Removed azure stuff

* add date_histogram

* add return result

.github/workflows/coverage.yml

@@ -2,9 +2,9 @@ name: Coverage
 
 on:
   push:
-    branches: [ main ]
+    branches: [main]
   pull_request:
-    branches: [ main ]
+    branches: [main]
 
 jobs:
   coverage:
@@ -16,7 +16,7 @@ jobs:
       - uses: Swatinem/rust-cache@v2
       - uses: taiki-e/install-action@cargo-llvm-cov
       - name: Generate code coverage
-        run: cargo +nightly llvm-cov --all-features --workspace --lcov --output-path lcov.info
+        run: cargo +nightly llvm-cov --all-features --workspace --doctests --lcov --output-path lcov.info
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v3
         continue-on-error: true

Cargo.toml

@@ -15,15 +15,14 @@ rust-version = "1.62"
 
 [dependencies]
 oneshot = "0.1.5"
-base64 = "0.20.0"
-byteorder = "1.4.3"
+base64 = "0.21.0"
 crc32fast = "1.3.2"
 once_cell = "1.10.0"
 regex = { version = "1.5.5", default-features = false, features = ["std", "unicode"] }
 aho-corasick = "0.7"
 tantivy-fst = "0.4.0"
 memmap2 = { version = "0.5.3", optional = true }
-lz4_flex = { version = "0.9.2", default-features = false, features = ["checked-decode"], optional = true }
+lz4_flex = { version = "0.10", default-features = false, features = ["checked-decode"], optional = true }
 brotli = { version = "3.3.4", optional = true }
 zstd = { version = "0.12", optional = true, default-features = false }
 snap = { version = "1.0.5", optional = true }
@@ -44,23 +43,24 @@ rustc-hash = "1.1.0"
 thiserror = "1.0.30"
 htmlescape = "0.3.1"
 fail = "0.5.0"
-murmurhash32 = "0.2.0"
+murmurhash32 = "0.3.0"
 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"
 async-trait = "0.1.53"
 arc-swap = "1.5.0"
 
+columnar = { version="0.1", path="./columnar", package ="tantivy-columnar" }
 sstable = { version="0.1", path="./sstable", package ="tantivy-sstable", optional = true }
 stacker = { version="0.1", path="./stacker", package ="tantivy-stacker" }
-tantivy-query-grammar = { version= "0.19.0", path="./query-grammar" }
-tantivy-bitpacker = 		{ version= "0.3", path="./bitpacker" }
-common = 								{ version= "0.5", path = "./common/", package = "tantivy-common" }
-fastfield_codecs = 			{ version= "0.3", path="./fastfield_codecs", default-features = false }
+query-grammar = { version= "0.19.0", path="./query-grammar", package = "tantivy-query-grammar" }
+tantivy-bitpacker = { version= "0.3", path="./bitpacker" }
+common = { version= "0.5", path = "./common/", package = "tantivy-common" }
+tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
 
 [target.'cfg(windows)'.dependencies]
 winapi = "0.3.9"
@@ -76,6 +76,7 @@ test-log = "0.2.10"
 env_logger = "0.10.0"
 pprof = { version = "0.11.0", features = ["flamegraph", "criterion"] }
 futures = "0.3.21"
+paste = "1.0.11"
 
 [dev-dependencies.fail]
 version = "0.5.0"
@@ -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", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
 
 # 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)
@@ -41,13 +41,13 @@ Your mileage WILL vary depending on the nature of queries and their load.
 - SIMD integer compression when the platform/CPU includes the SSE2 instruction set
 - 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
+- Text, i64, u64, f64, dates, ip, bool, and hierarchical facet fields
+- 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
@@ -80,10 +80,11 @@ There are many ways to support this project.
 # Contributing code
 
 We use the GitHub Pull Request workflow: reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
+Feel free to update CHANGELOG.md with your contribution.
 
-## Minimum supported Rust version
+## Tokenizer
 
-Tantivy currently requires at least Rust 1.62 or later to compile.
+When implementing a tokenizer for tantivy depend on the `tantivy-tokenizer-api` crate.
 
 ## Clone and build locally
 
@@ -91,41 +92,9 @@ Tantivy compiles on stable Rust.
 To check out and run tests, you can simply run:
 
 ```bash
-    git clone https://github.com/quickwit-oss/tantivy.git
-    cd tantivy
-    cargo build
-```
-
-## Run tests
-
-Some tests will not run with just `cargo test` because of `fail-rs`.
-To run the tests exhaustively, run `./run-tests.sh`.
-
-## Debug
-
-You might find it useful to step through the programme with a debugger.
-
-### A failing test
-
-Make sure you haven't run `cargo clean` after the most recent `cargo test` or `cargo build` to guarantee that the `target/` directory exists. Use this bash script to find the name of the most recent debug build of Tantivy and run it under `rust-gdb`:
-
-```bash
-find target/debug/ -maxdepth 1 -executable -type f -name "tantivy*" -printf '%TY-%Tm-%Td %TT %p\n' | sort -r | cut -d " " -f 3 | xargs -I RECENT_DBG_TANTIVY rust-gdb RECENT_DBG_TANTIVY
-```
-
-Now that you are in `rust-gdb`, you can set breakpoints on lines and methods that match your source code and run the debug executable with flags that you normally pass to `cargo test` like this:
-
-```bash
-$gdb run --test-threads 1 --test $NAME_OF_TEST
-```
-
-### An example
-
-By default, `rustc` compiles everything in the `examples/` directory in debug mode. This makes it easy for you to make examples to reproduce bugs:
-
-```bash
-rust-gdb target/debug/examples/$EXAMPLE_NAME
-$ gdb run
+git clone https://github.com/quickwit-oss/tantivy.git
+cd tantivy
+cargo test
 ```
 
 # Companies Using Tantivy

TODO.txt

@@ -0,0 +1,18 @@
+Make schema_builder API fluent.
+fix doc serialization and prevent compression problems
+
+u64 , etc. shoudl return Resutl<Option> now that we support optional missing a column is really not an error
+remove fastfield codecs
+ditch the first_or_default trick. if it is still useful, improve its implementation.
+rename FastFieldReaders::open to load
+
+
+remove fast field reader
+
+find a way to unify the two DateTime.
+readd type check in the filter wrapper
+
+add unit test on columnar list columns.
+
+make sure sort works
+

benches/index-bench.rs

@@ -34,7 +34,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let index = Index::create_in_ram(schema.clone());
             let index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let doc = schema.parse_document(doc_json).unwrap();
                     index_writer.add_document(doc).unwrap();
                 }
@@ -46,7 +46,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let index = Index::create_in_ram(schema.clone());
             let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let doc = schema.parse_document(doc_json).unwrap();
                     index_writer.add_document(doc).unwrap();
                 }
@@ -59,7 +59,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let index = Index::create_in_ram(schema_with_store.clone());
             let index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let doc = schema.parse_document(doc_json).unwrap();
                     index_writer.add_document(doc).unwrap();
                 }
@@ -71,7 +71,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let index = Index::create_in_ram(schema_with_store.clone());
             let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let doc = schema.parse_document(doc_json).unwrap();
                     index_writer.add_document(doc).unwrap();
                 }
@@ -85,7 +85,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let json_field = dynamic_schema.get_field("json").unwrap();
             let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let json_val: serde_json::Map<String, serde_json::Value> =
                         serde_json::from_str(doc_json).unwrap();
                     let doc = tantivy::doc!(json_field=>json_val);
@@ -101,7 +101,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
             let json_field = dynamic_schema.get_field("json").unwrap();
             let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
             for _ in 0..NUM_REPEATS {
-                for doc_json in HDFS_LOGS.trim().split("\n") {
+                for doc_json in HDFS_LOGS.trim().split('\n') {
                     let json_val: serde_json::Map<String, serde_json::Value> =
                         serde_json::from_str(doc_json).unwrap();
                     let doc = tantivy::doc!(json_field=>json_val);

bitpacker/Cargo.toml

@@ -15,3 +15,7 @@ homepage = "https://github.com/quickwit-oss/tantivy"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
+
+[dev-dependencies]
+rand = "0.8"
+proptest = "1"

bitpacker/benches/bench.rs

@@ -4,9 +4,39 @@ extern crate test;
 
 #[cfg(test)]
 mod tests {
-    use tantivy_bitpacker::BlockedBitpacker;
+    use rand::seq::IteratorRandom;
+    use rand::thread_rng;
+    use tantivy_bitpacker::{BitPacker, BitUnpacker, BlockedBitpacker};
     use test::Bencher;
 
+    #[inline(never)]
+    fn create_bitpacked_data(bit_width: u8, num_els: u32) -> Vec<u8> {
+        let mut bitpacker = BitPacker::new();
+        let mut buffer = Vec::new();
+        for _ in 0..num_els {
+            // the values do not matter.
+            bitpacker.write(0u64, bit_width, &mut buffer).unwrap();
+            bitpacker.flush(&mut buffer).unwrap();
+        }
+        buffer
+    }
+
+    #[bench]
+    fn bench_bitpacking_read(b: &mut Bencher) {
+        let bit_width = 3;
+        let num_els = 1_000_000u32;
+        let bit_unpacker = BitUnpacker::new(bit_width);
+        let data = create_bitpacked_data(bit_width, num_els);
+        let idxs: Vec<u32> = (0..num_els).choose_multiple(&mut thread_rng(), 100_000);
+        b.iter(|| {
+            let mut out = 0u64;
+            for &idx in &idxs {
+                out = out.wrapping_add(bit_unpacker.get(idx, &data[..]));
+            }
+            out
+        });
+    }
+
     #[bench]
     fn bench_blockedbitp_read(b: &mut Bencher) {
         let mut blocked_bitpacker = BlockedBitpacker::new();
@@ -14,9 +44,9 @@ mod tests {
             blocked_bitpacker.add(val * val);
         }
         b.iter(|| {
-            let mut out = 0;
+            let mut out = 0u64;
             for val in 0..=21500 {
-                out = blocked_bitpacker.get(val);
+                out = out.wrapping_add(blocked_bitpacker.get(val));
             }
             out
         });

bitpacker/src/bitpacker.rs

@@ -19,7 +19,7 @@ impl BitPacker {
     }
 
     #[inline]
-    pub fn write<TWrite: io::Write>(
+    pub fn write<TWrite: io::Write + ?Sized>(
         &mut self,
         val: u64,
         num_bits: u8,
@@ -43,7 +43,7 @@ impl BitPacker {
         Ok(())
     }
 
-    pub fn flush<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn flush<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
         if self.mini_buffer_written > 0 {
             let num_bytes = (self.mini_buffer_written + 7) / 8;
             let bytes = self.mini_buffer.to_le_bytes();
@@ -54,29 +54,33 @@ impl BitPacker {
         Ok(())
     }
 
-    pub fn close<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn close<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
         self.flush(output)?;
-        // Padding the write file to simplify reads.
-        output.write_all(&[0u8; 7])?;
         Ok(())
     }
 }
 
-#[derive(Clone, Debug, Default)]
+#[derive(Clone, Debug, Default, Copy)]
 pub struct BitUnpacker {
-    num_bits: u64,
+    num_bits: u32,
     mask: u64,
 }
 
 impl BitUnpacker {
+    /// Creates a bit unpacker, that assumes the same bitwidth for all values.
+    ///
+    /// The bitunpacker works by doing an unaligned read of 8 bytes.
+    /// For this reason, values of `num_bits` between
+    /// [57..63] are forbidden.
     pub fn new(num_bits: u8) -> BitUnpacker {
+        assert!(num_bits <= 7 * 8 || num_bits == 64);
         let mask: u64 = if num_bits == 64 {
             !0u64
         } else {
             (1u64 << num_bits) - 1u64
         };
         BitUnpacker {
-            num_bits: u64::from(num_bits),
+            num_bits: u32::from(num_bits),
             mask,
         }
     }
@@ -87,28 +91,40 @@ impl BitUnpacker {
 
     #[inline]
     pub fn get(&self, idx: u32, data: &[u8]) -> u64 {
-        if self.num_bits == 0 {
-            return 0u64;
-        }
-        let addr_in_bits = idx * self.num_bits as u32;
+        let addr_in_bits = idx * self.num_bits;
         let addr = (addr_in_bits >> 3) as usize;
+        if addr + 8 > data.len() {
+            if self.num_bits == 0 {
+                return 0;
+            }
+            let bit_shift = addr_in_bits & 7;
+            return self.get_slow_path(addr, bit_shift, data);
+        }
         let bit_shift = addr_in_bits & 7;
-        debug_assert!(
-            addr + 8 <= data.len(),
-            "The fast field field should have been padded with 7 bytes."
-        );
         let bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
         let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
         let val_shifted = val_unshifted_unmasked >> bit_shift;
         val_shifted & self.mask
     }
+
+    #[inline(never)]
+    fn get_slow_path(&self, addr: usize, bit_shift: u32, data: &[u8]) -> u64 {
+        let mut bytes: [u8; 8] = [0u8; 8];
+        let available_bytes = data.len() - addr;
+        // This function is meant to only be called if we did not have 8 bytes to load.
+        debug_assert!(available_bytes < 8);
+        bytes[..available_bytes].copy_from_slice(&data[addr..]);
+        let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
+        let val_shifted = val_unshifted_unmasked >> bit_shift;
+        val_shifted & self.mask
+    }
 }
 
 #[cfg(test)]
 mod test {
     use super::{BitPacker, BitUnpacker};
 
-    fn create_fastfield_bitpacker(len: usize, num_bits: u8) -> (BitUnpacker, Vec<u64>, Vec<u8>) {
+    fn create_bitpacker(len: usize, num_bits: u8) -> (BitUnpacker, Vec<u64>, Vec<u8>) {
         let mut data = Vec::new();
         let mut bitpacker = BitPacker::new();
         let max_val: u64 = (1u64 << num_bits as u64) - 1u64;
@@ -119,13 +135,13 @@ mod test {
             bitpacker.write(val, num_bits, &mut data).unwrap();
         }
         bitpacker.close(&mut data).unwrap();
-        assert_eq!(data.len(), ((num_bits as usize) * len + 7) / 8 + 7);
+        assert_eq!(data.len(), ((num_bits as usize) * len + 7) / 8);
         let bitunpacker = BitUnpacker::new(num_bits);
         (bitunpacker, vals, data)
     }
 
     fn test_bitpacker_util(len: usize, num_bits: u8) {
-        let (bitunpacker, vals, data) = create_fastfield_bitpacker(len, num_bits);
+        let (bitunpacker, vals, data) = create_bitpacker(len, num_bits);
         for (i, val) in vals.iter().enumerate() {
             assert_eq!(bitunpacker.get(i as u32, &data), *val);
         }
@@ -139,4 +155,49 @@ mod test {
         test_bitpacker_util(6, 14);
         test_bitpacker_util(1000, 14);
     }
+
+    use proptest::prelude::*;
+
+    fn num_bits_strategy() -> impl Strategy<Value = u8> {
+        prop_oneof!(Just(0), Just(1), 2u8..56u8, Just(56), Just(64),)
+    }
+
+    fn vals_strategy() -> impl Strategy<Value = (u8, Vec<u64>)> {
+        (num_bits_strategy(), 0usize..100usize).prop_flat_map(|(num_bits, len)| {
+            let max_val = if num_bits == 64 {
+                u64::MAX
+            } else {
+                (1u64 << num_bits as u32) - 1
+            };
+            let vals = proptest::collection::vec(0..=max_val, len);
+            vals.prop_map(move |vals| (num_bits, vals))
+        })
+    }
+
+    fn test_bitpacker_aux(num_bits: u8, vals: &[u64]) {
+        let mut buffer: Vec<u8> = Vec::new();
+        let mut bitpacker = BitPacker::new();
+        for &val in vals {
+            bitpacker.write(val, num_bits, &mut buffer).unwrap();
+        }
+        bitpacker.flush(&mut buffer).unwrap();
+        assert_eq!(buffer.len(), (vals.len() * num_bits as usize + 7) / 8);
+        let bitunpacker = BitUnpacker::new(num_bits);
+        let max_val = if num_bits == 64 {
+            u64::MAX
+        } else {
+            (1u64 << num_bits) - 1
+        };
+        for (i, val) in vals.iter().copied().enumerate() {
+            assert!(val <= max_val);
+            assert_eq!(bitunpacker.get(i as u32, &buffer), val);
+        }
+    }
+
+    proptest::proptest! {
+        #[test]
+        fn test_bitpacker_proptest((num_bits, vals) in vals_strategy()) {
+            test_bitpacker_aux(num_bits, &vals);
+        }
+    }
 }

ci/before_deploy.ps1

@@ -1,23 +0,0 @@
-# This script takes care of packaging the build artifacts that will go in the
-# release zipfile
-
-$SRC_DIR = $PWD.Path
-$STAGE = [System.Guid]::NewGuid().ToString()
-
-Set-Location $ENV:Temp
-New-Item -Type Directory -Name $STAGE
-Set-Location $STAGE
-
-$ZIP = "$SRC_DIR\$($Env:CRATE_NAME)-$($Env:APPVEYOR_REPO_TAG_NAME)-$($Env:TARGET).zip"
-
-# TODO Update this to package the right artifacts
-Copy-Item "$SRC_DIR\target\$($Env:TARGET)\release\hello.exe" '.\'
-
-7z a "$ZIP" *
-
-Push-AppveyorArtifact "$ZIP"
-
-Remove-Item *.* -Force
-Set-Location ..
-Remove-Item $STAGE
-Set-Location $SRC_DIR

ci/before_deploy.sh

@@ -1,33 +0,0 @@
-# This script takes care of building your crate and packaging it for release
-
-set -ex
-
-main() {
-    local src=$(pwd) \
-          stage=
-
-    case $TRAVIS_OS_NAME in
-        linux)
-            stage=$(mktemp -d)
-            ;;
-        osx)
-            stage=$(mktemp -d -t tmp)
-            ;;
-    esac
-
-    test -f Cargo.lock || cargo generate-lockfile
-
-    # TODO Update this to build the artifacts that matter to you
-    cross rustc --bin hello --target $TARGET --release -- -C lto
-
-    # TODO Update this to package the right artifacts
-    cp target/$TARGET/release/hello $stage/
-
-    cd $stage
-    tar czf $src/$CRATE_NAME-$TRAVIS_TAG-$TARGET.tar.gz *
-    cd $src
-
-    rm -rf $stage
-}
-
-main

ci/install.sh

@@ -1,47 +0,0 @@
-set -ex
-
-main() {
-    local target=
-    if [ $TRAVIS_OS_NAME = linux ]; then
-        target=x86_64-unknown-linux-musl
-        sort=sort
-    else
-        target=x86_64-apple-darwin
-        sort=gsort  # for `sort --sort-version`, from brew's coreutils.
-    fi
-
-    # Builds for iOS are done on OSX, but require the specific target to be
-    # installed.
-    case $TARGET in
-        aarch64-apple-ios)
-            rustup target install aarch64-apple-ios
-            ;;
-        armv7-apple-ios)
-            rustup target install armv7-apple-ios
-            ;;
-        armv7s-apple-ios)
-            rustup target install armv7s-apple-ios
-            ;;
-        i386-apple-ios)
-            rustup target install i386-apple-ios
-            ;;
-        x86_64-apple-ios)
-            rustup target install x86_64-apple-ios
-            ;;
-    esac
-
-    # This fetches latest stable release
-    local tag=$(git ls-remote --tags --refs --exit-code https://github.com/japaric/cross \
-                       | cut -d/ -f3 \
-                       | grep -E '^v[0.1.0-9.]+$' \
-                       | $sort --version-sort \
-                       | tail -n1)
-    curl -LSfs https://japaric.github.io/trust/install.sh | \
-        sh -s -- \
-           --force \
-           --git japaric/cross \
-           --tag $tag \
-           --target $target
-}
-
-main

ci/script.sh

@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-
-# This script takes care of testing your crate
-
-set -ex
-
-main() {
-    if [ ! -z $CODECOV ]; then
-        echo "Codecov"
-        cargo build --verbose && cargo coverage --verbose --all && bash <(curl -s https://codecov.io/bash) -s target/kcov
-    else
-        echo "Build"
-        cross build --target $TARGET
-        if [ ! -z $DISABLE_TESTS ]; then
-            return
-        fi
-        echo "Test"
-        cross test --target $TARGET --no-default-features --features mmap
-        cross test --target $TARGET --no-default-features --features mmap query-grammar
-    fi
-    for example in $(ls examples/*.rs)
-    do
-        cargo run --example  $(basename $example .rs)
-    done
-}
-
-# we don't run the "test phase" when doing deploys
-if [ -z $TRAVIS_TAG ]; then
-    main
-fi

columnar/Cargo.toml

@@ -0,0 +1,28 @@
+[package]
+name = "tantivy-columnar"
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+
+[dependencies]
+itertools = "0.10.5"
+log = "0.4.17"
+fnv = "1.0.7"
+fastdivide = "0.4.0"
+rand = { version = "0.8.5", optional = true }
+measure_time = { version = "0.8.2", optional = true }
+prettytable-rs = { version = "0.10.0", optional = true }
+
+stacker = { path = "../stacker", package="tantivy-stacker"}
+sstable = { path = "../sstable", package = "tantivy-sstable" }
+common = { path = "../common", package = "tantivy-common" }
+tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
+serde = "1.0.152"
+
+[dev-dependencies]
+proptest = "1"
+more-asserts = "0.3.1"
+rand = "0.8.5"
+
+[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/benches/bench_u128.rs

@@ -0,0 +1,124 @@
+#![feature(test)]
+
+use std::ops::RangeInclusive;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use rand::rngs::StdRng;
+use rand::seq::SliceRandom;
+use rand::{random, Rng, SeedableRng};
+use tantivy_columnar::ColumnValues;
+use test::Bencher;
+extern crate test;
+
+// TODO does this make sense for IPv6 ?
+fn generate_random() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64)
+        .map(|el| el + random::<u16>() as u64)
+        .collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+fn get_u128_column_random() -> Arc<dyn ColumnValues<u128>> {
+    let permutation = generate_random();
+    let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    get_u128_column_from_data(&permutation)
+}
+
+fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn ColumnValues<u128>> {
+    let mut out = vec![];
+    tantivy_columnar::column_values::serialize_column_values_u128(&data, &mut out).unwrap();
+    let out = OwnedBytes::new(out);
+    tantivy_columnar::column_values::open_u128_mapped::<u128>(out).unwrap()
+}
+
+const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
+const SINGLE_ITEM: u64 = 90;
+const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
+
+fn get_data_50percent_item() -> Vec<u128> {
+    let mut rng = StdRng::from_seed([1u8; 32]);
+
+    let mut data = vec![];
+    for _ in 0..300_000 {
+        let val = rng.gen_range(1..=100);
+        data.push(val);
+    }
+    data.push(SINGLE_ITEM);
+    data.shuffle(&mut rng);
+    let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    data
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            *FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(
+            *SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let column = get_u128_column_from_data(&data);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+// U128 RANGE END
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
+    let column = get_u128_column_random();
+
+    b.iter(|| {
+        let mut a = 0u128;
+        for i in 0u64..column.num_vals() as u64 {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
+    let column = get_u128_column_random();
+
+    b.iter(|| {
+        let n = column.num_vals();
+        let mut a = 0u128;
+        for i in (0..n / 5).map(|val| val * 5) {
+            a += column.get_val(i);
+        }
+        a
+    });
+}

columnar/benches/bench_u64.rs

@@ -0,0 +1,211 @@
+#![feature(test)]
+extern crate test;
+
+use std::ops::RangeInclusive;
+use std::sync::Arc;
+
+use rand::prelude::*;
+use tantivy_columnar::column_values::{serialize_and_load_u64_based_column_values, CodecType};
+use tantivy_columnar::*;
+use test::Bencher;
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+fn generate_random() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64)
+        .map(|el| el + random::<u16>() as u64)
+        .collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation_gcd() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+pub fn serialize_and_load(column: &[u64], codec_type: CodecType) -> Arc<dyn ColumnValues<u64>> {
+    serialize_and_load_u64_based_column_values(&column, &[codec_type])
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = permutation[a as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_fflookup_bitpacked(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = column.get_val(a as u32);
+        }
+        a
+    });
+}
+
+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 _ in 0..300_000 {
+        let val = rng.gen_range(1..=100);
+        data.push(val);
+    }
+    data.push(SINGLE_ITEM);
+
+    data.shuffle(&mut rng);
+    let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    data
+}
+
+// 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 ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_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 ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_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 ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_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 ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_row_ids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+// U64 RANGE END
+
+#[bench]
+fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += permutation[i as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    let column_ref = column.as_ref();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0u32..n as u32 {
+            a += column_ref.get_val(i);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
+    let permutation = generate_permutation_gcd();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..n {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..permutation.len() {
+            a += permutation[i as usize] as u64;
+        }
+        a
+    });
+}

columnar/columnar-cli/Cargo.toml

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

columnar/columnar-cli/src/main.rs

@@ -0,0 +1,134 @@
+use columnar::ColumnarWriter;
+use columnar::NumericalValue;
+use serde_json_borrow;
+use std::fs::File;
+use std::io;
+use std::io::BufRead;
+use std::io::BufReader;
+use std::time::Instant;
+
+#[derive(Default)]
+struct JsonStack {
+    path: String,
+    stack: Vec<usize>,
+}
+
+impl JsonStack {
+    fn push(&mut self, seg: &str) {
+        let len = self.path.len();
+        self.stack.push(len);
+        self.path.push('.');
+        self.path.push_str(seg);
+    }
+
+    fn pop(&mut self) {
+        if let Some(len) = self.stack.pop() {
+            self.path.truncate(len);
+        }
+    }
+
+    fn path(&self) -> &str {
+        &self.path[1..]
+    }
+}
+
+fn append_json_to_columnar(
+    doc: u32,
+    json_value: &serde_json_borrow::Value,
+    columnar: &mut ColumnarWriter,
+    stack: &mut JsonStack,
+) -> usize {
+    let mut count = 0;
+    match json_value {
+        serde_json_borrow::Value::Null => {}
+        serde_json_borrow::Value::Bool(val) => {
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                NumericalValue::from(if *val { 1u64 } else { 0u64 }),
+            );
+            count += 1;
+        }
+        serde_json_borrow::Value::Number(num) => {
+            let numerical_value: NumericalValue = if let Some(num_i64) = num.as_i64() {
+                num_i64.into()
+            } else if let Some(num_u64) = num.as_u64() {
+                num_u64.into()
+            } else if let Some(num_f64) = num.as_f64() {
+                num_f64.into()
+            } else {
+                panic!();
+            };
+            count += 1;
+            columnar.record_numerical(
+                doc,
+                stack.path(),
+                numerical_value,
+            );
+        }
+        serde_json_borrow::Value::Str(msg) => {
+            columnar.record_str(
+                doc,
+                stack.path(),
+                msg,
+            );
+            count += 1;
+        },
+        serde_json_borrow::Value::Array(vals) => {
+            for val in vals {
+                count += append_json_to_columnar(doc, val, columnar, stack);
+            }
+        },
+        serde_json_borrow::Value::Object(json_map) => {
+            for (child_key, child_val) in json_map {
+                stack.push(child_key);
+                count += append_json_to_columnar(doc, child_val, columnar, stack);
+                stack.pop();
+            }
+        },
+    }
+    count
+}
+
+fn main() -> io::Result<()> {
+    let file = File::open("gh_small.json")?;
+    let mut reader = BufReader::new(file);
+    let mut line = String::with_capacity(100);
+    let mut columnar = columnar::ColumnarWriter::default();
+    let mut doc = 0;
+    let start = Instant::now();
+    let mut stack = JsonStack::default();
+    let mut total_count = 0;
+
+    let start_build = Instant::now();
+    loop {
+        line.clear();
+        let len = reader.read_line(&mut line)?;
+        if len == 0 {
+            break;
+        }
+        let Ok(json_value) = serde_json::from_str::<serde_json_borrow::Value>(&line) else { continue; };
+        total_count += append_json_to_columnar(doc, &json_value, &mut columnar, &mut stack);
+        doc += 1;
+    }
+    println!("Build in {:?}", start_build.elapsed());
+
+    println!("value count {total_count}");
+
+    let mut buffer = Vec::new();
+    let start_serialize = Instant::now();
+    columnar.serialize(doc, None, &mut buffer)?;
+    println!("Serialized in {:?}", start_serialize.elapsed());
+    println!("num docs: {doc}, {:?}", start.elapsed());
+    println!("buffer len {} MB", buffer.len() / 1_000_000);
+    let columnar = columnar::ColumnarReader::open(buffer)?;
+    for (column_name, dynamic_column) in columnar.list_columns()? {
+        let num_bytes = dynamic_column.num_bytes();
+        let typ = dynamic_column.column_type();
+        if num_bytes > 1_000_000 {
+            println!("{column_name} {typ:?}  {} KB", num_bytes / 1_000);
+        }
+    }
+    println!("{} columns", columnar.num_columns());
+    Ok(())
+}

columnar/src/TODO.md

@@ -0,0 +1,47 @@
+# zero to one
+
+* revisit line codec
+* add columns from schema on merge
+* Plugging JSON
+* replug examples
+* move datetime to quickwit common
+* switch to nanos
+* reintroduce the gcd map.
+
+# Perf and Size
+* remove alloc in `ord_to_term`
++ multivaued range queries restrat frm the beginning all of the time.
+* 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
+rename columnar to something more explicit, like column_dictionary or columnar_table
+rename fastfield -> column
+document changes
+rationalization FastFieldValue, HasColumnType
+isolate u128_based and uniform naming
+
+# Other
+fix enhance column-cli
+
+# Santa claus
+autodetect datetime ipaddr, plug customizable tokenizer.

