100.0,
+                        fraction_field => i/100.0,
                     ))?;
                 }
                 index_writer.commit()?;

src/aggregation/segment_agg_result.rs

@@ -305,7 +305,7 @@ impl BucketCount {
     }
     pub(crate) fn add_count(&self, count: u32) {
         self.bucket_count
-            .fetch_add(count as u32, std::sync::atomic::Ordering::Relaxed);
+            .fetch_add(count, std::sync::atomic::Ordering::Relaxed);
     }
     pub(crate) fn get_count(&self) -> u32 {
         self.bucket_count.load(std::sync::atomic::Ordering::Relaxed)

src/collector/facet_collector.rs

@@ -357,7 +357,7 @@ impl SegmentCollector for FacetSegmentCollector {
             let mut facet = vec![];
             let facet_ord = self.collapse_facet_ords[collapsed_facet_ord];
             // TODO handle errors.
-            if facet_dict.ord_to_term(facet_ord as u64, &mut facet).is_ok() {
+            if facet_dict.ord_to_term(facet_ord, &mut facet).is_ok() {
                 if let Ok(facet) = Facet::from_encoded(facet) {
                     facet_counts.insert(facet, count);
                 }

src/collector/mod.rs

@@ -170,7 +170,7 @@ pub trait Collector: Sync + Send {
         segment_ord: u32,
         reader: &SegmentReader,
     ) -> crate::Result<<Self::Child as SegmentCollector>::Fruit> {
-        let mut segment_collector = self.for_segment(segment_ord as u32, reader)?;
+        let mut segment_collector = self.for_segment(segment_ord, reader)?;
 
         match (reader.alive_bitset(), self.requires_scoring()) {
             (Some(alive_bitset), true) => {

src/core/index.rs

@@ -813,7 +813,7 @@ mod tests {
             let field = schema.get_field("num_likes").unwrap();
             let tempdir = TempDir::new().unwrap();
             let tempdir_path = PathBuf::from(tempdir.path());
-            let index = Index::create_in_dir(&tempdir_path, schema).unwrap();
+            let index = Index::create_in_dir(tempdir_path, schema).unwrap();
             let reader = index
                 .reader_builder()
                 .reload_policy(ReloadPolicy::OnCommit)

src/core/inverted_index_reader.rs

@@ -200,10 +200,7 @@ impl InvertedIndexReader {
 
 #[cfg(feature = "quickwit")]
 impl InvertedIndexReader {
-    pub(crate) async fn get_term_info_async(
+    pub(crate) async fn get_term_info_async(&self, term: &Term) -> io::Result<Option<TermInfo>> {
-        &self,
-        term: &Term,
-    ) -> crate::AsyncIoResult<Option<TermInfo>> {
         self.termdict.get_async(term.value_bytes()).await
     }
 
@@ -211,12 +208,8 @@ impl InvertedIndexReader {
     /// This method is for an advanced usage only.
     ///
     /// Most users should prefer using [`Self::read_postings()`] instead.
-    pub async fn warm_postings(
+    pub async fn warm_postings(&self, term: &Term, with_positions: bool) -> io::Result<()> {
-        &self,
+        let term_info_opt: Option<TermInfo> = self.get_term_info_async(term).await?;
-        term: &Term,
-        with_positions: bool,
-    ) -> crate::AsyncIoResult<()> {
-        let term_info_opt = self.get_term_info_async(term).await?;
         if let Some(term_info) = term_info_opt {
             self.postings_file_slice
                 .read_bytes_slice_async(term_info.postings_range.clone())
@@ -234,7 +227,7 @@ impl InvertedIndexReader {
     /// This method is for an advanced usage only.
     ///
     /// If you know which terms to pre-load, prefer using [`Self::warm_postings`] instead.
-    pub async fn warm_postings_full(&self, with_positions: bool) -> crate::AsyncIoResult<()> {
+    pub async fn warm_postings_full(&self, with_positions: bool) -> io::Result<()> {
         self.postings_file_slice.read_bytes_async().await?;
         if with_positions {
             self.positions_file_slice.read_bytes_async().await?;
@@ -243,7 +236,7 @@ impl InvertedIndexReader {
     }
 
     /// Returns the number of documents containing the term asynchronously.
-    pub async fn doc_freq_async(&self, term: &Term) -> crate::AsyncIoResult<u32> {
+    pub async fn doc_freq_async(&self, term: &Term) -> io::Result<u32> {
         Ok(self
             .get_term_info_async(term)
             .await?

src/directory/composite_file.rs

@@ -75,7 +75,7 @@ impl<W: TerminatingWrite + Write> CompositeWrite<W> {
 
         let mut prev_offset = 0;
         for (file_addr, offset) in self.offsets {
-            VInt((offset - prev_offset) as u64).serialize(&mut self.write)?;
+            VInt(offset - prev_offset).serialize(&mut self.write)?;
             file_addr.serialize(&mut self.write)?;
             prev_offset = offset;
         }

src/directory/footer.rs

@@ -38,7 +38,7 @@ impl Footer {
         counting_write.write_all(serde_json::to_string(&self)?.as_ref())?;
         let footer_payload_len = counting_write.written_bytes();
         BinarySerializable::serialize(&(footer_payload_len as u32), write)?;
-        BinarySerializable::serialize(&(FOOTER_MAGIC_NUMBER as u32), write)?;
+        BinarySerializable::serialize(&FOOTER_MAGIC_NUMBER, write)?;
         Ok(())
     }
 
@@ -90,9 +90,10 @@ impl Footer {
             ));
         }
 
-        let footer: Footer = serde_json::from_slice(&file.read_bytes_slice(
+        let footer: Footer =
-            file.len() - total_footer_size..file.len() - footer_metadata_len as usize,
+            serde_json::from_slice(&file.read_bytes_slice(
-        )?)?;
+                file.len() - total_footer_size..file.len() - footer_metadata_len,
+            )?)?;
 
         let body = file.slice_to(file.len() - total_footer_size);
         Ok((footer, body))

src/directory/managed_directory.rs

@@ -388,7 +388,7 @@ mod tests_mmap_specific {
         let tempdir_path = PathBuf::from(tempdir.path());
         let living_files = HashSet::new();
 
-        let mmap_directory = MmapDirectory::open(&tempdir_path).unwrap();
+        let mmap_directory = MmapDirectory::open(tempdir_path).unwrap();
         let mut managed_directory = ManagedDirectory::wrap(Box::new(mmap_directory)).unwrap();
         let mut write = managed_directory.open_write(test_path1).unwrap();
         write.write_all(&[0u8, 1u8]).unwrap();

src/directory/mmap_directory.rs

@@ -6,10 +6,10 @@ use std::path::{Path, PathBuf};
 use std::sync::{Arc, RwLock, Weak};
 use std::{fmt, result};
 
+use common::StableDeref;
 use fs2::FileExt;
 use memmap2::Mmap;
 use serde::{Deserialize, Serialize};
-use stable_deref_trait::StableDeref;
 use tempfile::TempDir;
 
 use crate::core::META_FILEPATH;
@@ -341,7 +341,7 @@ impl Directory for MmapDirectory {
     /// removed before the file is deleted.
     fn delete(&self, path: &Path) -> result::Result<(), DeleteError> {
         let full_path = self.resolve_path(path);
-        fs::remove_file(&full_path).map_err(|e| {
+        fs::remove_file(full_path).map_err(|e| {
             if e.kind() == io::ErrorKind::NotFound {
                 DeleteError::FileDoesNotExist(path.to_owned())
             } else {
@@ -395,7 +395,7 @@ impl Directory for MmapDirectory {
     fn atomic_read(&self, path: &Path) -> Result<Vec<u8>, OpenReadError> {
         let full_path = self.resolve_path(path);
         let mut buffer = Vec::new();
-        match File::open(&full_path) {
+        match File::open(full_path) {
             Ok(mut file) => {
                 file.read_to_end(&mut buffer).map_err(|io_error| {
                     OpenReadError::wrap_io_error(io_error, path.to_path_buf())
@@ -425,7 +425,7 @@ impl Directory for MmapDirectory {
         let file: File = OpenOptions::new()
             .write(true)
             .create(true) //< if the file does not exist yet, create it.
-            .open(&full_path)
+            .open(full_path)
             .map_err(LockError::wrap_io_error)?;
         if lock.is_blocking {
             file.lock_exclusive().map_err(LockError::wrap_io_error)?;

src/directory/mod.rs

@@ -5,7 +5,6 @@ mod mmap_directory;
 
 mod directory;
 mod directory_lock;
-mod file_slice;
 mod file_watcher;
 mod footer;
 mod managed_directory;
@@ -20,13 +19,12 @@ mod composite_file;
 use std::io::BufWriter;
 use std::path::PathBuf;
 
-pub use common::{AntiCallToken, TerminatingWrite};
+pub use common::file_slice::{FileHandle, FileSlice};
-pub use ownedbytes::OwnedBytes;
+pub use common::{AntiCallToken, OwnedBytes, TerminatingWrite};
 
 pub(crate) use self::composite_file::{CompositeFile, CompositeWrite};
 pub use self::directory::{Directory, DirectoryClone, DirectoryLock};
 pub use self::directory_lock::{Lock, INDEX_WRITER_LOCK, META_LOCK};
-pub use self::file_slice::{FileHandle, FileSlice};
 pub use self::ram_directory::RamDirectory;
 pub use self::watch_event_router::{WatchCallback, WatchCallbackList, WatchHandle};
 

src/error.rs

@@ -104,28 +104,6 @@ pub enum TantivyError {
     InternalError(String),
 }
 
-#[cfg(feature = "quickwit")]
-#[derive(Error, Debug)]
-#[doc(hidden)]
-pub enum AsyncIoError {
-    #[error("io::Error `{0}`")]
-    Io(#[from] io::Error),
-    #[error("Asynchronous API is unsupported by this directory")]
-    AsyncUnsupported,
-}
-
-#[cfg(feature = "quickwit")]
-impl From<AsyncIoError> for TantivyError {
-    fn from(async_io_err: AsyncIoError) -> Self {
-        match async_io_err {
-            AsyncIoError::Io(io_err) => TantivyError::from(io_err),
-            AsyncIoError::AsyncUnsupported => {
-                TantivyError::SystemError(format!("{:?}", async_io_err))
-            }
-        }
-    }
-}
-
 impl From<io::Error> for TantivyError {
     fn from(io_err: io::Error) -> TantivyError {
         TantivyError::IoError(Arc::new(io_err))

src/fastfield/alive_bitset.rs

@@ -1,8 +1,7 @@
 use std::io;
 use std::io::Write;
 
