Added solution to force the type of a column.

Remove standard deviation from stats aggregation

Cargo.toml

@@ -15,7 +15,7 @@ rust-version = "1.62"
 
 [dependencies]
 oneshot = "0.1.5"
-base64 = "0.20.0"
+base64 = "0.21.0"
 byteorder = "1.4.3"
 crc32fast = "1.3.2"
 once_cell = "1.10.0"
@@ -48,7 +48,7 @@ murmurhash32 = "0.2.0"
 time = { version = "0.3.10", features = ["serde-well-known"] }
 smallvec = "1.8.0"
 rayon = "1.5.2"
-lru = "0.7.5"
+lru = "0.9.0"
 fastdivide = "0.4.0"
 itertools = "0.10.3"
 measure_time = "0.8.2"
@@ -61,6 +61,7 @@ tantivy-query-grammar = { version= "0.19.0", path="./query-grammar" }
 tantivy-bitpacker = 		{ version= "0.3", path="./bitpacker" }
 common = 								{ version= "0.5", path = "./common/", package = "tantivy-common" }
 fastfield_codecs = 			{ version= "0.3", path="./fastfield_codecs", default-features = false }
+tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
 
 [target.'cfg(windows)'.dependencies]
 winapi = "0.3.9"
@@ -106,7 +107,7 @@ unstable = [] # useful for benches.
 quickwit = ["sstable"]
 
 [workspace]
-members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable"]
+members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable", "tokenizer-api"]
 
 # Following the "fail" crate best practises, we isolate
 # tests that define specific behavior in fail check points

README.md

@@ -29,7 +29,7 @@ Your mileage WILL vary depending on the nature of queries and their load.
 # Features
 
 - Full-text search
-- Configurable tokenizer (stemming available for 17 Latin languages with third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)), Japanese ([lindera](https://github.com/lindera-morphology/lindera-tantivy), [Vaporetto](https://crates.io/crates/vaporetto_tantivy), and [tantivy-tokenizer-tiny-segmenter](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)) and Korean ([lindera](https://github.com/lindera-morphology/lindera-tantivy) + [lindera-ko-dic-builder](https://github.com/lindera-morphology/lindera-ko-dic-builder))
+- Configurable tokenizer (stemming available for 17 Latin languages) with third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)), Japanese ([lindera](https://github.com/lindera-morphology/lindera-tantivy), [Vaporetto](https://crates.io/crates/vaporetto_tantivy), and [tantivy-tokenizer-tiny-segmenter](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)) and Korean ([lindera](https://github.com/lindera-morphology/lindera-tantivy) + [lindera-ko-dic-builder](https://github.com/lindera-morphology/lindera-ko-dic-builder))
 - Fast (check out the :racehorse: :sparkles: [benchmark](https://tantivy-search.github.io/bench/) :sparkles: :racehorse:)
 - Tiny startup time (<10ms), perfect for command-line tools
 - BM25 scoring (the same as Lucene)
@@ -42,12 +42,12 @@ Your mileage WILL vary depending on the nature of queries and their load.
 - Single valued and multivalued u64, i64, and f64 fast fields (equivalent of doc values in Lucene)
 - `&[u8]` fast fields
 - Text, i64, u64, f64, dates, and hierarchical facet fields
-- LZ4 compressed document store
+- Compressed document store (LZ4, Zstd, None, Brotli, Snap)
 - Range queries
 - Faceted search
 - Configurable indexing (optional term frequency and position indexing)
 - JSON Field
-- Aggregation Collector: range buckets, average, and stats metrics
+- Aggregation Collector: histogram, range buckets, average, and stats metrics
 - LogMergePolicy with deletes
 - Searcher Warmer API
 - Cheesy logo with a horse
@@ -81,6 +81,10 @@ There are many ways to support this project.
 
 We use the GitHub Pull Request workflow: reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
 
+## Tokenizer
+
+When implementing a tokenizer for tantivy depend on the `tantivy-tokenizer-api` crate.
+
 ## Minimum supported Rust version
 
 Tantivy currently requires at least Rust 1.62 or later to compile.

columnar/Cargo.toml

@@ -0,0 +1,32 @@
+[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" }
+common = { path = "../common", package = "tantivy-common" }
+itertools = "0.10"
+log = "0.4"
+tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
+prettytable-rs = {version="0.10.0", optional= true}
+rand = {version="0.8.3", optional= true}
+fastdivide = "0.4"
+measure_time = { version="0.8.2", optional=true}
+
+[dev-dependencies]
+proptest = "1"
+more-asserts = "0.3.0"
+rand = "0.8.3"
+
+# temporary
+[workspace]
+members = []
+
+[features]
+unstable = []

columnar/README.md

@@ -0,0 +1,109 @@
+# Columnar format
+
+This crate describes columnar format used in tantivy.
+
+## Goals
+
+This format is special in the following way.
+- it needs to be compact
+- accessing a specific column does not require to load the entire columnar. It can be done in 2 to 3 random access.
+- 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.
+- it is designed to allow range queries.
+
+# Coercion rules
+
+Users can create a columnar by inserting rows to a `ColumnarWriter`,
+and serializing it into a `Write` object.
+Nothing prevents a user from recording values with different type to the same `column_name`.
+
+In that case, `tantivy-columnar`'s behavior is as follows:
+- JsonValues are grouped into 3 types (String, Number, bool).
+Values that corresponds to different groups 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
+
+This columnar format may have more than one column (with different types) associated to the same `column_name` (see [Coercion rules](#coercion-rules) above).
+The `(column_name, columne_type)` couple however uniquely identifies a column.
+That couple is serialized as a column `column_key`.  The format of that key is:
+`[column_name][ZERO_BYTE][column_type_header: u8]`
+
+```
+COLUMNAR:=
+    [COLUMNAR_DATA]
+    [COLUMNAR_KEY_TO_DATA_INDEX]
+    [COLUMNAR_FOOTER];
+
+
+# Columns are sorted by their column key.
+COLUMNAR_DATA:=
+    [COLUMN_DATA]+;
+
+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 sstable associates
+`(column name, column_cardinality, column_type) to range of bytes.
+
+Column name may not contain the zero byte `\0`.
+
+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
+
+This crate exposes a columnar format for tantivy.
+This format is described in README.md
+
+
+The crate introduces the following concepts.
+
+`Columnar` is an equivalent of a dataframe.
+It maps `column_key` to `Column`.
+
+A `Column<T>` asssociates a `RowId` (u32) to any
+number of values.
+
+This is made possible by wrapping a `ColumnIndex` and a `ColumnValue` object.
+The `ColumnValue<T>` represents a mapping that associates each `RowId` to
+exactly one single value.
+
+The `ColumnIndex` then maps each RowId to a set of `RowId` in the
+`ColumnValue`.
+
+For optimization, and compression purposes, the `ColumnIndex` has three
+possible representation, each for different cardinalities.
+
+- Full
+
+All RowId have exactly one value. The ColumnIndex is the trivial mapping.
+
+- Optional
+
+All RowIds can have at most one value. The ColumnIndex is the trivial mapping `ColumnRowId -> Option<ColumnValueRowId>`.
+
+- Multivalued
+
+All RowIds can have any number of values.
+The column index is mapping values to a range.
+
+
+All these objects are implemented an unit tested independently
+in their own module:
+
+- columnar
+- column_index
+- column_values
+- column

columnar/src/TODO.md

@@ -0,0 +1,45 @@
+# zero to one
+* merges
+* full still needs a num_values
+* replug u128
+* add dictionary encoded stuff
+* fix multivalued
+* find a way to make columnar work with strict types
+* plug to tantivy
+    - indexing
+    - aggregations
+    - merge
+
+# Perf and Size
+* re-add ZSTD compression for dictionaries
+no systematic monotonic mapping
+consider removing multilinear
+f32?
+adhoc solution for bool?
+
+add metrics helper for aggregate. sum(row_id)
+review inline absence/presence
+improv perf of select using PDEP
+compare with roaring bitmap/elias fano etc etc.
+SIMD range? (see blog post)
+Add alignment?
+Consider another codec to bridge the gap between few and 5k elements
+
+# Cleanup and rationalization
+in benchmark, unify percent vs ratio, f32 vs f64.
+investigate if should have better errors? io::Error is overused at the moment.
+rename rank/select in unit tests
+Review the public API via cargo doc
+go through TODOs
+remove all  doc_id occurences -> row_id
+use the rank & select naming in unit tests branch.
+multi-linear -> blockwise
+linear codec -> simply a multiplication for the index column
+
+# Other
+fix enhance column-cli
+
+# Santa claus
+
+autodetect datetime ipaddr, plug customizable tokenizer.
+

columnar/src/column/dictionary_encoded.rs

@@ -0,0 +1,40 @@
+use std::io;
+use std::ops::Deref;
+use std::sync::Arc;
+
+use sstable::{Dictionary, VoidSSTable};
+
+use crate::column::Column;
+use crate::column_index::ColumnIndex;
+
+/// Dictionary encoded column.
+#[derive(Clone)]
+pub struct BytesColumn {
+    pub(crate) dictionary: Arc<Dictionary<VoidSSTable>>,
+    pub(crate) term_ord_column: Column<u64>,
+}
+
+impl BytesColumn {
+    /// Returns `false` if the term does not exist (e.g. `term_ord` is greater or equal to the
+    /// overll number of terms).
+    pub fn term_ord_to_str(&self, term_ord: u64, output: &mut Vec<u8>) -> io::Result<bool> {
+        self.dictionary.ord_to_term(term_ord, output)
+    }
+
+    pub fn term_ords(&self) -> &Column<u64> {
+        &self.term_ord_column
+    }
+}
+
+impl Deref for BytesColumn {
+    type Target = ColumnIndex<'static>;
+
+    fn deref(&self) -> &Self::Target {
+        &**self.term_ords()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::{ColumnarReader, ColumnarWriter};
+}

columnar/src/column/mod.rs

@@ -0,0 +1,56 @@
+mod dictionary_encoded;
+mod serialize;
+
+use std::ops::Deref;
+use std::sync::Arc;
+
+use common::BinarySerializable;
+pub use dictionary_encoded::BytesColumn;
+pub use serialize::{open_column_bytes, open_column_u64, serialize_column_u64};
+
+use crate::column_index::ColumnIndex;
+use crate::column_values::ColumnValues;
+use crate::{Cardinality, RowId};
+
+#[derive(Clone)]
+pub struct Column<T> {
+    pub idx: ColumnIndex<'static>,
+    pub values: Arc<dyn ColumnValues<T>>,
+}
+
+use crate::column_index::Set;
+
+impl<T: PartialOrd> Column<T> {
+    pub fn first(&self, row_id: RowId) -> Option<T> {
+        match &self.idx {
+            ColumnIndex::Full => Some(self.values.get_val(row_id)),
+            ColumnIndex::Optional(opt_idx) => {
+                let value_row_idx = opt_idx.rank_if_exists(row_id)?;
+                Some(self.values.get_val(value_row_idx))
+            }
+            ColumnIndex::Multivalued(_multivalued_index) => {
+                todo!();
+            }
+        }
+    }
+}
+
+impl<T> Deref for Column<T> {
+    type Target = ColumnIndex<'static>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.idx
+    }
+}
+
+impl BinarySerializable for Cardinality {
+    fn serialize<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
+        self.to_code().serialize(writer)
+    }
+
+    fn deserialize<R: std::io::Read>(reader: &mut R) -> std::io::Result<Self> {
+        let cardinality_code = u8::deserialize(reader)?;
+        let cardinality = Cardinality::try_from_code(cardinality_code)?;
+        Ok(cardinality)
+    }
+}

columnar/src/column/serialize.rs

@@ -0,0 +1,54 @@
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::{CountingWriter, OwnedBytes};
+use sstable::Dictionary;
+
+use crate::column::{BytesColumn, Column};
+use crate::column_index::{serialize_column_index, SerializableColumnIndex};
+use crate::column_values::{
+    serialize_column_values, ColumnValues, MonotonicallyMappableToU64, ALL_CODEC_TYPES,
+};
+pub fn serialize_column_u64<T: MonotonicallyMappableToU64>(
+    column_index: SerializableColumnIndex<'_>,
+    column_values: &impl ColumnValues<T>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let mut counting_writer = CountingWriter::wrap(output);
+    serialize_column_index(column_index, &mut counting_writer)?;
+    let column_index_num_bytes = counting_writer.written_bytes() as u32;
+    let output = counting_writer.finish();
+    serialize_column_values(column_values, &ALL_CODEC_TYPES[..], output)?;
+    output.write_all(&column_index_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Column<T>> {
+    let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
+    let column_index_num_bytes = u32::from_le_bytes(
+        column_index_num_bytes_payload
+            .as_slice()
+            .try_into()
+            .unwrap(),
+    );
+    let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
+    let column_index = crate::column_index::open_column_index(column_index_data)?;
+    let column_values = crate::column_values::open_u64_mapped(column_values_data)?;
+    Ok(Column {
+        idx: column_index,
+        values: column_values,
+    })
+}
+
+pub fn open_column_bytes(data: OwnedBytes) -> io::Result<BytesColumn> {
+    let (body, dictionary_len_bytes) = data.rsplit(4);
+    let dictionary_len = u32::from_le_bytes(dictionary_len_bytes.as_slice().try_into().unwrap());
+    let (dictionary_bytes, column_bytes) = body.split(dictionary_len as usize);
+    let dictionary = Arc::new(Dictionary::from_bytes(dictionary_bytes)?);
+    let term_ord_column = crate::column::open_column_u64::<u64>(column_bytes)?;
+    Ok(BytesColumn {
+        dictionary,
+        term_ord_column,
+    })
+}

columnar/src/column_index/mod.rs

@@ -0,0 +1,40 @@
+mod multivalued_index;
+mod optional_index;
+mod serialize;
+
+use std::sync::Arc;
+
+pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
+pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
+
+use crate::column_values::ColumnValues;
+use crate::{Cardinality, RowId};
+
+#[derive(Clone)]
+pub enum ColumnIndex<'a> {
+    Full,
+    Optional(OptionalIndex),
+    // TODO remove the Arc<dyn> apart from serialization this is not
+    // dynamic at all.
+    Multivalued(Arc<dyn ColumnValues<RowId> + 'a>),
+}
+
+impl<'a> ColumnIndex<'a> {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            ColumnIndex::Full => Cardinality::Full,
+            ColumnIndex::Optional(_) => Cardinality::Optional,
+            ColumnIndex::Multivalued(_) => Cardinality::Multivalued,
+        }
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        match self {
+            ColumnIndex::Full => {
+                todo!()
+            }
+            ColumnIndex::Optional(optional_index) => optional_index.num_rows(),
+            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.num_vals() - 1,
+        }
+    }
+}

columnar/src/column_index/multivalued_index.rs

@@ -0,0 +1,27 @@
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+
+use crate::column_values::{ColumnValues, FastFieldCodecType};
+use crate::RowId;
+
+#[derive(Clone)]
+pub struct MultivaluedIndex(Arc<dyn ColumnValues<RowId>>);
+
+pub fn serialize_multivalued_index(
+    multivalued_index: MultivaluedIndex,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    crate::column_values::serialize_column_values(
+        &*multivalued_index.0,
+        &[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
+        output,
+    )?;
+    Ok(())
+}
+
+pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<Arc<dyn ColumnValues<RowId>>> {
+    todo!();
+}

columnar/src/column_index/optional_index/mod.rs

@@ -0,0 +1,453 @@
+use std::io::{self, Write};
+use std::ops::Range;
+use std::sync::Arc;
+
+mod set;
+mod set_block;
+
+use common::{BinarySerializable, GroupByIteratorExtended, OwnedBytes, VInt};
+pub use set::{Set, SetCodec};
+use set_block::{
+    DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
+};
+
+use crate::{InvalidData, RowId};
+
+/// The threshold for for number of elements after which we switch to dense block encoding.
+///
+/// We simply pick the value that minimize the size of the blocks.
+const DENSE_BLOCK_THRESHOLD: u32 =
+    set_block::DENSE_BLOCK_NUM_BYTES / std::mem::size_of::<u16>() as u32; //< 5_120
+
+const ELEMENTS_PER_BLOCK: u32 = u16::MAX as u32 + 1;
+
+const BLOCK_SIZE: RowId = 1 << 16;
+
+#[derive(Copy, Clone, Debug)]
+struct BlockMeta {
+    non_null_rows_before_block: u32,
+    start_byte_offset: u32,
+    block_variant: BlockVariant,
+}
+
+#[derive(Clone, Copy, Debug)]
+enum BlockVariant {
+    Dense,
+    Sparse { num_vals: u16 },
+}
+
+impl BlockVariant {
+    pub fn empty() -> Self {
+        Self::Sparse { num_vals: 0 }
+    }
+    pub fn num_bytes_in_block(&self) -> u32 {
+        match *self {
+            BlockVariant::Dense => set_block::DENSE_BLOCK_NUM_BYTES,
+            BlockVariant::Sparse { num_vals } => num_vals as u32 * 2,
+        }
+    }
+}
+
+/// This codec is inspired by roaring bitmaps.
+/// In the dense blocks, however, in order to accelerate `select`
+/// we interleave an offset over two bytes. (more on this lower)
+///
+/// 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.
+#[derive(Clone)]
+pub struct OptionalIndex {
+    num_rows: RowId,
+    num_non_null_rows: RowId,
+    block_data: OwnedBytes,
+    block_metas: Arc<[BlockMeta]>,
+}
+
+impl OptionalIndex {
+    pub fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
+    pub fn num_non_nulls(&self) -> RowId {
+        self.num_non_null_rows
+    }
+}
+
+/// 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(Copy, Debug, Clone)]
+struct RowAddr {
+    block_id: u16,
+    in_block_row_id: u16,
+}
+
+#[inline(always)]
+fn row_addr_from_row_id(row_id: RowId) -> RowAddr {
+    RowAddr {
+        block_id: (row_id / BLOCK_SIZE) as u16,
+        in_block_row_id: (row_id % BLOCK_SIZE) as u16,
+    }
+}
+
+impl Set<RowId> for OptionalIndex {
+    // Check if value at position is not null.
+    #[inline]
+    fn contains(&self, row_id: RowId) -> bool {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(row_id);
+        let block_meta = self.block_metas[block_id as usize];
+        match self.block(block_meta) {
+            Block::Dense(dense_block) => dense_block.contains(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.contains(in_block_row_id),
+        }
+    }
+
+    #[inline]
+    fn rank_if_exists(&self, row_id: RowId) -> Option<RowId> {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(row_id);
+        let block_meta = self.block_metas[block_id as usize];
+        let block = self.block(block_meta);
+        let block_offset_row_id = match block {
+            Block::Dense(dense_block) => dense_block.rank_if_exists(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank_if_exists(in_block_row_id),
+        }? as u32;
+        Some(block_meta.non_null_rows_before_block + block_offset_row_id)
+    }
+
+    #[inline]
+    fn select(&self, rank: RowId) -> RowId {
+        let block_pos = self.find_block(rank, 0);
+        let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
+        let block_meta = self.block_metas[block_pos as usize];
+        let block: Block<'_> = self.block(block_meta);
+        let index_in_block = (rank - block_meta.non_null_rows_before_block) as u16;
+        let in_block_rank = match block {
+            Block::Dense(dense_block) => dense_block.select(index_in_block),
+            Block::Sparse(sparse_block) => sparse_block.select(index_in_block),
+        };
+        block_doc_idx_start + in_block_rank as u32
+    }
+
+    fn select_batch(&self, ranks: &[u32], output_idxs: &mut [u32]) {
+        let mut block_pos = 0u32;
+        let mut start = 0;
+        let group_by_it = ranks.iter().copied().group_by(move |codec_idx| {
+            block_pos = self.find_block(*codec_idx, block_pos);
+            block_pos
+        });
+        for (block_pos, block_iter) in group_by_it {
+            let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
+            let block_meta = self.block_metas[block_pos as usize];
+            let block: Block<'_> = self.block(block_meta);
+            let offset = block_meta.non_null_rows_before_block;
+            let indexes_in_block_iter =
+                block_iter.map(move |codec_idx| (codec_idx - offset) as u16);
+            match block {
+                Block::Dense(dense_block) => {
+                    for in_offset in dense_block.select_iter(indexes_in_block_iter) {
+                        output_idxs[start] = in_offset as u32 + block_doc_idx_start;
+                        start += 1;
+                    }
+                }
+                Block::Sparse(sparse_block) => {
+                    for in_offset in sparse_block.select_iter(indexes_in_block_iter) {
+                        output_idxs[start] = in_offset as u32 + block_doc_idx_start;
+                        start += 1;
+                    }
+                }
+            };
+        }
+    }
+}
+
+impl OptionalIndex {
+    #[inline]
+    fn block<'a>(&'a self, block_meta: BlockMeta) -> Block<'a> {
+        let BlockMeta {
+            start_byte_offset,
+            block_variant,
+            ..
+        } = block_meta;
+        let start_byte_offset = start_byte_offset as usize;
+        let bytes = self.block_data.as_slice();
+        match block_variant {
+            BlockVariant::Dense => Block::Dense(DenseBlockCodec::open(
+                &bytes[start_byte_offset..start_byte_offset + DENSE_BLOCK_NUM_BYTES as usize],
+            )),
+            BlockVariant::Sparse { num_vals } => {
+                let end_byte_offset = start_byte_offset + num_vals as usize * 2;
+                let sparse_bytes = &bytes[start_byte_offset..end_byte_offset];
+                Block::Sparse(SparseBlockCodec::open(sparse_bytes))
+            }
+        }
+    }
+
+    #[inline]
+    fn find_block(&self, dense_idx: u32, start_block_pos: u32) -> u32 {
+        for block_pos in start_block_pos..self.block_metas.len() as u32 {
+            let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
+            if offset > dense_idx {
+                return block_pos - 1;
+            }
+        }
+        self.block_metas.len() as u32 - 1u32
+    }
+
+    // TODO Add a good API for the codec_idx to original_idx translation.
+    // The Iterator API is a probably a bad idea
+}
+
+#[derive(Copy, Clone)]
+enum Block<'a> {
+    Dense(DenseBlock<'a>),
+    Sparse(SparseBlock<'a>),
+}
+
+#[derive(Debug, Copy, Clone)]
+enum OptionalIndexCodec {
+    Dense = 0,
+    Sparse = 1,
+}
+
+impl OptionalIndexCodec {
+    fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        match code {
+            0 => Ok(Self::Dense),
+            1 => Ok(Self::Sparse),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+impl BinarySerializable for OptionalIndexCodec {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&[self.to_code()])
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let optional_codec_code = u8::deserialize(reader)?;
+        let optional_codec = Self::try_from_code(optional_codec_code)?;
+        Ok(optional_codec)
+    }
+}
+
+fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -> io::Result<()> {
+    let is_sparse = is_sparse(block_els.len() as u32);
+    if is_sparse {
+        SparseBlockCodec::serialize(block_els.iter().copied(), out)?;
+    } else {
+        DenseBlockCodec::serialize(block_els.iter().copied(), out)?;
+    }
+    Ok(())
+}
+
+pub fn serialize_optional_index<'a, W: io::Write>(
+    serializable_optional_index: &dyn SerializableOptionalIndex<'a>,
+    output: &mut W,
+) -> io::Result<()> {
+    VInt(serializable_optional_index.num_rows() as u64).serialize(output)?;
+
+    let mut rows_it = serializable_optional_index.non_null_rows();
+    let mut block_metadata: Vec<SerializedBlockMeta> = 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.
+    let Some(idx) = rows_it.next() else {
+        output.write_all(&0u16.to_le_bytes())?;
+        return Ok(());
+    };
+
+    let row_addr = row_addr_from_row_id(idx);
+
+    let mut current_block_id = row_addr.block_id;
+    current_block.push(row_addr.in_block_row_id);
+
+    for idx in rows_it {
+        let value_addr = row_addr_from_row_id(idx);
+        if current_block_id != value_addr.block_id {
+            serialize_optional_index_block(&current_block[..], output)?;
+            block_metadata.push(SerializedBlockMeta {
+                block_id: current_block_id,
+                num_non_null_rows: current_block.len() as u32,
+            });
+            current_block.clear();
+            current_block_id = value_addr.block_id;
+        }
+        current_block.push(value_addr.in_block_row_id);
+    }
+
+    // handle last block
+    serialize_optional_index_block(&current_block[..], output)?;
+
+    block_metadata.push(SerializedBlockMeta {
+        block_id: current_block_id,
+        num_non_null_rows: current_block.len() as u32,
+    });
+
+    for block in &block_metadata {
+        output.write_all(&block.to_bytes())?;
+    }
+
+    output.write_all((block_metadata.len() as u16).to_le_bytes().as_ref())?;
+
+    Ok(())
+}
+
+const SERIALIZED_BLOCK_META_NUM_BYTES: usize = 4;
+
+#[derive(Clone, Copy, Debug)]
+struct SerializedBlockMeta {
+    block_id: u16,
+    num_non_null_rows: u32, //< takes values in 1..=u16::MAX
+}
+
+// TODO unit tests
+impl SerializedBlockMeta {
+    #[inline]
+    fn from_bytes(bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES]) -> SerializedBlockMeta {
+        let block_id = u16::from_le_bytes(bytes[0..2].try_into().unwrap());
+        let num_non_null_rows: u32 =
+            u16::from_le_bytes(bytes[2..4].try_into().unwrap()) as u32 + 1u32;
+        SerializedBlockMeta {
+            block_id,
+            num_non_null_rows,
+        }
+    }
+
+    #[inline]
+    fn to_bytes(&self) -> [u8; SERIALIZED_BLOCK_META_NUM_BYTES] {
+        assert!(self.num_non_null_rows > 0);
+        let mut bytes = [0u8; SERIALIZED_BLOCK_META_NUM_BYTES];
+        bytes[0..2].copy_from_slice(&self.block_id.to_le_bytes());
+        // 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
+        bytes[2..4].copy_from_slice(&((self.num_non_null_rows - 1u32) as u16).to_le_bytes());
+        bytes
+    }
+}
+
+#[inline]
+fn is_sparse(num_rows_in_block: u32) -> bool {
+    num_rows_in_block < DENSE_BLOCK_THRESHOLD as u32
+}
+
+fn deserialize_optional_index_block_metadatas(
+    data: &[u8],
+    num_rows: u32,
+) -> (Box<[BlockMeta]>, u32) {
+    let num_blocks = data.len() / SERIALIZED_BLOCK_META_NUM_BYTES;
+    let mut block_metas = Vec::with_capacity(num_blocks as usize + 1);
+    let mut start_byte_offset = 0;
+    let mut non_null_rows_before_block = 0;
+    for block_meta_bytes in data.chunks_exact(SERIALIZED_BLOCK_META_NUM_BYTES) {
+        let block_meta_bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES] =
+            block_meta_bytes.try_into().unwrap();
+        let SerializedBlockMeta {
+            block_id,
+            num_non_null_rows,
+        } = SerializedBlockMeta::from_bytes(block_meta_bytes);
+        block_metas.resize(
+            block_id as usize,
+            BlockMeta {
+                non_null_rows_before_block,
+                start_byte_offset,
+                block_variant: BlockVariant::empty(),
+            },
+        );
+        let block_variant = if is_sparse(num_non_null_rows) {
+            BlockVariant::Sparse {
+                num_vals: num_non_null_rows as u16,
+            }
+        } else {
+            BlockVariant::Dense
+        };
+        block_metas.push(BlockMeta {
+            non_null_rows_before_block,
+            start_byte_offset,
+            block_variant,
+        });
+        start_byte_offset += block_variant.num_bytes_in_block();
+        non_null_rows_before_block += num_non_null_rows as u32;
+    }
+    block_metas.resize(
+        ((num_rows + BLOCK_SIZE - 1) / BLOCK_SIZE) as usize,
+        BlockMeta {
+            non_null_rows_before_block,
+            start_byte_offset,
+            block_variant: BlockVariant::empty(),
+        },
+    );
+    (block_metas.into_boxed_slice(), non_null_rows_before_block)
+}
+
+pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
+    let (mut bytes, num_non_empty_blocks_bytes) = bytes.rsplit(2);
+    let num_non_empty_block_bytes =
+        u16::from_le_bytes(num_non_empty_blocks_bytes.as_slice().try_into().unwrap());
+    let num_rows = VInt::deserialize_u64(&mut bytes)? as u32;
+    let block_metas_num_bytes =
+        num_non_empty_block_bytes as usize * SERIALIZED_BLOCK_META_NUM_BYTES;
+    let (block_data, block_metas) = bytes.rsplit(block_metas_num_bytes);
+    let (block_metas, num_non_null_rows) =
+        deserialize_optional_index_block_metadatas(block_metas.as_slice(), num_rows).into();
+    let optional_index = OptionalIndex {
+        num_rows,
+        num_non_null_rows,
+        block_data,
+        block_metas: block_metas.into(),
+    };
+    Ok(optional_index)
+}
+
+pub trait SerializableOptionalIndex<'a> {
+    fn num_rows(&self) -> RowId;
+    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a>;
+}
+
+impl SerializableOptionalIndex<'static> for Range<u32> {
+    fn num_rows(&self) -> RowId {
+        self.end
+    }
+    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'static> {
+        Box::new(self.clone())
+    }
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_index/optional_index/set.rs

@@ -0,0 +1,38 @@
+use std::io;
+
+/// A codec makes it possible to serialize a set of
+/// elements, and open the resulting Set representation.
+pub trait SetCodec {
+    type Item: Copy + TryFrom<usize> + Eq + std::hash::Hash + std::fmt::Debug;
+    type Reader<'a>: Set<Self::Item>;
+
+    /// Serializes a set of unique sorted u16 elements.
+    ///
+    /// May panic if the elements are not sorted.
+    fn serialize(els: impl Iterator<Item = Self::Item>, wrt: impl io::Write) -> io::Result<()>;
+    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a>;
+}
+
+pub trait Set<T> {
+    /// Returns true if the elements is contained in the Set
+    fn contains(&self, el: T) -> bool;
+
+    /// If the set contains `el` returns its position in the sortd set of elements.
+    /// If the set does not contain the element, it returns `None`.
+    fn rank_if_exists(&self, el: T) -> Option<T>;
+
+    /// Return the rank-th value stored in this bitmap.
+    ///
+    /// # Panics
+    ///
+    /// May panic if rank is greater than the number of elements in the Set.
+    fn select(&self, rank: T) -> T;
+
+    /// Batch version of select.
+    /// `ranks` is assumed to be sorted.
+    ///
+    /// # Panics
+    ///
+    /// May panic if rank is greater than the number of elements in the Set.
+    fn select_batch(&self, ranks: &[T], outputs: &mut [T]);
+}

columnar/src/column_index/optional_index/set_block/mod.rs

@@ -0,0 +1,8 @@
+mod set_block;
+mod sparse;
+
+pub use set_block::{DenseBlock, DenseBlockCodec, DENSE_BLOCK_NUM_BYTES};
+pub use sparse::{SparseBlock, SparseBlockCodec};
+
+#[cfg(test)]
+mod tests;

columnar/src/column_index/optional_index/set_block/set_block.rs

@@ -0,0 +1,271 @@
+use std::convert::TryInto;
+use std::io::{self, Write};
+
+use common::BinarySerializable;
+
+use crate::column_index::optional_index::{Set, SetCodec, ELEMENTS_PER_BLOCK};
+
+#[inline(always)]
+fn get_bit_at(input: u64, n: u16) -> bool {
+    input & (1 << n) != 0
+}
+
+#[inline]
+fn set_bit_at(input: &mut u64, n: u16) {
+    *input |= 1 << n;
+}
+
+/// 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.
+
+const ELEMENTS_PER_MINI_BLOCK: u16 = 64;
+const MINI_BLOCK_BITVEC_NUM_BYTES: usize = 8;
+const MINI_BLOCK_OFFSET_NUM_BYTES: usize = 2;
+pub const MINI_BLOCK_NUM_BYTES: usize = MINI_BLOCK_BITVEC_NUM_BYTES + MINI_BLOCK_OFFSET_NUM_BYTES;
+
+/// Number of bytes in a dense block.
+pub const DENSE_BLOCK_NUM_BYTES: u32 =
+    (ELEMENTS_PER_BLOCK as u32 / ELEMENTS_PER_MINI_BLOCK as u32) * MINI_BLOCK_NUM_BYTES as u32;
+
+pub struct DenseBlockCodec;
+
+impl SetCodec for DenseBlockCodec {
+    type Item = u16;
+    type Reader<'a> = DenseBlock<'a>;
+
+    fn serialize(els: impl Iterator<Item = u16>, wrt: impl io::Write) -> io::Result<()> {
+        serialize_dense_codec(els, wrt)
+    }
+
+    #[inline]
+    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
+        assert_eq!(data.len(), DENSE_BLOCK_NUM_BYTES as usize);
+        DenseBlock(data)
+    }
+}
+
+/// Interpreting the bitvec as a set of integer within 0..=63
+/// and given an element, returns the number of elements in the
+/// set lesser than the element.
+///
+/// # Panics
+///
+/// May panic or return a wrong result if el <= 64.
+#[inline(always)]
+fn rank_u64(bitvec: u64, el: u16) -> u16 {
+    debug_assert!(el < 64);
+    let mask = (1u64 << el) - 1;
+    let masked_bitvec = bitvec & mask;
+    masked_bitvec.count_ones() as u16
+}
+
+#[inline(always)]
+fn select_u64(mut bitvec: u64, rank: u16) -> u16 {
+    for _ in 0..rank {
+        bitvec &= bitvec - 1;
+    }
+    bitvec.trailing_zeros() as u16
+}
+
+// TODO test the following solution on Intel... on Ryzen Zen <3 it is a catastrophy.
+// #[target_feature(enable = "bmi2")]
+// unsafe fn select_bitvec_unsafe(bitvec: u64, rank: u16) -> u16 {
+//     let pdep = _pdep_u64(1u64 << rank, bitvec);
+//     pdep.trailing_zeros() as u16
+// }
+
+#[derive(Clone, Copy, Debug)]
+struct DenseMiniBlock {
+    bitvec: u64,
+    rank: u16,
+}
+
+impl DenseMiniBlock {
+    fn from_bytes(data: [u8; MINI_BLOCK_NUM_BYTES]) -> Self {
+        let bitvec = u64::from_le_bytes(data[..MINI_BLOCK_BITVEC_NUM_BYTES].try_into().unwrap());
+        let rank = u16::from_le_bytes(data[MINI_BLOCK_BITVEC_NUM_BYTES..].try_into().unwrap());
+        Self { bitvec, rank }
+    }
+
+    fn to_bytes(&self) -> [u8; MINI_BLOCK_NUM_BYTES] {
+        let mut bytes = [0u8; MINI_BLOCK_NUM_BYTES];
+        bytes[..MINI_BLOCK_BITVEC_NUM_BYTES].copy_from_slice(&self.bitvec.to_le_bytes());
+        bytes[MINI_BLOCK_BITVEC_NUM_BYTES..].copy_from_slice(&self.rank.to_le_bytes());
+        bytes
+    }
+}
+
+#[derive(Copy, Clone)]
+pub struct DenseBlock<'a>(&'a [u8]);
+
+impl<'a> Set<u16> for DenseBlock<'a> {
+    #[inline(always)]
+    fn contains(&self, el: u16) -> bool {
+        let mini_block_id = el / ELEMENTS_PER_MINI_BLOCK;
+        let bitvec = self.mini_block(mini_block_id).bitvec;
+        let pos_in_bitvec = el % ELEMENTS_PER_MINI_BLOCK;
+        get_bit_at(bitvec, pos_in_bitvec)
+    }
+
+    #[inline(always)]
+    fn rank_if_exists(&self, el: u16) -> Option<u16> {
+        let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
+        let index_block = self.mini_block(block_pos);
+        let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
+        let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
+        let rank = index_block.rank + ones_in_block;
+        if get_bit_at(index_block.bitvec, pos_in_block_bit_vec) {
+            Some(rank)
+        } else {
+            None
+        }
+    }
+
+    #[inline(always)]
+    fn select(&self, rank: u16) -> u16 {
+        let block_id = self.find_miniblock_containing_rank(rank, 0).unwrap();
+        let index_block = self.mini_block(block_id);
+        let in_block_rank = rank - index_block.rank;
+        block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+    }
+
+    fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
+        let orig_ids = self.select_iter(ranks.iter().copied());
+        for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
+            *output = original_id;
+        }
+    }
+}
+
+impl<'a> DenseBlock<'a> {
+    /// Iterator verison of select.
+    ///
+    /// # Panics
+    /// Panics if one of the rank is higher than the number of elements in the set.
+    pub fn select_iter<'b>(
+        &self,
+        rank_it: impl Iterator<Item = u16> + 'b,
+    ) -> impl Iterator<Item = u16> + 'b
+    where
+        Self: 'b,
+    {
+        let mut block_id = 0u16;
+        let me = *self;
+        rank_it.map(move |rank| {
+            block_id = me.find_miniblock_containing_rank(rank, block_id).unwrap();
+            let index_block = me.mini_block(block_id);
+            let in_block_rank = rank - index_block.rank;
+            block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+        })
+    }
+}
+
+impl<'a> DenseBlock<'a> {
+    #[inline]
+    fn mini_block(&self, mini_block_id: u16) -> DenseMiniBlock {
+        let data_start_pos = mini_block_id as usize * MINI_BLOCK_NUM_BYTES;
+        DenseMiniBlock::from_bytes(
+            self.0[data_start_pos..data_start_pos + MINI_BLOCK_NUM_BYTES]
+                .try_into()
+                .unwrap(),
+        )
+    }
+
+    #[inline]
+    fn iter_miniblocks(
+        &self,
+        from_block_id: u16,
+    ) -> impl Iterator<Item = (u16, DenseMiniBlock)> + '_ {
+        self.0
+            .chunks_exact(MINI_BLOCK_NUM_BYTES)
+            .enumerate()
+            .skip(from_block_id as usize)
+            .map(|(block_id, bytes)| {
+                let mini_block = DenseMiniBlock::from_bytes(bytes.try_into().unwrap());
+                (block_id as u16, mini_block)
+            })
+    }
+
+    /// 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.
+    #[inline]
+    fn find_miniblock_containing_rank(&self, rank: u16, from_block_id: u16) -> Option<u16> {
+        self.iter_miniblocks(from_block_id)
+            .take_while(|(_, block)| block.rank <= rank)
+            .map(|(block_id, _)| block_id)
+            .last()
+    }
+}
+
+/// Iterator over all values, true if set, otherwise false
+pub fn serialize_dense_codec(
+    els: impl Iterator<Item = u16>,
+    mut output: impl Write,
+) -> io::Result<()> {
+    let mut non_null_rows_before: u16 = 0u16;
+    let mut block = 0u64;
+    let mut current_block_id = 0u16;
+    for el in els {
+        let block_id = el / ELEMENTS_PER_MINI_BLOCK;
+        let in_offset = el % ELEMENTS_PER_MINI_BLOCK;
+        while block_id > current_block_id {
+            let dense_mini_block = DenseMiniBlock {
+                bitvec: block,
+                rank: non_null_rows_before as u16,
+            };
+            output.write_all(&dense_mini_block.to_bytes())?;
+            non_null_rows_before += block.count_ones() as u16;
+            block = 0u64;
+            current_block_id += 1u16;
+        }
+        set_bit_at(&mut block, in_offset);
+    }
+    while current_block_id <= u16::MAX / ELEMENTS_PER_MINI_BLOCK {
+        block.serialize(&mut output)?;
+        non_null_rows_before.serialize(&mut output)?;
+        // This will overflow to 0 exactly if all bits are set.
+        // This is however not problem as we won't use this last value.
+        non_null_rows_before = non_null_rows_before.wrapping_add(block.count_ones() as u16);
+        block = 0u64;
+        current_block_id += 1u16;
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_select_bitvec() {
+        assert_eq!(select_u64(1u64, 0), 0);
+        assert_eq!(select_u64(2u64, 0), 1);
+        assert_eq!(select_u64(4u64, 0), 2);
+        assert_eq!(select_u64(8u64, 0), 3);
+        assert_eq!(select_u64(1 | 8u64, 0), 0);
+        assert_eq!(select_u64(1 | 8u64, 1), 3);
+    }
+
+    #[test]
+    fn test_count_ones() {
+        for i in 0..=63 {
+            assert_eq!(rank_u64(u64::MAX, i), i);
+        }
+    }
+
+    #[test]
+    fn test_dense() {
+        assert_eq!(DENSE_BLOCK_NUM_BYTES, 10_240);
+    }
+}

columnar/src/column_index/optional_index/set_block/sparse.rs

@@ -0,0 +1,112 @@
+use crate::column_index::optional_index::{Set, SetCodec};
+
+pub struct SparseBlockCodec;
+
+impl SetCodec for SparseBlockCodec {
+    type Item = u16;
+    type Reader<'a> = SparseBlock<'a>;
+
+    fn serialize(
+        els: impl Iterator<Item = u16>,
+        mut wrt: impl std::io::Write,
+    ) -> std::io::Result<()> {
+        for el in els {
+            wrt.write_all(&el.to_le_bytes())?;
+        }
+        Ok(())
+    }
+
+    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
+        SparseBlock(data)
+    }
+}
+
+#[derive(Copy, Clone)]
+pub struct SparseBlock<'a>(&'a [u8]);
+
+impl<'a> Set<u16> for SparseBlock<'a> {
+    #[inline(always)]
+    fn contains(&self, el: u16) -> bool {
+        self.binary_search(el).is_ok()
+    }
+
+    #[inline(always)]
+    fn rank_if_exists(&self, el: u16) -> Option<u16> {
+        self.binary_search(el).ok()
+    }
+
+    #[inline(always)]
+    fn select(&self, rank: u16) -> u16 {
+        let offset = rank as usize * 2;
+        u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
+    }
+
+    fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
+        let orig_ids = self.select_iter(ranks.iter().copied());
+        for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
+            *output = original_id;
+        }
+    }
+}
+
+#[inline(always)]
+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)
+}
+
+impl<'a> SparseBlock<'a> {
+    #[inline(always)]
+    fn value_at_idx(&self, data: &[u8], idx: u16) -> u16 {
+        let start_offset: usize = idx as usize * 2;
+        get_u16(data, start_offset)
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u16 {
+        (self.0.len() / 2) as u16
+    }
+
+    #[inline]
+    #[allow(clippy::comparison_chain)]
+    // Looks for the element in the block. Returns the positions if found.
+    fn binary_search(&self, target: u16) -> Result<u16, u16> {
+        let data = &self.0;
+        let mut size = self.num_vals();
+        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 Ok(mid);
+            }
+
+            size = right - left;
+        }
+        Err(left)
+    }
+
+    pub fn select_iter<'b>(
+        &self,
+        iter: impl Iterator<Item = u16> + 'b,
+    ) -> impl Iterator<Item = u16> + 'b
+    where
+        Self: 'b,
+    {
+        iter.map(|codec_id| {
+            let offset = codec_id as usize * 2;
+            u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
+        })
+    }
+}