columnar/src/column/dictionary_encoded.rs

@@ -0,0 +1,100 @@
+use std::io;
+use std::ops::Deref;
+use std::sync::Arc;
+
+use sstable::{Dictionary, VoidSSTable};
+
+use crate::column::Column;
+use crate::RowId;
+
+/// Dictionary encoded column.
+///
+/// The column simply gives access to a regular u64-column that, in
+/// which the values are term-ordinals.
+///
+/// These ordinals are ids uniquely identify the bytes that are stored in
+/// the column. These ordinals are small, and sorted in the same order
+/// as the term_ord_column.
+#[derive(Clone)]
+pub struct BytesColumn {
+    pub(crate) dictionary: Arc<Dictionary<VoidSSTable>>,
+    pub(crate) term_ord_column: Column<u64>,
+}
+
+impl BytesColumn {
+    /// Fills the given `output` buffer with the term associated to the ordinal `ord`.
+    ///
+    /// Returns `false` if the term does not exist (e.g. `term_ord` is greater or equal to the
+    /// overll number of terms).
+    pub fn ord_to_bytes(&self, ord: u64, output: &mut Vec<u8>) -> io::Result<bool> {
+        self.dictionary.ord_to_term(ord, output)
+    }
+
+    /// Returns the number of rows in the column.
+    pub fn num_rows(&self) -> RowId {
+        self.term_ord_column.num_docs()
+    }
+
+    pub fn term_ords(&self, row_id: RowId) -> impl Iterator<Item = u64> + '_ {
+        self.term_ord_column.values_for_doc(row_id)
+    }
+
+    /// Returns the column of ordinals
+    pub fn ords(&self) -> &Column<u64> {
+        &self.term_ord_column
+    }
+
+    pub fn num_terms(&self) -> usize {
+        self.dictionary.num_terms()
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.dictionary.as_ref()
+    }
+}
+
+#[derive(Clone)]
+pub struct StrColumn(BytesColumn);
+
+impl From<StrColumn> for BytesColumn {
+    fn from(str_column: StrColumn) -> BytesColumn {
+        str_column.0
+    }
+}
+
+impl StrColumn {
+    pub(crate) fn wrap(bytes_column: BytesColumn) -> StrColumn {
+        StrColumn(bytes_column)
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.0.dictionary.as_ref()
+    }
+
+    /// Fills the buffer
+    pub fn ord_to_str(&self, term_ord: u64, output: &mut String) -> io::Result<bool> {
+        unsafe {
+            let buf = output.as_mut_vec();
+            if !self.0.dictionary.ord_to_term(term_ord, buf)? {
+                return Ok(false);
+            }
+            // TODO consider remove checks if it hurts performance.
+            if std::str::from_utf8(buf.as_slice()).is_err() {
+                buf.clear();
+                return Err(io::Error::new(
+                    io::ErrorKind::InvalidData,
+                    "Not valid utf-8",
+                ));
+            }
+        }
+        Ok(true)
+    }
+}
+
+impl Deref for StrColumn {
+    type Target = BytesColumn;
+
+    fn deref(&self) -> &Self::Target {
+        &self.0
+    }
+}

columnar/src/column/mod.rs

@@ -0,0 +1,161 @@
+mod dictionary_encoded;
+mod serialize;
+
+use std::fmt::Debug;
+use std::io::Write;
+use std::ops::{Deref, Range, RangeInclusive};
+use std::sync::Arc;
+
+use common::BinarySerializable;
+pub use dictionary_encoded::{BytesColumn, StrColumn};
+pub use serialize::{
+    open_column_bytes, open_column_str, open_column_u128, open_column_u64,
+    serialize_column_mappable_to_u128, serialize_column_mappable_to_u64,
+};
+
+use crate::column_index::ColumnIndex;
+use crate::column_values::monotonic_mapping::StrictlyMonotonicMappingToInternal;
+use crate::column_values::{monotonic_map_column, ColumnValues};
+use crate::{Cardinality, MonotonicallyMappableToU64, RowId};
+
+#[derive(Clone)]
+pub struct Column<T = u64> {
+    pub idx: ColumnIndex,
+    pub values: Arc<dyn ColumnValues<T>>,
+}
+
+impl<T: MonotonicallyMappableToU64> Column<T> {
+    pub fn to_u64_monotonic(self) -> Column<u64> {
+        let values = Arc::new(monotonic_map_column(
+            self.values,
+            StrictlyMonotonicMappingToInternal::<T>::new(),
+        ));
+        Column {
+            idx: self.idx,
+            values,
+        }
+    }
+}
+
+impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
+    #[inline]
+    pub fn get_cardinality(&self) -> Cardinality {
+        self.idx.get_cardinality()
+    }
+
+    pub fn num_docs(&self) -> RowId {
+        match &self.idx {
+            ColumnIndex::Empty { num_docs } => *num_docs,
+            ColumnIndex::Full => self.values.num_vals(),
+            ColumnIndex::Optional(optional_index) => optional_index.num_docs(),
+            ColumnIndex::Multivalued(col_index) => {
+                // The multivalued index contains all value start row_id,
+                // and one extra value at the end with the overall number of rows.
+                col_index.num_docs()
+            }
+        }
+    }
+
+    pub fn min_value(&self) -> T {
+        self.values.min_value()
+    }
+
+    pub fn max_value(&self) -> T {
+        self.values.max_value()
+    }
+
+    pub fn first(&self, row_id: RowId) -> Option<T> {
+        self.values_for_doc(row_id).next()
+    }
+
+    pub fn values_for_doc(&self, row_id: RowId) -> impl Iterator<Item = T> + '_ {
+        self.value_row_ids(row_id)
+            .map(|value_row_id: RowId| self.values.get_val(value_row_id))
+    }
+
+    /// Get the docids of values which are in the provided value range.
+    #[inline]
+    pub fn get_docids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        selected_docid_range: Range<u32>,
+        doc_ids: &mut Vec<u32>,
+    ) {
+        // convert passed docid range to row id range
+        let rowid_range = self.idx.docid_range_to_rowids(selected_docid_range.clone());
+
+        // Load rows
+        self.values
+            .get_row_ids_for_value_range(value_range, rowid_range, doc_ids);
+        // Convert rows to docids
+        self.idx
+            .select_batch_in_place(selected_docid_range.start, doc_ids);
+    }
+
+    /// Fils the output vector with the (possibly multiple values that are associated_with
+    /// `row_id`.
+    ///
+    /// This method clears the `output` vector.
+    pub fn fill_vals(&self, row_id: RowId, output: &mut Vec<T>) {
+        output.clear();
+        output.extend(self.values_for_doc(row_id));
+    }
+
+    pub fn first_or_default_col(self, default_value: T) -> Arc<dyn ColumnValues<T>> {
+        Arc::new(FirstValueWithDefault {
+            column: self,
+            default_value,
+        })
+    }
+}
+
+impl<T> Deref for Column<T> {
+    type Target = ColumnIndex;
+
+    fn deref(&self) -> &Self::Target {
+        &self.idx
+    }
+}
+
+impl BinarySerializable for Cardinality {
+    fn serialize<W: Write + ?Sized>(&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)
+    }
+}
+
+// TODO simplify or optimize
+struct FirstValueWithDefault<T: Copy> {
+    column: Column<T>,
+    default_value: T,
+}
+
+impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
+    for FirstValueWithDefault<T>
+{
+    fn get_val(&self, idx: u32) -> T {
+        self.column.first(idx).unwrap_or(self.default_value)
+    }
+
+    fn min_value(&self) -> T {
+        self.column.values.min_value()
+    }
+
+    fn max_value(&self) -> T {
+        self.column.values.max_value()
+    }
+
+    fn num_vals(&self) -> u32 {
+        match &self.column.idx {
+            ColumnIndex::Empty { .. } => 0u32,
+            ColumnIndex::Full => self.column.values.num_vals(),
+            ColumnIndex::Optional(optional_idx) => optional_idx.num_docs(),
+            ColumnIndex::Multivalued(multivalue_idx) => multivalue_idx.num_docs(),
+        }
+    }
+}

columnar/src/column/serialize.rs

@@ -0,0 +1,94 @@
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use sstable::Dictionary;
+
+use crate::column::{BytesColumn, Column};
+use crate::column_index::{serialize_column_index, SerializableColumnIndex};
+use crate::column_values::{
+    load_u64_based_column_values, serialize_column_values_u128, serialize_u64_based_column_values,
+    CodecType, MonotonicallyMappableToU128, MonotonicallyMappableToU64,
+};
+use crate::iterable::Iterable;
+use crate::StrColumn;
+
+pub fn serialize_column_mappable_to_u128<T: MonotonicallyMappableToU128>(
+    column_index: SerializableColumnIndex<'_>,
+    iterable: &dyn Iterable<T>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let column_index_num_bytes = serialize_column_index(column_index, output)?;
+    serialize_column_values_u128(iterable, output)?;
+    output.write_all(&column_index_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64>(
+    column_index: SerializableColumnIndex<'_>,
+    column_values: &impl Iterable<T>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    let column_index_num_bytes = serialize_column_index(column_index, output)?;
+    serialize_u64_based_column_values(
+        column_values,
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        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 = load_u64_based_column_values(column_values_data)?;
+    Ok(Column {
+        idx: column_index,
+        values: column_values,
+    })
+}
+
+pub fn open_column_u128<T: MonotonicallyMappableToU128>(
+    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_u128_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,
+    })
+}
+
+pub fn open_column_str(data: OwnedBytes) -> io::Result<StrColumn> {
+    let bytes_column = open_column_bytes(data)?;
+    Ok(StrColumn::wrap(bytes_column))
+}

columnar/src/column_index/merge/mod.rs

@@ -0,0 +1,136 @@
+mod shuffled;
+mod stacked;
+
+use shuffled::merge_column_index_shuffled;
+use stacked::merge_column_index_stacked;
+
+use crate::column_index::SerializableColumnIndex;
+use crate::{Cardinality, ColumnIndex, MergeRowOrder};
+
+// For simplification, we never have cardinality go down due to deletes.
+fn detect_cardinality(columns: &[Option<ColumnIndex>]) -> Cardinality {
+    columns
+        .iter()
+        .flatten()
+        .map(ColumnIndex::get_cardinality)
+        .max()
+        .unwrap_or(Cardinality::Full)
+}
+
+pub fn merge_column_index<'a>(
+    columns: &'a [Option<ColumnIndex>],
+    merge_row_order: &'a MergeRowOrder,
+) -> SerializableColumnIndex<'a> {
+    // For simplification, we do not try to detect whether the cardinality could be
+    // downgraded thanks to deletes.
+    let cardinality_after_merge = detect_cardinality(columns);
+    match merge_row_order {
+        MergeRowOrder::Stack(stack_merge_order) => {
+            merge_column_index_stacked(columns, cardinality_after_merge, stack_merge_order)
+        }
+        MergeRowOrder::Shuffled(complex_merge_order) => {
+            merge_column_index_shuffled(columns, cardinality_after_merge, complex_merge_order)
+        }
+    }
+}
+
+// TODO actually, the shuffled code path is a bit too general.
+// In practise, we do not really shuffle everything.
+// The merge order restricted to a specific column keeps the original row order.
+//
+// This may offer some optimization that we have not explored yet.
+
+#[cfg(test)]
+mod tests {
+    use crate::column_index::merge::detect_cardinality;
+    use crate::column_index::multivalued_index::MultiValueIndex;
+    use crate::column_index::{merge_column_index, OptionalIndex, SerializableColumnIndex};
+    use crate::{Cardinality, ColumnIndex, MergeRowOrder, RowAddr, RowId, ShuffleMergeOrder};
+
+    #[test]
+    fn test_detect_cardinality() {
+        assert_eq!(detect_cardinality(&[]), Cardinality::Full);
+        let optional_index: ColumnIndex = OptionalIndex::for_test(1, &[]).into();
+        let multivalued_index: ColumnIndex = MultiValueIndex::for_test(&[0, 1]).into();
+        assert_eq!(
+            detect_cardinality(&[Some(optional_index.clone()), None]),
+            Cardinality::Optional
+        );
+        assert_eq!(
+            detect_cardinality(&[Some(optional_index.clone()), Some(ColumnIndex::Full)]),
+            Cardinality::Optional
+        );
+        assert_eq!(
+            detect_cardinality(&[Some(multivalued_index.clone()), None]),
+            Cardinality::Multivalued
+        );
+        assert_eq!(
+            detect_cardinality(&[
+                Some(multivalued_index.clone()),
+                Some(optional_index.clone())
+            ]),
+            Cardinality::Multivalued
+        );
+        assert_eq!(
+            detect_cardinality(&[Some(optional_index), Some(multivalued_index)]),
+            Cardinality::Multivalued
+        );
+    }
+
+    #[test]
+    fn test_merge_index_multivalued_sorted() {
+        let column_indexes: Vec<Option<ColumnIndex>> =
+            vec![Some(MultiValueIndex::for_test(&[0, 2, 5]).into())];
+        let merge_row_order: MergeRowOrder = ShuffleMergeOrder::for_test(
+            &[2],
+            vec![
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 1u32,
+                },
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 0u32,
+                },
+            ],
+        )
+        .into();
+        let merged_column_index = merge_column_index(&column_indexes[..], &merge_row_order);
+        let SerializableColumnIndex::Multivalued(start_index_iterable) = merged_column_index
+        else { panic!("Excpected a multivalued index") };
+        let start_indexes: Vec<RowId> = start_index_iterable.boxed_iter().collect();
+        assert_eq!(&start_indexes, &[0, 3, 5]);
+    }
+
+    #[test]
+    fn test_merge_index_multivalued_sorted_several_segment() {
+        let column_indexes: Vec<Option<ColumnIndex>> = vec![
+            Some(MultiValueIndex::for_test(&[0, 2, 5]).into()),
+            None,
+            Some(MultiValueIndex::for_test(&[0, 1, 4]).into()),
+        ];
+        let merge_row_order: MergeRowOrder = ShuffleMergeOrder::for_test(
+            &[2, 0, 2],
+            vec![
+                RowAddr {
+                    segment_ord: 2u32,
+                    row_id: 1u32,
+                },
+                RowAddr {
+                    segment_ord: 0u32,
+                    row_id: 0u32,
+                },
+                RowAddr {
+                    segment_ord: 2u32,
+                    row_id: 0u32,
+                },
+            ],
+        )
+        .into();
+        let merged_column_index = merge_column_index(&column_indexes[..], &merge_row_order);
+        let SerializableColumnIndex::Multivalued(start_index_iterable) = merged_column_index
+        else { panic!("Excpected a multivalued index") };
+        let start_indexes: Vec<RowId> = start_index_iterable.boxed_iter().collect();
+        assert_eq!(&start_indexes, &[0, 3, 5, 6]);
+    }
+}

columnar/src/column_index/merge/shuffled.rs

@@ -0,0 +1,168 @@
+use std::iter;
+
+use crate::column_index::{SerializableColumnIndex, Set};
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, ShuffleMergeOrder};
+
+pub fn merge_column_index_shuffled<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    cardinality_after_merge: Cardinality,
+    shuffle_merge_order: &'a ShuffleMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => {
+            let non_null_row_ids =
+                merge_column_index_shuffled_optional(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Optional {
+                non_null_row_ids,
+                num_rows: shuffle_merge_order.num_rows(),
+            }
+        }
+        Cardinality::Multivalued => {
+            let multivalue_start_index =
+                merge_column_index_shuffled_multivalued(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Multivalued(multivalue_start_index)
+        }
+    }
+}
+
+/// Merge several column indexes into one, ordering rows according to the merge_order passed as
+/// argument. While it is true that the `merge_order` may imply deletes and hence could in theory a
+/// multivalued index into an optional one, this is not supported today for simplification.
+///
+/// In other words the column_indexes passed as argument may NOT be multivalued.
+fn merge_column_index_shuffled_optional<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledOptionalIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledOptionalIndex<'a> {
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+impl<'a> Iterable<u32> for ShuffledOptionalIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new(self.merge_order
+        .iter_new_to_old_row_addrs()
+        .enumerate()
+        .filter_map(|(new_row_id, old_row_addr)| {
+            let Some(column_index) = &self.column_indexes[old_row_addr.segment_ord as usize] else {
+                return None;
+            };
+            let row_id = new_row_id as u32;
+            if column_index.has_value(old_row_addr.row_id) {
+                Some(row_id)
+            } else {
+                None
+            }
+        }))
+    }
+}
+
+fn merge_column_index_shuffled_multivalued<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledMultivaluedIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledMultivaluedIndex<'a> {
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+fn iter_num_values<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> impl Iterator<Item = u32> + 'a {
+    merge_order.iter_new_to_old_row_addrs().map(|row_addr| {
+        let Some(column_index) = &column_indexes[row_addr.segment_ord as usize] else {
+                    // No values in the entire column. It surely means there are 0 values associated to this row.
+                    return 0u32;
+                };
+        match column_index {
+            ColumnIndex::Empty { .. } => 0u32,
+            ColumnIndex::Full => 1,
+            ColumnIndex::Optional(optional_index) => {
+                u32::from(optional_index.contains(row_addr.row_id))
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_addr.row_id).len() as u32
+            }
+        }
+    })
+}
+
+/// Transforms an iterator containing the number of vals per row (with `num_rows` elements)
+/// into a `start_offset` iterator starting at 0 and (with `num_rows + 1` element)
+fn integrate_num_vals(num_vals: impl Iterator<Item = u32>) -> impl Iterator<Item = RowId> {
+    iter::once(0u32).chain(num_vals.scan(0, |state, num_vals| {
+        *state += num_vals;
+        Some(*state)
+    }))
+}
+
+impl<'a> Iterable<u32> for ShuffledMultivaluedIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        let num_vals_per_row = iter_num_values(self.column_indexes, self.merge_order);
+        Box::new(integrate_num_vals(num_vals_per_row))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_index::OptionalIndex;
+    use crate::RowAddr;
+
+    #[test]
+    fn test_integrate_num_vals_empty() {
+        assert!(integrate_num_vals(iter::empty()).eq(iter::once(0)));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_one_el() {
+        assert!(integrate_num_vals(iter::once(10)).eq([0, 10].into_iter()));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_several() {
+        assert!(integrate_num_vals([3, 0, 10, 20].into_iter()).eq([0, 3, 3, 13, 33].into_iter()));
+    }
+
+    #[test]
+    fn test_merge_column_index_optional_shuffle() {
+        let optional_index: ColumnIndex = OptionalIndex::for_test(2, &[0]).into();
+        let column_indexes = vec![Some(optional_index), Some(ColumnIndex::Full)];
+        let row_addrs = vec![
+            RowAddr {
+                segment_ord: 0u32,
+                row_id: 1u32,
+            },
+            RowAddr {
+                segment_ord: 1u32,
+                row_id: 0u32,
+            },
+        ];
+        let shuffle_merge_order = ShuffleMergeOrder::for_test(&[2, 1], row_addrs);
+        let serializable_index = merge_column_index_shuffled(
+            &column_indexes[..],
+            Cardinality::Optional,
+            &shuffle_merge_order,
+        );
+        let SerializableColumnIndex::Optional { non_null_row_ids, num_rows } = serializable_index else { panic!() };
+        assert_eq!(num_rows, 2);
+        let non_null_rows: Vec<RowId> = non_null_row_ids.boxed_iter().collect();
+        assert_eq!(&non_null_rows, &[1]);
+    }
+}

columnar/src/column_index/merge/stacked.rs

@@ -0,0 +1,156 @@
+use std::iter;
+
+use crate::column_index::{SerializableColumnIndex, Set};
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, StackMergeOrder};
+
+/// Simple case:
+/// The new mapping just consists in stacking the different column indexes.
+///
+/// There are no sort nor deletes involved.
+pub fn merge_column_index_stacked<'a>(
+    columns: &'a [Option<ColumnIndex>],
+    cardinality_after_merge: Cardinality,
+    stack_merge_order: &'a StackMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => SerializableColumnIndex::Optional {
+            non_null_row_ids: Box::new(StackedOptionalIndex {
+                columns,
+                stack_merge_order,
+            }),
+            num_rows: stack_merge_order.num_rows(),
+        },
+        Cardinality::Multivalued => {
+            let stacked_multivalued_index = StackedMultivaluedIndex {
+                columns,
+                stack_merge_order,
+            };
+            SerializableColumnIndex::Multivalued(Box::new(stacked_multivalued_index))
+        }
+    }
+}
+
+struct StackedOptionalIndex<'a> {
+    columns: &'a [Option<ColumnIndex>],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+impl<'a> Iterable<RowId> for StackedOptionalIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(
+            self.columns
+                .iter()
+                .enumerate()
+                .flat_map(|(columnar_id, column_index_opt)| {
+                    let columnar_row_range = self.stack_merge_order.columnar_range(columnar_id);
+                    let rows_it: Box<dyn Iterator<Item = RowId>> = match column_index_opt {
+                        Some(ColumnIndex::Full) => Box::new(columnar_row_range),
+                        Some(ColumnIndex::Optional(optional_index)) => Box::new(
+                            optional_index
+                                .iter_rows()
+                                .map(move |row_id: RowId| columnar_row_range.start + row_id),
+                        ),
+                        Some(ColumnIndex::Multivalued(_)) => {
+                            panic!("No multivalued index is allowed when stacking column index");
+                        }
+                        None | Some(ColumnIndex::Empty { .. }) => Box::new(std::iter::empty()),
+                    };
+                    rows_it
+                }),
+        )
+    }
+}
+
+#[derive(Clone, Copy)]
+struct StackedMultivaluedIndex<'a> {
+    columns: &'a [Option<ColumnIndex>],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+fn convert_column_opt_to_multivalued_index<'a>(
+    column_index_opt: Option<&'a ColumnIndex>,
+    num_rows: RowId,
+) -> Box<dyn Iterator<Item = RowId> + 'a> {
+    match column_index_opt {
+        None | Some(ColumnIndex::Empty { .. }) => {
+            Box::new(iter::repeat(0u32).take(num_rows as usize + 1))
+        }
+        Some(ColumnIndex::Full) => Box::new(0..num_rows + 1),
+        Some(ColumnIndex::Optional(optional_index)) => {
+            Box::new(
+                (0..num_rows)
+                    // TODO optimize
+                    .map(|row_id| optional_index.rank(row_id))
+                    .chain(std::iter::once(optional_index.num_non_nulls())),
+            )
+        }
+        Some(ColumnIndex::Multivalued(multivalued_index)) => {
+            multivalued_index.start_index_column.iter()
+        }
+    }
+}
+
+impl<'a> Iterable<RowId> for StackedMultivaluedIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + '_> {
+        let multivalued_indexes =
+            self.columns
+                .iter()
+                .map(Option::as_ref)
+                .enumerate()
+                .map(|(columnar_id, column_opt)| {
+                    let num_rows =
+                        self.stack_merge_order.columnar_range(columnar_id).len() as RowId;
+                    convert_column_opt_to_multivalued_index(column_opt, num_rows)
+                });
+        stack_multivalued_indexes(multivalued_indexes)
+    }
+}
+
+// Refactor me
+fn stack_multivalued_indexes<'a>(
+    mut multivalued_indexes: impl Iterator<Item = Box<dyn Iterator<Item = RowId> + 'a>> + 'a,
+) -> Box<dyn Iterator<Item = RowId> + 'a> {
+    let mut offset = 0;
+    let mut last_row_id = 0;
+    let mut current_it = multivalued_indexes.next();
+    Box::new(std::iter::from_fn(move || loop {
+        let Some(multivalued_index) = current_it.as_mut() else {
+            return None;
+        };
+        if let Some(row_id) = multivalued_index.next() {
+            last_row_id = offset + row_id;
+            return Some(last_row_id);
+        }
+        offset = last_row_id;
+        loop {
+            current_it = multivalued_indexes.next();
+            if current_it.as_mut()?.next().is_some() {
+                break;
+            }
+        }
+    }))
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::RowId;
+
+    fn it<'a>(row_ids: &'a [RowId]) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(row_ids.iter().copied())
+    }
+
+    #[test]
+    fn test_stack() {
+        let columns = [
+            it(&[0u32, 0u32]),
+            it(&[0u32, 1u32, 1u32, 4u32]),
+            it(&[0u32, 3u32, 5u32]),
+            it(&[0u32, 4u32]),
+        ]
+        .into_iter();
+        let start_offsets: Vec<RowId> = super::stack_multivalued_indexes(columns).collect();
+        assert_eq!(start_offsets, &[0, 0, 1, 1, 4, 7, 9, 13]);
+    }
+}

columnar/src/column_index/mod.rs

@@ -0,0 +1,115 @@
+mod merge;
+mod multivalued_index;
+mod optional_index;
+mod serialize;
+
+use std::ops::Range;
+
+pub use merge::merge_column_index;
+pub use optional_index::{OptionalIndex, Set};
+pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
+
+use crate::column_index::multivalued_index::MultiValueIndex;
+use crate::{Cardinality, DocId, RowId};
+
+#[derive(Clone)]
+pub enum ColumnIndex {
+    Empty {
+        num_docs: u32,
+    },
+    Full,
+    Optional(OptionalIndex),
+    /// In addition, at index num_rows, an extra value is added
+    /// containing the overal number of values.
+    Multivalued(MultiValueIndex),
+}
+
+impl From<OptionalIndex> for ColumnIndex {
+    fn from(optional_index: OptionalIndex) -> ColumnIndex {
+        ColumnIndex::Optional(optional_index)
+    }
+}
+
+impl From<MultiValueIndex> for ColumnIndex {
+    fn from(multi_value_index: MultiValueIndex) -> ColumnIndex {
+        ColumnIndex::Multivalued(multi_value_index)
+    }
+}
+
+impl ColumnIndex {
+    #[inline]
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            ColumnIndex::Empty { .. } => Cardinality::Optional,
+            ColumnIndex::Full => Cardinality::Full,
+            ColumnIndex::Optional(_) => Cardinality::Optional,
+            ColumnIndex::Multivalued(_) => Cardinality::Multivalued,
+        }
+    }
+
+    /// Returns true if and only if there are at least one value associated to the row.
+    pub fn has_value(&self, doc_id: DocId) -> bool {
+        match self {
+            ColumnIndex::Empty { .. } => false,
+            ColumnIndex::Full => true,
+            ColumnIndex::Optional(optional_index) => optional_index.contains(doc_id),
+            ColumnIndex::Multivalued(multivalued_index) => {
+                !multivalued_index.range(doc_id).is_empty()
+            }
+        }
+    }
+
+    pub fn value_row_ids(&self, doc_id: DocId) -> Range<RowId> {
+        match self {
+            ColumnIndex::Empty { .. } => 0..0,
+            ColumnIndex::Full => doc_id..doc_id + 1,
+            ColumnIndex::Optional(optional_index) => {
+                if let Some(val) = optional_index.rank_if_exists(doc_id) {
+                    val..val + 1
+                } else {
+                    0..0
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(doc_id),
+        }
+    }
+
+    pub fn docid_range_to_rowids(&self, doc_id: Range<DocId>) -> Range<RowId> {
+        match self {
+            ColumnIndex::Empty { .. } => 0..0,
+            ColumnIndex::Full => doc_id,
+            ColumnIndex::Optional(optional_index) => {
+                let row_start = optional_index.rank(doc_id.start);
+                let row_end = optional_index.rank(doc_id.end);
+                row_start..row_end
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                let end_docid = doc_id.end.min(multivalued_index.num_docs() - 1) + 1;
+                let start_docid = doc_id.start.min(end_docid);
+
+                let row_start = multivalued_index.start_index_column.get_val(start_docid);
+                let row_end = multivalued_index.start_index_column.get_val(end_docid);
+
+                row_start..row_end
+            }
+        }
+    }
+
+    pub fn select_batch_in_place(&self, doc_id_start: DocId, rank_ids: &mut Vec<RowId>) {
+        match self {
+            ColumnIndex::Empty { .. } => {
+                rank_ids.clear();
+            }
+            ColumnIndex::Full => {
+                // No need to do anything:
+                // value_idx and row_idx are the same.
+            }
+            ColumnIndex::Optional(optional_index) => {
+                optional_index.select_batch(&mut rank_ids[..]);
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.select_batch_in_place(doc_id_start, rank_ids)
+            }
+        }
+    }
+}

columnar/src/column_index/multivalued_index.rs

@@ -0,0 +1,141 @@
+use std::io;
+use std::io::Write;
+use std::ops::Range;
+use std::sync::Arc;
+
+use common::OwnedBytes;
+
+use crate::column_values::{
+    load_u64_based_column_values, serialize_u64_based_column_values, CodecType, ColumnValues,
+};
+use crate::iterable::Iterable;
+use crate::{DocId, RowId};
+
+pub fn serialize_multivalued_index(
+    multivalued_index: &dyn Iterable<RowId>,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    serialize_u64_based_column_values(
+        multivalued_index,
+        &[CodecType::Bitpacked, CodecType::Linear],
+        output,
+    )?;
+    Ok(())
+}
+
+pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex> {
+    let start_index_column: Arc<dyn ColumnValues<RowId>> = load_u64_based_column_values(bytes)?;
+    Ok(MultiValueIndex { start_index_column })
+}
+
+#[derive(Clone)]
+/// Index to resolve value range for given doc_id.
+/// Starts at 0.
+pub struct MultiValueIndex {
+    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
+}
+
+impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {
+    fn from(start_index_column: Arc<dyn ColumnValues<RowId>>) -> Self {
+        MultiValueIndex { start_index_column }
+    }
+}
+
+impl MultiValueIndex {
+    pub fn for_test(start_offsets: &[RowId]) -> MultiValueIndex {
+        let mut buffer = Vec::new();
+        serialize_multivalued_index(&start_offsets, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_multivalued_index(bytes).unwrap()
+    }
+
+    /// Returns `[start, end)`, such that the values associated with
+    /// the given document are `start..end`.
+    #[inline]
+    pub(crate) fn range(&self, doc_id: DocId) -> Range<RowId> {
+        let start = self.start_index_column.get_val(doc_id);
+        let end = self.start_index_column.get_val(doc_id + 1);
+        start..end
+    }
+
+    /// Returns the number of documents in the index.
+    #[inline]
+    pub fn num_docs(&self) -> u32 {
+        self.start_index_column.num_vals() - 1
+    }
+
+    /// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
+    /// docids. Positions are converted inplace to docids.
+    ///
+    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
+    /// index.
+    ///
+    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
+    /// increasing positions.
+    ///
+    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
+    /// match a docid to its value position.
+    #[allow(clippy::bool_to_int_with_if)]
+    pub(crate) fn select_batch_in_place(&self, docid_start: DocId, ranks: &mut Vec<u32>) {
+        if ranks.is_empty() {
+            return;
+        }
+        let mut cur_doc = docid_start;
+        let mut last_doc = None;
+
+        assert!(self.start_index_column.get_val(docid_start) <= ranks[0]);
+
+        let mut write_doc_pos = 0;
+        for i in 0..ranks.len() {
+            let pos = ranks[i];
+            loop {
+                let end = self.start_index_column.get_val(cur_doc + 1);
+                if end > pos {
+                    ranks[write_doc_pos] = cur_doc;
+                    write_doc_pos += if last_doc == Some(cur_doc) { 0 } else { 1 };
+                    last_doc = Some(cur_doc);
+                    break;
+                }
+                cur_doc += 1;
+            }
+        }
+        ranks.truncate(write_doc_pos);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ops::Range;
+    use std::sync::Arc;
+
+    use super::MultiValueIndex;
+    use crate::column_values::IterColumn;
+    use crate::{ColumnValues, RowId};
+
+    fn index_to_pos_helper(
+        index: &MultiValueIndex,
+        doc_id_range: Range<u32>,
+        positions: &[u32],
+    ) -> Vec<u32> {
+        let mut positions = positions.to_vec();
+        index.select_batch_in_place(doc_id_range.start, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_positions_to_docid() {
+        let offsets: Vec<RowId> = vec![0, 10, 12, 15, 22, 23]; // docid values are [0..10, 10..12, 12..15, etc.]
+        let column: Arc<dyn ColumnValues<RowId>> = Arc::new(IterColumn::from(offsets.into_iter()));
+        let index = MultiValueIndex::from(column);
+        assert_eq!(index.num_docs(), 5);
+        let positions = &[10u32, 11, 15, 20, 21, 22];
+        assert_eq!(index_to_pos_helper(&index, 0..5, positions), vec![1, 3, 4]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, positions), vec![1, 3, 4]);
+        assert_eq!(index_to_pos_helper(&index, 0..5, &[9]), vec![0]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[10]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[11]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14, 15]), vec![2, 3]);
+    }
+}

columnar/src/column_index/optional_index/mod.rs

@@ -0,0 +1,515 @@
+use std::io::{self, Write};
+use std::sync::Arc;
+
+mod set;
+mod set_block;
+
+use common::{BinarySerializable, OwnedBytes, VInt};
+pub use set::{SelectCursor, Set, SetCodec};
+use set_block::{
+    DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
+};
+
+use crate::iterable::Iterable;
+use crate::{DocId, 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]>,
+}
+
+/// 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,
+    }
+}
+
+enum BlockSelectCursor<'a> {
+    Dense(<DenseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+    Sparse(<SparseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+}
+
+impl<'a> BlockSelectCursor<'a> {
+    fn select(&mut self, rank: u16) -> u16 {
+        match self {
+            BlockSelectCursor::Dense(dense_select_cursor) => dense_select_cursor.select(rank),
+            BlockSelectCursor::Sparse(sparse_select_cursor) => sparse_select_cursor.select(rank),
+        }
+    }
+}
+pub struct OptionalIndexSelectCursor<'a> {
+    current_block_cursor: BlockSelectCursor<'a>,
+    current_block_id: u16,
+    // The current block is guaranteed to contain ranks < end_rank.
+    current_block_end_rank: RowId,
+    optional_index: &'a OptionalIndex,
+    block_doc_idx_start: RowId,
+    num_null_rows_before_block: RowId,
+}
+
+impl<'a> OptionalIndexSelectCursor<'a> {
+    fn search_and_load_block(&mut self, rank: RowId) {
+        if rank < self.current_block_end_rank {
+            // we are already in the right block
+            return;
+        }
+        self.current_block_id = self.optional_index.find_block(rank, self.current_block_id);
+        self.current_block_end_rank = self
+            .optional_index
+            .block_metas
+            .get(self.current_block_id as usize + 1)
+            .map(|block_meta| block_meta.non_null_rows_before_block)
+            .unwrap_or(u32::MAX);
+        self.block_doc_idx_start = (self.current_block_id as u32) * ELEMENTS_PER_BLOCK;
+        let block_meta = self.optional_index.block_metas[self.current_block_id as usize];
+        self.num_null_rows_before_block = block_meta.non_null_rows_before_block;
+        let block: Block<'_> = self.optional_index.block(block_meta);
+        self.current_block_cursor = match block {
+            Block::Dense(dense_block) => BlockSelectCursor::Dense(dense_block.select_cursor()),
+            Block::Sparse(sparse_block) => BlockSelectCursor::Sparse(sparse_block.select_cursor()),
+        };
+    }
+}
+
+impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor<'a> {
+    fn select(&mut self, rank: RowId) -> RowId {
+        self.search_and_load_block(rank);
+        let index_in_block = (rank - self.num_null_rows_before_block) as u16;
+        self.current_block_cursor.select(index_in_block) as RowId + self.block_doc_idx_start
+    }
+}
+
+impl Set<RowId> for OptionalIndex {
+    type SelectCursor<'b> = OptionalIndexSelectCursor<'b> where Self: 'b;
+    // 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(&self, doc_id: DocId) -> RowId {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(doc_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(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank(in_block_row_id),
+        } as u32;
+        block_meta.non_null_rows_before_block + block_offset_row_id
+    }
+
+    #[inline]
+    fn rank_if_exists(&self, doc_id: DocId) -> Option<RowId> {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(doc_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 as u32) * 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_cursor(&self) -> OptionalIndexSelectCursor<'_> {
+        OptionalIndexSelectCursor {
+            current_block_cursor: BlockSelectCursor::Sparse(
+                SparseBlockCodec::open(b"").select_cursor(),
+            ),
+            current_block_id: 0u16,
+            current_block_end_rank: 0u32, //< this is sufficient to force the first load
+            optional_index: self,
+            block_doc_idx_start: 0u32,
+            num_null_rows_before_block: 0u32,
+        }
+    }
+}
+
+impl OptionalIndex {
+    pub fn for_test(num_rows: RowId, row_ids: &[RowId]) -> OptionalIndex {
+        assert!(row_ids
+            .last()
+            .copied()
+            .map(|last_row_id| last_row_id < num_rows)
+            .unwrap_or(true));
+        let mut buffer = Vec::new();
+        serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_optional_index(bytes).unwrap()
+    }
+
+    pub fn num_docs(&self) -> RowId {
+        self.num_rows
+    }
+
+    pub fn num_non_nulls(&self) -> RowId {
+        self.num_non_null_rows
+    }
+
+    pub fn iter_rows(&self) -> impl Iterator<Item = RowId> + '_ {
+        // TODO optimize
+        let mut select_batch = self.select_cursor();
+        (0..self.num_non_null_rows).map(move |rank| select_batch.select(rank))
+    }
+    pub fn select_batch(&self, ranks: &mut [RowId]) {
+        let mut select_cursor = self.select_cursor();
+        for rank in ranks.iter_mut() {
+            *rank = select_cursor.select(*rank);
+        }
+    }
+
+    #[inline]
+    fn block(&self, block_meta: BlockMeta) -> Block<'_> {
+        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: u16) -> u16 {
+        for block_pos in start_block_pos..self.block_metas.len() as u16 {
+            let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
+            if offset > dense_idx {
+                return block_pos - 1u16;
+            }
+        }
+        self.block_metas.len() as u16 - 1u16
+    }
+
+    // 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 + ?Sized>(&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<W: io::Write>(
+    non_null_rows: &dyn Iterable<RowId>,
+    num_rows: RowId,
+    output: &mut W,
+) -> io::Result<()> {
+    VInt(num_rows as u64).serialize(output)?;
+
+    let mut rows_it = non_null_rows.boxed_iter();
+    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
+}
+
+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 + 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;
+    }
+    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);
+    let optional_index = OptionalIndex {
+        num_rows,
+        num_non_null_rows,
+        block_data,
+        block_metas: block_metas.into(),
+    };
+    Ok(optional_index)
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_index/optional_index/set.rs

