Adding a special method to Arc<dyn ColumnValues>

* fix(CI) enable coverage on doctest
⚠️ Marked as [unstable](https://github.com/taiki-e/cargo-llvm-cov/issues/2)
refs #1761

* remove obsolete CI directory

.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

@@ -23,7 +23,7 @@ regex = { version = "1.5.5", default-features = false, features = ["std", "unico
 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 }
@@ -55,12 +55,12 @@ 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" }
-columnar = 		{ version= "0.1", path="./columnar", package="tantivy-columnar" }
-common = 								{ version= "0.5", path = "./common/", package = "tantivy-common" }
+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]

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

@@ -5,24 +5,23 @@ 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"}
-serde_json = "1"
-thiserror = "1"
-fnv = "1"
 sstable = { path = "../sstable", package = "tantivy-sstable" }
 common = { path = "../common", package = "tantivy-common" }
-itertools = "0.10"
-log = "0.4"
 tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-prettytable-rs = {version="0.10.0", optional= true}
-rand = {version="0.8.3", optional= true}
-fastdivide = "0.4"
-measure_time = { version="0.8.2", optional=true}
 
 [dev-dependencies]
 proptest = "1"
-more-asserts = "0.3.0"
-rand = "0.8.3"
+more-asserts = "0.3.1"
+rand = "0.8.5"
 
 [features]
 unstable = []

columnar/Makefile

@@ -1,6 +0,0 @@
-test:
-	echo "Run test only... No examples."
-	cargo test --tests --lib
-
-fmt:
-	cargo +nightly fmt --all

columnar/benches/bench_u128.rs

@@ -28,12 +28,7 @@ fn get_u128_column_random() -> Arc<dyn ColumnValues<u128>> {
 
 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.iter().copied()),
-        data.len() as u32,
-        &mut out,
-    )
-    .unwrap();
+    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()
 }
@@ -41,7 +36,7 @@ fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn ColumnValues<u128>> {
 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]);
 
@@ -63,7 +58,7 @@ fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(
+        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,
@@ -79,7 +74,7 @@ fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(
+        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,
@@ -95,7 +90,7 @@ fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
+        column.get_row_ids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
         positions
     });
 }

columnar/benches/bench_u64.rs

@@ -5,9 +5,7 @@ use std::ops::RangeInclusive;
 use std::sync::Arc;
 
 use rand::prelude::*;
-use tantivy_columnar::column_values::{
-    serialize_and_load_u64_based_column_values, CodecType, ALL_U64_CODEC_TYPES,
-};
+use tantivy_columnar::column_values::{serialize_and_load_u64_based_column_values, CodecType};
 use tantivy_columnar::*;
 use test::Bencher;
 
@@ -91,7 +89,7 @@ fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
     let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(
+        column.get_row_ids_for_value_range(
             FIFTY_PERCENT_RANGE,
             0..data.len() as u32,
             &mut positions,
@@ -108,7 +106,7 @@ fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(
+        column.get_row_ids_for_value_range(
             ONE_PERCENT_ITEM_RANGE,
             0..data.len() as u32,
             &mut positions,
@@ -125,7 +123,7 @@ fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(SINGLE_ITEM_RANGE, 0..data.len() as u32, &mut positions);
+        column.get_row_ids_for_value_range(SINGLE_ITEM_RANGE, 0..data.len() as u32, &mut positions);
         positions
     });
 }
@@ -138,7 +136,7 @@ fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
 
     b.iter(|| {
         let mut positions = Vec::new();
-        column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
+        column.get_row_ids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
         positions
     });
 }
@@ -180,7 +178,7 @@ fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
     b.iter(|| {
         let mut a = 0u64;
         for i in 0u32..n as u32 {
-            a += column.get_val(i);
+            a += column_ref.get_val(i);
         }
         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

@@ -1,14 +1,17 @@
 # zero to one
-* merges with non trivial mapping (deletes / sort)
-* emission of the sort mapping.
-+ muttivaued range queries restrat frm the beginning all of the time.
+
 * revisit line codec
 * removal of all rows of a column in the schema due to deletes
+* add columns from schema on merge
 * Plugging JSON
-replug examples
-
+* 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

columnar/src/column/dictionary_encoded.rs

@@ -32,7 +32,7 @@ impl BytesColumn {
 
     /// Returns the number of rows in the column.
     pub fn num_rows(&self) -> RowId {
-        self.term_ord_column.num_rows()
+        self.term_ord_column.num_docs()
     }
 
     pub fn term_ords(&self, row_id: RowId) -> impl Iterator<Item = u64> + '_ {
@@ -56,12 +56,6 @@ impl BytesColumn {
 #[derive(Clone)]
 pub struct StrColumn(BytesColumn);
 
-impl From<BytesColumn> for StrColumn {
-    fn from(bytes_col: BytesColumn) -> Self {
-        StrColumn(bytes_col)
-    }
-}
-
 impl From<StrColumn> for BytesColumn {
     fn from(str_column: StrColumn) -> BytesColumn {
         str_column.0
@@ -69,6 +63,10 @@ impl From<StrColumn> for BytesColumn {
 }
 
 impl StrColumn {
+    pub(crate) fn wrap(bytes_column: BytesColumn) -> StrColumn {
+        StrColumn(bytes_column)
+    }
+
     pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
         self.0.dictionary.as_ref()
     }

columnar/src/column/mod.rs

@@ -3,14 +3,14 @@ mod serialize;
 