columnar/src/column_index/optional_index/set_block/tests.rs

@@ -0,0 +1,110 @@
+use std::collections::HashMap;
+
+use crate::column_index::optional_index::set_block::set_block::DENSE_BLOCK_NUM_BYTES;
+use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
+use crate::column_index::optional_index::{Set, SetCodec};
+
+fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
+    let mut buffer = Vec::new();
+    C::serialize(vals.iter().copied(), &mut buffer).unwrap();
+    let tested_set = C::open(buffer.as_slice());
+    let hash_set: HashMap<C::Item, C::Item> = vals
+        .iter()
+        .copied()
+        .enumerate()
+        .map(|(ord, val)| (val, C::Item::try_from(ord).ok().unwrap()))
+        .collect();
+    for val in 0u16..=u16::MAX {
+        assert_eq!(tested_set.contains(val), hash_set.contains_key(&val));
+        assert_eq!(tested_set.rank_if_exists(val), hash_set.get(&val).copied());
+    }
+    for rank in 0..vals.len() {
+        assert_eq!(tested_set.select(rank as u16), vals[rank]);
+    }
+    buffer.len()
+}
+
+#[test]
+fn test_dense_block_set_u16_empty() {
+    let buffer_len = test_set_helper::<DenseBlockCodec>(&[]);
+    assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+}
+
+#[test]
+fn test_dense_block_set_u16_max() {
+    let buffer_len = test_set_helper::<DenseBlockCodec>(&[u16::MAX]);
+    assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+}
+
+#[test]
+fn test_sparse_block_set_u16_empty() {
+    let buffer_len = test_set_helper::<SparseBlockCodec>(&[]);
+    assert_eq!(buffer_len, 0);
+}
+
+#[test]
+fn test_sparse_block_set_u16_max() {
+    let buffer_len = test_set_helper::<SparseBlockCodec>(&[u16::MAX]);
+    assert_eq!(buffer_len, 2);
+}
+
+use proptest::prelude::*;
+
+proptest! {
+    #[test]
+    fn test_prop_test_dense(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
+        let vals: Vec<u16> = els.into_iter().collect();
+        let buffer_len = test_set_helper::<DenseBlockCodec>(&vals);
+        assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
+    }
+
+    #[test]
+    fn test_prop_test_sparse(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
+        let vals: Vec<u16> = els.into_iter().collect();
+        let buffer_len = test_set_helper::<SparseBlockCodec>(&vals);
+        assert_eq!(buffer_len, vals.len() * 2);
+    }
+}
+
+#[test]
+fn test_simple_translate_codec_codec_idx_to_original_idx_dense() {
+    let mut buffer = Vec::new();
+    DenseBlockCodec::serialize([1, 3, 17, 32, 30_000, 30_001].iter().copied(), &mut buffer)
+        .unwrap();
+    let tested_set = DenseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    assert_eq!(
+        &tested_set
+            .select_iter([0, 1, 2, 5].iter().copied())
+            .collect::<Vec<u16>>(),
+        &[1, 3, 17, 30_001]
+    );
+}
+
+#[test]
+fn test_simple_translate_codec_idx_to_original_idx_sparse() {
+    let mut buffer = Vec::new();
+    SparseBlockCodec::serialize([1, 3, 17].iter().copied(), &mut buffer).unwrap();
+    let tested_set = SparseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    assert_eq!(
+        &tested_set
+            .select_iter([0, 1, 2].iter().copied())
+            .collect::<Vec<u16>>(),
+        &[1, 3, 17]
+    );
+}
+
+#[test]
+fn test_simple_translate_codec_idx_to_original_idx_dense() {
+    let mut buffer = Vec::new();
+    DenseBlockCodec::serialize(0u16..150u16, &mut buffer).unwrap();
+    let tested_set = DenseBlockCodec::open(buffer.as_slice());
+    assert!(tested_set.contains(1));
+    let rg = 0u16..150u16;
+    let els: Vec<u16> = rg.clone().collect();
+    assert_eq!(
+        &tested_set.select_iter(rg.clone()).collect::<Vec<u16>>(),
+        &els
+    );
+}

columnar/src/column_index/optional_index/tests.rs

@@ -0,0 +1,327 @@
+use proptest::prelude::{any, prop, *};
+use proptest::strategy::Strategy;
+use proptest::{prop_oneof, proptest};
+
+use super::*;
+
+#[test]
+fn test_dense_block_threshold() {
+    assert_eq!(super::DENSE_BLOCK_THRESHOLD, 5_120);
+}
+
+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 test_with_random_sets_simple() {
+    let vals = 10..BLOCK_SIZE * 2;
+    let mut out: Vec<u8> = Vec::new();
+    serialize_optional_index(&vals.clone(), &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
+    let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
+    let mut output = vec![0u32; ranks.len()];
+    null_index.select_batch(&ranks[..], &mut output[..]);
+    assert_eq!(&output, &els);
+}
+
+#[test]
+fn test_optional_index_trailing_empty_blocks() {
+    test_null_index(&[false]);
+}
+
+#[test]
+fn test_optional_index_one_block_false() {
+    let mut iter = vec![false; ELEMENTS_PER_BLOCK as usize];
+    iter.push(true);
+    test_null_index(&iter[..]);
+}
+
+#[test]
+fn test_optional_index_one_block_true() {
+    let mut iter = vec![true; ELEMENTS_PER_BLOCK as usize];
+    iter.push(true);
+    test_null_index(&iter[..]);
+}
+
+impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
+    fn num_rows(&self) -> RowId {
+        self.len() as u32
+    }
+
+    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(
+            self.iter()
+                .cloned()
+                .enumerate()
+                .filter(|(_pos, val)| *val)
+                .map(|(pos, _val)| pos as u32),
+        )
+    }
+}
+
+fn test_null_index(data: &[bool]) {
+    let mut out: Vec<u8> = Vec::new();
+    serialize_optional_index(&data, &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    let orig_idx_with_value: Vec<u32> = data
+        .iter()
+        .enumerate()
+        .filter(|(_pos, val)| **val)
+        .map(|(pos, _val)| pos as u32)
+        .collect();
+    let ids: Vec<u32> = (0..orig_idx_with_value.len() as u32).collect();
+    let mut output = vec![0u32; ids.len()];
+    null_index.select_batch(&ids[..], &mut output);
+    // assert_eq!(&output[0..100], &orig_idx_with_value[0..100]);
+    assert_eq!(output, 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.rank_if_exists(*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.contains(pos as u32), *value);
+    }
+}
+
+#[test]
+fn test_optional_index_test_translation() {
+    let mut out = vec![];
+    let iter = &[true, false, true, false];
+    serialize_optional_index(&&iter[..], &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    let mut output = vec![0u32; 2];
+    null_index.select_batch(&[0, 1], &mut output);
+    assert_eq!(output, &[0, 2]);
+}
+
+#[test]
+fn test_optional_index_translate() {
+    let mut out = vec![];
+    let iter = &[true, false, true, false];
+    serialize_optional_index(&&iter[..], &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    assert_eq!(null_index.rank_if_exists(0), Some(0));
+    assert_eq!(null_index.rank_if_exists(2), Some(1));
+}
+
+#[test]
+fn test_optional_index_small() {
+    let mut out = vec![];
+    let iter = &[true, false, true, false];
+    serialize_optional_index(&&iter[..], &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    assert!(null_index.contains(0));
+    assert!(!null_index.contains(1));
+    assert!(null_index.contains(2));
+    assert!(!null_index.contains(3));
+}
+
+#[test]
+fn test_optional_index_large() {
+    let mut docs = vec![];
+    docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
+    docs.extend((0..=1).map(|_idx| true));
+
+    let mut out = vec![];
+    serialize_optional_index(&&docs[..], &mut out).unwrap();
+    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
+    assert!(!null_index.contains(0));
+    assert!(!null_index.contains(100));
+    assert!(!null_index.contains(ELEMENTS_PER_BLOCK - 1));
+    assert!(null_index.contains(ELEMENTS_PER_BLOCK));
+    assert!(null_index.contains(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) -> OptionalIndex {
+        let mut out = Vec::new();
+        let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
+        let vals: Vec<bool> = (0..TOTAL_NUM_VALUES)
+            .map(|_| rng.gen_bool(fill_ratio))
+            .collect();
+        serialize_optional_index(&&vals[..], &mut out).unwrap();
+
+        let codec = open_optional_index(OwnedBytes::new(out)).unwrap();
+        codec
+    }
+
+    fn random_range_iterator(
+        start: u32,
+        end: u32,
+        avg_step_size: u32,
+        avg_deviation: 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(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
+            if current >= end {
+                None
+            } else {
+                Some(current)
+            }
+        })
+    }
+
+    fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
+        let ratio = percent as f32 / 100.0;
+        let step_size = (1f32 / ratio) as u32;
+        let deviation = step_size - 1;
+        random_range_iterator(0, num_values, step_size, deviation)
+    }
+
+    fn walk_over_data(codec: &OptionalIndex, avg_step_size: u32) -> Option<u32> {
+        walk_over_data_from_positions(
+            codec,
+            random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
+        )
+    }
+
+    fn walk_over_data_from_positions(
+        codec: &OptionalIndex,
+        positions: impl Iterator<Item = u32>,
+    ) -> Option<u32> {
+        let mut dense_idx: Option<u32> = None;
+        for idx in positions {
+            dense_idx = dense_idx.or(codec.rank_if_exists(idx));
+        }
+        dense_idx
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 1000));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.5f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 0.005f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_10percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.1f64, 0.005f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 10f32, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.01f64, 100f32, bench);
+    }
+
+    fn bench_translate_codec_to_orig_util(
+        percent_filled: f64,
+        percent_hit: f32,
+        bench: &mut Bencher,
+    ) {
+        let codec = gen_bools(percent_filled);
+        let num_non_nulls = codec.num_non_nulls();
+        let idxs: Vec<u32> = if percent_hit == 100.0f32 {
+            (0..num_non_nulls).collect()
+        } else {
+            n_percent_step_iterator(percent_hit, num_non_nulls).collect()
+        };
+        let mut output = vec![0u32; idxs.len()];
+        bench.iter(|| {
+            codec.select_batch(&idxs[..], &mut output);
+        });
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.9f64, 0.005, bench);
+    }
+
+    #[bench]
+    fn bench_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
+        bench_translate_codec_to_orig_util(0.9f64, 100.0f32, bench);
+    }
+}

columnar/src/column_index/serialize.rs

@@ -0,0 +1,70 @@
+use std::io;
+use std::io::Write;
+
+use common::OwnedBytes;
+
+use crate::column_index::multivalued_index::{serialize_multivalued_index, MultivaluedIndex};
+use crate::column_index::optional_index::serialize_optional_index;
+use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
+use crate::Cardinality;
+
+pub enum SerializableColumnIndex<'a> {
+    Full,
+    Optional(Box<dyn SerializableOptionalIndex<'a> + 'a>),
+    // TODO remove the Arc<dyn> apart from serialization this is not
+    // dynamic at all.
+    Multivalued(MultivaluedIndex),
+}
+
+impl<'a> SerializableColumnIndex<'a> {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            SerializableColumnIndex::Full => Cardinality::Full,
+            SerializableColumnIndex::Optional(_) => Cardinality::Optional,
+            SerializableColumnIndex::Multivalued(_) => Cardinality::Multivalued,
+        }
+    }
+}
+
+pub fn serialize_column_index(
+    column_index: SerializableColumnIndex,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let cardinality = column_index.get_cardinality().to_code();
+    output.write_all(&[cardinality])?;
+    match column_index {
+        SerializableColumnIndex::Full => {}
+        SerializableColumnIndex::Optional(optional_index) => {
+            serialize_optional_index(&*optional_index, output)?
+        }
+        SerializableColumnIndex::Multivalued(multivalued_index) => {
+            serialize_multivalued_index(multivalued_index, output)?
+        }
+    }
+    Ok(())
+}
+
+pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex<'static>> {
+    if bytes.is_empty() {
+        return Err(io::Error::new(
+            io::ErrorKind::UnexpectedEof,
+            "Failed to deserialize column index. Empty buffer.",
+        ));
+    }
+    let cardinality_code = bytes[0];
+    let cardinality = Cardinality::try_from_code(cardinality_code)?;
+    bytes.advance(1);
+    match cardinality {
+        Cardinality::Full => Ok(ColumnIndex::Full),
+        Cardinality::Optional => {
+            let optional_index = super::optional_index::open_optional_index(bytes)?;
+            Ok(ColumnIndex::Optional(optional_index))
+        }
+        Cardinality::Multivalued => {
+            let multivalued_index = super::multivalued_index::open_multivalued_index(bytes)?;
+            Ok(ColumnIndex::Multivalued(multivalued_index))
+        }
+    }
+}
+
+// TODO unit tests

columnar/src/column_values/bitpacked.rs

@@ -0,0 +1,115 @@
+use std::io::{self, Write};
+
+use common::OwnedBytes;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use super::serialize::NormalizedHeader;
+use super::{ColumnValues, FastFieldCodec, FastFieldCodecType};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct BitpackedReader {
+    data: OwnedBytes,
+    bit_unpacker: BitUnpacker,
+    normalized_header: NormalizedHeader,
+}
+
+impl ColumnValues for BitpackedReader {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u64 {
+        self.bit_unpacker.get(doc, &self.data)
+    }
+    #[inline]
+    fn min_value(&self) -> u64 {
+        // The BitpackedReader assumes a normalized vector.
+        0
+    }
+    #[inline]
+    fn max_value(&self) -> u64 {
+        self.normalized_header.max_value
+    }
+    #[inline]
+    fn num_vals(&self) -> u32 {
+        self.normalized_header.num_vals
+    }
+}
+
+pub struct BitpackedCodec;
+
+impl FastFieldCodec for BitpackedCodec {
+    /// The CODEC_TYPE is an enum value used for serialization.
+    const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::Bitpacked;
+
+    type Reader = BitpackedReader;
+
+    /// Opens a fast field given a file.
+    fn open_from_bytes(
+        data: OwnedBytes,
+        normalized_header: NormalizedHeader,
+    ) -> io::Result<Self::Reader> {
+        let num_bits = compute_num_bits(normalized_header.max_value);
+        let bit_unpacker = BitUnpacker::new(num_bits);
+        Ok(BitpackedReader {
+            data,
+            bit_unpacker,
+            normalized_header,
+        })
+    }
+
+    /// Serializes data with the BitpackedFastFieldSerializer.
+    ///
+    /// The bitpacker assumes that the column has been normalized.
+    /// i.e. It has already been shifted by its minimum value, so that its
+    /// current minimum value is 0.
+    ///
+    /// Ideally, we made a shift upstream on the column so that `col.min_value() == 0`.
+    fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()> {
+        assert_eq!(column.min_value(), 0u64);
+        let num_bits = compute_num_bits(column.max_value());
+        let mut bit_packer = BitPacker::new();
+        for val in column.iter() {
+            bit_packer.write(val, num_bits, write)?;
+        }
+        bit_packer.close(write)?;
+        Ok(())
+    }
+
+    fn estimate(column: &dyn ColumnValues) -> Option<f32> {
+        let num_bits = compute_num_bits(column.max_value());
+        let num_bits_uncompressed = 64;
+        Some(num_bits as f32 / num_bits_uncompressed as f32)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::tests::create_and_validate;
+
+    fn create_and_validate_bitpacked_codec(data: &[u64], name: &str) {
+        create_and_validate::<BitpackedCodec>(data, name);
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate_bitpacked_codec(&data, name);
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn bitpacked_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate_bitpacked_codec(&data, "rand");
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/blockwise_linear.rs

@@ -0,0 +1,188 @@
+use std::sync::Arc;
+use std::{io, iter};
+
+use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::line::Line;
+use crate::column_values::serialize::NormalizedHeader;
+use crate::column_values::{ColumnValues, FastFieldCodec, FastFieldCodecType, VecColumn};
+
+const CHUNK_SIZE: usize = 512;
+
+#[derive(Debug, Default)]
+struct Block {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+    data_start_offset: usize,
+}
+
+impl BinarySerializable for Block {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Block {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+            data_start_offset: 0,
+        })
+    }
+}
+
+fn compute_num_blocks(num_vals: u32) -> usize {
+    (num_vals as usize + CHUNK_SIZE - 1) / CHUNK_SIZE
+}
+
+pub struct BlockwiseLinearCodec;
+
+impl FastFieldCodec for BlockwiseLinearCodec {
+    const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::BlockwiseLinear;
+    type Reader = BlockwiseLinearReader;
+
+    fn open_from_bytes(
+        bytes: common::OwnedBytes,
+        normalized_header: NormalizedHeader,
+    ) -> io::Result<Self::Reader> {
+        let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
+        let footer_offset = bytes.len() - 4 - footer_len as usize;
+        let (data, mut footer) = bytes.split(footer_offset);
+        let num_blocks = compute_num_blocks(normalized_header.num_vals);
+        let mut blocks: Vec<Block> = iter::repeat_with(|| Block::deserialize(&mut footer))
+            .take(num_blocks)
+            .collect::<io::Result<_>>()?;
+
+        let mut start_offset = 0;
+        for block in &mut blocks {
+            block.data_start_offset = start_offset;
+            start_offset += (block.bit_unpacker.bit_width() as usize) * CHUNK_SIZE / 8;
+        }
+        Ok(BlockwiseLinearReader {
+            blocks: Arc::new(blocks),
+            data,
+            normalized_header,
+        })
+    }
+
+    // Estimate first_chunk and extrapolate
+    fn estimate(column: &dyn ColumnValues) -> Option<f32> {
+        if column.num_vals() < 10 * CHUNK_SIZE as u32 {
+            return None;
+        }
+        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);
+            *buffer_val = buffer_val.wrapping_sub(interpolated_val);
+        }
+        let estimated_bit_width = first_chunk
+            .iter()
+            .map(|el| ((el + 1) as f32 * 3.0) as u64)
+            .map(compute_num_bits)
+            .max()
+            .unwrap();
+
+        let metadata_per_block = {
+            let mut out = vec![];
+            Block::default().serialize(&mut out).unwrap();
+            out.len()
+        };
+        let num_bits = estimated_bit_width as u64 * column.num_vals() as u64
+            // function metadata per block
+            + metadata_per_block as u64 * (column.num_vals() as u64 / CHUNK_SIZE as u64);
+        let num_bits_uncompressed = 64 * column.num_vals();
+        Some(num_bits as f32 / num_bits_uncompressed as f32)
+    }
+
+    fn serialize(column: &dyn ColumnValues, wrt: &mut impl io::Write) -> io::Result<()> {
+        // The BitpackedReader assumes a normalized vector.
+        assert_eq!(column.min_value(), 0);
+        let mut buffer = Vec::with_capacity(CHUNK_SIZE);
+        let num_vals = column.num_vals();
+
+        let num_blocks = compute_num_blocks(num_vals);
+        let mut blocks = Vec::with_capacity(num_blocks);
+
+        let mut vals = column.iter();
+
+        let mut bit_packer = BitPacker::new();
+
+        for _ in 0..num_blocks {
+            buffer.clear();
+            buffer.extend((&mut vals).take(CHUNK_SIZE));
+            let line = Line::train(&VecColumn::from(&buffer));
+
+            assert!(!buffer.is_empty());
+
+            for (i, buffer_val) in buffer.iter_mut().enumerate() {
+                let interpolated_val = line.eval(i as u32);
+                *buffer_val = buffer_val.wrapping_sub(interpolated_val);
+            }
+            let bit_width = buffer.iter().copied().map(compute_num_bits).max().unwrap();
+
+            for &buffer_val in &buffer {
+                bit_packer.write(buffer_val, bit_width, wrt)?;
+            }
+
+            blocks.push(Block {
+                line,
+                bit_unpacker: BitUnpacker::new(bit_width),
+                data_start_offset: 0,
+            });
+        }
+
+        bit_packer.close(wrt)?;
+
+        assert_eq!(blocks.len(), compute_num_blocks(num_vals));
+
+        let mut counting_wrt = CountingWriter::wrap(wrt);
+        for block in &blocks {
+            block.serialize(&mut counting_wrt)?;
+        }
+        let footer_len = counting_wrt.written_bytes();
+        (footer_len as u32).serialize(&mut counting_wrt)?;
+
+        Ok(())
+    }
+}
+
+#[derive(Clone)]
+pub struct BlockwiseLinearReader {
+    blocks: Arc<Vec<Block>>,
+    normalized_header: NormalizedHeader,
+    data: OwnedBytes,
+}
+
+impl ColumnValues for BlockwiseLinearReader {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> u64 {
+        let block_id = (idx / CHUNK_SIZE as u32) as usize;
+        let idx_within_block = idx % (CHUNK_SIZE as u32);
+        let block = &self.blocks[block_id];
+        let interpoled_val: u64 = block.line.eval(idx_within_block);
+        let block_bytes = &self.data[block.data_start_offset..];
+        let bitpacked_diff = block.bit_unpacker.get(idx_within_block, block_bytes);
+        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
+    }
+}

columnar/src/column_values/column.rs

@@ -0,0 +1,350 @@
+use std::marker::PhantomData;
+use std::ops::{Range, RangeInclusive};
+
+use tantivy_bitpacker::minmax;
+
+use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
+
+/// `ColumnValues` provides access to a dense field column.
+///
+/// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
+pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
+    /// Return the value associated with the given idx.
+    ///
+    /// This accessor should return as fast as possible.
+    ///
+    /// # Panics
+    ///
+    /// May panic if `idx` is greater than the column length.
+    fn get_val(&self, idx: u32) -> T;
+
+    /// Fills an output buffer with the fast field values
+    /// associated with the `DocId` going from
+    /// `start` to `start + output.len()`.
+    ///
+    /// # Panics
+    ///
+    /// Must panic if `start + output.len()` is greater than
+    /// the segment's `maxdoc`.
+    #[inline]
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        for (out, idx) in output.iter_mut().zip(start..) {
+            *out = self.get_val(idx as u32);
+        }
+    }
+
+    /// Get the positions of values which are in the provided value range.
+    ///
+    /// Note that position == docid for single value fast fields
+    #[inline]
+    fn get_docids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        let doc_id_range = doc_id_range.start..doc_id_range.end.min(self.num_vals());
+
+        for idx in doc_id_range.start..doc_id_range.end {
+            let val = self.get_val(idx);
+            if value_range.contains(&val) {
+                positions.push(idx);
+            }
+        }
+    }
+
+    /// Returns the minimum value for this fast field.
+    ///
+    /// This min_value may not be exact.
+    /// For instance, the min value does not take in account of possible
+    /// deleted document. All values are however guaranteed to be higher than
+    /// `.min_value()`.
+    fn min_value(&self) -> T;
+
+    /// Returns the maximum value for this fast field.
+    ///
+    /// This max_value may not be exact.
+    /// For instance, the max value does not take in account of possible
+    /// deleted document. All values are however guaranteed to be higher than
+    /// `.max_value()`.
+    fn max_value(&self) -> T;
+
+    /// The number of values in the column.
+    fn num_vals(&self) -> u32;
+
+    /// Returns a iterator over the data
+    fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = T> + 'a> {
+        Box::new((0..self.num_vals()).map(|idx| self.get_val(idx)))
+    }
+}
+
+impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd> ColumnValues<T> for &'a C {
+    fn get_val(&self, idx: u32) -> T {
+        (*self).get_val(idx)
+    }
+
+    fn min_value(&self) -> T {
+        (*self).min_value()
+    }
+
+    fn max_value(&self) -> T {
+        (*self).max_value()
+    }
+
+    fn num_vals(&self) -> u32 {
+        (*self).num_vals()
+    }
+
+    fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
+        (*self).iter()
+    }
+
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        (*self).get_range(start, output)
+    }
+}
+
+/// VecColumn provides `Column` over a slice.
+pub struct VecColumn<'a, T = u64> {
+    pub(crate) values: &'a [T],
+    pub(crate) min_value: T,
+    pub(crate) max_value: T,
+}
+
+impl<'a, T: Copy + PartialOrd + Send + Sync> ColumnValues<T> for VecColumn<'a, T> {
+    fn get_val(&self, position: u32) -> T {
+        self.values[position as usize]
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        Box::new(self.values.iter().copied())
+    }
+
+    fn min_value(&self) -> T {
+        self.min_value
+    }
+
+    fn max_value(&self) -> T {
+        self.max_value
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.values.len() as u32
+    }
+
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        output.copy_from_slice(&self.values[start as usize..][..output.len()])
+    }
+}
+
+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 {
+        let values = values.as_ref();
+        let (min_value, max_value) = minmax(values.iter().copied()).unwrap_or_default();
+        Self {
+            values,
+            min_value,
+            max_value,
+        }
+    }
+}
+
+struct MonotonicMappingColumn<C, T, Input> {
+    from_column: C,
+    monotonic_mapping: T,
+    _phantom: PhantomData<Input>,
+}
+
+/// Creates a view of a column transformed by a strictly monotonic mapping. See
+/// [`StrictlyMonotonicFn`].
+///
+/// E.g. apply a gcd monotonic_mapping([100, 200, 300]) == [1, 2, 3]
+/// monotonic_mapping.mapping() is expected to be injective, and we should always have
+/// monotonic_mapping.inverse(monotonic_mapping.mapping(el)) == el
+///
+/// The inverse of the mapping is required for:
+/// `fn get_positions_for_value_range(&self, range: RangeInclusive<T>) -> Vec<u64> `
+/// The user provides the original value range and we need to monotonic map them in the same way the
+/// serialization does before calling the underlying column.
+///
+/// Note that when opening a codec, the monotonic_mapping should be the inverse of the mapping
+/// during serialization. And therefore the monotonic_mapping_inv when opening is the same as
+/// monotonic_mapping during serialization.
+pub fn monotonic_map_column<C, T, Input, Output>(
+    from_column: C,
+    monotonic_mapping: T,
+) -> impl ColumnValues<Output>
+where
+    C: ColumnValues<Input>,
+    T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
+    Input: PartialOrd + Send + Sync + Clone,
+    Output: PartialOrd + Send + Sync + Clone,
+{
+    MonotonicMappingColumn {
+        from_column,
+        monotonic_mapping,
+        _phantom: PhantomData,
+    }
+}
+
+impl<C, T, Input, Output> ColumnValues<Output> for MonotonicMappingColumn<C, T, Input>
+where
+    C: ColumnValues<Input>,
+    T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
+    Input: PartialOrd + Send + Sync + Clone,
+    Output: PartialOrd + Send + Sync + Clone,
+{
+    #[inline]
+    fn get_val(&self, idx: u32) -> Output {
+        let from_val = self.from_column.get_val(idx);
+        self.monotonic_mapping.mapping(from_val)
+    }
+
+    fn min_value(&self) -> Output {
+        let from_min_value = self.from_column.min_value();
+        self.monotonic_mapping.mapping(from_min_value)
+    }
+
+    fn max_value(&self) -> Output {
+        let from_max_value = self.from_column.max_value();
+        self.monotonic_mapping.mapping(from_max_value)
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.from_column.num_vals()
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = Output> + '_> {
+        Box::new(
+            self.from_column
+                .iter()
+                .map(|el| self.monotonic_mapping.mapping(el)),
+        )
+    }
+
+    fn get_docids_for_value_range(
+        &self,
+        range: RangeInclusive<Output>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.from_column.get_docids_for_value_range(
+            self.monotonic_mapping.inverse(range.start().clone())
+                ..=self.monotonic_mapping.inverse(range.end().clone()),
+            doc_id_range,
+            positions,
+        )
+    }
+
+    // We voluntarily do not implement get_range as it yields a regression,
+    // and we do not have any specialized implementation anyway.
+}
+
+/// Wraps an iterator into a `Column`.
+pub struct IterColumn<T>(T);
+
+impl<T> From<T> for IterColumn<T>
+where T: Iterator + Clone + ExactSizeIterator
+{
+    fn from(iter: T) -> Self {
+        IterColumn(iter)
+    }
+}
+
+impl<T> ColumnValues<T::Item> for IterColumn<T>
+where
+    T: Iterator + Clone + ExactSizeIterator + Send + Sync,
+    T::Item: PartialOrd,
+{
+    fn get_val(&self, idx: u32) -> T::Item {
+        self.0.clone().nth(idx as usize).unwrap()
+    }
+
+    fn min_value(&self) -> T::Item {
+        self.0.clone().next().unwrap()
+    }
+
+    fn max_value(&self) -> T::Item {
+        self.0.clone().last().unwrap()
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.0.len() as u32
+    }
+
+    fn iter(&self) -> Box<dyn Iterator<Item = T::Item> + '_> {
+        Box::new(self.0.clone())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::monotonic_mapping::{
+        StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternalBaseval,
+        StrictlyMonotonicMappingToInternalGCDBaseval,
+    };
+
+    #[test]
+    fn test_monotonic_mapping() {
+        let vals = &[3u64, 5u64][..];
+        let col = VecColumn::from(vals);
+        let mapped = monotonic_map_column(col, StrictlyMonotonicMappingToInternalBaseval::new(2));
+        assert_eq!(mapped.min_value(), 1u64);
+        assert_eq!(mapped.max_value(), 3u64);
+        assert_eq!(mapped.num_vals(), 2);
+        assert_eq!(mapped.num_vals(), 2);
+        assert_eq!(mapped.get_val(0), 1);
+        assert_eq!(mapped.get_val(1), 3);
+    }
+
+    #[test]
+    fn test_range_as_col() {
+        let col = IterColumn::from(10..100);
+        assert_eq!(col.num_vals(), 90);
+        assert_eq!(col.max_value(), 99);
+    }
+
+    #[test]
+    fn test_monotonic_mapping_iter() {
+        let vals: Vec<u64> = (10..110u64).map(|el| el * 10).collect();
+        let col = VecColumn::from(&vals);
+        let mapped = monotonic_map_column(
+            col,
+            StrictlyMonotonicMappingInverter::from(
+                StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100),
+            ),
+        );
+        let val_i64s: Vec<u64> = mapped.iter().collect();
+        for i in 0..100 {
+            assert_eq!(val_i64s[i as usize], mapped.get_val(i));
+        }
+    }
+
+    #[test]
+    fn test_monotonic_mapping_get_range() {
+        let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
+        let col = VecColumn::from(&vals);
+        let mapped = monotonic_map_column(
+            col,
+            StrictlyMonotonicMappingInverter::from(
+                StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 0),
+            ),
+        );
+
+        assert_eq!(mapped.min_value(), 0u64);
+        assert_eq!(mapped.max_value(), 9900u64);
+        assert_eq!(mapped.num_vals(), 100);
+        let val_u64s: Vec<u64> = mapped.iter().collect();
+        assert_eq!(val_u64s.len(), 100);
+        for i in 0..100 {
+            assert_eq!(val_u64s[i as usize], mapped.get_val(i));
+            assert_eq!(val_u64s[i as usize], vals[i as usize] * 10);
+        }
+        let mut buf = [0u64; 20];
+        mapped.get_range(7, &mut buf[..]);
+        assert_eq!(&val_u64s[7..][..20], &buf);
+    }
+}

columnar/src/column_values/column_with_cardinality.rs

@@ -0,0 +1,19 @@
+// Copyright (C) 2022 Quickwit, Inc.
+//
+// Quickwit is offered under the AGPL v3.0 and as commercial software.
+// For commercial licensing, contact us at hello@quickwit.io.
+//
+// AGPL:
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as
+// published by the Free Software Foundation, either version 3 of the
+// License, or (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see <http://www.gnu.org/licenses/>.
+//

columnar/src/column_values/compact_space/blank_range.rs

@@ -0,0 +1,43 @@
+use std::ops::RangeInclusive;
+
+/// The range of a blank in value space.
+///
+/// A blank is an unoccupied space in the data.
+/// Use try_into() to construct.
+/// A range has to have at least length of 3. Invalid ranges will be rejected.
+///
+/// Ordered by range length.
+#[derive(Debug, Eq, PartialEq, Clone)]
+pub(crate) struct BlankRange {
+    blank_range: RangeInclusive<u128>,
+}
+impl TryFrom<RangeInclusive<u128>> for BlankRange {
+    type Error = &'static str;
+    fn try_from(range: RangeInclusive<u128>) -> Result<Self, Self::Error> {
+        let blank_size = range.end().saturating_sub(*range.start());
+        if blank_size < 2 {
+            Err("invalid range")
+        } else {
+            Ok(BlankRange { blank_range: range })
+        }
+    }
+}
+impl BlankRange {
+    pub(crate) fn blank_size(&self) -> u128 {
+        self.blank_range.end() - self.blank_range.start() + 1
+    }
+    pub(crate) fn blank_range(&self) -> RangeInclusive<u128> {
+        self.blank_range.clone()
+    }
+}
+
+impl Ord for BlankRange {
+    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+        self.blank_size().cmp(&other.blank_size())
+    }
+}
+impl PartialOrd for BlankRange {
+    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+        Some(self.blank_size().cmp(&other.blank_size()))
+    }
+}

columnar/src/column_values/compact_space/build_compact_space.rs

@@ -0,0 +1,231 @@
+use std::collections::{BTreeSet, BinaryHeap};
+use std::iter;
+use std::ops::RangeInclusive;
+
+use itertools::Itertools;
+
+use super::blank_range::BlankRange;
+use super::{CompactSpace, RangeMapping};
+
+/// Put the blanks for the sorted values into a binary heap
+fn get_blanks(values_sorted: &BTreeSet<u128>) -> BinaryHeap<BlankRange> {
+    let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
+    for (first, second) in values_sorted.iter().tuple_windows() {
+        // Correctness Overflow: the values are deduped and sorted (BTreeSet property), that means
+        // there's always space between two values.
+        let blank_range = first + 1..=second - 1;
+        let blank_range: Result<BlankRange, _> = blank_range.try_into();
+        if let Ok(blank_range) = blank_range {
+            blanks.push(blank_range);
+        }
+    }
+
+    blanks
+}
+
+struct BlankCollector {
+    blanks: Vec<BlankRange>,
+    staged_blanks_sum: u128,
+}
+impl BlankCollector {
+    fn new() -> Self {
+        Self {
+            blanks: vec![],
+            staged_blanks_sum: 0,
+        }
+    }
+    fn stage_blank(&mut self, blank: BlankRange) {
+        self.staged_blanks_sum += blank.blank_size();
+        self.blanks.push(blank);
+    }
+    fn drain(&mut self) -> impl Iterator<Item = BlankRange> + '_ {
+        self.staged_blanks_sum = 0;
+        self.blanks.drain(..)
+    }
+    fn staged_blanks_sum(&self) -> u128 {
+        self.staged_blanks_sum
+    }
+    fn num_staged_blanks(&self) -> usize {
+        self.blanks.len()
+    }
+}
+fn num_bits(val: u128) -> u8 {
+    (128u32 - val.leading_zeros()) as u8
+}
+
+/// Will collect blanks and add them to compact space if more bits are saved than cost from
+/// metadata.
+pub fn get_compact_space(
+    values_deduped_sorted: &BTreeSet<u128>,
+    total_num_values: u32,
+    cost_per_blank: usize,
+) -> CompactSpace {
+    let mut compact_space_builder = CompactSpaceBuilder::new();
+    if values_deduped_sorted.is_empty() {
+        return compact_space_builder.finish();
+    }
+
+    let mut blanks: BinaryHeap<BlankRange> = get_blanks(values_deduped_sorted);
+    // Replace after stabilization of https://github.com/rust-lang/rust/issues/62924
+
+    // We start by space that's limited to min_value..=max_value
+    let min_value = *values_deduped_sorted.iter().next().unwrap_or(&0);
+    let max_value = *values_deduped_sorted.iter().last().unwrap_or(&0);
+
+    // +1 for null, in case min and max covers the whole space, we are off by one.
+    let mut amplitude_compact_space = (max_value - min_value).saturating_add(1);
+    if min_value != 0 {
+        compact_space_builder.add_blanks(iter::once(0..=min_value - 1));
+    }
+    if max_value != u128::MAX {
+        compact_space_builder.add_blanks(iter::once(max_value + 1..=u128::MAX));
+    }
+
+    let mut amplitude_bits: u8 = num_bits(amplitude_compact_space);
+
+    let mut blank_collector = BlankCollector::new();
+    // We will stage blanks until they reduce the compact space by at least 1 bit and then flush
+    // them if the metadata cost is lower than the total number of saved bits.
+    // Binary heap to process the gaps by their size
+    while let Some(blank_range) = blanks.pop() {
+        blank_collector.stage_blank(blank_range);
+
+        let staged_spaces_sum: u128 = blank_collector.staged_blanks_sum();
+        let amplitude_new_compact_space = amplitude_compact_space - staged_spaces_sum;
+        let amplitude_new_bits = num_bits(amplitude_new_compact_space);
+        if amplitude_bits == amplitude_new_bits {
+            continue;
+        }
+        let saved_bits = (amplitude_bits - amplitude_new_bits) as usize * total_num_values as usize;
+        // TODO: Maybe calculate exact cost of blanks and run this more expensive computation only,
+        // when amplitude_new_bits changes
+        let cost = blank_collector.num_staged_blanks() * cost_per_blank;
+        if cost >= saved_bits {
+            // Continue here, since although we walk over the blanks by size,
+            // we can potentially save a lot at the last bits, which are smaller blanks
+            //
+            // E.g. if the first range reduces the compact space by 1000 from 2000 to 1000, which
+            // saves 11-10=1 bit and the next range reduces the compact space by 950 to
+            // 50, which saves 10-6=4 bit
+            continue;
+        }
+
+        amplitude_compact_space = amplitude_new_compact_space;
+        amplitude_bits = amplitude_new_bits;
+        compact_space_builder.add_blanks(blank_collector.drain().map(|blank| blank.blank_range()));
+    }
+
+    // special case, when we don't collected any blanks because:
+    // * the data is empty (early exit)
+    // * the algorithm did decide it's not worth the cost, which can be the case for single values
+    //
+    // We drain one collected blank unconditionally, so the empty case is reserved for empty
+    // data, and therefore empty compact_space means the data is empty and no data is covered
+    // (conversely to all data) and we can assign null to it.
+    if compact_space_builder.is_empty() {
+        compact_space_builder.add_blanks(
+            blank_collector
+                .drain()
+                .map(|blank| blank.blank_range())
+                .take(1),
+        );
+    }
+
+    let compact_space = compact_space_builder.finish();
+    if max_value - min_value != u128::MAX {
+        debug_assert_eq!(
+            compact_space.amplitude_compact_space(),
+            amplitude_compact_space
+        );
+    }
+    compact_space
+}
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+struct CompactSpaceBuilder {
+    blanks: Vec<RangeInclusive<u128>>,
+}
+
+impl CompactSpaceBuilder {
+    /// Creates a new compact space builder which will initially cover the whole space.
+    fn new() -> Self {
+        Self { blanks: Vec::new() }
+    }
+
+    /// Assumes that repeated add_blank calls don't overlap and are not adjacent,
+    /// e.g. [3..=5, 5..=10] is not allowed
+    ///
+    /// Both of those assumptions are true when blanks are produced from sorted values.
+    fn add_blanks(&mut self, blank: impl Iterator<Item = RangeInclusive<u128>>) {
+        self.blanks.extend(blank);
+    }
+
+    fn is_empty(&self) -> bool {
+        self.blanks.is_empty()
+    }
+
+    /// Convert blanks to covered space and assign null value
+    fn finish(mut self) -> CompactSpace {
+        // sort by start. ranges are not allowed to overlap
+        self.blanks.sort_unstable_by_key(|blank| *blank.start());
+
+        let mut covered_space = Vec::with_capacity(self.blanks.len());
+
+        // begining of the blanks
+        if let Some(first_blank_start) = self.blanks.first().map(RangeInclusive::start) {
+            if *first_blank_start != 0 {
+                covered_space.push(0..=first_blank_start - 1);
+            }
+        }
+
+        // Between the blanks
+        let between_blanks = self.blanks.iter().tuple_windows().map(|(left, right)| {
+            assert!(
+                left.end() < right.start(),
+                "overlapping or adjacent ranges detected"
+            );
+            *left.end() + 1..=*right.start() - 1
+        });
+        covered_space.extend(between_blanks);
+
+        // end of the blanks
+        if let Some(last_blank_end) = self.blanks.last().map(RangeInclusive::end) {
+            if *last_blank_end != u128::MAX {
+                covered_space.push(last_blank_end + 1..=u128::MAX);
+            }
+        }
+
+        if covered_space.is_empty() {
+            covered_space.push(0..=0); // empty data case
+        };
+
+        let mut compact_start: u64 = 1; // 0 is reserved for `null`
+        let mut ranges_mapping: Vec<RangeMapping> = Vec::with_capacity(covered_space.len());
+        for cov in covered_space {
+            let range_mapping = super::RangeMapping {
+                value_range: cov,
+                compact_start,
+            };
+            let covered_range_len = range_mapping.range_length();
+            ranges_mapping.push(range_mapping);
+            compact_start += covered_range_len;
+        }
+        // println!("num ranges {}", ranges_mapping.len());
+        CompactSpace { ranges_mapping }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_binary_heap_pop_order() {
+        let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
+        blanks.push((0..=10).try_into().unwrap());
+        blanks.push((100..=200).try_into().unwrap());
+        blanks.push((100..=110).try_into().unwrap());
+        assert_eq!(blanks.pop().unwrap().blank_size(), 101);
+        assert_eq!(blanks.pop().unwrap().blank_size(), 11);
+    }
+}