@@ -0,0 +1,47 @@
+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(data: &[u8]) -> Self::Reader<'_>;
+}
+
+/// Stateful object that makes it possible to compute several select in a row,
+/// provided the rank passed as argument are increasing.
+pub trait SelectCursor<T> {
+    // May panic if rank is greater than the number of elements in the Set,
+    // or if rank is < than value provided in the previous call.
+    fn select(&mut self, rank: T) -> T;
+}
+
+pub trait Set<T> {
+    type SelectCursor<'b>: SelectCursor<T>
+    where Self: 'b;
+
+    /// Returns true if the elements is contained in the Set
+    fn contains(&self, el: T) -> bool;
+
+    /// Returns the number of rows in the set that are < `el`
+    fn rank(&self, el: T) -> T;
+
+    /// If the set contains `el` returns the element rank.
+    /// 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;
+
+    /// Creates a brand new select cursor.
+    fn select_cursor(&self) -> Self::SelectCursor<'_>;
+}

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

@@ -0,0 +1,278 @@
+use std::convert::TryInto;
+use std::io::{self, Write};
+
+use common::BinarySerializable;
+
+use crate::column_index::optional_index::{SelectCursor, 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 / 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(data: &[u8]) -> Self::Reader<'_> {
+        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]);
+
+pub struct DenseBlockSelectCursor<'a> {
+    block_id: u16,
+    dense_block: DenseBlock<'a>,
+}
+
+impl<'a> SelectCursor<u16> for DenseBlockSelectCursor<'a> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        self.block_id = self
+            .dense_block
+            .find_miniblock_containing_rank(rank, self.block_id)
+            .unwrap();
+        let index_block = self.dense_block.mini_block(self.block_id);
+        let in_block_rank = rank - index_block.rank;
+        self.block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+    }
+}
+
+impl<'a> Set<u16> for DenseBlock<'a> {
+    type SelectCursor<'b> = DenseBlockSelectCursor<'a> where Self: 'b;
+
+    #[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 rank(&self, el: u16) -> 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);
+        index_block.rank + ones_in_block
+    }
+
+    #[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)
+    }
+
+    #[inline(always)]
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
+        DenseBlockSelectCursor {
+            block_id: 0,
+            dense_block: *self,
+        }
+    }
+}
+
+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,
+            };
+            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/mod.rs

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

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

@@ -0,0 +1,111 @@
+use crate::column_index::optional_index::{SelectCursor, 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(data: &[u8]) -> Self::Reader<'_> {
+        SparseBlock(data)
+    }
+}
+
+#[derive(Copy, Clone)]
+pub struct SparseBlock<'a>(&'a [u8]);
+
+impl<'a> SelectCursor<u16> for SparseBlock<'a> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        <SparseBlock<'a> as Set<u16>>::select(self, rank)
+    }
+}
+
+impl<'a> Set<u16> for SparseBlock<'a> {
+    type SelectCursor<'b> = Self where Self: 'b;
+
+    #[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 rank(&self, el: u16) -> u16 {
+        self.binary_search(el).unwrap_or_else(|el| el)
+    }
+
+    #[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())
+    }
+
+    #[inline(always)]
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
+        *self
+    }
+}
+
+#[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)
+    }
+}

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

@@ -0,0 +1,109 @@
+use std::collections::HashMap;
+
+use crate::column_index::optional_index::set_block::dense::DENSE_BLOCK_NUM_BYTES;
+use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
+use crate::column_index::optional_index::{SelectCursor, 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());
+        assert_eq!(
+            tested_set.rank(val),
+            vals.iter().cloned().take_while(|v| *v < val).count() as u16
+        );
+    }
+    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! {
+    #![proptest_config(ProptestConfig::with_cases(1))]
+    #[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));
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(select_cursor.select(0), 1);
+    assert_eq!(select_cursor.select(1), 3);
+    assert_eq!(select_cursor.select(2), 17);
+}
+
+#[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));
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(SelectCursor::select(&mut select_cursor, 0), 1);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 1), 3);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 2), 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 mut select_cursor = tested_set.select_cursor();
+    for i in 0..150 {
+        assert_eq!(i, select_cursor.select(i));
+    }
+}

columnar/src/column_index/optional_index/tests.rs

@@ -0,0 +1,371 @@
+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, 100, &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 select_cursor = null_index.select_cursor();
+    for (rank, el) in ranks.iter().copied().zip(els.iter().copied()) {
+        assert_eq!(select_cursor.select(rank), el);
+    }
+}
+
+#[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> Iterable<RowId> for &'a [bool] {
+    fn boxed_iter(&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, data.len() as RowId, &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 mut select_iter = null_index.select_cursor();
+    for i in 0..orig_idx_with_value.len() {
+        assert_eq!(select_iter.select(i as u32), orig_idx_with_value[i]);
+    }
+
+    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 optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    let mut select_cursor = optional_index.select_cursor();
+    assert_eq!(select_cursor.select(0), 0);
+    assert_eq!(select_cursor.select(1), 2);
+}
+
+#[test]
+fn test_optional_index_translate() {
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert_eq!(optional_index.rank_if_exists(0), Some(0));
+    assert_eq!(optional_index.rank_if_exists(2), Some(1));
+}
+
+#[test]
+fn test_optional_index_small() {
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert!(optional_index.contains(0));
+    assert!(!optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+}
+
+#[test]
+fn test_optional_index_large() {
+    let row_ids = &[ELEMENTS_PER_BLOCK, ELEMENTS_PER_BLOCK + 1];
+    let optional_index = OptionalIndex::for_test(ELEMENTS_PER_BLOCK + 2, row_ids);
+    assert!(!optional_index.contains(0));
+    assert!(!optional_index.contains(100));
+    assert!(!optional_index.contains(ELEMENTS_PER_BLOCK - 1));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK + 1));
+}
+
+fn test_optional_index_iter_aux(row_ids: &[RowId], num_rows: RowId) {
+    let optional_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(optional_index.num_docs(), num_rows);
+    assert!(optional_index.iter_rows().eq(row_ids.iter().copied()));
+}
+
+#[test]
+fn test_optional_index_iter_empty() {
+    test_optional_index_iter_aux(&[], 0u32);
+}
+
+fn test_optional_index_rank_aux(row_ids: &[RowId]) {
+    let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
+    let null_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(null_index.num_docs(), num_rows);
+    for (row_id, row_val) in row_ids.iter().copied().enumerate() {
+        assert_eq!(null_index.rank(row_val), row_id as u32);
+        assert_eq!(null_index.rank_if_exists(row_val), Some(row_id as u32));
+        if row_val > 0 && !null_index.contains(&row_val - 1) {
+            assert_eq!(null_index.rank(row_val - 1), row_id as u32);
+        }
+        assert_eq!(null_index.rank(row_val + 1), row_id as u32 + 1);
+    }
+}
+
+#[test]
+fn test_optional_index_rank() {
+    test_optional_index_rank_aux(&[1u32]);
+    test_optional_index_rank_aux(&[0u32, 1u32]);
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_rank_aux(&block);
+}
+
+#[test]
+fn test_optional_index_iter_empty_one() {
+    test_optional_index_iter_aux(&[1], 2u32);
+    test_optional_index_iter_aux(&[100_000], 200_000u32);
+}
+
+#[test]
+fn test_optional_index_iter_dense_block() {
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_iter_aux(&block, 3 * BLOCK_SIZE);
+}
+
+#[test]
+fn test_optional_index_for_tests() {
+    let optional_index = OptionalIndex::for_test(4, &[1, 2]);
+    assert!(!optional_index.contains(0));
+    assert!(optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+    assert_eq!(optional_index.num_docs(), 4);
+}
+
+#[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<RowId> = (0..TOTAL_NUM_VALUES)
+            .map(|_| rng.gen_bool(fill_ratio))
+            .enumerate()
+            .filter(|(pos, val)| *val)
+            .map(|(pos, _)| pos as RowId)
+            .collect();
+        serialize_optional_index(&&vals[..], TOTAL_NUM_VALUES, &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(|| {
+            output.copy_from_slice(&idxs[..]);
+            codec.select_batch(&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 +1,77 @@
+use std::io;
+use std::io::Write;
+
+use common::{CountingWriter, OwnedBytes};
+
+use crate::column_index::multivalued_index::serialize_multivalued_index;
+use crate::column_index::optional_index::serialize_optional_index;
+use crate::column_index::ColumnIndex;
+use crate::iterable::Iterable;
+use crate::{Cardinality, RowId};
+
+pub enum SerializableColumnIndex<'a> {
+    Full,
+    Optional {
+        non_null_row_ids: Box<dyn Iterable<RowId> + 'a>,
+        num_rows: RowId,
+    },
+    // TODO remove the Arc<dyn> apart from serialization this is not
+    // dynamic at all.
+    Multivalued(Box<dyn Iterable<RowId> + 'a>),
+}
+
+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<u32> {
+    let mut output = CountingWriter::wrap(output);
+    let cardinality = column_index.get_cardinality().to_code();
+    output.write_all(&[cardinality])?;
+    match column_index {
+        SerializableColumnIndex::Full => {}
+        SerializableColumnIndex::Optional {
+            non_null_row_ids,
+            num_rows,
+        } => serialize_optional_index(non_null_row_ids.as_ref(), num_rows, &mut output)?,
+        SerializableColumnIndex::Multivalued(multivalued_index) => {
+            serialize_multivalued_index(&*multivalued_index, &mut output)?
+        }
+    }
+    let column_index_num_bytes = output.written_bytes() as u32;
+    Ok(column_index_num_bytes)
+}
+
+pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex> {
+    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 multivalue_index = super::multivalued_index::open_multivalued_index(bytes)?;
+            Ok(ColumnIndex::Multivalued(multivalue_index))
+        }
+    }
+}
+
+// TODO unit tests

columnar/src/column_values/bench.rs

@@ -0,0 +1,135 @@
+use std::sync::Arc;
+
+use common::OwnedBytes;
+use rand::rngs::StdRng;
+use rand::{Rng, SeedableRng};
+use test::{self, Bencher};
+
+use super::*;
+use crate::column_values::u64_based::*;
+
+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
+}
+
+fn compute_stats(vals: impl Iterator<Item = u64>) -> ColumnStats {
+    let mut stats_collector = StatsCollector::default();
+    for val in vals {
+        stats_collector.collect(val);
+    }
+    stats_collector.stats()
+}
+
+#[inline(never)]
+fn value_iter() -> impl Iterator<Item = u64> {
+    0..20_000
+}
+fn get_reader_for_bench<Codec: ColumnCodec>(data: &[u64]) -> Codec::ColumnValues {
+    let mut bytes = Vec::new();
+    let stats = compute_stats(data.iter().cloned());
+    let mut codec_serializer = Codec::estimator();
+    for val in data {
+        codec_serializer.collect(*val);
+    }
+    codec_serializer.serialize(&stats, Box::new(data.iter().copied()).as_mut(), &mut bytes);
+
+    Codec::load(OwnedBytes::new(bytes)).unwrap()
+}
+fn bench_get<Codec: ColumnCodec>(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: ColumnCodec>(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: ColumnCodec>(b: &mut Bencher, data: &[u64]) {
+    let stats = compute_stats(data.iter().cloned());
+
+    let mut bytes = Vec::new();
+    b.iter(|| {
+        bytes.clear();
+        let mut codec_serializer = Codec::estimator();
+        for val in data.iter().take(1024) {
+            codec_serializer.collect(*val);
+        }
+
+        codec_serializer.serialize(&stats, Box::new(data.iter().copied()).as_mut(), &mut bytes)
+    });
+}
+
+#[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/merge.rs

@@ -0,0 +1,41 @@
+use std::fmt::Debug;
+use std::sync::Arc;
+
+use crate::iterable::Iterable;
+use crate::{ColumnIndex, ColumnValues, MergeRowOrder};
+
+pub(crate) struct MergedColumnValues<'a, T> {
+    pub(crate) column_indexes: &'a [Option<ColumnIndex>],
+    pub(crate) column_values: &'a [Option<Arc<dyn ColumnValues<T>>>],
+    pub(crate) merge_row_order: &'a MergeRowOrder,
+}
+
+impl<'a, T: Copy + PartialOrd + Debug> Iterable<T> for MergedColumnValues<'a, T> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        match self.merge_row_order {
+            MergeRowOrder::Stack(_) => Box::new(
+                self.column_values
+                    .iter()
+                    .flatten()
+                    .flat_map(|column_value| column_value.iter()),
+            ),
+            MergeRowOrder::Shuffled(shuffle_merge_order) => Box::new(
+                shuffle_merge_order
+                    .iter_new_to_old_row_addrs()
+                    .flat_map(|row_addr| {
+                        let column_index =
+                            self.column_indexes[row_addr.segment_ord as usize].as_ref()?;
+                        let column_values =
+                            self.column_values[row_addr.segment_ord as usize].as_ref()?;
+                        let value_range = column_index.value_row_ids(row_addr.row_id);
+                        Some((value_range, column_values))
+                    })
+                    .flat_map(|(value_range, column_values)| {
+                        value_range
+                            .into_iter()
+                            .map(|val| column_values.get_val(val))
+                    }),
+            ),
+        }
+    }
+}

columnar/src/column_values/mod.rs

@@ -0,0 +1,220 @@
+#![warn(missing_docs)]
+
+//! # `fastfield_codecs`
+//!
+//! - Columnar storage of data for tantivy [`Column`].
+//! - Encode data in different codecs.
+//! - Monotonically map values to u64/u128
+
+use std::fmt::Debug;
+use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
+
+pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
+pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
+
+mod merge;
+pub(crate) mod monotonic_mapping;
+pub(crate) mod monotonic_mapping_u128;
+mod stats;
+mod u128_based;
+mod u64_based;
+mod vec_column;
+
+mod monotonic_column;
+
+pub(crate) use merge::MergedColumnValues;
+pub use stats::ColumnStats;
+pub use u128_based::{open_u128_mapped, serialize_column_values_u128};
+pub use u64_based::{
+    load_u64_based_column_values, serialize_and_load_u64_based_column_values,
+    serialize_u64_based_column_values, CodecType, ALL_U64_CODEC_TYPES,
+};
+pub use vec_column::VecColumn;
+
+pub use self::monotonic_column::monotonic_map_column;
+use crate::RowId;
+
+/// `ColumnValues` provides access to a dense field column.
+///
+/// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
+///
+/// Any methods with a default and specialized implementation need to be called in the
+/// wrappers that implement the trait: Arc and MonotonicMappingColumn
+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;
+
+    /// Allows to push down multiple fetch calls, to avoid dynamic dispatch overhead.
+    ///
+    /// idx and output should have the same length
+    ///
+    /// # Panics
+    ///
+    /// May panic if `idx` is greater than the column length.
+    fn get_vals(&self, idx: &[u32], output: &mut [T]) {
+        assert!(idx.len() == output.len());
+        for (out, idx) in output.iter_mut().zip(idx.iter()) {
+            *out = self.get_val(*idx as u32);
+        }
+    }
+
+    /// 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(always)]
+    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 row ids of values which are in the provided value range.
+    ///
+    /// Note that position == docid for single value fast fields
+    #[inline(always)]
+    fn get_row_ids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        row_id_range: Range<RowId>,
+        row_id_hits: &mut Vec<RowId>,
+    ) {
+        let row_id_range = row_id_range.start..row_id_range.end.min(self.num_vals());
+        for idx in row_id_range.start..row_id_range.end {
+            let val = self.get_val(idx);
+            if value_range.contains(&val) {
+                row_id_hits.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<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> T {
+        self.as_ref().get_val(idx)
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> T {
+        self.as_ref().min_value()
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> T {
+        self.as_ref().max_value()
+    }
+
+    #[inline(always)]
+    fn num_vals(&self) -> u32 {
+        self.as_ref().num_vals()
+    }
+
+    #[inline(always)]
+    fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
+        self.as_ref().iter()
+    }
+
+    #[inline(always)]
+    fn get_range(&self, start: u64, output: &mut [T]) {
+        self.as_ref().get_range(start, output)
+    }
+
+    #[inline(always)]
+    fn get_row_ids_for_value_range(
+        &self,
+        range: RangeInclusive<T>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.as_ref()
+            .get_row_ids_for_value_range(range, doc_id_range, positions)
+    }
+}
+
+/// Wraps an cloneable 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 + Debug,
+{
+    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(all(test, feature = "unstable"))]
+mod bench;
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_range_as_col() {
+        let col = IterColumn::from(10..100);
+        assert_eq!(col.num_vals(), 90);
+        assert_eq!(col.max_value(), 99);
+    }
+}

columnar/src/column_values/monotonic_column.rs

@@ -0,0 +1,120 @@
+use std::fmt::Debug;
+use std::marker::PhantomData;
+use std::ops::{Range, RangeInclusive};
+
+use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
+use crate::ColumnValues;
+
+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 + Debug + Send + Sync + Clone,
+    Output: PartialOrd + Debug + 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 + Debug + Sync + Clone,
+    Output: PartialOrd + Send + Debug + 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_row_ids_for_value_range(
+        &self,
+        range: RangeInclusive<Output>,
+        doc_id_range: Range<u32>,
+        positions: &mut Vec<u32>,
+    ) {
+        self.from_column.get_row_ids_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.
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::monotonic_mapping::{
+        StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+    };
+    use crate::column_values::VecColumn;
+
+    #[test]
+    fn test_monotonic_mapping_iter() {
+        let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
+        let col = VecColumn::from(&vals);
+        let mapped = monotonic_map_column(
+            col,
+            StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<i64>::new()),
+        );
+        let val_i64s: Vec<u64> = mapped.iter().collect();
+        for i in 0..100 {
+            assert_eq!(val_i64s[i as usize], mapped.get_val(i));
+        }
+    }
+}

columnar/src/column_values/monotonic_mapping.rs

@@ -1,12 +1,14 @@
+use std::fmt::Debug;
 use std::marker::PhantomData;
 
-use fastdivide::DividerU64;
+use common::DateTime;
 
-use crate::MonotonicallyMappableToU128;
+use super::MonotonicallyMappableToU128;
+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 {
+pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Debug + Copy + Send + Sync {
     /// Converts a value to u64.
     ///
     /// Internally all fast field values are encoded as u64.
@@ -110,65 +112,6 @@ where T: MonotonicallyMappableToU64
     }
 }
 
-/// 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 {
@@ -193,6 +136,18 @@ impl MonotonicallyMappableToU64 for i64 {
     }
 }
 
+impl MonotonicallyMappableToU64 for DateTime {
+    #[inline(always)]
+    fn to_u64(self) -> u64 {
+        common::i64_to_u64(self.into_timestamp_micros())
+    }
+
+    #[inline(always)]
+    fn from_u64(val: u64) -> Self {
+        DateTime::from_timestamp_micros(common::u64_to_i64(val))
+    }
+}
+
 impl MonotonicallyMappableToU64 for bool {
     #[inline(always)]
     fn to_u64(self) -> u64 {
@@ -205,6 +160,18 @@ impl MonotonicallyMappableToU64 for bool {
     }
 }
 
+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 {
@@ -230,15 +197,9 @@ mod tests {
         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);
+        // test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
     }
 
     fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L>(

columnar/src/column_values/monotonic_mapping_u128.rs

@@ -1,8 +1,9 @@
+use std::fmt::Debug;
 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 + Debug + Send + Sync {
     /// Converts a value to u128.
     ///
     /// Internally all fast field values are encoded as u64.

columnar/src/column_values/stats.rs

@@ -0,0 +1,103 @@
+use std::io;
+use std::io::Write;
+use std::num::NonZeroU64;
+
+use common::{BinarySerializable, VInt};
+
+use crate::RowId;
+
+/// Column statistics.
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct ColumnStats {
+    /// GCD of the elements `el - min(column)`.
+    pub gcd: NonZeroU64,
+    /// Minimum value of the column.
+    pub min_value: u64,
+    /// Maximum value of the column.
+    pub max_value: u64,
+    /// Number of rows in the column.
+    pub num_rows: RowId,
+}
+
+impl ColumnStats {
+    /// Amplitude of value.
+    /// Difference between the maximum and the minimum value.
+    pub fn amplitude(&self) -> u64 {
+        self.max_value - self.min_value
+    }
+}
+
+impl BinarySerializable for ColumnStats {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.min_value).serialize(writer)?;
+        VInt(self.gcd.get()).serialize(writer)?;
+        VInt(self.amplitude() / self.gcd).serialize(writer)?;
+        VInt(self.num_rows as u64).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let min_value = VInt::deserialize(reader)?.0;
+        let gcd = VInt::deserialize(reader)?.0;
+        let gcd = NonZeroU64::new(gcd)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "GCD of 0 is forbidden"))?;
+        let amplitude = VInt::deserialize(reader)?.0 * gcd.get();
+        let max_value = min_value + amplitude;
+        let num_rows = VInt::deserialize(reader)?.0 as RowId;
+        Ok(ColumnStats {
+            min_value,
+            max_value,
+            num_rows,
+            gcd,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use common::BinarySerializable;
+
+    use crate::column_values::ColumnStats;
+
+    #[track_caller]
+    fn test_stats_ser_deser_aux(stats: &ColumnStats, num_bytes: usize) {
+        let mut buffer: Vec<u8> = Vec::new();
+        stats.serialize(&mut buffer).unwrap();
+        assert_eq!(buffer.len(), num_bytes);
+        let deser_stats = ColumnStats::deserialize(&mut &buffer[..]).unwrap();
+        assert_eq!(stats, &deser_stats);
+    }
+
+    #[test]
+    fn test_stats_serialization() {
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(3).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(1_000).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0,
+            }),
+            4,
+        );
+    }
+}

columnar/src/column_values/u128_based/compact_space/mod.rs

@@ -17,14 +17,15 @@ use std::{
     ops::{Range, RangeInclusive},
 };
 
+mod blank_range;
+mod build_compact_space;
+
+use build_compact_space::get_compact_space;
 use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
 use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
 
-use crate::compact_space::build_compact_space::get_compact_space;
-use crate::Column;
-
-mod blank_range;
-mod build_compact_space;
+use crate::column_values::ColumnValues;
+use crate::RowId;
 
 /// The cost per blank is quite hard actually, since blanks are delta encoded, the actual cost of
 /// blanks depends on the number of blanks.
@@ -55,7 +56,7 @@ impl RangeMapping {
 }
 
 impl BinarySerializable for CompactSpace {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
 
         let mut prev_value = 0;
@@ -158,23 +159,30 @@ impl CompactSpace {
 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_vals: RowId,
     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;
+    pub fn num_vals(&self) -> RowId {
+        self.params.num_vals
+    }
 
+    /// 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>) -> Self {
+        let mut values_sorted = BTreeSet::new();
+        let mut total_num_values = 0u32;
+        for val in iter {
+            total_num_values += 1u32;
+            values_sorted.insert(val);
+        }
         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();
@@ -247,7 +255,7 @@ pub struct CompactSpaceDecompressor {
 }
 
 impl BinarySerializable for IPCodecParams {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         // header flags for future optional dictionary encoding
         let footer_flags = 0u64;
         footer_flags.serialize(writer)?;
@@ -281,7 +289,7 @@ impl BinarySerializable for IPCodecParams {
     }
 }
 
-impl Column<u128> for CompactSpaceDecompressor {
+impl ColumnValues<u128> for CompactSpaceDecompressor {
     #[inline]
     fn get_val(&self, doc: u32) -> u128 {
         self.get(doc)
@@ -305,7 +313,7 @@ impl Column<u128> for CompactSpaceDecompressor {
     }
 
     #[inline]
-    fn get_docids_for_value_range(
+    fn get_row_ids_for_value_range(
         &self,
         value_range: RangeInclusive<u128>,
         positions_range: Range<u32>,
@@ -453,11 +461,11 @@ impl CompactSpaceDecompressor {
 #[cfg(test)]
 mod tests {
 
+    use itertools::Itertools;
+
     use super::*;
-    use crate::format_version::read_format_version;
-    use crate::null_index_footer::read_null_index_footer;
-    use crate::serialize::U128Header;
-    use crate::{open_u128, serialize_u128};
+    use crate::column_values::u128_based::U128Header;
+    use crate::column_values::{open_u128_mapped, serialize_column_values_u128};
 
     #[test]
     fn compact_space_test() {
@@ -533,18 +541,9 @@ mod tests {
 
     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();
-
+        serialize_column_values_u128(&u128_vals, &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
     }
 
@@ -575,8 +574,8 @@ mod tests {
         }
 
         // handle docid range out of bounds
-        let positions = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
-        assert_eq!(positions, vec![]);
+        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());
@@ -612,61 +611,59 @@ mod tests {
             vec![3, 4]
         );
         assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 99998u128..=99999u128,
                 complete_range.clone()
             ),
-            vec![3]
+            &[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,
-                99998u128..=99998u128,
-                complete_range.clone()
-            ),
-            vec![]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 333u128..=333u128,
                 complete_range.clone()
             ),
-            vec![8]
+            &[8]
         );
         assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 332u128..=333u128,
                 complete_range.clone()
             ),
-            vec![8]
+            &[8]
         );
         assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 332u128..=334u128,
                 complete_range.clone()
             ),
-            vec![8]
+            &[8]
         );
         assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 333u128..=334u128,
                 complete_range.clone()
             ),
-            vec![8]
+            &[8]
         );
 
         assert_eq!(
-            get_positions_for_value_range_helper(
+            &get_positions_for_value_range_helper(
                 &decomp,
                 4_000_211_221u128..=5_000_000_000u128,
                 complete_range
             ),
-            vec![6, 7]
+            &[6, 7]
         );
     }
 
@@ -692,27 +689,27 @@ mod tests {
         let _header = U128Header::deserialize(&mut data);
         let decomp = CompactSpaceDecompressor::open(data).unwrap();
         let complete_range = 0..vals.len() as u32;
-        assert_eq!(
-            get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone()),
-            vec![]
+        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()),
-            vec![0]
+            &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),
-            vec![0]
+            &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>(
+    fn get_positions_for_value_range_helper<C: ColumnValues<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);
+        column.get_row_ids_for_value_range(value_range, doc_id_range, &mut positions);
         positions
     }
 
@@ -734,8 +731,8 @@ mod tests {
             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();
+        serialize_column_values_u128(&&vals[..], &mut out).unwrap();
+        let decomp = open_u128_mapped(OwnedBytes::new(out)).unwrap();
         let complete_range = 0..vals.len() as u32;
 
         assert_eq!(
@@ -788,7 +785,7 @@ mod tests {
         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> {
@@ -804,10 +801,9 @@ mod tests {
     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);
-            }
+        #[test]
+        fn compress_decompress_random(vals in proptest::collection::vec(num_strategy() , 1..1000)) {
+            let _data = test_aux_vals(&vals);
         }
+    }
 }

columnar/src/column_values/u128_based/mod.rs

@@ -0,0 +1,178 @@
+use std::fmt::Debug;
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+mod compact_space;
+
+use common::{BinarySerializable, OwnedBytes, VInt};
+use compact_space::{CompactSpaceCompressor, CompactSpaceDecompressor};
+
+use crate::column_values::monotonic_map_column;
+use crate::column_values::monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+};
+use crate::iterable::Iterable;
+use crate::{ColumnValues, MonotonicallyMappableToU128};
+
+#[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 + ?Sized>(&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,
+        })
+    }
+}
+
+/// Serializes u128 values with the compact space codec.
+pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
+    iterable: &dyn Iterable<T>,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let compressor = CompactSpaceCompressor::train_from(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+    );
+    let header = U128Header {
+        num_vals: compressor.num_vals(),
+        codec_type: U128FastFieldCodecType::CompactSpace,
+    };
+    header.serialize(output)?;
+    compressor.compress_into(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+        output,
+    )?;
+    Ok(())
+}
+
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+/// Available codecs to use to encode the u128 (via [`MonotonicallyMappableToU128`]) converted data.
+pub(crate) 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 + ?Sized>(&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_mapped<T: MonotonicallyMappableToU128 + Debug>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let header = U128Header::deserialize(&mut bytes)?;
+    assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
+    let reader = CompactSpaceDecompressor::open(bytes)?;
+    let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<T>> =
+        StrictlyMonotonicMappingToInternal::<T>::new().into();
+    Ok(Arc::new(monotonic_map_column(reader, inverted)))
+}
+#[cfg(test)]
+pub mod tests {
+    use super::*;
+    use crate::column_values::u64_based::{
+        serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
+        ALL_U64_CODEC_TYPES,
+    };
+    use crate::column_values::CodecType;
+
+    #[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_u64_based_column_values(&&original[..], &ALL_U64_CODEC_TYPES)
+                .iter()
+                .collect();
+        assert_eq!(&restored, &original[..]);
+    }
+
+    #[test]
+    fn test_fastfield_bool_size_bitwidth_1() {
+        let mut buffer = Vec::new();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_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);
+    }
+
+    #[test]
+    fn test_fastfield_bool_bit_size_bitwidth_0() {
+        let mut buffer = Vec::new();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
+        // 6 bytes of header, 0 bytes of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 6);
+    }
+
+    #[test]
+    fn test_fastfield_gcd() {
+        let mut buffer = Vec::new();
+        let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
+        serialize_u64_based_column_values(&&vals[..], &[CodecType::Bitpacked], &mut buffer)
+            .unwrap();
+        // Values are stored over 3 bits.
+        assert_eq!(buffer.len(), 6 + (3 * 80 / 8));
+    }
+}

columnar/src/column_values/u64_based/bitpacked.rs

@@ -0,0 +1,127 @@
+use std::io::{self, Write};
+
+use common::{BinarySerializable, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+use crate::{ColumnValues, RowId};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct BitpackedReader {
+    data: OwnedBytes,
+    bit_unpacker: BitUnpacker,
+    stats: ColumnStats,
+}
+
+impl ColumnValues for BitpackedReader {
+    #[inline(always)]
+    fn get_val(&self, doc: u32) -> u64 {
+        self.stats.min_value + self.stats.gcd.get() * self.bit_unpacker.get(doc, &self.data)
+    }
+
+    #[inline]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+    #[inline]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+    #[inline]
+    fn num_vals(&self) -> RowId {
+        self.stats.num_rows
+    }
+}
+
+fn num_bits(stats: &ColumnStats) -> u8 {
+    compute_num_bits(stats.amplitude() / stats.gcd)
+}
+
+#[derive(Default)]
+pub struct BitpackedCodecEstimator;
+
+impl ColumnCodecEstimator for BitpackedCodecEstimator {
+    fn collect(&mut self, _value: u64) {}
+
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let num_bits_per_value = num_bits(stats);
+        Some(stats.num_bytes() + (stats.num_rows as u64 * (num_bits_per_value as u64) + 7) / 8)
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let num_bits = num_bits(stats);
+        let mut bit_packer = BitPacker::new();
+        let divider = DividerU64::divide_by(stats.gcd.get());
+        for val in vals {
+            bit_packer.write(divider.divide(val - stats.min_value), num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+}
+
+pub struct BitpackedCodec;
+
+impl ColumnCodec for BitpackedCodec {
+    type ColumnValues = BitpackedReader;
+    type Estimator = BitpackedCodecEstimator;
+
+    /// Opens a fast field given a file.
+    fn load(mut data: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut data)?;
+        let num_bits = num_bits(&stats);
+        let bit_unpacker = BitUnpacker::new(num_bits);
+        Ok(BitpackedReader {
+            data,
+            bit_unpacker,
+            stats,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BitpackedCodec>(&[4, 3, 12], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        create_and_validate::<BitpackedCodec>(&[1000, 2000, 3000], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BitpackedCodec>(&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::<BitpackedCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/blockwise_linear.rs