-use common::{intersect_bitsets, BitSet, ReadOnlyBitSet};
+use common::{intersect_bitsets, BitSet, OwnedBytes, ReadOnlyBitSet};
-use ownedbytes::OwnedBytes;
 
 use crate::space_usage::ByteCount;
 use crate::DocId;

src/fastfield/facet_reader.rs

@@ -64,9 +64,7 @@ impl FacetReader {
         facet_ord: TermOrdinal,
         output: &mut Facet,
     ) -> crate::Result<()> {
-        let found_term = self
+        let found_term = self.term_dict.ord_to_term(facet_ord, &mut self.buffer)?;
-            .term_dict
-            .ord_to_term(facet_ord as u64, &mut self.buffer)?;
         assert!(found_term, "Term ordinal {} no found.", facet_ord);
         let facet_str = str::from_utf8(&self.buffer[..])
             .map_err(|utf8_err| DataCorruption::comment_only(utf8_err.to_string()))?;

src/fastfield/mod.rs

@@ -473,7 +473,7 @@ mod tests {
             let fast_field_reader = open::<u64>(data)?;
 
             for a in 0..n {
-                assert_eq!(fast_field_reader.get_val(a as u32), permutation[a as usize]);
+                assert_eq!(fast_field_reader.get_val(a as u32), permutation[a]);
             }
         }
         Ok(())

src/fastfield/multivalued/index.rs

@@ -80,6 +80,7 @@ impl MultiValueIndex {
     ///
     /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
     /// match a docid to its value position.
+    #[allow(clippy::bool_to_int_with_if)]
     pub(crate) fn positions_to_docids(&self, doc_id_range: Range<u32>, positions: &mut Vec<u32>) {
         if positions.is_empty() {
             return;

src/fastfield/multivalued/writer.rs

@@ -264,7 +264,7 @@ fn iter_remapped_multivalue_index<'a, C: Column>(
     std::iter::once(0).chain(doc_id_map.iter_old_doc_ids().map(move |old_doc| {
         let num_vals_for_doc = column.get_val(old_doc + 1) - column.get_val(old_doc);
         offset += num_vals_for_doc;
-        offset as u64
+        offset
     }))
 }
 

src/fastfield/writer.rs

@@ -360,20 +360,10 @@ impl U128FastFieldWriter {
                     .map(|idx| self.vals[idx as usize])
             };
 
-            serializer.create_u128_fast_field_with_idx(
+            serializer.create_u128_fast_field_with_idx(self.field, iter_gen, self.val_count, 0)?;
-                self.field,
-                iter_gen,
-                self.val_count as u32,
-                0,
-            )?;
         } else {
             let iter_gen = || self.vals.iter().cloned();
-            serializer.create_u128_fast_field_with_idx(
+            serializer.create_u128_fast_field_with_idx(self.field, iter_gen, self.val_count, 0)?;
-                self.field,
-                iter_gen,
-                self.val_count as u32,
-                0,
-            )?;
         }
 
         Ok(())

src/indexer/demuxer.rs

@@ -252,8 +252,8 @@ mod tests {
             &demux_mapping,
             target_settings,
             vec![
-                Box::new(RamDirectory::default()),
+                Box::<RamDirectory>::default(),
-                Box::new(RamDirectory::default()),
+                Box::<RamDirectory>::default(),
             ],
         )?;
 

src/indexer/index_writer.rs

@@ -152,7 +152,7 @@ pub(crate) fn advance_deletes(
     let num_deleted_docs = max_doc - num_alive_docs;
     if num_deleted_docs > num_deleted_docs_before {
         // There are new deletes. We need to write a new delete file.
-        segment = segment.with_delete_meta(num_deleted_docs as u32, target_opstamp);
+        segment = segment.with_delete_meta(num_deleted_docs, target_opstamp);
         let mut alive_doc_file = segment.open_write(SegmentComponent::Delete)?;
         write_alive_bitset(&alive_bitset, &mut alive_doc_file)?;
         alive_doc_file.terminate()?;
@@ -984,7 +984,7 @@ mod tests {
             "LogMergePolicy { min_num_segments: 8, max_docs_before_merge: 10000000, \
              min_layer_size: 10000, level_log_size: 0.75, del_docs_ratio_before_merge: 1.0 }"
         );
-        let merge_policy = Box::new(NoMergePolicy::default());
+        let merge_policy = Box::<NoMergePolicy>::default();
         index_writer.set_merge_policy(merge_policy);
         assert_eq!(
             format!("{:?}", index_writer.get_merge_policy()),
@@ -1813,8 +1813,8 @@ mod tests {
         }
 
         let num_docs_expected = expected_ids_and_num_occurrences
-            .iter()
+            .values()
-            .map(|(_, id_occurrences)| *id_occurrences as usize)
+            .map(|id_occurrences| *id_occurrences as usize)
             .sum::<usize>();
         assert_eq!(searcher.num_docs() as usize, num_docs_expected);
         assert_eq!(old_searcher.num_docs() as usize, num_docs_expected);

src/indexer/merger.rs

@@ -366,7 +366,7 @@ impl IndexMerger {
                         .map(|doc| reader.num_vals(doc))
                         .sum()
                 } else {
-                    reader.total_num_vals() as u32
+                    reader.total_num_vals()
                 }
             })
             .sum();
@@ -968,7 +968,7 @@ impl IndexMerger {
                     let doc_bytes = doc_bytes_res?;
                     store_writer.store_bytes(&doc_bytes)?;
                 } else {
-                    return Err(DataCorruption::comment_only(&format!(
+                    return Err(DataCorruption::comment_only(format!(
                         "unexpected missing document in docstore on merge, doc address \
                          {old_doc_addr:?}",
                     ))

src/indexer/segment_updater.rs

@@ -866,7 +866,7 @@ mod tests {
         }
 
         assert_eq!(indices.len(), 3);
-        let output_directory: Box<dyn Directory> = Box::new(RamDirectory::default());
+        let output_directory: Box<dyn Directory> = Box::<RamDirectory>::default();
         let index = merge_indices(&indices, output_directory)?;
         assert_eq!(index.schema(), schema);
 

src/indexer/stamper.rs

@@ -16,11 +16,11 @@ mod atomic_impl {
 
     impl AtomicU64Wrapper {
         pub fn new(first_opstamp: Opstamp) -> AtomicU64Wrapper {
-            AtomicU64Wrapper(AtomicU64::new(first_opstamp as u64))
+            AtomicU64Wrapper(AtomicU64::new(first_opstamp))
         }
 
         pub fn fetch_add(&self, val: u64, order: Ordering) -> u64 {
-            self.0.fetch_add(val as u64, order) as u64
+            self.0.fetch_add(val, order)
         }
 
         pub fn revert(&self, val: u64, order: Ordering) -> u64 {
@@ -77,7 +77,7 @@ impl Stamper {
     }
 
     pub fn stamp(&self) -> Opstamp {
-        self.0.fetch_add(1u64, Ordering::SeqCst) as u64
+        self.0.fetch_add(1u64, Ordering::SeqCst)
     }
 
     /// Given a desired count `n`, `stamps` returns an iterator that

src/lib.rs

@@ -177,7 +177,7 @@ impl DateTime {
     /// The given date/time is converted to UTC and the actual
     /// time zone is discarded.
     pub const fn from_utc(dt: OffsetDateTime) -> Self {
-        let timestamp_micros = dt.unix_timestamp() as i64 * 1_000_000 + dt.microsecond() as i64;
+        let timestamp_micros = dt.unix_timestamp() * 1_000_000 + dt.microsecond() as i64;
         Self { timestamp_micros }
     }
 
@@ -259,10 +259,6 @@ pub use crate::future_result::FutureResult;
 /// and instead, refer to this as `crate::Result<T>`.
 pub type Result<T> = std::result::Result<T, TantivyError>;
 
-/// Result for an Async io operation.
-#[cfg(feature = "quickwit")]
-pub type AsyncIoResult<T> = std::result::Result<T, crate::error::AsyncIoError>;
-
 mod core;
 mod indexer;
 

src/positions/reader.rs

@@ -71,7 +71,7 @@ impl PositionReader {
             .map(|num_bits| num_bits as usize)
             .sum();
         let num_bytes_to_skip = num_bits * COMPRESSION_BLOCK_SIZE / 8;
-        self.bit_widths.advance(num_blocks as usize);
+        self.bit_widths.advance(num_blocks);
         self.positions.advance(num_bytes_to_skip);
         self.anchor_offset += (num_blocks * COMPRESSION_BLOCK_SIZE) as u64;
     }

src/postings/indexing_context.rs

@@ -1,11 +1,11 @@
-use crate::postings::stacker::{MemoryArena, TermHashMap};
+use stacker::{ArenaHashMap, MemoryArena};
 