columnar/src/column_values/compact_space/mod.rs

@@ -0,0 +1,813 @@
+/// This codec takes a large number space (u128) and reduces it to a compact number space.
+///
+/// It will find spaces in the number range. For example:
+///
+/// 100, 101, 102, 103, 104, 50000, 50001
+/// could be mapped to
+/// 100..104 -> 0..4
+/// 50000..50001 -> 5..6
+///
+/// Compact space 0..=6 requires much less bits than 100..=50001
+///
+/// The codec is created to compress ip addresses, but may be employed in other use cases.
+use std::{
+    cmp::Ordering,
+    collections::BTreeSet,
+    io::{self, Write},
+    ops::{Range, RangeInclusive},
+};
+
+use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
+use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
+
+use crate::column_values::compact_space::build_compact_space::get_compact_space;
+use crate::column_values::ColumnValues;
+
+mod blank_range;
+mod build_compact_space;
+
+/// The cost per blank is quite hard actually, since blanks are delta encoded, the actual cost of
+/// blanks depends on the number of blanks.
+///
+/// The number is taken by looking at a real dataset. It is optimized for larger datasets.
+const COST_PER_BLANK_IN_BITS: usize = 36;
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct CompactSpace {
+    ranges_mapping: Vec<RangeMapping>,
+}
+
+/// Maps the range from the original space to compact_start + range.len()
+#[derive(Debug, Clone, Eq, PartialEq)]
+struct RangeMapping {
+    value_range: RangeInclusive<u128>,
+    compact_start: u64,
+}
+impl RangeMapping {
+    fn range_length(&self) -> u64 {
+        (self.value_range.end() - self.value_range.start()) as u64 + 1
+    }
+
+    // The last value of the compact space in this range
+    fn compact_end(&self) -> u64 {
+        self.compact_start + self.range_length() - 1
+    }
+}
+
+impl BinarySerializable for CompactSpace {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
+
+        let mut prev_value = 0;
+        for value_range in self
+            .ranges_mapping
+            .iter()
+            .map(|range_mapping| &range_mapping.value_range)
+        {
+            let blank_delta_start = value_range.start() - prev_value;
+            VIntU128(blank_delta_start).serialize(writer)?;
+            prev_value = *value_range.start();
+
+            let blank_delta_end = value_range.end() - prev_value;
+            VIntU128(blank_delta_end).serialize(writer)?;
+            prev_value = *value_range.end();
+        }
+
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let num_ranges = VInt::deserialize(reader)?.0;
+        let mut ranges_mapping: Vec<RangeMapping> = vec![];
+        let mut value = 0u128;
+        let mut compact_start = 1u64; // 0 is reserved for `null`
+        for _ in 0..num_ranges {
+            let blank_delta_start = VIntU128::deserialize(reader)?.0;
+            value += blank_delta_start;
+            let blank_start = value;
+
+            let blank_delta_end = VIntU128::deserialize(reader)?.0;
+            value += blank_delta_end;
+            let blank_end = value;
+
+            let range_mapping = RangeMapping {
+                value_range: blank_start..=blank_end,
+                compact_start,
+            };
+            let range_length = range_mapping.range_length();
+            ranges_mapping.push(range_mapping);
+            compact_start += range_length;
+        }
+
+        Ok(Self { ranges_mapping })
+    }
+}
+
+impl CompactSpace {
+    /// Amplitude is the value range of the compact space including the sentinel value used to
+    /// identify null values. The compact space is 0..=amplitude .
+    ///
+    /// It's only used to verify we don't exceed u64 number space, which would indicate a bug.
+    fn amplitude_compact_space(&self) -> u128 {
+        self.ranges_mapping
+            .last()
+            .map(|last_range| last_range.compact_end() as u128)
+            .unwrap_or(1) // compact space starts at 1, 0 == null
+    }
+
+    fn get_range_mapping(&self, pos: usize) -> &RangeMapping {
+        &self.ranges_mapping[pos]
+    }
+
+    /// Returns either Ok(the value in the compact space) or if it is outside the compact space the
+    /// Err(position where it would be inserted)
+    fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
+        self.ranges_mapping
+            .binary_search_by(|probe| {
+                let value_range = &probe.value_range;
+                if value < *value_range.start() {
+                    Ordering::Greater
+                } else if value > *value_range.end() {
+                    Ordering::Less
+                } else {
+                    Ordering::Equal
+                }
+            })
+            .map(|pos| {
+                let range_mapping = &self.ranges_mapping[pos];
+                let pos_in_range = (value - range_mapping.value_range.start()) as u64;
+                range_mapping.compact_start + pos_in_range
+            })
+    }
+
+    /// Unpacks a value from compact space u64 to u128 space
+    fn compact_to_u128(&self, compact: u64) -> u128 {
+        let pos = self
+            .ranges_mapping
+            .binary_search_by_key(&compact, |range_mapping| range_mapping.compact_start)
+            // Correctness: Overflow. The first range starts at compact space 0, the error from
+            // binary search can never be 0
+            .map_or_else(|e| e - 1, |v| v);
+
+        let range_mapping = &self.ranges_mapping[pos];
+        let diff = compact - range_mapping.compact_start;
+        range_mapping.value_range.start() + diff as u128
+    }
+}
+
+pub struct CompactSpaceCompressor {
+    params: IPCodecParams,
+}
+#[derive(Debug, Clone)]
+pub struct IPCodecParams {
+    compact_space: CompactSpace,
+    bit_unpacker: BitUnpacker,
+    min_value: u128,
+    max_value: u128,
+    num_vals: u32,
+    num_bits: u8,
+}
+
+impl CompactSpaceCompressor {
+    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
+    pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
+        let mut values_sorted = BTreeSet::new();
+        values_sorted.extend(iter);
+        let total_num_values = num_vals;
+
+        let compact_space =
+            get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
+        let amplitude_compact_space = compact_space.amplitude_compact_space();
+
+        assert!(
+            amplitude_compact_space <= u64::MAX as u128,
+            "case unsupported."
+        );
+
+        let num_bits = tantivy_bitpacker::compute_num_bits(amplitude_compact_space as u64);
+        let min_value = *values_sorted.iter().next().unwrap_or(&0);
+        let max_value = *values_sorted.iter().last().unwrap_or(&0);
+        assert_eq!(
+            compact_space
+                .u128_to_compact(max_value)
+                .expect("could not convert max value to compact space"),
+            amplitude_compact_space as u64
+        );
+        CompactSpaceCompressor {
+            params: IPCodecParams {
+                compact_space,
+                bit_unpacker: BitUnpacker::new(num_bits),
+                min_value,
+                max_value,
+                num_vals: total_num_values,
+                num_bits,
+            },
+        }
+    }
+
+    fn write_footer(self, writer: &mut impl Write) -> io::Result<()> {
+        let writer = &mut CountingWriter::wrap(writer);
+        self.params.serialize(writer)?;
+
+        let footer_len = writer.written_bytes() as u32;
+        footer_len.serialize(writer)?;
+
+        Ok(())
+    }
+
+    pub fn compress_into(
+        self,
+        vals: impl Iterator<Item = u128>,
+        write: &mut impl Write,
+    ) -> io::Result<()> {
+        let mut bitpacker = BitPacker::default();
+        for val in vals {
+            let compact = self
+                .params
+                .compact_space
+                .u128_to_compact(val)
+                .map_err(|_| {
+                    io::Error::new(
+                        io::ErrorKind::InvalidData,
+                        "Could not convert value to compact_space. This is a bug.",
+                    )
+                })?;
+            bitpacker.write(compact, self.params.num_bits, write)?;
+        }
+        bitpacker.close(write)?;
+        self.write_footer(write)?;
+        Ok(())
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct CompactSpaceDecompressor {
+    data: OwnedBytes,
+    params: IPCodecParams,
+}
+
+impl BinarySerializable for IPCodecParams {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        // header flags for future optional dictionary encoding
+        let footer_flags = 0u64;
+        footer_flags.serialize(writer)?;
+
+        VIntU128(self.min_value).serialize(writer)?;
+        VIntU128(self.max_value).serialize(writer)?;
+        VIntU128(self.num_vals as u128).serialize(writer)?;
+        self.num_bits.serialize(writer)?;
+
+        self.compact_space.serialize(writer)?;
+
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let _header_flags = u64::deserialize(reader)?;
+        let min_value = VIntU128::deserialize(reader)?.0;
+        let max_value = VIntU128::deserialize(reader)?.0;
+        let num_vals = VIntU128::deserialize(reader)?.0 as u32;
+        let num_bits = u8::deserialize(reader)?;
+        let compact_space = CompactSpace::deserialize(reader)?;
+
+        Ok(Self {
+            compact_space,
+            bit_unpacker: BitUnpacker::new(num_bits),
+            min_value,
+            max_value,
+            num_vals,
+            num_bits,
+        })
+    }
+}
+
+impl ColumnValues<u128> for CompactSpaceDecompressor {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u128 {
+        self.get(doc)
+    }
+
+    fn min_value(&self) -> u128 {
+        self.min_value()
+    }
+
+    fn max_value(&self) -> u128 {
+        self.max_value()
+    }
+
+    fn num_vals(&self) -> u32 {
+        self.params.num_vals
+    }
+
+    #[inline]
+    fn iter(&self) -> Box<dyn Iterator<Item = u128> + '_> {
+        Box::new(self.iter())
+    }
+
+    #[inline]
+    fn get_docids_for_value_range(
+        &self,
+        value_range: RangeInclusive<u128>,
+        positions_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.get_positions_for_value_range(value_range, positions_range, positions)
+    }
+}
+
+impl CompactSpaceDecompressor {
+    pub fn open(data: OwnedBytes) -> io::Result<CompactSpaceDecompressor> {
+        let (data_slice, footer_len_bytes) = data.split_at(data.len() - 4);
+        let footer_len = u32::deserialize(&mut &footer_len_bytes[..])?;
+
+        let data_footer = &data_slice[data_slice.len() - footer_len as usize..];
+        let params = IPCodecParams::deserialize(&mut &data_footer[..])?;
+        let decompressor = CompactSpaceDecompressor { data, params };
+
+        Ok(decompressor)
+    }
+
+    /// Converting to compact space for the decompressor is more complex, since we may get values
+    /// which are outside the compact space. e.g. if we map
+    /// 1000 => 5
+    /// 2000 => 6
+    ///
+    /// and we want a mapping for 1005, there is no equivalent compact space. We instead return an
+    /// error with the index of the next range.
+    fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
+        self.params.compact_space.u128_to_compact(value)
+    }
+
+    fn compact_to_u128(&self, compact: u64) -> u128 {
+        self.params.compact_space.compact_to_u128(compact)
+    }
+
+    /// Comparing on compact space: Random dataset 0,24 (50% random hit) - 1.05 GElements/s
+    /// Comparing on compact space: Real dataset 1.08 GElements/s
+    ///
+    /// Comparing on original space: Real dataset .06 GElements/s (not completely optimized)
+    #[inline]
+    pub fn get_positions_for_value_range(
+        &self,
+        value_range: RangeInclusive<u128>,
+        position_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        if value_range.start() > value_range.end() {
+            return;
+        }
+        let position_range = position_range.start..position_range.end.min(self.num_vals());
+        let from_value = *value_range.start();
+        let to_value = *value_range.end();
+        assert!(to_value >= from_value);
+        let compact_from = self.u128_to_compact(from_value);
+        let compact_to = self.u128_to_compact(to_value);
+
+        // Quick return, if both ranges fall into the same non-mapped space, the range can't cover
+        // any values, so we can early exit
+        match (compact_to, compact_from) {
+            (Err(pos1), Err(pos2)) if pos1 == pos2 => return,
+            _ => {}
+        }
+
+        let compact_from = compact_from.unwrap_or_else(|pos| {
+            // Correctness: Out of bounds, if this value is Err(last_index + 1), we early exit,
+            // since the to_value also mapps into the same non-mapped space
+            let range_mapping = self.params.compact_space.get_range_mapping(pos);
+            range_mapping.compact_start
+        });
+        // If there is no compact space, we go to the closest upperbound compact space
+        let compact_to = compact_to.unwrap_or_else(|pos| {
+            // Correctness: Overflow, if this value is Err(0), we early exit,
+            // since the from_value also mapps into the same non-mapped space
+
+            // Get end of previous range
+            let pos = pos - 1;
+            let range_mapping = self.params.compact_space.get_range_mapping(pos);
+            range_mapping.compact_end()
+        });
+
+        let range = compact_from..=compact_to;
+
+        let scan_num_docs = position_range.end - position_range.start;
+
+        let step_size = 4;
+        let cutoff = position_range.start + scan_num_docs - scan_num_docs % step_size;
+
+        let mut push_if_in_range = |idx, val| {
+            if range.contains(&val) {
+                positions.push(idx);
+            }
+        };
+        let get_val = |idx| self.params.bit_unpacker.get(idx, &self.data);
+        // unrolled loop
+        for idx in (position_range.start..cutoff).step_by(step_size as usize) {
+            let idx1 = idx;
+            let idx2 = idx + 1;
+            let idx3 = idx + 2;
+            let idx4 = idx + 3;
+            let val1 = get_val(idx1);
+            let val2 = get_val(idx2);
+            let val3 = get_val(idx3);
+            let val4 = get_val(idx4);
+            push_if_in_range(idx1, val1);
+            push_if_in_range(idx2, val2);
+            push_if_in_range(idx3, val3);
+            push_if_in_range(idx4, val4);
+        }
+
+        // handle rest
+        for idx in cutoff..position_range.end {
+            push_if_in_range(idx, get_val(idx));
+        }
+    }
+
+    #[inline]
+    fn iter_compact(&self) -> impl Iterator<Item = u64> + '_ {
+        (0..self.params.num_vals).map(move |idx| self.params.bit_unpacker.get(idx, &self.data))
+    }
+
+    #[inline]
+    fn iter(&self) -> impl Iterator<Item = u128> + '_ {
+        // TODO: Performance. It would be better to iterate on the ranges and check existence via
+        // the bit_unpacker.
+        self.iter_compact()
+            .map(|compact| self.compact_to_u128(compact))
+    }
+
+    #[inline]
+    pub fn get(&self, idx: u32) -> u128 {
+        let compact = self.params.bit_unpacker.get(idx, &self.data);
+        self.compact_to_u128(compact)
+    }
+
+    pub fn min_value(&self) -> u128 {
+        self.params.min_value
+    }
+
+    pub fn max_value(&self) -> u128 {
+        self.params.max_value
+    }
+}
+
+// TODO reenable what can be reenabled.
+// #[cfg(test)]
+// mod tests {
+//
+// use super::*;
+// use crate::column::format_version::read_format_version;
+// use crate::column::column_footer::read_null_index_footer;
+// use crate::column::serialize::U128Header;
+// use crate::column::{open_u128, serialize_u128};
+//
+// #[test]
+// fn compact_space_test() {
+// let ips = &[
+// 2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
+// ]
+// .into_iter()
+// .collect();
+// let compact_space = get_compact_space(ips, ips.len() as u32, 11);
+// let amplitude = compact_space.amplitude_compact_space();
+// assert_eq!(amplitude, 17);
+// assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
+// assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
+// assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
+//
+// for (num1, num2) in (0..3).tuple_windows() {
+// assert_eq!(
+// compact_space.get_range_mapping(num1).compact_end() + 1,
+// compact_space.get_range_mapping(num2).compact_start
+// );
+// }
+//
+// let mut output: Vec<u8> = Vec::new();
+// compact_space.serialize(&mut output).unwrap();
+//
+// assert_eq!(
+// compact_space,
+// CompactSpace::deserialize(&mut &output[..]).unwrap()
+// );
+//
+// for ip in ips {
+// let compact = compact_space.u128_to_compact(*ip).unwrap();
+// assert_eq!(compact_space.compact_to_u128(compact), *ip);
+// }
+// }
+//
+// #[test]
+// fn compact_space_amplitude_test() {
+// let ips = &[100000u128, 1000000].into_iter().collect();
+// let compact_space = get_compact_space(ips, ips.len() as u32, 1);
+// let amplitude = compact_space.amplitude_compact_space();
+// assert_eq!(amplitude, 2);
+// }
+//
+// fn test_all(mut data: OwnedBytes, expected: &[u128]) {
+// let _header = U128Header::deserialize(&mut data);
+// let decompressor = CompactSpaceDecompressor::open(data).unwrap();
+// for (idx, expected_val) in expected.iter().cloned().enumerate() {
+// let val = decompressor.get(idx as u32);
+// assert_eq!(val, expected_val);
+//
+// let test_range = |range: RangeInclusive<u128>| {
+// let expected_positions = expected
+// .iter()
+// .positions(|val| range.contains(val))
+// .map(|pos| pos as u32)
+// .collect::<Vec<_>>();
+// let mut positions = Vec::new();
+// decompressor.get_positions_for_value_range(
+// range,
+// 0..decompressor.num_vals(),
+// &mut positions,
+// );
+// assert_eq!(positions, expected_positions);
+// };
+//
+// test_range(expected_val.saturating_sub(1)..=expected_val);
+// test_range(expected_val..=expected_val);
+// test_range(expected_val..=expected_val.saturating_add(1));
+// test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
+// }
+// }
+//
+// fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
+// let mut out = Vec::new();
+// serialize_u128(
+// || u128_vals.iter().cloned(),
+// u128_vals.len() as u32,
+// &mut out,
+// )
+// .unwrap();
+//
+// let data = OwnedBytes::new(out);
+// let (data, _format_version) = read_format_version(data).unwrap();
+// let (data, _null_index_footer) = read_null_index_footer(data).unwrap();
+// test_all(data.clone(), u128_vals);
+//
+// data
+// }
+//
+// #[test]
+// fn test_range_1() {
+// let vals = &[
+// 1u128,
+// 100u128,
+// 3u128,
+// 99999u128,
+// 100000u128,
+// 100001u128,
+// 4_000_211_221u128,
+// 4_000_211_222u128,
+// 333u128,
+// ];
+// let mut data = test_aux_vals(vals);
+//
+// let _header = U128Header::deserialize(&mut data);
+// let decomp = CompactSpaceDecompressor::open(data).unwrap();
+// let complete_range = 0..vals.len() as u32;
+// for (pos, val) in vals.iter().enumerate() {
+// 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);
+// assert_eq!(positions, vec![pos]);
+// }
+//
+// handle docid range out of bounds
+// let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
+// assert!(positions.is_empty());
+//
+// let positions =
+// get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
+// assert_eq!(positions, vec![0]);
+// let positions =
+// get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
+// assert_eq!(positions, vec![0]);
+// let positions =
+// get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
+// assert_eq!(positions, vec![0, 2]);
+// assert_eq!(
+// get_positions_for_value_range_helper(
+// &decomp,
+// 99999u128..=99999u128,
+// complete_range.clone()
+// ),
+// vec![3]
+// );
+// assert_eq!(
+// get_positions_for_value_range_helper(
+// &decomp,
+// 99999u128..=100000u128,
+// complete_range.clone()
+// ),
+// vec![3, 4]
+// );
+// assert_eq!(
+// get_positions_for_value_range_helper(
+// &decomp,
+// 99998u128..=100000u128,
+// complete_range.clone()
+// ),
+// vec![3, 4]
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 99998u128..=99999u128,
+// complete_range.clone()
+// ),
+// &[3]
+// );
+// assert!(get_positions_for_value_range_helper(
+// &decomp,
+// 99998u128..=99998u128,
+// complete_range.clone()
+// )
+// .is_empty());
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 333u128..=333u128,
+// complete_range.clone()
+// ),
+// &[8]
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 332u128..=333u128,
+// complete_range.clone()
+// ),
+// &[8]
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 332u128..=334u128,
+// complete_range.clone()
+// ),
+// &[8]
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 333u128..=334u128,
+// complete_range.clone()
+// ),
+// &[8]
+// );
+//
+// assert_eq!(
+// &get_positions_for_value_range_helper(
+// &decomp,
+// 4_000_211_221u128..=5_000_000_000u128,
+// complete_range
+// ),
+// &[6, 7]
+// );
+// }
+//
+// #[test]
+// fn test_empty() {
+// let vals = &[];
+// let data = test_aux_vals(vals);
+// let _decomp = CompactSpaceDecompressor::open(data).unwrap();
+// }
+//
+// #[test]
+// fn test_range_2() {
+// let vals = &[
+// 100u128,
+// 99999u128,
+// 100000u128,
+// 100001u128,
+// 4_000_211_221u128,
+// 4_000_211_222u128,
+// 333u128,
+// ];
+// let mut data = test_aux_vals(vals);
+// let _header = U128Header::deserialize(&mut data);
+// let decomp = CompactSpaceDecompressor::open(data).unwrap();
+// let complete_range = 0..vals.len() as u32;
+// assert!(
+// &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
+// .is_empty(),
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
+// &[0]
+// );
+// assert_eq!(
+// &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
+// &[0]
+// );
+// }
+//
+// fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd>(
+// column: &C,
+// value_range: RangeInclusive<T>,
+// doc_id_range: Range<u32>,
+// ) -> Vec<u32> {
+// let mut positions = Vec::new();
+// column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
+// positions
+// }
+//
+// #[test]
+// fn test_range_3() {
+// let vals = &[
+// 200u128,
+// 201,
+// 202,
+// 203,
+// 204,
+// 204,
+// 206,
+// 207,
+// 208,
+// 209,
+// 210,
+// 1_000_000,
+// 5_000_000_000,
+// ];
+// let mut out = Vec::new();
+// serialize_u128(|| vals.iter().cloned(), vals.len() as u32, &mut out).unwrap();
+// let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
+// let complete_range = 0..vals.len() as u32;
+//
+// assert_eq!(
+// get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
+// vec![0]
+// );
+//
+// assert_eq!(
+// get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
+// vec![0, 1]
+// );
+//
+// assert_eq!(
+// get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
+// vec![0]
+// );
+//
+// assert_eq!(
+// get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
+// vec![11]
+// );
+// }
+//
+// #[test]
+// fn test_bug1() {
+// let vals = &[9223372036854775806];
+// let _data = test_aux_vals(vals);
+// }
+//
+// #[test]
+// fn test_bug2() {
+// let vals = &[340282366920938463463374607431768211455u128];
+// let _data = test_aux_vals(vals);
+// }
+//
+// #[test]
+// fn test_bug3() {
+// let vals = &[340282366920938463463374607431768211454];
+// let _data = test_aux_vals(vals);
+// }
+//
+// #[test]
+// fn test_bug4() {
+// let vals = &[340282366920938463463374607431768211455, 0];
+// let _data = test_aux_vals(vals);
+// }
+//
+// #[test]
+// fn test_first_large_gaps() {
+// let vals = &[1_000_000_000u128; 100];
+// let _data = test_aux_vals(vals);
+// }
+// use itertools::Itertools;
+// use proptest::prelude::*;
+//
+// fn num_strategy() -> impl Strategy<Value = u128> {
+// prop_oneof![
+// 1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
+// 1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
+// 1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
+// 1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
+// 20 => prop::num::u128::ANY,
+// ]
+// }
+//
+// proptest! {
+// #![proptest_config(ProptestConfig::with_cases(10))]
+//
+// #[test]
+// fn compress_decompress_random(vals in proptest::collection::vec(num_strategy()
+// , 1..1000)) {
+// let _data = test_aux_vals(&vals);
+// }
+// }
+// }
+//

columnar/src/column_values/gcd.rs

@@ -0,0 +1,75 @@
+use std::num::NonZeroU64;
+
+use fastdivide::DividerU64;
+
+/// Compute the gcd of two non null numbers.
+///
+/// It is recommended, but not required, to feed values such that `large >= small`.
+fn compute_gcd(mut large: NonZeroU64, mut small: NonZeroU64) -> NonZeroU64 {
+    loop {
+        let rem: u64 = large.get() % small;
+        if let Some(new_small) = NonZeroU64::new(rem) {
+            (large, small) = (small, new_small);
+        } else {
+            return small;
+        }
+    }
+}
+
+// Find GCD for iterator of numbers
+pub fn find_gcd(numbers: impl Iterator<Item = u64>) -> Option<NonZeroU64> {
+    let mut numbers = numbers.flat_map(NonZeroU64::new);
+    let mut gcd: NonZeroU64 = numbers.next()?;
+    if gcd.get() == 1 {
+        return Some(gcd);
+    }
+
+    let mut gcd_divider = DividerU64::divide_by(gcd.get());
+    for val in numbers {
+        let remainder = val.get() - (gcd_divider.divide(val.get())) * gcd.get();
+        if remainder == 0 {
+            continue;
+        }
+        gcd = compute_gcd(val, gcd);
+        if gcd.get() == 1 {
+            return Some(gcd);
+        }
+
+        gcd_divider = DividerU64::divide_by(gcd.get());
+    }
+    Some(gcd)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use crate::column_values::gcd::{compute_gcd, find_gcd};
+
+    #[test]
+    fn test_compute_gcd() {
+        let test_compute_gcd_aux = |large, small, expected| {
+            let large = NonZeroU64::new(large).unwrap();
+            let small = NonZeroU64::new(small).unwrap();
+            let expected = NonZeroU64::new(expected).unwrap();
+            assert_eq!(compute_gcd(small, large), expected);
+            assert_eq!(compute_gcd(large, small), expected);
+        };
+        test_compute_gcd_aux(1, 4, 1);
+        test_compute_gcd_aux(2, 4, 2);
+        test_compute_gcd_aux(10, 25, 5);
+        test_compute_gcd_aux(25, 25, 25);
+    }
+
+    #[test]
+    fn find_gcd_test() {
+        assert_eq!(find_gcd([0].into_iter()), None);
+        assert_eq!(find_gcd([0, 10].into_iter()), NonZeroU64::new(10));
+        assert_eq!(find_gcd([10, 0].into_iter()), NonZeroU64::new(10));
+        assert_eq!(find_gcd([].into_iter()), None);
+        assert_eq!(find_gcd([15, 30, 5, 10].into_iter()), NonZeroU64::new(5));
+        assert_eq!(find_gcd([15, 16, 10].into_iter()), NonZeroU64::new(1));
+        assert_eq!(find_gcd([0, 5, 5, 5].into_iter()), NonZeroU64::new(5));
+        assert_eq!(find_gcd([0, 0].into_iter()), None);
+    }
+}

columnar/src/column_values/line.rs

@@ -0,0 +1,222 @@
+use std::io;
+use std::num::NonZeroU32;
+
+use common::{BinarySerializable, VInt};
+
+use crate::column_values::ColumnValues;
+
+const MID_POINT: u64 = (1u64 << 32) - 1u64;
+
+/// `Line` describes a line function `y: ax + b` using integer
+/// arithmetics.
+///
+/// The slope is in fact a decimal split into a 32 bit integer value,
+/// and a 32-bit decimal value.
+///
+/// The multiplication then becomes.
+/// `y = m * x >> 32 + b`
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Line {
+    slope: u64,
+    intercept: u64,
+}
+
+/// Compute the line slope.
+///
+/// This function has the nice property of being
+/// invariant by translation.
+/// `
+///   compute_slope(y0, y1)
+/// = compute_slope(y0 + X % 2^64, y1 + X % 2^64)
+/// `
+fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU32) -> u64 {
+    let dy = y1.wrapping_sub(y0);
+    let sign = dy <= (1 << 63);
+    let abs_dy = if sign {
+        y1.wrapping_sub(y0)
+    } else {
+        y0.wrapping_sub(y1)
+    };
+    if abs_dy >= 1 << 32 {
+        // This is outside of realm we handle.
+        // Let's just bail.
+        return 0u64;
+    }
+
+    let abs_slope = (abs_dy << 32) / num_vals.get() as u64;
+    if sign {
+        abs_slope
+    } else {
+        // The complement does indeed create the
+        // opposite decreasing slope...
+        //
+        // Intuitively (without the bitshifts and % u64::MAX)
+        // ```
+        //    (x + shift)*(u64::MAX - abs_slope)
+        // -  (x * (u64::MAX - abs_slope))
+        // = - shift * abs_slope
+        // ```
+        u64::MAX - abs_slope
+    }
+}
+
+impl Line {
+    #[inline(always)]
+    pub fn eval(&self, x: u32) -> u64 {
+        let linear_part = ((x as u64).wrapping_mul(self.slope) >> 32) as i32 as u64;
+        self.intercept.wrapping_add(linear_part)
+    }
+
+    // Same as train, but the intercept is only estimated from provided sample positions
+    pub fn estimate(sample_positions_and_values: &[(u64, u64)]) -> Self {
+        let first_val = sample_positions_and_values[0].1;
+        let last_val = sample_positions_and_values[sample_positions_and_values.len() - 1].1;
+        let num_vals = sample_positions_and_values[sample_positions_and_values.len() - 1].0 + 1;
+        Self::train_from(
+            first_val,
+            last_val,
+            num_vals as u32,
+            sample_positions_and_values.iter().cloned(),
+        )
+    }
+
+    // Intercept is only computed from provided positions
+    fn train_from(
+        first_val: u64,
+        last_val: u64,
+        num_vals: u32,
+        positions_and_values: impl Iterator<Item = (u64, u64)>,
+    ) -> Self {
+        // TODO replace with let else
+        let idx_last_val = if let Some(idx_last_val) = NonZeroU32::new(num_vals - 1) {
+            idx_last_val
+        } else {
+            return Line::default();
+        };
+
+        let y0 = first_val;
+        let y1 = last_val;
+
+        // We first independently pick our slope.
+        let slope = compute_slope(y0, y1, idx_last_val);
+
+        // We picked our slope. Note that it does not have to be perfect.
+        // Now we need to compute the best intercept.
+        //
+        // Intuitively, the best intercept is such that line passes through one of the
+        // `(i, ys[])`.
+        //
+        // The best intercept therefore has the form
+        // `y[i] - line.eval(i)` (using wrapping arithmetics).
+        // In other words, the best intercept is one of the `y - Line::eval(ys[i])`
+        // and our task is just to pick the one that minimizes our error.
+        //
+        // Without sorting our values, this is a difficult problem.
+        // We however rely on the following trick...
+        //
+        // We only focus on the case where the interpolation is half decent.
+        // If the line interpolation is doing its job on a dataset suited for it,
+        // we can hope that the maximum error won't be larger than `u64::MAX / 2`.
+        //
+        // In other words, even without the intercept the values `y - Line::eval(ys[i])` will all be
+        // within an interval that takes less than half of the modulo space of `u64`.
+        //
+        // Our task is therefore to identify this interval.
+        // Here we simply translate all of our values by `y0 - 2^63` and pick the min.
+        let mut line = Line {
+            slope,
+            intercept: 0,
+        };
+        let heuristic_shift = y0.wrapping_sub(MID_POINT);
+        line.intercept = positions_and_values
+            .map(|(pos, y)| y.wrapping_sub(line.eval(pos as u32)))
+            .min_by_key(|&val| val.wrapping_sub(heuristic_shift))
+            .unwrap_or(0u64); //< Never happens.
+        line
+    }
+
+    /// Returns a line that attemps to approximate a function
+    /// f: i in 0..[ys.num_vals()) -> ys[i].
+    ///
+    /// - The approximation is always lower than the actual value.
+    /// Or more rigorously, formally `f(i).wrapping_sub(ys[i])` is small
+    /// for any i in [0..ys.len()).
+    /// - It computes without panicking for any value of it.
+    ///
+    /// This function is only invariable by translation if all of the
+    /// `ys` are packaged into half of the space. (See heuristic below)
+    pub fn train(ys: &dyn ColumnValues) -> Self {
+        let first_val = ys.iter().next().unwrap();
+        let last_val = ys.iter().nth(ys.num_vals() as usize - 1).unwrap();
+        Self::train_from(
+            first_val,
+            last_val,
+            ys.num_vals(),
+            ys.iter().enumerate().map(|(pos, val)| (pos as u64, val)),
+        )
+    }
+}
+
+impl BinarySerializable for Line {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.slope).serialize(writer)?;
+        VInt(self.intercept).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let slope = VInt::deserialize(reader)?.0;
+        let intercept = VInt::deserialize(reader)?.0;
+        Ok(Line { slope, intercept })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::VecColumn;
+
+    /// Test training a line and ensuring that the maximum difference between
+    /// the data points and the line is `expected`.
+    ///
+    /// This function operates translation over the data for better coverage.
+    #[track_caller]
+    fn test_line_interpol_with_translation(ys: &[u64], expected: Option<u64>) {
+        let mut translations = vec![0, 100, u64::MAX / 2, u64::MAX, u64::MAX - 1];
+        translations.extend_from_slice(ys);
+        for translation in translations {
+            let translated_ys: Vec<u64> = ys
+                .iter()
+                .copied()
+                .map(|y| y.wrapping_add(translation))
+                .collect();
+            let largest_err = test_eval_max_err(&translated_ys);
+            assert_eq!(largest_err, expected);
+        }
+    }
+
+    fn test_eval_max_err(ys: &[u64]) -> Option<u64> {
+        let line = Line::train(&VecColumn::from(&ys));
+        ys.iter()
+            .enumerate()
+            .map(|(x, y)| y.wrapping_sub(line.eval(x as u32)))
+            .max()
+    }
+
+    #[test]
+    fn test_train() {
+        test_line_interpol_with_translation(&[11, 11, 11, 12, 12, 13], Some(1));
+        test_line_interpol_with_translation(&[13, 12, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
+        test_line_interpol_with_translation(&[u64::MAX - 1, 0, 0, 1], Some(1));
+        test_line_interpol_with_translation(&[u64::MAX - 1, u64::MAX, 0, 1], Some(0));
+        test_line_interpol_with_translation(&[0, 1, 2, 3, 5], Some(0));
+        test_line_interpol_with_translation(&[1, 2, 3, 4], Some(0));
+
+        let data: Vec<u64> = (0..255).collect();
+        test_line_interpol_with_translation(&data, Some(0));
+        let data: Vec<u64> = (0..255).map(|el| el * 2).collect();
+        test_line_interpol_with_translation(&data, Some(0));
+    }
+}

columnar/src/column_values/linear.rs

@@ -0,0 +1,230 @@
+use std::io::{self, Write};
+
+use common::{BinarySerializable, OwnedBytes};
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use super::line::Line;
+use super::serialize::NormalizedHeader;
+use super::{ColumnValues, FastFieldCodec, FastFieldCodecType};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct LinearReader {
+    data: OwnedBytes,
+    linear_params: LinearParams,
+    header: NormalizedHeader,
+}
+
+impl ColumnValues for LinearReader {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u64 {
+        let interpoled_val: u64 = self.linear_params.line.eval(doc);
+        let bitpacked_diff = self.linear_params.bit_unpacker.get(doc, &self.data);
+        interpoled_val.wrapping_add(bitpacked_diff)
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        // The LinearReader assumes a normalized vector.
+        0u64
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.header.max_value
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u32 {
+        self.header.num_vals
+    }
+}
+
+/// Fastfield serializer, which tries to guess values by linear interpolation
+/// and stores the difference bitpacked.
+pub struct LinearCodec;
+
+#[derive(Debug, Clone)]
+struct LinearParams {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+}
+
+impl BinarySerializable for LinearParams {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Self {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+        })
+    }
+}
+
+impl FastFieldCodec for LinearCodec {
+    const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::Linear;
+
+    type Reader = LinearReader;
+
+    /// Opens a fast field given a file.
+    fn open_from_bytes(mut data: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader> {
+        let linear_params = LinearParams::deserialize(&mut data)?;
+        Ok(LinearReader {
+            data,
+            linear_params,
+            header,
+        })
+    }
+
+    /// Creates a new fast field serializer.
+    fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()> {
+        assert_eq!(column.min_value(), 0);
+        let line = Line::train(column);
+
+        let max_offset_from_line = column
+            .iter()
+            .enumerate()
+            .map(|(pos, actual_value)| {
+                let calculated_value = line.eval(pos as u32);
+                actual_value.wrapping_sub(calculated_value)
+            })
+            .max()
+            .unwrap();
+
+        let num_bits = compute_num_bits(max_offset_from_line);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        linear_params.serialize(write)?;
+
+        let mut bit_packer = BitPacker::new();
+        for (pos, actual_value) in column.iter().enumerate() {
+            let calculated_value = line.eval(pos as u32);
+            let offset = actual_value.wrapping_sub(calculated_value);
+            bit_packer.write(offset, num_bits, write)?;
+        }
+        bit_packer.close(write)?;
+
+        Ok(())
+    }
+
+    /// estimation for linear interpolation is hard because, you don't know
+    /// where the local maxima for the deviation of the calculated value are and
+    /// the offset to shift all values to >=0 is also unknown.
+    #[allow(clippy::question_mark)]
+    fn estimate(column: &dyn ColumnValues) -> Option<f32> {
+        if column.num_vals() < 3 {
+            return None; // disable compressor for this case
+        }
+
+        let limit_num_vals = column.num_vals().min(100_000);
+
+        let num_samples = 100;
+        let step_size = (limit_num_vals / num_samples).max(1); // 20 samples
+        let mut sample_positions_and_values: Vec<_> = Vec::new();
+        for (pos, val) in column.iter().enumerate().step_by(step_size as usize) {
+            sample_positions_and_values.push((pos as u64, val));
+        }
+
+        let line = Line::estimate(&sample_positions_and_values);
+
+        let estimated_bit_width = sample_positions_and_values
+            .into_iter()
+            .map(|(pos, actual_value)| {
+                let interpolated_val = line.eval(pos as u32);
+                actual_value.wrapping_sub(interpolated_val)
+            })
+            .map(|diff| ((diff as f32 * 1.5) * 2.0) as u64)
+            .map(compute_num_bits)
+            .max()
+            .unwrap_or(0);
+
+        // Extrapolate to whole column
+        let num_bits = (estimated_bit_width as u64 * column.num_vals() as u64) + 64;
+        let num_bits_uncompressed = 64 * column.num_vals();
+        Some(num_bits as f32 / num_bits_uncompressed as f32)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use rand::RngCore;
+
+    use super::*;
+    use crate::column_values::tests;
+
+    fn create_and_validate(data: &[u64], name: &str) -> Option<(f32, f32)> {
+        tests::create_and_validate::<LinearCodec>(data, name)
+    }
+
+    #[test]
+    fn test_compression() {
+        let data = (10..=6_000_u64).collect::<Vec<_>>();
+        let (estimate, actual_compression) =
+            create_and_validate(&data, "simple monotonically large").unwrap();
+
+        assert_le!(actual_compression, 0.001);
+        assert_le!(estimate, 0.02);
+    }
+
+    #[test]
+    fn test_with_codec_datasets() {
+        let data_sets = tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate(&data, name);
+            data.reverse();
+            create_and_validate(&data, name);
+        }
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_large_amplitude() {
+        let data = vec![
+            i64::MAX as u64 / 2,
+            i64::MAX as u64 / 3,
+            i64::MAX as u64 / 2,
+        ];
+
+        create_and_validate(&data, "large amplitude");
+    }
+
+    #[test]
+    fn overflow_error_test() {
+        let data = vec![1572656989877777, 1170935903116329, 720575940379279, 0];
+        create_and_validate(&data, "overflow test");
+    }
+
+    #[test]
+    fn linear_interpol_fast_concave_data() {
+        let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
+        create_and_validate(&data, "concave data");
+    }
+    #[test]
+    fn linear_interpol_fast_convex_data() {
+        let data = vec![0, 40, 60, 70, 75, 77];
+        create_and_validate(&data, "convex data");
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_simple() {
+        let data = (10..=20_u64).collect::<Vec<_>>();
+        create_and_validate(&data, "simple monotonically");
+    }
+
+    #[test]
+    fn linear_interpol_fast_field_rand() {
+        let mut rng = rand::thread_rng();
+        for _ in 0..50 {
+            let mut data = (0..10_000).map(|_| rng.next_u64()).collect::<Vec<_>>();
+            create_and_validate(&data, "random");
+            data.reverse();
+            create_and_validate(&data, "random");
+        }
+    }
+}