@@ -0,0 +1,281 @@
+use std::io::Write;
+use std::sync::Arc;
+use std::{io, iter};
+
+use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::u64_based::line::Line;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+use crate::column_values::{ColumnValues, VecColumn};
+use crate::MonotonicallyMappableToU64;
+
+const BLOCK_SIZE: u32 = 512u32;
+
+#[derive(Debug, Default)]
+struct Block {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+    data_start_offset: usize,
+}
+
+impl BinarySerializable for Block {
+    fn serialize<W: Write + ?Sized>(&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) -> u32 {
+    (num_vals + BLOCK_SIZE - 1) / BLOCK_SIZE
+}
+
+pub struct BlockwiseLinearEstimator {
+    block: Vec<u64>,
+    values_num_bytes: u64,
+    meta_num_bytes: u64,
+}
+
+impl Default for BlockwiseLinearEstimator {
+    fn default() -> Self {
+        Self {
+            block: Vec::with_capacity(BLOCK_SIZE as usize),
+            values_num_bytes: 0u64,
+            meta_num_bytes: 0u64,
+        }
+    }
+}
+
+impl BlockwiseLinearEstimator {
+    fn flush_block_estimate(&mut self) {
+        if self.block.is_empty() {
+            return;
+        }
+        let line = Line::train(&VecColumn::from(&self.block));
+        let mut max_value = 0u64;
+        for (i, buffer_val) in self.block.iter().enumerate() {
+            let interpolated_val = line.eval(i as u32);
+            let val = buffer_val.wrapping_sub(interpolated_val);
+            max_value = val.max(max_value);
+        }
+        let bit_width = compute_num_bits(max_value) as usize;
+        self.values_num_bytes += (bit_width * self.block.len() + 7) as u64 / 8;
+        self.meta_num_bytes += 1 + line.num_bytes();
+    }
+}
+
+impl ColumnCodecEstimator for BlockwiseLinearEstimator {
+    fn collect(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == BLOCK_SIZE as usize {
+            self.flush_block_estimate();
+            self.block.clear();
+        }
+    }
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let mut estimate = 4 + stats.num_bytes() + self.meta_num_bytes + self.values_num_bytes;
+        if stats.gcd.get() > 1 {
+            let estimate_gain_from_gcd =
+                (stats.gcd.get() as f32).log2().floor() * stats.num_rows as f32 / 8.0f32;
+            estimate = estimate.saturating_sub(estimate_gain_from_gcd as u64);
+        }
+        Some(estimate)
+    }
+
+    fn finalize(&mut self) {
+        self.flush_block_estimate();
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        mut vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let mut buffer = Vec::with_capacity(BLOCK_SIZE as usize);
+        let num_blocks = compute_num_blocks(stats.num_rows) as usize;
+        let mut blocks = Vec::with_capacity(num_blocks);
+
+        let mut bit_packer = BitPacker::new();
+
+        let gcd_divider = DividerU64::divide_by(stats.gcd.get());
+
+        for _ in 0..num_blocks {
+            buffer.clear();
+            buffer.extend(
+                (&mut vals)
+                    .map(MonotonicallyMappableToU64::to_u64)
+                    .take(BLOCK_SIZE as usize),
+            );
+
+            for buffer_val in buffer.iter_mut() {
+                *buffer_val = gcd_divider.divide(*buffer_val - stats.min_value);
+            }
+
+            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(), num_blocks);
+
+        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(())
+    }
+}
+
+pub struct BlockwiseLinearCodec;
+
+impl ColumnCodec<u64> for BlockwiseLinearCodec {
+    type ColumnValues = BlockwiseLinearReader;
+
+    type Estimator = BlockwiseLinearEstimator;
+
+    fn load(mut bytes: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut bytes)?;
+        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(stats.num_rows);
+        let mut blocks: Vec<Block> = iter::repeat_with(|| Block::deserialize(&mut footer))
+            .take(num_blocks as usize)
+            .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) * BLOCK_SIZE as usize / 8;
+        }
+        Ok(BlockwiseLinearReader {
+            blocks: blocks.into_boxed_slice().into(),
+            data,
+            stats,
+        })
+    }
+}
+
+#[derive(Clone)]
+pub struct BlockwiseLinearReader {
+    blocks: Arc<[Block]>,
+    data: OwnedBytes,
+    stats: ColumnStats,
+}
+
+impl ColumnValues for BlockwiseLinearReader {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> u64 {
+        let block_id = (idx / BLOCK_SIZE) as usize;
+        let idx_within_block = idx % BLOCK_SIZE;
+        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);
+        // TODO optimize me! the line parameters could be tweaked to include the multiplication and
+        // remove the dependency.
+        self.stats.min_value
+            + self
+                .stats
+                .gcd
+                .get()
+                .wrapping_mul(interpoled_val.wrapping_add(bitpacked_diff))
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline(always)]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BlockwiseLinearCodec>(
+            &[11, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "simple test",
+        )
+        .unwrap();
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        let (_, actual_compression_rate) = create_and_validate::<BlockwiseLinearCodec>(
+            &[10, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "name",
+        )
+        .unwrap();
+        assert_eq!(actual_compression_rate, 0.175);
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn test_blockwise_linear_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::<BlockwiseLinearCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/line.rs

@@ -3,7 +3,7 @@ use std::num::NonZeroU32;
 
 use common::{BinarySerializable, VInt};
 
-use crate::Column;
+use crate::column_values::ColumnValues;
 
 const MID_POINT: u64 = (1u64 << 32) - 1u64;
 
@@ -17,8 +17,8 @@ const MID_POINT: u64 = (1u64 << 32) - 1u64;
 /// `y = m * x >> 32 + b`
 #[derive(Debug, Clone, Copy, Default)]
 pub struct Line {
-    slope: u64,
-    intercept: u64,
+    pub(crate) slope: u64,
+    pub(crate) intercept: u64,
 }
 
 /// Compute the line slope.
@@ -67,21 +67,8 @@ impl Line {
         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(
+    pub fn train_from(
         first_val: u64,
         last_val: u64,
         num_vals: u32,
@@ -145,7 +132,8 @@ impl Line {
     ///
     /// 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 Column) -> Self {
+    /// TODO USE array
+    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(
@@ -158,7 +146,7 @@ impl Line {
 }
 
 impl BinarySerializable for Line {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.slope).serialize(writer)?;
         VInt(self.intercept).serialize(writer)?;
         Ok(())
@@ -174,7 +162,7 @@ impl BinarySerializable for Line {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::VecColumn;
+    use crate::column_values::VecColumn;
 
     /// Test training a line and ensuring that the maximum difference between
     /// the data points and the line is `expected`.

columnar/src/column_values/u64_based/linear.rs

@@ -0,0 +1,277 @@
+use std::io;
+
+use common::{BinarySerializable, OwnedBytes};
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use super::line::Line;
+use super::ColumnValues;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
+use crate::column_values::VecColumn;
+use crate::RowId;
+
+const HALF_SPACE: u64 = u64::MAX / 2;
+const LINE_ESTIMATION_BLOCK_LEN: usize = 512;
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct LinearReader {
+    data: OwnedBytes,
+    linear_params: LinearParams,
+    stats: ColumnStats,
+}
+
+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 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+/// 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 + ?Sized>(&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),
+        })
+    }
+}
+
+pub struct LinearCodecEstimator {
+    block: Vec<u64>,
+    line: Option<Line>,
+    row_id: RowId,
+    min_deviation: u64,
+    max_deviation: u64,
+    first_val: u64,
+    last_val: u64,
+}
+
+impl Default for LinearCodecEstimator {
+    fn default() -> LinearCodecEstimator {
+        LinearCodecEstimator {
+            block: Vec::with_capacity(LINE_ESTIMATION_BLOCK_LEN),
+            line: None,
+            row_id: 0,
+            min_deviation: u64::MAX,
+            max_deviation: u64::MIN,
+            first_val: 0u64,
+            last_val: 0u64,
+        }
+    }
+}
+
+impl ColumnCodecEstimator for LinearCodecEstimator {
+    fn finalize(&mut self) {
+        if let Some(line) = self.line.as_mut() {
+            line.intercept = line
+                .intercept
+                .wrapping_add(self.min_deviation)
+                .wrapping_sub(HALF_SPACE);
+        }
+    }
+
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64> {
+        let line = self.line?;
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        Some(
+            stats.num_bytes()
+                + linear_params.num_bytes()
+                + (num_bits as u64 * stats.num_rows as u64 + 7) / 8,
+        )
+    }
+
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let line = self.line.unwrap();
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        linear_params.serialize(wrt)?;
+        let mut bit_packer = BitPacker::new();
+        for (pos, value) in vals.enumerate() {
+            let calculated_value = line.eval(pos as u32);
+            let offset = value.wrapping_sub(calculated_value);
+            bit_packer.write(offset, num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+
+    fn collect(&mut self, value: u64) {
+        if let Some(line) = self.line {
+            self.collect_after_line_estimation(&line, value);
+        } else {
+            self.collect_before_line_estimation(value);
+        }
+    }
+}
+
+impl LinearCodecEstimator {
+    #[inline]
+    fn collect_after_line_estimation(&mut self, line: &Line, value: u64) {
+        let interpoled_val: u64 = line.eval(self.row_id);
+        let deviation = value.wrapping_add(HALF_SPACE).wrapping_sub(interpoled_val);
+        self.min_deviation = self.min_deviation.min(deviation);
+        self.max_deviation = self.max_deviation.max(deviation);
+        if self.row_id == 0 {
+            self.first_val = value;
+        }
+        self.last_val = value;
+        self.row_id += 1u32;
+    }
+
+    #[inline]
+    fn collect_before_line_estimation(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == LINE_ESTIMATION_BLOCK_LEN {
+            let line = Line::train(&VecColumn::from(&self.block));
+            let block = std::mem::take(&mut self.block);
+            for val in block {
+                self.collect_after_line_estimation(&line, val);
+            }
+            self.line = Some(line);
+        }
+    }
+}
+
+impl ColumnCodec for LinearCodec {
+    type ColumnValues = LinearReader;
+
+    type Estimator = LinearCodecEstimator;
+
+    fn load(mut data: OwnedBytes) -> io::Result<Self::ColumnValues> {
+        let stats = ColumnStats::deserialize(&mut data)?;
+        let linear_params = LinearParams::deserialize(&mut data)?;
+        Ok(LinearReader {
+            stats,
+            linear_params,
+            data,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use rand::RngCore;
+
+    use super::*;
+    use crate::column_values::u64_based::tests::{create_and_validate, get_codec_test_datasets};
+
+    #[test]
+    fn test_compression_simple() {
+        let vals = (100u64..)
+            .take(super::LINE_ESTIMATION_BLOCK_LEN)
+            .collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&vals, "simple monotonically large").unwrap();
+    }
+
+    #[test]
+    fn test_compression() {
+        let data = (10..=6_000_u64).collect::<Vec<_>>();
+        let (estimate, actual_compression) =
+            create_and_validate::<LinearCodec>(&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 = get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<LinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<LinearCodec>(&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::<LinearCodec>(&data, "large amplitude");
+    }
+
+    #[test]
+    fn overflow_error_test() {
+        let data = vec![1572656989877777, 1170935903116329, 720575940379279, 0];
+        create_and_validate::<LinearCodec>(&data, "overflow test");
+    }
+
+    #[test]
+    fn linear_interpol_fast_concave_data() {
+        let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
+        create_and_validate::<LinearCodec>(&data, "concave data");
+    }
+    #[test]
+    fn linear_interpol_fast_convex_data() {
+        let data = vec![0, 40, 60, 70, 75, 77];
+        create_and_validate::<LinearCodec>(&data, "convex data");
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_simple() {
+        let data = (10..=20_u64).collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&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::<LinearCodec>(&data, "random");
+            data.reverse();
+            create_and_validate::<LinearCodec>(&data, "random");
+        }
+    }
+}

columnar/src/column_values/u64_based/mod.rs

@@ -0,0 +1,214 @@
+mod bitpacked;
+mod blockwise_linear;
+mod line;
+mod linear;
+mod stats_collector;
+
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::{BinarySerializable, OwnedBytes};
+
+use crate::column_values::monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+};
+pub use crate::column_values::u64_based::bitpacked::BitpackedCodec;
+pub use crate::column_values::u64_based::blockwise_linear::BlockwiseLinearCodec;
+pub use crate::column_values::u64_based::linear::LinearCodec;
+pub use crate::column_values::u64_based::stats_collector::StatsCollector;
+use crate::column_values::{monotonic_map_column, ColumnStats};
+use crate::iterable::Iterable;
+use crate::{ColumnValues, MonotonicallyMappableToU64};
+
+/// A `ColumnCodecEstimator` is in charge of gathering all
+/// data required to serialize a column.
+///
+/// This happens during a first pass on data of the column elements.
+/// During that pass, all column estimators receive a call to their
+/// `.collect(el)`.
+///
+/// After this first pass, finalize is called.
+/// `.estimate(..)` then should return an accurate estimation of the
+/// size of the serialized column (were we to pick this codec.).
+/// `.serialize(..)` then serializes the column using this codec.
+pub trait ColumnCodecEstimator<T = u64>: 'static {
+    /// Records a new value for estimation.
+    /// This method will be called for each element of the column during
+    /// `estimation`.
+    fn collect(&mut self, value: u64);
+    /// Finalizes the first pass phase.
+    fn finalize(&mut self) {}
+    /// Returns an accurate estimation of the number of bytes that will
+    /// be used to represent this column.
+    fn estimate(&self, stats: &ColumnStats) -> Option<u64>;
+    /// Serializes the column using the given codec.
+    /// This constitutes a second pass over the columns values.
+    fn serialize(
+        &self,
+        stats: &ColumnStats,
+        vals: &mut dyn Iterator<Item = T>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()>;
+}
+
+/// A column codec describes a colunm serialization format.
+pub trait ColumnCodec<T: PartialOrd = u64> {
+    /// Specialized `ColumnValues` type.
+    type ColumnValues: ColumnValues<T> + 'static;
+    /// `Estimator` for the given codec.
+    type Estimator: ColumnCodecEstimator + Default;
+
+    /// Loads a column that has been serialized using this codec.
+    fn load(bytes: OwnedBytes) -> io::Result<Self::ColumnValues>;
+
+    /// Returns an estimator.
+    fn estimator() -> Self::Estimator {
+        Self::Estimator::default()
+    }
+
+    /// Returns a boxed estimator.
+    fn boxed_estimator() -> Box<dyn ColumnCodecEstimator> {
+        Box::new(Self::estimator())
+    }
+}
+
+/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+pub enum CodecType {
+    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
+    /// `column.max_value() - column.min_value()`
+    Bitpacked = 0u8,
+    /// 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 = 1u8,
+    /// Same as [`CodecType::Linear`], but encodes in blocks of 512 elements.
+    BlockwiseLinear = 2u8,
+}
+
+/// List of all available u64-base codecs.
+pub const ALL_U64_CODEC_TYPES: [CodecType; 3] = [
+    CodecType::Bitpacked,
+    CodecType::Linear,
+    CodecType::BlockwiseLinear,
+];
+
+impl CodecType {
+    fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    fn try_from_code(code: u8) -> Option<CodecType> {
+        match code {
+            0u8 => Some(CodecType::Bitpacked),
+            1u8 => Some(CodecType::Linear),
+            2u8 => Some(CodecType::BlockwiseLinear),
+            _ => None,
+        }
+    }
+
+    fn load<T: MonotonicallyMappableToU64>(
+        &self,
+        bytes: OwnedBytes,
+    ) -> io::Result<Arc<dyn ColumnValues<T>>> {
+        match self {
+            CodecType::Bitpacked => load_specific_codec::<BitpackedCodec, T>(bytes),
+            CodecType::Linear => load_specific_codec::<LinearCodec, T>(bytes),
+            CodecType::BlockwiseLinear => load_specific_codec::<BlockwiseLinearCodec, T>(bytes),
+        }
+    }
+}
+
+fn load_specific_codec<C: ColumnCodec, T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let reader = C::load(bytes)?;
+    let reader_typed = monotonic_map_column(
+        reader,
+        StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<T>::new()),
+    );
+    Ok(Arc::new(reader_typed))
+}
+
+impl CodecType {
+    /// Returns a boxed codec estimator associated to a given `CodecType`.
+    pub fn estimator(&self) -> Box<dyn ColumnCodecEstimator> {
+        match self {
+            CodecType::Bitpacked => BitpackedCodec::boxed_estimator(),
+            CodecType::Linear => LinearCodec::boxed_estimator(),
+            CodecType::BlockwiseLinear => BlockwiseLinearCodec::boxed_estimator(),
+        }
+    }
+}
+
+/// Serializes a given column of u64-mapped values.
+pub fn serialize_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable<T>,
+    codec_types: &[CodecType],
+    wrt: &mut dyn Write,
+) -> io::Result<()> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimators: Vec<(CodecType, Box<dyn ColumnCodecEstimator>)> =
+        Vec::with_capacity(codec_types.len());
+    for &codec_type in codec_types {
+        estimators.push((codec_type, codec_type.estimator()));
+    }
+    for val in vals.boxed_iter() {
+        let val_u64 = val.to_u64();
+        stats_collector.collect(val_u64);
+        for (_, estimator) in &mut estimators {
+            estimator.collect(val_u64);
+        }
+    }
+    for (_, estimator) in &mut estimators {
+        estimator.finalize();
+    }
+    let stats = stats_collector.stats();
+    let (_, best_codec, best_codec_estimator) = estimators
+        .into_iter()
+        .flat_map(|(codec_type, estimator)| {
+            let num_bytes = estimator.estimate(&stats)?;
+            Some((num_bytes, codec_type, estimator))
+        })
+        .min_by_key(|(num_bytes, _, _)| *num_bytes)
+        .ok_or_else(|| {
+            io::Error::new(io::ErrorKind::InvalidData, "No available applicable codec.")
+        })?;
+    best_codec.to_code().serialize(wrt)?;
+    best_codec_estimator.serialize(
+        &stats,
+        &mut vals.boxed_iter().map(MonotonicallyMappableToU64::to_u64),
+        wrt,
+    )?;
+    Ok(())
+}
+
+/// Load u64-based column values.
+///
+/// This method first identifies the codec off the first byte.
+pub fn load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let codec_type: CodecType = bytes
+        .first()
+        .copied()
+        .and_then(CodecType::try_from_code)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Failed to read codec type"))?;
+    bytes.advance(1);
+    codec_type.load(bytes)
+}
+
+/// Helper function to serialize a column (autodetect from all codecs) and then open it
+pub fn serialize_and_load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable,
+    codec_types: &[CodecType],
+) -> Arc<dyn ColumnValues<T>> {
+    let mut buffer = Vec::new();
+    serialize_u64_based_column_values(vals, codec_types, &mut buffer).unwrap();
+    load_u64_based_column_values::<T>(OwnedBytes::new(buffer)).unwrap()
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_values/u64_based/stats_collector.rs

@@ -0,0 +1,200 @@
+use std::num::NonZeroU64;
+
+use fastdivide::DividerU64;
+
+use crate::column_values::ColumnStats;
+use crate::RowId;
+
+/// 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;
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct StatsCollector {
+    min_max_opt: Option<(u64, u64)>,
+    num_rows: RowId,
+    // We measure the GCD of the difference between the values and the minimal value.
+    // This is the same as computing the difference between the values and the first value.
+    //
+    // This way, we can compress i64-converted-to-u64 (e.g. timestamp that were supplied in
+    // seconds, only to be converted in microseconds).
+    increment_gcd_opt: Option<(NonZeroU64, DividerU64)>,
+    first_value_opt: Option<u64>,
+}
+
+impl StatsCollector {
+    pub fn stats(&self) -> ColumnStats {
+        let (min_value, max_value) = self.min_max_opt.unwrap_or((0u64, 0u64));
+        let increment_gcd = if let Some((increment_gcd, _)) = self.increment_gcd_opt {
+            increment_gcd
+        } else {
+            NonZeroU64::new(1u64).unwrap()
+        };
+        ColumnStats {
+            min_value,
+            max_value,
+            num_rows: self.num_rows,
+            gcd: increment_gcd,
+        }
+    }
+
+    #[inline]
+    fn update_increment_gcd(&mut self, value: u64) {
+        let Some(first_value) = self.first_value_opt else {
+            // We set the first value and just quit.
+            self.first_value_opt = Some(value);
+            return;
+        };
+        let Some(non_zero_value) = NonZeroU64::new(value.abs_diff(first_value)) else {
+            // We can simply skip 0 values.
+            return;
+        };
+        let Some((gcd, gcd_divider)) = self.increment_gcd_opt else {
+            self.set_increment_gcd(non_zero_value);
+            return;
+        };
+        if gcd.get() == 1 {
+            // It won't see any update now.
+            return;
+        }
+        let remainder =
+            non_zero_value.get() - (gcd_divider.divide(non_zero_value.get())) * gcd.get();
+        if remainder == 0 {
+            return;
+        }
+        let new_gcd = compute_gcd(non_zero_value, gcd);
+        self.set_increment_gcd(new_gcd);
+    }
+
+    fn set_increment_gcd(&mut self, gcd: NonZeroU64) {
+        let new_divider = DividerU64::divide_by(gcd.get());
+        self.increment_gcd_opt = Some((gcd, new_divider));
+    }
+
+    pub fn collect(&mut self, value: u64) {
+        self.min_max_opt = Some(if let Some((min, max)) = self.min_max_opt {
+            (min.min(value), max.max(value))
+        } else {
+            (value, value)
+        });
+        self.num_rows += 1;
+        self.update_increment_gcd(value);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use crate::column_values::u64_based::stats_collector::{compute_gcd, StatsCollector};
+    use crate::column_values::u64_based::ColumnStats;
+
+    fn compute_stats(vals: impl Iterator<Item = u64>) -> ColumnStats {
+        let mut stats_collector = StatsCollector::default();
+        for val in vals {
+            stats_collector.collect(val);
+        }
+        stats_collector.stats()
+    }
+
+    fn find_gcd(vals: impl Iterator<Item = u64>) -> u64 {
+        compute_stats(vals).gcd.get()
+    }
+
+    #[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 test_gcd() {
+        assert_eq!(find_gcd([0].into_iter()), 1);
+        assert_eq!(find_gcd([0, 10].into_iter()), 10);
+        assert_eq!(find_gcd([10, 0].into_iter()), 10);
+        assert_eq!(find_gcd([].into_iter()), 1);
+        assert_eq!(find_gcd([15, 30, 5, 10].into_iter()), 5);
+        assert_eq!(find_gcd([15, 16, 10].into_iter()), 1);
+        assert_eq!(find_gcd([0, 5, 5, 5].into_iter()), 5);
+        assert_eq!(find_gcd([0, 0].into_iter()), 1);
+        assert_eq!(find_gcd([1, 10, 4, 1, 7, 10].into_iter()), 3);
+        assert_eq!(find_gcd([1, 10, 0, 4, 1, 7, 10].into_iter()), 1);
+    }
+
+    #[test]
+    fn test_stats() {
+        assert_eq!(
+            compute_stats([].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 20, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 10,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 50, 10, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(20).unwrap(),
+                min_value: 10,
+                max_value: 50,
+                num_rows: 4
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 0, 30].into_iter()),
+            ColumnStats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 0,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+    }
+}

columnar/src/column_values/u64_based/tests.rs

@@ -0,0 +1,401 @@
+use proptest::prelude::*;
+use proptest::strategy::Strategy;
+use proptest::{num, prop_oneof, proptest};
+
+#[test]
+fn test_serialize_and_load_simple() {
+    let mut buffer = Vec::new();
+    let vals = &[1u64, 2u64, 5u64];
+    serialize_u64_based_column_values(
+        &&vals[..],
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        &mut buffer,
+    )
+    .unwrap();
+    assert_eq!(buffer.len(), 7);
+    let col = load_u64_based_column_values::<u64>(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(col.num_vals(), 3);
+    assert_eq!(col.get_val(0), 1);
+    assert_eq!(col.get_val(1), 2);
+    assert_eq!(col.get_val(2), 5);
+}
+
+#[test]
+fn test_empty_column_i64() {
+    let vals: [i64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<i64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        assert_eq!(col.min_value(), i64::MIN);
+        assert_eq!(col.max_value(), i64::MIN);
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+#[test]
+fn test_empty_column_u64() {
+    let vals: [u64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<u64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        assert_eq!(col.min_value(), u64::MIN);
+        assert_eq!(col.max_value(), u64::MIN);
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+#[test]
+fn test_empty_column_f64() {
+    let vals: [f64; 0] = [];
+    let mut num_acceptable_codecs = 0;
+    for codec in ALL_U64_CODEC_TYPES {
+        let mut buffer = Vec::new();
+        if serialize_u64_based_column_values(&&vals[..], &[codec], &mut buffer).is_err() {
+            continue;
+        }
+        num_acceptable_codecs += 1;
+        let col = load_u64_based_column_values::<f64>(OwnedBytes::new(buffer)).unwrap();
+        assert_eq!(col.num_vals(), 0);
+        // FIXME. f64::MIN would be better!
+        assert!(col.min_value().is_nan());
+        assert!(col.max_value().is_nan());
+    }
+    assert!(num_acceptable_codecs > 0);
+}
+
+pub(crate) fn create_and_validate<TColumnCodec: ColumnCodec>(
+    vals: &[u64],
+    name: &str,
+) -> Option<(f32, f32)> {
+    let mut stats_collector = StatsCollector::default();
+    let mut codec_estimator: TColumnCodec::Estimator = Default::default();
+
+    for val in vals.boxed_iter() {
+        stats_collector.collect(val);
+        codec_estimator.collect(val);
+    }
+    codec_estimator.finalize();
+    let stats = stats_collector.stats();
+    let estimation = codec_estimator.estimate(&stats)?;
+
+    let mut buffer = Vec::new();
+    codec_estimator
+        .serialize(&stats, vals.boxed_iter().as_mut(), &mut buffer)
+        .unwrap();
+
+    let actual_compression = buffer.len() as u64;
+
+    let reader = TColumnCodec::load(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(reader.num_vals(), vals.len() as u32);
+    for (doc, orig_val) in vals.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 `{vals:?}`",
+        );
+    }
+
+    if !vals.is_empty() {
+        let test_rand_idx = rand::thread_rng().gen_range(0..=vals.len() - 1);
+        let expected_positions: Vec<u32> = vals
+            .iter()
+            .enumerate()
+            .filter(|(_, el)| **el == vals[test_rand_idx])
+            .map(|(pos, _)| pos as u32)
+            .collect();
+        let mut positions = Vec::new();
+        reader.get_row_ids_for_value_range(
+            vals[test_rand_idx]..=vals[test_rand_idx],
+            0..vals.len() as u32,
+            &mut positions,
+        );
+        assert_eq!(expected_positions, positions);
+    }
+    if actual_compression > 1000 {
+        assert!(relative_difference(estimation, actual_compression) < 0.10f32);
+    }
+    Some((
+        compression_rate(estimation, stats.num_rows),
+        compression_rate(actual_compression, stats.num_rows),
+    ))
+}
+
+fn compression_rate(num_bytes: u64, num_values: u32) -> f32 {
+    num_bytes as f32 / (num_values as f32 * 8.0)
+}
+
+fn relative_difference(left: u64, right: u64) -> f32 {
+    let left = left as f32;
+    let right = right as f32;
+    2.0f32 * (left - right).abs() / (left + right)
+}
+
+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");
+    }
+}
+
+#[test]
+fn test_small_blockwise_linear_example() {
+    create_and_validate::<BlockwiseLinearCodec>(
+        &[9223372036854775808, 9223370937344622593],
+        "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: ColumnCodec>() {
+    let codec_name = std::any::type_name::<C>();
+    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::*;
+
+fn estimate<C: ColumnCodec>(vals: &[u64]) -> Option<f32> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimator = C::Estimator::default();
+    for &val in vals {
+        stats_collector.collect(val);
+        estimator.collect(val);
+    }
+    estimator.finalize();
+    let stats = stats_collector.stats();
+    let num_bytes = estimator.estimate(&stats)?;
+    if stats.num_rows == 0 {
+        return None;
+    }
+    Some(num_bytes as f32 / (8.0 * stats.num_rows as f32))
+}
+
+#[test]
+fn estimation_good_interpolation_case() {
+    let data = (10..=20000_u64).collect::<Vec<_>>();
+
+    let linear_interpol_estimation = estimate::<LinearCodec>(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.01);
+
+    let multi_linear_interpol_estimation = estimate::<BlockwiseLinearCodec>(&data).unwrap();
+    assert_le!(multi_linear_interpol_estimation, 0.2);
+    assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
+
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
+    assert_lt!(linear_interpol_estimation, bitpacked_estimation);
+}
+
+#[test]
+fn estimation_test_bad_interpolation_case_monotonically_increasing() {
+    let mut data: Vec<u64> = (201..=20000_u64).collect();
+    data.push(1_000_000);
+
+    // 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 = estimate::<LinearCodec>(&data[..]).unwrap();
+    assert_le!(linear_interpol_estimation, 0.35);
+
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&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) = CodecType::try_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: CodecType, 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_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::load_u64_based_column_values::<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_u64_based_column_values(
+        &&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 &[
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::Linear,
+    ] {
+        test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
+    }
+    Ok(())
+}
+
+fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, 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_u64_based_column_values(
+        &&vals[..],
+        &[codec_type],
+        &mut buffer,
+    )?;
+    let buffer = OwnedBytes::new(buffer);
+    let column = crate::column_values::load_u64_based_column_values::<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_u64_based_column_values(
+        &&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 &[
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::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_u64_based_column_values::<u64>(
+        &&[100u64, 200u64, 300u64][..],
+        &ALL_U64_CODEC_TYPES,
+    );
+    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/column_values/vec_column.rs

@@ -0,0 +1,52 @@
+use std::fmt::Debug;
+
+use tantivy_bitpacker::minmax;
+
+use crate::ColumnValues;
+
+/// 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 + Debug> 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,
+        }
+    }
+}

columnar/src/columnar/column_type.rs

@@ -0,0 +1,163 @@
+use std::fmt::Debug;
+use std::net::Ipv6Addr;
+
+use serde::{Deserialize, Serialize};
+
+use crate::value::NumericalType;
+use crate::InvalidData;
+
+/// The column type represents the column type.
+/// Any changes need to be propagated to `COLUMN_TYPES`.
+#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy, Ord, PartialOrd, Serialize, Deserialize)]
+#[repr(u8)]
+pub enum ColumnType {
+    I64 = 0u8,
+    U64 = 1u8,
+    F64 = 2u8,
+    Bytes = 3u8,
+    Str = 4u8,
+    Bool = 5u8,
+    IpAddr = 6u8,
+    DateTime = 7u8,
+}
+
+// The order needs to match _exactly_ the order in the enum
+const COLUMN_TYPES: [ColumnType; 8] = [
+    ColumnType::I64,
+    ColumnType::U64,
+    ColumnType::F64,
+    ColumnType::Bytes,
+    ColumnType::Str,
+    ColumnType::Bool,
+    ColumnType::IpAddr,
+    ColumnType::DateTime,
+];
+
+impl ColumnType {
+    pub fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
+        COLUMN_TYPES.get(code as usize).copied().ok_or(InvalidData)
+    }
+}
+
+impl From<NumericalType> for ColumnType {
+    fn from(numerical_type: NumericalType) -> Self {
+        match numerical_type {
+            NumericalType::I64 => ColumnType::I64,
+            NumericalType::U64 => ColumnType::U64,
+            NumericalType::F64 => ColumnType::F64,
+        }
+    }
+}
+
+impl ColumnType {
+    pub fn numerical_type(&self) -> Option<NumericalType> {
+        match self {
+            ColumnType::I64 => Some(NumericalType::I64),
+            ColumnType::U64 => Some(NumericalType::U64),
+            ColumnType::F64 => Some(NumericalType::F64),
+            ColumnType::Bytes
+            | ColumnType::Str
+            | ColumnType::Bool
+            | ColumnType::IpAddr
+            | ColumnType::DateTime => None,
+        }
+    }
+}
+
+// TODO remove if possible
+pub trait HasAssociatedColumnType: 'static + Debug + Send + Sync + Copy + PartialOrd {
+    fn column_type() -> ColumnType;
+    fn default_value() -> Self;
+}
+
+impl HasAssociatedColumnType for u64 {
+    fn column_type() -> ColumnType {
+        ColumnType::U64
+    }
+
+    fn default_value() -> Self {
+        0u64
+    }
+}
+
+impl HasAssociatedColumnType for i64 {
+    fn column_type() -> ColumnType {
+        ColumnType::I64
+    }
+
+    fn default_value() -> Self {
+        0i64
+    }
+}
+
+impl HasAssociatedColumnType for f64 {
+    fn column_type() -> ColumnType {
+        ColumnType::F64
+    }
+
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for bool {
+    fn column_type() -> ColumnType {
+        ColumnType::Bool
+    }
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for common::DateTime {
+    fn column_type() -> ColumnType {
+        ColumnType::DateTime
+    }
+    fn default_value() -> Self {
+        Default::default()
+    }
+}
+
+impl HasAssociatedColumnType for Ipv6Addr {
+    fn column_type() -> ColumnType {
+        ColumnType::IpAddr
+    }
+
+    fn default_value() -> Self {
+        Ipv6Addr::from([0u8; 16])
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::Cardinality;
+
+    #[test]
+    fn test_column_type_to_code() {
+        for (code, expected_column_type) in super::COLUMN_TYPES.iter().copied().enumerate() {
+            if let Ok(column_type) = ColumnType::try_from_code(code as u8) {
+                assert_eq!(column_type, expected_column_type);
+            }
+        }
+        for code in COLUMN_TYPES.len() as u8..=u8::MAX {
+            assert!(ColumnType::try_from_code(code).is_err());
+        }
+    }
+
+    #[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, 66];
+
+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/merge/merge_dict_column.rs