 use std::fmt::Debug;
 use std::io::Write;
-use std::ops::Deref;
+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_u128, open_column_u64, serialize_column_mappable_to_u128,
-    serialize_column_mappable_to_u64,
+    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;
@@ -41,14 +41,15 @@ impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
     pub fn get_cardinality(&self) -> Cardinality {
         self.idx.get_cardinality()
     }
-    pub fn num_rows(&self) -> RowId {
+
+    pub fn num_docs(&self) -> RowId {
         match &self.idx {
             ColumnIndex::Full => self.values.num_vals() as u32,
-            ColumnIndex::Optional(optional_index) => optional_index.num_rows(),
+            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_rows()
+                col_index.num_docs()
             }
         }
     }
@@ -70,6 +71,34 @@ impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
             .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>,
+        docids: &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, docids);
+        // Convert rows to docids
+        self.idx
+            .select_batch_in_place(docids, selected_docid_range.start);
+    }
+
+    /// 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(row_id));
+    }
+
     pub fn first_or_default_col(self, default_value: T) -> Arc<dyn ColumnValues<T>> {
         Arc::new(FirstValueWithDefault {
             column: self,
@@ -122,8 +151,8 @@ impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
     fn num_vals(&self) -> u32 {
         match &self.column.idx {
             ColumnIndex::Full => self.column.values.num_vals(),
-            ColumnIndex::Optional(optional_idx) => optional_idx.num_rows(),
-            ColumnIndex::Multivalued(_) => todo!(),
+            ColumnIndex::Optional(optional_idx) => optional_idx.num_docs(),
+            ColumnIndex::Multivalued(multivalue_idx) => multivalue_idx.num_docs(),
         }
     }
 }

columnar/src/column/serialize.rs

@@ -1,4 +1,3 @@
-use std::fmt::Debug;
 use std::io;
 use std::io::Write;
 use std::sync::Arc;
@@ -12,20 +11,20 @@ use crate::column_values::serialize::serialize_column_values_u128;
 use crate::column_values::u64_based::{serialize_u64_based_column_values, CodecType};
 use crate::column_values::{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>,
-    num_vals: u32,
     output: &mut impl Write,
 ) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_column_values_u128(iterable, num_vals, 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 + Debug>(
+pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64>(
     column_index: SerializableColumnIndex<'_>,
     column_values: &impl Iterable<T>,
     output: &mut impl Write,
@@ -77,15 +76,19 @@ pub fn open_column_u128<T: MonotonicallyMappableToU128>(
     })
 }
 
-pub fn open_column_bytes<T: From<BytesColumn>>(data: OwnedBytes) -> io::Result<T> {
+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)?;
-    let bytes_column = BytesColumn {
+    Ok(BytesColumn {
         dictionary,
         term_ord_column,
-    };
-    Ok(bytes_column.into())
+    })
+}
+
+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 +1,167 @@
+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::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

@@ -1,29 +1,19 @@
 use std::iter;
 
-use crate::column_index::{
-    multivalued_index, serialize_column_index, SerializableColumnIndex, Set,
-};
+use crate::column_index::{SerializableColumnIndex, Set};
 use crate::iterable::Iterable;
-use crate::{Cardinality, ColumnIndex, MergeRowOrder, RowId, StackMergeOrder};
+use crate::{Cardinality, ColumnIndex, RowId, StackMergeOrder};
 
-fn detect_cardinality(columns: &[Option<ColumnIndex>]) -> Cardinality {
-    columns
-        .iter()
-        .flatten()
-        .map(ColumnIndex::get_cardinality)
-        .max()
-        .unwrap_or(Cardinality::Full)
-}
-
-pub fn stack_column_index<'a>(
+/// 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>],
-    merge_row_order: &'a MergeRowOrder,
+    cardinality_after_merge: Cardinality,
+    stack_merge_order: &'a StackMergeOrder,
 ) -> SerializableColumnIndex<'a> {
-    let MergeRowOrder::Stack(stack_merge_order) = merge_row_order else {
-        panic!("only supporting stacking at the moment.");
-    };
-    let cardinality = detect_cardinality(columns);
-    match cardinality {
+    match cardinality_after_merge {
         Cardinality::Full => SerializableColumnIndex::Full,
         Cardinality::Optional => SerializableColumnIndex::Optional {
             non_null_row_ids: Box::new(StackedOptionalIndex {
@@ -60,7 +50,7 @@ impl<'a> Iterable<RowId> for StackedOptionalIndex<'a> {
                         Some(ColumnIndex::Optional(optional_index)) => Box::new(
                             optional_index
                                 .iter_rows()
-                                .map(move |row_id: RowId| row_id + columnar_row_range.start),
+                                .map(move |row_id: RowId| columnar_row_range.start + row_id),
                         ),
                         Some(ColumnIndex::Multivalued(_)) => {
                             panic!("No multivalued index is allowed when stacking column index");
@@ -141,13 +131,6 @@ fn stack_multivalued_indexes<'a>(
     }))
 }
 
-fn stack_multivalued_index<'a>(
-    columns: &'a [Option<ColumnIndex>],
-    stack_merge_order: &StackMergeOrder,
-) -> Box<dyn Iterable<RowId> + 'a> {
-    todo!()
-}
-
 #[cfg(test)]
 mod tests {
     use crate::RowId;

columnar/src/column_index/mod.rs

@@ -5,12 +5,12 @@ mod serialize;
 
 use std::ops::Range;
 
-pub use merge::stack_column_index;
+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, RowId};
+use crate::{Cardinality, DocId, RowId};
 
 #[derive(Clone)]
 pub enum ColumnIndex {
@@ -21,6 +21,18 @@ pub enum ColumnIndex {
     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 {
     pub fn get_cardinality(&self) -> Cardinality {
         match self {
@@ -30,21 +42,52 @@ impl ColumnIndex {
         }
     }
 
-    pub fn value_row_ids(&self, row_id: RowId) -> Range<RowId> {
+    /// 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::Full => row_id..row_id + 1,
+            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::Full => doc_id..doc_id + 1,
             ColumnIndex::Optional(optional_index) => {
-                if let Some(val) = optional_index.rank_if_exists(row_id) {
+                if let Some(val) = optional_index.rank_if_exists(doc_id) {
                     val..val + 1
                 } else {
                     0..0
                 }
             }
-            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(row_id),
+            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(doc_id),
         }
     }
 