columnar/src/column_values/main.rs

@@ -0,0 +1,222 @@
+#[macro_use]
+extern crate prettytable;
+use std::collections::HashSet;
+use std::env;
+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 prettytable::{Cell, Row, Table};
+
+fn print_set_stats(ip_addrs: &[u128]) {
+    println!("NumIps\t{}", ip_addrs.len());
+    let ip_addr_set: HashSet<u128> = ip_addrs.iter().cloned().collect();
+    println!("NumUniqueIps\t{}", ip_addr_set.len());
+    let ratio_unique = ip_addr_set.len() as f64 / ip_addrs.len() as f64;
+    println!("RatioUniqueOverTotal\t{ratio_unique:.4}");
+
+    // histogram
+    let mut ip_addrs = ip_addrs.to_vec();
+    ip_addrs.sort();
+    let mut cnts: Vec<usize> = ip_addrs
+        .into_iter()
+        .dedup_with_count()
+        .map(|(cnt, _)| cnt)
+        .collect();
+    cnts.sort();
+
+    let top_256_cnt: usize = cnts.iter().rev().take(256).sum();
+    let top_128_cnt: usize = cnts.iter().rev().take(128).sum();
+    let top_64_cnt: usize = cnts.iter().rev().take(64).sum();
+    let top_8_cnt: usize = cnts.iter().rev().take(8).sum();
+    let total: usize = cnts.iter().sum();
+
+    println!("{}", total);
+    println!("{}", top_256_cnt);
+    println!("{}", top_128_cnt);
+    println!("Percentage Top8 {:02}", top_8_cnt as f32 / total as f32);
+    println!("Percentage Top64 {:02}", top_64_cnt as f32 / total as f32);
+    println!("Percentage Top128 {:02}", top_128_cnt as f32 / total as f32);
+    println!("Percentage Top256 {:02}", top_256_cnt as f32 / total as f32);
+
+    let mut cnts: Vec<(usize, usize)> = cnts.into_iter().dedup_with_count().collect();
+    cnts.sort_by(|a, b| {
+        if a.1 == b.1 {
+            a.0.cmp(&b.0)
+        } else {
+            b.1.cmp(&a.1)
+        }
+    });
+}
+
+fn ip_dataset() -> Vec<u128> {
+    let mut ip_addr_v4 = 0;
+
+    let stdin = std::io::stdin();
+    let ip_addrs: Vec<u128> = stdin
+        .lock()
+        .lines()
+        .flat_map(|line| {
+            let line = line.unwrap();
+            let line = line.trim();
+            let ip_addr = IpAddr::from_str(line.trim()).ok()?;
+            if ip_addr.is_ipv4() {
+                ip_addr_v4 += 1;
+            }
+            let ip_addr_v6: Ipv6Addr = match ip_addr {
+                IpAddr::V4(v4) => v4.to_ipv6_mapped(),
+                IpAddr::V6(v6) => v6,
+            };
+            Some(ip_addr_v6)
+        })
+        .map(|ip_v6| u128::from_be_bytes(ip_v6.octets()))
+        .collect();
+
+    println!("IpAddrsAny\t{}", ip_addrs.len());
+    println!("IpAddrsV4\t{}", ip_addr_v4);
+
+    ip_addrs
+}
+
+fn bench_ip() {
+    let dataset = ip_dataset();
+    print_set_stats(&dataset);
+
+    // Chunks
+    {
+        let mut data = vec![];
+        for dataset in dataset.chunks(500_000) {
+            serialize_u128(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
+        }
+        let compression = data.len() as f64 / (dataset.len() * 16) as f64;
+        println!("Compression 50_000 chunks {:.4}", compression);
+        println!(
+            "Num Bits per elem {:.2}",
+            (data.len() * 8) as f32 / dataset.len() as f32
+        );
+    }
+
+    let mut data = vec![];
+    {
+        print_time!("creation");
+        serialize_u128(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
+    }
+
+    let compression = data.len() as f64 / (dataset.len() * 16) as f64;
+    println!("Compression {:.2}", compression);
+    println!(
+        "Num Bits per elem {:.2}",
+        (data.len() * 8) as f32 / dataset.len() as f32
+    );
+
+    let decompressor = open_u128::<u128>(OwnedBytes::new(data)).unwrap();
+    // Sample some ranges
+    let mut doc_values = Vec::new();
+    for value in dataset.iter().take(1110).skip(1100).cloned() {
+        doc_values.clear();
+        print_time!("get range");
+        decompressor.get_docids_for_value_range(
+            value..=value,
+            0..decompressor.num_vals(),
+            &mut doc_values,
+        );
+        println!("{:?}", doc_values.len());
+    }
+}
+
+fn main() {
+    if env::args().nth(1).unwrap() == "bench_ip" {
+        bench_ip();
+        return;
+    }
+
+    let mut table = Table::new();
+
+    // Add a row per time
+    table.add_row(row!["", "Compression Ratio", "Compression Estimation"]);
+
+    for (data, data_set_name) in get_codec_test_data_sets() {
+        let results: Vec<(f32, f32, FastFieldCodecType)> = [
+            serialize_with_codec(&data, FastFieldCodecType::Bitpacked),
+            serialize_with_codec(&data, FastFieldCodecType::Linear),
+            serialize_with_codec(&data, FastFieldCodecType::BlockwiseLinear),
+        ]
+        .into_iter()
+        .flatten()
+        .collect();
+        let best_compression_ratio_codec = results
+            .iter()
+            .min_by(|&res1, &res2| res1.partial_cmp(res2).unwrap())
+            .cloned()
+            .unwrap();
+
+        table.add_row(Row::new(vec![Cell::new(data_set_name).style_spec("Bbb")]));
+        for (est, comp, codec_type) in results {
+            let est_cell = est.to_string();
+            let ratio_cell = comp.to_string();
+            let style = if comp == best_compression_ratio_codec.1 {
+                "Fb"
+            } else {
+                ""
+            };
+            table.add_row(Row::new(vec![
+                Cell::new(&format!("{codec_type:?}")).style_spec("bFg"),
+                Cell::new(&ratio_cell).style_spec(style),
+                Cell::new(&est_cell).style_spec(""),
+            ]));
+        }
+    }
+
+    table.printstd();
+}
+
+pub fn get_codec_test_data_sets() -> Vec<(Vec<u64>, &'static str)> {
+    let mut data_and_names = vec![];
+
+    let data = (1000..=200_000_u64).collect::<Vec<_>>();
+    data_and_names.push((data, "Autoincrement"));
+
+    let mut current_cumulative = 0;
+    let data = (1..=200_000_u64)
+        .map(|num| {
+            let num = (num as f32 + num as f32).log10() as u64;
+            current_cumulative += num;
+            current_cumulative
+        })
+        .collect::<Vec<_>>();
+    // let data = (1..=200000_u64).map(|num| num + num).collect::<Vec<_>>();
+    data_and_names.push((data, "Monotonically increasing concave"));
+
+    let mut current_cumulative = 0;
+    let data = (1..=200_000_u64)
+        .map(|num| {
+            let num = (200_000.0 - num as f32).log10() as u64;
+            current_cumulative += num;
+            current_cumulative
+        })
+        .collect::<Vec<_>>();
+    data_and_names.push((data, "Monotonically increasing convex"));
+
+    let data = (1000..=200_000_u64)
+        .map(|num| num + rand::random::<u8>() as u64)
+        .collect::<Vec<_>>();
+    data_and_names.push((data, "Almost monotonically increasing"));
+
+    data_and_names
+}
+
+pub fn serialize_with_codec(
+    data: &[u64],
+    codec_type: FastFieldCodecType,
+) -> Option<(f32, f32, FastFieldCodecType)> {
+    let col = VecColumn::from(data);
+    let estimation = fastfield_codecs::estimate(&col, codec_type)?;
+    let mut out = Vec::new();
+    fastfield_codecs::serialize(&col, &mut out, &[codec_type]).ok()?;
+    let actual_compression = out.len() as f32 / (col.num_vals() * 8) as f32;
+    Some((estimation, actual_compression, codec_type))
+}

columnar/src/column_values/mod.rs

@@ -0,0 +1,333 @@
+#![warn(missing_docs)]
+#![cfg_attr(all(feature = "unstable", test), feature(test))]
+
+//! # `fastfield_codecs`
+//!
+//! - Columnar storage of data for tantivy [`Column`].
+//! - Encode data in different codecs.
+//! - Monotonically map values to u64/u128
+
+#[cfg(test)]
+mod tests;
+
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::{BinarySerializable, OwnedBytes};
+use compact_space::CompactSpaceDecompressor;
+use monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+    StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
+};
+use serialize::{Header, U128Header};
+
+mod bitpacked;
+mod blockwise_linear;
+mod compact_space;
+mod line;
+mod linear;
+pub(crate) mod monotonic_mapping;
+// mod monotonic_mapping_u128;
+
+mod column;
+mod column_with_cardinality;
+mod gcd;
+pub mod serialize;
+
+pub use self::column::{monotonic_map_column, ColumnValues, IterColumn, VecColumn};
+pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
+// pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
+pub use self::serialize::{serialize_and_load, serialize_column_values, NormalizedHeader};
+use crate::column_values::bitpacked::BitpackedCodec;
+use crate::column_values::blockwise_linear::BlockwiseLinearCodec;
+use crate::column_values::linear::LinearCodec;
+
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
+pub enum FastFieldCodecType {
+    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
+    /// `column.max_value() - column.min_value()`
+    Bitpacked = 1,
+    /// Linear interpolation puts a line between the first and last value and then bitpacks the
+    /// values by the offset from the line. The number of bits is defined by the max deviation from
+    /// the line.
+    Linear = 2,
+    /// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
+    BlockwiseLinear = 3,
+}
+
+impl BinarySerializable for FastFieldCodecType {
+    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
+        self.to_code().serialize(wrt)
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let code = u8::deserialize(reader)?;
+        let codec_type: Self = Self::from_code(code)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
+        Ok(codec_type)
+    }
+}
+
+impl FastFieldCodecType {
+    pub(crate) fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn from_code(code: u8) -> Option<Self> {
+        match code {
+            1 => Some(Self::Bitpacked),
+            2 => Some(Self::Linear),
+            3 => Some(Self::BlockwiseLinear),
+            _ => None,
+        }
+    }
+}
+
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+/// Available codecs to use to encode the u128 (via [`MonotonicallyMappableToU128`]) converted data.
+pub enum U128FastFieldCodecType {
+    /// This codec takes a large number space (u128) and reduces it to a compact number space, by
+    /// removing the holes.
+    CompactSpace = 1,
+}
+
+impl BinarySerializable for U128FastFieldCodecType {
+    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
+        self.to_code().serialize(wrt)
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let code = u8::deserialize(reader)?;
+        let codec_type: Self = Self::from_code(code)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
+        Ok(codec_type)
+    }
+}
+
+impl U128FastFieldCodecType {
+    pub(crate) fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn from_code(code: u8) -> Option<Self> {
+        match code {
+            1 => Some(Self::CompactSpace),
+            _ => None,
+        }
+    }
+}
+
+/// Returns the correct codec reader wrapped in the `Arc` for the data.
+// pub fn open_u128<Item: MonotonicallyMappableToU128>(
+//     bytes: OwnedBytes,
+// ) -> io::Result<Arc<dyn Column<Item>>> {
+//     todo!();
+//     // let (bytes, _format_version) = read_format_version(bytes)?;
+//     // let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
+//     // let header = U128Header::deserialize(&mut bytes)?;
+//     // assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
+//     // let reader = CompactSpaceDecompressor::open(bytes)?;
+//     // let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<Item>> =
+//     //     StrictlyMonotonicMappingToInternal::<Item>::new().into();
+//     // Ok(Arc::new(monotonic_map_column(reader, inverted)))
+// }
+
+/// Returns the correct codec reader wrapped in the `Arc` for the data.
+pub fn open_u64_mapped<T: MonotonicallyMappableToU64>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let header = Header::deserialize(&mut bytes)?;
+    match header.codec_type {
+        FastFieldCodecType::Bitpacked => open_specific_codec::<BitpackedCodec, _>(bytes, &header),
+        FastFieldCodecType::Linear => open_specific_codec::<LinearCodec, _>(bytes, &header),
+        FastFieldCodecType::BlockwiseLinear => {
+            open_specific_codec::<BlockwiseLinearCodec, _>(bytes, &header)
+        }
+    }
+}
+
+fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+    header: &Header,
+) -> io::Result<Arc<dyn ColumnValues<Item>>> {
+    let normalized_header = header.normalized();
+    let reader = C::open_from_bytes(bytes, normalized_header)?;
+    let min_value = header.min_value;
+    if let Some(gcd) = header.gcd {
+        let mapping = StrictlyMonotonicMappingInverter::from(
+            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
+        );
+        Ok(Arc::new(monotonic_map_column(reader, mapping)))
+    } else {
+        let mapping = StrictlyMonotonicMappingInverter::from(
+            StrictlyMonotonicMappingToInternalBaseval::new(min_value),
+        );
+        Ok(Arc::new(monotonic_map_column(reader, mapping)))
+    }
+}
+
+/// The FastFieldSerializerEstimate trait is required on all variants
+/// of fast field compressions, to decide which one to choose.
+pub(crate) trait FastFieldCodec: 'static {
+    /// A codex needs to provide a unique name and id, which is
+    /// used for debugging and de/serialization.
+    const CODEC_TYPE: FastFieldCodecType;
+
+    type Reader: ColumnValues<u64> + 'static;
+
+    /// Reads the metadata and returns the CodecReader
+    fn open_from_bytes(bytes: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader>;
+
+    /// Serializes the data using the serializer into write.
+    ///
+    /// The column iterator should be preferred over using column `get_val` method for
+    /// performance reasons.
+    fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()>;
+
+    /// Returns an estimate of the compression ratio.
+    /// If the codec is not applicable, returns `None`.
+    ///
+    /// The baseline is uncompressed 64bit data.
+    ///
+    /// It could make sense to also return a value representing
+    /// computational complexity.
+    fn estimate(column: &dyn ColumnValues) -> Option<f32>;
+}
+
+/// The list of all available codecs for u64 convertible data.
+pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
+    FastFieldCodecType::Bitpacked,
+    FastFieldCodecType::BlockwiseLinear,
+    FastFieldCodecType::Linear,
+];
+
+#[cfg(all(test, feature = "unstable"))]
+mod bench {
+    use std::sync::Arc;
+
+    use common::OwnedBytes;
+    use rand::rngs::StdRng;
+    use rand::{Rng, SeedableRng};
+    use test::{self, Bencher};
+
+    use super::*;
+
+    fn get_data() -> Vec<u64> {
+        let mut rng = StdRng::seed_from_u64(2u64);
+        let mut data: Vec<_> = (100..55000_u64)
+            .map(|num| num + rng.gen::<u8>() as u64)
+            .collect();
+        data.push(99_000);
+        data.insert(1000, 2000);
+        data.insert(2000, 100);
+        data.insert(3000, 4100);
+        data.insert(4000, 100);
+        data.insert(5000, 800);
+        data
+    }
+
+    #[inline(never)]
+    fn value_iter() -> impl Iterator<Item = u64> {
+        0..20_000
+    }
+    fn get_reader_for_bench<Codec: FastFieldCodec>(data: &[u64]) -> Codec::Reader {
+        let mut bytes = Vec::new();
+        let min_value = *data.iter().min().unwrap();
+        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
+        let col = VecColumn::from(&data);
+        let normalized_header = NormalizedHeader {
+            num_vals: col.num_vals(),
+            max_value: col.max_value(),
+        };
+        Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
+        Codec::open_from_bytes(OwnedBytes::new(bytes), normalized_header).unwrap()
+    }
+    fn bench_get<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
+        let col = get_reader_for_bench::<Codec>(data);
+        b.iter(|| {
+            let mut sum = 0u64;
+            for pos in value_iter() {
+                let val = col.get_val(pos as u32);
+                sum = sum.wrapping_add(val);
+            }
+            sum
+        });
+    }
+
+    #[inline(never)]
+    fn bench_get_dynamic_helper(b: &mut Bencher, col: Arc<dyn ColumnValues>) {
+        b.iter(|| {
+            let mut sum = 0u64;
+            for pos in value_iter() {
+                let val = col.get_val(pos as u32);
+                sum = sum.wrapping_add(val);
+            }
+            sum
+        });
+    }
+
+    fn bench_get_dynamic<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
+        let col = Arc::new(get_reader_for_bench::<Codec>(data));
+        bench_get_dynamic_helper(b, col);
+    }
+    fn bench_create<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
+        let min_value = *data.iter().min().unwrap();
+        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
+
+        let mut bytes = Vec::new();
+        b.iter(|| {
+            bytes.clear();
+            Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
+        });
+    }
+
+    #[bench]
+    fn bench_fastfield_bitpack_create(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_create::<BitpackedCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_linearinterpol_create(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_create::<LinearCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_multilinearinterpol_create(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_create::<BlockwiseLinearCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_bitpack_get(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get::<BitpackedCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_bitpack_get_dynamic(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get_dynamic::<BitpackedCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_linearinterpol_get(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get::<LinearCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_linearinterpol_get_dynamic(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get_dynamic::<LinearCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_multilinearinterpol_get(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get::<BlockwiseLinearCodec>(b, &data);
+    }
+    #[bench]
+    fn bench_fastfield_multilinearinterpol_get_dynamic(b: &mut Bencher) {
+        let data: Vec<_> = get_data();
+        bench_get_dynamic::<BlockwiseLinearCodec>(b, &data);
+    }
+}

columnar/src/column_values/monotonic_mapping.rs

@@ -0,0 +1,264 @@
+use std::marker::PhantomData;
+
+use fastdivide::DividerU64;
+
+use crate::RowId;
+
+/// Monotonic maps a value to u64 value space.
+/// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
+pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
+    /// Converts a value to u64.
+    ///
+    /// Internally all fast field values are encoded as u64.
+    fn to_u64(self) -> u64;
+
+    /// Converts a value from u64
+    ///
+    /// Internally all fast field values are encoded as u64.
+    /// **Note: To be used for converting encoded Term, Posting values.**
+    fn from_u64(val: u64) -> Self;
+}
+
+/// Values need to be strictly monotonic mapped to a `Internal` value (u64 or u128) that can be
+/// used in fast field codecs.
+///
+/// The monotonic mapping is required so that `PartialOrd` can be used on `Internal` without
+/// converting to `External`.
+///
+/// All strictly monotonic functions are invertible because they are guaranteed to have a one-to-one
+/// mapping from their range to their domain. The `inverse` method is required when opening a codec,
+/// so a value can be converted back to its original domain (e.g. ip address or f64) from its
+/// internal representation.
+pub trait StrictlyMonotonicFn<External, Internal> {
+    /// Strictly monotonically maps the value from External to Internal.
+    fn mapping(&self, inp: External) -> Internal;
+    /// Inverse of `mapping`. Maps the value from Internal to External.
+    fn inverse(&self, out: Internal) -> External;
+}
+
+/// Inverts a strictly monotonic mapping from `StrictlyMonotonicFn<A, B>` to
+/// `StrictlyMonotonicFn<B, A>`.
+///
+/// # Warning
+///
+/// This type comes with a footgun. A type being strictly monotonic does not impose that the inverse
+/// mapping is strictly monotonic over the entire space External. e.g. a -> a * 2. Use at your own
+/// risks.
+pub(crate) struct StrictlyMonotonicMappingInverter<T> {
+    orig_mapping: T,
+}
+impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
+    fn from(orig_mapping: T) -> Self {
+        Self { orig_mapping }
+    }
+}
+
+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)
+    }
+}
+
+/// Applies the strictly monotonic mapping from `T` without any additional changes.
+pub(crate) struct StrictlyMonotonicMappingToInternal<T> {
+    _phantom: PhantomData<T>,
+}
+
+impl<T> StrictlyMonotonicMappingToInternal<T> {
+    pub(crate) fn new() -> StrictlyMonotonicMappingToInternal<T> {
+        Self {
+            _phantom: PhantomData,
+        }
+    }
+}
+
+// TODO
+// 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)
+//     }
+// }
+
+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)
+    }
+}
+
+/// Mapping dividing by  gcd and a base value.
+///
+/// The function is assumed to be only called on values divided by passed
+/// gcd value. (It is necessary for the function to be monotonic.)
+pub(crate) struct StrictlyMonotonicMappingToInternalGCDBaseval {
+    gcd_divider: DividerU64,
+    gcd: u64,
+    min_value: u64,
+}
+impl StrictlyMonotonicMappingToInternalGCDBaseval {
+    pub(crate) fn new(gcd: u64, min_value: u64) -> Self {
+        let gcd_divider = DividerU64::divide_by(gcd);
+        Self {
+            gcd_divider,
+            gcd,
+            min_value,
+        }
+    }
+}
+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)
+    }
+}
+
+/// Strictly monotonic mapping with a base value.
+pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
+    min_value: u64,
+}
+impl StrictlyMonotonicMappingToInternalBaseval {
+    #[inline(always)]
+    pub(crate) fn new(min_value: u64) -> Self {
+        Self { min_value }
+    }
+}
+
+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
+    }
+}
+
+impl MonotonicallyMappableToU64 for i64 {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        common::i64_to_u64(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        common::u64_to_i64(val)
+    }
+}
+
+impl MonotonicallyMappableToU64 for bool {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        u64::from(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        val > 0
+    }
+}
+
+impl MonotonicallyMappableToU64 for RowId {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        u64::from(self)
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> RowId {
+        val as RowId
+    }
+}
+
+// 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)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::*;
+
+    #[test]
+    fn strictly_monotonic_test() {
+        // identity mapping
+        test_round_trip(&StrictlyMonotonicMappingToInternal::<u64>::new(), 100u64);
+        // round trip to i64
+        test_round_trip(&StrictlyMonotonicMappingToInternal::<i64>::new(), 100u64);
+        // TODO
+        // identity mapping
+        // test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
+
+        // base value to i64 round trip
+        let mapping = StrictlyMonotonicMappingToInternalBaseval::new(100);
+        test_round_trip::<_, _, u64>(&mapping, 100i64);
+        // base value and gcd to u64 round trip
+        let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100);
+        test_round_trip::<_, _, u64>(&mapping, 100u64);
+    }
+
+    fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L>(
+        mapping: &T,
+        test_val: K,
+    ) {
+        assert_eq!(mapping.inverse(mapping.mapping(test_val)), test_val);
+    }
+}

columnar/src/column_values/monotonic_mapping_u128.rs

@@ -0,0 +1,40 @@
+use std::net::Ipv6Addr;
+
+/// Montonic maps a value to u128 value space
+/// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
+pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Send + Sync {
+    /// Converts a value to u128.
+    ///
+    /// Internally all fast field values are encoded as u64.
+    fn to_u128(self) -> u128;
+
+    /// Converts a value from u128
+    ///
+    /// Internally all fast field values are encoded as u64.
+    /// **Note: To be used for converting encoded Term, Posting values.**
+    fn from_u128(val: u128) -> Self;
+}
+
+impl MonotonicallyMappableToU128 for u128 {
+    fn to_u128(self) -> u128 {
+        self
+    }
+
+    fn from_u128(val: u128) -> Self {
+        val
+    }
+}
+
+impl MonotonicallyMappableToU128 for Ipv6Addr {
+    fn to_u128(self) -> u128 {
+        ip_to_u128(self)
+    }
+
+    fn from_u128(val: u128) -> Self {
+        Ipv6Addr::from(val.to_be_bytes())
+    }
+}
+
+fn ip_to_u128(ip_addr: Ipv6Addr) -> u128 {
+    u128::from_be_bytes(ip_addr.octets())
+}

columnar/src/column_values/serialize.rs

@@ -0,0 +1,343 @@
+// Copyright (C) 2022 Quickwit, Inc.
+//
+// Quickwit is offered under the AGPL v3.0 and as commercial software.
+// For commercial licensing, contact us at hello@quickwit.io.
+//
+// AGPL:
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as
+// published by the Free Software Foundation, either version 3 of the
+// License, or (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see <http://www.gnu.org/licenses/>.
+
+use std::io;
+use std::num::NonZeroU64;
+use std::sync::Arc;
+
+use common::{BinarySerializable, OwnedBytes, VInt};
+use log::warn;
+
+use super::bitpacked::BitpackedCodec;
+use super::blockwise_linear::BlockwiseLinearCodec;
+use super::linear::LinearCodec;
+use super::monotonic_mapping::{
+    StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
+    StrictlyMonotonicMappingToInternalGCDBaseval,
+};
+use super::{
+    monotonic_map_column, ColumnValues, FastFieldCodec, FastFieldCodecType,
+    MonotonicallyMappableToU64, U128FastFieldCodecType, VecColumn, ALL_CODEC_TYPES,
+};
+
+/// The normalized header gives some parameters after applying the following
+/// normalization of the vector:
+/// `val -> (val - min_value) / gcd`
+///
+/// By design, after normalization, `min_value = 0` and `gcd = 1`.
+#[derive(Debug, Copy, Clone)]
+pub struct NormalizedHeader {
+    /// The number of values in the underlying column.
+    pub num_vals: u32,
+    /// The max value of the underlying column.
+    pub max_value: u64,
+}
+
+#[derive(Debug, Copy, Clone)]
+pub(crate) struct Header {
+    pub num_vals: u32,
+    pub min_value: u64,
+    pub max_value: u64,
+    pub gcd: Option<NonZeroU64>,
+    pub codec_type: FastFieldCodecType,
+}
+
+impl Header {
+    pub fn normalized(self) -> NormalizedHeader {
+        let gcd = self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
+        let gcd_min_val_mapping =
+            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, self.min_value);
+
+        let max_value = gcd_min_val_mapping.mapping(self.max_value);
+        NormalizedHeader {
+            num_vals: self.num_vals,
+            max_value,
+        }
+    }
+
+    pub(crate) fn normalize_column<C: ColumnValues>(&self, from_column: C) -> impl ColumnValues {
+        normalize_column(from_column, self.min_value, self.gcd)
+    }
+
+    pub fn compute_header(
+        column: impl ColumnValues<u64>,
+        codecs: &[FastFieldCodecType],
+    ) -> Option<Header> {
+        let num_vals = column.num_vals();
+        let min_value = column.min_value();
+        let max_value = column.max_value();
+        let gcd = super::gcd::find_gcd(column.iter().map(|val| val - min_value))
+            .filter(|gcd| gcd.get() > 1u64);
+        let normalized_column = normalize_column(column, min_value, gcd);
+        let codec_type = detect_codec(normalized_column, codecs)?;
+        Some(Header {
+            num_vals,
+            min_value,
+            max_value,
+            gcd,
+            codec_type,
+        })
+    }
+}
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+pub(crate) struct U128Header {
+    pub num_vals: u32,
+    pub codec_type: U128FastFieldCodecType,
+}
+
+impl BinarySerializable for U128Header {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.num_vals as u64).serialize(writer)?;
+        self.codec_type.serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let num_vals = VInt::deserialize(reader)?.0 as u32;
+        let codec_type = U128FastFieldCodecType::deserialize(reader)?;
+        Ok(U128Header {
+            num_vals,
+            codec_type,
+        })
+    }
+}
+
+fn normalize_column<C: ColumnValues>(
+    from_column: C,
+    min_value: u64,
+    gcd: Option<NonZeroU64>,
+) -> impl ColumnValues {
+    let gcd = gcd.map(|gcd| gcd.get()).unwrap_or(1);
+    let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, min_value);
+    monotonic_map_column(from_column, mapping)
+}
+
+impl BinarySerializable for Header {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.num_vals as u64).serialize(writer)?;
+        VInt(self.min_value).serialize(writer)?;
+        VInt(self.max_value - self.min_value).serialize(writer)?;
+        if let Some(gcd) = self.gcd {
+            VInt(gcd.get()).serialize(writer)?;
+        } else {
+            VInt(0u64).serialize(writer)?;
+        }
+        self.codec_type.serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let num_vals = VInt::deserialize(reader)?.0 as u32;
+        let min_value = VInt::deserialize(reader)?.0;
+        let amplitude = VInt::deserialize(reader)?.0;
+        let max_value = min_value + amplitude;
+        let gcd_u64 = VInt::deserialize(reader)?.0;
+        let codec_type = FastFieldCodecType::deserialize(reader)?;
+        Ok(Header {
+            num_vals,
+            min_value,
+            max_value,
+            gcd: NonZeroU64::new(gcd_u64),
+            codec_type,
+        })
+    }
+}
+
+/// Return estimated compression for given codec in the value range [0.0..1.0], where 1.0 means no
+/// compression.
+pub(crate) fn estimate<T: MonotonicallyMappableToU64>(
+    typed_column: impl ColumnValues<T>,
+    codec_type: FastFieldCodecType,
+) -> Option<f32> {
+    let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
+    let min_value = column.min_value();
+    let gcd = super::gcd::find_gcd(column.iter().map(|val| val - min_value))
+        .filter(|gcd| gcd.get() > 1u64);
+    let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(
+        gcd.map(|gcd| gcd.get()).unwrap_or(1u64),
+        min_value,
+    );
+    let normalized_column = monotonic_map_column(&column, mapping);
+    match codec_type {
+        FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&normalized_column),
+        FastFieldCodecType::Linear => LinearCodec::estimate(&normalized_column),
+        FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&normalized_column),
+    }
+}
+
+// TODO
+/// Serializes u128 values with the compact space codec.
+// pub fn serialize_u128_new<F: Fn() -> I, I: Iterator<Item = u128>>(
+//     value_index: ColumnIndex,
+//     iter_gen: F,
+//     num_vals: u32,
+//     output: &mut impl io::Write,
+// ) -> io::Result<()> {
+//     let header = U128Header {
+//         num_vals,
+//         codec_type: U128FastFieldCodecType::CompactSpace,
+//     };
+//     header.serialize(output)?;
+//     let compressor = CompactSpaceCompressor::train_from(iter_gen(), num_vals);
+//     compressor.compress_into(iter_gen(), output).unwrap();
+
+//     let null_index_footer = ColumnFooter {
+//         cardinality: value_index.get_cardinality(),
+//         null_index_codec: NullIndexCodec::Full,
+//         null_index_byte_range: 0..0,
+//     };
+//     append_null_index_footer(output, null_index_footer)?;
+//     append_format_version(output)?;
+
+//     Ok(())
+// }
+
+/// Serializes the column with the codec with the best estimate on the data.
+pub fn serialize_column_values<T: MonotonicallyMappableToU64>(
+    typed_column: impl ColumnValues<T>,
+    codecs: &[FastFieldCodecType],
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
+    let header = Header::compute_header(&column, codecs).ok_or_else(|| {
+        io::Error::new(
+            io::ErrorKind::InvalidInput,
+            format!(
+                "Data cannot be serialized with this list of codec. {:?}",
+                codecs
+            ),
+        )
+    })?;
+    header.serialize(output)?;
+    let normalized_column = header.normalize_column(column);
+    assert_eq!(normalized_column.min_value(), 0u64);
+    serialize_given_codec(normalized_column, header.codec_type, output)?;
+    Ok(())
+}
+
+fn detect_codec(
+    column: impl ColumnValues<u64>,
+    codecs: &[FastFieldCodecType],
+) -> Option<FastFieldCodecType> {
+    let mut estimations = Vec::new();
+    for &codec in codecs {
+        let estimation_opt = match codec {
+            FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&column),
+            FastFieldCodecType::Linear => LinearCodec::estimate(&column),
+            FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&column),
+        };
+        if let Some(estimation) = estimation_opt {
+            estimations.push((estimation, codec));
+        }
+    }
+    if let Some(broken_estimation) = estimations.iter().find(|estimation| estimation.0.is_nan()) {
+        warn!(
+            "broken estimation for fast field codec {:?}",
+            broken_estimation.1
+        );
+    }
+    // removing nan values for codecs with broken calculations, and max values which disables
+    // codecs
+    estimations.retain(|estimation| !estimation.0.is_nan() && estimation.0 != f32::MAX);
+    estimations.sort_by(|(score_left, _), (score_right, _)| score_left.total_cmp(score_right));
+    Some(estimations.first()?.1)
+}
+
+pub(crate) fn serialize_given_codec(
+    column: impl ColumnValues<u64>,
+    codec_type: FastFieldCodecType,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    match codec_type {
+        FastFieldCodecType::Bitpacked => {
+            BitpackedCodec::serialize(&column, output)?;
+        }
+        FastFieldCodecType::Linear => {
+            LinearCodec::serialize(&column, output)?;
+        }
+        FastFieldCodecType::BlockwiseLinear => {
+            BlockwiseLinearCodec::serialize(&column, output)?;
+        }
+    }
+    Ok(())
+}
+
+/// Helper function to serialize a column (autodetect from all codecs) and then open it
+pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
+    column: &[T],
+) -> Arc<dyn ColumnValues<T>> {
+    let mut buffer = Vec::new();
+    super::serialize_column_values(&VecColumn::from(&column), &ALL_CODEC_TYPES, &mut buffer)
+        .unwrap();
+    super::open_u64_mapped(OwnedBytes::new(buffer)).unwrap()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_serialize_deserialize_u128_header() {
+        let original = U128Header {
+            num_vals: 11,
+            codec_type: U128FastFieldCodecType::CompactSpace,
+        };
+        let mut out = Vec::new();
+        original.serialize(&mut out).unwrap();
+        let restored = U128Header::deserialize(&mut &out[..]).unwrap();
+        assert_eq!(restored, original);
+    }
+
+    #[test]
+    fn test_serialize_deserialize() {
+        let original = [1u64, 5u64, 10u64];
+        let restored: Vec<u64> = serialize_and_load(&original[..]).iter().collect();
+        assert_eq!(&restored, &original[..]);
+    }
+
+    #[test]
+    fn test_fastfield_bool_size_bitwidth_1() {
+        let mut buffer = Vec::new();
+        let col = VecColumn::from(&[false, true][..]);
+        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
+        // TODO put the header as a footer so that it serves as a padding.
+        // 5 bytes of header, 1 byte of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 5 + 1 + 7);
+    }
+
+    #[test]
+    fn test_fastfield_bool_bit_size_bitwidth_0() {
+        let mut buffer = Vec::new();
+        let col = VecColumn::from(&[true][..]);
+        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
+        // 5 bytes of header, 0 bytes of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 5 + 7);
+    }
+
+    #[test]
+    fn test_fastfield_gcd() {
+        let mut buffer = Vec::new();
+        let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
+        let col = VecColumn::from(&vals[..]);
+        serialize_column_values(&col, &[FastFieldCodecType::Bitpacked], &mut buffer).unwrap();
+        // Values are stored over 3 bits.
+        assert_eq!(buffer.len(), 7 + (3 * 80 / 8) + 7);
+    }
+}

columnar/src/column_values/tests.rs

@@ -0,0 +1,309 @@
+use proptest::prelude::*;
+use proptest::strategy::Strategy;
+use proptest::{prop_oneof, proptest};
+
+use super::bitpacked::BitpackedCodec;
+use super::blockwise_linear::BlockwiseLinearCodec;
+use super::linear::LinearCodec;
+use super::serialize::Header;
+
+pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
+    data: &[u64],
+    name: &str,
+) -> Option<(f32, f32)> {
+    let col = &VecColumn::from(data);
+    let header = Header::compute_header(col, &[Codec::CODEC_TYPE])?;
+    let normalized_col = header.normalize_column(col);
+    let estimation = Codec::estimate(&normalized_col)?;
+
+    let mut out = Vec::new();
+    let col = VecColumn::from(data);
+    serialize_column_values(&col, &[Codec::CODEC_TYPE], &mut out).unwrap();
+
+    let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
+
+    let reader = super::open_u64_mapped::<u64>(OwnedBytes::new(out)).unwrap();
+    assert_eq!(reader.num_vals(), data.len() as u32);
+    for (doc, orig_val) in data.iter().copied().enumerate() {
+        let val = reader.get_val(doc as u32);
+        assert_eq!(
+            val, orig_val,
+            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{data:?}`",
+        );
+    }
+
+    if !data.is_empty() {
+        let test_rand_idx = rand::thread_rng().gen_range(0..=data.len() - 1);
+        let expected_positions: Vec<u32> = data
+            .iter()
+            .enumerate()
+            .filter(|(_, el)| **el == data[test_rand_idx])
+            .map(|(pos, _)| pos as u32)
+            .collect();
+        let mut positions = Vec::new();
+        reader.get_docids_for_value_range(
+            data[test_rand_idx]..=data[test_rand_idx],
+            0..data.len() as u32,
+            &mut positions,
+        );
+        assert_eq!(expected_positions, positions);
+    }
+    Some((estimation, actual_compression))
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(100))]
+
+    #[test]
+    fn test_proptest_small_bitpacked(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
+    }
+
+    #[test]
+    fn test_proptest_small_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
+    }
+
+    #[test]
+    fn test_proptest_small_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
+        create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
+    }
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(10))]
+
+    #[test]
+    fn test_proptest_large_bitpacked(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
+    }
+
+    #[test]
+    fn test_proptest_large_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
+    }
+
+    #[test]
+    fn test_proptest_large_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
+        create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
+    }
+}
+
+fn num_strategy() -> impl Strategy<Value = u64> {
+    prop_oneof![
+        1 => prop::num::u64::ANY.prop_map(|num| u64::MAX - (num % 10) ),
+        1 => prop::num::u64::ANY.prop_map(|num| num % 10 ),
+        20 => prop::num::u64::ANY,
+    ]
+}
+
+pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
+    let mut data_and_names = vec![];
+
+    let data = (10..=10_000_u64).collect::<Vec<_>>();
+    data_and_names.push((data, "simple monotonically increasing"));
+
+    data_and_names.push((
+        vec![5, 6, 7, 8, 9, 10, 99, 100],
+        "offset in linear interpol",
+    ));
+    data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
+    data_and_names.push((vec![10], "single value"));
+
+    data_and_names.push((
+        vec![1572656989877777, 1170935903116329, 720575940379279, 0],
+        "overflow error",
+    ));
+
+    data_and_names
+}
+
+fn test_codec<C: FastFieldCodec>() {
+    let codec_name = format!("{:?}", C::CODEC_TYPE);
+    for (data, dataset_name) in get_codec_test_datasets() {
+        let estimate_actual_opt: Option<(f32, f32)> =
+            tests::create_and_validate::<C>(&data, dataset_name);
+        let result = if let Some((estimate, actual)) = estimate_actual_opt {
+            format!("Estimate `{estimate}` Actual `{actual}`")
+        } else {
+            "Disabled".to_string()
+        };
+        println!("Codec {codec_name}, DataSet {dataset_name}, {result}");
+    }
+}
+#[test]
+fn test_codec_bitpacking() {
+    test_codec::<BitpackedCodec>();
+}
+#[test]
+fn test_codec_interpolation() {
+    test_codec::<LinearCodec>();
+}
+#[test]
+fn test_codec_multi_interpolation() {
+    test_codec::<BlockwiseLinearCodec>();
+}
+
+use super::*;
+
+#[test]
+fn estimation_good_interpolation_case() {
+    let data = (10..=20000_u64).collect::<Vec<_>>();
+    let data: VecColumn = data.as_slice().into();
+
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.01);
+
+    let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
+    assert_le!(multi_linear_interpol_estimation, 0.2);
+    assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
+
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_lt!(linear_interpol_estimation, bitpacked_estimation);
+}
+#[test]
+fn estimation_test_bad_interpolation_case() {
+    let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
+
+    let data: VecColumn = data.into();
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.34);
+
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
+}
+
+#[test]
+fn estimation_prefer_bitpacked() {
+    let data = VecColumn::from(&[10, 10, 10, 10]);
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
+}
+
+#[test]
+fn estimation_test_bad_interpolation_case_monotonically_increasing() {
+    let mut data: Vec<u64> = (201..=20000_u64).collect();
+    data.push(1_000_000);
+    let data: VecColumn = data.as_slice().into();
+
+    // in this case the linear interpolation can't in fact not be worse than bitpacking,
+    // but the estimator adds some threshold, which leads to estimated worse behavior
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.35);
+
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_le!(bitpacked_estimation, 0.32);
+    assert_le!(bitpacked_estimation, linear_interpol_estimation);
+}
+
+#[test]
+fn test_fast_field_codec_type_to_code() {
+    let mut count_codec = 0;
+    for code in 0..=255 {
+        if let Some(codec_type) = FastFieldCodecType::from_code(code) {
+            assert_eq!(codec_type.to_code(), code);
+            count_codec += 1;
+        }
+    }
+    assert_eq!(count_codec, 3);
+}
+
+fn test_fastfield_gcd_i64_with_codec(
+    codec_type: FastFieldCodecType,
+    num_vals: usize,
+) -> io::Result<()> {
+    let mut vals: Vec<i64> = (-4..=(num_vals as i64) - 5).map(|val| val * 1000).collect();
+    let mut buffer: Vec<u8> = Vec::new();
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::open_u64_mapped::<i64>(buffer.clone())?;
+    assert_eq!(column.get_val(0), -4000i64);
+    assert_eq!(column.get_val(1), -3000i64);
+    assert_eq!(column.get_val(2), -2000i64);
+    assert_eq!(column.max_value(), (num_vals as i64 - 5) * 1000);
+    assert_eq!(column.min_value(), -4000i64);
+
+    // Can't apply gcd
+    let mut buffer_without_gcd = Vec::new();
+    vals.pop();
+    vals.push(1001i64);
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
+        &[codec_type],
+        &mut buffer_without_gcd,
+    )?;
+    let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
+    assert!(buffer_without_gcd.len() > buffer.len());
+
+    Ok(())
+}
+
+#[test]
+fn test_fastfield_gcd_i64() -> io::Result<()> {
+    for &codec_type in &[
+        FastFieldCodecType::Bitpacked,
+        FastFieldCodecType::BlockwiseLinear,
+        FastFieldCodecType::Linear,
+    ] {
+        test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
+    }
+    Ok(())
+}
+
+fn test_fastfield_gcd_u64_with_codec(
+    codec_type: FastFieldCodecType,
+    num_vals: usize,
+) -> io::Result<()> {
+    let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
+    let mut buffer: Vec<u8> = Vec::new();
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::open_u64_mapped::<u64>(buffer.clone())?;
+    assert_eq!(column.get_val(0), 1000u64);
+    assert_eq!(column.get_val(1), 2000u64);
+    assert_eq!(column.get_val(2), 3000u64);
+    assert_eq!(column.max_value(), num_vals as u64 * 1000);
+    assert_eq!(column.min_value(), 1000u64);
+
+    // Can't apply gcd
+    let mut buffer_without_gcd = Vec::new();
+    vals.pop();
+    vals.push(1001u64);
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
+        &[codec_type],
+        &mut buffer_without_gcd,
+    )?;
+    let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
+    assert!(buffer_without_gcd.len() > buffer.len());
+    Ok(())
+}
+
+#[test]
+fn test_fastfield_gcd_u64() -> io::Result<()> {
+    for &codec_type in &[
+        FastFieldCodecType::Bitpacked,
+        FastFieldCodecType::BlockwiseLinear,
+        FastFieldCodecType::Linear,
+    ] {
+        test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
+    }
+    Ok(())
+}
+
+#[test]
+pub fn test_fastfield2() {
+    let test_fastfield = crate::column_values::serialize_and_load(&[100u64, 200u64, 300u64]);
+    assert_eq!(test_fastfield.get_val(0), 100);
+    assert_eq!(test_fastfield.get_val(1), 200);
+    assert_eq!(test_fastfield.get_val(2), 300);
+}