@@ -0,0 +1,204 @@
+use std::io::{self, Write};
+
+use common::{BitSet, CountingWriter, ReadOnlyBitSet};
+use sstable::{SSTable, TermOrdinal};
+
+use super::term_merger::TermMerger;
+use crate::column::serialize_column_mappable_to_u64;
+use crate::column_index::SerializableColumnIndex;
+use crate::iterable::Iterable;
+use crate::{BytesColumn, MergeRowOrder, ShuffleMergeOrder};
+
+// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
+// Column: [Column Index, Column Values, column index num bytes U32::LE]
+pub fn merge_bytes_or_str_column(
+    column_index: SerializableColumnIndex<'_>,
+    bytes_columns: &[Option<BytesColumn>],
+    merge_row_order: &MergeRowOrder,
+    output: &mut impl Write,
+) -> io::Result<()> {
+    // Serialize dict and generate mapping for values
+    let mut output = CountingWriter::wrap(output);
+    // TODO !!! Remove useless terms.
+    let term_ord_mapping = serialize_merged_dict(bytes_columns, merge_row_order, &mut output)?;
+    let dictionary_num_bytes: u32 = output.written_bytes() as u32;
+    let output = output.finish();
+    let remapped_term_ordinals_values = RemappedTermOrdinalsValues {
+        bytes_columns,
+        term_ord_mapping: &term_ord_mapping,
+        merge_row_order,
+    };
+    serialize_column_mappable_to_u64(column_index, &remapped_term_ordinals_values, output)?;
+    output.write_all(&dictionary_num_bytes.to_le_bytes())?;
+    Ok(())
+}
+
+struct RemappedTermOrdinalsValues<'a> {
+    bytes_columns: &'a [Option<BytesColumn>],
+    term_ord_mapping: &'a TermOrdinalMapping,
+    merge_row_order: &'a MergeRowOrder,
+}
+
+impl<'a> Iterable for RemappedTermOrdinalsValues<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u64> + '_> {
+        match self.merge_row_order {
+            MergeRowOrder::Stack(_) => self.boxed_iter_stacked(),
+            MergeRowOrder::Shuffled(shuffle_merge_order) => {
+                self.boxed_iter_shuffled(shuffle_merge_order)
+            }
+        }
+    }
+}
+
+impl<'a> RemappedTermOrdinalsValues<'a> {
+    fn boxed_iter_stacked(&self) -> Box<dyn Iterator<Item = u64> + '_> {
+        let iter = self
+            .bytes_columns
+            .iter()
+            .enumerate()
+            .flat_map(|(segment_ord, byte_column)| {
+                let segment_ord = self.term_ord_mapping.get_segment(segment_ord as u32);
+                byte_column.iter().flat_map(move |bytes_column| {
+                    bytes_column
+                        .ords()
+                        .values
+                        .iter()
+                        .map(move |term_ord| segment_ord[term_ord as usize])
+                })
+            });
+        // TODO see if we can better decompose the mapping / and the stacking
+        Box::new(iter)
+    }
+
+    fn boxed_iter_shuffled<'b>(
+        &'b self,
+        shuffle_merge_order: &'b ShuffleMergeOrder,
+    ) -> Box<dyn Iterator<Item = u64> + 'b> {
+        Box::new(
+            shuffle_merge_order
+                .iter_new_to_old_row_addrs()
+                .flat_map(move |old_addr| {
+                    let segment_ord = self.term_ord_mapping.get_segment(old_addr.segment_ord);
+                    self.bytes_columns[old_addr.segment_ord as usize]
+                        .as_ref()
+                        .into_iter()
+                        .flat_map(move |bytes_column| {
+                            bytes_column
+                                .term_ords(old_addr.row_id)
+                                .map(|old_term_ord: u64| segment_ord[old_term_ord as usize])
+                        })
+                }),
+        )
+    }
+}
+
+fn compute_term_bitset(column: &BytesColumn, row_bitset: &ReadOnlyBitSet) -> BitSet {
+    let num_terms = column.dictionary().num_terms();
+    let mut term_bitset = BitSet::with_max_value(num_terms as u32);
+    for row_id in row_bitset.iter() {
+        for term_ord in column.term_ord_column.values_for_doc(row_id) {
+            term_bitset.insert(term_ord as u32);
+        }
+    }
+    term_bitset
+}
+
+fn is_term_present(bitsets: &[Option<BitSet>], term_merger: &TermMerger) -> bool {
+    for (segment_ord, from_term_ord) in term_merger.matching_segments() {
+        if let Some(bitset) = bitsets[segment_ord].as_ref() {
+            if bitset.contains(from_term_ord as u32) {
+                return true;
+            }
+        } else {
+            return true;
+        }
+    }
+    false
+}
+
+fn serialize_merged_dict(
+    bytes_columns: &[Option<BytesColumn>],
+    merge_row_order: &MergeRowOrder,
+    output: &mut impl Write,
+) -> io::Result<TermOrdinalMapping> {
+    let mut term_ord_mapping = TermOrdinalMapping::default();
+
+    let mut field_term_streams = Vec::new();
+    for column in bytes_columns.iter().flatten() {
+        term_ord_mapping.add_segment(column.dictionary.num_terms());
+        let terms = column.dictionary.stream()?;
+        field_term_streams.push(terms);
+    }
+
+    let mut merged_terms = TermMerger::new(field_term_streams);
+    let mut sstable_builder = sstable::VoidSSTable::writer(output);
+
+    // TODO support complex `merge_row_order`.
+    match merge_row_order {
+        MergeRowOrder::Stack(_) => {
+            let mut current_term_ord = 0;
+            while merged_terms.advance() {
+                let term_bytes: &[u8] = merged_terms.key();
+                sstable_builder.insert(term_bytes, &())?;
+                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
+                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, current_term_ord);
+                }
+                current_term_ord += 1;
+            }
+            sstable_builder.finish()?;
+        }
+        MergeRowOrder::Shuffled(shuffle_merge_order) => {
+            assert_eq!(shuffle_merge_order.alive_bitsets.len(), bytes_columns.len());
+            let mut term_bitsets: Vec<Option<BitSet>> = Vec::with_capacity(bytes_columns.len());
+            for (alive_bitset_opt, bytes_column_opt) in shuffle_merge_order
+                .alive_bitsets
+                .iter()
+                .zip(bytes_columns.iter())
+            {
+                match (alive_bitset_opt, bytes_column_opt) {
+                    (Some(alive_bitset), Some(bytes_column)) => {
+                        let term_bitset = compute_term_bitset(bytes_column, alive_bitset);
+                        term_bitsets.push(Some(term_bitset));
+                    }
+                    _ => {
+                        term_bitsets.push(None);
+                    }
+                }
+            }
+            let mut current_term_ord = 0;
+            while merged_terms.advance() {
+                let term_bytes: &[u8] = merged_terms.key();
+                if !is_term_present(&term_bitsets[..], &merged_terms) {
+                    continue;
+                }
+                sstable_builder.insert(term_bytes, &())?;
+                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
+                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, current_term_ord);
+                }
+                current_term_ord += 1;
+            }
+            sstable_builder.finish()?;
+        }
+    }
+    Ok(term_ord_mapping)
+}
+
+#[derive(Default, Debug)]
+struct TermOrdinalMapping {
+    per_segment_new_term_ordinals: Vec<Vec<TermOrdinal>>,
+}
+
+impl TermOrdinalMapping {
+    fn add_segment(&mut self, max_term_ord: usize) {
+        self.per_segment_new_term_ordinals
+            .push(vec![TermOrdinal::default(); max_term_ord]);
+    }
+
+    fn register_from_to(&mut self, segment_ord: usize, from_ord: TermOrdinal, to_ord: TermOrdinal) {
+        self.per_segment_new_term_ordinals[segment_ord][from_ord as usize] = to_ord;
+    }
+
+    fn get_segment(&self, segment_ord: u32) -> &[TermOrdinal] {
+        &(self.per_segment_new_term_ordinals[segment_ord as usize])[..]
+    }
+}

columnar/src/columnar/merge/merge_mapping.rs

@@ -0,0 +1,118 @@
+use std::ops::Range;
+
+use common::{BitSet, OwnedBytes, ReadOnlyBitSet};
+
+use crate::{ColumnarReader, RowAddr, RowId};
+
+pub struct StackMergeOrder {
+    // This does not start at 0. The first row is the number of
+    // rows in the first columnar.
+    cumulated_row_ids: Vec<RowId>,
+}
+
+impl StackMergeOrder {
+    pub fn stack(columnars: &[&ColumnarReader]) -> StackMergeOrder {
+        let mut cumulated_row_ids: Vec<RowId> = Vec::with_capacity(columnars.len());
+        let mut cumulated_row_id = 0;
+        for columnar in columnars {
+            cumulated_row_id += columnar.num_rows();
+            cumulated_row_ids.push(cumulated_row_id);
+        }
+        StackMergeOrder { cumulated_row_ids }
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.cumulated_row_ids.last().copied().unwrap_or(0)
+    }
+
+    pub fn offset(&self, columnar_id: usize) -> RowId {
+        if columnar_id == 0 {
+            return 0;
+        }
+        self.cumulated_row_ids[columnar_id - 1]
+    }
+
+    pub fn columnar_range(&self, columnar_id: usize) -> Range<RowId> {
+        self.offset(columnar_id)..self.offset(columnar_id + 1)
+    }
+}
+
+pub enum MergeRowOrder {
+    /// Columnar tables are simply stacked one above the other.
+    /// If the i-th columnar_readers has n_rows_i rows, then
+    /// in the resulting columnar,
+    /// rows [r0..n_row_0) contains the row of columnar_readers[0], in ordder
+    /// rows [n_row_0..n_row_0 + n_row_1 contains the row of columnar_readers[1], in order.
+    /// ..
+    /// No documents is deleted.
+    Stack(StackMergeOrder),
+    /// Some more complex mapping, that may interleaves rows from the different readers and
+    /// drop rows, or do both.
+    Shuffled(ShuffleMergeOrder),
+}
+
+impl From<StackMergeOrder> for MergeRowOrder {
+    fn from(stack_merge_order: StackMergeOrder) -> MergeRowOrder {
+        MergeRowOrder::Stack(stack_merge_order)
+    }
+}
+
+impl From<ShuffleMergeOrder> for MergeRowOrder {
+    fn from(shuffle_merge_order: ShuffleMergeOrder) -> MergeRowOrder {
+        MergeRowOrder::Shuffled(shuffle_merge_order)
+    }
+}
+
+impl MergeRowOrder {
+    pub fn num_rows(&self) -> RowId {
+        match self {
+            MergeRowOrder::Stack(stack_row_order) => stack_row_order.num_rows(),
+            MergeRowOrder::Shuffled(complex_mapping) => complex_mapping.num_rows(),
+        }
+    }
+}
+
+pub struct ShuffleMergeOrder {
+    pub new_row_id_to_old_row_id: Vec<RowAddr>,
+    pub alive_bitsets: Vec<Option<ReadOnlyBitSet>>,
+}
+
+impl ShuffleMergeOrder {
+    pub fn for_test(
+        segment_num_rows: &[RowId],
+        new_row_id_to_old_row_id: Vec<RowAddr>,
+    ) -> ShuffleMergeOrder {
+        let mut alive_bitsets: Vec<BitSet> = segment_num_rows
+            .iter()
+            .map(|&num_rows| BitSet::with_max_value(num_rows))
+            .collect();
+        for &RowAddr {
+            segment_ord,
+            row_id,
+        } in &new_row_id_to_old_row_id
+        {
+            alive_bitsets[segment_ord as usize].insert(row_id);
+        }
+        let alive_bitsets: Vec<Option<ReadOnlyBitSet>> = alive_bitsets
+            .into_iter()
+            .map(|alive_bitset| {
+                let mut buffer = Vec::new();
+                alive_bitset.serialize(&mut buffer).unwrap();
+                let data = OwnedBytes::new(buffer);
+                Some(ReadOnlyBitSet::open(data))
+            })
+            .collect();
+        ShuffleMergeOrder {
+            new_row_id_to_old_row_id,
+            alive_bitsets,
+        }
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.new_row_id_to_old_row_id.len() as RowId
+    }
+
+    pub fn iter_new_to_old_row_addrs(&self) -> impl Iterator<Item = RowAddr> + '_ {
+        self.new_row_id_to_old_row_id.iter().copied()
+    }
+}

columnar/src/columnar/merge/mod.rs

@@ -0,0 +1,375 @@
+mod merge_dict_column;
+mod merge_mapping;
+mod term_merger;
+
+use std::collections::{BTreeMap, HashMap, HashSet};
+use std::io;
+use std::net::Ipv6Addr;
+use std::sync::Arc;
+
+pub use merge_mapping::{MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
+
+use super::writer::ColumnarSerializer;
+use crate::column::{serialize_column_mappable_to_u128, serialize_column_mappable_to_u64};
+use crate::column_values::MergedColumnValues;
+use crate::columnar::merge::merge_dict_column::merge_bytes_or_str_column;
+use crate::columnar::writer::CompatibleNumericalTypes;
+use crate::columnar::ColumnarReader;
+use crate::dynamic_column::DynamicColumn;
+use crate::{
+    BytesColumn, Column, ColumnIndex, ColumnType, ColumnValues, NumericalType, NumericalValue,
+};
+
+/// Column types are grouped into different categories.
+/// After merge, all columns belonging to the same category are coerced to
+/// the same column type.
+///
+/// In practise, today, only Numerical colummns are coerced into one type today.
+///
+/// See also [README.md].
+#[derive(Copy, Clone, Eq, PartialEq, Hash, Debug)]
+enum ColumnTypeCategory {
+    Bool,
+    Str,
+    Numerical,
+    DateTime,
+    Bytes,
+    IpAddr,
+}
+
+impl From<ColumnType> for ColumnTypeCategory {
+    fn from(column_type: ColumnType) -> Self {
+        match column_type {
+            ColumnType::I64 => ColumnTypeCategory::Numerical,
+            ColumnType::U64 => ColumnTypeCategory::Numerical,
+            ColumnType::F64 => ColumnTypeCategory::Numerical,
+            ColumnType::Bytes => ColumnTypeCategory::Bytes,
+            ColumnType::Str => ColumnTypeCategory::Str,
+            ColumnType::Bool => ColumnTypeCategory::Bool,
+            ColumnType::IpAddr => ColumnTypeCategory::IpAddr,
+            ColumnType::DateTime => ColumnTypeCategory::DateTime,
+        }
+    }
+}
+
+/// Merge several columnar table together.
+///
+/// If several columns with the same name are conflicting with the numerical types in the
+/// input columnars, the first type compatible out of i64, u64, f64 in that order will be used.
+///
+/// `require_columns` makes it possible to ensure that some columns will be present in the
+/// resulting columnar. When a required column is a numerical column type, one of two things can
+/// happen:
+/// - If the required column type is compatible with all of the input columnar, the resulsting
+///   merged
+/// columnar will simply coerce the input column and use the required column type.
+/// - If the required column type is incompatible with one of the input columnar, the merged
+/// will fail with an InvalidData error.
+///
+/// `merge_row_order` makes it possible to remove or reorder row in the resulting
+/// `Columnar` table.
+///
+/// Reminder: a string and a numerical column may bare the same column name. This is not
+/// considered a conflict.
+pub fn merge_columnar(
+    columnar_readers: &[&ColumnarReader],
+    required_columns: &[(String, ColumnType)],
+    merge_row_order: MergeRowOrder,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let mut serializer = ColumnarSerializer::new(output);
+
+    let columns_to_merge = group_columns_for_merge(columnar_readers, required_columns)?;
+    for ((column_name, column_type), columns) in columns_to_merge {
+        let mut column_serializer =
+            serializer.serialize_column(column_name.as_bytes(), column_type);
+        merge_column(
+            column_type,
+            columns,
+            &merge_row_order,
+            &mut column_serializer,
+        )?;
+    }
+    serializer.finalize(merge_row_order.num_rows())?;
+
+    Ok(())
+}
+
+fn dynamic_column_to_u64_monotonic(dynamic_column: DynamicColumn) -> Option<Column<u64>> {
+    match dynamic_column {
+        DynamicColumn::Bool(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::I64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::U64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::F64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::DateTime(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::IpAddr(_) | DynamicColumn::Bytes(_) | DynamicColumn::Str(_) => None,
+    }
+}
+
+fn merge_column(
+    column_type: ColumnType,
+    columns: Vec<Option<DynamicColumn>>,
+    merge_row_order: &MergeRowOrder,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    match column_type {
+        ColumnType::I64
+        | ColumnType::U64
+        | ColumnType::F64
+        | ColumnType::DateTime
+        | ColumnType::Bool => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues>>> =
+                Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                if let Some(Column { idx, values }) =
+                    dynamic_column_opt.and_then(dynamic_column_to_u64_monotonic)
+                {
+                    column_indexes.push(Some(idx));
+                    column_values.push(Some(values));
+                } else {
+                    column_indexes.push(None);
+                    column_values.push(None);
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values[..],
+                merge_row_order,
+            };
+            serialize_column_mappable_to_u64(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::IpAddr => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues<Ipv6Addr>>>> =
+                Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                if let Some(DynamicColumn::IpAddr(Column { idx, values })) = dynamic_column_opt {
+                    column_indexes.push(Some(idx));
+                    column_values.push(Some(values));
+                } else {
+                    column_indexes.push(None);
+                    column_values.push(None);
+                }
+            }
+
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values,
+                merge_row_order,
+            };
+
+            serialize_column_mappable_to_u128(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::Bytes | ColumnType::Str => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut bytes_columns: Vec<Option<BytesColumn>> = Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                match dynamic_column_opt {
+                    Some(DynamicColumn::Str(str_column)) => {
+                        column_indexes.push(Some(str_column.term_ord_column.idx.clone()));
+                        bytes_columns.push(Some(str_column.into()));
+                    }
+                    Some(DynamicColumn::Bytes(bytes_column)) => {
+                        column_indexes.push(Some(bytes_column.term_ord_column.idx.clone()));
+                        bytes_columns.push(Some(bytes_column));
+                    }
+                    _ => {
+                        column_indexes.push(None);
+                        bytes_columns.push(None);
+                    }
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            merge_bytes_or_str_column(merged_column_index, &bytes_columns, merge_row_order, wrt)?;
+        }
+    }
+    Ok(())
+}
+
+struct GroupedColumns {
+    required_column_type: Option<ColumnType>,
+    columns: Vec<Option<DynamicColumn>>,
+    column_category: ColumnTypeCategory,
+}
+
+impl GroupedColumns {
+    fn for_category(column_category: ColumnTypeCategory, num_columnars: usize) -> Self {
+        GroupedColumns {
+            required_column_type: None,
+            columns: vec![None; num_columnars],
+            column_category,
+        }
+    }
+
+    /// Set the dynamic column for a given columnar.
+    fn set_column(&mut self, columnar_id: usize, column: DynamicColumn) {
+        self.columns[columnar_id] = Some(column);
+    }
+
+    /// Force the existence of a column, as well as its type.
+    fn require_type(&mut self, required_type: ColumnType) -> io::Result<()> {
+        if let Some(existing_required_type) = self.required_column_type {
+            if existing_required_type == required_type {
+                // This was just a duplicate in the `required_columns`.
+                // Nothing to do.
+                return Ok(());
+            } else {
+                return Err(io::Error::new(
+                    io::ErrorKind::InvalidInput,
+                    "Required column conflicts with another required column of the same type \
+                     category.",
+                ));
+            }
+        }
+        self.required_column_type = Some(required_type);
+        Ok(())
+    }
+
+    /// Returns the column type after merge.
+    ///
+    /// This method does not check if the column types can actually be coerced to
+    /// this type.
+    fn column_type_after_merge(&self) -> ColumnType {
+        if let Some(required_type) = self.required_column_type {
+            return required_type;
+        }
+        let column_type: HashSet<ColumnType> = self
+            .columns
+            .iter()
+            .flatten()
+            .map(|column| column.column_type())
+            .collect();
+        if column_type.len() == 1 {
+            return column_type.into_iter().next().unwrap();
+        }
+        // At the moment, only the numerical categorical column type has more than one possible
+        // column type.
+        assert_eq!(self.column_category, ColumnTypeCategory::Numerical);
+        merged_numerical_columns_type(self.columns.iter().flatten()).into()
+    }
+}
+
+/// Returns the type of the merged numerical column.
+///
+/// This function picks the first numerical type out of i64, u64, f64 (order matters
+/// here), that is compatible with all the `columns`.
+///
+/// # Panics
+/// Panics if one of the column is not numerical.
+fn merged_numerical_columns_type<'a>(
+    columns: impl Iterator<Item = &'a DynamicColumn>,
+) -> NumericalType {
+    let mut compatible_numerical_types = CompatibleNumericalTypes::default();
+    for column in columns {
+        let (min_value, max_value) =
+            min_max_if_numerical(column).expect("All columns re required to be numerical");
+        compatible_numerical_types.accept_value(min_value);
+        compatible_numerical_types.accept_value(max_value);
+    }
+    compatible_numerical_types.to_numerical_type()
+}
+
+#[allow(clippy::type_complexity)]
+fn group_columns_for_merge(
+    columnar_readers: &[&ColumnarReader],
+    required_columns: &[(String, ColumnType)],
+) -> io::Result<BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>>> {
+    // Each column name may have multiple types of column associated.
+    // For merging we are interested in the same column type category since they can be merged.
+    let mut columns_grouped: HashMap<(String, ColumnTypeCategory), GroupedColumns> = HashMap::new();
+
+    for &(ref column_name, column_type) in required_columns {
+        columns_grouped
+            .entry((column_name.clone(), column_type.into()))
+            .or_insert_with(|| {
+                GroupedColumns::for_category(column_type.into(), columnar_readers.len())
+            })
+            .require_type(column_type)?;
+    }
+
+    for (columnar_id, columnar_reader) in columnar_readers.iter().enumerate() {
+        let column_name_and_handle = columnar_reader.list_columns()?;
+        for (column_name, handle) in column_name_and_handle {
+            let column_category: ColumnTypeCategory = handle.column_type().into();
+            let column = handle.open()?;
+            columns_grouped
+                .entry((column_name, column_category))
+                .or_insert_with(|| {
+                    GroupedColumns::for_category(column_category, columnar_readers.len())
+                })
+                .set_column(columnar_id, column);
+        }
+    }
+
+    let mut merge_columns: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        Default::default();
+
+    for ((column_name, _), mut grouped_columns) in columns_grouped {
+        let column_type = grouped_columns.column_type_after_merge();
+        coerce_columns(column_type, &mut grouped_columns.columns)?;
+        merge_columns.insert((column_name, column_type), grouped_columns.columns);
+    }
+
+    Ok(merge_columns)
+}
+
+fn coerce_columns(
+    column_type: ColumnType,
+    columns: &mut [Option<DynamicColumn>],
+) -> io::Result<()> {
+    for column_opt in columns.iter_mut() {
+        if let Some(column) = column_opt.take() {
+            *column_opt = Some(coerce_column(column_type, column)?);
+        }
+    }
+    Ok(())
+}
+
+fn coerce_column(column_type: ColumnType, column: DynamicColumn) -> io::Result<DynamicColumn> {
+    if let Some(numerical_type) = column_type.numerical_type() {
+        column
+            .coerce_numerical(numerical_type)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, ""))
+    } else {
+        if column.column_type() != column_type {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidInput,
+                format!(
+                    "Cannot coerce column of type `{:?}` to `{column_type:?}`",
+                    column.column_type()
+                ),
+            ));
+        }
+        Ok(column)
+    }
+}
+
+/// Returns the (min, max) of a column provided it is numerical (i64, u64. f64).
+///
+/// The min and the max are simply the numerical value as defined by `ColumnValue::min_value()`,
+/// and `ColumnValue::max_value()`.
+///
+/// It is important to note that these values are only guaranteed to be lower/upper bound
+/// (as opposed to min/max value).
+/// If a column is empty, the min and max values are currently set to 0.
+fn min_max_if_numerical(column: &DynamicColumn) -> Option<(NumericalValue, NumericalValue)> {
+    match column {
+        DynamicColumn::I64(column) => Some((column.min_value().into(), column.max_value().into())),
+        DynamicColumn::U64(column) => Some((column.min_value().into(), column.min_value().into())),
+        DynamicColumn::F64(column) => Some((column.min_value().into(), column.min_value().into())),
+        DynamicColumn::Bool(_)
+        | DynamicColumn::IpAddr(_)
+        | DynamicColumn::DateTime(_)
+        | DynamicColumn::Bytes(_)
+        | DynamicColumn::Str(_) => None,
+    }
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/columnar/merge/term_merger.rs

@@ -0,0 +1,107 @@
+use std::cmp::Ordering;
+use std::collections::BinaryHeap;
+
+use sstable::TermOrdinal;
+
+use crate::Streamer;
+
+pub struct HeapItem<'a> {
+    pub streamer: Streamer<'a>,
+    pub segment_ord: usize,
+}
+
+impl<'a> PartialEq for HeapItem<'a> {
+    fn eq(&self, other: &Self) -> bool {
+        self.segment_ord == other.segment_ord
+    }
+}
+
+impl<'a> Eq for HeapItem<'a> {}
+
+impl<'a> PartialOrd for HeapItem<'a> {
+    fn partial_cmp(&self, other: &HeapItem<'a>) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl<'a> Ord for HeapItem<'a> {
+    fn cmp(&self, other: &HeapItem<'a>) -> Ordering {
+        (&other.streamer.key(), &other.segment_ord).cmp(&(&self.streamer.key(), &self.segment_ord))
+    }
+}
+
+/// Given a list of sorted term streams,
+/// returns an iterator over sorted unique terms.
+///
+/// The item yield is actually a pair with
+/// - the term
+/// - a slice with the ordinal of the segments containing
+/// the terms.
+pub struct TermMerger<'a> {
+    heap: BinaryHeap<HeapItem<'a>>,
+    current_streamers: Vec<HeapItem<'a>>,
+}
+
+impl<'a> TermMerger<'a> {
+    /// Stream of merged term dictionary
+    pub fn new(streams: Vec<Streamer<'a>>) -> TermMerger<'a> {
+        TermMerger {
+            heap: BinaryHeap::new(),
+            current_streamers: streams
+                .into_iter()
+                .enumerate()
+                .map(|(ord, streamer)| HeapItem {
+                    streamer,
+                    segment_ord: ord,
+                })
+                .collect(),
+        }
+    }
+
+    pub(crate) fn matching_segments<'b: 'a>(
+        &'b self,
+    ) -> impl 'b + Iterator<Item = (usize, TermOrdinal)> {
+        self.current_streamers
+            .iter()
+            .map(|heap_item| (heap_item.segment_ord, heap_item.streamer.term_ord()))
+    }
+
+    fn advance_segments(&mut self) {
+        let streamers = &mut self.current_streamers;
+        let heap = &mut self.heap;
+        for mut heap_item in streamers.drain(..) {
+            if heap_item.streamer.advance() {
+                heap.push(heap_item);
+            }
+        }
+    }
+
+    /// Advance the term iterator to the next term.
+    /// Returns true if there is indeed another term
+    /// False if there is none.
+    pub fn advance(&mut self) -> bool {
+        self.advance_segments();
+        if let Some(head) = self.heap.pop() {
+            self.current_streamers.push(head);
+            while let Some(next_streamer) = self.heap.peek() {
+                if self.current_streamers[0].streamer.key() != next_streamer.streamer.key() {
+                    break;
+                }
+                let next_heap_it = self.heap.pop().unwrap(); // safe : we peeked beforehand
+                self.current_streamers.push(next_heap_it);
+            }
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Returns the current term.
+    ///
+    /// This method may be called
+    /// if and only if advance() has been called before
+    /// and "true" was returned.
+    pub fn key(&self) -> &[u8] {
+        self.current_streamers[0].streamer.key()
+    }
+}

columnar/src/columnar/merge/tests.rs

@@ -0,0 +1,318 @@
+use super::*;
+use crate::{Cardinality, ColumnarWriter, HasAssociatedColumnType, RowId};
+
+fn make_columnar<T: Into<NumericalValue> + HasAssociatedColumnType + Copy>(
+    column_name: &str,
+    vals: &[T],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_column_type(column_name, T::column_type(), false);
+    for (row_id, val) in vals.iter().copied().enumerate() {
+        dataframe_writer.record_numerical(row_id as RowId, column_name, val.into());
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(vals.len() as RowId, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_column_coercion_to_u64() {
+    // i64 type
+    let columnar1 = make_columnar("numbers", &[1i64]);
+    // u64 type
+    let columnar2 = make_columnar("numbers", &[u64::MAX]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2], &[]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_column_no_coercion_if_all_the_same() {
+    let columnar1 = make_columnar("numbers", &[1u64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2], &[]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_column_coercion_to_i64() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2], &[]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::I64)));
+}
+
+#[test]
+fn test_impossible_coercion_returns_an_error() {
+    let columnar1 = make_columnar("numbers", &[u64::MAX]);
+    let group_error =
+        group_columns_for_merge(&[&columnar1], &[("numbers".to_string(), ColumnType::I64)])
+            .map(|_| ())
+            .unwrap_err();
+    assert_eq!(group_error.kind(), io::ErrorKind::InvalidInput);
+}
+
+#[test]
+fn test_group_columns_with_required_column() {
+    let columnar1 = make_columnar("numbers", &[1i64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(
+            &[&columnar1, &columnar2],
+            &[("numbers".to_string(), ColumnType::U64)],
+        )
+        .unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_group_columns_required_column_with_no_existing_columns() {
+    let columnar1 = make_columnar("numbers", &[2u64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(
+            &[&columnar1, &columnar2],
+            &[("required_col".to_string(), ColumnType::Str)],
+        )
+        .unwrap();
+    assert_eq!(column_map.len(), 2);
+    let columns = column_map
+        .get(&("required_col".to_string(), ColumnType::Str))
+        .unwrap();
+    assert_eq!(columns.len(), 2);
+    assert!(columns[0].is_none());
+    assert!(columns[1].is_none());
+}
+
+#[test]
+fn test_group_columns_required_column_is_above_all_columns_have_the_same_type_rule() {
+    let columnar1 = make_columnar("numbers", &[2i64]);
+    let columnar2 = make_columnar("numbers", &[2i64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(
+            &[&columnar1, &columnar2],
+            &[("numbers".to_string(), ColumnType::U64)],
+        )
+        .unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_missing_column() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers2", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2], &[]).unwrap();
+    assert_eq!(column_map.len(), 2);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::I64)));
+    {
+        let columns = column_map
+            .get(&("numbers".to_string(), ColumnType::I64))
+            .unwrap();
+        assert!(columns[0].is_some());
+        assert!(columns[1].is_none());
+    }
+    {
+        let columns = column_map
+            .get(&("numbers2".to_string(), ColumnType::U64))
+            .unwrap();
+        assert!(columns[0].is_none());
+        assert!(columns[1].is_some());
+    }
+}
+
+fn make_numerical_columnar_multiple_columns(
+    columns: &[(&str, &[&[NumericalValue]])],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_numerical(row_id as u32, column_name, *val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+fn make_byte_columnar_multiple_columns(columns: &[(&str, &[&[&[u8]]])]) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_bytes(row_id as u32, column_name, val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+fn make_text_columnar_multiple_columns(columns: &[(&str, &[&[&str]])]) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_str(row_id as u32, column_name, val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_merge_columnar_numbers() {
+    let columnar1 =
+        make_numerical_columnar_multiple_columns(&[("numbers", &[&[NumericalValue::from(-1f64)]])]);
+    let columnar2 = make_numerical_columnar_multiple_columns(&[(
+        "numbers",
+        &[&[], &[NumericalValue::from(-3f64)]],
+    )]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("numbers").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::F64(vals) = dynamic_column else { panic!() };
+    assert_eq!(vals.get_cardinality(), Cardinality::Optional);
+    assert_eq!(vals.first(0u32), Some(-1f64));
+    assert_eq!(vals.first(1u32), None);
+    assert_eq!(vals.first(2u32), Some(-3f64));
+}
+
+#[test]
+fn test_merge_columnar_texts() {
+    let columnar1 = make_text_columnar_multiple_columns(&[("texts", &[&["a"]])]);
+    let columnar2 = make_text_columnar_multiple_columns(&[("texts", &[&[], &["b"]])]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("texts").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Str(vals) = dynamic_column else { panic!() };
+    let get_str_for_ord = |ord| {
+        let mut out = String::new();
+        vals.ord_to_str(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 2);
+    assert_eq!(get_str_for_ord(0), "a");
+    assert_eq!(get_str_for_ord(1), "b");
+
+    let get_str_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = String::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_str(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_str_for_row(0), "a");
+    assert_eq!(get_str_for_row(1), "");
+    assert_eq!(get_str_for_row(2), "b");
+}
+
+#[test]
+fn test_merge_columnar_byte() {
+    let columnar1 = make_byte_columnar_multiple_columns(&[("bytes", &[&[b"bbbb"], &[b"baaa"]])]);
+    let columnar2 = make_byte_columnar_multiple_columns(&[("bytes", &[&[], &[b"a"]])]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        &[],
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 4);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("bytes").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Bytes(vals) = dynamic_column else { panic!() };
+    let get_bytes_for_ord = |ord| {
+        let mut out = Vec::new();
+        vals.ord_to_bytes(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 3);
+    assert_eq!(get_bytes_for_ord(0), b"a");
+    assert_eq!(get_bytes_for_ord(1), b"baaa");
+    assert_eq!(get_bytes_for_ord(2), b"bbbb");
+
+    let get_bytes_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = Vec::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_bytes(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_bytes_for_row(0), b"bbbb");
+    assert_eq!(get_bytes_for_row(1), b"baaa");
+    assert_eq!(get_bytes_for_row(2), b"");
+    assert_eq!(get_bytes_for_row(3), b"a");
+}