-    pub fn select_batch_in_place(&self, rank_ids: &mut Vec<RowId>) {
+    pub fn docid_range_to_rowids(&self, doc_id: Range<DocId>) -> Range<RowId> {
+        match self {
+            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, rank_ids: &mut Vec<RowId>, doc_id_start: DocId) {
         match self {
             ColumnIndex::Full => {
                 // No need to do anything:
@@ -54,8 +97,7 @@ impl ColumnIndex {
                 optional_index.select_batch(&mut rank_ids[..]);
             }
             ColumnIndex::Multivalued(multivalued_index) => {
-                // TODO important: avoid using 0u32, and restart from the beginning all of the time.
-                multivalued_index.select_batch_in_place(0u32, rank_ids)
+                multivalued_index.select_batch_in_place(doc_id_start, rank_ids)
             }
         }
     }

columnar/src/column_index/multivalued_index.rs

@@ -8,7 +8,7 @@ use common::OwnedBytes;
 use crate::column_values::u64_based::CodecType;
 use crate::column_values::ColumnValues;
 use crate::iterable::Iterable;
-use crate::RowId;
+use crate::{DocId, RowId};
 
 pub fn serialize_multivalued_index(
     multivalued_index: &dyn Iterable<RowId>,
@@ -42,23 +42,30 @@ impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {
 }
 
 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, row_id: RowId) -> Range<RowId> {
-        let start = self.start_index_column.get_val(row_id);
-        let end = self.start_index_column.get_val(row_id + 1);
+    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_rows(&self) -> u32 {
+    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
-    /// row_ids. Positions are converted inplace to docids.
+    /// 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.
@@ -69,14 +76,14 @@ impl MultiValueIndex {
     /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
     /// match a docid to its value position.
     #[allow(clippy::bool_to_int_with_if)]
-    pub(crate) fn select_batch_in_place(&self, row_start: RowId, ranks: &mut Vec<u32>) {
+    pub(crate) fn select_batch_in_place(&self, docid_start: DocId, ranks: &mut Vec<u32>) {
         if ranks.is_empty() {
             return;
         }
-        let mut cur_doc = row_start;
+        let mut cur_doc = docid_start;
         let mut last_doc = None;
 
-        assert!(self.start_index_column.get_val(row_start) as u32 <= ranks[0]);
+        assert!(self.start_index_column.get_val(docid_start) as u32 <= ranks[0]);
 
         let mut write_doc_pos = 0;
         for i in 0..ranks.len() {
@@ -120,7 +127,7 @@ mod tests {
         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_rows(), 5);
+        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]);

columnar/src/column_index/optional_index/mod.rs

@@ -11,7 +11,7 @@ use set_block::{
 };
 
 use crate::iterable::Iterable;
-use crate::{InvalidData, RowId};
+use crate::{DocId, InvalidData, RowId};
 
 /// The threshold for for number of elements after which we switch to dense block encoding.
 ///
@@ -88,22 +88,6 @@ pub struct OptionalIndex {
     block_metas: Arc<[BlockMeta]>,
 }
 
-impl OptionalIndex {
-    pub fn num_rows(&self) -> RowId {
-        self.num_rows
-    }
-
-    pub fn num_non_nulls(&self) -> RowId {
-        self.num_non_null_rows
-    }
-
-    pub fn iter_rows<'a>(&'a self) -> impl Iterator<Item = RowId> + 'a {
-        // TODO optimize
-        let mut select_batch = self.select_cursor();
-        (0..self.num_non_null_rows).map(move |rank| select_batch.select(rank))
-    }
-}
-
 /// 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
@@ -193,11 +177,11 @@ impl Set<RowId> for OptionalIndex {
     }
 
     #[inline]
-    fn rank(&self, row_id: RowId) -> RowId {
+    fn rank(&self, doc_id: DocId) -> RowId {
         let RowAddr {
             block_id,
             in_block_row_id,
-        } = row_addr_from_row_id(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 {
@@ -208,11 +192,11 @@ impl Set<RowId> for OptionalIndex {
     }
 
     #[inline]
-    fn rank_if_exists(&self, row_id: RowId) -> Option<RowId> {
+    fn rank_if_exists(&self, doc_id: DocId) -> Option<RowId> {
         let RowAddr {
             block_id,
             in_block_row_id,
-        } = row_addr_from_row_id(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 {
@@ -236,7 +220,7 @@ impl Set<RowId> for OptionalIndex {
         block_doc_idx_start + in_block_rank as u32
     }
 
-    fn select_cursor<'b>(&'b self) -> OptionalIndexSelectCursor<'b> {
+    fn select_cursor(&self) -> OptionalIndexSelectCursor<'_> {
         OptionalIndexSelectCursor {
             current_block_cursor: BlockSelectCursor::Sparse(
                 SparseBlockCodec::open(b"").select_cursor(),
@@ -251,6 +235,31 @@ impl Set<RowId> for OptionalIndex {
 }
 
 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() {
@@ -259,7 +268,7 @@ impl OptionalIndex {
     }
 
     #[inline]
-    fn block<'a>(&'a self, block_meta: BlockMeta) -> Block<'a> {
+    fn block(&self, block_meta: BlockMeta) -> Block<'_> {
         let BlockMeta {
             start_byte_offset,
             block_variant,
@@ -342,7 +351,7 @@ fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -
     Ok(())
 }
 
-pub fn serialize_optional_index<'a, W: io::Write>(
+pub fn serialize_optional_index<W: io::Write>(
     non_null_rows: &dyn Iterable<RowId>,
     num_rows: RowId,
     output: &mut W,
@@ -418,7 +427,7 @@ impl SerializedBlockMeta {
     }
 