columnar/src/columnar/column_type.rs

@@ -0,0 +1,126 @@
+use crate::utils::{place_bits, select_bits};
+use crate::value::NumericalType;
+use crate::InvalidData;
+
+/// The column type represents the column type and can fit on 6-bits.
+///
+/// - bits[0..3]: Column category type.
+/// - bits[3..6]: Numerical type if necessary.
+#[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 column_type_category;
+        let numerical_type_code: u8;
+        match self {
+            ColumnType::Bytes => {
+                column_type_category = ColumnTypeCategory::Str;
+                numerical_type_code = 0u8;
+            }
+            ColumnType::Numerical(numerical_type) => {
+                column_type_category = ColumnTypeCategory::Numerical;
+                numerical_type_code = numerical_type.to_code();
+            }
+            ColumnType::Bool => {
+                column_type_category = ColumnTypeCategory::Bool;
+                numerical_type_code = 0u8;
+            }
+        }
+        place_bits::<0, 3>(column_type_category.to_code()) | place_bits::<3, 6>(numerical_type_code)
+    }
+
+    pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
+        if select_bits::<6, 8>(code) != 0u8 {
+            return Err(InvalidData);
+        }
+        let column_type_category_code = select_bits::<0, 3>(code);
+        let numerical_type_code = select_bits::<3, 6>(code);
+        let column_type_category = ColumnTypeCategory::try_from_code(column_type_category_code)?;
+        match column_type_category {
+            ColumnTypeCategory::Bool => {
+                if numerical_type_code != 0u8 {
+                    return Err(InvalidData);
+                }
+                Ok(ColumnType::Bool)
+            }
+            ColumnTypeCategory::Str => {
+                if numerical_type_code != 0u8 {
+                    return Err(InvalidData);
+                }
+                Ok(ColumnType::Bytes)
+            }
+            ColumnTypeCategory::Numerical => {
+                let numerical_type = NumericalType::try_from_code(numerical_type_code)?;
+                Ok(ColumnType::Numerical(numerical_type))
+            }
+        }
+    }
+}
+
+/// Column types are grouped into different categories that
+/// corresponds to the different types of `JsonValue` types.
+///
+/// The columnar writer will apply coercion rules to make sure that
+/// at most one column exist per `ColumnTypeCategory`.
+///
+/// See also [README.md].
+#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq, Debug)]
+#[repr(u8)]
+pub(crate) enum ColumnTypeCategory {
+    Bool = 0u8,
+    Str = 1u8,
+    Numerical = 2u8,
+}
+
+impl ColumnTypeCategory {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        match code {
+            0u8 => Ok(Self::Bool),
+            1u8 => Ok(Self::Str),
+            2u8 => Ok(Self::Numerical),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashSet;
+
+    use super::*;
+    use crate::Cardinality;
+
+    #[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 Ok(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 {
+            if let Ok(cardinality) = Cardinality::try_from_code(code) {
+                assert_eq!(cardinality.to_code(), code);
+                num_cardinality += 1;
+            }
+        }
+        assert_eq!(num_cardinality, 3);
+    }
+}

columnar/src/columnar/format_version.rs

@@ -0,0 +1,73 @@
+use crate::InvalidData;
+
+pub const VERSION_FOOTER_NUM_BYTES: usize = MAGIC_BYTES.len() + std::mem::size_of::<u32>();
+
+/// We end the file by these 4 bytes just to somewhat identify that
+/// this is indeed a columnar file.
+const MAGIC_BYTES: [u8; 4] = [2, 113, 119, 066];
+
+pub fn footer() -> [u8; VERSION_FOOTER_NUM_BYTES] {
+    let mut footer_bytes = [0u8; VERSION_FOOTER_NUM_BYTES];
+    footer_bytes[0..4].copy_from_slice(&Version::V1.to_bytes());
+    footer_bytes[4..8].copy_from_slice(&MAGIC_BYTES[..]);
+    footer_bytes
+}
+
+pub fn parse_footer(footer_bytes: [u8; VERSION_FOOTER_NUM_BYTES]) -> Result<Version, InvalidData> {
+    if footer_bytes[4..8] != MAGIC_BYTES {
+        return Err(InvalidData);
+    }
+    Version::try_from_bytes(footer_bytes[0..4].try_into().unwrap())
+}
+
+#[derive(Debug, Copy, Clone, Eq, PartialEq)]
+#[repr(u32)]
+pub enum Version {
+    V1 = 1u32,
+}
+
+impl Version {
+    fn to_bytes(&self) -> [u8; 4] {
+        (*self as u32).to_le_bytes()
+    }
+
+    fn try_from_bytes(bytes: [u8; 4]) -> Result<Version, InvalidData> {
+        let code = u32::from_le_bytes(bytes);
+        match code {
+            1u32 => Ok(Version::V1),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashSet;
+
+    use super::*;
+
+    #[test]
+    fn test_footer_dserialization() {
+        let parsed_version: Version = parse_footer(footer()).unwrap();
+        assert_eq!(Version::V1, parsed_version);
+    }
+
+    #[test]
+    fn test_version_serialization() {
+        let version_to_tests: Vec<u32> = [0, 1 << 8, 1 << 16, 1 << 24]
+            .iter()
+            .copied()
+            .flat_map(|offset| (0..255).map(move |el| el + offset))
+            .collect();
+        let mut valid_versions: HashSet<u32> = HashSet::default();
+        for &i in &version_to_tests {
+            let version_res = Version::try_from_bytes(i.to_le_bytes());
+            if let Ok(version) = version_res {
+                assert_eq!(version, Version::V1);
+                assert_eq!(version.to_bytes(), i.to_le_bytes());
+                valid_versions.insert(i);
+            }
+        }
+        assert_eq!(valid_versions.len(), 1);
+    }
+}

columnar/src/columnar/mod.rs

@@ -0,0 +1,28 @@
+// Copyright (C) 2022 Quickwit, Inc.
+//
+// Quickwit is offered under the AGPL v3.0 and as commercial software.
+// For commercial licensing, contact us at hello@quickwit.io.
+//
+// AGPL:
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as
+// published by the Free Software Foundation, either version 3 of the
+// License, or (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see <http://www.gnu.org/licenses/>.
+//
+
+mod column_type;
+mod format_version;
+mod reader;
+mod writer;
+
+pub use column_type::ColumnType;
+pub use reader::ColumnarReader;
+pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -0,0 +1,121 @@
+use std::{io, mem};
+
+use common::file_slice::FileSlice;
+use common::BinarySerializable;
+use sstable::{Dictionary, RangeSSTable};
+
+use crate::columnar::{format_version, ColumnType};
+use crate::dynamic_column::DynamicColumnHandle;
+
+fn io_invalid_data(msg: String) -> io::Error {
+    io::Error::new(io::ErrorKind::InvalidData, msg)
+}
+
+/// 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, footer_slice) = file_slice
+            .split_from_end(mem::size_of::<u64>() + format_version::VERSION_FOOTER_NUM_BYTES);
+        let footer_bytes = footer_slice.read_bytes()?;
+        let (mut sstable_len_bytes, version_footer_bytes) =
+            footer_bytes.rsplit(format_version::VERSION_FOOTER_NUM_BYTES);
+        let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
+            version_footer_bytes.as_slice().try_into().unwrap();
+        let _version = format_version::parse_footer(version_footer_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, DynamicColumnHandle)>> {
+        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: ColumnType = ColumnType::try_from_code(column_code)
+                .map_err(|_| 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]).to_string();
+            let file_slice = self
+                .column_data
+                .slice(range.start as usize..range.end as usize);
+            let column_handle = DynamicColumnHandle {
+                file_slice,
+                column_type,
+            };
+            results.push((column_name, column_handle));
+        }
+        Ok(results)
+    }
+
+    /// Get all columns for the given column name.
+    ///
+    /// There can be more than one column associated to a given column name, provided they have
+    /// different types.
+    // TODO fix ugly API
+    pub fn read_columns(&self, column_name: &str) -> io::Result<Vec<DynamicColumnHandle>> {
+        // Each column is a associated to a given `column_key`,
+        // that starts by `column_name\0column_header`.
+        //
+        // Listing the columns associated to the given column name is therefore equivalent to
+        // listing `column_key` with the prefix `column_name\0`.
+        //
+        // This is in turn equivalent to searching for the range
+        // `[column_name,\0`..column_name\1)`.
+
+        // TODO can we get some more generic `prefix(..)` logic in the dictioanry.
+        let mut start_key = column_name.to_string();
+        start_key.push('\0');
+        let mut end_key = column_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();
+            assert!(key_bytes.starts_with(start_key.as_bytes()));
+            let column_code: u8 = key_bytes.last().cloned().unwrap();
+            let column_type = ColumnType::try_from_code(column_code)
+                .map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
+            let range = stream.value().clone();
+            let file_slice = self
+                .column_data
+                .slice(range.start as usize..range.end as usize);
+            let dynamic_column_handle = DynamicColumnHandle {
+                file_slice,
+                column_type,
+            };
+            results.push(dynamic_column_handle);
+        }
+        Ok(results)
+    }
+
+    /// Return the number of columns in the columnar.
+    pub fn num_columns(&self) -> usize {
+        self.column_dictionary.num_terms()
+    }
+}

columnar/src/columnar/writer/column_operation.rs

@@ -0,0 +1,346 @@
+use crate::dictionary::UnorderedId;
+use crate::utils::{place_bits, pop_first_byte, select_bits};
+use crate::value::NumericalValue;
+use crate::{InvalidData, NumericalType, RowId};
+
+/// When we build a columnar dataframe, we first just group
+/// all mutations per column, and appends them in append-only buffer
+/// in the stacker.
+///
+/// These ColumnOperation<T> are therefore serialize/deserialized
+/// in memory.
+///
+/// We represents all of these operations as `ColumnOperation`.
+#[derive(Eq, PartialEq, Debug, Clone, Copy)]
+pub(super) enum ColumnOperation<T> {
+    NewDoc(RowId),
+    Value(T),
+}
+
+#[derive(Copy, Clone, Eq, PartialEq, Debug)]
+struct ColumnOperationMetadata {
+    op_type: ColumnOperationType,
+    len: u8,
+}
+
+impl ColumnOperationMetadata {
+    fn to_code(self) -> u8 {
+        place_bits::<0, 4>(self.len) | place_bits::<4, 8>(self.op_type.to_code())
+    }
+
+    fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        let len = select_bits::<0, 4>(code);
+        let typ_code = select_bits::<4, 8>(code);
+        let column_type = ColumnOperationType::try_from_code(typ_code)?;
+        Ok(ColumnOperationMetadata {
+            op_type: column_type,
+            len,
+        })
+    }
+}
+
+#[derive(Copy, Clone, Eq, PartialEq, Debug)]
+#[repr(u8)]
+enum ColumnOperationType {
+    NewDoc = 0u8,
+    AddValue = 1u8,
+}
+
+impl ColumnOperationType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        match code {
+            0 => Ok(Self::NewDoc),
+            1 => Ok(Self::AddValue),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+impl<V: SymbolValue> ColumnOperation<V> {
+    pub(super) fn serialize(self) -> impl AsRef<[u8]> {
+        let mut minibuf = MiniBuffer::default();
+        let column_op_metadata = match self {
+            ColumnOperation::NewDoc(new_doc) => {
+                let symbol_len = new_doc.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationMetadata {
+                    op_type: ColumnOperationType::NewDoc,
+                    len: symbol_len,
+                }
+            }
+            ColumnOperation::Value(val) => {
+                let symbol_len = val.serialize(&mut minibuf.bytes[1..]);
+                ColumnOperationMetadata {
+                    op_type: ColumnOperationType::AddValue,
+                    len: symbol_len,
+                }
+            }
+        };
+        minibuf.bytes[0] = column_op_metadata.to_code();
+        // +1 for the metadata
+        minibuf.len = 1 + column_op_metadata.len;
+        minibuf
+    }
+
+    /// Deserialize a colummn operation.
+    /// Returns None if the buffer is empty.
+    ///
+    /// Panics if the payload is invalid:
+    /// this deserialize method is meant to target in memory.
+    pub(super) fn deserialize(bytes: &mut &[u8]) -> Option<Self> {
+        let column_op_metadata_byte = pop_first_byte(bytes)?;
+        let column_op_metadata = ColumnOperationMetadata::try_from_code(column_op_metadata_byte)
+            .expect("Invalid op metadata byte");
+        let symbol_bytes: &[u8];
+        (symbol_bytes, *bytes) = bytes.split_at(column_op_metadata.len as usize);
+        match column_op_metadata.op_type {
+            ColumnOperationType::NewDoc => {
+                let new_doc = u32::deserialize(symbol_bytes);
+                Some(ColumnOperation::NewDoc(new_doc))
+            }
+            ColumnOperationType::AddValue => {
+                let value = V::deserialize(symbol_bytes);
+                Some(ColumnOperation::Value(value))
+            }
+        }
+    }
+}
+
+impl<T> From<T> for ColumnOperation<T> {
+    fn from(value: T) -> Self {
+        ColumnOperation::Value(value)
+    }
+}
+
+// Serialization trait very local to the writer.
+// As we write fast fields, we accumulate them in "in memory".
+// In order to limit memory usage, and in order
+// to benefit from the stacker, we do this by serialization our data
+// as "Symbols".
+#[allow(clippy::from_over_into)]
+pub(super) trait SymbolValue: Clone + Copy {
+    // Serializes the symbol into the given buffer.
+    // Returns the number of bytes written into the buffer.
+    /// # Panics
+    /// May not exceed 9bytes
+    fn serialize(self, buffer: &mut [u8]) -> u8;
+    // Panics if invalid
+    fn deserialize(bytes: &[u8]) -> Self;
+}
+
+impl SymbolValue for bool {
+    fn serialize(self, buffer: &mut [u8]) -> u8 {
+        buffer[0] = u8::from(self);
+        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)
+            }
+        }
+    }
+
+    /// F64: Serialize with a fixed size of 9 bytes
+    /// U64: Serialize without leading zeroes
+    /// I64: ZigZag encoded and serialize without leading zeroes
+    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_column_op_metadata_byte_serialization() {
+        for len in 0..=15 {
+            for op_type in [ColumnOperationType::AddValue, ColumnOperationType::NewDoc] {
+                let column_op_metadata = ColumnOperationMetadata { op_type, len };
+                let column_op_metadata_code = column_op_metadata.to_code();
+                let serdeser_metadata =
+                    ColumnOperationMetadata::try_from_code(column_op_metadata_code).unwrap();
+                assert_eq!(column_op_metadata, serdeser_metadata);
+            }
+        }
+    }
+
+    #[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/columnar/writer/column_writers.rs

@@ -0,0 +1,322 @@
+use std::cmp::Ordering;
+
+use stacker::{ExpUnrolledLinkedList, MemoryArena};
+
+use crate::columnar::writer::column_operation::{ColumnOperation, SymbolValue};
+use crate::dictionary::{DictionaryBuilder, UnorderedId};
+use crate::{Cardinality, NumericalType, NumericalValue, RowId};
+
+#[derive(Copy, Clone, Debug, Eq, PartialEq)]
+#[repr(u8)]
+enum DocumentStep {
+    Same = 0,
+    Next = 1,
+    Skipped = 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::Same,
+        Ordering::Equal => DocumentStep::Next,
+        Ordering::Greater => DocumentStep::Skipped,
+    }
+}
+
+#[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(super) 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(super) fn record<S: SymbolValue>(&mut self, doc: RowId, value: S, arena: &mut MemoryArena) {
+        // Difference between `doc` and the last doc.
+        match delta_with_last_doc(self.last_doc_opt, doc) {
+            DocumentStep::Same => {
+                // This is the last encounterred document.
+                self.cardinality = Cardinality::Multivalued;
+            }
+            DocumentStep::Next => {
+                self.last_doc_opt = Some(doc);
+                self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
+            }
+            DocumentStep::Skipped => {
+                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(crate) fn get_cardinality(&self, num_docs: RowId) -> Cardinality {
+        match delta_with_last_doc(self.last_doc_opt, num_docs) {
+            DocumentStep::Same | DocumentStep::Next => self.cardinality,
+            DocumentStep::Skipped => 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,
+}
+
+impl NumericalColumnWriter {
+    pub fn force_numerical_type(&mut self, numerical_type: NumericalType) {
+        assert!(self
+            .compatible_numerical_types
+            .is_type_accepted(numerical_type));
+        self.compatible_numerical_types = CompatibleNumericalTypes::StaticType(numerical_type);
+    }
+}
+
+/// State used to store what types are still acceptable
+/// after having seen a set of numerical values.
+#[derive(Clone, Copy)]
+enum CompatibleNumericalTypes {
+    Dynamic {
+        all_values_within_i64_range: bool,
+        all_values_within_u64_range: bool,
+    },
+    StaticType(NumericalType),
+}
+
+impl Default for CompatibleNumericalTypes {
+    fn default() -> CompatibleNumericalTypes {
+        CompatibleNumericalTypes::Dynamic {
+            all_values_within_i64_range: true,
+            all_values_within_u64_range: true,
+        }
+    }
+}
+
+impl CompatibleNumericalTypes {
+    fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
+        match self {
+            CompatibleNumericalTypes::Dynamic {
+                all_values_within_i64_range,
+                all_values_within_u64_range,
+            } => match numerical_type {
+                NumericalType::I64 => *all_values_within_i64_range,
+                NumericalType::U64 => *all_values_within_u64_range,
+                NumericalType::F64 => true,
+            },
+            CompatibleNumericalTypes::StaticType(static_numerical_type) => {
+                *static_numerical_type == numerical_type
+            }
+        }
+    }
+
+    fn accept_value(&mut self, numerical_value: NumericalValue) {
+        match self {
+            CompatibleNumericalTypes::Dynamic {
+                all_values_within_i64_range,
+                all_values_within_u64_range,
+            } => match numerical_value {
+                NumericalValue::I64(val_i64) => {
+                    let value_within_u64_range = val_i64 >= 0i64;
+                    *all_values_within_u64_range &= value_within_u64_range;
+                }
+                NumericalValue::U64(val_u64) => {
+                    let value_within_i64_range = val_u64 < i64::MAX as u64;
+                    *all_values_within_i64_range &= value_within_i64_range;
+                }
+                NumericalValue::F64(_) => {
+                    *all_values_within_i64_range = false;
+                    *all_values_within_u64_range = false;
+                }
+            },
+            CompatibleNumericalTypes::StaticType(typ) => {
+                assert_eq!(numerical_value.numerical_type(), *typ);
+            }
+        }
+    }
+
+    pub fn to_numerical_type(self) -> NumericalType {
+        for numerical_type in [NumericalType::I64, NumericalType::U64] {
+            if self.is_type_accepted(numerical_type) {
+                return numerical_type;
+            }
+        }
+        NumericalType::F64
+    }
+}
+
+impl NumericalColumnWriter {
+    pub fn column_type_and_cardinality(&self, num_docs: RowId) -> (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: RowId,
+        value: NumericalValue,
+        arena: &mut MemoryArena,
+    ) {
+        self.compatible_numerical_types.accept_value(value);
+        self.column_writer.record(doc, value, arena);
+    }
+
+    pub(super) 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(crate) struct StrColumnWriter {
+    pub(crate) dictionary_id: u32,
+    pub(crate) column_writer: ColumnWriter,
+}
+
+impl StrColumnWriter {
+    pub(crate) fn with_dictionary_id(dictionary_id: u32) -> StrColumnWriter {
+        StrColumnWriter {
+            dictionary_id,
+            column_writer: Default::default(),
+        }
+    }
+
+    pub(crate) fn record_bytes(
+        &mut self,
+        doc: RowId,
+        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(super) 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::Next);
+        assert_eq!(delta_with_last_doc(None, 1u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(None, 2u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(Some(0u32), 0u32), DocumentStep::Same);
+        assert_eq!(delta_with_last_doc(Some(1u32), 1u32), DocumentStep::Same);
+        assert_eq!(delta_with_last_doc(Some(1u32), 2u32), DocumentStep::Next);
+        assert_eq!(delta_with_last_doc(Some(1u32), 3u32), DocumentStep::Skipped);
+        assert_eq!(delta_with_last_doc(Some(1u32), 4u32), DocumentStep::Skipped);
+    }
+
+    #[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);
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_compatible_numerical_types_static_incompatible_type() {
+        let mut compatible_numerical_types =
+            CompatibleNumericalTypes::StaticType(NumericalType::U64);
+        compatible_numerical_types.accept_value(NumericalValue::I64(1i64));
+    }
+
+    #[test]
+    fn test_compatible_numerical_types_static_different_type_forbidden() {
+        let mut compatible_numerical_types =
+            CompatibleNumericalTypes::StaticType(NumericalType::U64);
+        compatible_numerical_types.accept_value(NumericalValue::U64(u64::MAX));
+    }
+
+    #[test]
+    fn test_compatible_numerical_types_static() {
+        for typ in [NumericalType::I64, NumericalType::I64, NumericalType::F64] {
+            let compatible_numerical_types = CompatibleNumericalTypes::StaticType(typ);
+            assert_eq!(compatible_numerical_types.to_numerical_type(), typ);
+        }
+    }
+}

columnar/src/columnar/writer/mod.rs

@@ -0,0 +1,559 @@
+mod column_operation;
+mod column_writers;
+mod serializer;
+mod value_index;
+
+use std::io;
+
+use column_operation::ColumnOperation;
+use common::CountingWriter;
+use serializer::ColumnarSerializer;
+use stacker::{Addr, ArenaHashMap, MemoryArena};
+
+use crate::column_index::SerializableColumnIndex;
+use crate::column_values::{ColumnValues, MonotonicallyMappableToU64, VecColumn};
+use crate::columnar::column_type::{ColumnType, ColumnTypeCategory};
+use crate::columnar::writer::column_writers::{
+    ColumnWriter, NumericalColumnWriter, StrColumnWriter,
+};
+use crate::columnar::writer::value_index::{IndexBuilder, PreallocatedIndexBuilders};
+use crate::dictionary::{DictionaryBuilder, TermIdMapping, UnorderedId};
+use crate::value::{Coerce, NumericalType, NumericalValue};
+use crate::{Cardinality, RowId};
+
+/// This is a set of buffers that are used to temporarily write the values into before passing them
+/// to the fast field codecs.
+#[derive(Default)]
+struct SpareBuffers {
+    value_index_builders: PreallocatedIndexBuilders,
+    i64_values: Vec<i64>,
+    u64_values: Vec<u64>,
+    f64_values: Vec<f64>,
+    bool_values: Vec<bool>,
+}
+
+/// Makes it possible to create a new columnar.
+///
+/// ```rust
+/// use tantivy_columnar::ColumnarWriter;
+///
+/// let mut columnar_writer = ColumnarWriter::default();
+/// columnar_writer.record_str(0u32 /* doc id */, "product_name", "Red backpack");
+/// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10u64);
+/// columnar_writer.record_str(1u32 /* doc id */, "product_name", "Apple");
+/// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10.5f64); //< uh oh we ended up mixing integer and floats.
+/// let mut wrt: Vec<u8> =  Vec::new();
+/// columnar_writer.serialize(2u32, &mut wrt).unwrap();
+/// ```
+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(),
+        }
+    }
+}
+
+#[inline]
+fn mutate_or_create_column<V, TMutator>(
+    arena_hash_map: &mut ArenaHashMap,
+    column_name: &str,
+    updater: TMutator,
+) where
+    V: Copy + 'static,
+    TMutator: FnMut(Option<V>) -> V,
+{
+    assert!(
+        !column_name.as_bytes().contains(&0u8),
+        "key may not contain the 0 byte"
+    );
+    arena_hash_map.mutate_or_create(column_name.as_bytes(), updater);
+}
+
+impl ColumnarWriter {
+    pub fn force_numerical_type(&mut self, column_name: &str, numerical_type: NumericalType) {
+        let (hash_map, _) = (&mut self.numerical_field_hash_map, &mut self.arena);
+        mutate_or_create_column(
+            hash_map,
+            column_name,
+            |column_opt: Option<NumericalColumnWriter>| {
+                let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
+                column.force_numerical_type(numerical_type);
+                column
+            },
+        );
+    }
+
+    pub fn record_numerical<T: Into<NumericalValue> + Copy>(
+        &mut self,
+        doc: RowId,
+        column_name: &str,
+        numerical_value: T,
+    ) {
+        let (hash_map, arena) = (&mut self.numerical_field_hash_map, &mut self.arena);
+        mutate_or_create_column(
+            hash_map,
+            column_name,
+            |column_opt: Option<NumericalColumnWriter>| {
+                let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
+                column.record_numerical_value(doc, numerical_value.into(), arena);
+                column
+            },
+        );
+    }
+
+    pub fn record_bool(&mut self, doc: RowId, column_name: &str, val: bool) {
+        let (hash_map, arena) = (&mut self.bool_field_hash_map, &mut self.arena);
+        mutate_or_create_column(hash_map, column_name, |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: RowId, column_name: &str, value: &str) {
+        let (hash_map, arena, dictionaries) = (
+            &mut self.bytes_field_hash_map,
+            &mut self.arena,
+            &mut self.dictionaries,
+        );
+        mutate_or_create_column(
+            hash_map,
+            column_name,
+            |column_opt: Option<StrColumnWriter>| {
+                let mut column: StrColumnWriter = column_opt.unwrap_or_else(|| {
+                    // Each column has its own dictionary
+                    let dictionary_id = dictionaries.len() as u32;
+                    dictionaries.push(DictionaryBuilder::default());
+                    StrColumnWriter::with_dictionary_id(dictionary_id)
+                });
+                column.record_bytes(doc, value.as_bytes(), dictionaries, arena);
+                column
+            },
+        );
+    }
+
+    pub fn serialize(&mut self, num_docs: RowId, wrt: &mut dyn io::Write) -> io::Result<()> {
+        let mut serializer = ColumnarSerializer::new(wrt);
+        let mut field_columns: Vec<(&[u8], ColumnTypeCategory, Addr)> = self
+            .numerical_field_hash_map
+            .iter()
+            .map(|(term, addr, _)| (term, ColumnTypeCategory::Numerical, addr))
+            .collect();
+        field_columns.extend(
+            self.bytes_field_hash_map
+                .iter()
+                .map(|(term, addr, _)| (term, ColumnTypeCategory::Str, addr)),
+        );
+        field_columns.extend(
+            self.bool_field_hash_map
+                .iter()
+                .map(|(term, addr, _)| (term, ColumnTypeCategory::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 {
+                ColumnTypeCategory::Bool => {
+                    let column_writer: ColumnWriter = self.bool_field_hash_map.read(addr);
+                    let cardinality = column_writer.get_cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, ColumnType::Bool);
+                    serialize_bool_column(
+                        cardinality,
+                        num_docs,
+                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnTypeCategory::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 mut column_serializer =
+                        serializer.serialize_column(column_name, ColumnType::Bytes);
+                    serialize_bytes_column(
+                        cardinality,
+                        num_docs,
+                        dictionary_builder,
+                        str_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnTypeCategory::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 mut column_serializer = serializer
+                        .serialize_column(column_name, ColumnType::Numerical(numerical_type));
+                    serialize_numerical_column(
+                        cardinality,
+                        num_docs,
+                        numerical_type,
+                        numerical_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+            };
+        }
+        serializer.finalize()?;
+        Ok(())
+    }
+}
+
+fn serialize_bytes_column(
+    cardinality: Cardinality,
+    num_docs: RowId,
+    dictionary_builder: &DictionaryBuilder,
+    operation_it: impl Iterator<Item = ColumnOperation<UnorderedId>>,
+    buffers: &mut SpareBuffers,
+    wrt: impl io::Write,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        u64_values,
+        ..
+    } = buffers;
+    let mut counting_writer = CountingWriter::wrap(wrt);
+    let term_id_mapping: TermIdMapping = dictionary_builder.serialize(&mut counting_writer)?;
+    let dictionary_num_bytes: u32 = counting_writer.written_bytes() as u32;
+    let mut wrt = counting_writer.finish();
+    let operation_iterator = operation_it.map(|symbol: ColumnOperation<UnorderedId>| {
+        // We map unordered ids to ordered ids.
+        match symbol {
+            ColumnOperation::Value(unordered_id) => {
+                let ordered_id = term_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,
+        &mut wrt,
+    )?;
+    wrt.write_all(&dictionary_num_bytes.to_le_bytes()[..])?;
+    Ok(())
+}
+
+fn serialize_numerical_column(
+    cardinality: Cardinality,
+    num_docs: RowId,
+    numerical_type: NumericalType,
+    op_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
+    buffers: &mut SpareBuffers,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        u64_values,
+        i64_values,
+        f64_values,
+        ..
+    } = buffers;
+    match numerical_type {
+        NumericalType::I64 => {
+            serialize_column(
+                coerce_numerical_symbol::<i64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                i64_values,
+                wrt,
+            )?;
+        }
+        NumericalType::U64 => {
+            serialize_column(
+                coerce_numerical_symbol::<u64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                u64_values,
+                wrt,
+            )?;
+        }
+        NumericalType::F64 => {
+            serialize_column(
+                coerce_numerical_symbol::<f64>(op_iterator),
+                cardinality,
+                num_docs,
+                value_index_builders,
+                f64_values,
+                wrt,
+            )?;
+        }
+    };
+    Ok(())
+}
+
+fn serialize_bool_column(
+    cardinality: Cardinality,
+    num_docs: RowId,
+    column_operations_it: impl Iterator<Item = ColumnOperation<bool>>,
+    buffers: &mut SpareBuffers,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        bool_values,
+        ..
+    } = buffers;
+    serialize_column(
+        column_operations_it,
+        cardinality,
+        num_docs,
+        value_index_builders,
+        bool_values,
+        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: RowId,
+    value_index_builders: &mut PreallocatedIndexBuilders,
+    values: &mut Vec<T>,
+    mut wrt: impl io::Write,
+) -> io::Result<()>
+where
+    for<'a> VecColumn<'a, T>: ColumnValues<T>,
+{
+    values.clear();
+    let serializable_column_index = match cardinality {
+        Cardinality::Full => {
+            consume_operation_iterator(
+                op_iterator,
+                value_index_builders.borrow_required_index_builder(),
+                values,
+            );
+            SerializableColumnIndex::Full
+        }
+        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);
+            SerializableColumnIndex::Optional(Box::new(optional_index))
+        }
+        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);
+            todo!();
+            // SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
+        }
+    };
+    crate::column::serialize_column_u64(
+        serializable_column_index,
+        &VecColumn::from(&values[..]),
+        &mut wrt,
+    )?;
+    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_row(doc);
+            }
+            ColumnOperation::Value(value) => {
+                index_builder.record_value();
+                values.push(value);
+            }
+        }
+    }
+}
+
+// /// Serializes the column with the codec with the best estimate on the data.
+// fn serialize_numerical<T: MonotonicallyMappableToU64>(
+//     value_index: ValueIndexInfo,
+//     typed_column: impl Column<T>,
+//     output: &mut impl io::Write,
+//     codecs: &[FastFieldCodecType],
+// ) -> io::Result<()> {
+
+//     let counting_writer = CountingWriter::wrap(output);
+//     serialize_value_index(value_index, output)?;
+//     let value_index_len = counting_writer.written_bytes();
+//     let output = counting_writer.finish();
+
+//     serialize_column(value_index, output)?;
+//     let column = monotonic_map_column(
+//         typed_column,
+//         crate::column::monotonic_mapping::StrictlyMonotonicMappingToInternal::<T>::new(),
+//     );
+//     let header = Header::compute_header(&column, codecs).ok_or_else(|| {
+//         io::Error::new(
+//             io::ErrorKind::InvalidInput,
+//             format!(
+//                 "Data cannot be serialized with this list of codec. {:?}",
+//                 codecs
+//             ),
+//         )
+//     })?;
+//     header.serialize(output)?;
+//     let normalized_column = header.normalize_column(column);
+//     assert_eq!(normalized_column.min_value(), 0u64);
+//     serialize_given_codec(normalized_column, header.codec_type, output)?;
+
+//     let column_header = ColumnFooter {
+//         value_index_len: todo!(),
+//         cardinality: todo!(),
+//     };
+
+//     let null_index_footer = NullIndexFooter {
+//         cardinality: value_index.get_cardinality(),
+//         null_index_codec: NullIndexCodec::Full,
+//         null_index_byte_range: 0..0,
+//     };
+//     append_null_index_footer(output, null_index_footer)?;
+//     Ok(())
+// }
+
+#[cfg(test)]
+mod tests {
+    use column_operation::ColumnOperation;
+    use stacker::MemoryArena;
+
+    use super::*;
+    use crate::value::NumericalValue;
+
+    #[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::Full);
+        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/columnar/writer/serializer.rs

@@ -0,0 +1,106 @@
+use std::io;
+use std::io::Write;
+
+use common::CountingWriter;
+use sstable::value::RangeValueWriter;
+use sstable::RangeSSTable;
+
+use crate::columnar::ColumnType;
+
+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(key: &[u8], column_type: ColumnType, buffer: &mut Vec<u8>) {
+    buffer.clear();
+    buffer.extend_from_slice(key);
+    buffer.push(0u8);
+    buffer.push(column_type.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: ColumnType,
+    ) -> impl io::Write + 'a {
+        let start_offset = self.wrt.written_bytes();
+        prepare_key(column_name, column_type, &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()[..])?;
+        self.wrt
+            .write_all(&super::super::format_version::footer())?;
+        self.wrt.flush()?;
+        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::columnar::column_type::ColumnType;
+
+    #[test]
+    fn test_prepare_key_bytes() {
+        let mut buffer: Vec<u8> = b"somegarbage".to_vec();
+        prepare_key(b"root\0child", ColumnType::Bytes, &mut buffer);
+        assert_eq!(buffer.len(), 12);
+        assert_eq!(&buffer[..10], b"root\0child");
+        assert_eq!(buffer[10], 0u8);
+        assert_eq!(buffer[11], ColumnType::Bytes.to_code());
+    }
+}

columnar/src/columnar/writer/value_index.rs

@@ -0,0 +1,205 @@
+use crate::column_index::SerializableOptionalIndex;
+use crate::column_values::{ColumnValues, VecColumn};
+use crate::RowId;
+
+/// The `IndexBuilder` interprets a sequence of
+/// calls of the form:
+/// (record_doc,record_value+)*
+/// and can then serialize the results into an index to associate docids with their value[s].
+///
+/// It has different implementation depending on whether the
+/// cardinality is required, optional, or multivalued.
+pub(crate) trait IndexBuilder {
+    fn record_row(&mut self, doc: RowId);
+    #[inline]
+    fn record_value(&mut self) {}
+}
+
+/// The FullIndexBuilder does nothing.
+#[derive(Default)]
+pub struct FullIndexBuilder;
+
+impl IndexBuilder for FullIndexBuilder {
+    #[inline(always)]
+    fn record_row(&mut self, _doc: RowId) {}
+}
+
+#[derive(Default)]
+pub struct OptionalIndexBuilder {
+    docs: Vec<RowId>,
+}
+
+struct SingleValueArrayIndex<'a> {
+    // RowIds with a value, in a strictly increasing order
+    row_ids: &'a [RowId],
+    num_rows: RowId,
+}
+
+impl<'a> SerializableOptionalIndex<'a> for SingleValueArrayIndex<'a> {
+    fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
+    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(self.row_ids.iter().copied())
+    }
+}
+
+impl OptionalIndexBuilder {
+    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<'a>(&'a mut self, num_rows: RowId) -> impl SerializableOptionalIndex + 'a {
+        debug_assert!(self
+            .docs
+            .last()
+            .copied()
+            .map(|last_doc| last_doc < num_rows)
+            .unwrap_or(true));
+        SingleValueArrayIndex {
+            row_ids: &self.docs[..],
+            num_rows,
+        }
+    }
+
+    fn reset(&mut self) {
+        self.docs.clear();
+    }
+}
+
+impl IndexBuilder for OptionalIndexBuilder {
+    #[inline(always)]
+    fn record_row(&mut self, doc: RowId) {
+        debug_assert!(self
+            .docs
+            .last()
+            .copied()
+            .map(|prev_doc| doc > prev_doc)
+            .unwrap_or(true));
+        self.docs.push(doc);
+    }
+}
+
+#[derive(Default)]
+pub struct MultivaluedIndexBuilder {
+    start_offsets: Vec<RowId>,
+    total_num_vals_seen: u32,
+}
+
+impl MultivaluedIndexBuilder {
+    pub fn finish(&mut self, num_docs: RowId) -> impl ColumnValues<u32> + '_ {
+        self.start_offsets
+            .resize(num_docs as usize, self.total_num_vals_seen);
+        VecColumn {
+            values: &&self.start_offsets[..],
+            min_value: 0,
+            max_value: self.start_offsets.last().copied().unwrap_or(0),
+        }
+    }
+
+    fn reset(&mut self) {
+        self.start_offsets.clear();
+        self.start_offsets.push(0u32);
+        self.total_num_vals_seen = 0;
+    }
+}
+
+impl IndexBuilder for MultivaluedIndexBuilder {
+    fn record_row(&mut self, row_id: RowId) {
+        self.start_offsets
+            .resize(row_id as usize + 1, 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 PreallocatedIndexBuilders {
+    required_index_builder: FullIndexBuilder,
+    optional_index_builder: OptionalIndexBuilder,
+    multivalued_index_builder: MultivaluedIndexBuilder,
+}
+
+impl PreallocatedIndexBuilders {
+    pub fn borrow_required_index_builder(&mut self) -> &mut FullIndexBuilder {
+        &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_row(0u32);
+        opt_value_index_builder.record_value();
+        assert_eq!(
+            &opt_value_index_builder
+                .finish(1u32)
+                .non_null_rows()
+                .collect::<Vec<u32>>(),
+            &[0]
+        );
+        opt_value_index_builder.reset();
+        opt_value_index_builder.record_row(1u32);
+        opt_value_index_builder.record_value();
+        assert_eq!(
+            &opt_value_index_builder
+                .finish(2u32)
+                .non_null_rows()
+                .collect::<Vec<u32>>(),
+            &[1]
+        );
+    }
+
+    #[test]
+    fn test_multivalued_value_index_builder() {
+        let mut multivalued_value_index_builder = MultivaluedIndexBuilder::default();
+        multivalued_value_index_builder.record_row(1u32);
+        multivalued_value_index_builder.record_value();
+        multivalued_value_index_builder.record_value();
+        multivalued_value_index_builder.record_row(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_row(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]
+        );
+    }
+}

columnar/src/dictionary.rs

@@ -0,0 +1,84 @@
+use std::io;
+
+use fnv::FnvHashMap;
+use sstable::SSTable;
+
+pub(crate) struct TermIdMapping {
+    unordered_to_ord: Vec<OrderedId>,
+}
+
+impl TermIdMapping {
+    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<TermIdMapping> {
+        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(TermIdMapping { 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/dynamic_column.rs