 /// IndexingContext contains all of the transient memory arenas
 /// required for building the inverted index.
 pub(crate) struct IndexingContext {
     /// The term index is an adhoc hashmap,
     /// itself backed by a dedicated memory arena.
-    pub term_index: TermHashMap,
+    pub term_index: ArenaHashMap,
     /// Arena is a memory arena that stores posting lists / term frequencies / positions.
     pub arena: MemoryArena,
 }
@@ -13,9 +13,9 @@ pub(crate) struct IndexingContext {
 impl IndexingContext {
     /// Create a new IndexingContext given the size of the term hash map.
     pub(crate) fn new(table_size: usize) -> IndexingContext {
-        let term_index = TermHashMap::new(table_size);
+        let term_index = ArenaHashMap::new(table_size);
         IndexingContext {
-            arena: MemoryArena::new(),
+            arena: MemoryArena::default(),
             term_index,
         }
     }

src/postings/json_postings_writer.rs

@@ -1,10 +1,11 @@
 use std::io;
 
+use stacker::Addr;
+
 use crate::fastfield::MultiValuedFastFieldWriter;
 use crate::indexer::doc_id_mapping::DocIdMapping;
 use crate::postings::postings_writer::SpecializedPostingsWriter;
 use crate::postings::recorder::{BufferLender, DocIdRecorder, Recorder};
-use crate::postings::stacker::Addr;
 use crate::postings::{
     FieldSerializer, IndexingContext, IndexingPosition, PostingsWriter, UnorderedTermId,
 };

src/postings/mod.rs

@@ -15,9 +15,10 @@ mod recorder;
 mod segment_postings;
 mod serializer;
 mod skip;
-mod stacker;
 mod term_info;
 
+pub(crate) use stacker::compute_table_size;
+
 pub use self::block_segment_postings::BlockSegmentPostings;
 pub(crate) use self::indexing_context::IndexingContext;
 pub(crate) use self::per_field_postings_writer::PerFieldPostingsWriter;
@@ -26,10 +27,9 @@ pub(crate) use self::postings_writer::{serialize_postings, IndexingPosition, Pos
 pub use self::segment_postings::SegmentPostings;
 pub use self::serializer::{FieldSerializer, InvertedIndexSerializer};
 pub(crate) use self::skip::{BlockInfo, SkipReader};
-pub(crate) use self::stacker::compute_table_size;
 pub use self::term_info::TermInfo;
 
-pub(crate) type UnorderedTermId = u64;
+pub(crate) type UnorderedTermId = stacker::UnorderedId;
 
 #[allow(clippy::enum_variant_names)]
 #[derive(Debug, PartialEq, Clone, Copy, Eq)]

src/postings/per_field_postings_writer.rs

@@ -51,7 +51,7 @@ fn posting_writer_from_field_entry(field_entry: &FieldEntry) -> Box<dyn Postings
         | FieldType::Date(_)
         | FieldType::Bytes(_)
         | FieldType::IpAddr(_)
-        | FieldType::Facet(_) => Box::new(SpecializedPostingsWriter::<DocIdRecorder>::default()),
+        | FieldType::Facet(_) => Box::<SpecializedPostingsWriter<DocIdRecorder>>::default(),
         FieldType::JsonObject(ref json_object_options) => {
             if let Some(text_indexing_option) = json_object_options.get_text_indexing_options() {
                 match text_indexing_option.index_option() {

src/postings/postings_writer.rs

@@ -4,8 +4,8 @@ use std::marker::PhantomData;
 use std::ops::Range;
 
 use rustc_hash::FxHashMap;
+use stacker::Addr;
 
-use super::stacker::Addr;
 use crate::fastfield::MultiValuedFastFieldWriter;
 use crate::fieldnorm::FieldNormReaders;
 use crate::indexer::doc_id_mapping::DocIdMapping;
@@ -59,7 +59,11 @@ pub(crate) fn serialize_postings(
 ) -> crate::Result<HashMap<Field, FxHashMap<UnorderedTermId, TermOrdinal>>> {
     let mut term_offsets: Vec<(Term<&[u8]>, Addr, UnorderedTermId)> =
         Vec::with_capacity(ctx.term_index.len());
-    term_offsets.extend(ctx.term_index.iter());
+    term_offsets.extend(
+        ctx.term_index
+            .iter()
+            .map(|(bytes, addr, unordered_id)| (Term::wrap(bytes), addr, unordered_id)),
+    );
     term_offsets.sort_unstable_by_key(|(k, _, _)| k.clone());
     let mut unordered_term_mappings: HashMap<Field, FxHashMap<UnorderedTermId, TermOrdinal>> =
         HashMap::new();

src/postings/recorder.rs

@@ -1,6 +1,6 @@
 use common::read_u32_vint;
+use stacker::{ExpUnrolledLinkedList, MemoryArena};
 
-use super::stacker::{ExpUnrolledLinkedList, MemoryArena};
 use crate::indexer::doc_id_mapping::DocIdMapping;
 use crate::postings::FieldSerializer;
 use crate::DocId;
@@ -91,7 +91,7 @@ pub struct DocIdRecorder {
 impl Default for DocIdRecorder {
     fn default() -> Self {
         DocIdRecorder {
-            stack: ExpUnrolledLinkedList::new(),
+            stack: ExpUnrolledLinkedList::default(),
             current_doc: u32::MAX,
         }
     }
@@ -144,7 +144,7 @@ impl Recorder for DocIdRecorder {
 }
 
 /// Recorder encoding document ids, and term frequencies
-#[derive(Clone, Copy)]
+#[derive(Clone, Copy, Default)]
 pub struct TermFrequencyRecorder {
     stack: ExpUnrolledLinkedList,
     current_doc: DocId,
@@ -152,17 +152,6 @@ pub struct TermFrequencyRecorder {
     term_doc_freq: u32,
 }
 
-impl Default for TermFrequencyRecorder {
-    fn default() -> Self {
-        TermFrequencyRecorder {
-            stack: ExpUnrolledLinkedList::new(),
-            current_doc: 0,
-            current_tf: 0u32,
-            term_doc_freq: 0u32,
-        }
-    }
-}
-
 impl Recorder for TermFrequencyRecorder {
     fn current_doc(&self) -> DocId {
         self.current_doc
@@ -229,7 +218,7 @@ pub struct TfAndPositionRecorder {
 impl Default for TfAndPositionRecorder {
     fn default() -> Self {
         TfAndPositionRecorder {
-            stack: ExpUnrolledLinkedList::new(),
+            stack: ExpUnrolledLinkedList::default(),
             current_doc: u32::MAX,
             term_doc_freq: 0u32,
         }

src/postings/serializer.rs

@@ -465,7 +465,7 @@ impl<W: Write> PostingsSerializer<W> {
     /// When called after writing the postings of a term, this value is used as a
     /// end offset.
     fn written_bytes(&self) -> u64 {
-        self.output_write.written_bytes() as u64
+        self.output_write.written_bytes()
     }
 
     fn clear(&mut self) {

src/postings/stacker/mod.rs

@@ -1,7 +0,0 @@
-mod expull;
-mod memory_arena;
-mod term_hashmap;
-
-pub(crate) use self::expull::ExpUnrolledLinkedList;
-pub(crate) use self::memory_arena::{Addr, MemoryArena};
-pub(crate) use self::term_hashmap::{compute_table_size, TermHashMap};

src/query/bitset/mod.rs

@@ -47,7 +47,7 @@ impl From<BitSet> for BitSetDocSet {
 impl DocSet for BitSetDocSet {
     fn advance(&mut self) -> DocId {
         if let Some(lower) = self.cursor_tinybitset.pop_lowest() {
-            self.doc = (self.cursor_bucket as u32 * 64u32) | lower;
+            self.doc = (self.cursor_bucket * 64u32) | lower;
             return self.doc;
         }
         if let Some(cursor_bucket) = self.docs.first_non_empty_bucket(self.cursor_bucket + 1) {

src/query/disjunction_max_query.rs

@@ -3,7 +3,7 @@ use tantivy_query_grammar::Occur;
 use crate::query::{BooleanWeight, DisjunctionMaxCombiner, EnableScoring, Query, Weight};
 use crate::{Score, Term};
 
-/// The disjunction max query кeturns documents matching one or more wrapped queries,
+/// The disjunction max query returns documents matching one or more wrapped queries,
 /// called query clauses or clauses.
 ///
 /// If a returned document matches multiple query clauses,

src/query/range_query_ip_fastfield.rs

@@ -126,7 +126,7 @@ impl VecCursor {
     }
     #[inline]
     fn current(&self) -> Option<u32> {
-        self.docs.get(self.current_pos).map(|el| *el as u32)
+        self.docs.get(self.current_pos).copied()
     }
     fn get_cleared_data(&mut self) -> &mut Vec<u32> {
         self.docs.clear();
@@ -268,9 +268,9 @@ impl DocSet for IpRangeDocSet {
     #[inline]
     fn advance(&mut self) -> DocId {
         if let Some(docid) = self.loaded_docs.next() {
-            docid as u32
+            docid
         } else {
-            if self.next_fetch_start >= self.ip_addr_fast_field.num_docs() as u32 {
+            if self.next_fetch_start >= self.ip_addr_fast_field.num_docs() {
                 return TERMINATED;
             }
             self.fetch_block();
@@ -280,10 +280,7 @@ impl DocSet for IpRangeDocSet {
 
     #[inline]
     fn doc(&self) -> DocId {
-        self.loaded_docs
+        self.loaded_docs.current().unwrap_or(TERMINATED)
-            .current()
-            .map(|el| el as u32)
-            .unwrap_or(TERMINATED)
     }
 
     /// Advances the `DocSet` forward until reaching the target, or going to the

src/query/union.rs

@@ -43,7 +43,7 @@ fn refill<TScorer: Scorer, TScoreCombiner: ScoreCombiner>(
     min_doc: DocId,
 ) {
     unordered_drain_filter(scorers, |scorer| {
-        let horizon = min_doc + HORIZON as u32;
+        let horizon = min_doc + HORIZON;
         loop {
             let doc = scorer.doc();
             if doc >= horizon {

src/schema/date_time_options.rs

@@ -236,7 +236,7 @@ mod tests {
         )
         .unwrap();
 