     #[inline]
-    fn to_bytes(&self) -> [u8; SERIALIZED_BLOCK_META_NUM_BYTES] {
+    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());
@@ -492,7 +501,7 @@ pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
         num_non_empty_block_bytes as usize * SERIALIZED_BLOCK_META_NUM_BYTES;
     let (block_data, block_metas) = bytes.rsplit(block_metas_num_bytes);
     let (block_metas, num_non_null_rows) =
-        deserialize_optional_index_block_metadatas(block_metas.as_slice(), num_rows).into();
+        deserialize_optional_index_block_metadatas(block_metas.as_slice(), num_rows);
     let optional_index = OptionalIndex {
         num_rows,
         num_non_null_rows,

columnar/src/column_index/optional_index/set.rs

@@ -10,7 +10,7 @@ pub trait SetCodec {
     ///
     /// May panic if the elements are not sorted.
     fn serialize(els: impl Iterator<Item = Self::Item>, wrt: impl io::Write) -> io::Result<()>;
-    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a>;
+    fn open(data: &[u8]) -> Self::Reader<'_>;
 }
 
 /// Stateful object that makes it possible to compute several select in a row,
@@ -43,5 +43,5 @@ pub trait Set<T> {
     fn select(&self, rank: T) -> T;
 
     /// Creates a brand new select cursor.
-    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b>;
+    fn select_cursor(&self) -> Self::SelectCursor<'_>;
 }

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

@@ -45,7 +45,7 @@ impl SetCodec for DenseBlockCodec {
     }
 
     #[inline]
-    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
+    fn open(data: &[u8]) -> Self::Reader<'_> {
         assert_eq!(data.len(), DENSE_BLOCK_NUM_BYTES as usize);
         DenseBlock(data)
     }
@@ -94,7 +94,7 @@ impl DenseMiniBlock {
         Self { bitvec, rank }
     }
 
-    fn to_bytes(&self) -> [u8; MINI_BLOCK_NUM_BYTES] {
+    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());
@@ -166,7 +166,7 @@ impl<'a> Set<u16> for DenseBlock<'a> {
     }
 
     #[inline(always)]
-    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
         DenseBlockSelectCursor {
             block_id: 0,
             dense_block: *self,

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

@@ -16,7 +16,7 @@ impl SetCodec for SparseBlockCodec {
         Ok(())
     }
 
-    fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
+    fn open(data: &[u8]) -> Self::Reader<'_> {
         SparseBlock(data)
     }
 }
@@ -56,7 +56,7 @@ impl<'a> Set<u16> for SparseBlock<'a> {
     }
 
     #[inline(always)]
-    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
+    fn select_cursor(&self) -> Self::SelectCursor<'_> {
         *self
     }
 }

columnar/src/column_index/optional_index/tests.rs

@@ -107,59 +107,43 @@ fn test_null_index(data: &[bool]) {
 
 #[test]
 fn test_optional_index_test_translation() {
-    let mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as u32, &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    let mut select_cursor = null_index.select_cursor();
+    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 mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as RowId, &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert_eq!(null_index.rank_if_exists(0), Some(0));
-    assert_eq!(null_index.rank_if_exists(2), Some(1));
+    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 mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as RowId, &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert!(null_index.contains(0));
-    assert!(!null_index.contains(1));
-    assert!(null_index.contains(2));
-    assert!(!null_index.contains(3));
+    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 mut docs = vec![];
-    docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
-    docs.extend((0..=1).map(|_idx| true));
-
-    let mut out = vec![];
-    serialize_optional_index(&&docs[..], docs.len() as RowId, &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert!(!null_index.contains(0));
-    assert!(!null_index.contains(100));
-    assert!(!null_index.contains(ELEMENTS_PER_BLOCK - 1));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK + 1));
+    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 mut buffer: Vec<u8> = Vec::new();
-    serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(buffer)).unwrap();
-    assert_eq!(null_index.num_rows(), num_rows);
-    assert!(null_index.iter_rows().eq(row_ids.iter().copied()));
+    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]
@@ -168,11 +152,9 @@ fn test_optional_index_iter_empty() {
 }
 
 fn test_optional_index_rank_aux(row_ids: &[RowId]) {
-    let mut buffer: Vec<u8> = Vec::new();
     let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
-    serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(buffer)).unwrap();
-    assert_eq!(null_index.num_rows(), num_rows);
+    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));
@@ -207,6 +189,16 @@ fn test_optional_index_iter_dense_block() {
     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 {
 
@@ -220,10 +212,13 @@ mod bench {
     fn gen_bools(fill_ratio: f64) -> OptionalIndex {
         let mut out = Vec::new();
         let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
-        let vals: Vec<bool> = (0..TOTAL_NUM_VALUES)
+        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[..], &mut out).unwrap();
+        serialize_optional_index(&&vals[..], TOTAL_NUM_VALUES, &mut out).unwrap();
         let codec = open_optional_index(OwnedBytes::new(out)).unwrap();
         codec
     }

columnar/src/column_values/column.rs

@@ -6,7 +6,7 @@ use std::sync::Arc;
 use tantivy_bitpacker::minmax;
 
 use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
-use crate::iterable::Iterable;
+use crate::RowId;
 
 /// `ColumnValues` provides access to a dense field column.
 ///
@@ -36,21 +36,21 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
         }
     }
 
-    /// Get the positions of values which are in the provided value range.
+    /// 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_docids_for_value_range(
+    fn get_row_ids_for_value_range(
         &self,
         value_range: RangeInclusive<T>,
-        doc_id_range: Range<u32>,
-        positions: &mut Vec<u32>,
+        row_id_range: Range<RowId>,
+        row_id_hits: &mut Vec<RowId>,
     ) {
-        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 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) {
-                positions.push(idx);
+                row_id_hits.push(idx);
             }
         }
     }
@@ -80,12 +80,6 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     }
 }
 