@@ -0,0 +1,95 @@
+use std::io;
+use std::net::IpAddr;
+
+use common::file_slice::FileSlice;
+use common::{HasLen, OwnedBytes};
+
+use crate::column::{BytesColumn, Column};
+use crate::columnar::ColumnType;
+use crate::DateTime;
+
+#[derive(Clone)]
+pub enum DynamicColumn {
+    Bool(Column<bool>),
+    I64(Column<i64>),
+    U64(Column<u64>),
+    F64(Column<f64>),
+    IpAddr(Column<IpAddr>),
+    DateTime(Column<DateTime>),
+    Str(BytesColumn),
+}
+
+impl From<Column<i64>> for DynamicColumn {
+    fn from(column_i64: Column<i64>) -> Self {
+        DynamicColumn::I64(column_i64)
+    }
+}
+
+impl From<Column<u64>> for DynamicColumn {
+    fn from(column_u64: Column<u64>) -> Self {
+        DynamicColumn::U64(column_u64)
+    }
+}
+
+impl From<Column<f64>> for DynamicColumn {
+    fn from(column_f64: Column<f64>) -> Self {
+        DynamicColumn::F64(column_f64)
+    }
+}
+
+impl From<Column<bool>> for DynamicColumn {
+    fn from(bool_column: Column<bool>) -> Self {
+        DynamicColumn::Bool(bool_column)
+    }
+}
+
+impl From<BytesColumn> for DynamicColumn {
+    fn from(dictionary_encoded_col: BytesColumn) -> Self {
+        DynamicColumn::Str(dictionary_encoded_col)
+    }
+}
+
+#[derive(Clone)]
+pub struct DynamicColumnHandle {
+    pub(crate) file_slice: FileSlice,
+    pub(crate) column_type: ColumnType,
+}
+
+impl DynamicColumnHandle {
+    pub fn open(&self) -> io::Result<DynamicColumn> {
+        let column_bytes: OwnedBytes = self.file_slice.read_bytes()?;
+        self.open_internal(column_bytes)
+    }
+
+    pub async fn open_async(&self) -> io::Result<DynamicColumn> {
+        let column_bytes: OwnedBytes = self.file_slice.read_bytes_async().await?;
+        self.open_internal(column_bytes)
+    }
+
+    fn open_internal(&self, column_bytes: OwnedBytes) -> io::Result<DynamicColumn> {
+        let dynamic_column: DynamicColumn = match self.column_type {
+            ColumnType::Bytes => crate::column::open_column_bytes(column_bytes)?.into(),
+            ColumnType::Numerical(numerical_type) => match numerical_type {
+                crate::NumericalType::I64 => {
+                    crate::column::open_column_u64::<i64>(column_bytes)?.into()
+                }
+                crate::NumericalType::U64 => {
+                    crate::column::open_column_u64::<u64>(column_bytes)?.into()
+                }
+                crate::NumericalType::F64 => {
+                    crate::column::open_column_u64::<f64>(column_bytes)?.into()
+                }
+            },
+            ColumnType::Bool => crate::column::open_column_u64::<bool>(column_bytes)?.into(),
+        };
+        Ok(dynamic_column)
+    }
+
+    pub fn num_bytes(&self) -> usize {
+        self.file_slice.len()
+    }
+
+    pub fn column_type(&self) -> ColumnType {
+        self.column_type
+    }
+}

columnar/src/lib.rs

@@ -0,0 +1,75 @@
+#![cfg_attr(all(feature = "unstable", test), feature(test))]
+
+#[cfg(test)]
+#[macro_use]
+extern crate more_asserts;
+
+#[cfg(all(test, feature = "unstable"))]
+extern crate test;
+
+use std::io;
+
+mod column;
+mod column_index;
+mod column_values;
+mod columnar;
+mod dictionary;
+mod dynamic_column;
+pub(crate) mod utils;
+mod value;
+
+pub use columnar::{ColumnarReader, ColumnarWriter};
+pub use value::{NumericalType, NumericalValue};
+
+// pub use self::dynamic_column::DynamicColumnHandle;
+
+pub type RowId = u32;
+
+#[derive(Clone, Copy)]
+pub struct DateTime {
+    timestamp_micros: i64,
+}
+
+#[derive(Copy, Clone, Debug)]
+pub struct InvalidData;
+
+impl From<InvalidData> for io::Error {
+    fn from(_: InvalidData) -> Self {
+        io::Error::new(io::ErrorKind::InvalidData, "Invalid data")
+    }
+}
+
+/// Enum describing the number of values that can exist per document
+/// (or per row if you will).
+///
+/// The cardinality must fit on 2 bits.
+#[derive(Clone, Copy, Hash, Default, Debug, PartialEq, Eq, PartialOrd, Ord)]
+#[repr(u8)]
+pub enum Cardinality {
+    /// All documents contain exactly one value.
+    /// `Full` is the default for auto-detecting the Cardinality, since it is the most strict.
+    #[default]
+    Full = 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) -> Result<Cardinality, InvalidData> {
+        match code {
+            0 => Ok(Cardinality::Full),
+            1 => Ok(Cardinality::Optional),
+            2 => Ok(Cardinality::Multivalued),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/tests.rs

@@ -0,0 +1,84 @@
+use crate::columnar::ColumnType;
+use crate::dynamic_column::{DynamicColumn, DynamicColumnHandle};
+use crate::value::NumericalValue;
+use crate::{Cardinality, ColumnarReader, ColumnarWriter};
+
+#[test]
+fn test_dataframe_writer_bytes() {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_str(1u32, "my_string", "hello");
+    dataframe_writer.record_str(3u32, "my_string", "helloeee");
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
+    assert_eq!(cols.len(), 1);
+    assert_eq!(cols[0].num_bytes(), 165);
+}
+
+#[test]
+fn test_dataframe_writer_bool() {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_bool(1u32, "bool.value", false);
+    dataframe_writer.record_bool(3u32, "bool.value", true);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("bool.value").unwrap();
+    assert_eq!(cols.len(), 1);
+    assert_eq!(cols[0].num_bytes(), 29);
+    assert_eq!(cols[0].column_type(), ColumnType::Bool);
+    let dyn_bool_col = cols[0].open().unwrap();
+    let DynamicColumn::Bool(bool_col) = dyn_bool_col else { panic!(); };
+    let vals: Vec<Option<bool>> = (0..5).map(|row_id| bool_col.first(row_id)).collect();
+    assert_eq!(&vals, &[None, Some(false), None, Some(true), None,]);
+}
+
+#[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(6, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = 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
+    assert_eq!(cols[0].num_bytes(), 40);
+    let column = cols[0].open().unwrap();
+    let DynamicColumn::I64(column_i64) = column else { panic!(); };
+    assert_eq!(column_i64.idx.get_cardinality(), Cardinality::Optional);
+    assert_eq!(column_i64.first(0), None);
+    assert_eq!(column_i64.first(1), Some(12i64));
+    assert_eq!(column_i64.first(2), Some(13i64));
+    assert_eq!(column_i64.first(3), None);
+    assert_eq!(column_i64.first(4), Some(15i64));
+    assert_eq!(column_i64.first(5), None);
+    assert_eq!(column_i64.first(6), None); //< we can change the spec for that one.
+}
+
+#[test]
+fn test_dictionary_encoded() {
+    let mut buffer = Vec::new();
+    let mut columnar_writer = ColumnarWriter::default();
+    columnar_writer.record_str(1, "my.column", "my.key");
+    columnar_writer.record_str(3, "my.column", "my.key2");
+    columnar_writer.record_str(3, "my.column2", "different_column!");
+    columnar_writer.serialize(5, &mut buffer).unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_columns(), 2);
+    let col_handles = columnar_reader.read_columns("my.column").unwrap();
+    assert_eq!(col_handles.len(), 1);
+    let DynamicColumn::Str(str_col) = col_handles[0].open().unwrap() else  { panic!(); };
+    assert_eq!(str_col.num_rows(), 5);
+    // let term_ords = (0..)
+}

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 +1,124 @@
+use crate::InvalidData;
+
+#[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) -> Result<NumericalType, InvalidData> {
+        match code {
+            0 => Ok(NumericalType::I64),
+            1 => Ok(NumericalType::U64),
+            2 => Ok(NumericalType::F64),
+            _ => Err(InvalidData),
+        }
+    }
+}
+
+/// We voluntarily avoid using `Into` here to keep this
+/// implementation quirk as private as possible.
+///
+/// # Panics
+/// 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 Ok(numerical_type) = NumericalType::try_from_code(code) {
+                assert_eq!(numerical_type.to_code(), code);
+                num_numerical_type += 1;
+            }
+        }
+        assert_eq!(num_numerical_type, 3);
+    }
+}

common/src/group_by.rs

@@ -0,0 +1,166 @@
+use std::cell::RefCell;
+use std::iter::Peekable;
+use std::rc::Rc;
+
+pub trait GroupByIteratorExtended: Iterator {
+    /// Return an `Iterator` that groups iterator elements. Consecutive elements that map to the
+    /// same key are assigned to the same group.
+    ///
+    /// The returned Iterator item is `(K, impl Iterator)`, where Iterator are the items of the
+    /// group.
+    ///
+    /// ```
+    /// use tantivy_common::GroupByIteratorExtended;
+    ///
+    /// // group data into blocks of larger than zero or not.
+    /// let data: Vec<i32> = vec![1, 3, -2, -2, 1, 0, 1, 2];
+    /// // groups:               |---->|------>|--------->|
+    ///
+    /// let mut data_grouped = Vec::new();
+    /// // Note: group is an iterator
+    /// for (key, group) in data.into_iter().group_by(|val| *val >= 0) {
+    ///     data_grouped.push((key, group.collect()));
+    /// }
+    /// assert_eq!(data_grouped, vec![(true, vec![1, 3]), (false, vec![-2, -2]), (true, vec![1, 0, 1, 2])]);
+    /// ```
+    fn group_by<K, F>(self, key: F) -> GroupByIterator<Self, F, K>
+    where
+        Self: Sized,
+        F: FnMut(&Self::Item) -> K,
+        K: PartialEq + Copy,
+        Self::Item: Copy,
+    {
+        GroupByIterator::new(self, key)
+    }
+}
+impl<I: Iterator> GroupByIteratorExtended for I {}
+
+pub struct GroupByIterator<I, F, K: Copy>
+where
+    I: Iterator,
+    F: FnMut(&I::Item) -> K,
+{
+    // I really would like to avoid the Rc<RefCell>, but the Iterator is shared between
+    // `GroupByIterator` and `GroupIter`. In practice they are used consecutive and
+    // `GroupByIter` is finished before calling next on `GroupByIterator`. I'm not sure there
+    // is a solution with lifetimes for that, because we would need to enforce it in the usage
+    // somehow.
+    //
+    // One potential solution would be to replace the iterator approach with something similar.
+    inner: Rc<RefCell<GroupByShared<I, F, K>>>,
+}
+
+struct GroupByShared<I, F, K: Copy>
+where
+    I: Iterator,
+    F: FnMut(&I::Item) -> K,
+{
+    iter: Peekable<I>,
+    group_by_fn: F,
+}
+
+impl<I, F, K> GroupByIterator<I, F, K>
+where
+    I: Iterator,
+    F: FnMut(&I::Item) -> K,
+    K: Copy,
+{
+    fn new(inner: I, group_by_fn: F) -> Self {
+        let inner = GroupByShared {
+            iter: inner.peekable(),
+            group_by_fn,
+        };
+
+        Self {
+            inner: Rc::new(RefCell::new(inner)),
+        }
+    }
+}
+
+impl<I, F, K> Iterator for GroupByIterator<I, F, K>
+where
+    I: Iterator,
+    I::Item: Copy,
+    F: FnMut(&I::Item) -> K,
+    K: Copy,
+{
+    type Item = (K, GroupIterator<I, F, K>);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let mut inner = self.inner.borrow_mut();
+        let value = *inner.iter.peek()?;
+        let key = (inner.group_by_fn)(&value);
+
+        let inner = self.inner.clone();
+
+        let group_iter = GroupIterator {
+            inner,
+            group_key: key,
+        };
+        Some((key, group_iter))
+    }
+}
+
+pub struct GroupIterator<I, F, K: Copy>
+where
+    I: Iterator,
+    F: FnMut(&I::Item) -> K,
+{
+    inner: Rc<RefCell<GroupByShared<I, F, K>>>,
+    group_key: K,
+}
+
+impl<I, F, K: PartialEq + Copy> Iterator for GroupIterator<I, F, K>
+where
+    I: Iterator,
+    I::Item: Copy,
+    F: FnMut(&I::Item) -> K,
+{
+    type Item = I::Item;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let mut inner = self.inner.borrow_mut();
+        // peek if next value is in group
+        let peek_val = *inner.iter.peek()?;
+        if (inner.group_by_fn)(&peek_val) == self.group_key {
+            inner.iter.next()
+        } else {
+            None
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn group_by_collect<I: Iterator<Item = u32>>(iter: I) -> Vec<(I::Item, Vec<I::Item>)> {
+        iter.group_by(|val| val / 10)
+            .map(|(el, iter)| (el, iter.collect::<Vec<_>>()))
+            .collect::<Vec<_>>()
+    }
+
+    #[test]
+    fn group_by_two_groups() {
+        let vals = vec![1u32, 4, 15];
+        let grouped_vals = group_by_collect(vals.into_iter());
+        assert_eq!(grouped_vals, vec![(0, vec![1, 4]), (1, vec![15])]);
+    }
+
+    #[test]
+    fn group_by_test_empty() {
+        let vals = vec![];
+        let grouped_vals = group_by_collect(vals.into_iter());
+        assert_eq!(grouped_vals, vec![]);
+    }
+
+    #[test]
+    fn group_by_three_groups() {
+        let vals = vec![1u32, 4, 15, 1];
+        let grouped_vals = group_by_collect(vals.into_iter());
+        assert_eq!(
+            grouped_vals,
+            vec![(0, vec![1, 4]), (1, vec![15]), (0, vec![1])]
+        );
+    }
+}

common/src/lib.rs

@@ -6,10 +6,12 @@ pub use byteorder::LittleEndian as Endianness;
 
 mod bitset;
 pub mod file_slice;
+mod group_by;
 mod serialize;
 mod vint;
 mod writer;
 pub use bitset::*;
+pub use group_by::GroupByIteratorExtended;
 pub use ownedbytes::{OwnedBytes, StableDeref};
 pub use serialize::{BinarySerializable, DeserializeFrom, FixedSize};
 pub use vint::{

examples/faceted_search.rs

@@ -1,15 +1,17 @@
-// # Basic Example
+// # Faceted Search
 //
-// This example covers the basic functionalities of
+// This example covers the faceted search functionalities of
 // tantivy.
 //
 // We will :
-// - define our schema
-// = create an index in a directory
-// - index few documents in our index
-// - search for the best document matchings "sea whale"
-// - retrieve the best document original content.
-
+// - define a text field "name" in our schema
+// - define a facet field "classification" in our schema
+// - create an index in memory
+// - index few documents with respective facets in our index
+// - search and count the number of documents that the classifications start the facet "/Felidae"
+// - Search the facet "/Felidae/Pantherinae" and count the number of documents that the
+//   classifications include the facet.
+//
 // ---
 // Importing tantivy...
 use tantivy::collector::FacetCollector;
@@ -21,7 +23,7 @@ fn main() -> tantivy::Result<()> {
     // Let's create a temporary directory for the sake of this example
     let mut schema_builder = Schema::builder();
 
-    let name = schema_builder.add_text_field("felin_name", TEXT | STORED);
+    let name = schema_builder.add_text_field("name", TEXT | STORED);
     // this is our faceted field: its scientific classification
     let classification = schema_builder.add_facet_field("classification", FacetOptions::default());
 

examples/ip_field.rs

@@ -0,0 +1,73 @@
+// # IP Address example
+//
+// This example shows how the ip field can be used
+// with IpV6 and IpV4.
+
+use tantivy::collector::{Count, TopDocs};
+use tantivy::query::QueryParser;
+use tantivy::schema::{Schema, FAST, INDEXED, STORED, STRING};
+use tantivy::Index;
+
+fn main() -> tantivy::Result<()> {
+    // # Defining the schema
+    let mut schema_builder = Schema::builder();
+    let event_type = schema_builder.add_text_field("event_type", STRING | STORED);
+    let ip = schema_builder.add_ip_addr_field("ip", STORED | INDEXED | FAST);
+    let schema = schema_builder.build();
+
+    // # Indexing documents
+    let index = Index::create_in_ram(schema.clone());
+
+    let mut index_writer = index.writer(50_000_000)?;
+    let doc = schema.parse_document(
+        r#"{
+        "ip": "192.168.0.33",
+        "event_type": "login"
+    }"#,
+    )?;
+    index_writer.add_document(doc)?;
+    let doc = schema.parse_document(
+        r#"{
+        "ip": "192.168.0.80",
+        "event_type": "checkout"
+    }"#,
+    )?;
+    index_writer.add_document(doc)?;
+    let doc = schema.parse_document(
+        r#"{
+        "ip": "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
+        "event_type": "checkout"
+    }"#,
+    )?;
+
+    index_writer.add_document(doc)?;
+    index_writer.commit()?;
+
+    let reader = index.reader()?;
+    let searcher = reader.searcher();
+
+    let query_parser = QueryParser::for_index(&index, vec![event_type, ip]);
+    {
+        let query = query_parser.parse_query("ip:[192.168.0.0 TO 192.168.0.100]")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(5))?;
+        assert_eq!(count_docs.len(), 2);
+    }
+    {
+        let query = query_parser.parse_query("ip:[192.168.1.0 TO 192.168.1.100]")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(2))?;
+        assert_eq!(count_docs.len(), 0);
+    }
+    {
+        let query = query_parser.parse_query("ip:192.168.0.80")?;
+        let count_docs = searcher.search(&*query, &Count)?;
+        assert_eq!(count_docs, 1);
+    }
+    {
+        // IpV6 needs to be escaped because it contains `:`
+        let query = query_parser.parse_query("ip:\"2001:0db8:85a3:0000:0000:8a2e:0370:7334\"")?;
+        let count_docs = searcher.search(&*query, &Count)?;
+        assert_eq!(count_docs, 1);
+    }
+
+    Ok(())
+}

fastfield_codecs/Cargo.toml

@@ -14,7 +14,7 @@ repository = "https://github.com/quickwit-oss/tantivy"
 [dependencies]
 common = { version = "0.5", path = "../common/", package = "tantivy-common" }
 tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-prettytable-rs = {version="0.9.0", optional= true}
+prettytable-rs = {version="0.10.0", optional= true}
 rand = {version="0.8.3", optional= true}
 fastdivide = "0.4"
 log = "0.4"

fastfield_codecs/benches/bench.rs

@@ -4,7 +4,7 @@ extern crate test;
 
 #[cfg(test)]
 mod tests {
-    use std::iter;
+    use std::ops::RangeInclusive;
     use std::sync::Arc;
 
     use common::OwnedBytes;
@@ -71,27 +71,24 @@ mod tests {
         });
     }
 
-    fn get_exp_data() -> Vec<u64> {
+    const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
+    const SINGLE_ITEM: u64 = 90;
+    const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
+    const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
+    fn get_data_50percent_item() -> Vec<u128> {
+        let mut rng = StdRng::from_seed([1u8; 32]);
+
         let mut data = vec![];
-        for i in 0..100 {
-            let num = i * i;
-            data.extend(iter::repeat(i as u64).take(num));
+        for _ in 0..300_000 {
+            let val = rng.gen_range(1..=100);
+            data.push(val);
         }
-        data.shuffle(&mut StdRng::from_seed([1u8; 32]));
+        data.push(SINGLE_ITEM);
 
-        // lengt = 328350
+        data.shuffle(&mut rng);
+        let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
         data
     }
-
-    fn get_data_50percent_item() -> (u128, u128, Vec<u128>) {
-        let mut permutation = get_exp_data();
-        let major_item = 20;
-        let minor_item = 10;
-        permutation.extend(iter::repeat(major_item).take(permutation.len()));
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
-        (major_item as u128, minor_item as u128, permutation)
-    }
     fn get_u128_column_random() -> Arc<dyn Column<u128>> {
         let permutation = generate_random();
         let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
@@ -106,15 +103,82 @@ mod tests {
         open_u128::<u128>(out).unwrap()
     }
 
+    // U64 RANGE START
+    #[bench]
+    fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
+        let data = get_data_50percent_item();
+        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
+
+        b.iter(|| {
+            let mut positions = Vec::new();
+            column.get_docids_for_value_range(
+                FIFTY_PERCENT_RANGE,
+                0..data.len() as u32,
+                &mut positions,
+            );
+            positions
+        });
+    }
+
+    #[bench]
+    fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
+        let data = get_data_50percent_item();
+        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
+
+        b.iter(|| {
+            let mut positions = Vec::new();
+            column.get_docids_for_value_range(
+                ONE_PERCENT_ITEM_RANGE,
+                0..data.len() as u32,
+                &mut positions,
+            );
+            positions
+        });
+    }
+
+    #[bench]
+    fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
+        let data = get_data_50percent_item();
+        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
+
+        b.iter(|| {
+            let mut positions = Vec::new();
+            column.get_docids_for_value_range(
+                SINGLE_ITEM_RANGE,
+                0..data.len() as u32,
+                &mut positions,
+            );
+            positions
+        });
+    }
+
+    #[bench]
+    fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
+        let data = get_data_50percent_item();
+        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
+
+        b.iter(|| {
+            let mut positions = Vec::new();
+            column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
+            positions
+        });
+    }
+    // U64 RANGE END
+
+    // U128 RANGE START
     #[bench]
     fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
-        let (major_item, _minor_item, data) = get_data_50percent_item();
+        let data = get_data_50percent_item();
         let column = get_u128_column_from_data(&data);
 
         b.iter(|| {
             let mut positions = Vec::new();
             column.get_docids_for_value_range(
-                major_item..=major_item,
+                *FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
                 0..data.len() as u32,
                 &mut positions,
             );
@@ -124,13 +188,13 @@ mod tests {
 
     #[bench]
     fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
-        let (_major_item, minor_item, data) = get_data_50percent_item();
+        let data = get_data_50percent_item();
         let column = get_u128_column_from_data(&data);
 
         b.iter(|| {
             let mut positions = Vec::new();
             column.get_docids_for_value_range(
-                minor_item..=minor_item,
+                *SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
                 0..data.len() as u32,
                 &mut positions,
             );
@@ -140,7 +204,7 @@ mod tests {
 
     #[bench]
     fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
-        let (_major_item, _minor_item, data) = get_data_50percent_item();
+        let data = get_data_50percent_item();
         let column = get_u128_column_from_data(&data);
 
         b.iter(|| {
@@ -149,6 +213,7 @@ mod tests {
             positions
         });
     }
+    // U128 RANGE END
 
     #[bench]
     fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {

fastfield_codecs/src/column.rs

@@ -1,3 +1,4 @@
+use std::fmt::{self, Debug};
 use std::marker::PhantomData;
 use std::ops::{Range, RangeInclusive};
 
@@ -6,7 +7,7 @@ use tantivy_bitpacker::minmax;
 use crate::monotonic_mapping::StrictlyMonotonicFn;
 
 /// `Column` provides columnar access on a field.
-pub trait Column<T: PartialOrd = u64>: Send + Sync {
+pub trait Column<T: PartialOrd + Debug = u64>: Send + Sync {
     /// Return the value associated with the given idx.
     ///
     /// This accessor should return as fast as possible.
@@ -83,7 +84,7 @@ pub struct VecColumn<'a, T = u64> {
     max_value: T,
 }
 
-impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
+impl<'a, C: Column<T>, T: Copy + PartialOrd + fmt::Debug> Column<T> for &'a C {
     fn get_val(&self, idx: u32) -> T {
         (*self).get_val(idx)
     }
@@ -109,7 +110,7 @@ impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
     }
 }
 
-impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
+impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> Column<T> for VecColumn<'a, T> {
     fn get_val(&self, position: u32) -> T {
         self.values[position as usize]
     }
@@ -177,8 +178,8 @@ pub fn monotonic_map_column<C, T, Input, Output>(
 where
     C: Column<Input>,
     T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Clone,
-    Output: PartialOrd + Send + Sync + Clone,
+    Input: PartialOrd + Send + Sync + Copy + Debug,
+    Output: PartialOrd + Send + Sync + Copy + Debug,
 {
     MonotonicMappingColumn {
         from_column,
@@ -191,8 +192,8 @@ impl<C, T, Input, Output> Column<Output> for MonotonicMappingColumn<C, T, Input>
 where
     C: Column<Input>,
     T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Clone,
-    Output: PartialOrd + Send + Sync + Clone,
+    Input: PartialOrd + Send + Sync + Copy + Debug,
+    Output: PartialOrd + Send + Sync + Copy + Debug,
 {
     #[inline]
     fn get_val(&self, idx: u32) -> Output {
@@ -228,12 +229,15 @@ where
         doc_id_range: Range<u32>,
         positions: &mut Vec<u32>,
     ) {
-        self.from_column.get_docids_for_value_range(
-            self.monotonic_mapping.inverse(range.start().clone())
-                ..=self.monotonic_mapping.inverse(range.end().clone()),
-            doc_id_range,
-            positions,
-        )
+        if range.start() > &self.max_value() || range.end() < &self.min_value() {
+            return;
+        }
+        let range = self.monotonic_mapping.inverse_coerce(range);
+        if range.start() > range.end() {
+            return;
+        }
+        self.from_column
+            .get_docids_for_value_range(range, doc_id_range, positions)
     }
 
     // We voluntarily do not implement get_range as it yields a regression,
@@ -254,7 +258,7 @@ where T: Iterator + Clone + ExactSizeIterator
 impl<T> Column<T::Item> for IterColumn<T>
 where
     T: Iterator + Clone + ExactSizeIterator + Send + Sync,
-    T::Item: PartialOrd,
+    T::Item: PartialOrd + fmt::Debug,
 {
     fn get_val(&self, idx: u32) -> T::Item {
         self.0.clone().nth(idx as usize).unwrap()

fastfield_codecs/src/compact_space/mod.rs

@@ -453,6 +453,8 @@ impl CompactSpaceDecompressor {
 #[cfg(test)]
 mod tests {
 
+    use std::fmt;
+
     use super::*;
     use crate::format_version::read_format_version;
     use crate::null_index_footer::read_null_index_footer;
@@ -706,7 +708,7 @@ mod tests {
         );
     }
 
-    fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd>(
+    fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd + fmt::Debug>(
         column: &C,
         value_range: RangeInclusive<T>,
         doc_id_range: Range<u32>,

fastfield_codecs/src/lib.rs

@@ -14,9 +14,9 @@ extern crate more_asserts;
 #[cfg(all(test, feature = "unstable"))]
 extern crate test;
 
-use std::io;
 use std::io::Write;
 use std::sync::Arc;
+use std::{fmt, io};
 
 use common::{BinarySerializable, OwnedBytes};
 use compact_space::CompactSpaceDecompressor;
@@ -42,7 +42,7 @@ mod null_index_footer;
 
 mod column;
 mod gcd;
-mod serialize;
+pub mod serialize;
 
 use self::bitpacked::BitpackedCodec;
 use self::blockwise_linear::BlockwiseLinearCodec;
@@ -133,7 +133,7 @@ impl U128FastFieldCodecType {
 }
 
 /// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open_u128<Item: MonotonicallyMappableToU128>(
+pub fn open_u128<Item: MonotonicallyMappableToU128 + fmt::Debug>(
     bytes: OwnedBytes,
 ) -> io::Result<Arc<dyn Column<Item>>> {
     let (bytes, _format_version) = read_format_version(bytes)?;
@@ -147,7 +147,9 @@ pub fn open_u128<Item: MonotonicallyMappableToU128>(
 }
 
 /// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Arc<dyn Column<T>>> {
+pub fn open<T: MonotonicallyMappableToU64 + fmt::Debug>(
+    bytes: OwnedBytes,
+) -> io::Result<Arc<dyn Column<T>>> {
     let (bytes, _format_version) = read_format_version(bytes)?;
     let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
     let header = Header::deserialize(&mut bytes)?;
@@ -160,7 +162,7 @@ pub fn open<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Arc<
     }
 }
 
-fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
+fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64 + fmt::Debug>(
     bytes: OwnedBytes,
     header: &Header,
 ) -> io::Result<Arc<dyn Column<Item>>> {
@@ -321,6 +323,9 @@ mod tests {
     pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
         let mut data_and_names = vec![];
 
+        let data = vec![10];
+        data_and_names.push((data, "minimal test"));
+
         let data = (10..=10_000_u64).collect::<Vec<_>>();
         data_and_names.push((data, "simple monotonically increasing"));
 
@@ -328,6 +333,9 @@ mod tests {
             vec![5, 6, 7, 8, 9, 10, 99, 100],
             "offset in linear interpol",
         ));
+
+        data_and_names.push((vec![3, 18446744073709551613, 5], "docid range regression"));
+
         data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
         data_and_names.push((vec![10], "single value"));
 

fastfield_codecs/src/monotonic_mapping.rs

@@ -1,4 +1,6 @@
+use std::fmt;
 use std::marker::PhantomData;
+use std::ops::RangeInclusive;
 
 use fastdivide::DividerU64;
 
@@ -6,7 +8,9 @@ use crate::MonotonicallyMappableToU128;
 
 /// Monotonic maps a value to u64 value space.
 /// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
-pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
+pub trait MonotonicallyMappableToU64:
+    'static + PartialOrd + Copy + Send + Sync + fmt::Debug
+{
     /// Converts a value to u64.
     ///
     /// Internally all fast field values are encoded as u64.
@@ -29,11 +33,29 @@ pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync
 /// mapping from their range to their domain. The `inverse` method is required when opening a codec,
 /// so a value can be converted back to its original domain (e.g. ip address or f64) from its
 /// internal representation.
-pub trait StrictlyMonotonicFn<External, Internal> {
+pub trait StrictlyMonotonicFn<External: Copy, Internal: Copy> {
     /// Strictly monotonically maps the value from External to Internal.
     fn mapping(&self, inp: External) -> Internal;
     /// Inverse of `mapping`. Maps the value from Internal to External.
     fn inverse(&self, out: Internal) -> External;
+
+    /// Maps a user provded value from External to Internal.
+    /// It may be necessary to coerce the value if it is outside the value space.
+    /// In that case it tries to find the next greater value in the value space.
+    ///
+    /// Returns a bool to mark if a value was outside the value space and had to be coerced _up_.
+    /// With that information we can detect if two values in a range both map outside the same value
+    /// space.
+    ///
+    /// coerce_up means the next valid upper value in the value space will be chosen if the value
+    /// has to be coerced.
+    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<Internal> {
+        self.mapping(*inp.start())..=self.mapping(*inp.end())
+    }
+    /// Inverse of `mapping_coerce`.
+    fn inverse_coerce(&self, out: RangeInclusive<Internal>) -> RangeInclusive<External> {
+        self.inverse(*out.start())..=self.inverse(*out.end())
+    }
 }
 
 /// Inverts a strictly monotonic mapping from `StrictlyMonotonicFn<A, B>` to
@@ -54,7 +76,10 @@ impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
 }
 
 impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
-where T: StrictlyMonotonicFn<From, To>
+where
+    T: StrictlyMonotonicFn<From, To>,
+    From: Copy,
+    To: Copy,
 {
     #[inline(always)]
     fn mapping(&self, val: To) -> From {
@@ -65,6 +90,15 @@ where T: StrictlyMonotonicFn<From, To>
     fn inverse(&self, val: From) -> To {
         self.orig_mapping.mapping(val)
     }
+
+    #[inline]
+    fn mapping_coerce(&self, inp: RangeInclusive<To>) -> RangeInclusive<From> {
+        self.orig_mapping.inverse_coerce(inp)
+    }
+    #[inline]
+    fn inverse_coerce(&self, out: RangeInclusive<From>) -> RangeInclusive<To> {
+        self.orig_mapping.mapping_coerce(out)
+    }
 }
 
 /// Applies the strictly monotonic mapping from `T` without any additional changes.
@@ -142,6 +176,31 @@ impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
     fn inverse(&self, out: u64) -> External {
         External::from_u64(self.min_value + out * self.gcd)
     }
+
+    #[inline]
+    #[allow(clippy::reversed_empty_ranges)]
+    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
+        let end = External::to_u64(*inp.end());
+        if end < self.min_value || inp.end() < inp.start() {
+            return 1..=0;
+        }
+        let map_coerce = |mut inp, coerce_up| {
+            let inp_lower_bound = self.inverse(0);
+            if inp < inp_lower_bound {
+                inp = inp_lower_bound;
+            }
+            let val = External::to_u64(inp);
+            let need_coercion = coerce_up && (val - self.min_value) % self.gcd != 0;
+            let mut mapped_val = self.mapping(inp);
+            if need_coercion {
+                mapped_val += 1;
+            }
+            mapped_val
+        };
+        let start = map_coerce(*inp.start(), true);
+        let end = map_coerce(*inp.end(), false);
+        start..=end
+    }
 }
 
 /// Strictly monotonic mapping with a base value.
@@ -158,6 +217,17 @@ impl StrictlyMonotonicMappingToInternalBaseval {
 impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
     for StrictlyMonotonicMappingToInternalBaseval
 {
+    #[inline]
+    #[allow(clippy::reversed_empty_ranges)]
+    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
+        if External::to_u64(*inp.end()) < self.min_value {
+            return 1..=0;
+        }
+        let start = self.mapping(External::to_u64(*inp.start()).max(self.min_value));
+        let end = self.mapping(External::to_u64(*inp.end()));
+        start..=end
+    }
+
     #[inline(always)]
     fn mapping(&self, val: External) -> u64 {
         External::to_u64(val) - self.min_value
@@ -241,7 +311,7 @@ mod tests {
         test_round_trip::<_, _, u64>(&mapping, 100u64);
     }
 
-    fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L>(
+    fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L: Copy>(
         mapping: &T,
         test_val: K,
     ) {

fastfield_codecs/src/monotonic_mapping_u128.rs

@@ -1,8 +1,11 @@
+use std::fmt;
 use std::net::Ipv6Addr;
 
 /// Montonic maps a value to u128 value space
 /// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
-pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Send + Sync {
+pub trait MonotonicallyMappableToU128:
+    'static + PartialOrd + Copy + Send + Sync + fmt::Debug
+{
     /// Converts a value to u128.
     ///
     /// Internally all fast field values are encoded as u64.

fastfield_codecs/src/null_index/dense.rs

@@ -31,15 +31,16 @@ const BLOCK_BITVEC_SIZE: usize = 8;
 const BLOCK_OFFSET_SIZE: usize = 4;
 const SERIALIZED_BLOCK_SIZE: usize = BLOCK_BITVEC_SIZE + BLOCK_OFFSET_SIZE;
 
+/// Interpreting the bitvec as a list of 64 bits from the low weight to the
+/// high weight.
+///
+/// This function returns the number of bits set to 1 within
+/// `[0..pos_in_vec)`.
 #[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()
-    }
+    let mask = (1u64 << pos_in_bitvec) - 1;
+    let masked_bitvec = bitvec & mask;
+    masked_bitvec.count_ones()
 }
 
 #[derive(Clone, Copy)]
@@ -66,9 +67,7 @@ impl DenseCodec {
     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]
@@ -90,8 +89,7 @@ impl DenseCodec {
         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)
+            Some(index_block.offset + ones_in_block)
         } else {
             None
         }
@@ -319,9 +317,10 @@ mod tests {
         set_bit_at(&mut block, 0);
         set_bit_at(&mut block, 2);
 
-        assert_eq!(count_ones(block, 0), 1);
+        assert_eq!(count_ones(block, 0), 0);
         assert_eq!(count_ones(block, 1), 1);
-        assert_eq!(count_ones(block, 2), 2);
+        assert_eq!(count_ones(block, 2), 1);
+        assert_eq!(count_ones(block, 3), 2);
     }
 }
 
@@ -347,11 +346,16 @@ mod bench {
         codec
     }
 
-    fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
+    fn random_range_iterator(
+        start: u32,
+        end: u32,
+        avg_step_size: u32,
+        avg_deviation: 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);
+            current += rng.gen_range(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
             if current >= end {
                 None
             } else {
@@ -360,10 +364,17 @@ mod bench {
         })
     }
 
-    fn walk_over_data(codec: &DenseCodec, max_step_size: u32) -> Option<u32> {
+    fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
+        let ratio = percent as f32 / 100.0;
+        let step_size = (1f32 / ratio) as u32;
+        let deviation = step_size - 1;
+        random_range_iterator(0, num_values, step_size, deviation)
+    }
+
+    fn walk_over_data(codec: &DenseCodec, avg_step_size: u32) -> Option<u32> {
         walk_over_data_from_positions(
             codec,
-            random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
+            random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
         )
     }
 
@@ -379,69 +390,105 @@ mod bench {
     }
 
     #[bench]
-    fn bench_dense_codec_translate_orig_to_codec_90percent_filled_random_stride(
-        bench: &mut Bencher,
-    ) {
-        let codec = gen_bools(0.9f64);
+    fn bench_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
         bench.iter(|| walk_over_data(&codec, 100));
     }
 
     #[bench]
-    fn bench_dense_codec_translate_orig_to_codec_50percent_filled_random_stride(
-        bench: &mut Bencher,
-    ) {
+    fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 1000));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.9f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.1f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(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) {
+    fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(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();
+    fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        let num_non_nulls = codec.num_non_nulls();
         bench.iter(|| {
             codec
-                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
                 .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();
+    fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        let num_non_nulls = codec.num_non_nulls();
         bench.iter(|| {
             codec
-                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(10.0, num_non_nulls))
                 .last()
         });
     }
 
     #[bench]
-    fn bench_dense_codec_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
+    fn bench_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_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.90f64);
+        let num_non_nulls = codec.num_non_nulls();
+        bench.iter(|| {
+            codec
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
+                .last()
+        });
+    }
+
+    #[bench]
+    fn bench_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(|| {

fastfield_codecs/src/null_index/sparse.rs

@@ -1,6 +1,6 @@
 use std::io::{self, Write};
 
-use common::{BitSet, OwnedBytes};
+use common::{BitSet, GroupByIteratorExtended, OwnedBytes};
 
 use super::{serialize_dense_codec, DenseCodec};
 
@@ -78,12 +78,22 @@ struct DenseBlock {
 }
 
 impl DenseBlock {
+    #[inline]
     pub fn exists(&self, idx: u32) -> bool {
         self.codec.exists(idx)
     }
+    #[inline]
     pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
         self.codec.translate_to_codec_idx(idx)
     }
+    #[inline]
+    pub fn translate_codec_idx_to_original_idx_iter<'a>(
+        &'a self,
+        iter: impl Iterator<Item = u32> + 'a,
+    ) -> impl Iterator<Item = u32> + 'a {
+        self.codec.translate_codec_idx_to_original_idx(iter)
+    }
+    #[inline]
     pub fn translate_codec_idx_to_original_idx(&self, idx: u32) -> u32 {
         self.codec
             .translate_codec_idx_to_original_idx(idx..=idx)
@@ -207,6 +217,7 @@ struct ValueAddr {
 }
 
 /// Splits a idx into block index and value in the block
+#[inline]
 fn value_addr(idx: u32) -> ValueAddr {
     /// Static assert number elements per block this method expects
     #[allow(clippy::assertions_on_constants)]
@@ -273,6 +284,7 @@ impl SparseCodec {
         }
     }
 
+    #[inline]
     fn find_block(&self, dense_idx: u32, mut block_pos: u32) -> u32 {
         loop {
             let offset = self.blocks[block_pos as usize].offset();
@@ -284,6 +296,7 @@ impl SparseCodec {
     }
 
     /// Translate positions from the codec index to the original index.
+    /// Correctness: Provided values must be in increasing values
     ///
     /// # Panics
     ///
@@ -292,35 +305,41 @@ impl SparseCodec {
         &'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);
+        iter.group_by(move |codec_idx| {
+            block_pos = self.find_block(*codec_idx, block_pos);
+            block_pos
+        })
+        .flat_map(move |(block_pos, block_iter)| {
             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();
+            let offset = block.offset();
+            let indexes_in_block_iter = block_iter.map(move |codec_idx| codec_idx - 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
+                         block with dense idx. block_pos {}, idx_in_block {:?}",
+                        block_pos,
+                        indexes_in_block_iter.collect::<Vec<_>>()
                     )
                 }
                 SparseCodecBlockVariant::Dense(dense) => {
-                    dense.translate_codec_idx_to_original_idx(idx_in_block) + block_doc_idx_start
+                    Box::new(dense.translate_codec_idx_to_original_idx_iter(indexes_in_block_iter))
+                        as Box<dyn Iterator<Item = u32>>
                 }
                 SparseCodecBlockVariant::Sparse(block) => {
-                    block.value_at_idx(&self.data, idx_in_block as u16) as u32 + block_doc_idx_start
+                    Box::new(indexes_in_block_iter.map(move |idx_in_block| {
+                        block.value_at_idx(&self.data, idx_in_block as u16) as u32
+                    }))
                 }
             }
+            .map(move |idx| idx + block_doc_idx_start)
         })
     }
 }
 
+#[inline]
 fn is_sparse(num_elem_in_block: u32) -> bool {
     num_elem_in_block < DENSE_BLOCK_THRESHOLD
 }
@@ -595,11 +614,16 @@ mod bench {
         codec
     }
 
-    fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
+    fn random_range_iterator(
+        start: u32,
+        end: u32,
+        avg_step_size: u32,
+        avg_deviation: 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);
+            current += rng.gen_range(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
             if current >= end {
                 None
             } else {
@@ -608,10 +632,17 @@ mod bench {
         })
     }
 
-    fn walk_over_data(codec: &SparseCodec, max_step_size: u32) -> Option<u32> {
+    fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
+        let ratio = percent as f32 / 100.0;
+        let step_size = (1f32 / ratio) as u32;
+        let deviation = step_size - 1;
+        random_range_iterator(0, num_values, step_size, deviation)
+    }
+
+    fn walk_over_data(codec: &SparseCodec, avg_step_size: u32) -> Option<u32> {
         walk_over_data_from_positions(
             codec,
-            random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
+            random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
         )
     }
 
@@ -627,83 +658,83 @@ mod bench {
     }
 
     #[bench]
-    fn bench_sparse_codec_translate_orig_to_codec_1percent_filled_random_stride(
-        bench: &mut Bencher,
-    ) {
+    fn bench_translate_orig_to_codec_1percent_filled_10percent_hit(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,
-    ) {
+    fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(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) {
+    fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.05f64);
+        bench.iter(|| walk_over_data(&codec, 1000));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
+        let codec = gen_bools(0.01f64);
+        bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_full_scan_10percent_filled(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) {
+    fn bench_translate_orig_to_codec_full_scan_90percent_filled(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,
-    ) {
+    fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(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,
-    ) {
+    fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
+        let codec = gen_bools(0.5f64);
+        bench.iter(|| walk_over_data(&codec, 100));
+    }
+
+    #[bench]
+    fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(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,
-    ) {
+    fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
         let codec = gen_bools(0.01f64);
-        let num_vals = codec.num_non_nulls();
+        let num_non_nulls = codec.num_non_nulls();
         bench.iter(|| {
             codec
-                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
                 .last()
         });
     }
 
     #[bench]
-    fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_random_stride(
-        bench: &mut Bencher,
-    ) {
+    fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
         let codec = gen_bools(0.01f64);
-        let num_vals = codec.num_non_nulls();
+        let num_non_nulls = codec.num_non_nulls();
         bench.iter(|| {
             codec
-                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(10.0, num_non_nulls))
                 .last()
         });
     }
 
     #[bench]
-    fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
+    fn bench_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(|| {
@@ -714,33 +745,18 @@ mod bench {
     }
 
     #[bench]
-    fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_random_stride_big_step(
-        bench: &mut Bencher,
-    ) {
+    fn bench_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
         let codec = gen_bools(0.90f64);
-        let num_vals = codec.num_non_nulls();
+        let num_non_nulls = codec.num_non_nulls();
         bench.iter(|| {
             codec
-                .translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
+                .translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
                 .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) {
+    fn bench_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(|| {

fastfield_codecs/src/serialize.rs

@@ -17,9 +17,9 @@
 // You should have received a copy of the GNU Affero General Public License
 // along with this program. If not, see <http://www.gnu.org/licenses/>.
 