-        let date_options_json = serde_json::to_value(&date_options).unwrap();
+        let date_options_json = serde_json::to_value(date_options).unwrap();
         assert_eq!(
             date_options_json,
             serde_json::json!({

src/store/index/mod.rs

@@ -193,8 +193,8 @@ mod tests {
         (0..max_len)
             .prop_flat_map(move |len: usize| {
                 (
-                    proptest::collection::vec(1usize..20, len as usize).prop_map(integrate_delta),
+                    proptest::collection::vec(1usize..20, len).prop_map(integrate_delta),
-                    proptest::collection::vec(1usize..26, len as usize).prop_map(integrate_delta),
+                    proptest::collection::vec(1usize..26, len).prop_map(integrate_delta),
                 )
                     .prop_map(|(docs, offsets)| {
                         (0..docs.len() - 1)

src/store/reader.rs

@@ -4,9 +4,8 @@ use std::ops::{AddAssign, Range};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
 
-use common::{BinarySerializable, HasLen};
+use common::{BinarySerializable, HasLen, OwnedBytes};
 use lru::LruCache;
-use ownedbytes::OwnedBytes;
 
 use super::footer::DocStoreFooter;
 use super::index::SkipIndex;
@@ -66,11 +65,7 @@ impl BlockCache {
 
     #[cfg(test)]
     fn peek_lru(&self) -> Option<usize> {
-        self.cache
+        self.cache.lock().unwrap().peek_lru().map(|(&k, _)| k)
-            .lock()
-            .unwrap()
-            .peek_lru()
-            .map(|(&k, _)| k as usize)
     }
 }
 
@@ -323,7 +318,7 @@ impl StoreReader {
     /// In most cases use [`get_async`](Self::get_async)
     ///
     /// Loads and decompresses a block asynchronously.
-    async fn read_block_async(&self, checkpoint: &Checkpoint) -> crate::AsyncIoResult<Block> {
+    async fn read_block_async(&self, checkpoint: &Checkpoint) -> io::Result<Block> {
         let cache_key = checkpoint.byte_range.start;
         if let Some(block) = self.cache.get_from_cache(checkpoint.byte_range.start) {
             return Ok(block);

src/store/store_compressor.rs

@@ -142,7 +142,7 @@ impl BlockCompressorImpl {
     }
 
     fn close(mut self) -> io::Result<()> {
-        let header_offset: u64 = self.writer.written_bytes() as u64;
+        let header_offset: u64 = self.writer.written_bytes();
         let docstore_footer =
             DocStoreFooter::new(header_offset, Decompressor::from(self.compressor));
         self.offset_index_writer.serialize_into(&mut self.writer)?;

src/termdict/fst_termdict/term_info_store.rs

@@ -69,7 +69,7 @@ impl TermInfoBlockMeta {
         let posting_end_addr = posting_start_addr + num_bits;
         let positions_start_addr = posting_start_addr + self.postings_offset_nbits as usize;
         // the position_end is the positions_start of the next term info.
-        let positions_end_addr = positions_start_addr + num_bits as usize;
+        let positions_end_addr = positions_start_addr + num_bits;
 
         let doc_freq_addr = positions_start_addr + self.positions_offset_nbits as usize;
 
@@ -121,7 +121,7 @@ fn extract_bits(data: &[u8], addr_bits: usize, num_bits: u8) -> u64 {
 }
 
 impl TermInfoStore {
-    pub fn open(term_info_store_file: FileSlice) -> crate::Result<TermInfoStore> {
+    pub fn open(term_info_store_file: FileSlice) -> io::Result<TermInfoStore> {
         let (len_slice, main_slice) = term_info_store_file.split(16);
         let mut bytes = len_slice.read_bytes()?;
         let len = u64::deserialize(&mut bytes)? as usize;

src/termdict/fst_termdict/termdict.rs

@@ -8,7 +8,6 @@ use tantivy_fst::Automaton;
 use super::term_info_store::{TermInfoStore, TermInfoStoreWriter};
 use super::{TermStreamer, TermStreamerBuilder};
 use crate::directory::{FileSlice, OwnedBytes};
-use crate::error::DataCorruption;
 use crate::postings::TermInfo;
 use crate::termdict::TermOrdinal;
 
@@ -55,7 +54,7 @@ where W: Write
     /// to insert_key and insert_value.
     ///
     /// Prefer using `.insert(key, value)`
-    pub(crate) fn insert_key(&mut self, key: &[u8]) -> io::Result<()> {
+    pub fn insert_key(&mut self, key: &[u8]) -> io::Result<()> {
         self.fst_builder
             .insert(key, self.term_ord)
             .map_err(convert_fst_error)?;
@@ -66,7 +65,7 @@ where W: Write
     /// # Warning
     ///
     /// Horribly dangerous internal API. See `.insert_key(...)`.
-    pub(crate) fn insert_value(&mut self, term_info: &TermInfo) -> io::Result<()> {
+    pub fn insert_value(&mut self, term_info: &TermInfo) -> io::Result<()> {
         self.term_info_store_writer.write_term_info(term_info)?;
         Ok(())
     }
@@ -80,16 +79,20 @@ where W: Write
             self.term_info_store_writer
                 .serialize(&mut counting_writer)?;
             let footer_size = counting_writer.written_bytes();
-            (footer_size as u64).serialize(&mut counting_writer)?;
+            footer_size.serialize(&mut counting_writer)?;
         }
         Ok(file)
     }
 }
 
-fn open_fst_index(fst_file: FileSlice) -> crate::Result<tantivy_fst::Map<OwnedBytes>> {
+fn open_fst_index(fst_file: FileSlice) -> io::Result<tantivy_fst::Map<OwnedBytes>> {
     let bytes = fst_file.read_bytes()?;
-    let fst = Fst::new(bytes)
+    let fst = Fst::new(bytes).map_err(|err| {
-        .map_err(|err| DataCorruption::comment_only(format!("Fst data is corrupted: {:?}", err)))?;
+        io::Error::new(
+            io::ErrorKind::InvalidData,
+            format!("Fst data is corrupted: {:?}", err),
+        )
+    })?;
     Ok(tantivy_fst::Map::from(fst))
 }
 
@@ -114,7 +117,7 @@ pub struct TermDictionary {
 
 impl TermDictionary {
     /// Opens a `TermDictionary`.
-    pub fn open(file: FileSlice) -> crate::Result<Self> {
+    pub fn open(file: FileSlice) -> io::Result<Self> {
         let (main_slice, footer_len_slice) = file.split_from_end(8);
         let mut footer_len_bytes = footer_len_slice.read_bytes()?;
         let footer_size = u64::deserialize(&mut footer_len_bytes)?;

src/termdict/sstable_termdict/mod.rs

@@ -1,50 +1,64 @@
 use std::io;
 
 mod merger;
-mod sstable;
-mod streamer;
-mod termdict;
 
 use std::iter::ExactSizeIterator;
 
 use common::VInt;
+use sstable::value::{ValueReader, ValueWriter};
+use sstable::SSTable;
+use tantivy_fst::automaton::AlwaysMatch;
 
 pub use self::merger::TermMerger;
-use self::sstable::value::{ValueReader, ValueWriter};
-use self::sstable::{BlockReader, SSTable};
-pub use self::streamer::{TermStreamer, TermStreamerBuilder};
-pub use self::termdict::{TermDictionary, TermDictionaryBuilder};
 use crate::postings::TermInfo;
 
+/// The term dictionary contains all of the terms in
+/// `tantivy index` in a sorted manner.
+///
+/// The `Fst` crate is used to associate terms to their
+/// respective `TermOrdinal`. The `TermInfoStore` then makes it
+/// possible to fetch the associated `TermInfo`.
+pub type TermDictionary = sstable::Dictionary<TermSSTable>;
+
+/// Builder for the new term dictionary.
+pub type TermDictionaryBuilder<W> = sstable::Writer<W, TermInfoValueWriter>;
+
+/// `TermStreamer` acts as a cursor over a range of terms of a segment.
+/// Terms are guaranteed to be sorted.
+pub type TermStreamer<'a, A = AlwaysMatch> = sstable::Streamer<'a, TermSSTable, A>;
+
+/// SSTable used to store TermInfo objects.
 pub struct TermSSTable;
 
 impl SSTable for TermSSTable {
     type Value = TermInfo;
-    type Reader = TermInfoReader;
+    type ValueReader = TermInfoValueReader;
-    type Writer = TermInfoWriter;
+    type ValueWriter = TermInfoValueWriter;
 }
 
 #[derive(Default)]
-pub struct TermInfoReader {
+pub struct TermInfoValueReader {
     term_infos: Vec<TermInfo>,
 }
 
-impl ValueReader for TermInfoReader {
+impl ValueReader for TermInfoValueReader {
     type Value = TermInfo;
 
+    #[inline(always)]
     fn value(&self, idx: usize) -> &TermInfo {
         &self.term_infos[idx]
     }
 