-impl<'a, T: PartialOrd> Iterable<T> for &'a [Arc<dyn ColumnValues<T>>] {
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
-        Box::new(self.iter().flat_map(|column_value| column_value.iter()))
-    }
-}
-
 impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
     #[inline(always)]
     fn get_val(&self, idx: u32) -> T {
@@ -116,31 +110,14 @@ impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>>
     fn get_range(&self, start: u64, output: &mut [T]) {
         self.as_ref().get_range(start, output)
     }
-}
 
-impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd + Debug> ColumnValues<T> for &'a C {
-    fn get_val(&self, idx: u32) -> T {
-        (*self).get_val(idx)
-    }
-
-    fn min_value(&self) -> T {
-        (*self).min_value()
-    }
-
-    fn max_value(&self) -> T {
-        (*self).max_value()
-    }
-
-    fn num_vals(&self) -> u32 {
-        (*self).num_vals()
-    }
-
-    fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
-        (*self).iter()
-    }
-
-    fn get_range(&self, start: u64, output: &mut [T]) {
-        (*self).get_range(start, output)
+    fn get_row_ids_for_value_range(
+        &self,
+        value_range: RangeInclusive<T>,
+        row_id_range: Range<RowId>,
+        row_id_hits: &mut Vec<RowId>,
+    ) {
+        self.as_ref().get_row_ids_for_value_range(value_range, row_id_range, row_id_hits)
     }
 }
 
@@ -264,13 +241,13 @@ where
         )
     }
 
-    fn get_docids_for_value_range(
+    fn get_row_ids_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.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,

columnar/src/column_values/compact_space/mod.rs

@@ -22,6 +22,7 @@ use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
 
 use crate::column_values::compact_space::build_compact_space::get_compact_space;
 use crate::column_values::ColumnValues;
+use crate::RowId;
 
 mod blank_range;
 mod build_compact_space;
@@ -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();
@@ -305,7 +313,7 @@ impl ColumnValues<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>,
@@ -450,364 +458,352 @@ impl CompactSpaceDecompressor {
     }
 }
 