columnar/src/columnar/mod.rs

@@ -0,0 +1,10 @@
+mod column_type;
+mod format_version;
+mod merge;
+mod reader;
+mod writer;
+
+pub use column_type::{ColumnType, HasAssociatedColumnType};
+pub use merge::{merge_columnar, MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
+pub use reader::ColumnarReader;
+pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -0,0 +1,192 @@
+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;
+use crate::RowId;
+
+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.
+#[derive(Clone)]
+pub struct ColumnarReader {
+    column_dictionary: Dictionary<RangeSSTable>,
+    column_data: FileSlice,
+    num_rows: RowId,
+}
+
+/// Functions by both the async/sync code listing columns.
+/// It takes a stream from the column sstable and return the list of
+/// `DynamicColumn` available in it.
+fn read_all_columns_in_stream(
+    mut stream: sstable::Streamer<'_, RangeSSTable>,
+    column_data: &FileSlice,
+) -> io::Result<Vec<DynamicColumnHandle>> {
+    let mut results = Vec::new();
+    while stream.advance() {
+        let key_bytes: &[u8] = stream.key();
+        let Some(column_code) = key_bytes.last().copied() else {
+            return Err(io_invalid_data("Empty column name.".to_string()));
+        };
+        let column_type = ColumnType::try_from_code(column_code)
+            .map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
+        let range = stream.value();
+        let file_slice = 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)
+}
+
+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>() + 4 + format_version::VERSION_FOOTER_NUM_BYTES);
+        let footer_bytes = footer_slice.read_bytes()?;
+        let sstable_len = u64::deserialize(&mut &footer_bytes[0..8])?;
+        let num_rows = u32::deserialize(&mut &footer_bytes[8..12])?;
+        let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
+            footer_bytes[12..].try_into().unwrap();
+        let _version = format_version::parse_footer(version_footer_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,
+            num_rows,
+        })
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
+    // TODO Add unit tests
+    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 =
+                // The last two bytes are respectively the 0u8 separator and the column_type.
+                String::from_utf8_lossy(&key_bytes[..key_bytes.len() - 2]).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)
+    }
+
+    fn stream_for_column_range(&self, column_name: &str) -> sstable::StreamerBuilder<RangeSSTable> {
+        // 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 dictionary.
+        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);
+        self.column_dictionary
+            .range()
+            .ge(start_key.as_bytes())
+            .lt(end_key.as_bytes())
+    }
+
+    pub async fn read_columns_async(
+        &self,
+        column_name: &str,
+    ) -> io::Result<Vec<DynamicColumnHandle>> {
+        let stream = self
+            .stream_for_column_range(column_name)
+            .into_stream_async()
+            .await?;
+        read_all_columns_in_stream(stream, &self.column_data)
+    }
+
+    /// 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.
+    pub fn read_columns(&self, column_name: &str) -> io::Result<Vec<DynamicColumnHandle>> {
+        let stream = self.stream_for_column_range(column_name).into_stream()?;
+        read_all_columns_in_stream(stream, &self.column_data)
+    }
+
+    /// Return the number of columns in the columnar.
+    pub fn num_columns(&self) -> usize {
+        self.column_dictionary.num_terms()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::{ColumnType, ColumnarReader, ColumnarWriter};
+
+    #[test]
+    fn test_list_columns() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("col1", ColumnType::Str, false);
+        columnar_writer.record_column_type("col2", ColumnType::U64, false);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(1, None, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 2);
+        assert_eq!(&columns[0].0, "col1");
+        assert_eq!(columns[0].1.column_type(), ColumnType::Str);
+        assert_eq!(&columns[1].0, "col2");
+        assert_eq!(columns[1].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    fn test_list_columns_strict_typing_prevents_coercion() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1u64);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(2, None, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 1);
+        assert_eq!(&columns[0].0, "count");
+        assert_eq!(columns[0].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    #[should_panic(expected = "Input type forbidden")]
+    fn test_list_columns_strict_typing_panics_on_wrong_types() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1i64);
+    }
+}

columnar/src/columnar/writer/column_operation.rs

@@ -0,0 +1,360 @@
+use std::net::Ipv6Addr;
+
+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, 6>(self.len) | place_bits::<6, 8>(self.op_type.to_code())
+    }
+
+    fn try_from_code(code: u8) -> Result<Self, InvalidData> {
+        let len = select_bits::<0, 6>(code);
+        let typ_code = select_bits::<6, 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
+    }
+}
+
+impl SymbolValue for Ipv6Addr {
+    fn serialize(self, buffer: &mut [u8]) -> u8 {
+        buffer[0..16].copy_from_slice(&self.octets());
+        16
+    }
+
+    fn deserialize(bytes: &[u8]) -> Self {
+        let octets: [u8; 16] = bytes[0..16].try_into().unwrap();
+        Ipv6Addr::from(octets)
+    }
+}
+
+#[derive(Default)]
+struct MiniBuffer {
+    pub bytes: [u8; 17],
+    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(), 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() }, 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 +1,363 @@
+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,
+        old_to_new_ids_opt: Option<&[RowId]>,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<V>> + 'a {
+        buffer.clear();
+        self.values.read_to_end(arena, buffer);
+        if let Some(old_to_new_ids) = old_to_new_ids_opt {
+            // TODO avoid the extra deserialization / serialization.
+            let mut sorted_ops: Vec<(RowId, ColumnOperation<V>)> = Vec::new();
+            let mut new_doc = 0u32;
+            let mut cursor = &buffer[..];
+            for op in std::iter::from_fn(|| ColumnOperation::<V>::deserialize(&mut cursor)) {
+                if let ColumnOperation::NewDoc(doc) = &op {
+                    new_doc = old_to_new_ids[*doc as usize];
+                    sorted_ops.push((new_doc, ColumnOperation::NewDoc(new_doc)));
+                } else {
+                    sorted_ops.push((new_doc, op));
+                }
+            }
+            // stable sort is crucial here.
+            sorted_ops.sort_by_key(|(new_doc_id, _)| *new_doc_id);
+            buffer.clear();
+            for (_, op) in sorted_ops {
+                buffer.extend_from_slice(op.serialize().as_ref());
+            }
+        }
+        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)]
+pub(crate) 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 {
+    pub 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
+            }
+        }
+    }
+
+    pub 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,
+                    "Input type forbidden. This column has been forced to type {typ:?}, received \
+                     {numerical_value:?}"
+                );
+            }
+        }
+    }
+
+    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 numerical_type(&self) -> NumericalType {
+        self.compatible_numerical_types.to_numerical_type()
+    }
+
+    pub fn cardinality(&self, num_docs: RowId) -> Cardinality {
+        self.column_writer.get_cardinality(num_docs)
+    }
+
+    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,
+        old_to_new_ids: Option<&[RowId]>,
+        buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<NumericalValue>> + 'a {
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, buffer)
+    }
+}
+
+#[derive(Copy, Clone)]
+pub(crate) struct StrOrBytesColumnWriter {
+    pub(crate) dictionary_id: u32,
+    pub(crate) column_writer: ColumnWriter,
+    // If true, when facing a multivalued cardinality,
+    // values associated to a given document will be sorted.
+    //
+    // This is useful for facets.
+    //
+    // If false, the order of appearance in the document will be
+    // observed.
+    pub(crate) sort_values_within_row: bool,
+}
+
+impl StrOrBytesColumnWriter {
+    pub(crate) fn with_dictionary_id(dictionary_id: u32) -> StrOrBytesColumnWriter {
+        StrOrBytesColumnWriter {
+            dictionary_id,
+            column_writer: Default::default(),
+            sort_values_within_row: false,
+        }
+    }
+
+    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,
+        old_to_new_ids: Option<&[RowId]>,
+        byte_buffer: &'a mut Vec<u8>,
+    ) -> impl Iterator<Item = ColumnOperation<UnorderedId>> + 'a {
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, 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 +1,848 @@
+mod column_operation;
+mod column_writers;
+mod serializer;
+mod value_index;
+
+use std::io;
+use std::net::Ipv6Addr;
+
+use column_operation::ColumnOperation;
+pub(crate) use column_writers::CompatibleNumericalTypes;
+use common::CountingWriter;
+pub(crate) use serializer::ColumnarSerializer;
+use stacker::{Addr, ArenaHashMap, MemoryArena};
+
+use crate::column_index::SerializableColumnIndex;
+use crate::column_values::{
+    ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64, VecColumn,
+};
+use crate::columnar::column_type::ColumnType;
+use crate::columnar::writer::column_writers::{
+    ColumnWriter, NumericalColumnWriter, StrOrBytesColumnWriter,
+};
+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,
+    u64_values: Vec<u64>,
+    ip_addr_values: Vec<Ipv6Addr>,
+}
+
+/// 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, None, &mut wrt).unwrap();
+/// ```
+#[derive(Default)]
+pub struct ColumnarWriter {
+    numerical_field_hash_map: ArenaHashMap,
+    datetime_field_hash_map: ArenaHashMap,
+    bool_field_hash_map: ArenaHashMap,
+    ip_addr_field_hash_map: ArenaHashMap,
+    bytes_field_hash_map: ArenaHashMap,
+    str_field_hash_map: ArenaHashMap,
+    arena: MemoryArena,
+    // Dictionaries used to store dictionary-encoded values.
+    dictionaries: Vec<DictionaryBuilder>,
+    buffers: SpareBuffers,
+}
+
+#[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 mem_usage(&self) -> usize {
+        // TODO add dictionary builders.
+        self.arena.mem_usage()
+            + self.numerical_field_hash_map.mem_usage()
+            + self.bool_field_hash_map.mem_usage()
+            + self.bytes_field_hash_map.mem_usage()
+            + self.str_field_hash_map.mem_usage()
+            + self.ip_addr_field_hash_map.mem_usage()
+            + self.datetime_field_hash_map.mem_usage()
+    }
+
+    /// Returns the list of doc ids from 0..num_docs sorted by the `sort_field`
+    /// column.
+    ///
+    /// If the column is multivalued, use the first value for scoring.
+    /// If no value is associated to a specific row, the document is assigned
+    /// the lowest possible score.
+    ///
+    /// The sort applied is stable.
+    pub fn sort_order(&self, sort_field: &str, num_docs: RowId, reversed: bool) -> Vec<u32> {
+        let Some(numerical_col_writer) =
+            self.numerical_field_hash_map.get::<NumericalColumnWriter>(sort_field.as_bytes()) else {
+                return Vec::new();
+        };
+        let mut symbols_buffer = Vec::new();
+        let mut values = Vec::new();
+        let mut last_doc_opt: Option<RowId> = None;
+        for op in numerical_col_writer.operation_iterator(&self.arena, None, &mut symbols_buffer) {
+            match op {
+                ColumnOperation::NewDoc(doc) => {
+                    last_doc_opt = Some(doc);
+                }
+                ColumnOperation::Value(numerical_value) => {
+                    if let Some(last_doc) = last_doc_opt {
+                        let score: f32 = f64::coerce(numerical_value) as f32;
+                        values.push((score, last_doc));
+                    }
+                }
+            }
+        }
+        for doc in values.len() as u32..num_docs {
+            values.push((0.0f32, doc));
+        }
+        values.sort_by(|(left_score, _), (right_score, _)| {
+            if reversed {
+                right_score.partial_cmp(left_score).unwrap()
+            } else {
+                left_score.partial_cmp(right_score).unwrap()
+            }
+        });
+        values.into_iter().map(|(_score, doc)| doc).collect()
+    }
+
+    /// Records a column type. This is useful to bypass the coercion process,
+    /// makes sure the empty is present in the resulting columnar, or set
+    /// the `sort_values_within_row`.
+    ///
+    /// `sort_values_within_row` is only allowed for `Bytes` or `Str` columns.
+    pub fn record_column_type(
+        &mut self,
+        column_name: &str,
+        column_type: ColumnType,
+        sort_values_within_row: bool,
+    ) {
+        if sort_values_within_row {
+            assert!(
+                column_type == ColumnType::Bytes || column_type == ColumnType::Str,
+                "sort_values_within_row is only allowed for Bytes and Str columns",
+            );
+        }
+        match column_type {
+            ColumnType::Str | ColumnType::Bytes => {
+                let (hash_map, dictionaries) = (
+                    if column_type == ColumnType::Str {
+                        &mut self.str_field_hash_map
+                    } else {
+                        &mut self.bytes_field_hash_map
+                    },
+                    &mut self.dictionaries,
+                );
+                mutate_or_create_column(
+                    hash_map,
+                    column_name,
+                    |column_opt: Option<StrOrBytesColumnWriter>| {
+                        let mut column_writer = if let Some(column_writer) = column_opt {
+                            column_writer
+                        } else {
+                            let dictionary_id = dictionaries.len() as u32;
+                            dictionaries.push(DictionaryBuilder::default());
+                            StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
+                        };
+                        column_writer.sort_values_within_row = sort_values_within_row;
+                        column_writer
+                    },
+                );
+            }
+            ColumnType::Bool => {
+                mutate_or_create_column(
+                    &mut self.bool_field_hash_map,
+                    column_name,
+                    |column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
+                );
+            }
+            ColumnType::DateTime => {
+                mutate_or_create_column(
+                    &mut self.datetime_field_hash_map,
+                    column_name,
+                    |column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
+                );
+            }
+            ColumnType::I64 | ColumnType::F64 | ColumnType::U64 => {
+                let numerical_type = column_type.numerical_type().unwrap();
+                mutate_or_create_column(
+                    &mut self.numerical_field_hash_map,
+                    column_name,
+                    |column_opt: Option<NumericalColumnWriter>| {
+                        let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
+                        column.force_numerical_type(numerical_type);
+                        column
+                    },
+                );
+            }
+            ColumnType::IpAddr => mutate_or_create_column(
+                &mut self.ip_addr_field_hash_map,
+                column_name,
+                |column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
+            ),
+        }
+    }
+
+    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_ip_addr(&mut self, doc: RowId, column_name: &str, ip_addr: Ipv6Addr) {
+        assert!(
+            !column_name.as_bytes().contains(&0u8),
+            "key may not contain the 0 byte"
+        );
+        let (hash_map, arena) = (&mut self.ip_addr_field_hash_map, &mut self.arena);
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<ColumnWriter>| {
+                let mut column: ColumnWriter = column_opt.unwrap_or_default();
+                column.record(doc, ip_addr, 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_datetime(&mut self, doc: RowId, column_name: &str, datetime: common::DateTime) {
+        let (hash_map, arena) = (&mut self.datetime_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,
+                NumericalValue::I64(datetime.into_timestamp_micros()),
+                arena,
+            );
+            column
+        });
+    }
+
+    pub fn record_str(&mut self, doc: RowId, column_name: &str, value: &str) {
+        let (hash_map, arena, dictionaries) = (
+            &mut self.str_field_hash_map,
+            &mut self.arena,
+            &mut self.dictionaries,
+        );
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<StrOrBytesColumnWriter>| {
+                let mut column: StrOrBytesColumnWriter = column_opt.unwrap_or_else(|| {
+                    // Each column has its own dictionary
+                    let dictionary_id = dictionaries.len() as u32;
+                    dictionaries.push(DictionaryBuilder::default());
+                    StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
+                });
+                column.record_bytes(doc, value.as_bytes(), dictionaries, arena);
+                column
+            },
+        );
+    }
+
+    pub fn record_bytes(&mut self, doc: RowId, column_name: &str, value: &[u8]) {
+        assert!(
+            !column_name.as_bytes().contains(&0u8),
+            "key may not contain the 0 byte"
+        );
+        let (hash_map, arena, dictionaries) = (
+            &mut self.bytes_field_hash_map,
+            &mut self.arena,
+            &mut self.dictionaries,
+        );
+        hash_map.mutate_or_create(
+            column_name.as_bytes(),
+            |column_opt: Option<StrOrBytesColumnWriter>| {
+                let mut column: StrOrBytesColumnWriter = column_opt.unwrap_or_else(|| {
+                    // Each column has its own dictionary
+                    let dictionary_id = dictionaries.len() as u32;
+                    dictionaries.push(DictionaryBuilder::default());
+                    StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
+                });
+                column.record_bytes(doc, value, dictionaries, arena);
+                column
+            },
+        );
+    }
+    pub fn serialize(
+        &mut self,
+        num_docs: RowId,
+        old_to_new_row_ids: Option<&[RowId]>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()> {
+        let mut serializer = ColumnarSerializer::new(wrt);
+        let mut columns: Vec<(&[u8], ColumnType, Addr)> = self
+            .numerical_field_hash_map
+            .iter()
+            .map(|(column_name, addr, _)| {
+                let numerical_column_writer: NumericalColumnWriter =
+                    self.numerical_field_hash_map.read(addr);
+                let column_type = numerical_column_writer.numerical_type().into();
+                (column_name, column_type, addr)
+            })
+            .collect();
+        columns.extend(
+            self.bytes_field_hash_map
+                .iter()
+                .map(|(term, addr, _)| (term, ColumnType::Bytes, addr)),
+        );
+        columns.extend(
+            self.str_field_hash_map
+                .iter()
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Str, addr)),
+        );
+        columns.extend(
+            self.bool_field_hash_map
+                .iter()
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Bool, addr)),
+        );
+        columns.extend(
+            self.ip_addr_field_hash_map
+                .iter()
+                .map(|(column_name, addr, _)| (column_name, ColumnType::IpAddr, addr)),
+        );
+        columns.extend(
+            self.datetime_field_hash_map
+                .iter()
+                .map(|(column_name, addr, _)| (column_name, ColumnType::DateTime, addr)),
+        );
+        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, column_type, addr) in columns {
+            match column_type {
+                ColumnType::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, column_type);
+                    serialize_bool_column(
+                        cardinality,
+                        num_docs,
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnType::IpAddr => {
+                    let column_writer: ColumnWriter = self.ip_addr_field_hash_map.read(addr);
+                    let cardinality = column_writer.get_cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, ColumnType::IpAddr);
+                    serialize_ip_addr_column(
+                        cardinality,
+                        num_docs,
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnType::Bytes | ColumnType::Str => {
+                    let str_or_bytes_column_writer: StrOrBytesColumnWriter =
+                        if column_type == ColumnType::Bytes {
+                            self.bytes_field_hash_map.read(addr)
+                        } else {
+                            self.str_field_hash_map.read(addr)
+                        };
+                    let dictionary_builder =
+                        &dictionaries[str_or_bytes_column_writer.dictionary_id as usize];
+                    let cardinality = str_or_bytes_column_writer
+                        .column_writer
+                        .get_cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, column_type);
+                    serialize_bytes_or_str_column(
+                        cardinality,
+                        num_docs,
+                        str_or_bytes_column_writer.sort_values_within_row,
+                        dictionary_builder,
+                        str_or_bytes_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnType::F64 | ColumnType::I64 | ColumnType::U64 => {
+                    let numerical_column_writer: NumericalColumnWriter =
+                        self.numerical_field_hash_map.read(addr);
+                    let cardinality = numerical_column_writer.cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, column_type);
+                    let numerical_type = column_type.numerical_type().unwrap();
+                    serialize_numerical_column(
+                        cardinality,
+                        num_docs,
+                        numerical_type,
+                        numerical_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnType::DateTime => {
+                    let column_writer: ColumnWriter = self.datetime_field_hash_map.read(addr);
+                    let cardinality = column_writer.get_cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, ColumnType::DateTime);
+                    serialize_numerical_column(
+                        cardinality,
+                        num_docs,
+                        NumericalType::I64,
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+            };
+        }
+        serializer.finalize(num_docs)?;
+        Ok(())
+    }
+}
+
+// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
+// Column: [Column Index, Column Values, column index num bytes U32::LE]
+fn serialize_bytes_or_str_column(
+    cardinality: Cardinality,
+    num_docs: RowId,
+    sort_values_within_row: bool,
+    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),
+        }
+    });
+    send_to_serialize_column_mappable_to_u64(
+        operation_iterator,
+        cardinality,
+        num_docs,
+        sort_values_within_row,
+        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,
+        ..
+    } = buffers;
+    match numerical_type {
+        NumericalType::I64 => {
+            send_to_serialize_column_mappable_to_u64(
+                coerce_numerical_symbol::<i64>(op_iterator),
+                cardinality,
+                num_docs,
+                false,
+                value_index_builders,
+                u64_values,
+                wrt,
+            )?;
+        }
+        NumericalType::U64 => {
+            send_to_serialize_column_mappable_to_u64(
+                coerce_numerical_symbol::<u64>(op_iterator),
+                cardinality,
+                num_docs,
+                false,
+                value_index_builders,
+                u64_values,
+                wrt,
+            )?;
+        }
+        NumericalType::F64 => {
+            send_to_serialize_column_mappable_to_u64(
+                coerce_numerical_symbol::<f64>(op_iterator),
+                cardinality,
+                num_docs,
+                false,
+                value_index_builders,
+                u64_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,
+        u64_values,
+        ..
+    } = buffers;
+    send_to_serialize_column_mappable_to_u64(
+        column_operations_it.map(|bool_column_operation| match bool_column_operation {
+            ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
+            ColumnOperation::Value(bool_val) => ColumnOperation::Value(bool_val.to_u64()),
+        }),
+        cardinality,
+        num_docs,
+        false,
+        value_index_builders,
+        u64_values,
+        wrt,
+    )?;
+    Ok(())
+}
+
+fn serialize_ip_addr_column(
+    cardinality: Cardinality,
+    num_docs: RowId,
+    column_operations_it: impl Iterator<Item = ColumnOperation<Ipv6Addr>>,
+    buffers: &mut SpareBuffers,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    let SpareBuffers {
+        value_index_builders,
+        ip_addr_values,
+        ..
+    } = buffers;
+    send_to_serialize_column_mappable_to_u128(
+        column_operations_it,
+        cardinality,
+        num_docs,
+        value_index_builders,
+        ip_addr_values,
+        wrt,
+    )?;
+    Ok(())
+}
+
+fn send_to_serialize_column_mappable_to_u128<
+    T: Copy + Ord + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU128 + PartialOrd,
+>(
+    op_iterator: impl Iterator<Item = ColumnOperation<T>>,
+    cardinality: Cardinality,
+    num_rows: 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();
+    // TODO: split index and values
+    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_rows);
+            SerializableColumnIndex::Optional {
+                num_rows,
+                non_null_row_ids: 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_rows);
+            SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
+        }
+    };
+    crate::column::serialize_column_mappable_to_u128(
+        serializable_column_index,
+        &&values[..],
+        &mut wrt,
+    )?;
+    Ok(())
+}
+
+fn sort_values_within_row_in_place(multivalued_index: &[RowId], values: &mut [u64]) {
+    let mut start_index: usize = 0;
+    for end_index in multivalued_index.iter().copied() {
+        let end_index = end_index as usize;
+        values[start_index..end_index].sort_unstable();
+        start_index = end_index;
+    }
+}
+
+fn send_to_serialize_column_mappable_to_u64(
+    op_iterator: impl Iterator<Item = ColumnOperation<u64>>,
+    cardinality: Cardinality,
+    num_rows: RowId,
+    sort_values_within_row: bool,
+    value_index_builders: &mut PreallocatedIndexBuilders,
+    values: &mut Vec<u64>,
+    mut wrt: impl io::Write,
+) -> io::Result<()>
+where
+    for<'a> VecColumn<'a, u64>: ColumnValues<u64>,
+{
+    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_rows);
+            SerializableColumnIndex::Optional {
+                non_null_row_ids: Box::new(optional_index),
+                num_rows,
+            }
+        }
+        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_rows);
+            if sort_values_within_row {
+                sort_values_within_row_in_place(multivalued_index, values);
+            }
+            SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
+        }
+    };
+    crate::column::serialize_column_mappable_to_u64(
+        serializable_column_index,
+        &&values[..],
+        &mut wrt,
+    )?;
+    Ok(())
+}
+
+fn coerce_numerical_symbol<T>(
+    operation_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
+) -> impl Iterator<Item = ColumnOperation<u64>>
+where T: Coerce + MonotonicallyMappableToU64 {
+    operation_iterator.map(|symbol| match symbol {
+        ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
+        ColumnOperation::Value(numerical_value) => {
+            ColumnOperation::Value(T::coerce(numerical_value).to_u64())
+        }
+    })
+}
+
+fn consume_operation_iterator<T: Ord, 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);
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use stacker::MemoryArena;
+
+    use crate::columnar::writer::column_operation::ColumnOperation;
+    use crate::{Cardinality, 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(&arena, None, &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(&arena, None, &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(&arena, None, &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(&arena, None, &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 +1,108 @@
+use std::io;
+use std::io::Write;
+
+use common::{BinarySerializable, CountingWriter};
+use sstable::value::RangeValueWriter;
+use sstable::RangeSSTable;
+
+use crate::columnar::ColumnType;
+use crate::RowId;
+
+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, num_rows: RowId) -> 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()[..])?;
+        num_rows.serialize(&mut self.wrt)?;
+        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::Str, &mut buffer);
+        assert_eq!(buffer.len(), 12);
+        assert_eq!(&buffer[..10], b"root\0child");
+        assert_eq!(buffer[10], 0u8);
+        assert_eq!(buffer[11], ColumnType::Str.to_code());
+    }
+}

columnar/src/columnar/writer/value_index.rs

@@ -0,0 +1,165 @@
+use crate::iterable::Iterable;
+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>,
+}
+
+impl OptionalIndexBuilder {
+    pub fn finish(&mut self, num_rows: RowId) -> impl Iterable<RowId> + '_ {
+        debug_assert!(self
+            .docs
+            .last()
+            .copied()
+            .map(|last_doc| last_doc < num_rows)
+            .unwrap_or(true));
+        &self.docs[..]
+    }
+
+    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) -> &[u32] {
+        self.start_offsets
+            .resize(num_docs as usize + 1, self.total_num_vals_seen);
+        &self.start_offsets[..]
+    }
+
+    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)
+                .boxed_iter()
+                .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)
+                .boxed_iter()
+                .collect::<Vec<u32>>(),
+            &[1]
+        );
+    }
+
+    #[test]
+    fn test_multivalued_value_index_builder() {
+        let mut multivalued_value_index_builder = MultivaluedIndexBuilder::default();
+        multivalued_value_index_builder.record_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).to_vec(),
+            vec![0, 0, 2, 3, 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).to_vec(),
+            vec![0, 0, 0, 2, 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 +1,258 @@
+use std::io;
+use std::net::Ipv6Addr;
+use std::sync::Arc;
+
+use common::file_slice::FileSlice;
+use common::{DateTime, HasLen, OwnedBytes};
+
+use crate::column::{BytesColumn, Column, StrColumn};
+use crate::column_values::{monotonic_map_column, StrictlyMonotonicFn};
+use crate::columnar::ColumnType;
+use crate::{Cardinality, NumericalType};
+
+#[derive(Clone)]
+pub enum DynamicColumn {
+    Bool(Column<bool>),
+    I64(Column<i64>),
+    U64(Column<u64>),
+    F64(Column<f64>),
+    IpAddr(Column<Ipv6Addr>),
+    DateTime(Column<DateTime>),
+    Bytes(BytesColumn),
+    Str(StrColumn),
+}
+
+impl DynamicColumn {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            DynamicColumn::Bool(c) => c.get_cardinality(),
+            DynamicColumn::I64(c) => c.get_cardinality(),
+            DynamicColumn::U64(c) => c.get_cardinality(),
+            DynamicColumn::F64(c) => c.get_cardinality(),
+            DynamicColumn::IpAddr(c) => c.get_cardinality(),
+            DynamicColumn::DateTime(c) => c.get_cardinality(),
+            DynamicColumn::Bytes(c) => c.ords().get_cardinality(),
+            DynamicColumn::Str(c) => c.ords().get_cardinality(),
+        }
+    }
+    pub fn column_type(&self) -> ColumnType {
+        match self {
+            DynamicColumn::Bool(_) => ColumnType::Bool,
+            DynamicColumn::I64(_) => ColumnType::I64,
+            DynamicColumn::U64(_) => ColumnType::U64,
+            DynamicColumn::F64(_) => ColumnType::F64,
+            DynamicColumn::IpAddr(_) => ColumnType::IpAddr,
+            DynamicColumn::DateTime(_) => ColumnType::DateTime,
+            DynamicColumn::Bytes(_) => ColumnType::Bytes,
+            DynamicColumn::Str(_) => ColumnType::Str,
+        }
+    }
+
+    pub fn coerce_numerical(self, target_numerical_type: NumericalType) -> Option<Self> {
+        match target_numerical_type {
+            NumericalType::I64 => self.coerce_to_i64(),
+            NumericalType::U64 => self.coerce_to_u64(),
+            NumericalType::F64 => self.coerce_to_f64(),
+        }
+    }
+
+    pub fn is_numerical(&self) -> bool {
+        self.column_type().numerical_type().is_some()
+    }
+
+    pub fn is_f64(&self) -> bool {
+        self.column_type().numerical_type() == Some(NumericalType::F64)
+    }
+    pub fn is_i64(&self) -> bool {
+        self.column_type().numerical_type() == Some(NumericalType::I64)
+    }
+    pub fn is_u64(&self) -> bool {
+        self.column_type().numerical_type() == Some(NumericalType::U64)
+    }
+
+    fn coerce_to_f64(self) -> Option<DynamicColumn> {
+        match self {
+            DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
+                idx: column.idx,
+                values: Arc::new(monotonic_map_column(column.values, MapI64ToF64)),
+            })),
+            DynamicColumn::U64(column) => Some(DynamicColumn::F64(Column {
+                idx: column.idx,
+                values: Arc::new(monotonic_map_column(column.values, MapU64ToF64)),
+            })),
+            DynamicColumn::F64(_) => Some(self),
+            _ => None,
+        }
+    }
+    fn coerce_to_i64(self) -> Option<DynamicColumn> {
+        match self {
+            DynamicColumn::U64(column) => {
+                if column.max_value() > i64::MAX as u64 {
+                    return None;
+                }
+                Some(DynamicColumn::I64(Column {
+                    idx: column.idx,
+                    values: Arc::new(monotonic_map_column(column.values, MapU64ToI64)),
+                }))
+            }
+            DynamicColumn::I64(_) => Some(self),
+            _ => None,
+        }
+    }
+    fn coerce_to_u64(self) -> Option<DynamicColumn> {
+        match self {
+            DynamicColumn::I64(column) => {
+                if column.min_value() < 0 {
+                    return None;
+                }
+                Some(DynamicColumn::U64(Column {
+                    idx: column.idx,
+                    values: Arc::new(monotonic_map_column(column.values, MapI64ToU64)),
+                }))
+            }
+            DynamicColumn::U64(_) => Some(self),
+            _ => None,
+        }
+    }
+}
+
+struct MapI64ToF64;
+impl StrictlyMonotonicFn<i64, f64> for MapI64ToF64 {
+    #[inline(always)]
+    fn mapping(&self, inp: i64) -> f64 {
+        inp as f64
+    }
+    #[inline(always)]
+    fn inverse(&self, out: f64) -> i64 {
+        out as i64
+    }
+}
+
+struct MapU64ToF64;
+impl StrictlyMonotonicFn<u64, f64> for MapU64ToF64 {
+    #[inline(always)]
+    fn mapping(&self, inp: u64) -> f64 {
+        inp as f64
+    }
+    #[inline(always)]
+    fn inverse(&self, out: f64) -> u64 {
+        out as u64
+    }
+}
+
+struct MapU64ToI64;
+impl StrictlyMonotonicFn<u64, i64> for MapU64ToI64 {
+    #[inline(always)]
+    fn mapping(&self, inp: u64) -> i64 {
+        inp as i64
+    }
+    #[inline(always)]
+    fn inverse(&self, out: i64) -> u64 {
+        out as u64
+    }
+}
+
+struct MapI64ToU64;
+impl StrictlyMonotonicFn<i64, u64> for MapI64ToU64 {
+    #[inline(always)]
+    fn mapping(&self, inp: i64) -> u64 {
+        inp as u64
+    }
+    #[inline(always)]
+    fn inverse(&self, out: u64) -> i64 {
+        out as i64
+    }
+}
+
+macro_rules! static_dynamic_conversions {
+    ($typ:ty, $enum_name:ident) => {
+        impl From<DynamicColumn> for Option<$typ> {
+            fn from(dynamic_column: DynamicColumn) -> Option<$typ> {
+                if let DynamicColumn::$enum_name(col) = dynamic_column {
+                    Some(col)
+                } else {
+                    None
+                }
+            }
+        }
+
+        impl From<$typ> for DynamicColumn {
+            fn from(typed_column: $typ) -> Self {
+                DynamicColumn::$enum_name(typed_column)
+            }
+        }
+    };
+}
+
+static_dynamic_conversions!(Column<bool>, Bool);
+static_dynamic_conversions!(Column<u64>, U64);
+static_dynamic_conversions!(Column<i64>, I64);
+static_dynamic_conversions!(Column<f64>, F64);
+static_dynamic_conversions!(Column<DateTime>, DateTime);
+static_dynamic_conversions!(StrColumn, Str);
+static_dynamic_conversions!(BytesColumn, Bytes);
+static_dynamic_conversions!(Column<Ipv6Addr>, IpAddr);
+
+#[derive(Clone)]
+pub struct DynamicColumnHandle {
+    pub(crate) file_slice: FileSlice,
+    pub(crate) column_type: ColumnType,
+}
+
+impl DynamicColumnHandle {
+    // TODO rename load
+    pub fn open(&self) -> io::Result<DynamicColumn> {
+        let column_bytes: OwnedBytes = self.file_slice.read_bytes()?;
+        self.open_internal(column_bytes)
+    }
+
+    #[doc(hidden)]
+    pub fn file_slice(&self) -> &FileSlice {
+        &self.file_slice
+    }
+
+    /// Returns the `u64` fast field reader reader associated with `fields` of types
+    /// Str, u64, i64, f64, or datetime.
+    ///
+    /// If not, the fastfield reader will returns the u64-value associated with the original
+    /// FastValue.
+    pub fn open_u64_lenient(&self) -> io::Result<Option<Column<u64>>> {
+        let column_bytes = self.file_slice.read_bytes()?;
+        match self.column_type {
+            ColumnType::Str | ColumnType::Bytes => {
+                let column: BytesColumn = crate::column::open_column_bytes(column_bytes)?;
+                Ok(Some(column.term_ord_column))
+            }
+            ColumnType::Bool => Ok(None),
+            ColumnType::IpAddr => Ok(None),
+            ColumnType::I64 | ColumnType::U64 | ColumnType::F64 | ColumnType::DateTime => {
+                let column = crate::column::open_column_u64::<u64>(column_bytes)?;
+                Ok(Some(column))
+            }
+        }
+    }
+
+    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::Str => crate::column::open_column_str(column_bytes)?.into(),
+            ColumnType::I64 => crate::column::open_column_u64::<i64>(column_bytes)?.into(),
+            ColumnType::U64 => crate::column::open_column_u64::<u64>(column_bytes)?.into(),
+            ColumnType::F64 => crate::column::open_column_u64::<f64>(column_bytes)?.into(),
+            ColumnType::Bool => crate::column::open_column_u64::<bool>(column_bytes)?.into(),
+            ColumnType::IpAddr => crate::column::open_column_u128::<Ipv6Addr>(column_bytes)?.into(),
+            ColumnType::DateTime => {
+                crate::column::open_column_u64::<DateTime>(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/iterable.rs