-    fn read(&mut self, reader: &mut BlockReader) -> io::Result<()> {
+    fn load(&mut self, mut data: &[u8]) -> io::Result<usize> {
+        let len_before = data.len();
         self.term_infos.clear();
-        let num_els = VInt::deserialize_u64(reader)?;
+        let num_els = VInt::deserialize_u64(&mut data)?;
-        let mut postings_start = VInt::deserialize_u64(reader)? as usize;
+        let mut postings_start = VInt::deserialize_u64(&mut data)? as usize;
-        let mut positions_start = VInt::deserialize_u64(reader)? as usize;
+        let mut positions_start = VInt::deserialize_u64(&mut data)? as usize;
         for _ in 0..num_els {
-            let doc_freq = VInt::deserialize_u64(reader)? as u32;
+            let doc_freq = VInt::deserialize_u64(&mut data)? as u32;
-            let postings_num_bytes = VInt::deserialize_u64(reader)?;
+            let postings_num_bytes = VInt::deserialize_u64(&mut data)?;
-            let positions_num_bytes = VInt::deserialize_u64(reader)?;
+            let positions_num_bytes = VInt::deserialize_u64(&mut data)?;
             let postings_end = postings_start + postings_num_bytes as usize;
             let positions_end = positions_start + positions_num_bytes as usize;
             let term_info = TermInfo {
@@ -56,23 +70,24 @@ impl ValueReader for TermInfoReader {
             postings_start = postings_end;
             positions_start = positions_end;
         }
-        Ok(())
+        let consumed_len = len_before - data.len();
+        Ok(consumed_len)
     }
 }
 
 #[derive(Default)]
-pub struct TermInfoWriter {
+pub struct TermInfoValueWriter {
     term_infos: Vec<TermInfo>,
 }
 
-impl ValueWriter for TermInfoWriter {
+impl ValueWriter for TermInfoValueWriter {
     type Value = TermInfo;
 
     fn write(&mut self, term_info: &TermInfo) {
         self.term_infos.push(term_info.clone());
     }
 
-    fn write_block(&mut self, buffer: &mut Vec<u8>) {
+    fn serialize_block(&self, buffer: &mut Vec<u8>) {
         VInt(self.term_infos.len() as u64).serialize_into_vec(buffer);
         if self.term_infos.is_empty() {
             return;
@@ -84,23 +99,23 @@ impl ValueWriter for TermInfoWriter {
             VInt(term_info.postings_range.len() as u64).serialize_into_vec(buffer);
             VInt(term_info.positions_range.len() as u64).serialize_into_vec(buffer);
         }
+    }
+
+    fn clear(&mut self) {
         self.term_infos.clear();
     }
 }
 
 #[cfg(test)]
 mod tests {
-    use std::io;
+    use sstable::value::{ValueReader, ValueWriter};
 
-    use super::BlockReader;
-    use crate::directory::OwnedBytes;
     use crate::postings::TermInfo;
-    use crate::termdict::sstable_termdict::sstable::value::{ValueReader, ValueWriter};
+    use crate::termdict::sstable_termdict::TermInfoValueReader;
-    use crate::termdict::sstable_termdict::TermInfoReader;
 
     #[test]
-    fn test_block_terminfos() -> io::Result<()> {
+    fn test_block_terminfos() {
-        let mut term_info_writer = super::TermInfoWriter::default();
+        let mut term_info_writer = super::TermInfoValueWriter::default();
         term_info_writer.write(&TermInfo {
             doc_freq: 120u32,
             postings_range: 17..45,
@@ -117,10 +132,9 @@ mod tests {
             positions_range: 1100..1302,
         });
         let mut buffer = Vec::new();
-        term_info_writer.write_block(&mut buffer);
+        term_info_writer.serialize_block(&mut buffer);
-        let mut block_reader = make_block_reader(&buffer[..]);
+        let mut term_info_reader = TermInfoValueReader::default();
-        let mut term_info_reader = TermInfoReader::default();
+        let num_bytes: usize = term_info_reader.load(&buffer[..]).unwrap();
-        term_info_reader.read(&mut block_reader)?;
         assert_eq!(
             term_info_reader.value(0),
             &TermInfo {
@@ -129,16 +143,6 @@ mod tests {
                 positions_range: 10..122
             }
         );
-        assert!(block_reader.buffer().is_empty());
+        assert_eq!(buffer.len(), num_bytes);
-        Ok(())
-    }
-
-    fn make_block_reader(data: &[u8]) -> BlockReader {
-        let mut buffer = (data.len() as u32).to_le_bytes().to_vec();
-        buffer.extend_from_slice(data);
-        let owned_bytes = OwnedBytes::new(buffer);
-        let mut block_reader = BlockReader::new(Box::new(owned_bytes));
-        block_reader.read_block().unwrap();
-        block_reader
     }
 }

src/termdict/sstable_termdict/sstable/value.rs

@@ -1,95 +0,0 @@
-use std::io;
-
-use super::{vint, BlockReader};
-
-pub trait ValueReader: Default {
-    type Value;
-
-    fn value(&self, idx: usize) -> &Self::Value;
-
-    fn read(&mut self, reader: &mut BlockReader) -> io::Result<()>;
-}
-
-pub trait ValueWriter: Default {
-    type Value;
-
-    fn write(&mut self, val: &Self::Value);
-
-    fn write_block(&mut self, writer: &mut Vec<u8>);
-}
-
-#[derive(Default)]
-pub struct VoidReader;
-
-impl ValueReader for VoidReader {
-    type Value = ();
-
-    fn value(&self, _idx: usize) -> &() {
-        &()
-    }
-
-    fn read(&mut self, _reader: &mut BlockReader) -> io::Result<()> {
-        Ok(())
-    }
-}
-
-#[derive(Default)]
-pub struct VoidWriter;
-
-impl ValueWriter for VoidWriter {
-    type Value = ();
-
-    fn write(&mut self, _val: &()) {}
-
-    fn write_block(&mut self, _writer: &mut Vec<u8>) {}
-}
-
-#[derive(Default)]
-pub struct U64MonotonicWriter {
-    vals: Vec<u64>,
-}
-
-impl ValueWriter for U64MonotonicWriter {
-    type Value = u64;
-
-    fn write(&mut self, val: &Self::Value) {
-        self.vals.push(*val);
-    }
-
-    fn write_block(&mut self, writer: &mut Vec<u8>) {
-        let mut prev_val = 0u64;
-        vint::serialize_into_vec(self.vals.len() as u64, writer);
-        for &val in &self.vals {
-            let delta = val - prev_val;
-            vint::serialize_into_vec(delta, writer);
-            prev_val = val;
-        }
-        self.vals.clear();
-    }
-}
-
-#[derive(Default)]
-pub struct U64MonotonicReader {
-    vals: Vec<u64>,
-}
-
-impl ValueReader for U64MonotonicReader {
-    type Value = u64;
-
-    fn value(&self, idx: usize) -> &Self::Value {
-        &self.vals[idx]
-    }
-
-    fn read(&mut self, reader: &mut BlockReader) -> io::Result<()> {
-        let len = reader.deserialize_u64() as usize;
-        self.vals.clear();
-        let mut prev_val = 0u64;
-        for _ in 0..len {
-            let delta = reader.deserialize_u64() as u64;
-            let val = prev_val + delta;
-            self.vals.push(val);
-            prev_val = val;
-        }
-        Ok(())
-    }
-}