-// TODO reenable what can be reenabled.
-// #[cfg(test)]
-// mod tests {
-//
-// use super::*;
-// use crate::column::format_version::read_format_version;
-// use crate::column::column_footer::read_null_index_footer;
-// use crate::column::serialize::U128Header;
-// use crate::column::{open_u128, serialize_u128};
-//
-// #[test]
-// fn compact_space_test() {
-// let ips = &[
-// 2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
-// ]
-// .into_iter()
-// .collect();
-// let compact_space = get_compact_space(ips, ips.len() as u32, 11);
-// let amplitude = compact_space.amplitude_compact_space();
-// assert_eq!(amplitude, 17);
-// assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
-// assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
-// assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
-//
-// for (num1, num2) in (0..3).tuple_windows() {
-// assert_eq!(
-// compact_space.get_range_mapping(num1).compact_end() + 1,
-// compact_space.get_range_mapping(num2).compact_start
-// );
-// }
-//
-// let mut output: Vec<u8> = Vec::new();
-// compact_space.serialize(&mut output).unwrap();
-//
-// assert_eq!(
-// compact_space,
-// CompactSpace::deserialize(&mut &output[..]).unwrap()
-// );
-//
-// for ip in ips {
-// let compact = compact_space.u128_to_compact(*ip).unwrap();
-// assert_eq!(compact_space.compact_to_u128(compact), *ip);
-// }
-// }
-//
-// #[test]
-// fn compact_space_amplitude_test() {
-// let ips = &[100000u128, 1000000].into_iter().collect();
-// let compact_space = get_compact_space(ips, ips.len() as u32, 1);
-// let amplitude = compact_space.amplitude_compact_space();
-// assert_eq!(amplitude, 2);
-// }
-//
-// fn test_all(mut data: OwnedBytes, expected: &[u128]) {
-// let _header = U128Header::deserialize(&mut data);
-// let decompressor = CompactSpaceDecompressor::open(data).unwrap();
-// for (idx, expected_val) in expected.iter().cloned().enumerate() {
-// let val = decompressor.get(idx as u32);
-// assert_eq!(val, expected_val);
-//
-// let test_range = |range: RangeInclusive<u128>| {
-// let expected_positions = expected
-// .iter()
-// .positions(|val| range.contains(val))
-// .map(|pos| pos as u32)
-// .collect::<Vec<_>>();
-// let mut positions = Vec::new();
-// decompressor.get_positions_for_value_range(
-// range,
-// 0..decompressor.num_vals(),
-// &mut positions,
-// );
-// assert_eq!(positions, expected_positions);
-// };
-//
-// test_range(expected_val.saturating_sub(1)..=expected_val);
-// test_range(expected_val..=expected_val);
-// test_range(expected_val..=expected_val.saturating_add(1));
-// test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
-// }
-// }
-//
-// fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
-// let mut out = Vec::new();
-// serialize_u128(
-// || u128_vals.iter().cloned(),
-// u128_vals.len() as u32,
-// &mut out,
-// )
-// .unwrap();
-//
-// let data = OwnedBytes::new(out);
-// let (data, _format_version) = read_format_version(data).unwrap();
-// let (data, _null_index_footer) = read_null_index_footer(data).unwrap();
-// test_all(data.clone(), u128_vals);
-//
-// data
-// }
-//
-// #[test]
-// fn test_range_1() {
-// let vals = &[
-// 1u128,
-// 100u128,
-// 3u128,
-// 99999u128,
-// 100000u128,
-// 100001u128,
-// 4_000_211_221u128,
-// 4_000_211_222u128,
-// 333u128,
-// ];
-// let mut data = test_aux_vals(vals);
-//
-// let _header = U128Header::deserialize(&mut data);
-// let decomp = CompactSpaceDecompressor::open(data).unwrap();
-// let complete_range = 0..vals.len() as u32;
-// for (pos, val) in vals.iter().enumerate() {
-// let val = *val;
-// let pos = pos as u32;
-// let mut positions = Vec::new();
-// decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
-// assert_eq!(positions, vec![pos]);
-// }
-//
-// handle docid range out of bounds
-// let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
-// assert!(positions.is_empty());
-//
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
-// assert_eq!(positions, vec![0]);
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
-// assert_eq!(positions, vec![0]);
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
-// assert_eq!(positions, vec![0, 2]);
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99999u128..=99999u128,
-// complete_range.clone()
-// ),
-// vec![3]
-// );
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99999u128..=100000u128,
-// complete_range.clone()
-// ),
-// vec![3, 4]
-// );
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=100000u128,
-// complete_range.clone()
-// ),
-// vec![3, 4]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=99999u128,
-// complete_range.clone()
-// ),
-// &[3]
-// );
-// assert!(get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=99998u128,
-// complete_range.clone()
-// )
-// .is_empty());
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 333u128..=333u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 332u128..=333u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 332u128..=334u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 333u128..=334u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-//
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 4_000_211_221u128..=5_000_000_000u128,
-// complete_range
-// ),
-// &[6, 7]
-// );
-// }
-//
-// #[test]
-// fn test_empty() {
-// let vals = &[];
-// let data = test_aux_vals(vals);
-// let _decomp = CompactSpaceDecompressor::open(data).unwrap();
-// }
-//
-// #[test]
-// fn test_range_2() {
-// let vals = &[
-// 100u128,
-// 99999u128,
-// 100000u128,
-// 100001u128,
-// 4_000_211_221u128,
-// 4_000_211_222u128,
-// 333u128,
-// ];
-// let mut data = test_aux_vals(vals);
-// let _header = U128Header::deserialize(&mut data);
-// let decomp = CompactSpaceDecompressor::open(data).unwrap();
-// let complete_range = 0..vals.len() as u32;
-// assert!(
-// &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
-// .is_empty(),
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
-// &[0]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
-// &[0]
-// );
-// }
-//
-// fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd>(
-// column: &C,
-// value_range: RangeInclusive<T>,
-// doc_id_range: Range<u32>,
-// ) -> Vec<u32> {
-// let mut positions = Vec::new();
-// column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
-// positions
-// }
-//
-// #[test]
-// fn test_range_3() {
-// let vals = &[
-// 200u128,
-// 201,
-// 202,
-// 203,
-// 204,
-// 204,
-// 206,
-// 207,
-// 208,
-// 209,
-// 210,
-// 1_000_000,
-// 5_000_000_000,
-// ];
-// let mut out = Vec::new();
-// serialize_u128(|| vals.iter().cloned(), vals.len() as u32, &mut out).unwrap();
-// let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
-// let complete_range = 0..vals.len() as u32;
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
-// vec![0]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
-// vec![0, 1]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
-// vec![0]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
-// vec![11]
-// );
-// }
-//
-// #[test]
-// fn test_bug1() {
-// let vals = &[9223372036854775806];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug2() {
-// let vals = &[340282366920938463463374607431768211455u128];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug3() {
-// let vals = &[340282366920938463463374607431768211454];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug4() {
-// let vals = &[340282366920938463463374607431768211455, 0];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_first_large_gaps() {
-// let vals = &[1_000_000_000u128; 100];
-// let _data = test_aux_vals(vals);
-// }
-// use itertools::Itertools;
-// use proptest::prelude::*;
-//
-// fn num_strategy() -> impl Strategy<Value = u128> {
-// prop_oneof![
-// 1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
-// 20 => prop::num::u128::ANY,
-// ]
-// }
-//
-// proptest! {
-// #![proptest_config(ProptestConfig::with_cases(10))]
-//
-// #[test]
-// fn compress_decompress_random(vals in proptest::collection::vec(num_strategy()
-// , 1..1000)) {
-// let _data = test_aux_vals(&vals);
-// }
-// }
-// }
-//
+#[cfg(test)]
+mod tests {
+
+    use itertools::Itertools;
+
+    use super::*;
+    use crate::column_values::serialize::U128Header;
+    use crate::column_values::{open_u128_mapped, serialize_column_values_u128};
+
+    #[test]
+    fn compact_space_test() {
+        let ips = &[
+            2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
+        ]
+        .into_iter()
+        .collect();
+        let compact_space = get_compact_space(ips, ips.len() as u32, 11);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 17);
+        assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
+        assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
+        assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
+
+        for (num1, num2) in (0..3).tuple_windows() {
+            assert_eq!(
+                compact_space.get_range_mapping(num1).compact_end() + 1,
+                compact_space.get_range_mapping(num2).compact_start
+            );
+        }
+
+        let mut output: Vec<u8> = Vec::new();
+        compact_space.serialize(&mut output).unwrap();
+
+        assert_eq!(
+            compact_space,
+            CompactSpace::deserialize(&mut &output[..]).unwrap()
+        );
+
+        for ip in ips {
+            let compact = compact_space.u128_to_compact(*ip).unwrap();
+            assert_eq!(compact_space.compact_to_u128(compact), *ip);
+        }
+    }
+
+    #[test]
+    fn compact_space_amplitude_test() {
+        let ips = &[100000u128, 1000000].into_iter().collect();
+        let compact_space = get_compact_space(ips, ips.len() as u32, 1);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 2);
+    }
+
+    fn test_all(mut data: OwnedBytes, expected: &[u128]) {
+        let _header = U128Header::deserialize(&mut data);
+        let decompressor = CompactSpaceDecompressor::open(data).unwrap();
+        for (idx, expected_val) in expected.iter().cloned().enumerate() {
+            let val = decompressor.get(idx as u32);
+            assert_eq!(val, expected_val);
+
+            let test_range = |range: RangeInclusive<u128>| {
+                let expected_positions = expected
+                    .iter()
+                    .positions(|val| range.contains(val))
+                    .map(|pos| pos as u32)
+                    .collect::<Vec<_>>();
+                let mut positions = Vec::new();
+                decompressor.get_positions_for_value_range(
+                    range,
+                    0..decompressor.num_vals(),
+                    &mut positions,
+                );
+                assert_eq!(positions, expected_positions);
+            };
+
+            test_range(expected_val.saturating_sub(1)..=expected_val);
+            test_range(expected_val..=expected_val);
+            test_range(expected_val..=expected_val.saturating_add(1));
+            test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
+        }
+    }
+
+    fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
+        let mut out = Vec::new();
+        serialize_column_values_u128(&u128_vals, &mut out).unwrap();
+        let data = OwnedBytes::new(out);
+        test_all(data.clone(), u128_vals);
+        data
+    }
+
+    #[test]
+    fn test_range_1() {
+        let vals = &[
+            1u128,
+            100u128,
+            3u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        for (pos, val) in vals.iter().enumerate() {
+            let val = *val;
+            let pos = pos as u32;
+            let mut positions = Vec::new();
+            decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
+            assert_eq!(positions, vec![pos]);
+        }
+
+        // handle docid range out of bounds
+        let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
+        assert!(positions.is_empty());
+
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
+        assert_eq!(positions, vec![0, 2]);
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=99999u128,
+                complete_range.clone()
+            ),
+            vec![3]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=99999u128,
+                complete_range.clone()
+            ),
+            &[3]
+        );
+        assert!(get_positions_for_value_range_helper(
+            &decomp,
+            99998u128..=99998u128,
+            complete_range.clone()
+        )
+        .is_empty());
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                4_000_211_221u128..=5_000_000_000u128,
+                complete_range
+            ),
+            &[6, 7]
+        );
+    }
+
+    #[test]
+    fn test_empty() {
+        let vals = &[];
+        let data = test_aux_vals(vals);
+        let _decomp = CompactSpaceDecompressor::open(data).unwrap();
+    }
+
+    #[test]
+    fn test_range_2() {
+        let vals = &[
+            100u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        assert!(
+            &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
+                .is_empty(),
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
+            &[0]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
+            &[0]
+        );
+    }
+
+    fn get_positions_for_value_range_helper<C: 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_row_ids_for_value_range(value_range, doc_id_range, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_range_3() {
+        let vals = &[
+            200u128,
+            201,
+            202,
+            203,
+            204,
+            204,
+            206,
+            207,
+            208,
+            209,
+            210,
+            1_000_000,
+            5_000_000_000,
+        ];
+        let mut out = Vec::new();
+        serialize_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!(
+            get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
+            vec![0, 1]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
+            vec![11]
+        );
+    }
+
+    #[test]
+    fn test_bug1() {
+        let vals = &[9223372036854775806];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug2() {
+        let vals = &[340282366920938463463374607431768211455u128];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug3() {
+        let vals = &[340282366920938463463374607431768211454];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug4() {
+        let vals = &[340282366920938463463374607431768211455, 0];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_first_large_gaps() {
+        let vals = &[1_000_000_000u128; 100];
+        let _data = test_aux_vals(vals);
+    }
+
+    use proptest::prelude::*;
+
+    fn num_strategy() -> impl Strategy<Value = u128> {
+        prop_oneof![
+            1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
+            20 => prop::num::u128::ANY,
+        ]
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(10))]
+
+        #[test]
+        fn compress_decompress_random(vals in proptest::collection::vec(num_strategy() , 1..1000)) {
+            let _data = test_aux_vals(&vals);
+        }
+    }
+}