-use std::io;
 use std::num::NonZeroU64;
 use std::sync::Arc;
+use std::{fmt, io};
 
 use common::{BinarySerializable, OwnedBytes, VInt};
 use log::warn;
@@ -167,7 +167,7 @@ impl BinarySerializable for Header {
 
 /// Return estimated compression for given codec in the value range [0.0..1.0], where 1.0 means no
 /// compression.
-pub fn estimate<T: MonotonicallyMappableToU64>(
+pub fn estimate<T: MonotonicallyMappableToU64 + fmt::Debug>(
     typed_column: impl Column<T>,
     codec_type: FastFieldCodecType,
 ) -> Option<f32> {
@@ -276,7 +276,7 @@ pub fn serialize_u128_new<F: Fn() -> I, I: Iterator<Item = u128>>(
 }
 
 /// Serializes the column with the codec with the best estimate on the data.
-pub fn serialize<T: MonotonicallyMappableToU64>(
+pub fn serialize<T: MonotonicallyMappableToU64 + fmt::Debug>(
     typed_column: impl Column<T>,
     output: &mut impl io::Write,
     codecs: &[FastFieldCodecType],
@@ -285,7 +285,7 @@ pub fn serialize<T: MonotonicallyMappableToU64>(
 }
 
 /// Serializes the column with the codec with the best estimate on the data.
-pub fn serialize_new<T: MonotonicallyMappableToU64>(
+pub fn serialize_new<T: MonotonicallyMappableToU64 + fmt::Debug>(
     value_index: ValueIndexInfo,
     typed_column: impl Column<T>,
     output: &mut impl io::Write,
@@ -366,7 +366,7 @@ fn serialize_given_codec(
 }
 
 /// Helper function to serialize a column (autodetect from all codecs) and then open it
-pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
+pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default + fmt::Debug>(
     column: &[T],
 ) -> Arc<dyn Column<T>> {
     let mut buffer = Vec::new();

src/aggregation/agg_req.rs

@@ -51,7 +51,10 @@ use serde::{Deserialize, Serialize};
 
 pub use super::bucket::RangeAggregation;
 use super::bucket::{HistogramAggregation, TermsAggregation};
-use super::metric::{AverageAggregation, StatsAggregation};
+use super::metric::{
+    AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
+    SumAggregation,
+};
 use super::VecWithNames;
 
 /// The top-level aggregation request structure, which contains [`Aggregation`] and their user
@@ -237,20 +240,38 @@ impl BucketAggregationType {
 /// called multi-value numeric metrics aggregation.
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub enum MetricAggregation {
-    /// Calculates the average.
+    /// Computes the average of the extracted values.
     #[serde(rename = "avg")]
     Average(AverageAggregation),
-    /// Calculates stats sum, average, min, max, standard_deviation on a field.
+    /// Counts the number of extracted values.
+    #[serde(rename = "value_count")]
+    Count(CountAggregation),
+    /// Finds the maximum value.
+    #[serde(rename = "max")]
+    Max(MaxAggregation),
+    /// Finds the minimum value.
+    #[serde(rename = "min")]
+    Min(MinAggregation),
+    /// Computes a collection of statistics (`min`, `max`, `sum`, `count`, and `avg`) over the
+    /// extracted values.
     #[serde(rename = "stats")]
     Stats(StatsAggregation),
+    /// Computes the sum of the extracted values.
+    #[serde(rename = "sum")]
+    Sum(SumAggregation),
 }
 
 impl MetricAggregation {
     fn get_fast_field_names(&self, fast_field_names: &mut HashSet<String>) {
-        match self {
-            MetricAggregation::Average(avg) => fast_field_names.insert(avg.field.to_string()),
-            MetricAggregation::Stats(stats) => fast_field_names.insert(stats.field.to_string()),
+        let fast_field_name = match self {
+            MetricAggregation::Average(avg) => avg.field_name(),
+            MetricAggregation::Count(count) => count.field_name(),
+            MetricAggregation::Max(max) => max.field_name(),
+            MetricAggregation::Min(min) => min.field_name(),
+            MetricAggregation::Stats(stats) => stats.field_name(),
+            MetricAggregation::Sum(sum) => sum.field_name(),
         };
+        fast_field_names.insert(fast_field_name.to_string());
     }
 }
 
@@ -258,6 +279,38 @@ impl MetricAggregation {
 mod tests {
     use super::*;
 
+    #[test]
+    fn test_metric_aggregations_deser() {
+        let agg_req_json = r#"{
+            "price_avg": { "avg": { "field": "price" } },
+            "price_count": { "value_count": { "field": "price" } },
+            "price_max": { "max": { "field": "price" } },
+            "price_min": { "min": { "field": "price" } },
+            "price_stats": { "stats": { "field": "price" } },
+            "price_sum": { "sum": { "field": "price" } }
+        }"#;
+        let agg_req: Aggregations = serde_json::from_str(agg_req_json).unwrap();
+
+        assert!(
+            matches!(agg_req.get("price_avg").unwrap(), Aggregation::Metric(MetricAggregation::Average(avg)) if avg.field == "price")
+        );
+        assert!(
+            matches!(agg_req.get("price_count").unwrap(), Aggregation::Metric(MetricAggregation::Count(count)) if count.field == "price")
+        );
+        assert!(
+            matches!(agg_req.get("price_max").unwrap(), Aggregation::Metric(MetricAggregation::Max(max)) if max.field == "price")
+        );
+        assert!(
+            matches!(agg_req.get("price_min").unwrap(), Aggregation::Metric(MetricAggregation::Min(min)) if min.field == "price")
+        );
+        assert!(
+            matches!(agg_req.get("price_stats").unwrap(), Aggregation::Metric(MetricAggregation::Stats(stats)) if stats.field == "price")
+        );
+        assert!(
+            matches!(agg_req.get("price_sum").unwrap(), Aggregation::Metric(MetricAggregation::Sum(sum)) if sum.field == "price")
+        );
+    }
+
     #[test]
     fn serialize_to_json_test() {
         let agg_req1: Aggregations = vec![(

src/aggregation/agg_req_with_accessor.rs

@@ -8,7 +8,10 @@ use fastfield_codecs::Column;
 
 use super::agg_req::{Aggregation, Aggregations, BucketAggregationType, MetricAggregation};
 use super::bucket::{HistogramAggregation, RangeAggregation, TermsAggregation};
-use super::metric::{AverageAggregation, StatsAggregation};
+use super::metric::{
+    AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
+    SumAggregation,
+};
 use super::segment_agg_result::BucketCount;
 use super::VecWithNames;
 use crate::fastfield::{type_and_cardinality, MultiValuedFastFieldReader};
@@ -134,7 +137,11 @@ impl MetricAggregationWithAccessor {
     ) -> crate::Result<MetricAggregationWithAccessor> {
         match &metric {
             MetricAggregation::Average(AverageAggregation { field: field_name })
-            | MetricAggregation::Stats(StatsAggregation { field: field_name }) => {
+            | MetricAggregation::Count(CountAggregation { field: field_name })
+            | MetricAggregation::Max(MaxAggregation { field: field_name })
+            | MetricAggregation::Min(MinAggregation { field: field_name })
+            | MetricAggregation::Stats(StatsAggregation { field: field_name })
+            | MetricAggregation::Sum(SumAggregation { field: field_name }) => {
                 let (accessor, field_type) =
                     get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?;
 

src/aggregation/agg_result.rs

@@ -30,7 +30,7 @@ impl AggregationResults {
         } else {
             // Validation is be done during request parsing, so we can't reach this state.
             Err(TantivyError::InternalError(format!(
-                "Can't find aggregation {:?} in sub_aggregations",
+                "Can't find aggregation {:?} in sub-aggregations",
                 name
             )))
         }
@@ -70,27 +70,51 @@ impl AggregationResult {
 pub enum MetricResult {
     /// Average metric result.
     Average(SingleMetricResult),
+    /// Count metric result.
+    Count(SingleMetricResult),
+    /// Max metric result.
+    Max(SingleMetricResult),
+    /// Min metric result.
+    Min(SingleMetricResult),
     /// Stats metric result.
     Stats(Stats),
+    /// Sum metric result.
+    Sum(SingleMetricResult),
 }
 
 impl MetricResult {
     fn get_value(&self, agg_property: &str) -> crate::Result<Option<f64>> {
         match self {
             MetricResult::Average(avg) => Ok(avg.value),
+            MetricResult::Count(count) => Ok(count.value),
+            MetricResult::Max(max) => Ok(max.value),
+            MetricResult::Min(min) => Ok(min.value),
             MetricResult::Stats(stats) => stats.get_value(agg_property),
+            MetricResult::Sum(sum) => Ok(sum.value),
         }
     }
 }
 impl From<IntermediateMetricResult> for MetricResult {
     fn from(metric: IntermediateMetricResult) -> Self {
         match metric {
-            IntermediateMetricResult::Average(avg_data) => {
-                MetricResult::Average(avg_data.finalize().into())
+            IntermediateMetricResult::Average(intermediate_avg) => {
+                MetricResult::Average(intermediate_avg.finalize().into())
+            }
+            IntermediateMetricResult::Count(intermediate_count) => {
+                MetricResult::Count(intermediate_count.finalize().into())
+            }
+            IntermediateMetricResult::Max(intermediate_max) => {
+                MetricResult::Max(intermediate_max.finalize().into())
+            }
+            IntermediateMetricResult::Min(intermediate_min) => {
+                MetricResult::Min(intermediate_min.finalize().into())
             }
             IntermediateMetricResult::Stats(intermediate_stats) => {
                 MetricResult::Stats(intermediate_stats.finalize())
             }
+            IntermediateMetricResult::Sum(intermediate_sum) => {
+                MetricResult::Sum(intermediate_sum.finalize().into())
+            }
         }
     }
 }
@@ -100,13 +124,13 @@ impl From<IntermediateMetricResult> for MetricResult {
 #[serde(untagged)]
 pub enum BucketResult {
     /// This is the range entry for a bucket, which contains a key, count, from, to, and optionally
-    /// sub_aggregations.
+    /// sub-aggregations.
     Range {
         /// The range buckets sorted by range.
         buckets: BucketEntries<RangeBucketEntry>,
     },
     /// This is the histogram entry for a bucket, which contains a key, count, and optionally
-    /// sub_aggregations.
+    /// sub-aggregations.
     Histogram {
         /// The buckets.
         ///
@@ -151,7 +175,7 @@ pub enum BucketEntries<T> {
 }
 
 /// This is the default entry for a bucket, which contains a key, count, and optionally
-/// sub_aggregations.
+/// sub-aggregations.
 ///
 /// # JSON Format
 /// ```json
@@ -201,7 +225,7 @@ impl GetDocCount for BucketEntry {
 }
 
 /// This is the range entry for a bucket, which contains a key, count, and optionally
-/// sub_aggregations.
+/// sub-aggregations.
 ///
 /// # JSON Format
 /// ```json
@@ -237,7 +261,7 @@ pub struct RangeBucketEntry {
     /// Number of documents in the bucket.
     pub doc_count: u64,
     #[serde(flatten)]
-    /// sub-aggregations in this bucket.
+    /// Sub-aggregations in this bucket.
     pub sub_aggregation: AggregationResults,
     /// The from range of the bucket. Equals `f64::MIN` when `None`.
     #[serde(skip_serializing_if = "Option::is_none")]

src/aggregation/bucket/histogram/histogram.rs

@@ -1366,7 +1366,6 @@ mod tests {
                     "min": Value::Null,
                     "max": Value::Null,
                     "avg": Value::Null,
-                    "standard_deviation": Value::Null,
                 }
             })
         );

src/aggregation/intermediate_agg_result.rs

@@ -17,7 +17,10 @@ use super::bucket::{
     cut_off_buckets, get_agg_name_and_property, intermediate_histogram_buckets_to_final_buckets,
     GetDocCount, Order, OrderTarget, SegmentHistogramBucketEntry, TermsAggregation,
 };
-use super::metric::{IntermediateAverage, IntermediateStats};
+use super::metric::{
+    IntermediateAverage, IntermediateCount, IntermediateMax, IntermediateMin, IntermediateStats,
+    IntermediateSum,
+};
 use super::segment_agg_result::SegmentMetricResultCollector;
 use super::{format_date, Key, SerializedKey, VecWithNames};
 use crate::aggregation::agg_result::{AggregationResults, BucketEntries, BucketEntry};
@@ -204,21 +207,43 @@ pub enum IntermediateAggregationResult {
 /// Holds the intermediate data for metric results
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub enum IntermediateMetricResult {
-    /// Average containing intermediate average data result
+    /// Intermediate average result.
     Average(IntermediateAverage),
-    /// AverageData variant
+    /// Intermediate count result.
+    Count(IntermediateCount),
+    /// Intermediate max result.
+    Max(IntermediateMax),
+    /// Intermediate min result.
+    Min(IntermediateMin),
+    /// Intermediate stats result.
     Stats(IntermediateStats),
+    /// Intermediate sum result.
+    Sum(IntermediateSum),
 }
 
 impl From<SegmentMetricResultCollector> for IntermediateMetricResult {
     fn from(tree: SegmentMetricResultCollector) -> Self {
         match tree {
-            SegmentMetricResultCollector::Average(collector) => {
-                IntermediateMetricResult::Average(IntermediateAverage::from_collector(collector))
-            }
-            SegmentMetricResultCollector::Stats(collector) => {
-                IntermediateMetricResult::Stats(collector.stats)
-            }
+            SegmentMetricResultCollector::Stats(collector) => match collector.collecting_for {
+                super::metric::SegmentStatsType::Average => IntermediateMetricResult::Average(
+                    IntermediateAverage::from_collector(collector),
+                ),
+                super::metric::SegmentStatsType::Count => {
+                    IntermediateMetricResult::Count(IntermediateCount::from_collector(collector))
+                }
+                super::metric::SegmentStatsType::Max => {
+                    IntermediateMetricResult::Max(IntermediateMax::from_collector(collector))
+                }
+                super::metric::SegmentStatsType::Min => {
+                    IntermediateMetricResult::Min(IntermediateMin::from_collector(collector))
+                }
+                super::metric::SegmentStatsType::Stats => {
+                    IntermediateMetricResult::Stats(collector.stats)
+                }
+                super::metric::SegmentStatsType::Sum => {
+                    IntermediateMetricResult::Sum(IntermediateSum::from_collector(collector))
+                }
+            },
         }
     }
 }
@@ -229,18 +254,36 @@ impl IntermediateMetricResult {
             MetricAggregation::Average(_) => {
                 IntermediateMetricResult::Average(IntermediateAverage::default())
             }
+            MetricAggregation::Count(_) => {
+                IntermediateMetricResult::Count(IntermediateCount::default())
+            }
+            MetricAggregation::Max(_) => IntermediateMetricResult::Max(IntermediateMax::default()),
+            MetricAggregation::Min(_) => IntermediateMetricResult::Min(IntermediateMin::default()),
             MetricAggregation::Stats(_) => {
                 IntermediateMetricResult::Stats(IntermediateStats::default())
             }
+            MetricAggregation::Sum(_) => IntermediateMetricResult::Sum(IntermediateSum::default()),
         }
     }
     fn merge_fruits(&mut self, other: IntermediateMetricResult) {
         match (self, other) {
             (
-                IntermediateMetricResult::Average(avg_data_left),
-                IntermediateMetricResult::Average(avg_data_right),
+                IntermediateMetricResult::Average(avg_left),
+                IntermediateMetricResult::Average(avg_right),
             ) => {
-                avg_data_left.merge_fruits(avg_data_right);
+                avg_left.merge_fruits(avg_right);
+            }
+            (
+                IntermediateMetricResult::Count(count_left),
+                IntermediateMetricResult::Count(count_right),
+            ) => {
+                count_left.merge_fruits(count_right);
+            }
+            (IntermediateMetricResult::Max(max_left), IntermediateMetricResult::Max(max_right)) => {
+                max_left.merge_fruits(max_right);
+            }
+            (IntermediateMetricResult::Min(min_left), IntermediateMetricResult::Min(min_right)) => {
+                min_left.merge_fruits(min_right);
             }
             (
                 IntermediateMetricResult::Stats(stats_left),
@@ -248,6 +291,9 @@ impl IntermediateMetricResult {
             ) => {
                 stats_left.merge_fruits(stats_right);
             }
+            (IntermediateMetricResult::Sum(sum_left), IntermediateMetricResult::Sum(sum_right)) => {
+                sum_left.merge_fruits(sum_right);
+            }
             _ => {
                 panic!("incompatible fruit types in tree");
             }

src/aggregation/metric/average.rs

@@ -1,13 +1,9 @@
 use std::fmt::Debug;
 
-use fastfield_codecs::Column;
 use serde::{Deserialize, Serialize};
 
-use crate::aggregation::f64_from_fastfield_u64;
-use crate::schema::Type;
-use crate::DocId;
+use super::{IntermediateStats, SegmentStatsCollector};
 
-#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 /// A single-value metric aggregation that computes the average of numeric values that are
 /// extracted from the aggregated documents.
 /// Supported field types are u64, i64, and f64.
@@ -21,94 +17,43 @@ use crate::DocId;
 ///     }
 /// }
 /// ```
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct AverageAggregation {
-    /// The field name to compute the stats on.
+    /// The field name to compute the average on.
     pub field: String,
 }
+
 impl AverageAggregation {
-    /// Create new AverageAggregation from a field.
+    /// Creates a new [`AverageAggregation`] instance from a field name.
     pub fn from_field_name(field_name: String) -> Self {
-        AverageAggregation { field: field_name }
+        Self { field: field_name }
     }
-    /// Return the field name.
+    /// Returns the field name the aggregation is computed on.
     pub fn field_name(&self) -> &str {
         &self.field
     }
 }
 
-#[derive(Clone, PartialEq)]
-pub(crate) struct SegmentAverageCollector {
-    pub data: IntermediateAverage,
-    field_type: Type,
-}
-
-impl Debug for SegmentAverageCollector {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("AverageCollector")
-            .field("data", &self.data)
-            .finish()
-    }
-}
-
-impl SegmentAverageCollector {
-    pub fn from_req(field_type: Type) -> Self {
-        Self {
-            field_type,
-            data: Default::default(),
-        }
-    }
-    pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &dyn Column<u64>) {
-        let mut iter = doc.chunks_exact(4);
-        for docs in iter.by_ref() {
-            let val1 = field.get_val(docs[0]);
-            let val2 = field.get_val(docs[1]);
-            let val3 = field.get_val(docs[2]);
-            let val4 = field.get_val(docs[3]);
-            let val1 = f64_from_fastfield_u64(val1, &self.field_type);
-            let val2 = f64_from_fastfield_u64(val2, &self.field_type);
-            let val3 = f64_from_fastfield_u64(val3, &self.field_type);
-            let val4 = f64_from_fastfield_u64(val4, &self.field_type);
-            self.data.collect(val1);
-            self.data.collect(val2);
-            self.data.collect(val3);
-            self.data.collect(val4);
-        }
-        for &doc in iter.remainder() {
-            let val = field.get_val(doc);
-            let val = f64_from_fastfield_u64(val, &self.field_type);
-            self.data.collect(val);
-        }
-    }
-}
-
-/// Contains mergeable version of average data.
+/// Intermediate result of the average aggregation that can be combined with other intermediate
+/// results.
 #[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct IntermediateAverage {
-    pub(crate) sum: f64,
-    pub(crate) doc_count: u64,
+    stats: IntermediateStats,
 }
 
 impl IntermediateAverage {
-    pub(crate) fn from_collector(collector: SegmentAverageCollector) -> Self {
-        collector.data
-    }
-
-    /// Merge average data into this instance.
-    pub fn merge_fruits(&mut self, other: IntermediateAverage) {
-        self.sum += other.sum;
-        self.doc_count += other.doc_count;
-    }
-    /// compute final result
-    pub fn finalize(&self) -> Option<f64> {
-        if self.doc_count == 0 {
-            None
-        } else {
-            Some(self.sum / self.doc_count as f64)
+    /// Creates a new [`IntermediateAverage`] instance from a [`SegmentStatsCollector`].
+    pub(crate) fn from_collector(collector: SegmentStatsCollector) -> Self {
+        Self {
+            stats: collector.stats,
         }
     }
-    #[inline]
-    fn collect(&mut self, val: f64) {
-        self.doc_count += 1;
-        self.sum += val;
+    /// Merges the other intermediate result into self.
+    pub fn merge_fruits(&mut self, other: IntermediateAverage) {
+        self.stats.merge_fruits(other.stats);
+    }
+    /// Computes the final average value.
+    pub fn finalize(&self) -> Option<f64> {
+        self.stats.finalize().avg
     }
 }

src/aggregation/metric/count.rs

@@ -0,0 +1,59 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+use super::{IntermediateStats, SegmentStatsCollector};
+
+/// A single-value metric aggregation that counts the number of values that are
+/// extracted from the aggregated documents.
+/// Supported field types are u64, i64, and f64.
+/// See [super::SingleMetricResult] for return value.
+///
+/// # JSON Format
+/// ```json
+/// {
+///     "value_count": {
+///         "field": "score",
+///     }
+/// }
+/// ```
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct CountAggregation {
+    /// The field name to compute the minimum on.
+    pub field: String,
+}
+
+impl CountAggregation {
+    /// Creates a new [`CountAggregation`] instance from a field name.
+    pub fn from_field_name(field_name: String) -> Self {
+        Self { field: field_name }
+    }
+    /// Returns the field name the aggregation is computed on.
+    pub fn field_name(&self) -> &str {
+        &self.field
+    }
+}
+
+/// Intermediate result of the count aggregation that can be combined with other intermediate
+/// results.
+#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct IntermediateCount {
+    stats: IntermediateStats,
+}
+
+impl IntermediateCount {
+    /// Creates a new [`IntermediateCount`] instance from a [`SegmentStatsCollector`].
+    pub(crate) fn from_collector(collector: SegmentStatsCollector) -> Self {
+        Self {
+            stats: collector.stats,
+        }
+    }
+    /// Merges the other intermediate result into self.
+    pub fn merge_fruits(&mut self, other: IntermediateCount) {
+        self.stats.merge_fruits(other.stats);
+    }
+    /// Computes the final minimum value.
+    pub fn finalize(&self) -> Option<f64> {
+        Some(self.stats.finalize().count as f64)
+    }
+}

src/aggregation/metric/max.rs

@@ -0,0 +1,59 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+use super::{IntermediateStats, SegmentStatsCollector};
+
+/// A single-value metric aggregation that computes the maximum of numeric values that are
+/// extracted from the aggregated documents.
+/// Supported field types are u64, i64, and f64.
+/// See [super::SingleMetricResult] for return value.
+///
+/// # JSON Format
+/// ```json
+/// {
+///     "max": {
+///         "field": "score",
+///     }
+/// }
+/// ```
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct MaxAggregation {
+    /// The field name to compute the maximum on.
+    pub field: String,
+}
+
+impl MaxAggregation {
+    /// Creates a new [`MaxAggregation`] instance from a field name.
+    pub fn from_field_name(field_name: String) -> Self {
+        Self { field: field_name }
+    }
+    /// Returns the field name the aggregation is computed on.
+    pub fn field_name(&self) -> &str {
+        &self.field
+    }
+}
+
+/// Intermediate result of the maximum aggregation that can be combined with other intermediate
+/// results.
+#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct IntermediateMax {
+    stats: IntermediateStats,
+}
+
+impl IntermediateMax {
+    /// Creates a new [`IntermediateMax`] instance from a [`SegmentStatsCollector`].
+    pub(crate) fn from_collector(collector: SegmentStatsCollector) -> Self {
+        Self {
+            stats: collector.stats,
+        }
+    }
+    /// Merges the other intermediate result into self.
+    pub fn merge_fruits(&mut self, other: IntermediateMax) {
+        self.stats.merge_fruits(other.stats);
+    }
+    /// Computes the final maximum value.
+    pub fn finalize(&self) -> Option<f64> {
+        self.stats.finalize().max
+    }
+}

src/aggregation/metric/min.rs

@@ -0,0 +1,59 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+use super::{IntermediateStats, SegmentStatsCollector};
+
+/// A single-value metric aggregation that computes the minimum of numeric values that are
+/// extracted from the aggregated documents.
+/// Supported field types are u64, i64, and f64.
+/// See [super::SingleMetricResult] for return value.
+///
+/// # JSON Format
+/// ```json
+/// {
+///     "min": {
+///         "field": "score",
+///     }
+/// }
+/// ```
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct MinAggregation {
+    /// The field name to compute the minimum on.
+    pub field: String,
+}
+
+impl MinAggregation {
+    /// Creates a new [`MinAggregation`] instance from a field name.
+    pub fn from_field_name(field_name: String) -> Self {
+        Self { field: field_name }
+    }
+    /// Returns the field name the aggregation is computed on.
+    pub fn field_name(&self) -> &str {
+        &self.field
+    }
+}
+
+/// Intermediate result of the minimum aggregation that can be combined with other intermediate
+/// results.
+#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct IntermediateMin {
+    stats: IntermediateStats,
+}
+
+impl IntermediateMin {
+    /// Creates a new [`IntermediateMin`] instance from a [`SegmentStatsCollector`].
+    pub(crate) fn from_collector(collector: SegmentStatsCollector) -> Self {
+        Self {
+            stats: collector.stats,
+        }
+    }
+    /// Merges the other intermediate result into self.
+    pub fn merge_fruits(&mut self, other: IntermediateMin) {
+        self.stats.merge_fruits(other.stats);
+    }
+    /// Computes the final minimum value.
+    pub fn finalize(&self) -> Option<f64> {
+        self.stats.finalize().min
+    }
+}

src/aggregation/metric/mod.rs

@@ -3,10 +3,18 @@
 //! The aggregations in this family compute metrics, see [super::agg_req::MetricAggregation] for
 //! details.
 mod average;
+mod count;
+mod max;
+mod min;
 mod stats;
+mod sum;
 pub use average::*;
+pub use count::*;
+pub use max::*;
+pub use min::*;
 use serde::{Deserialize, Serialize};
 pub use stats::*;
+pub use sum::*;
 
 /// Single-metric aggregations use this common result structure.
 ///
@@ -28,3 +36,61 @@ impl From<Option<f64>> for SingleMetricResult {
         Self { value }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::aggregation::agg_req::Aggregations;
+    use crate::aggregation::agg_result::AggregationResults;
+    use crate::aggregation::AggregationCollector;
+    use crate::query::AllQuery;
+    use crate::schema::{Cardinality, NumericOptions, Schema};
+    use crate::Index;
+
+    #[test]
+    fn test_metric_aggregations() {
+        let mut schema_builder = Schema::builder();
+        let field_options = NumericOptions::default().set_fast(Cardinality::SingleValue);
+        let field = schema_builder.add_f64_field("price", field_options);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer = index.writer_for_tests().unwrap();
+
+        for i in 0..3 {
+            index_writer
+                .add_document(doc!(
+                    field => i as f64,
+                ))
+                .unwrap();
+        }
+        index_writer.commit().unwrap();
+
+        for i in 3..6 {
+            index_writer
+                .add_document(doc!(
+                    field => i as f64,
+                ))
+                .unwrap();
+        }
+        index_writer.commit().unwrap();
+
+        let aggregations_json = r#"{
+            "price_avg": { "avg": { "field": "price" } },
+            "price_count": { "value_count": { "field": "price" } },
+            "price_max": { "max": { "field": "price" } },
+            "price_min": { "min": { "field": "price" } },
+            "price_stats": { "stats": { "field": "price" } },
+            "price_sum": { "sum": { "field": "price" } }
+        }"#;
+        let aggregations: Aggregations = serde_json::from_str(&aggregations_json).unwrap();
+        let collector = AggregationCollector::from_aggs(aggregations, None, index.schema());
+        let reader = index.reader().unwrap();
+        let searcher = reader.searcher();
+        let aggregations_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
+        let aggregations_res_json = serde_json::to_value(&aggregations_res).unwrap();
+
+        assert_eq!(aggregations_res_json["price_avg"]["value"], 2.5);
+        assert_eq!(aggregations_res_json["price_count"]["value"], 6.0);
+        assert_eq!(aggregations_res_json["price_max"]["value"], 5.0);
+        assert_eq!(aggregations_res_json["price_min"]["value"], 0.0);
+        assert_eq!(aggregations_res_json["price_sum"]["value"], 15.0);
+    }
+}

src/aggregation/metric/stats.rs

@@ -5,8 +5,8 @@ use crate::aggregation::f64_from_fastfield_u64;
 use crate::schema::Type;
 use crate::{DocId, TantivyError};
 
-/// A multi-value metric aggregation that computes stats of numeric values that are
-/// extracted from the aggregated documents.
+/// A multi-value metric aggregation that computes a collection of statistics on numeric values that
+/// are extracted from the aggregated documents.
 /// Supported field types are `u64`, `i64`, and `f64`.
 /// See [`Stats`] for returned statistics.
 ///
@@ -26,11 +26,11 @@ pub struct StatsAggregation {
 }
 
 impl StatsAggregation {
-    /// Create new StatsAggregation from a field.
+    /// Creates a new [`StatsAggregation`] instance from a field name.
     pub fn from_field_name(field_name: String) -> Self {
         StatsAggregation { field: field_name }
     }
-    /// Return the field name.
+    /// Returns the field name the aggregation is computed on.
     pub fn field_name(&self) -> &str {
         &self.field
     }
@@ -40,16 +40,14 @@ impl StatsAggregation {
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct Stats {
     /// The number of documents.
-    pub count: usize,
+    pub count: u64,
     /// The sum of the fast field values.
     pub sum: f64,
-    /// The standard deviation of the fast field values. `None` for count == 0.
-    pub standard_deviation: Option<f64>,
     /// The min value of the fast field values.
     pub min: Option<f64>,
     /// The max value of the fast field values.
     pub max: Option<f64>,
-    /// The average of the values. `None` for count == 0.
+    /// The average of the fast field values. `None` if count equals zero.
     pub avg: Option<f64>,
 }
 
@@ -58,33 +56,36 @@ impl Stats {
         match agg_property {
             "count" => Ok(Some(self.count as f64)),
             "sum" => Ok(Some(self.sum)),
-            "standard_deviation" => Ok(self.standard_deviation),
             "min" => Ok(self.min),
             "max" => Ok(self.max),
             "avg" => Ok(self.avg),
             _ => Err(TantivyError::InvalidArgument(format!(
-                "unknown property {} on stats metric aggregation",
+                "Unknown property {} on stats metric aggregation",
                 agg_property
             ))),
         }
     }
 }
 
-/// `IntermediateStats` contains the mergeable version for stats.
+/// Intermediate result of the stats aggregation that can be combined with other intermediate
+/// results.
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct IntermediateStats {
-    count: usize,
+    /// The number of extracted values.
+    count: u64,
+    /// The sum of the extracted values.
     sum: f64,
-    squared_sum: f64,
+    /// The min value.
     min: f64,
+    /// The max value.
     max: f64,
 }
+
 impl Default for IntermediateStats {
     fn default() -> Self {
         Self {
             count: 0,
             sum: 0.0,
-            squared_sum: 0.0,
             min: f64::MAX,
             max: f64::MIN,
         }
@@ -92,33 +93,15 @@ impl Default for IntermediateStats {
 }
 
 impl IntermediateStats {
-    pub(crate) fn avg(&self) -> Option<f64> {
-        if self.count == 0 {
-            None
-        } else {
-            Some(self.sum / (self.count as f64))
-        }
-    }
-
-    fn square_mean(&self) -> f64 {
-        self.squared_sum / (self.count as f64)
-    }
-
-    pub(crate) fn standard_deviation(&self) -> Option<f64> {
-        self.avg()
-            .map(|average| (self.square_mean() - average * average).sqrt())
-    }
-
-    /// Merge data from other stats into this instance.
+    /// Merges the other stats intermediate result into self.
     pub fn merge_fruits(&mut self, other: IntermediateStats) {
         self.count += other.count;
         self.sum += other.sum;
-        self.squared_sum += other.squared_sum;
         self.min = self.min.min(other.min);
         self.max = self.max.max(other.max);
     }
 
-    /// compute final resultimprove_docs
+    /// Computes the final stats value.
     pub fn finalize(&self) -> Stats {
         let min = if self.count == 0 {
             None
@@ -130,13 +113,17 @@ impl IntermediateStats {
         } else {
             Some(self.max)
         };
+        let avg = if self.count == 0 {
+            None
+        } else {
+            Some(self.sum / (self.count as f64))
+        };
         Stats {
             count: self.count,
             sum: self.sum,
-            standard_deviation: self.standard_deviation(),
             min,
             max,
-            avg: self.avg(),
+            avg,
         }
     }
 
@@ -144,22 +131,33 @@ impl IntermediateStats {
     fn collect(&mut self, value: f64) {
         self.count += 1;
         self.sum += value;
-        self.squared_sum += value * value;
         self.min = self.min.min(value);
         self.max = self.max.max(value);
     }
 }
 
+#[derive(Clone, Debug, PartialEq)]
+pub(crate) enum SegmentStatsType {
+    Average,
+    Count,
+    Max,
+    Min,
+    Stats,
+    Sum,
+}
+
 #[derive(Clone, Debug, PartialEq)]
 pub(crate) struct SegmentStatsCollector {
-    pub(crate) stats: IntermediateStats,
     field_type: Type,
+    pub(crate) collecting_for: SegmentStatsType,
+    pub(crate) stats: IntermediateStats,
 }
 
 impl SegmentStatsCollector {
-    pub fn from_req(field_type: Type) -> Self {
+    pub fn from_req(field_type: Type, collecting_for: SegmentStatsType) -> Self {
         Self {
             field_type,
+            collecting_for,
             stats: IntermediateStats::default(),
         }
     }
@@ -236,7 +234,6 @@ mod tests {
                 "count": 0,
                 "max": Value::Null,
                 "min": Value::Null,
-                "standard_deviation": Value::Null,
                 "sum": 0.0
             })
         );
@@ -313,7 +310,6 @@ mod tests {
                 "count": 7,
                 "max": 44.0,
                 "min": 1.0,
-                "standard_deviation": 13.65313748796613,
                 "sum": 85.0
             })
         );
@@ -325,7 +321,6 @@ mod tests {
                 "count": 7,
                 "max": 44.0,
                 "min": 1.0,
-                "standard_deviation": 13.65313748796613,
                 "sum": 85.0
             })
         );
@@ -337,7 +332,6 @@ mod tests {
                 "count": 7,
                 "max": 44.5,
                 "min": 1.0,
-                "standard_deviation": 13.819905785437443,
                 "sum": 85.5
             })
         );
@@ -349,7 +343,6 @@ mod tests {
                 "count": 3,
                 "max": 14.0,
                 "min": 7.0,
-                "standard_deviation": 2.867441755680877,
                 "sum": 32.0
             })
         );
@@ -361,7 +354,6 @@ mod tests {
                 "count": 0,
                 "max": serde_json::Value::Null,
                 "min": serde_json::Value::Null,
-                "standard_deviation": serde_json::Value::Null,
                 "sum": 0.0,
             })
         );

src/aggregation/metric/sum.rs

@@ -0,0 +1,59 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+use super::{IntermediateStats, SegmentStatsCollector};
+
+/// A single-value metric aggregation that sums up numeric values that are
+/// extracted from the aggregated documents.
+/// Supported field types are u64, i64, and f64.
+/// See [super::SingleMetricResult] for return value.
+///
+/// # JSON Format
+/// ```json
+/// {
+///     "sum": {
+///         "field": "score",
+///     }
+/// }
+/// ```
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct SumAggregation {
+    /// The field name to compute the minimum on.
+    pub field: String,
+}
+
+impl SumAggregation {
+    /// Creates a new [`SumAggregation`] instance from a field name.
+    pub fn from_field_name(field_name: String) -> Self {
+        Self { field: field_name }
+    }
+    /// Returns the field name the aggregation is computed on.
+    pub fn field_name(&self) -> &str {
+        &self.field
+    }
+}
+
+/// Intermediate result of the minimum aggregation that can be combined with other intermediate
+/// results.
+#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct IntermediateSum {
+    stats: IntermediateStats,
+}
+
+impl IntermediateSum {
+    /// Creates a new [`IntermediateSum`] instance from a [`SegmentStatsCollector`].
+    pub(crate) fn from_collector(collector: SegmentStatsCollector) -> Self {
+        Self {
+            stats: collector.stats,
+        }
+    }
+    /// Merges the other intermediate result into self.
+    pub fn merge_fruits(&mut self, other: IntermediateSum) {
+        self.stats.merge_fruits(other.stats);
+    }
+    /// Computes the final minimum value.
+    pub fn finalize(&self) -> Option<f64> {
+        Some(self.stats.finalize().sum)
+    }
+}

src/aggregation/mod.rs

@@ -216,8 +216,8 @@ impl<T: Clone> VecWithNames<T> {
     fn from_entries(mut entries: Vec<(String, T)>) -> Self {
         // Sort to ensure order of elements match across multiple instances
         entries.sort_by(|left, right| left.0.cmp(&right.0));
-        let mut data = vec![];
-        let mut data_names = vec![];
+        let mut data = Vec::with_capacity(entries.len());
+        let mut data_names = Vec::with_capacity(entries.len());
         for entry in entries {
             data_names.push(entry.0);
             data.push(entry.1);
@@ -1208,7 +1208,7 @@ mod tests {
                         text_field_many_terms => many_terms_data.choose(&mut rng).unwrap().to_string(),
                         text_field_few_terms => few_terms_data.choose(&mut rng).unwrap().to_string(),
                         score_field => val as u64,
-                        score_field_f64 => val as f64,
+                        score_field_f64 => val,
                         score_field_i64 => val as i64,
                     ))?;
                 }
@@ -1250,10 +1250,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&term_query, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&term_query, &collector).unwrap()
             });
         }
 
@@ -1281,10 +1278,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&term_query, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&term_query, &collector).unwrap()
             });
         }
 
@@ -1312,10 +1306,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&term_query, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&term_query, &collector).unwrap()
             });
         }
 
@@ -1351,10 +1342,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&term_query, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&term_query, &collector).unwrap()
             });
         }
 
@@ -1380,10 +1368,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1409,10 +1394,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1446,10 +1428,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1481,10 +1460,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1520,10 +1496,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1550,10 +1523,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&AllQuery, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&AllQuery, &collector).unwrap()
             });
         }
 
@@ -1597,7 +1567,7 @@ mod tests {
                                 ],
                                 ..Default::default()
                             }),
-                            sub_aggregation: sub_agg_req_1.clone(),
+                            sub_aggregation: sub_agg_req_1,
                         }),
                     ),
                 ]
@@ -1607,10 +1577,7 @@ mod tests {
                 let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
 
                 let searcher = reader.searcher();
-                let agg_res: AggregationResults =
-                    searcher.search(&term_query, &collector).unwrap().into();
-
-                agg_res
+                searcher.search(&term_query, &collector).unwrap()
             });
         }
     }

src/aggregation/segment_agg_result.rs

@@ -15,7 +15,8 @@ use super::bucket::{SegmentHistogramCollector, SegmentRangeCollector, SegmentTer
 use super::collector::MAX_BUCKET_COUNT;
 use super::intermediate_agg_result::{IntermediateAggregationResults, IntermediateBucketResult};
 use super::metric::{
-    AverageAggregation, SegmentAverageCollector, SegmentStatsCollector, StatsAggregation,
+    AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, SegmentStatsCollector,
+    SegmentStatsType, StatsAggregation, SumAggregation,
 };
 use super::VecWithNames;
 use crate::aggregation::agg_req::BucketAggregationType;
@@ -163,30 +164,46 @@ impl SegmentAggregationResultsCollector {
 
 #[derive(Clone, Debug, PartialEq)]
 pub(crate) enum SegmentMetricResultCollector {
-    Average(SegmentAverageCollector),
     Stats(SegmentStatsCollector),
 }
 
 impl SegmentMetricResultCollector {
     pub fn from_req_and_validate(req: &MetricAggregationWithAccessor) -> crate::Result<Self> {
         match &req.metric {
-            MetricAggregation::Average(AverageAggregation { field: _ }) => {
-                Ok(SegmentMetricResultCollector::Average(
-                    SegmentAverageCollector::from_req(req.field_type),
+            MetricAggregation::Average(AverageAggregation { .. }) => {
+                Ok(SegmentMetricResultCollector::Stats(
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Average),
                 ))
             }
-            MetricAggregation::Stats(StatsAggregation { field: _ }) => {
+            MetricAggregation::Count(CountAggregation { .. }) => {
                 Ok(SegmentMetricResultCollector::Stats(
-                    SegmentStatsCollector::from_req(req.field_type),
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Count),
+                ))
+            }
+            MetricAggregation::Max(MaxAggregation { .. }) => {
+                Ok(SegmentMetricResultCollector::Stats(
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Max),
+                ))
+            }
+            MetricAggregation::Min(MinAggregation { .. }) => {
+                Ok(SegmentMetricResultCollector::Stats(
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Min),
+                ))
+            }
+            MetricAggregation::Stats(StatsAggregation { .. }) => {
+                Ok(SegmentMetricResultCollector::Stats(
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Stats),
+                ))
+            }
+            MetricAggregation::Sum(SumAggregation { .. }) => {
+                Ok(SegmentMetricResultCollector::Stats(
+                    SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Sum),
                 ))
             }
         }
     }
     pub(crate) fn collect_block(&mut self, doc: &[DocId], metric: &MetricAggregationWithAccessor) {
         match self {
-            SegmentMetricResultCollector::Average(avg_collector) => {
-                avg_collector.collect_block(doc, &*metric.accessor);
-            }
             SegmentMetricResultCollector::Stats(stats_collector) => {
                 stats_collector.collect_block(doc, &*metric.accessor);
             }

src/core/searcher.rs

@@ -198,11 +198,10 @@ impl Searcher {
         collector: &C,
         executor: &Executor,
     ) -> crate::Result<C::Fruit> {
-        let scoring_enabled = collector.requires_scoring();
-        let enabled_scoring = if scoring_enabled {
-            EnableScoring::Enabled(self)
+        let enabled_scoring = if collector.requires_scoring() {
+            EnableScoring::enabled_from_searcher(self)
         } else {
-            EnableScoring::Disabled(self.schema())
+            EnableScoring::disabled_from_searcher(self)
         };
         let weight = query.weight(enabled_scoring)?;
         let segment_readers = self.segment_readers();

src/directory/error.rs

@@ -32,7 +32,7 @@ impl LockError {
 /// Error that may occur when opening a directory
 #[derive(Debug, Clone, Error)]
 pub enum OpenDirectoryError {
-    /// The underlying directory does not exists.
+    /// The underlying directory does not exist.
     #[error("Directory does not exist: '{0}'.")]
     DoesNotExist(PathBuf),
     /// The path exists but is not a directory.
@@ -151,8 +151,8 @@ impl fmt::Debug for Incompatibility {
 /// Error that may occur when accessing a file read
 #[derive(Debug, Clone, Error)]
 pub enum OpenReadError {
-    /// The file does not exists.
-    #[error("Files does not exists: {0:?}")]
+    /// The file does not exist.
+    #[error("Files does not exist: {0:?}")]
     FileDoesNotExist(PathBuf),
     /// Any kind of io::Error.
     #[error(
@@ -181,8 +181,8 @@ impl OpenReadError {
 /// Error that may occur when trying to delete a file
 #[derive(Debug, Clone, Error)]
 pub enum DeleteError {
-    /// The file does not exists.
-    #[error("File does not exists: '{0}'.")]
+    /// The file does not exist.
+    #[error("File does not exist: '{0}'.")]
     FileDoesNotExist(PathBuf),
     /// Any kind of IO error that happens when
     /// interacting with the underlying IO device.

src/directory/ram_directory.rs

@@ -232,7 +232,7 @@ impl Directory for RamDirectory {
         let path_buf = PathBuf::from(path);
         self.fs.write().unwrap().write(path_buf, data);
         if path == *META_FILEPATH {
-            let _ = self.fs.write().unwrap().watch_router.broadcast();
+            drop(self.fs.write().unwrap().watch_router.broadcast());
         }
         Ok(())
     }

src/directory/watch_event_router.rs

@@ -168,7 +168,7 @@ mod tests {
         watch_event_router.broadcast().wait().unwrap();
         assert_eq!(2, counter.load(Ordering::SeqCst));
         mem::drop(handle_a);
-        let _ = watch_event_router.broadcast();
+        drop(watch_event_router.broadcast());
         watch_event_router.broadcast().wait().unwrap();
         assert_eq!(2, counter.load(Ordering::SeqCst));
     }

src/fastfield/alive_bitset.rs

@@ -175,7 +175,7 @@ mod bench {
 
     fn get_alive() -> Vec<u32> {
         let mut data = (0..1_000_000_u32).collect::<Vec<u32>>();
-        for _ in 0..(1_000_000) * 1 / 8 {
+        for _ in 0..1_000_000 / 8 {
             remove_rand(&mut data);
         }
         data

src/fastfield/bytes/mod.rs

@@ -96,7 +96,7 @@ mod tests {
         let term = Term::from_field_bytes(field, b"lucene".as_ref());
         let term_query = TermQuery::new(term, IndexRecordOption::Basic);
         let term_weight_err =
-            term_query.specialized_weight(EnableScoring::Disabled(searcher.schema()));
+            term_query.specialized_weight(EnableScoring::disabled_from_schema(searcher.schema()));
         assert!(matches!(
             term_weight_err,
             Err(crate::TantivyError::SchemaError(_))

src/fastfield/mod.rs

@@ -12,13 +12,15 @@
 //!
 //!
 //! Fields have to be declared as `FAST` in the schema.
-//! Currently supported fields are: u64, i64, f64, bytes and text.
+//! Currently supported fields are: u64, i64, f64, bytes, ip and text.
 //!
 //! Fast fields are stored in with [different codecs](fastfield_codecs). The best codec is detected
 //! automatically, when serializing.
 //!
 //! Read access performance is comparable to that of an array lookup.
 