src/termdict/sstable_termdict/termdict.rs

@@ -1,258 +0,0 @@
-use std::io;
-use std::sync::Arc;
-
-use common::BinarySerializable;
-use once_cell::sync::Lazy;
-use tantivy_fst::automaton::AlwaysMatch;
-use tantivy_fst::Automaton;
-
-use crate::directory::{FileSlice, OwnedBytes};
-use crate::postings::TermInfo;
-use crate::termdict::sstable_termdict::sstable::sstable_index::BlockAddr;
-use crate::termdict::sstable_termdict::sstable::{
-    DeltaReader, Reader, SSTable, SSTableIndex, Writer,
-};
-use crate::termdict::sstable_termdict::{
-    TermInfoReader, TermInfoWriter, TermSSTable, TermStreamer, TermStreamerBuilder,
-};
-use crate::termdict::TermOrdinal;
-use crate::AsyncIoResult;
-
-pub struct TermInfoSSTable;
-impl SSTable for TermInfoSSTable {
-    type Value = TermInfo;
-    type Reader = TermInfoReader;
-    type Writer = TermInfoWriter;
-}
-
-/// Builder for the new term dictionary.
-pub struct TermDictionaryBuilder<W: io::Write> {
-    sstable_writer: Writer<W, TermInfoWriter>,
-}
-
-impl<W: io::Write> TermDictionaryBuilder<W> {
-    /// Creates a new `TermDictionaryBuilder`
-    pub fn create(w: W) -> io::Result<Self> {
-        let sstable_writer = TermSSTable::writer(w);
-        Ok(TermDictionaryBuilder { sstable_writer })
-    }
-
-    /// Inserts a `(key, value)` pair in the term dictionary.
-    ///
-    /// *Keys have to be inserted in order.*
-    pub fn insert<K: AsRef<[u8]>>(&mut self, key_ref: K, value: &TermInfo) -> io::Result<()> {
-        let key = key_ref.as_ref();
-        self.insert_key(key)?;
-        self.insert_value(value)?;
-        Ok(())
-    }
-
-    /// # Warning
-    /// Horribly dangerous internal API
-    ///
-    /// If used, it must be used by systematically alternating calls
-    /// to insert_key and insert_value.
-    ///
-    /// Prefer using `.insert(key, value)`
-    #[allow(clippy::unnecessary_wraps)]
-    pub(crate) fn insert_key(&mut self, key: &[u8]) -> io::Result<()> {
-        self.sstable_writer.write_key(key);
-        Ok(())
-    }
-
-    /// # Warning
-    ///
-    /// Horribly dangerous internal API. See `.insert_key(...)`.
-    pub(crate) fn insert_value(&mut self, term_info: &TermInfo) -> io::Result<()> {
-        self.sstable_writer.write_value(term_info)
-    }
-
-    /// Finalize writing the builder, and returns the underlying
-    /// `Write` object.
-    pub fn finish(self) -> io::Result<W> {
-        self.sstable_writer.finalize()
-    }
-}
-
-static EMPTY_TERM_DICT_FILE: Lazy<FileSlice> = Lazy::new(|| {
-    let term_dictionary_data: Vec<u8> = TermDictionaryBuilder::create(Vec::<u8>::new())
-        .expect("Creating a TermDictionaryBuilder in a Vec<u8> should never fail")
-        .finish()
-        .expect("Writing in a Vec<u8> should never fail");
-    FileSlice::from(term_dictionary_data)
-});
-
-/// The term dictionary contains all of the terms in
-/// `tantivy index` in a sorted manner.
-///
-/// The `Fst` crate is used to associate terms to their
-/// respective `TermOrdinal`. The `TermInfoStore` then makes it
-/// possible to fetch the associated `TermInfo`.
-pub struct TermDictionary {
-    sstable_slice: FileSlice,
-    sstable_index: SSTableIndex,
-    num_terms: u64,
-}
-
-impl TermDictionary {
-    pub(crate) fn sstable_reader(&self) -> io::Result<Reader<'static, TermInfoReader>> {
-        let data = self.sstable_slice.read_bytes()?;
-        Ok(TermInfoSSTable::reader(data))
-    }
-
-    pub(crate) fn sstable_reader_block(
-        &self,
-        block_addr: BlockAddr,
-    ) -> io::Result<Reader<'static, TermInfoReader>> {
-        let data = self.sstable_slice.read_bytes_slice(block_addr.byte_range)?;
-        Ok(TermInfoSSTable::reader(data))
-    }
-
-    pub(crate) async fn sstable_reader_block_async(
-        &self,
-        block_addr: BlockAddr,
-    ) -> AsyncIoResult<Reader<'static, TermInfoReader>> {
-        let data = self
-            .sstable_slice
-            .read_bytes_slice_async(block_addr.byte_range)
-            .await?;
-        Ok(TermInfoSSTable::reader(data))
-    }
-
-    pub(crate) fn sstable_delta_reader(&self) -> io::Result<DeltaReader<'static, TermInfoReader>> {
-        let data = self.sstable_slice.read_bytes()?;
-        Ok(TermInfoSSTable::delta_reader(data))
-    }
-
-    /// Opens a `TermDictionary`.
-    pub fn open(term_dictionary_file: FileSlice) -> crate::Result<Self> {
-        let (main_slice, footer_len_slice) = term_dictionary_file.split_from_end(16);
-        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 (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.as_slice())?;
-        Ok(TermDictionary {
-            sstable_slice,
-            sstable_index,
-            num_terms,
-        })
-    }
-
-    /// Creates a term dictionary from the supplied bytes.
-    pub fn from_bytes(owned_bytes: OwnedBytes) -> crate::Result<TermDictionary> {
-        TermDictionary::open(FileSlice::new(Arc::new(owned_bytes)))
-    }
-
-    /// Creates an empty term dictionary which contains no terms.
-    pub fn empty() -> Self {
-        TermDictionary::open(EMPTY_TERM_DICT_FILE.clone()).unwrap()
-    }
-
-    /// Returns the number of terms in the dictionary.
-    /// Term ordinals range from 0 to `num_terms() - 1`.
-    pub fn num_terms(&self) -> usize {
-        self.num_terms as usize
-    }
-
-    /// Returns the ordinal associated with a given term.
-    pub fn term_ord<K: AsRef<[u8]>>(&self, key: K) -> io::Result<Option<TermOrdinal>> {
-        let mut term_ord = 0u64;
-        let key_bytes = key.as_ref();
-        let mut sstable_reader = self.sstable_reader()?;
-        while sstable_reader.advance().unwrap_or(false) {
-            if sstable_reader.key() == key_bytes {
-                return Ok(Some(term_ord));
-            }
-            term_ord += 1;
-        }
-        Ok(None)
-    }
-
-    /// Returns the term associated with a given term ordinal.
-    ///
-    /// Term ordinals are defined as the position of the term in
-    /// the sorted list of terms.
-    ///
-    /// Returns true if and only if the term has been found.
-    ///
-    /// Regardless of whether the term is found or not,
-    /// the buffer may be modified.
-    pub fn ord_to_term(&self, ord: TermOrdinal, bytes: &mut Vec<u8>) -> io::Result<bool> {
-        let mut sstable_reader = self.sstable_reader()?;
-        bytes.clear();
-        for _ in 0..(ord + 1) {
-            if !sstable_reader.advance().unwrap_or(false) {
-                return Ok(false);
-            }
-        }
-        bytes.extend_from_slice(sstable_reader.key());
-        Ok(true)
-    }
-
-    /// Returns the number of terms in the dictionary.
-    pub fn term_info_from_ord(&self, term_ord: TermOrdinal) -> io::Result<TermInfo> {
-        let mut sstable_reader = self.sstable_reader()?;
-        for _ in 0..(term_ord + 1) {
-            if !sstable_reader.advance().unwrap_or(false) {
-                return Ok(TermInfo::default());
-            }
-        }
-        Ok(sstable_reader.value().clone())
-    }
-
-    /// Lookups the value corresponding to the key.
-    pub fn get<K: AsRef<[u8]>>(&self, key: K) -> io::Result<Option<TermInfo>> {
-        if let Some(block_addr) = self.sstable_index.search(key.as_ref()) {
-            let mut sstable_reader = self.sstable_reader_block(block_addr)?;
-            let key_bytes = key.as_ref();
-            while sstable_reader.advance().unwrap_or(false) {
-                if sstable_reader.key() == key_bytes {
-                    let term_info = sstable_reader.value().clone();
-                    return Ok(Some(term_info));
-                }
-            }
-        }
-        Ok(None)
-    }
-
-    /// Lookups the value corresponding to the key.
-    pub async fn get_async<K: AsRef<[u8]>>(&self, key: K) -> AsyncIoResult<Option<TermInfo>> {
-        if let Some(block_addr) = self.sstable_index.search(key.as_ref()) {
-            let mut sstable_reader = self.sstable_reader_block_async(block_addr).await?;
-            let key_bytes = key.as_ref();
-            while sstable_reader.advance().unwrap_or(false) {
-                if sstable_reader.key() == key_bytes {
-                    let term_info = sstable_reader.value().clone();
-                    return Ok(Some(term_info));
-                }
-            }
-        }
-        Ok(None)
-    }
-
-    /// Returns a range builder, to stream all of the terms
-    /// within an interval.
-    pub fn range(&self) -> TermStreamerBuilder<'_> {
-        TermStreamerBuilder::new(self, AlwaysMatch)
-    }
-
-    /// A stream of all the sorted terms.
-    pub fn stream(&self) -> io::Result<TermStreamer<'_>> {
-        self.range().into_stream()
-    }
-
-    /// Returns a search builder, to stream all of the terms
-    /// within the Automaton
-    pub fn search<'a, A: Automaton + 'a>(&'a self, automaton: A) -> TermStreamerBuilder<'a, A>
-    where A::State: Clone {
-        TermStreamerBuilder::<A>::new(self, automaton)
-    }
-
-    #[doc(hidden)]
-    pub async fn warm_up_dictionary(&self) -> AsyncIoResult<()> {
-        self.sstable_slice.read_bytes_async().await?;
-        Ok(())
-    }
-}

src/termdict/tests.rs

@@ -1,5 +1,5 @@
 use std::path::PathBuf;
-use std::str;
+use std::{io, str};
 
 use super::{TermDictionary, TermDictionaryBuilder, TermStreamer};
 use crate::directory::{Directory, FileSlice, RamDirectory, TerminatingWrite};
@@ -247,7 +247,7 @@ fn test_empty_string() -> crate::Result<()> {
     Ok(())
 }
 
-fn stream_range_test_dict() -> crate::Result<TermDictionary> {
+fn stream_range_test_dict() -> io::Result<TermDictionary> {
     let buffer: Vec<u8> = {
         let mut term_dictionary_builder = TermDictionaryBuilder::create(Vec::new())?;
         for i in 0u8..10u8 {

sstable/Cargo.toml

@@ -0,0 +1,21 @@
+[package]
+name = "tantivy-sstable"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+common = {path="../common", package="tantivy-common"}
+ciborium = "0.2"
+serde = "1"
+tantivy-fst = "0.4"
+
+[dev-dependencies]
+proptest = "1"
+criterion = "0.4"
+names = "0.14"
+rand = "0.8"
+
+[[bench]]
+name = "stream_bench"
+harness = false

sstable/README.md

@@ -0,0 +1,28 @@
+# SSTable
+
+The `tantivy-sstable` crate is yet another sstable crate.
+
+It has been designed to be used in `quickwit`:
+- as an alternative to the default tantivy fst dictionary.
+- as a way to store the column index for dynamic fast fields.
+
+The benefit compared to the fst crate is locality.
+Searching a key in the fst crate requires downloading the entire dictionary.
+
+Once the sstable index is downloaded, running a `get` in the sstable
+crate only requires a single fetch.
+
+Right now, the block index and the default block size have been thought
+for quickwit, and the performance of a get is very bad.
+
+# Sorted strings?
+
+SSTable stands for Sorted String Table.
+Strings have to be insert in sorted order.
+
+That sorted order is used in different ways:
+- it makes gets and streaming ranges of keys
+possible.
+- it allows incremental encoding of the keys
+- the front compression is leveraged to optimize
+the intersection with an automaton