columnar/src/column_values/main.rs

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

columnar/src/column_values/mod.rs

@@ -1,5 +1,4 @@
 #![warn(missing_docs)]
-#![cfg_attr(all(feature = "unstable", test), feature(test))]
 
 //! # `fastfield_codecs`
 //!
@@ -26,16 +25,58 @@ mod stats;
 pub(crate) mod u64_based;
 
 mod column;
-pub mod serialize;
+pub(crate) mod serialize;
 
 pub use serialize::serialize_column_values_u128;
-pub use stats::Stats;
+pub use stats::ColumnStats;
 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 self::column::{monotonic_map_column, ColumnValues, IterColumn, VecColumn};
+use crate::iterable::Iterable;
+use crate::{ColumnIndex, 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 Some(column_index) = self.column_indexes[row_addr.segment_ord as usize].as_ref() else {
+                            return None;
+                        };
+                        let Some(column_values) = self.column_values[row_addr.segment_ord as usize].as_ref() else {
+                            return None;
+                        };
+                        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))
+                    })
+                )
+            },
+        }
+    }
+}
 
 #[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
 #[repr(u8)]
@@ -95,6 +136,7 @@ mod bench {
     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);
@@ -110,23 +152,30 @@ mod bench {
         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: FastFieldCodec>(data: &[u64]) -> Codec::Reader {
+    fn get_reader_for_bench<Codec: ColumnCodec>(data: &[u64]) -> Codec::ColumnValues {
         let mut bytes = Vec::new();
-        let min_value = *data.iter().min().unwrap();
-        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
-        let col = VecColumn::from(&data);
-        let normalized_header = NormalizedHeader {
-            num_vals: col.num_vals(),
-            max_value: col.max_value(),
-        };
-        Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
-        Codec::open_from_bytes(OwnedBytes::new(bytes), normalized_header).unwrap()
+        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: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
+    fn bench_get<Codec: ColumnCodec>(b: &mut Bencher, data: &[u64]) {
         let col = get_reader_for_bench::<Codec>(data);
         b.iter(|| {
             let mut sum = 0u64;
@@ -150,18 +199,22 @@ mod bench {
         });
     }
 