+use std::net::Ipv6Addr;
+
 use fastfield_codecs::MonotonicallyMappableToU64;
 
 pub use self::alive_bitset::{intersect_alive_bitsets, write_alive_bitset, AliveBitSet};
@@ -28,7 +30,7 @@ pub use self::facet_reader::FacetReader;
 pub(crate) use self::multivalued::{get_fastfield_codecs_for_multivalue, MultivalueStartIndex};
 pub use self::multivalued::{
     MultiValueIndex, MultiValueU128FastFieldWriter, MultiValuedFastFieldReader,
-    MultiValuedFastFieldWriter, MultiValuedU128FastFieldReader,
+    MultiValuedFastFieldWriter,
 };
 pub(crate) use self::readers::type_and_cardinality;
 pub use self::readers::FastFieldReaders;
@@ -47,6 +49,33 @@ mod readers;
 mod serializer;
 mod writer;
 
+/// Trait for types that provide a zero value.
+///
+/// The resulting value is never used, just as placeholder, e.g. for `vec.resize()`.
+pub trait MakeZero {
+    /// Build a default value. This default value is never used, so the value does not
+    /// really matter.
+    fn make_zero() -> Self;
+}
+
+impl<T: FastValue> MakeZero for T {
+    fn make_zero() -> Self {
+        T::from_u64(0)
+    }
+}
+
+impl MakeZero for u128 {
+    fn make_zero() -> Self {
+        0
+    }
+}
+
+impl MakeZero for Ipv6Addr {
+    fn make_zero() -> Self {
+        Ipv6Addr::from(0u128.to_be_bytes())
+    }
+}
+
 /// Trait for types that are allowed for fast fields:
 /// (u64, i64 and f64, bool, DateTime).
 pub trait FastValue:
@@ -54,12 +83,6 @@ pub trait FastValue:
 {
     /// Returns the `schema::Type` for this FastValue.
     fn to_type() -> Type;
-
-    /// Build a default value. This default value is never used, so the value does not
-    /// really matter.
-    fn make_zero() -> Self {
-        Self::from_u64(0u64)
-    }
 }
 
 impl FastValue for u64 {
@@ -101,12 +124,6 @@ impl FastValue for DateTime {
     fn to_type() -> Type {
         Type::Date
     }
-
-    fn make_zero() -> Self {
-        DateTime {
-            timestamp_micros: 0,
-        }
-    }
 }
 
 fn value_to_u64(value: &Value) -> crate::Result<u64> {
@@ -145,7 +162,7 @@ impl FastFieldType {
 mod tests {
 
     use std::collections::HashMap;
-    use std::ops::Range;
+    use std::ops::{Range, RangeInclusive};
     use std::path::Path;
     use std::sync::Arc;
 
@@ -159,7 +176,9 @@ mod tests {
     use super::*;
     use crate::directory::{CompositeFile, Directory, RamDirectory, WritePtr};
     use crate::merge_policy::NoMergePolicy;
-    use crate::schema::{Cardinality, Document, Field, Schema, SchemaBuilder, FAST, STRING, TEXT};
+    use crate::schema::{
+        Cardinality, Document, Field, Schema, SchemaBuilder, FAST, INDEXED, STRING, TEXT,
+    };
     use crate::time::OffsetDateTime;
     use crate::{DateOptions, DatePrecision, Index, SegmentId, SegmentReader};
 
@@ -520,11 +539,6 @@ mod tests {
         Ok(())
     }
 
-    #[test]
-    fn test_default_date() {
-        assert_eq!(0, DateTime::make_zero().into_timestamp_secs());
-    }
-
     fn get_vals_for_docs(ff: &MultiValuedFastFieldReader<u64>, docs: Range<u32>) -> Vec<u64> {
         let mut all = vec![];
 
@@ -969,4 +983,117 @@ mod tests {
         }
         Ok(len)
     }
+
+    #[test]
+    fn test_gcd_bug_regression_1757() {
+        let mut schema_builder = Schema::builder();
+        let num_field = schema_builder.add_u64_field("url_norm_hash", FAST | INDEXED);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        {
+            let mut writer = index.writer_for_tests().unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 100u64,
+                })
+                .unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 200u64,
+                })
+                .unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 300u64,
+                })
+                .unwrap();
+
+            writer.commit().unwrap();
+        }
+
+        let reader = index.reader().unwrap();
+        let searcher = reader.searcher();
+        let segment = &searcher.segment_readers()[0];
+        let field = segment.fast_fields().u64(num_field).unwrap();
+
+        let numbers = vec![100, 200, 300];
+        let test_range = |range: RangeInclusive<u64>| {
+            let expexted_count = numbers.iter().filter(|num| range.contains(num)).count();
+            let mut vec = vec![];
+            field.get_docids_for_value_range(range, 0..u32::MAX, &mut vec);
+            assert_eq!(vec.len(), expexted_count);
+        };
+        test_range(50..=50);
+        test_range(150..=150);
+        test_range(350..=350);
+        test_range(100..=250);
+        test_range(101..=200);
+        test_range(101..=199);
+        test_range(100..=300);
+        test_range(100..=299);
+    }
+
+    #[test]
+    fn test_mapping_bug_docids_for_value_range() {
+        let mut schema_builder = Schema::builder();
+        let num_field = schema_builder.add_u64_field("url_norm_hash", FAST | INDEXED);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        {
+            // Values without gcd, but with min_value
+            let mut writer = index.writer_for_tests().unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 1000u64,
+                })
+                .unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 1001u64,
+                })
+                .unwrap();
+            writer
+                .add_document(doc! {
+                    num_field => 1003u64,
+                })
+                .unwrap();
+            writer.commit().unwrap();
+        }
+
+        let reader = index.reader().unwrap();
+        let searcher = reader.searcher();
+        let segment = &searcher.segment_readers()[0];
+        let field = segment.fast_fields().u64(num_field).unwrap();
+
+        let numbers = vec![1000, 1001, 1003];
+        let test_range = |range: RangeInclusive<u64>| {
+            let expexted_count = numbers.iter().filter(|num| range.contains(num)).count();
+            let mut vec = vec![];
+            field.get_docids_for_value_range(range, 0..u32::MAX, &mut vec);
+            assert_eq!(vec.len(), expexted_count);
+        };
+        let test_range_variant = |start, stop| {
+            let start_range = start..=stop;
+            test_range(start_range);
+            let start_range = start..=(stop - 1);
+            test_range(start_range);
+            let start_range = start..=(stop + 1);
+            test_range(start_range);
+            let start_range = (start - 1)..=stop;
+            test_range(start_range);
+            let start_range = (start - 1)..=(stop - 1);
+            test_range(start_range);
+            let start_range = (start - 1)..=(stop + 1);
+            test_range(start_range);
+            let start_range = (start + 1)..=stop;
+            test_range(start_range);
+            let start_range = (start + 1)..=(stop - 1);
+            test_range(start_range);
+            let start_range = (start + 1)..=(stop + 1);
+            test_range(start_range);
+        };
+        test_range_variant(50, 50);
+        test_range_variant(1000, 1000);
+        test_range_variant(1000, 1002);
+    }
 }

src/fastfield/multivalued/mod.rs

@@ -5,7 +5,7 @@ mod writer;
 use fastfield_codecs::FastFieldCodecType;
 pub use index::MultiValueIndex;
 
-pub use self::reader::{MultiValuedFastFieldReader, MultiValuedU128FastFieldReader};
+pub use self::reader::MultiValuedFastFieldReader;
 pub(crate) use self::writer::MultivalueStartIndex;
 pub use self::writer::{MultiValueU128FastFieldWriter, MultiValuedFastFieldWriter};
 
@@ -525,7 +525,7 @@ mod bench {
             serializer.close().unwrap();
             field
         };
-        let file = directory.open_read(&path).unwrap();
+        let file = directory.open_read(path).unwrap();
         {
             let fast_fields_composite = CompositeFile::open(&file).unwrap();
             let data_idx = fast_fields_composite

src/fastfield/multivalued/reader.rs

@@ -1,107 +1,31 @@
+use core::fmt;
 use std::ops::{Range, RangeInclusive};
 use std::sync::Arc;
 
-use fastfield_codecs::{Column, MonotonicallyMappableToU128};
+use fastfield_codecs::Column;
 
 use super::MultiValueIndex;
-use crate::fastfield::FastValue;
+use crate::fastfield::MakeZero;
 use crate::DocId;
 
-/// Reader for a multivalued `u64` fast field.
+/// Reader for a multivalued fast field.
 ///
-/// The reader is implemented as two `u64` fast field.
+/// The reader is implemented as two fast fields, one u64 fast field for the index and one for the
+/// values.
 ///
-/// The `vals_reader` will access the concatenated list of all
-/// values for all reader.
-/// The `idx_reader` associated, for each document, the index of its first value.
-/// Stores the start position for each document.
+/// The `vals_reader` will access the concatenated list of all values.
+/// The `idx_reader` associates, for each document, the index of its first value.
 #[derive(Clone)]
-pub struct MultiValuedFastFieldReader<Item: FastValue> {
-    idx_reader: MultiValueIndex,
-    vals_reader: Arc<dyn Column<Item>>,
-}
-
-impl<Item: FastValue> MultiValuedFastFieldReader<Item> {
-    pub(crate) fn open(
-        idx_reader: Arc<dyn Column<u64>>,
-        vals_reader: Arc<dyn Column<Item>>,
-    ) -> MultiValuedFastFieldReader<Item> {
-        MultiValuedFastFieldReader {
-            idx_reader: MultiValueIndex::new(idx_reader),
-            vals_reader,
-        }
-    }
-
-    /// Returns the array of values associated with the given `doc`.
-    #[inline]
-    fn get_vals_for_range(&self, range: Range<u32>, vals: &mut Vec<Item>) {
-        let len = (range.end - range.start) as usize;
-        vals.resize(len, Item::make_zero());
-        self.vals_reader
-            .get_range(range.start as u64, &mut vals[..]);
-    }
-
-    /// Returns the array of values associated with the given `doc`.
-    #[inline]
-    pub fn get_vals(&self, doc: DocId, vals: &mut Vec<Item>) {
-        let range = self.idx_reader.range(doc);
-        self.get_vals_for_range(range, vals);
-    }
-
-    /// returns the multivalue index
-    pub fn get_index_reader(&self) -> &MultiValueIndex {
-        &self.idx_reader
-    }
-
-    /// Returns the minimum value for this fast field.
-    ///
-    /// The min value does not take in account of possible
-    /// deleted document, and should be considered as a lower bound
-    /// of the actual minimum value.
-    pub fn min_value(&self) -> Item {
-        self.vals_reader.min_value()
-    }
-
-    /// Returns the maximum value for this fast field.
-    ///
-    /// The max value does not take in account of possible
-    /// deleted document, and should be considered as an upper bound
-    /// of the actual maximum value.
-    pub fn max_value(&self) -> Item {
-        self.vals_reader.max_value()
-    }
-
-    /// Returns the number of values associated with the document `DocId`.
-    #[inline]
-    pub fn num_vals(&self, doc: DocId) -> u32 {
-        self.idx_reader.num_vals_for_doc(doc)
-    }
-
-    /// Returns the overall number of values in this field.
-    #[inline]
-    pub fn total_num_vals(&self) -> u32 {
-        self.idx_reader.total_num_vals()
-    }
-}
-
-/// Reader for a multivalued `u128` fast field.
-///
-/// The reader is implemented as a `u64` fast field for the index and a `u128` fast field.
-///
-/// The `vals_reader` will access the concatenated list of all
-/// values for all reader.
-/// The `idx_reader` associated, for each document, the index of its first value.
-#[derive(Clone)]
-pub struct MultiValuedU128FastFieldReader<T: MonotonicallyMappableToU128> {
+pub struct MultiValuedFastFieldReader<T> {
     idx_reader: MultiValueIndex,
     vals_reader: Arc<dyn Column<T>>,
 }
 
-impl<T: MonotonicallyMappableToU128> MultiValuedU128FastFieldReader<T> {
+impl<T: PartialOrd + MakeZero + Copy + fmt::Debug> MultiValuedFastFieldReader<T> {
     pub(crate) fn open(
         idx_reader: Arc<dyn Column<u64>>,
         vals_reader: Arc<dyn Column<T>>,
-    ) -> MultiValuedU128FastFieldReader<T> {
+    ) -> MultiValuedFastFieldReader<T> {
         Self {
             idx_reader: MultiValueIndex::new(idx_reader),
             vals_reader,
@@ -122,7 +46,7 @@ impl<T: MonotonicallyMappableToU128> MultiValuedU128FastFieldReader<T> {
     #[inline]
     fn get_vals_for_range(&self, range: Range<u32>, vals: &mut Vec<T>) {
         let len = (range.end - range.start) as usize;
-        vals.resize(len, T::from_u128(0));
+        vals.resize(len, T::make_zero());
         self.vals_reader
             .get_range(range.start as u64, &mut vals[..]);
     }
@@ -199,8 +123,131 @@ impl<T: MonotonicallyMappableToU128> MultiValuedU128FastFieldReader<T> {
 #[cfg(test)]
 mod tests {
 
+    use time::{Duration, OffsetDateTime};
+
+    use crate::collector::Count;
     use crate::core::Index;
+    use crate::query::RangeQuery;
     use crate::schema::{Cardinality, Facet, FacetOptions, NumericOptions, Schema};
+    use crate::{DateOptions, DatePrecision, DateTime};
+
+    #[test]
+    fn test_multivalued_date_docids_for_value_range_1() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let date_field = schema_builder.add_date_field(
+            "multi_date_field",
+            DateOptions::default()
+                .set_fast(Cardinality::MultiValues)
+                .set_indexed()
+                .set_fieldnorm()
+                .set_precision(DatePrecision::Microseconds)
+                .set_stored(),
+        );
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_for_tests()?;
+        let first_time_stamp = OffsetDateTime::now_utc();
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_utc(first_time_stamp),
+            date_field => DateTime::from_utc(first_time_stamp),
+        ))?;
+        // add another second
+        let two_secs_ahead = first_time_stamp + Duration::seconds(2);
+        index_writer.commit()?;
+
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+        let reader = searcher.segment_reader(0);
+
+        let date_ff_reader = reader.fast_fields().dates(date_field).unwrap();
+        let mut docids = vec![];
+        date_ff_reader.get_docids_for_value_range(
+            DateTime::from_utc(first_time_stamp)..=DateTime::from_utc(two_secs_ahead),
+            0..5,
+            &mut docids,
+        );
+        assert_eq!(docids, vec![0]);
+
+        let count_multiples =
+            |range_query: RangeQuery| searcher.search(&range_query, &Count).unwrap();
+
+        assert_eq!(
+            count_multiples(RangeQuery::new_date(
+                date_field,
+                DateTime::from_utc(first_time_stamp)..DateTime::from_utc(two_secs_ahead)
+            )),
+            1
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_multivalued_date_docids_for_value_range_2() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let date_field = schema_builder.add_date_field(
+            "multi_date_field",
+            DateOptions::default()
+                .set_fast(Cardinality::MultiValues)
+                // TODO: Test different precision after fixing https://github.com/quickwit-oss/tantivy/issues/1783
+                .set_precision(DatePrecision::Microseconds)
+                .set_indexed()
+                .set_fieldnorm()
+                .set_stored(),
+        );
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_for_tests()?;
+        let first_time_stamp = OffsetDateTime::now_utc();
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_utc(first_time_stamp),
+            date_field => DateTime::from_utc(first_time_stamp),
+        ))?;
+        index_writer.add_document(doc!())?;
+        // add one second
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_utc(first_time_stamp + Duration::seconds(1)),
+        ))?;
+        // add another second
+        let two_secs_ahead = first_time_stamp + Duration::seconds(2);
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_utc(two_secs_ahead),
+            date_field => DateTime::from_utc(two_secs_ahead),
+            date_field => DateTime::from_utc(two_secs_ahead),
+        ))?;
+        // add three seconds
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_utc(first_time_stamp + Duration::seconds(3)),
+        ))?;
+        index_writer.commit()?;
+
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+        let reader = searcher.segment_reader(0);
+        assert_eq!(reader.num_docs(), 5);
+
+        let date_ff_reader = reader.fast_fields().dates(date_field).unwrap();
+        let mut docids = vec![];
+        date_ff_reader.get_docids_for_value_range(
+            DateTime::from_utc(first_time_stamp)..=DateTime::from_utc(two_secs_ahead),
+            0..5,
+            &mut docids,
+        );
+        assert_eq!(docids, vec![0, 2, 3]);
+
+        let count_multiples =
+            |range_query: RangeQuery| searcher.search(&range_query, &Count).unwrap();
+
+        assert_eq!(
+            count_multiples(RangeQuery::new_date(
+                date_field,
+                DateTime::from_utc(first_time_stamp)..DateTime::from_utc(two_secs_ahead)
+            )),
+            2
+        );
+
+        Ok(())
+    }
 
     #[test]
     fn test_multifastfield_reader() -> crate::Result<()> {

src/fastfield/readers.rs

@@ -3,11 +3,9 @@ use std::sync::Arc;
 
 use fastfield_codecs::{open, open_u128, Column};
 
-use super::multivalued::MultiValuedU128FastFieldReader;
+use super::multivalued::MultiValuedFastFieldReader;
 use crate::directory::{CompositeFile, FileSlice};
-use crate::fastfield::{
-    BytesFastFieldReader, FastFieldNotAvailableError, FastValue, MultiValuedFastFieldReader,
-};
+use crate::fastfield::{BytesFastFieldReader, FastFieldNotAvailableError, FastValue};
 use crate::schema::{Cardinality, Field, FieldType, Schema};
 use crate::space_usage::PerFieldSpaceUsage;
 use crate::{DateTime, TantivyError};
@@ -161,20 +159,14 @@ impl FastFieldReaders {
     /// Returns the `ip` fast field reader reader associated to `field`.
     ///
     /// If `field` is not a u128 fast field, this method returns an Error.
-    pub fn ip_addrs(
-        &self,
-        field: Field,
-    ) -> crate::Result<MultiValuedU128FastFieldReader<Ipv6Addr>> {
+    pub fn ip_addrs(&self, field: Field) -> crate::Result<MultiValuedFastFieldReader<Ipv6Addr>> {
         self.check_type(field, FastType::U128, Cardinality::MultiValues)?;
         let idx_reader: Arc<dyn Column<u64>> = self.typed_fast_field_reader(field)?;
 
         let bytes = self.fast_field_data(field, 1)?.read_bytes()?;
         let vals_reader = open_u128::<Ipv6Addr>(bytes)?;
 
-        Ok(MultiValuedU128FastFieldReader::open(
-            idx_reader,
-            vals_reader,
-        ))
+        Ok(MultiValuedFastFieldReader::open(idx_reader, vals_reader))
     }
 
     /// Returns the `u128` fast field reader reader associated to `field`.
@@ -189,17 +181,14 @@ impl FastFieldReaders {
     /// Returns the `u128` multi-valued fast field reader reader associated to `field`.
     ///
     /// If `field` is not a u128 multi-valued fast field, this method returns an Error.
-    pub fn u128s(&self, field: Field) -> crate::Result<MultiValuedU128FastFieldReader<u128>> {
+    pub fn u128s(&self, field: Field) -> crate::Result<MultiValuedFastFieldReader<u128>> {
         self.check_type(field, FastType::U128, Cardinality::MultiValues)?;
         let idx_reader: Arc<dyn Column<u64>> = self.typed_fast_field_reader(field)?;
 
         let bytes = self.fast_field_data(field, 1)?.read_bytes()?;
         let vals_reader = open_u128::<u128>(bytes)?;
 
-        Ok(MultiValuedU128FastFieldReader::open(
-            idx_reader,
-            vals_reader,
-        ))
+        Ok(MultiValuedFastFieldReader::open(idx_reader, vals_reader))
     }
 
     /// Returns the `u64` fast field reader reader associated with `field`, regardless of whether

src/fastfield/serializer/mod.rs

@@ -1,3 +1,4 @@
+use std::fmt;
 use std::io::{self, Write};
 
 pub use fastfield_codecs::Column;
@@ -49,7 +50,7 @@ impl CompositeFastFieldSerializer {
 
     /// Serialize data into a new u64 fast field. The best compression codec will be chosen
     /// automatically.
-    pub fn create_auto_detect_u64_fast_field<T: MonotonicallyMappableToU64>(
+    pub fn create_auto_detect_u64_fast_field<T: MonotonicallyMappableToU64 + fmt::Debug>(
         &mut self,
         field: Field,
         fastfield_accessor: impl Column<T>,
@@ -59,7 +60,9 @@ impl CompositeFastFieldSerializer {
 
     /// Serialize data into a new u64 fast field. The best compression codec will be chosen
     /// automatically.
-    pub fn create_auto_detect_u64_fast_field_with_idx<T: MonotonicallyMappableToU64>(
+    pub fn create_auto_detect_u64_fast_field_with_idx<
+        T: MonotonicallyMappableToU64 + fmt::Debug,
+    >(
         &mut self,
         field: Field,
         fastfield_accessor: impl Column<T>,
@@ -72,7 +75,9 @@ impl CompositeFastFieldSerializer {
 
     /// Serialize data into a new u64 fast field. The best compression codec of the the provided
     /// will be chosen.
-    pub fn create_auto_detect_u64_fast_field_with_idx_and_codecs<T: MonotonicallyMappableToU64>(
+    pub fn create_auto_detect_u64_fast_field_with_idx_and_codecs<
+        T: MonotonicallyMappableToU64 + fmt::Debug,
+    >(
         &mut self,
         field: Field,
         fastfield_accessor: impl Column<T>,

src/indexer/index_writer.rs

@@ -678,7 +678,7 @@ impl IndexWriter {
     /// only after calling `commit()`.
     #[doc(hidden)]
     pub fn delete_query(&self, query: Box<dyn Query>) -> crate::Result<Opstamp> {
-        let weight = query.weight(EnableScoring::Disabled(&self.index.schema()))?;
+        let weight = query.weight(EnableScoring::disabled_from_schema(&self.index.schema()))?;
         let opstamp = self.stamper.stamp();
         let delete_operation = DeleteOperation {
             opstamp,
@@ -759,7 +759,8 @@ impl IndexWriter {
             match user_op {
                 UserOperation::Delete(term) => {
                     let query = TermQuery::new(term, IndexRecordOption::Basic);
-                    let weight = query.weight(EnableScoring::Disabled(&self.index.schema()))?;
+                    let weight =
+                        query.weight(EnableScoring::disabled_from_schema(&self.index.schema()))?;
                     let delete_operation = DeleteOperation {
                         opstamp,
                         target: weight,

src/indexer/json_term_writer.rs

@@ -89,11 +89,11 @@ pub(crate) fn index_json_values<'a>(
     Ok(())
 }
 
-fn index_json_object<'a>(
+fn index_json_object(
     doc: DocId,
     json_value: &serde_json::Map<String, serde_json::Value>,
     text_analyzer: &TextAnalyzer,
-    json_term_writer: &mut JsonTermWriter<'a>,
+    json_term_writer: &mut JsonTermWriter,
     postings_writer: &mut dyn PostingsWriter,
     ctx: &mut IndexingContext,
     positions_per_path: &mut IndexingPositionsPerPath,
@@ -113,11 +113,11 @@ fn index_json_object<'a>(
     }
 }
 
-fn index_json_value<'a>(
+fn index_json_value(
     doc: DocId,
     json_value: &serde_json::Value,
     text_analyzer: &TextAnalyzer,
-    json_term_writer: &mut JsonTermWriter<'a>,
+    json_term_writer: &mut JsonTermWriter,
     postings_writer: &mut dyn PostingsWriter,
     ctx: &mut IndexingContext,
     positions_per_path: &mut IndexingPositionsPerPath,

src/indexer/merger.rs

@@ -13,7 +13,7 @@ use crate::docset::{DocSet, TERMINATED};
 use crate::error::DataCorruption;
 use crate::fastfield::{
     get_fastfield_codecs_for_multivalue, AliveBitSet, Column, CompositeFastFieldSerializer,
-    MultiValueIndex, MultiValuedFastFieldReader, MultiValuedU128FastFieldReader,
+    MultiValueIndex, MultiValuedFastFieldReader,
 };
 use crate::fieldnorm::{FieldNormReader, FieldNormReaders, FieldNormsSerializer, FieldNormsWriter};
 use crate::indexer::doc_id_mapping::{expect_field_id_for_sort_field, SegmentDocIdMapping};
@@ -331,18 +331,18 @@ impl IndexMerger {
         fast_field_serializer: &mut CompositeFastFieldSerializer,
         doc_id_mapping: &SegmentDocIdMapping,
     ) -> crate::Result<()> {
-        let segment_and_ff_readers: Vec<(&SegmentReader, MultiValuedU128FastFieldReader<u128>)> =
-            self.readers
-                .iter()
-                .map(|segment_reader| {
-                    let ff_reader: MultiValuedU128FastFieldReader<u128> =
-                        segment_reader.fast_fields().u128s(field).expect(
-                            "Failed to find index for multivalued field. This is a bug in \
-                             tantivy, please report.",
-                        );
-                    (segment_reader, ff_reader)
-                })
-                .collect::<Vec<_>>();
+        let segment_and_ff_readers: Vec<(&SegmentReader, MultiValuedFastFieldReader<u128>)> = self
+            .readers
+            .iter()
+            .map(|segment_reader| {
+                let ff_reader: MultiValuedFastFieldReader<u128> =
+                    segment_reader.fast_fields().u128s(field).expect(
+                        "Failed to find index for multivalued field. This is a bug in tantivy, \
+                         please report.",
+                    );
+                (segment_reader, ff_reader)
+            })
+            .collect::<Vec<_>>();
 
         Self::write_1_n_fast_field_idx_generic(
             field,

src/indexer/segment_updater.rs

@@ -577,7 +577,7 @@ impl SegmentUpdater {
         for merge_operation in merge_candidates {
             // If a merge cannot be started this is not a fatal error.
             // We do log a warning in `start_merge`.
-            let _ = self.start_merge(merge_operation);
+            drop(self.start_merge(merge_operation));
         }
     }
 

src/lib.rs

@@ -1,17 +1,14 @@
 #![doc(html_logo_url = "http://fulmicoton.com/tantivy-logo/tantivy-logo.png")]
 #![cfg_attr(all(feature = "unstable", test), feature(test))]
-#![cfg_attr(
-    feature = "cargo-clippy",
-    allow(
-        clippy::module_inception,
-        clippy::needless_range_loop,
-        clippy::bool_assert_comparison
-    )
-)]
 #![doc(test(attr(allow(unused_variables), deny(warnings))))]
 #![warn(missing_docs)]
-#![allow(clippy::len_without_is_empty)]
-#![allow(clippy::derive_partial_eq_without_eq)]
+#![allow(
+    clippy::len_without_is_empty,
+    clippy::derive_partial_eq_without_eq,
+    clippy::module_inception,
+    clippy::needless_range_loop,
+    clippy::bool_assert_comparison
+)]
 
 //! # `tantivy`
 //!
@@ -144,7 +141,7 @@ use crate::time::{OffsetDateTime, PrimitiveDateTime, UtcOffset};
 /// All constructors and conversions are provided as explicit
 /// functions and not by implementing any `From`/`Into` traits
 /// to prevent unintended usage.
-#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+#[derive(Clone, Default, Copy, PartialEq, Eq, PartialOrd, Ord)]
 pub struct DateTime {
     // Timestamp in microseconds.
     pub(crate) timestamp_micros: i64,

src/postings/compression/mod.rs

@@ -291,7 +291,7 @@ pub mod tests {
         const PADDING_VALUE: u32 = 234_234_345u32;
         let expected_length = 154;
         let mut encoder = BlockEncoder::new();
-        let input: Vec<u32> = (0u32..123u32).map(|i| 4 + i * 7 / 2).into_iter().collect();
+        let input: Vec<u32> = (0u32..123u32).map(|i| 4 + i * 7 / 2).collect();
         for offset in &[0u32, 1u32, 2u32] {
             let encoded_data = encoder.compress_vint_sorted(&input, *offset);
             assert!(encoded_data.len() <= expected_length);

src/postings/mod.rs

@@ -631,7 +631,7 @@ mod bench {
             let mut segment_postings = segment_reader
                 .inverted_index(TERM_A.field())
                 .unwrap()
-                .read_postings(&*TERM_A, IndexRecordOption::Basic)
+                .read_postings(&TERM_A, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             while segment_postings.advance() != TERMINATED {}
@@ -647,25 +647,25 @@ mod bench {
             let segment_postings_a = segment_reader
                 .inverted_index(TERM_A.field())
                 .unwrap()
-                .read_postings(&*TERM_A, IndexRecordOption::Basic)
+                .read_postings(&TERM_A, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             let segment_postings_b = segment_reader
                 .inverted_index(TERM_B.field())
                 .unwrap()
-                .read_postings(&*TERM_B, IndexRecordOption::Basic)
+                .read_postings(&TERM_B, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             let segment_postings_c = segment_reader
                 .inverted_index(TERM_C.field())
                 .unwrap()
-                .read_postings(&*TERM_C, IndexRecordOption::Basic)
+                .read_postings(&TERM_C, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             let segment_postings_d = segment_reader
                 .inverted_index(TERM_D.field())
                 .unwrap()
-                .read_postings(&*TERM_D, IndexRecordOption::Basic)
+                .read_postings(&TERM_D, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             let mut intersection = Intersection::new(vec![
@@ -687,7 +687,7 @@ mod bench {
         let mut segment_postings = segment_reader
             .inverted_index(TERM_A.field())
             .unwrap()
-            .read_postings(&*TERM_A, IndexRecordOption::Basic)
+            .read_postings(&TERM_A, IndexRecordOption::Basic)
             .unwrap()
             .unwrap();
 
@@ -705,7 +705,7 @@ mod bench {
             let mut segment_postings = segment_reader
                 .inverted_index(TERM_A.field())
                 .unwrap()
-                .read_postings(&*TERM_A, IndexRecordOption::Basic)
+                .read_postings(&TERM_A, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             for doc in &existing_docs {
@@ -746,7 +746,7 @@ mod bench {
             let mut segment_postings = segment_reader
                 .inverted_index(TERM_A.field())
                 .unwrap()
-                .read_postings(&*TERM_A, IndexRecordOption::Basic)
+                .read_postings(&TERM_A, IndexRecordOption::Basic)
                 .unwrap()
                 .unwrap();
             let mut s = 0u32;

src/query/all_query.rs

@@ -95,7 +95,7 @@ mod tests {
         let index = create_test_index()?;
         let reader = index.reader()?;
         let searcher = reader.searcher();
-        let weight = AllQuery.weight(EnableScoring::Disabled(&index.schema()))?;
+        let weight = AllQuery.weight(EnableScoring::disabled_from_schema(&index.schema()))?;
         {
             let reader = searcher.segment_reader(0);
             let mut scorer = weight.scorer(reader, 1.0)?;
@@ -118,7 +118,7 @@ mod tests {
         let index = create_test_index()?;
         let reader = index.reader()?;
         let searcher = reader.searcher();
-        let weight = AllQuery.weight(EnableScoring::Disabled(searcher.schema()))?;
+        let weight = AllQuery.weight(EnableScoring::disabled_from_schema(searcher.schema()))?;
         let reader = searcher.segment_reader(0);
         {
             let mut scorer = weight.scorer(reader, 2.0)?;

src/query/boolean_query/boolean_query.rs

@@ -146,7 +146,7 @@ impl Query for BooleanQuery {
         let sub_weights = self
             .subqueries
             .iter()
-            .map(|&(ref occur, ref subquery)| Ok((*occur, subquery.weight(enable_scoring)?)))
+            .map(|(occur, subquery)| Ok((*occur, subquery.weight(enable_scoring)?)))
             .collect::<crate::Result<_>>()?;
         Ok(Box::new(BooleanWeight::new(
             sub_weights,

src/query/boolean_query/boolean_weight.rs

@@ -91,7 +91,7 @@ impl<TScoreCombiner: ScoreCombiner> BooleanWeight<TScoreCombiner> {
         boost: Score,
     ) -> crate::Result<HashMap<Occur, Vec<Box<dyn Scorer>>>> {
         let mut per_occur_scorers: HashMap<Occur, Vec<Box<dyn Scorer>>> = HashMap::new();
-        for &(ref occur, ref subweight) in &self.weights {
+        for (occur, subweight) in &self.weights {
             let sub_scorer: Box<dyn Scorer> = subweight.scorer(reader, boost)?;
             per_occur_scorers
                 .entry(*occur)
@@ -191,7 +191,7 @@ impl<TScoreCombiner: ScoreCombiner + Sync> Weight for BooleanWeight<TScoreCombin
         }
 
         let mut explanation = Explanation::new("BooleanClause. Sum of ...", scorer.score());
-        for &(ref occur, ref subweight) in &self.weights {
+        for (occur, subweight) in &self.weights {
             if is_positive_occur(*occur) {
                 if let Ok(child_explanation) = subweight.explain(reader, doc) {
                     explanation.add_detail(child_explanation);

src/query/boolean_query/mod.rs

@@ -98,7 +98,7 @@ mod tests {
         }
         {
             let query = query_parser.parse_query("+a b")?;
-            let weight = query.weight(EnableScoring::Disabled(searcher.schema()))?;
+            let weight = query.weight(EnableScoring::disabled_from_schema(searcher.schema()))?;
             let scorer = weight.scorer(searcher.segment_reader(0u32), 1.0)?;
             assert!(scorer.is::<TermScorer>());
         }

src/query/fuzzy_query.rs

@@ -1,8 +1,5 @@
-use std::collections::HashMap;
-use std::ops::Range;
-
 use levenshtein_automata::{Distance, LevenshteinAutomatonBuilder, DFA};
-use once_cell::sync::Lazy;
+use once_cell::sync::OnceCell;
 use tantivy_fst::Automaton;
 
 use crate::query::{AutomatonWeight, EnableScoring, Query, Weight};
@@ -34,22 +31,6 @@ impl Automaton for DfaWrapper {
     }
 }
 
-/// A range of Levenshtein distances that we will build DFAs for our terms
-/// The computation is exponential, so best keep it to low single digits
-const VALID_LEVENSHTEIN_DISTANCE_RANGE: Range<u8> = 0..3;
-
-static LEV_BUILDER: Lazy<HashMap<(u8, bool), LevenshteinAutomatonBuilder>> = Lazy::new(|| {
-    let mut lev_builder_cache = HashMap::new();
-    // TODO make population lazy on a `(distance, val)` basis
-    for distance in VALID_LEVENSHTEIN_DISTANCE_RANGE {
-        for &transposition in &[false, true] {
-            let lev_automaton_builder = LevenshteinAutomatonBuilder::new(distance, transposition);
-            lev_builder_cache.insert((distance, transposition), lev_automaton_builder);
-        }
-    }
-    lev_builder_cache
-});
-
 /// A Fuzzy Query matches all of the documents
 /// containing a specific term that is within
 /// Levenshtein distance
@@ -129,30 +110,39 @@ impl FuzzyTermQuery {
     }
 
     fn specialized_weight(&self) -> crate::Result<AutomatonWeight<DfaWrapper>> {
-        // LEV_BUILDER is a HashMap, whose `get` method returns an Option
-        match LEV_BUILDER.get(&(self.distance, self.transposition_cost_one)) {
-            // Unwrap the option and build the Ok(AutomatonWeight)
-            Some(automaton_builder) => {
-                let term_text = self.term.as_str().ok_or_else(|| {
-                    crate::TantivyError::InvalidArgument(
-                        "The fuzzy term query requires a string term.".to_string(),
-                    )
-                })?;
-                let automaton = if self.prefix {
-                    automaton_builder.build_prefix_dfa(term_text)
-                } else {
-                    automaton_builder.build_dfa(term_text)
-                };
-                Ok(AutomatonWeight::new(
-                    self.term.field(),
-                    DfaWrapper(automaton),
+        static AUTOMATON_BUILDER: [[OnceCell<LevenshteinAutomatonBuilder>; 2]; 3] = [
+            [OnceCell::new(), OnceCell::new()],
+            [OnceCell::new(), OnceCell::new()],
+            [OnceCell::new(), OnceCell::new()],
+        ];
+
+        let automaton_builder = AUTOMATON_BUILDER
+            .get(self.distance as usize)
+            .ok_or_else(|| {
+                InvalidArgument(format!(
+                    "Levenshtein distance of {} is not allowed. Choose a value less than {}",
+                    self.distance,
+                    AUTOMATON_BUILDER.len()
                 ))
-            }
-            None => Err(InvalidArgument(format!(
-                "Levenshtein distance of {} is not allowed. Choose a value in the {:?} range",
-                self.distance, VALID_LEVENSHTEIN_DISTANCE_RANGE
-            ))),
-        }
+            })?
+            .get(self.transposition_cost_one as usize)
+            .unwrap()
+            .get_or_init(|| {
+                LevenshteinAutomatonBuilder::new(self.distance, self.transposition_cost_one)
+            });
+
+        let term_text = self.term.as_str().ok_or_else(|| {
+            InvalidArgument("The fuzzy term query requires a string term.".to_string())
+        })?;
+        let automaton = if self.prefix {
+            automaton_builder.build_prefix_dfa(term_text)
+        } else {
+            automaton_builder.build_dfa(term_text)
+        };
+        Ok(AutomatonWeight::new(
+            self.term.field(),
+            DfaWrapper(automaton),
+        ))
     }
 }