@@ -0,0 +1,19 @@
+use std::ops::Range;
+
+pub trait Iterable<T = u64> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_>;
+}
+
+impl<'a, T: Copy> Iterable<T> for &'a [T] {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        Box::new(self.iter().copied())
+    }
+}
+
+impl<T: Copy> Iterable<T> for Range<T>
+where Range<T>: Iterator<Item = T>
+{
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        Box::new(self.clone())
+    }
+}

columnar/src/lib.rs

@@ -0,0 +1,96 @@
+#![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;
+pub mod column_values;
+mod columnar;
+mod dictionary;
+mod dynamic_column;
+mod iterable;
+pub(crate) mod utils;
+mod value;
+
+pub use column::{BytesColumn, Column, StrColumn};
+pub use column_index::ColumnIndex;
+pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
+pub use columnar::{
+    merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
+    MergeRowOrder, ShuffleMergeOrder, StackMergeOrder,
+};
+use sstable::VoidSSTable;
+pub use value::{NumericalType, NumericalValue};
+
+pub use self::dynamic_column::{DynamicColumn, DynamicColumnHandle};
+
+pub type RowId = u32;
+pub type DocId = u32;
+
+#[derive(Clone, Copy)]
+pub struct RowAddr {
+    pub segment_ord: u32,
+    pub row_id: RowId,
+}
+
+pub use sstable::Dictionary;
+pub type Streamer<'a> = sstable::Streamer<'a, VoidSSTable>;
+
+pub use common::DateTime;
+
+#[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 fn is_optional(&self) -> bool {
+        matches!(self, Cardinality::Optional)
+    }
+    pub fn is_multivalue(&self) -> bool {
+        matches!(self, Cardinality::Multivalued)
+    }
+    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 +1,212 @@
+use std::net::Ipv6Addr;
+
+use crate::column_values::MonotonicallyMappableToU128;
+use crate::columnar::ColumnType;
+use crate::dynamic_column::{DynamicColumn, DynamicColumnHandle};
+use crate::value::NumericalValue;
+use crate::{Cardinality, ColumnarReader, ColumnarWriter};
+
+#[test]
+fn test_dataframe_writer_str() {
+    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, None, &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(), 158);
+}
+
+#[test]
+fn test_dataframe_writer_bytes() {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_bytes(1u32, "my_string", b"hello");
+    dataframe_writer.record_bytes(3u32, "my_string", b"helloeee");
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(5, None, &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(), 158);
+}
+
+#[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, None, &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(), 22);
+    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_u64_multivalued() {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_numerical(2u32, "divisor", 2u64);
+    dataframe_writer.record_numerical(3u32, "divisor", 3u64);
+    dataframe_writer.record_numerical(4u32, "divisor", 2u64);
+    dataframe_writer.record_numerical(5u32, "divisor", 5u64);
+    dataframe_writer.record_numerical(6u32, "divisor", 2u64);
+    dataframe_writer.record_numerical(6u32, "divisor", 3u64);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(7, None, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("divisor").unwrap();
+    assert_eq!(cols.len(), 1);
+    assert_eq!(cols[0].num_bytes(), 29);
+    let dyn_i64_col = cols[0].open().unwrap();
+    let DynamicColumn::I64(divisor_col) = dyn_i64_col else { panic!(); };
+    assert_eq!(
+        divisor_col.get_cardinality(),
+        crate::Cardinality::Multivalued
+    );
+    assert_eq!(divisor_col.num_docs(), 7);
+}
+
+#[test]
+fn test_dataframe_writer_ip_addr() {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_ip_addr(1, "ip_addr", Ipv6Addr::from_u128(1001));
+    dataframe_writer.record_ip_addr(3, "ip_addr", Ipv6Addr::from_u128(1050));
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer.serialize(5, None, &mut buffer).unwrap();
+    let columnar = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar.num_columns(), 1);
+    let cols: Vec<DynamicColumnHandle> = columnar.read_columns("ip_addr").unwrap();
+    assert_eq!(cols.len(), 1);
+    assert_eq!(cols[0].num_bytes(), 42);
+    assert_eq!(cols[0].column_type(), ColumnType::IpAddr);
+    let dyn_bool_col = cols[0].open().unwrap();
+    let DynamicColumn::IpAddr(ip_col) = dyn_bool_col else { panic!(); };
+    let vals: Vec<Option<Ipv6Addr>> = (0..5).map(|row_id| ip_col.first(row_id)).collect();
+    assert_eq!(
+        &vals,
+        &[
+            None,
+            Some(Ipv6Addr::from_u128(1001)),
+            None,
+            Some(Ipv6Addr::from_u128(1050)),
+            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, None, &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(), 33);
+    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_str() {
+    let mut buffer = Vec::new();
+    let mut columnar_writer = ColumnarWriter::default();
+    columnar_writer.record_str(1, "my.column", "a");
+    columnar_writer.record_str(3, "my.column", "c");
+    columnar_writer.record_str(3, "my.column2", "different_column!");
+    columnar_writer.record_str(4, "my.column", "b");
+    columnar_writer.serialize(5, None, &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!(); };
+    let index: Vec<Option<u64>> = (0..5).map(|row_id| str_col.ords().first(row_id)).collect();
+    assert_eq!(index, &[None, Some(0), None, Some(2), Some(1)]);
+    assert_eq!(str_col.num_rows(), 5);
+    let mut term_buffer = String::new();
+    let term_ords = str_col.ords();
+    assert_eq!(term_ords.first(0), None);
+    assert_eq!(term_ords.first(1), Some(0));
+    str_col.ord_to_str(0u64, &mut term_buffer).unwrap();
+    assert_eq!(term_buffer, "a");
+    assert_eq!(term_ords.first(2), None);
+    assert_eq!(term_ords.first(3), Some(2));
+    str_col.ord_to_str(2u64, &mut term_buffer).unwrap();
+    assert_eq!(term_buffer, "c");
+    assert_eq!(term_ords.first(4), Some(1));
+    str_col.ord_to_str(1u64, &mut term_buffer).unwrap();
+    assert_eq!(term_buffer, "b");
+}
+
+#[test]
+fn test_dictionary_encoded_bytes() {
+    let mut buffer = Vec::new();
+    let mut columnar_writer = ColumnarWriter::default();
+    columnar_writer.record_bytes(1, "my.column", b"a");
+    columnar_writer.record_bytes(3, "my.column", b"c");
+    columnar_writer.record_bytes(3, "my.column2", b"different_column!");
+    columnar_writer.record_bytes(4, "my.column", b"b");
+    columnar_writer.serialize(5, None, &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::Bytes(bytes_col) = col_handles[0].open().unwrap() else  { panic!(); };
+    let index: Vec<Option<u64>> = (0..5)
+        .map(|row_id| bytes_col.ords().first(row_id))
+        .collect();
+    assert_eq!(index, &[None, Some(0), None, Some(2), Some(1)]);
+    assert_eq!(bytes_col.num_rows(), 5);
+    let mut term_buffer = Vec::new();
+    let term_ords = bytes_col.ords();
+    assert_eq!(term_ords.first(0), None);
+    assert_eq!(term_ords.first(1), Some(0));
+    bytes_col
+        .dictionary
+        .ord_to_term(0u64, &mut term_buffer)
+        .unwrap();
+    assert_eq!(term_buffer, b"a");
+    assert_eq!(term_ords.first(2), None);
+    assert_eq!(term_ords.first(3), Some(2));
+    bytes_col
+        .dictionary
+        .ord_to_term(2u64, &mut term_buffer)
+        .unwrap();
+    assert_eq!(term_buffer, b"c");
+    assert_eq!(term_ords.first(4), Some(1));
+    bytes_col
+        .dictionary
+        .ord_to_term(1u64, &mut term_buffer)
+        .unwrap();
+    assert_eq!(term_buffer, b"b");
+}

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 +1,131 @@
+use common::DateTime;
+
+use crate::InvalidData;
+
+#[derive(Copy, Clone, PartialEq, Debug)]
+pub enum NumericalValue {
+    I64(i64),
+    U64(u64),
+    F64(f64),
+}
+
+impl NumericalValue {
+    pub fn numerical_type(&self) -> NumericalType {
+        match self {
+            NumericalValue::I64(_) => NumericalType::I64,
+            NumericalValue::U64(_) => NumericalType::U64,
+            NumericalValue::F64(_) => NumericalType::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)
+    }
+}
+
+#[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,
+        }
+    }
+}
+
+impl Coerce for DateTime {
+    fn coerce(value: NumericalValue) -> Self {
+        let timestamp_micros = i64::coerce(value);
+        DateTime::from_timestamp_micros(timestamp_micros)
+    }
+}
+
+#[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/Cargo.toml

@@ -13,9 +13,10 @@ repository = "https://github.com/quickwit-oss/tantivy"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-byteorder = "1.4.3"
 ownedbytes = { version= "0.5", path="../ownedbytes" }
 async-trait = "0.1"
+time = { version = "0.3.10", features = ["serde-well-known"] }
+serde = { version = "1.0.136", features = ["derive"] }
 
 [dev-dependencies]
 proptest = "1.0.0"

common/src/datetime.rs

@@ -0,0 +1,136 @@
+use std::fmt;
+
+use serde::{Deserialize, Serialize};
+use time::format_description::well_known::Rfc3339;
+use time::{OffsetDateTime, PrimitiveDateTime, UtcOffset};
+
+/// DateTime Precision
+#[derive(
+    Clone, Copy, Debug, Hash, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize, Default,
+)]
+#[serde(rename_all = "lowercase")]
+pub enum DatePrecision {
+    /// Seconds precision
+    #[default]
+    Seconds,
+    /// Milli-seconds precision.
+    Milliseconds,
+    /// Micro-seconds precision.
+    Microseconds,
+}
+
+/// A date/time value with microsecond precision.
+///
+/// This timestamp does not carry any explicit time zone information.
+/// Users are responsible for applying the provided conversion
+/// functions consistently. Internally the time zone is assumed
+/// to be UTC, which is also used implicitly for JSON serialization.
+///
+/// All constructors and conversions are provided as explicit
+/// functions and not by implementing any `From`/`Into` traits
+/// to prevent unintended usage.
+#[derive(Clone, Default, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct DateTime {
+    // Timestamp in microseconds.
+    pub(crate) timestamp_micros: i64,
+}
+
+impl DateTime {
+    /// Create new from UNIX timestamp in seconds
+    pub const fn from_timestamp_secs(seconds: i64) -> Self {
+        Self {
+            timestamp_micros: seconds * 1_000_000,
+        }
+    }
+
+    /// Create new from UNIX timestamp in milliseconds
+    pub const fn from_timestamp_millis(milliseconds: i64) -> Self {
+        Self {
+            timestamp_micros: milliseconds * 1_000,
+        }
+    }
+
+    /// Create new from UNIX timestamp in microseconds.
+    pub const fn from_timestamp_micros(microseconds: i64) -> Self {
+        Self {
+            timestamp_micros: microseconds,
+        }
+    }
+
+    /// Create new from `OffsetDateTime`
+    ///
+    /// The given date/time is converted to UTC and the actual
+    /// time zone is discarded.
+    pub const fn from_utc(dt: OffsetDateTime) -> Self {
+        let timestamp_micros = dt.unix_timestamp() * 1_000_000 + dt.microsecond() as i64;
+        Self { timestamp_micros }
+    }
+
+    /// Create new from `PrimitiveDateTime`
+    ///
+    /// Implicitly assumes that the given date/time is in UTC!
+    /// Otherwise the original value must only be reobtained with
+    /// [`Self::into_primitive()`].
+    pub fn from_primitive(dt: PrimitiveDateTime) -> Self {
+        Self::from_utc(dt.assume_utc())
+    }
+
+    /// Convert to UNIX timestamp in seconds.
+    pub const fn into_timestamp_secs(self) -> i64 {
+        self.timestamp_micros / 1_000_000
+    }
+
+    /// Convert to UNIX timestamp in milliseconds.
+    pub const fn into_timestamp_millis(self) -> i64 {
+        self.timestamp_micros / 1_000
+    }
+
+    /// Convert to UNIX timestamp in microseconds.
+    pub const fn into_timestamp_micros(self) -> i64 {
+        self.timestamp_micros
+    }
+
+    /// Convert to UTC `OffsetDateTime`
+    pub fn into_utc(self) -> OffsetDateTime {
+        let timestamp_nanos = self.timestamp_micros as i128 * 1000;
+        let utc_datetime = OffsetDateTime::from_unix_timestamp_nanos(timestamp_nanos)
+            .expect("valid UNIX timestamp");
+        debug_assert_eq!(UtcOffset::UTC, utc_datetime.offset());
+        utc_datetime
+    }
+
+    /// Convert to `OffsetDateTime` with the given time zone
+    pub fn into_offset(self, offset: UtcOffset) -> OffsetDateTime {
+        self.into_utc().to_offset(offset)
+    }
+
+    /// Convert to `PrimitiveDateTime` without any time zone
+    ///
+    /// The value should have been constructed with [`Self::from_primitive()`].
+    /// Otherwise the time zone is implicitly assumed to be UTC.
+    pub fn into_primitive(self) -> PrimitiveDateTime {
+        let utc_datetime = self.into_utc();
+        // Discard the UTC time zone offset
+        debug_assert_eq!(UtcOffset::UTC, utc_datetime.offset());
+        PrimitiveDateTime::new(utc_datetime.date(), utc_datetime.time())
+    }
+
+    /// Truncates the microseconds value to the corresponding precision.
+    pub fn truncate(self, precision: DatePrecision) -> Self {
+        let truncated_timestamp_micros = match precision {
+            DatePrecision::Seconds => (self.timestamp_micros / 1_000_000) * 1_000_000,
+            DatePrecision::Milliseconds => (self.timestamp_micros / 1_000) * 1_000,
+            DatePrecision::Microseconds => self.timestamp_micros,
+        };
+        Self {
+            timestamp_micros: truncated_timestamp_micros,
+        }
+    }
+}
+
+impl fmt::Debug for DateTime {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        let utc_rfc3339 = self.into_utc().format(&Rfc3339).map_err(|_| fmt::Error)?;
+        f.write_str(&utc_rfc3339)
+    }
+}

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

@@ -2,14 +2,16 @@
 
 use std::ops::Deref;
 
-pub use byteorder::LittleEndian as Endianness;
-
 mod bitset;
+mod datetime;
 pub mod file_slice;
+mod group_by;
 mod serialize;
 mod vint;
 mod writer;
 pub use bitset::*;
+pub use datetime::{DatePrecision, DateTime};
+pub use group_by::GroupByIteratorExtended;
 pub use ownedbytes::{OwnedBytes, StableDeref};
 pub use serialize::{BinarySerializable, DeserializeFrom, FixedSize};
 pub use vint::{
@@ -105,6 +107,21 @@ pub fn u64_to_f64(val: u64) -> f64 {
     })
 }
 
+/// Replaces a given byte in the `bytes` slice of bytes.
+///
+/// This function assumes that the needle is rarely contained in the bytes string
+/// and offers a fast path if the needle is not present.
+pub fn replace_in_place(needle: u8, replacement: u8, bytes: &mut [u8]) {
+    if !bytes.contains(&needle) {
+        return;
+    }
+    for b in bytes {
+        if *b == needle {
+            *b = replacement;
+        }
+    }
+}
+
 #[cfg(test)]
 pub mod test {
 
@@ -169,4 +186,20 @@ pub mod test {
         assert!(f64_to_u64(-2.0) < f64_to_u64(1.0));
         assert!(f64_to_u64(-2.0) < f64_to_u64(-1.5));
     }
+
+    #[test]
+    fn test_replace_in_place() {
+        let test_aux = |before_replacement: &[u8], expected: &[u8]| {
+            let mut bytes: Vec<u8> = before_replacement.to_vec();
+            super::replace_in_place(b'b', b'c', &mut bytes);
+            assert_eq!(&bytes[..], expected);
+        };
+        test_aux(b"", b"");
+        test_aux(b"b", b"c");
+        test_aux(b"baaa", b"caaa");
+        test_aux(b"aaab", b"aaac");
+        test_aux(b"aaabaa", b"aaacaa");
+        test_aux(b"aaaaaa", b"aaaaaa");
+        test_aux(b"bbbb", b"cccc");
+    }
 }

common/src/serialize.rs

@@ -1,16 +1,39 @@
 use std::io::{Read, Write};
 use std::{fmt, io};
 
-use byteorder::{ReadBytesExt, WriteBytesExt};
+use crate::VInt;
 
-use crate::{Endianness, VInt};
+#[derive(Default)]
+struct Counter(u64);
+
+impl io::Write for Counter {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        self.0 += buf.len() as u64;
+        Ok(buf.len())
+    }
+
+    fn write_all(&mut self, buf: &[u8]) -> io::Result<()> {
+        self.0 += buf.len() as u64;
+        Ok(())
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
 
 /// Trait for a simple binary serialization.
 pub trait BinarySerializable: fmt::Debug + Sized {
     /// Serialize
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()>;
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()>;
     /// Deserialize
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self>;
+
+    fn num_bytes(&self) -> u64 {
+        let mut counter = Counter::default();
+        self.serialize(&mut counter).unwrap();
+        counter.0
+    }
 }
 
 pub trait DeserializeFrom<T: BinarySerializable> {
@@ -34,7 +57,7 @@ pub trait FixedSize: BinarySerializable {
 }
 
 impl BinarySerializable for () {
-    fn serialize<W: Write>(&self, _: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, _: &mut W) -> io::Result<()> {
         Ok(())
     }
     fn deserialize<R: Read>(_: &mut R) -> io::Result<Self> {
@@ -47,7 +70,7 @@ impl FixedSize for () {
 }
 
 impl<T: BinarySerializable> BinarySerializable for Vec<T> {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.len() as u64).serialize(writer)?;
         for it in self {
             it.serialize(writer)?;
@@ -66,7 +89,7 @@ impl<T: BinarySerializable> BinarySerializable for Vec<T> {
 }
 
 impl<Left: BinarySerializable, Right: BinarySerializable> BinarySerializable for (Left, Right) {
-    fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, write: &mut W) -> io::Result<()> {
         self.0.serialize(write)?;
         self.1.serialize(write)
     }
@@ -81,12 +104,14 @@ impl<Left: BinarySerializable + FixedSize, Right: BinarySerializable + FixedSize
 }
 
 impl BinarySerializable for u32 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u32::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
 
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<u32> {
-        reader.read_u32::<Endianness>()
+        let mut buf = [0u8; 4];
+        reader.read_exact(&mut buf)?;
+        Ok(u32::from_le_bytes(buf))
     }
 }
 
@@ -95,12 +120,14 @@ impl FixedSize for u32 {
 }
 
 impl BinarySerializable for u16 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u16::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
 
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<u16> {
-        reader.read_u16::<Endianness>()
+        let mut buf = [0u8; 2];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -109,11 +136,13 @@ impl FixedSize for u16 {
 }
 
 impl BinarySerializable for u64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u64::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        reader.read_u64::<Endianness>()
+        let mut buf = [0u8; 8];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -122,11 +151,13 @@ impl FixedSize for u64 {
 }
 
 impl BinarySerializable for u128 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u128::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        reader.read_u128::<Endianness>()
+        let mut buf = [0u8; 16];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -135,11 +166,13 @@ impl FixedSize for u128 {
 }
 
 impl BinarySerializable for f32 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_f32::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        reader.read_f32::<Endianness>()
+        let mut buf = [0u8; 4];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -148,11 +181,13 @@ impl FixedSize for f32 {
 }
 
 impl BinarySerializable for i64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_i64::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        reader.read_i64::<Endianness>()
+        let mut buf = [0u8; Self::SIZE_IN_BYTES];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -161,11 +196,13 @@ impl FixedSize for i64 {
 }
 
 impl BinarySerializable for f64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_f64::<Endianness>(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
-        reader.read_f64::<Endianness>()
+        let mut buf = [0u8; Self::SIZE_IN_BYTES];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -174,11 +211,13 @@ impl FixedSize for f64 {
 }
 
 impl BinarySerializable for u8 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u8(*self)
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        writer.write_all(&self.to_le_bytes())
     }
-    fn deserialize<R: Read>(reader: &mut R) -> io::Result<u8> {
-        reader.read_u8()
+    fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
+        let mut buf = [0u8; Self::SIZE_IN_BYTES];
+        reader.read_exact(&mut buf)?;
+        Ok(Self::from_le_bytes(buf))
     }
 }
 
@@ -187,11 +226,11 @@ impl FixedSize for u8 {
 }
 
 impl BinarySerializable for bool {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
-        writer.write_u8(u8::from(*self))
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        (*self as u8).serialize(writer)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<bool> {
-        let val = reader.read_u8()?;
+        let val = u8::deserialize(reader)?;
         match val {
             0 => Ok(false),
             1 => Ok(true),
@@ -208,7 +247,7 @@ impl FixedSize for bool {
 }
 
 impl BinarySerializable for String {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let data: &[u8] = self.as_bytes();
         VInt(data.len() as u64).serialize(writer)?;
         writer.write_all(data)

common/src/vint.rs

@@ -1,8 +1,6 @@
 use std::io;
 use std::io::{Read, Write};
 
-use byteorder::{ByteOrder, LittleEndian};
-
 use super::BinarySerializable;
 
 /// Variable int serializes a u128 number
@@ -44,7 +42,7 @@ pub fn deserialize_vint_u128(data: &[u8]) -> io::Result<(u128, &[u8])> {
 pub struct VIntU128(pub u128);
 
 impl BinarySerializable for VIntU128 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let mut buffer = vec![];
         serialize_vint_u128(self.0, &mut buffer);
         writer.write_all(&buffer)
@@ -127,7 +125,7 @@ pub fn serialize_vint_u32(val: u32, buf: &mut [u8; 8]) -> &[u8] {
             5,
         ),
     };
-    LittleEndian::write_u64(&mut buf[..], res);
+    buf.copy_from_slice(&res.to_le_bytes());
     &buf[0..num_bytes]
 }
 
@@ -211,7 +209,7 @@ impl VInt {
 }
 
 impl BinarySerializable for VInt {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let mut buffer = [0u8; 10];
         let num_bytes = self.serialize_into(&mut buffer);
         writer.write_all(&buffer[0..num_bytes])

examples/aggregation.rs

@@ -1,130 +1,319 @@
 // # Aggregation example
 //
 // This example shows how you can use built-in aggregations.
-// We will use range buckets and compute the average in each bucket.
-//
+// We will use nested aggregations with buckets and metrics:
+// - Range buckets and compute the average in each bucket.
+// - Term aggregation and compute the min price in each bucket
+// ---
 
-use serde_json::Value;
+use serde_json::{Deserializer, Value};
 use tantivy::aggregation::agg_req::{
     Aggregation, Aggregations, BucketAggregation, BucketAggregationType, MetricAggregation,
     RangeAggregation,
 };
 use tantivy::aggregation::agg_result::AggregationResults;
+use tantivy::aggregation::bucket::RangeAggregationRange;
 use tantivy::aggregation::metric::AverageAggregation;
 use tantivy::aggregation::AggregationCollector;
-use tantivy::query::TermQuery;
-use tantivy::schema::{self, Cardinality, IndexRecordOption, Schema, TextFieldIndexing};
-use tantivy::{doc, Index, Term};
+use tantivy::query::AllQuery;
+use tantivy::schema::{self, IndexRecordOption, Schema, TextFieldIndexing, FAST};
+use tantivy::Index;
 
 fn main() -> tantivy::Result<()> {
+    // # Create Schema
+    //
+    // Lets create a schema for a footwear shop, with 4 fields: name, category, stock and price.
+    // category, stock and price will be fast fields as that's the requirement
+    // for aggregation queries.
+    //
+
     let mut schema_builder = Schema::builder();
+    // In preparation of the `TermsAggregation`, the category field is configured with:
+    // - `set_fast`
+    // - `raw` tokenizer
+    //
+    // The tokenizer is set to "raw", because the fast field uses the same dictionary as the
+    // inverted index. (This behaviour will change in tantivy 0.20, where the fast field will
+    // always be raw tokenized independent from the regular tokenizing)
+    //
     let text_fieldtype = schema::TextOptions::default()
         .set_indexing_options(
-            TextFieldIndexing::default().set_index_option(IndexRecordOption::WithFreqs),
+            TextFieldIndexing::default()
+                .set_index_option(IndexRecordOption::WithFreqs)
+                .set_tokenizer("raw"),
         )
+        .set_fast()
         .set_stored();
-    let text_field = schema_builder.add_text_field("text", text_fieldtype);
-    let score_fieldtype =
-        crate::schema::NumericOptions::default().set_fast(Cardinality::SingleValue);
-    let highscore_field = schema_builder.add_f64_field("highscore", score_fieldtype.clone());
-    let price_field = schema_builder.add_f64_field("price", score_fieldtype.clone());
+    schema_builder.add_text_field("category", text_fieldtype);
+    schema_builder.add_f64_field("stock", FAST);
+    schema_builder.add_f64_field("price", FAST);
 
     let schema = schema_builder.build();
 
     // # Indexing documents
     //
     // Lets index a bunch of documents for this example.
-    let index = Index::create_in_ram(schema);
+    let index = Index::create_in_ram(schema.clone());
+
+    let data = r#"{
+        "name": "Almond Toe Court Shoes, Patent Black",
+        "category": "Womens Footwear",
+        "price": 99.00,
+        "stock": 5
+    }
+    {
+        "name": "Suede Shoes, Blue",
+        "category": "Womens Footwear",
+        "price": 42.00,
+        "stock": 4
+    }
+    {
+        "name": "Leather Driver Saddle Loafers, Tan",
+        "category": "Mens Footwear",
+        "price": 34.00,
+        "stock": 12
+    }
+    {
+        "name": "Flip Flops, Red",
+        "category": "Mens Footwear",
+        "price": 19.00,
+        "stock": 6
+    }
+    {
+        "name": "Flip Flops, Blue",
+        "category": "Mens Footwear",
+        "price": 19.00,
+        "stock": 0
+    }
+    {
+        "name": "Gold Button Cardigan, Black",
+        "category": "Womens Casualwear",
+        "price": 167.00,
+        "stock": 6
+    }
+    {
+        "name": "Cotton Shorts, Medium Red",
+        "category": "Womens Casualwear",
+        "price": 30.00,
+        "stock": 5
+    }
+    {
+        "name": "Fine Stripe Short SleeveShirt, Grey",
+        "category": "Mens Casualwear",
+        "price": 49.99,
+        "stock": 9
+    }
+    {
+        "name": "Fine Stripe Short SleeveShirt, Green",
+        "category": "Mens Casualwear",
+        "price": 49.99,
+        "offer": 39.99,
+        "stock": 9
+    }
+    {
+        "name": "Sharkskin Waistcoat, Charcoal",
+        "category": "Mens Formalwear",
+        "price": 75.00,
+        "stock": 2
+    }
+    {
+        "name": "Lightweight Patch PocketBlazer, Deer",
+        "category": "Mens Formalwear",
+        "price": 175.50,
+        "stock": 1
+    }
+    {
+        "name": "Bird Print Dress, Black",
+        "category": "Womens Formalwear",
+        "price": 270.00,
+        "stock": 10
+    }
+    {
+        "name": "Mid Twist Cut-Out Dress, Pink",
+        "category": "Womens Formalwear",
+        "price": 540.00,
+        "stock": 5
+    }"#;
+
+    let stream = Deserializer::from_str(data).into_iter::<Value>();
 
     let mut index_writer = index.writer(50_000_000)?;
-    // writing the segment
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 1f64,
-        price_field => 0f64,
-    ))?;
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 3f64,
-        price_field => 1f64,
-    ))?;
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 5f64,
-        price_field => 1f64,
-    ))?;
-    index_writer.add_document(doc!(
-        text_field => "nohit",
-        highscore_field => 6f64,
-        price_field => 2f64,
-    ))?;
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 7f64,
-        price_field => 2f64,
-    ))?;
-    index_writer.commit()?;
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 11f64,
-        price_field => 10f64,
-    ))?;
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 14f64,
-        price_field => 15f64,
-    ))?;
-
-    index_writer.add_document(doc!(
-        text_field => "cool",
-        highscore_field => 15f64,
-        price_field => 20f64,
-    ))?;
+    let mut num_indexed = 0;
+    for value in stream {
+        let doc = schema.parse_document(&serde_json::to_string(&value.unwrap())?)?;
+        index_writer.add_document(doc)?;
+        num_indexed += 1;
+        if num_indexed > 4 {
+            // Writing the first segment
+            index_writer.commit()?;
+        }
+    }
 
+    // Writing the second segment
     index_writer.commit()?;
 
+    // We have two segments now. The `AggregationCollector` will run the aggregation on each
+    // segment and then merge the results into an `IntermediateAggregationResult`.
+
     let reader = index.reader()?;
-    let text_field = reader.searcher().schema().get_field("text").unwrap();
+    let searcher = reader.searcher();
+    // ---
+    // # Aggregation Query
+    //
+    //
+    // We can construct the query by building the request structure or by deserializing from JSON.
+    // The JSON API is more stable and therefore recommended.
+    //
+    // ## Request 1
 
-    let term_query = TermQuery::new(
-        Term::from_field_text(text_field, "cool"),
-        IndexRecordOption::Basic,
-    );
+    let agg_req_str = r#"
+    {
+      "group_by_stock": {
+        "aggs": {
+          "average_price": { "avg": { "field": "price" } }
+        },
+        "range": {
+          "field": "stock",
+          "ranges": [
+            { "key": "few", "to": 1.0 },
+            { "key": "some", "from": 1.0, "to": 10.0 },
+            { "key": "many", "from": 10.0 }
+          ]
+        }
+      }
+    } "#;
 
-    let sub_agg_req_1: Aggregations = vec![(
-        "average_price".to_string(),
-        Aggregation::Metric(MetricAggregation::Average(
-            AverageAggregation::from_field_name("price".to_string()),
-        )),
-    )]
-    .into_iter()
-    .collect();
+    // In this Aggregation we want to get the average price for different groups, depending on how
+    // many items are in stock. We define custom ranges `few`, `some`, `many` via the
+    // range aggregation.
+    // For every bucket we want the average price, so we create a nested metric aggregation on the
+    // range bucket aggregation. Only buckets support nested aggregations.
+    // ### Request JSON API
+    //
 
-    let agg_req_1: Aggregations = vec![(
-        "score_ranges".to_string(),
+    let agg_req: Aggregations = serde_json::from_str(agg_req_str)?;
+    let collector = AggregationCollector::from_aggs(agg_req, None);
+
+    let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
+    let res2: Value = serde_json::to_value(agg_res)?;
+
+    // ### Request Rust API
+    //
+    // This is exactly the same request as above, but via the rust structures.
+    //
+
+    let agg_req: Aggregations = vec![(
+        "group_by_stock".to_string(),
         Aggregation::Bucket(BucketAggregation {
             bucket_agg: BucketAggregationType::Range(RangeAggregation {
-                field: "highscore".to_string(),
+                field: "stock".to_string(),
                 ranges: vec![
-                    (-1f64..9f64).into(),
-                    (9f64..14f64).into(),
-                    (14f64..20f64).into(),
+                    RangeAggregationRange {
+                        key: Some("few".into()),
+                        from: None,
+                        to: Some(1f64),
+                    },
+                    RangeAggregationRange {
+                        key: Some("some".into()),
+                        from: Some(1f64),
+                        to: Some(10f64),
+                    },
+                    RangeAggregationRange {
+                        key: Some("many".into()),
+                        from: Some(10f64),
+                        to: None,
+                    },
                 ],
                 ..Default::default()
             }),
-            sub_aggregation: sub_agg_req_1.clone(),
+            sub_aggregation: vec![(
+                "average_price".to_string(),
+                Aggregation::Metric(MetricAggregation::Average(
+                    AverageAggregation::from_field_name("price".to_string()),
+                )),
+            )]
+            .into_iter()
+            .collect(),
         }),
     )]
     .into_iter()
     .collect();
 
-    let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
+    let collector = AggregationCollector::from_aggs(agg_req, None);
+    // We use the `AllQuery` which will pass all documents to the AggregationCollector.
+    let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
 
-    let searcher = reader.searcher();
-    let agg_res: AggregationResults = searcher.search(&term_query, &collector).unwrap();
+    let res1: Value = serde_json::to_value(agg_res)?;
 
-    let res: Value = serde_json::to_value(&agg_res)?;
-    println!("{}", serde_json::to_string_pretty(&res)?);
+    // ### Aggregation Result
+    //
+    // The resulting structure deserializes in the same JSON format as elastic search.
+    //
+    let expected_res = r#"
+    {
+        "group_by_stock":{
+            "buckets":[
+                {"average_price":{"value":19.0},"doc_count":1,"key":"few","to":1.0},
+                {"average_price":{"value":124.748},"doc_count":10,"from":1.0,"key":"some","to":10.0},
+                {"average_price":{"value":152.0},"doc_count":2,"from":10.0,"key":"many"}
+            ]
+        }
+    }
+    "#;
+    let expected_json: Value = serde_json::from_str(expected_res)?;
+    assert_eq!(expected_json, res1);
+    assert_eq!(expected_json, res2);
+
+    // ### Request 2
+    //
+    // Now we are interested in the minimum price per category, so we create a bucket per
+    // category via `TermsAggregation`. We are interested in the highest minimum prices, and set the
+    // order of the buckets `"order": { "min_price": "desc" }` to be sorted by the the metric of
+    // the sub aggregation. (awesome)
+    //
+    let agg_req_str = r#"
+    {
+      "min_price_per_category": {
+        "aggs": {
+          "min_price": { "min": { "field": "price" } }
+        },
+        "terms": {
+          "field": "category",
+          "min_doc_count": 1,
+          "order": { "min_price": "desc" }
+        }
+      }
+    } "#;
+
+    let agg_req: Aggregations = serde_json::from_str(agg_req_str)?;
+
+    let collector = AggregationCollector::from_aggs(agg_req, None);
+
+    let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
+    let res: Value = serde_json::to_value(agg_res)?;
+
+    // Minimum price per category, sorted by minimum price descending
+    //
+    // As you can see, the starting prices for `Formalwear` are higher than `Casualwear`.
+    //
+    let expected_res = r#"
+    {
+      "min_price_per_category": {
+        "buckets": [
+          { "doc_count": 2, "key": "Womens Formalwear", "min_price": { "value": 270.0 } },
+          { "doc_count": 2, "key": "Mens Formalwear", "min_price": { "value": 75.0 } },
+          { "doc_count": 2, "key": "Mens Casualwear", "min_price": { "value": 49.99 } },
+          { "doc_count": 2, "key": "Womens Footwear", "min_price": { "value": 42.0 } },
+          { "doc_count": 2, "key": "Womens Casualwear", "min_price": { "value": 30.0 } },
+          { "doc_count": 3, "key": "Mens Footwear", "min_price": { "value": 19.0 } }
+        ],
+        "sum_other_doc_count": 0
+      }
+    }
+    "#;
+    let expected_json: Value = serde_json::from_str(expected_res)?;
+
+    assert_eq!(expected_json, res);
 
     Ok(())
 }

examples/custom_collector.rs

@@ -7,14 +7,12 @@
 // Of course, you can have a look at the tantivy's built-in collectors
 // such as the `CountCollector` for more examples.
 