sstable/benches/stream_bench.rs

@@ -0,0 +1,87 @@
+use std::collections::BTreeSet;
+use std::io;
+
+use common::file_slice::FileSlice;
+use criterion::{criterion_group, criterion_main, Criterion};
+use rand::rngs::StdRng;
+use rand::{Rng, SeedableRng};
+use tantivy_sstable::{self, Dictionary, MonotonicU64SSTable};
+
+const CHARSET: &'static [u8] = b"abcdefghij";
+
+fn generate_key(rng: &mut impl Rng) -> String {
+    let len = rng.gen_range(3..12);
+    std::iter::from_fn(|| {
+        let idx = rng.gen_range(0..CHARSET.len());
+        Some(CHARSET[idx] as char)
+    })
+    .take(len)
+    .collect()
+}
+
+fn prepare_sstable() -> io::Result<Dictionary<MonotonicU64SSTable>> {
+    let mut rng = StdRng::from_seed([3u8; 32]);
+    let mut els = BTreeSet::new();
+    while els.len() < 100_000 {
+        els.insert(generate_key(&mut rng));
+    }
+    let mut dictionary_builder = Dictionary::<MonotonicU64SSTable>::builder(Vec::new())?;
+    for (ord, word) in els.iter().enumerate() {
+        dictionary_builder.insert(word, &(ord as u64))?;
+    }
+    let buffer = dictionary_builder.finish()?;
+    let dictionary = Dictionary::open(FileSlice::from(buffer))?;
+    Ok(dictionary)
+}
+
+fn stream_bench(
+    dictionary: &Dictionary<MonotonicU64SSTable>,
+    lower: &[u8],
+    upper: &[u8],
+    do_scan: bool,
+) -> usize {
+    let mut stream = dictionary
+        .range()
+        .ge(lower)
+        .lt(upper)
+        .into_stream()
+        .unwrap();
+    if !do_scan {
+        return 0;
+    }
+    let mut count = 0;
+    while stream.advance() {
+        count += 1;
+    }
+    count
+}
+
+pub fn criterion_benchmark(c: &mut Criterion) {
+    let dict = prepare_sstable().unwrap();
+    c.bench_function("short_scan_init", |b| {
+        b.iter(|| stream_bench(&dict, b"fa", b"fana", false))
+    });
+    c.bench_function("short_scan_init_and_scan", |b| {
+        b.iter(|| {
+            assert_eq!(stream_bench(&dict, b"fa", b"faz", true), 971);
+        })
+    });
+    c.bench_function("full_scan_init_and_scan_full_with_bound", |b| {
+        b.iter(|| {
+            assert_eq!(stream_bench(&dict, b"", b"z", true), 100_000);
+        })
+    });
+    c.bench_function("full_scan_init_and_scan_full_no_bounds", |b| {
+        b.iter(|| {
+            let mut stream = dict.stream().unwrap();
+            let mut count = 0;
+            while stream.advance() {
+                count += 1;
+            }
+            count
+        })
+    });
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);

sstable/src/block_reader.rs

@@ -1,6 +1,5 @@
-use std::io::{self, Read};
+use std::io;
-
+use std::ops::Range;
-use byteorder::{LittleEndian, ReadBytesExt};
 
 pub struct BlockReader<'a> {
     buffer: Vec<u8>,
@@ -8,6 +7,13 @@ pub struct BlockReader<'a> {
     offset: usize,
 }
 
+#[inline]
+fn read_u32(read: &mut dyn io::Read) -> io::Result<u32> {
+    let mut buf = [0u8; 4];
+    read.read_exact(&mut buf)?;
+    Ok(u32::from_le_bytes(buf))
+}
+
 impl<'a> BlockReader<'a> {
     pub fn new(reader: Box<dyn io::Read + 'a>) -> BlockReader<'a> {
         BlockReader {
@@ -24,13 +30,13 @@ impl<'a> BlockReader<'a> {
     }
 
     #[inline(always)]
-    pub fn buffer_from_to(&self, start: usize, end: usize) -> &[u8] {
+    pub fn buffer_from_to(&self, range: Range<usize>) -> &[u8] {
-        &self.buffer[start..end]
+        &self.buffer[range]
     }
 
     pub fn read_block(&mut self) -> io::Result<bool> {
         self.offset = 0;
-        let block_len_res = self.reader.read_u32::<LittleEndian>();
+        let block_len_res = read_u32(self.reader.as_mut());
         if let Err(err) = &block_len_res {
             if err.kind() == io::ErrorKind::UnexpectedEof {
                 return Ok(false);
@@ -46,14 +52,17 @@ impl<'a> BlockReader<'a> {
         Ok(true)
     }
 
+    #[inline(always)]
     pub fn offset(&self) -> usize {
         self.offset
     }
 
+    #[inline(always)]
     pub fn advance(&mut self, num_bytes: usize) {
         self.offset += num_bytes;
     }
 
+    #[inline(always)]
     pub fn buffer(&self) -> &[u8] {
         &self.buffer[self.offset..]
     }

sstable/src/delta.rs

@@ -16,6 +16,8 @@ where W: io::Write
     block: Vec<u8>,
     write: CountingWriter<BufWriter<W>>,
     value_writer: TValueWriter,
+    // Only here to avoid allocations.
+    stateless_buffer: Vec<u8>,
 }
 
 impl<W, TValueWriter> DeltaWriter<W, TValueWriter>
@@ -28,6 +30,7 @@ where
             block: Vec::with_capacity(BLOCK_LEN * 2),
             write: CountingWriter::wrap(BufWriter::new(wrt)),
             value_writer: TValueWriter::default(),
+            stateless_buffer: Vec::new(),
         }
     }
 }
@@ -42,15 +45,16 @@ where
             return Ok(None);
         }
         let start_offset = self.write.written_bytes() as usize;
-        // TODO avoid buffer allocation
+        let buffer: &mut Vec<u8> = &mut self.stateless_buffer;
-        let mut buffer = Vec::new();
+        self.value_writer.serialize_block(buffer);
-        self.value_writer.write_block(&mut buffer);
+        self.value_writer.clear();
         let block_len = buffer.len() + self.block.len();
         self.write.write_all(&(block_len as u32).to_le_bytes())?;
         self.write.write_all(&buffer[..])?;
         self.write.write_all(&self.block[..])?;
         let end_offset = self.write.written_bytes() as usize;
         self.block.clear();
+        buffer.clear();
         Ok(Some(start_offset..end_offset))
     }
 
@@ -84,15 +88,14 @@ where
         Ok(None)
     }
 
-    pub fn finalize(self) -> CountingWriter<BufWriter<W>> {
+    pub fn finish(self) -> CountingWriter<BufWriter<W>> {
         self.write
     }
 }
 
 pub struct DeltaReader<'a, TValueReader> {
     common_prefix_len: usize,
-    suffix_start: usize,
+    suffix_range: Range<usize>,
-    suffix_end: usize,
     value_reader: TValueReader,
     block_reader: BlockReader<'a>,
     idx: usize,
@@ -105,13 +108,16 @@ where TValueReader: value::ValueReader
         DeltaReader {
             idx: 0,
             common_prefix_len: 0,
-            suffix_start: 0,
+            suffix_range: 0..0,
-            suffix_end: 0,
             value_reader: TValueReader::default(),
             block_reader: BlockReader::new(Box::new(reader)),
         }
     }
 
+    pub fn empty() -> Self {
+        DeltaReader::new(&b""[..])
+    }
+
     fn deserialize_vint(&mut self) -> u64 {
         self.block_reader.deserialize_u64()
     }
@@ -140,15 +146,14 @@ where TValueReader: value::ValueReader
     }
 
     fn read_delta_key(&mut self) -> bool {
-        if let Some((keep, add)) = self.read_keep_add() {
+        let Some((keep, add)) = self.read_keep_add() else {
-            self.common_prefix_len = keep;
+            return false;
-            self.suffix_start = self.block_reader.offset();
+        };
-            self.suffix_end = self.suffix_start + add;
+        self.common_prefix_len = keep;
-            self.block_reader.advance(add);
+        let suffix_start = self.block_reader.offset();
-            true
+        self.suffix_range = suffix_start..(suffix_start + add);
-        } else {
+        self.block_reader.advance(add);
-            false
+        true
-        }
     }
 
     pub fn advance(&mut self) -> io::Result<bool> {
@@ -156,7 +161,8 @@ where TValueReader: value::ValueReader
             if !self.block_reader.read_block()? {
                 return Ok(false);
             }
-            self.value_reader.read(&mut self.block_reader)?;
+            let consumed_len = self.value_reader.load(self.block_reader.buffer())?;
+            self.block_reader.advance(consumed_len);
             self.idx = 0;
         } else {
             self.idx += 1;
@@ -167,16 +173,30 @@ where TValueReader: value::ValueReader
         Ok(true)
     }
 
+    #[inline(always)]
     pub fn common_prefix_len(&self) -> usize {
         self.common_prefix_len
     }
 
+    #[inline(always)]
     pub fn suffix(&self) -> &[u8] {
-        self.block_reader
+        self.block_reader.buffer_from_to(self.suffix_range.clone())
-            .buffer_from_to(self.suffix_start, self.suffix_end)
     }
 
+    #[inline(always)]
     pub fn value(&self) -> &TValueReader::Value {
         self.value_reader.value(self.idx)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::DeltaReader;
+    use crate::value::U64MonotonicValueReader;
+
+    #[test]
+    fn test_empty() {
+        let mut delta_reader: DeltaReader<U64MonotonicValueReader> = DeltaReader::empty();
+        assert!(!delta_reader.advance().unwrap());
+    }
+}

sstable/src/dictionary.rs

@@ -0,0 +1,261 @@
+use std::io;
+use std::marker::PhantomData;
+use std::ops::{Bound, RangeBounds};
+use std::sync::Arc;
+
+use common::file_slice::FileSlice;
+use common::{BinarySerializable, OwnedBytes};
+use tantivy_fst::automaton::AlwaysMatch;
+use tantivy_fst::Automaton;
+
+use crate::streamer::{Streamer, StreamerBuilder};
+use crate::{BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, TermOrdinal};
+
+/// An SSTable is a sorted map that associates sorted `&[u8]` keys
+/// to any kind of typed values.
+///
+/// The SSTable is organized in blocks.
+/// In each block, keys and values are encoded separately.
+///
+/// The keys are encoded using incremental encoding.
+/// The values on the other hand, are encoded according to a value-specific
+/// codec defined in the TSSTable generic argument.
+///
+/// Finally, an index is joined to the Dictionary to make it possible,
+/// given a key to identify which block contains this key.
+///
+/// The codec was designed in such a way that the sstable
+/// reader is not aware of block, and yet can read any sequence of blocks,
+/// as long as the slice of bytes it is given starts and stops at
+/// block boundary.
+///
+/// (See also README.md)
+pub struct Dictionary<TSSTable: SSTable> {
+    pub sstable_slice: FileSlice,
+    pub sstable_index: SSTableIndex,
+    num_terms: u64,
+    phantom_data: PhantomData<TSSTable>,
+}
+
+impl<TSSTable: SSTable> Dictionary<TSSTable> {
+    pub fn builder<W: io::Write>(wrt: W) -> io::Result<crate::Writer<W, TSSTable::ValueWriter>> {
+        Ok(TSSTable::writer(wrt))
+    }
+
+    pub(crate) fn sstable_reader(&self) -> io::Result<Reader<'static, TSSTable::ValueReader>> {
+        let data = self.sstable_slice.read_bytes()?;
+        Ok(TSSTable::reader(data))
+    }
+
+    pub(crate) fn sstable_reader_block(
+        &self,
+        block_addr: BlockAddr,
+    ) -> io::Result<Reader<'static, TSSTable::ValueReader>> {
+        let data = self.sstable_slice.read_bytes_slice(block_addr.byte_range)?;
+        Ok(TSSTable::reader(data))
+    }
+
+    pub(crate) async fn sstable_reader_block_async(
+        &self,
+        block_addr: BlockAddr,
+    ) -> io::Result<Reader<'static, TSSTable::ValueReader>> {
+        let data = self
+            .sstable_slice
+            .read_bytes_slice_async(block_addr.byte_range)
+            .await?;
+        Ok(TSSTable::reader(data))
+    }
+
+    pub(crate) fn sstable_delta_reader_for_key_range(
+        &self,
+        key_range: impl RangeBounds<[u8]>,
+    ) -> io::Result<DeltaReader<'static, TSSTable::ValueReader>> {
+        let slice = self.file_slice_for_range(key_range);
+        let data = slice.read_bytes()?;
+        Ok(TSSTable::delta_reader(data))
+    }
+
+    /// This function returns a file slice covering a set of sstable blocks
+    /// that include the key range passed in arguments.
+    ///
+    /// It works by identifying
+    /// - `first_block`: the block containing the start boudary key
+    /// - `last_block`: the block containing the end boundary key.
+    ///
+    /// And then returning the range that spans over all blocks between.
+    /// and including first_block and last_block, aka:
+    /// `[first_block.start_offset .. last_block.end_offset)`
+    ///
+    /// Technically this function does not provide the tightest fit, as
+    /// for simplification, it treats the start bound of the `key_range`
+    /// as if it was inclusive, even if it is exclusive.
+    /// On the rare edge case where a user asks for `(start_key, end_key]`
+    /// and `start_key` happens to be the last key of a block, we return a
+    /// slice that is the first block was not necessary.
+    fn file_slice_for_range(&self, key_range: impl RangeBounds<[u8]>) -> FileSlice {
+        let start_bound: Bound<usize> = match key_range.start_bound() {
+            Bound::Included(key) | Bound::Excluded(key) => {
+                let Some(first_block_addr) = self.sstable_index.search_block(key) else {
+                    return FileSlice::empty();
+                };
+                Bound::Included(first_block_addr.byte_range.start)
+            }
+            Bound::Unbounded => Bound::Unbounded,
+        };
+        let end_bound: Bound<usize> = match key_range.end_bound() {
+            Bound::Included(key) | Bound::Excluded(key) => {
+                if let Some(block_addr) = self.sstable_index.search_block(key) {
+                    Bound::Excluded(block_addr.byte_range.end)
+                } else {
+                    Bound::Unbounded
+                }
+            }
+            Bound::Unbounded => Bound::Unbounded,
+        };
+        self.sstable_slice.slice((start_bound, end_bound))
+    }
+
+    /// Opens a `TermDictionary`.
+    pub fn open(term_dictionary_file: FileSlice) -> io::Result<Self> {
+        let (main_slice, footer_len_slice) = term_dictionary_file.split_from_end(16);
+        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 (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.as_slice())
+            .map_err(|_| io::Error::new(io::ErrorKind::InvalidData, "SSTable corruption"))?;
+        Ok(Dictionary {
+            sstable_slice,
+            sstable_index,
+            num_terms,
+            phantom_data: PhantomData,
+        })
+    }
+
+    /// Creates a term dictionary from the supplied bytes.
+    pub fn from_bytes(owned_bytes: OwnedBytes) -> io::Result<Self> {
+        Dictionary::open(FileSlice::new(Arc::new(owned_bytes)))
+    }
+
+    /// Creates an empty term dictionary which contains no terms.
+    pub fn empty() -> Self {
+        let term_dictionary_data: Vec<u8> = Self::builder(Vec::<u8>::new())
+            .expect("Creating a TermDictionaryBuilder in a Vec<u8> should never fail")
+            .finish()
+            .expect("Writing in a Vec<u8> should never fail");
+        let empty_dict_file = FileSlice::from(term_dictionary_data);
+        Dictionary::open(empty_dict_file).unwrap()
+    }
+
+    /// Returns the number of terms in the dictionary.
+    /// Term ordinals range from 0 to `num_terms() - 1`.
+    pub fn num_terms(&self) -> usize {
+        self.num_terms as usize
+    }
+
+    /// Returns the ordinal associated with a given term.
+    pub fn term_ord<K: AsRef<[u8]>>(&self, key: K) -> io::Result<Option<TermOrdinal>> {
+        let mut term_ord = 0u64;
+        let key_bytes = key.as_ref();
+        let mut sstable_reader = self.sstable_reader()?;
+        while sstable_reader.advance().unwrap_or(false) {
+            if sstable_reader.key() == key_bytes {
+                return Ok(Some(term_ord));
+            }
+            term_ord += 1;
+        }
+        Ok(None)
+    }
+
+    /// Returns the term associated with a given term ordinal.
+    ///
+    /// Term ordinals are defined as the position of the term in
+    /// the sorted list of terms.
+    ///
+    /// Returns true if and only if the term has been found.
+    ///
+    /// Regardless of whether the term is found or not,
+    /// the buffer may be modified.
+    pub fn ord_to_term(&self, ord: TermOrdinal, bytes: &mut Vec<u8>) -> io::Result<bool> {
+        let mut sstable_reader = self.sstable_reader()?;
+        bytes.clear();
+        for _ in 0..(ord + 1) {
+            if !sstable_reader.advance().unwrap_or(false) {
+                return Ok(false);
+            }
+        }
+        bytes.extend_from_slice(sstable_reader.key());
+        Ok(true)
+    }
+
+    /// Returns the number of terms in the dictionary.
+    pub fn term_info_from_ord(&self, term_ord: TermOrdinal) -> io::Result<Option<TSSTable::Value>> {
+        let mut sstable_reader = self.sstable_reader()?;
+        for _ in 0..(term_ord + 1) {
+            if !sstable_reader.advance().unwrap_or(false) {
+                return Ok(None);
+            }
+        }
+        Ok(Some(sstable_reader.value().clone()))
+    }
+
+    /// Lookups the value corresponding to the key.
+    pub fn get<K: AsRef<[u8]>>(&self, key: K) -> io::Result<Option<TSSTable::Value>> {
+        if let Some(block_addr) = self.sstable_index.search_block(key.as_ref()) {
+            let mut sstable_reader = self.sstable_reader_block(block_addr)?;
+            let key_bytes = key.as_ref();
+            while sstable_reader.advance().unwrap_or(false) {
+                if sstable_reader.key() == key_bytes {
+                    let value = sstable_reader.value().clone();
+                    return Ok(Some(value));
+                }
+            }
+        }
+        Ok(None)
+    }
+
+    /// Lookups the value corresponding to the key.
+    pub async fn get_async<K: AsRef<[u8]>>(&self, key: K) -> io::Result<Option<TSSTable::Value>> {
+        if let Some(block_addr) = self.sstable_index.search_block(key.as_ref()) {
+            let mut sstable_reader = self.sstable_reader_block_async(block_addr).await?;
+            let key_bytes = key.as_ref();
+            while sstable_reader.advance().unwrap_or(false) {
+                if sstable_reader.key() == key_bytes {
+                    let value = sstable_reader.value().clone();
+                    return Ok(Some(value));
+                }
+            }
+        }
+        Ok(None)
+    }
+
+    /// Returns a range builder, to stream all of the terms
+    /// within an interval.
+    pub fn range(&self) -> StreamerBuilder<'_, TSSTable> {
+        StreamerBuilder::new(self, AlwaysMatch)
+    }
+
+    /// A stream of all the sorted terms.
+    pub fn stream(&self) -> io::Result<Streamer<'_, TSSTable>> {
+        self.range().into_stream()
+    }
+
+    /// Returns a search builder, to stream all of the terms
+    /// within the Automaton
+    pub fn search<'a, A: Automaton + 'a>(
+        &'a self,
+        automaton: A,
+    ) -> StreamerBuilder<'a, TSSTable, A>
+    where
+        A::State: Clone,
+    {
+        StreamerBuilder::<TSSTable, A>::new(self, automaton)
+    }
+
+    #[doc(hidden)]
+    pub async fn warm_up_dictionary(&self) -> io::Result<()> {
+        self.sstable_slice.read_bytes_async().await?;
+        Ok(())
+    }
+}