-use std::sync::Arc;
-
-use fastfield_codecs::Column;
+use columnar::Column;
 // ---
 // Importing tantivy...
 use tantivy::collector::{Collector, SegmentCollector};
 use tantivy::query::QueryParser;
-use tantivy::schema::{Field, Schema, FAST, INDEXED, TEXT};
+use tantivy::schema::{Schema, FAST, INDEXED, TEXT};
 use tantivy::{doc, Index, Score, SegmentReader};
 
 #[derive(Default)]
@@ -52,11 +50,11 @@ impl Stats {
 }
 
 struct StatsCollector {
-    field: Field,
+    field: String,
 }
 
 impl StatsCollector {
-    fn with_field(field: Field) -> StatsCollector {
+    fn with_field(field: String) -> StatsCollector {
         StatsCollector { field }
     }
 }
@@ -73,7 +71,7 @@ impl Collector for StatsCollector {
         _segment_local_id: u32,
         segment_reader: &SegmentReader,
     ) -> tantivy::Result<StatsSegmentCollector> {
-        let fast_field_reader = segment_reader.fast_fields().u64(self.field)?;
+        let fast_field_reader = segment_reader.fast_fields().u64(&self.field)?;
         Ok(StatsSegmentCollector {
             fast_field_reader,
             stats: Stats::default(),
@@ -97,7 +95,7 @@ impl Collector for StatsCollector {
 }
 
 struct StatsSegmentCollector {
-    fast_field_reader: Arc<dyn Column<u64>>,
+    fast_field_reader: Column,
     stats: Stats,
 }
 
@@ -105,10 +103,14 @@ impl SegmentCollector for StatsSegmentCollector {
     type Fruit = Option<Stats>;
 
     fn collect(&mut self, doc: u32, _score: Score) {
-        let value = self.fast_field_reader.get_val(doc) as f64;
-        self.stats.count += 1;
-        self.stats.sum += value;
-        self.stats.squared_sum += value * value;
+        // Since we know the values are single value, we could call `first_or_default_col` on the
+        // column and fetch single values.
+        for value in self.fast_field_reader.values_for_doc(doc) {
+            let value = value as f64;
+            self.stats.count += 1;
+            self.stats.sum += value;
+            self.stats.squared_sum += value * value;
+        }
     }
 
     fn harvest(self) -> <Self as SegmentCollector>::Fruit {
@@ -169,9 +171,11 @@ fn main() -> tantivy::Result<()> {
     let searcher = reader.searcher();
     let query_parser = QueryParser::for_index(&index, vec![product_name, product_description]);
 
-    // here we want to get a hit on the 'ken' in Frankenstein
+    // here we want to search for `broom` and use `StatsCollector` on the hits.
     let query = query_parser.parse_query("broom")?;
-    if let Some(stats) = searcher.search(&query, &StatsCollector::with_field(price))? {
+    if let Some(stats) =
+        searcher.search(&query, &StatsCollector::with_field("price".to_string()))?
+    {
         println!("count: {}", stats.count());
         println!("mean: {}", stats.mean());
         println!("standard deviation: {}", stats.standard_deviation());

examples/custom_tokenizer.rs

@@ -1,7 +1,7 @@
 // # Defining a tokenizer pipeline
 //
-// In this example, we'll see how to define a tokenizer pipeline
-// by aligning a bunch of `TokenFilter`.
+// In this example, we'll see how to define a tokenizer
+// by creating a custom `NgramTokenizer`.
 use tantivy::collector::TopDocs;
 use tantivy::query::QueryParser;
 use tantivy::schema::*;

examples/date_time_field.rs

@@ -4,7 +4,7 @@
 
 use tantivy::collector::TopDocs;
 use tantivy::query::QueryParser;
-use tantivy::schema::{Cardinality, DateOptions, Schema, Value, INDEXED, STORED, STRING};
+use tantivy::schema::{DateOptions, Schema, Value, INDEXED, STORED, STRING};
 use tantivy::Index;
 
 fn main() -> tantivy::Result<()> {
@@ -12,8 +12,9 @@ fn main() -> tantivy::Result<()> {
     let mut schema_builder = Schema::builder();
     let opts = DateOptions::from(INDEXED)
         .set_stored()
-        .set_fast(Cardinality::SingleValue)
+        .set_fast()
         .set_precision(tantivy::DatePrecision::Seconds);
+    // Add `occurred_at` date field type
     let occurred_at = schema_builder.add_date_field("occurred_at", opts);
     let event_type = schema_builder.add_text_field("event", STRING | STORED);
     let schema = schema_builder.build();
@@ -22,6 +23,7 @@ fn main() -> tantivy::Result<()> {
     let index = Index::create_in_ram(schema.clone());
 
     let mut index_writer = index.writer(50_000_000)?;
+    // The dates are passed as string in the RFC3339 format
     let doc = schema.parse_document(
         r#"{
         "occurred_at": "2022-06-22T12:53:50.53Z",
@@ -41,14 +43,16 @@ fn main() -> tantivy::Result<()> {
     let reader = index.reader()?;
     let searcher = reader.searcher();
 
-    // # Default fields: event_type
+    // # Search
     let query_parser = QueryParser::for_index(&index, vec![event_type]);
     {
-        let query = query_parser.parse_query("event:comment")?;
+        // Simple exact search on the date
+        let query = query_parser.parse_query("occurred_at:\"2022-06-22T12:53:50.53Z\"")?;
         let count_docs = searcher.search(&*query, &TopDocs::with_limit(5))?;
         assert_eq!(count_docs.len(), 1);
     }
     {
+        // Range query on the date field
         let query = query_parser
             .parse_query(r#"occurred_at:[2022-06-22T12:58:00Z TO 2022-06-23T00:00:00Z}"#)?;
         let count_docs = searcher.search(&*query, &TopDocs::with_limit(4))?;

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());
 
@@ -69,7 +71,7 @@ fn main() -> tantivy::Result<()> {
     let reader = index.reader()?;
     let searcher = reader.searcher();
     {
-        let mut facet_collector = FacetCollector::for_field(classification);
+        let mut facet_collector = FacetCollector::for_field("classification");
         facet_collector.add_facet("/Felidae");
         let facet_counts = searcher.search(&AllQuery, &facet_collector)?;
         // This lists all of the facet counts, right below "/Felidae".
@@ -95,7 +97,7 @@ fn main() -> tantivy::Result<()> {
         let facet = Facet::from("/Felidae/Pantherinae");
         let facet_term = Term::from_facet(classification, &facet);
         let facet_term_query = TermQuery::new(facet_term, IndexRecordOption::Basic);
-        let mut facet_collector = FacetCollector::for_field(classification);
+        let mut facet_collector = FacetCollector::for_field("classification");
         facet_collector.add_facet("/Felidae/Pantherinae");
         let facet_counts = searcher.search(&facet_term_query, &facet_collector)?;
         let facets: Vec<(&Facet, u64)> = facet_counts.get("/Felidae/Pantherinae").collect();

examples/faceted_search_with_tweaked_score.rs

@@ -1,3 +1,12 @@
+// # Faceted Search With Tweak Score
+//
+// This example covers the faceted search functionalities of
+// tantivy.
+//
+// We will :
+// - define a text field "name" in our schema
+// - define a facet field "classification" in our schema
+
 use std::collections::HashSet;
 
 use tantivy::collector::TopDocs;
@@ -55,8 +64,9 @@ fn main() -> tantivy::Result<()> {
                 .collect(),
         );
         let top_docs_by_custom_score =
+            // Call TopDocs with a custom tweak score
             TopDocs::with_limit(2).tweak_score(move |segment_reader: &SegmentReader| {
-                let ingredient_reader = segment_reader.facet_reader(ingredient).unwrap();
+                let ingredient_reader = segment_reader.facet_reader("ingredient").unwrap();
                 let facet_dict = ingredient_reader.facet_dict();
 
                 let query_ords: HashSet<u64> = facets
@@ -64,12 +74,10 @@ fn main() -> tantivy::Result<()> {
                     .filter_map(|key| facet_dict.term_ord(key.encoded_str()).unwrap())
                     .collect();
 
-                let mut facet_ords_buffer: Vec<u64> = Vec::with_capacity(20);
-
                 move |doc: DocId, original_score: Score| {
-                    ingredient_reader.facet_ords(doc, &mut facet_ords_buffer);
-                    let missing_ingredients = facet_ords_buffer
-                        .iter()
+                    // Update the original score with a tweaked score
+                    let missing_ingredients = ingredient_reader
+                        .facet_ords(doc)
                         .filter(|ord| !query_ords.contains(ord))
                         .count();
                     let tweak = 1.0 / 4_f32.powi(missing_ingredients as i32);

examples/fuzzy_search.rs

@@ -0,0 +1,167 @@
+// # Basic Example
+//
+// This example covers the basic functionalities of
+// tantivy.
+//
+// We will :
+// - define our schema
+// - create an index in a directory
+// - index a few documents into our index
+// - search for the best document matching a basic query
+// - retrieve the best document's original content.
+// ---
+// Importing tantivy...
+use tantivy::collector::{Count, TopDocs};
+use tantivy::query::FuzzyTermQuery;
+use tantivy::schema::*;
+use tantivy::{doc, Index, ReloadPolicy};
+use tempfile::TempDir;
+
+fn main() -> tantivy::Result<()> {
+    // Let's create a temporary directory for the
+    // sake of this example
+    let index_path = TempDir::new()?;
+
+    // # Defining the schema
+    //
+    // The Tantivy index requires a very strict schema.
+    // The schema declares which fields are in the index,
+    // and for each field, its type and "the way it should
+    // be indexed".
+
+    // First we need to define a schema ...
+    let mut schema_builder = Schema::builder();
+
+    // Our first field is title.
+    // We want full-text search for it, and we also want
+    // to be able to retrieve the document after the search.
+    //
+    // `TEXT | STORED` is some syntactic sugar to describe
+    // that.
+    //
+    // `TEXT` means the field should be tokenized and indexed,
+    // along with its term frequency and term positions.
+    //
+    // `STORED` means that the field will also be saved
+    // in a compressed, row-oriented key-value store.
+    // This store is useful for reconstructing the
+    // documents that were selected during the search phase.
+    let title = schema_builder.add_text_field("title", TEXT | STORED);
+
+    let schema = schema_builder.build();
+
+    // # Indexing documents
+    //
+    // Let's create a brand new index.
+    //
+    // This will actually just save a meta.json
+    // with our schema in the directory.
+    let index = Index::create_in_dir(&index_path, schema.clone())?;
+
+    // To insert a document we will need an index writer.
+    // There must be only one writer at a time.
+    // This single `IndexWriter` is already
+    // multithreaded.
+    //
+    // Here we give tantivy a budget of `50MB`.
+    // Using a bigger memory_arena for the indexer may increase
+    // throughput, but 50 MB is already plenty.
+    let mut index_writer = index.writer(50_000_000)?;
+
+    // Let's index our documents!
+    // We first need a handle on the title and the body field.
+
+    // ### Adding documents
+    //
+    index_writer.add_document(doc!(
+        title => "The Name of the Wind",
+    ))?;
+    index_writer.add_document(doc!(
+        title => "The Diary of Muadib",
+    ))?;
+    index_writer.add_document(doc!(
+        title => "A Dairy Cow",
+    ))?;
+    index_writer.add_document(doc!(
+        title => "The Diary of a Young Girl",
+    ))?;
+    index_writer.commit()?;
+
+    // ### Committing
+    //
+    // At this point our documents are not searchable.
+    //
+    //
+    // We need to call `.commit()` explicitly to force the
+    // `index_writer` to finish processing the documents in the queue,
+    // flush the current index to the disk, and advertise
+    // the existence of new documents.
+    //
+    // This call is blocking.
+    index_writer.commit()?;
+
+    // If `.commit()` returns correctly, then all of the
+    // documents that have been added are guaranteed to be
+    // persistently indexed.
+    //
+    // In the scenario of a crash or a power failure,
+    // tantivy behaves as if it has rolled back to its last
+    // commit.
+
+    // # Searching
+    //
+    // ### Searcher
+    //
+    // A reader is required first in order to search an index.
+    // It acts as a `Searcher` pool that reloads itself,
+    // depending on a `ReloadPolicy`.
+    //
+    // For a search server you will typically create one reader for the entire lifetime of your
+    // program, and acquire a new searcher for every single request.
+    //
+    // In the code below, we rely on the 'ON_COMMIT' policy: the reader
+    // will reload the index automatically after each commit.
+    let reader = index
+        .reader_builder()
+        .reload_policy(ReloadPolicy::OnCommit)
+        .try_into()?;
+
+    // We now need to acquire a searcher.
+    //
+    // A searcher points to a snapshotted, immutable version of the index.
+    //
+    // Some search experience might require more than
+    // one query. Using the same searcher ensures that all of these queries will run on the
+    // same version of the index.
+    //
+    // Acquiring a `searcher` is very cheap.
+    //
+    // You should acquire a searcher every time you start processing a request and
+    // and release it right after your query is finished.
+    let searcher = reader.searcher();
+
+    // ### FuzzyTermQuery
+    {
+        let term = Term::from_field_text(title, "Diary");
+        let query = FuzzyTermQuery::new(term, 2, true);
+
+        let (top_docs, count) = searcher
+            .search(&query, &(TopDocs::with_limit(5), Count))
+            .unwrap();
+        assert_eq!(count, 3);
+        assert_eq!(top_docs.len(), 3);
+        for (score, doc_address) in top_docs {
+            let retrieved_doc = searcher.doc(doc_address)?;
+            // Note that the score is not lower for the fuzzy hit.
+            // There's an issue open for that: https://github.com/quickwit-oss/tantivy/issues/563
+            println!("score {score:?} doc {}", schema.to_json(&retrieved_doc));
+            // score 1.0 doc {"title":["The Diary of Muadib"]}
+            //
+            // score 1.0 doc {"title":["The Diary of a Young Girl"]}
+            //
+            // score 1.0 doc {"title":["A Dairy Cow"]}
+        }
+    }
+
+    Ok(())
+}

examples/integer_range_search.rs

@@ -27,7 +27,7 @@ fn main() -> Result<()> {
     reader.reload()?;
     let searcher = reader.searcher();
     // The end is excluded i.e. here we are searching up to 1969
-    let docs_in_the_sixties = RangeQuery::new_u64(year_field, 1960..1970);
+    let docs_in_the_sixties = RangeQuery::new_u64("year".to_string(), 1960..1970);
     // Uses a Count collector to sum the total number of docs in the range
     let num_60s_books = searcher.search(&docs_in_the_sixties, &Count)?;
     assert_eq!(num_60s_books, 10);

examples/ip_field.rs

@@ -0,0 +1,107 @@
+// # 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
+    // We set the IP field as `INDEXED`, so it can be searched
+    // `FAST` will create a fast field. The fast field will be used to execute search queries.
+    // `FAST` is not a requirement for range queries, it can also be executed on the inverted index
+    // which is created by `INDEXED`.
+    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)?;
+
+    // ### IPv4
+    // Adding documents that contain an IPv4 address. Notice that the IP addresses are passed as
+    // `String`. Since the field is of type ip, we parse the IP address from the string and store it
+    // internally as IPv6.
+    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)?;
+    // ### IPv6
+    // Adding a document that contains an IPv6 address.
+    let doc = schema.parse_document(
+        r#"{
+            "ip": "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
+            "event_type": "checkout"
+        }"#,
+    )?;
+
+    index_writer.add_document(doc)?;
+    // Commit will create a segment containing our documents.
+    index_writer.commit()?;
+
+    let reader = index.reader()?;
+    let searcher = reader.searcher();
+
+    // # Search
+    // Range queries on IPv4. Since we created a fast field, the fast field will be used to execute
+    // the search.
+    // ### Range Queries
+    let query_parser = QueryParser::for_index(&index, vec![event_type, ip]);
+    {
+        // Inclusive range queries
+        let query = query_parser.parse_query("ip:[192.168.0.80 TO 192.168.0.100]")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(5))?;
+        assert_eq!(count_docs.len(), 1);
+    }
+    {
+        // Exclusive range queries
+        let query = query_parser.parse_query("ip:{192.168.0.80 TO 192.168.1.100]")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(2))?;
+        assert_eq!(count_docs.len(), 0);
+    }
+    {
+        // Find docs with IP addresses smaller equal 192.168.1.100
+        let query = query_parser.parse_query("ip:[* TO 192.168.1.100]")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(2))?;
+        assert_eq!(count_docs.len(), 2);
+    }
+    {
+        // Find docs with IP addresses smaller than 192.168.1.100
+        let query = query_parser.parse_query("ip:[* TO 192.168.1.100}")?;
+        let count_docs = searcher.search(&*query, &TopDocs::with_limit(2))?;
+        assert_eq!(count_docs.len(), 2);
+    }
+
+    // ### Exact Queries
+    // Exact search on IPv4.
+    {
+        let query = query_parser.parse_query("ip:192.168.0.80")?;
+        let count_docs = searcher.search(&*query, &Count)?;
+        assert_eq!(count_docs, 1);
+    }
+    // Exact search on IPv6.
+    // IpV6 addresses need to be quoted because they contain `:`
+    {
+        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(())
+}

examples/warmer.rs

@@ -4,7 +4,7 @@ use std::sync::{Arc, RwLock, Weak};
 
 use tantivy::collector::TopDocs;
 use tantivy::query::QueryParser;
-use tantivy::schema::{Field, Schema, FAST, TEXT};
+use tantivy::schema::{Schema, FAST, TEXT};
 use tantivy::{
     doc, DocAddress, DocId, Index, IndexReader, Opstamp, Searcher, SearcherGeneration, SegmentId,
     SegmentReader, Warmer,
@@ -17,7 +17,6 @@ use tantivy::{
 
 type ProductId = u64;
 
-/// Price
 type Price = u32;
 
 pub trait PriceFetcher: Send + Sync + 'static {
@@ -25,13 +24,13 @@ pub trait PriceFetcher: Send + Sync + 'static {
 }
 
 struct DynamicPriceColumn {
-    field: Field,
+    field: String,
     price_cache: RwLock<HashMap<(SegmentId, Option<Opstamp>), Arc<Vec<Price>>>>,
     price_fetcher: Box<dyn PriceFetcher>,
 }
 
 impl DynamicPriceColumn {
-    pub fn with_product_id_field<T: PriceFetcher>(field: Field, price_fetcher: T) -> Self {
+    pub fn with_product_id_field<T: PriceFetcher>(field: String, price_fetcher: T) -> Self {
         DynamicPriceColumn {
             field,
             price_cache: Default::default(),
@@ -48,7 +47,10 @@ impl Warmer for DynamicPriceColumn {
     fn warm(&self, searcher: &Searcher) -> tantivy::Result<()> {
         for segment in searcher.segment_readers() {
             let key = (segment.segment_id(), segment.delete_opstamp());
-            let product_id_reader = segment.fast_fields().u64(self.field)?;
+            let product_id_reader = segment
+                .fast_fields()
+                .u64(&self.field)?
+                .first_or_default_col(0);
             let product_ids: Vec<ProductId> = segment
                 .doc_ids_alive()
                 .map(|doc| product_id_reader.get_val(doc))
@@ -87,10 +89,10 @@ impl Warmer for DynamicPriceColumn {
     }
 }
 
-/// For the sake of this example, the table is just an editable HashMap behind a RwLock.
-/// This map represents a map (ProductId -> Price)
-///
-/// In practise, it could be fetching things from an external service, like a SQL table.
+// For the sake of this example, the table is just an editable HashMap behind a RwLock.
+// This map represents a map (ProductId -> Price)
+//
+// In practise, it could be fetching things from an external service, like a SQL table.
 #[derive(Default, Clone)]
 pub struct ExternalPriceTable {
     prices: Arc<RwLock<HashMap<ProductId, Price>>>,
@@ -123,7 +125,7 @@ fn main() -> tantivy::Result<()> {
 
     let price_table = ExternalPriceTable::default();
     let price_dynamic_column = Arc::new(DynamicPriceColumn::with_product_id_field(
-        product_id,
+        "product_id".to_string(),
         price_table.clone(),
     ));
     price_table.update_price(OLIVE_OIL, 12);

fastfield_codecs/Cargo.toml

@@ -1,33 +0,0 @@
-[package]
-name = "fastfield_codecs"
-version = "0.3.0"
-authors = ["Pascal Seitz <pascal@quickwit.io>"]
-license = "MIT"
-edition = "2021"
-description = "Fast field codecs used by tantivy"
-documentation = "https://docs.rs/fastfield_codecs/"
-homepage = "https://github.com/quickwit-oss/tantivy"
-repository = "https://github.com/quickwit-oss/tantivy"
-
-# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[dependencies]
-common = { version = "0.5", path = "../common/", package = "tantivy-common" }
-tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-prettytable-rs = {version="0.9.0", optional= true}
-rand = {version="0.8.3", optional= true}
-fastdivide = "0.4"
-log = "0.4"
-itertools = { version = "0.10.3" }
-measure_time = { version="0.8.2", optional=true}
-
-[dev-dependencies]
-more-asserts = "0.3.0"
-proptest = "1.0.0"
-rand = "0.8.3"
-
-[features]
-bin = ["prettytable-rs", "rand", "measure_time"]
-default = ["bin"]
-unstable = []
-

fastfield_codecs/README.md

@@ -1,68 +0,0 @@
-
-
-# Fast Field Codecs
-
-This crate contains various fast field codecs, used to compress/decompress fast field data in tantivy.
-
-## Contributing
-
-Contributing is pretty straightforward. Since the bitpacking is the simplest compressor, you can check it for reference.
-
-A codec needs to implement 2 traits:
-
-- A reader implementing `FastFieldCodecReader` to read the codec.
-- A serializer implementing `FastFieldCodecSerializer` for compression estimation and codec name + id.
-
-### Tests
-
-Once the traits are implemented test and benchmark integration is pretty easy (see `test_with_codec_data_sets` and `bench.rs`).
-
-Make sure to add the codec to the main.rs, which tests the compression ratio and estimation against different data sets. You can run it with:
-```
-cargo run --features bin
-```
-
-### TODO
-- Add real world data sets in comparison
-- Add codec to cover sparse data sets
-
-
-### Codec Comparison
-```
-+----------------------------------+-------------------+------------------------+
-|                                  | Compression Ratio | Compression Estimation |
-+----------------------------------+-------------------+------------------------+
-| Autoincrement                    |                   |                        |
-+----------------------------------+-------------------+------------------------+
-| LinearInterpol                   | 0.000039572664    | 0.000004396963         |
-+----------------------------------+-------------------+------------------------+
-| MultiLinearInterpol              | 0.1477348         | 0.17275847             |
-+----------------------------------+-------------------+------------------------+
-| Bitpacked                        | 0.28126493        | 0.28125                |
-+----------------------------------+-------------------+------------------------+
-| Monotonically increasing concave |                   |                        |
-+----------------------------------+-------------------+------------------------+
-| LinearInterpol                   | 0.25003937        | 0.26562938             |
-+----------------------------------+-------------------+------------------------+
-| MultiLinearInterpol              | 0.190665          | 0.1883836              |
-+----------------------------------+-------------------+------------------------+
-| Bitpacked                        | 0.31251436        | 0.3125                 |
-+----------------------------------+-------------------+------------------------+
-| Monotonically increasing convex  |                   |                        |
-+----------------------------------+-------------------+------------------------+
-| LinearInterpol                   | 0.25003937        | 0.28125438             |
-+----------------------------------+-------------------+------------------------+
-| MultiLinearInterpol              | 0.18676           | 0.2040086              |
-+----------------------------------+-------------------+------------------------+
-| Bitpacked                        | 0.31251436        | 0.3125                 |
-+----------------------------------+-------------------+------------------------+
-| Almost monotonically increasing  |                   |                        |
-+----------------------------------+-------------------+------------------------+
-| LinearInterpol                   | 0.14066513        | 0.1562544              |
-+----------------------------------+-------------------+------------------------+
-| MultiLinearInterpol              | 0.16335973        | 0.17275847             |
-+----------------------------------+-------------------+------------------------+
-| Bitpacked                        | 0.28126493        | 0.28125                |
-+----------------------------------+-------------------+------------------------+
-
-```

fastfield_codecs/benches/bench.rs

@@ -1,246 +0,0 @@
-#![feature(test)]
-
-extern crate test;
-
-#[cfg(test)]
-mod tests {
-    use std::iter;
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-    use fastfield_codecs::*;
-    use rand::prelude::*;
-    use test::Bencher;
-
-    use super::*;
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    fn generate_random() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64)
-            .map(|el| el + random::<u16>() as u64)
-            .collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation_gcd() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
-        column: &[T],
-    ) -> Arc<dyn Column<T>> {
-        let mut buffer = Vec::new();
-        serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
-        open(OwnedBytes::new(buffer)).unwrap()
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = permutation[a as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = column.get_val(a as u32);
-            }
-            a
-        });
-    }
-
-    fn get_exp_data() -> Vec<u64> {
-        let mut data = vec![];
-        for i in 0..100 {
-            let num = i * i;
-            data.extend(iter::repeat(i as u64).take(num));
-        }
-        data.shuffle(&mut StdRng::from_seed([1u8; 32]));
-
-        // lengt = 328350
-        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<_>>();
-        get_u128_column_from_data(&permutation)
-    }
-
-    fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
-        let mut out = vec![];
-        let iter_gen = || data.iter().cloned();
-        serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
-        let out = OwnedBytes::new(out);
-        open_u128::<u128>(out).unwrap()
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
-        let (major_item, _minor_item, 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,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
-        let (_major_item, minor_item, 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,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
-        let (_major_item, _minor_item, 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(0..=u128::MAX, 0..data.len() as u32, &mut positions);
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let mut a = 0u128;
-            for i in 0u64..column.num_vals() as u64 {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let n = column.num_vals();
-            let mut a = 0u128;
-            for i in (0..n / 5).map(|val| val * 5) {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += permutation[i as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0u32..n as u32 {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
-        let permutation = generate_permutation_gcd();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..n {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..permutation.len() {
-                a += permutation[i as usize] as u64;
-            }
-            a
-        });
-    }
-}

fastfield_codecs/src/bitpacked.rs

@@ -1,116 +0,0 @@
-use std::io::{self, Write};
-
-use common::OwnedBytes;
-use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
-
-use crate::serialize::NormalizedHeader;
-use crate::{Column, 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 Column 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 Column, 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 Column) -> 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::tests::get_codec_test_datasets;
-
-    fn create_and_validate(data: &[u64], name: &str) {
-        crate::tests::create_and_validate::<BitpackedCodec>(data, name);
-    }
-
-    #[test]
-    fn test_with_codec_data_sets() {
-        let data_sets = 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 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(&data, "rand");
-
-            data.reverse();
-            create_and_validate(&data, "rand");
-        }
-    }
-}

fastfield_codecs/src/blockwise_linear.rs

@@ -1,188 +0,0 @@
-use std::sync::Arc;
-use std::{io, iter};
-
-use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
-use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
-
-use crate::line::Line;
-use crate::serialize::NormalizedHeader;
-use crate::{Column, 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: crate::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 crate::Column) -> 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 Column, 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 Column 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
-    }
-}

fastfield_codecs/src/column.rs

@@ -1,348 +0,0 @@
-use std::marker::PhantomData;
-use std::ops::{Range, RangeInclusive};
-
-use tantivy_bitpacker::minmax;
-
-use crate::monotonic_mapping::StrictlyMonotonicFn;
-
-/// `Column` provides columnar access on a field.
-pub trait Column<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)))
-    }
-}
-
-/// VecColumn provides `Column` over a slice.
-pub struct VecColumn<'a, T = u64> {
-    values: &'a [T],
-    min_value: T,
-    max_value: T,
-}
-
-impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<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)
-    }
-}
-
-impl<'a, T: Copy + PartialOrd + Send + Sync> Column<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 Column<Output>
-where
-    C: Column<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> 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,
-{
-    #[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> Column<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::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);
-    }
-}

fastfield_codecs/src/format_version.rs

@@ -1,38 +0,0 @@
-use std::io;
-
-use common::{BinarySerializable, OwnedBytes};
-
-const MAGIC_NUMBER: u16 = 4335u16;
-const FASTFIELD_FORMAT_VERSION: u8 = 1;
-
-pub(crate) fn append_format_version(output: &mut impl io::Write) -> io::Result<()> {
-    FASTFIELD_FORMAT_VERSION.serialize(output)?;
-    MAGIC_NUMBER.serialize(output)?;
-
-    Ok(())
-}
-
-pub(crate) fn read_format_version(data: OwnedBytes) -> io::Result<(OwnedBytes, u8)> {
-    let (data, magic_number_bytes) = data.rsplit(2);
-
-    let magic_number = u16::deserialize(&mut magic_number_bytes.as_slice())?;
-    if magic_number != MAGIC_NUMBER {
-        return Err(io::Error::new(
-            io::ErrorKind::InvalidData,
-            format!("magic number mismatch {} != {}", magic_number, MAGIC_NUMBER),
-        ));
-    }
-    let (data, format_version_bytes) = data.rsplit(1);
-    let format_version = u8::deserialize(&mut format_version_bytes.as_slice())?;
-    if format_version > FASTFIELD_FORMAT_VERSION {
-        return Err(io::Error::new(
-            io::ErrorKind::InvalidData,
-            format!(
-                "Unsupported fastfield format version: {}. Max supported version: {}",
-                format_version, FASTFIELD_FORMAT_VERSION
-            ),
-        ));
-    }
-
-    Ok((data, format_version))
-}