-    fn bench_get_dynamic<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
+    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: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
-        let min_value = *data.iter().min().unwrap();
-        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
+    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();
-            Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
+            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)
         });
     }
 

columnar/src/column_values/monotonic_mapping.rs

@@ -1,6 +1,7 @@
 use std::fmt::Debug;
 use std::marker::PhantomData;
 
+use common::DateTime;
 use fastdivide::DividerU64;
 
 use super::MonotonicallyMappableToU128;
@@ -122,6 +123,7 @@ pub(crate) struct StrictlyMonotonicMappingToInternalGCDBaseval {
     min_value: u64,
 }
 impl StrictlyMonotonicMappingToInternalGCDBaseval {
+    /// Creates a linear mapping `x -> gcd*x + min_value`.
     pub(crate) fn new(gcd: u64, min_value: u64) -> Self {
         let gcd_divider = DividerU64::divide_by(gcd);
         Self {
@@ -150,7 +152,9 @@ impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
 pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
     min_value: u64,
 }
+
 impl StrictlyMonotonicMappingToInternalBaseval {
+    /// Creates a linear mapping `x -> x + min_value`.
     #[inline(always)]
     pub(crate) fn new(min_value: u64) -> Self {
         Self { min_value }
@@ -195,17 +199,15 @@ impl MonotonicallyMappableToU64 for i64 {
     }
 }
 
-impl MonotonicallyMappableToU64 for crate::DateTime {
+impl MonotonicallyMappableToU64 for DateTime {
     #[inline(always)]
     fn to_u64(self) -> u64 {
-        common::i64_to_u64(self.timestamp_micros)
+        common::i64_to_u64(self.into_timestamp_micros())
     }
 
     #[inline(always)]
     fn from_u64(val: u64) -> Self {
-        crate::DateTime {
-            timestamp_micros: common::u64_to_i64(val),
-        }
+        DateTime::from_timestamp_micros(common::u64_to_i64(val))
     }
 }
 

columnar/src/column_values/serialize.rs

@@ -8,19 +8,6 @@ use crate::column_values::U128FastFieldCodecType;
 use crate::iterable::Iterable;
 use crate::MonotonicallyMappableToU128;
 
-/// The normalized header gives some parameters after applying the following
-/// normalization of the vector:
-/// `val -> (val - min_value) / gcd`
-///
-/// By design, after normalization, `min_value = 0` and `gcd = 1`.
-#[derive(Debug, Copy, Clone)]
-pub struct NormalizedHeader {
-    /// The number of values in the underlying column.
-    pub num_vals: u32,
-    /// The max value of the underlying column.
-    pub max_value: u64,
-}
-
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
 pub(crate) struct U128Header {
     pub num_vals: u32,
@@ -47,20 +34,18 @@ impl BinarySerializable for U128Header {
 /// Serializes u128 values with the compact space codec.
 pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
     iterable: &dyn Iterable<T>,
-    num_vals: u32,
     output: &mut impl io::Write,
 ) -> io::Result<()> {
-    let header = U128Header {
-        num_vals,
-        codec_type: U128FastFieldCodecType::CompactSpace,
-    };
-    header.serialize(output)?;
     let compressor = CompactSpaceCompressor::train_from(
         iterable
             .boxed_iter()
             .map(MonotonicallyMappableToU128::to_u128),
-        num_vals,
     );
+    let header = U128Header {
+        num_vals: compressor.num_vals(),
+        codec_type: U128FastFieldCodecType::CompactSpace,
+    };
+    header.serialize(output)?;
     compressor.compress_into(
         iterable
             .boxed_iter()
@@ -74,7 +59,7 @@ pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
 pub mod tests {
     use super::*;
     use crate::column_values::u64_based::{
-        self, serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
+        serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
         ALL_U64_CODEC_TYPES,
     };
     use crate::column_values::CodecType;

columnar/src/column_values/stats.rs

@@ -6,21 +6,28 @@ use common::{BinarySerializable, VInt};
 
 use crate::RowId;
 
+/// Column statistics.
 #[derive(Debug, Clone, Eq, PartialEq)]
-pub struct Stats {
+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 Stats {
+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 Stats {
+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)?;
@@ -37,7 +44,7 @@ impl BinarySerializable for Stats {
         let amplitude = VInt::deserialize(reader)?.0 * gcd.get();
         let max_value = min_value + amplitude;
         let num_rows = VInt::deserialize(reader)?.0 as RowId;
-        Ok(Stats {
+        Ok(ColumnStats {
             min_value,
             max_value,
             num_rows,
@@ -52,21 +59,21 @@ mod tests {
 
     use common::BinarySerializable;
 
-    use crate::column_values::Stats;
+    use crate::column_values::ColumnStats;
 
     #[track_caller]
-    fn test_stats_ser_deser_aux(stats: &Stats, num_bytes: usize) {
+    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 = Stats::deserialize(&mut &buffer[..]).unwrap();
+        let deser_stats = ColumnStats::deserialize(&mut &buffer[..]).unwrap();
         assert_eq!(stats, &deser_stats);
     }
 
     #[test]
     fn test_stats_serialization() {
         test_stats_ser_deser_aux(
-            &(Stats {
+            &(ColumnStats {
                 gcd: NonZeroU64::new(3).unwrap(),
                 min_value: 1,
                 max_value: 3001,
@@ -75,7 +82,7 @@ mod tests {
             5,
         );
         test_stats_ser_deser_aux(
-            &(Stats {
+            &(ColumnStats {
                 gcd: NonZeroU64::new(1_000).unwrap(),
                 min_value: 1,
                 max_value: 3001,
@@ -84,7 +91,7 @@ mod tests {
             5,
         );
         test_stats_ser_deser_aux(
-            &(Stats {
+            &(ColumnStats {
                 gcd: NonZeroU64::new(1).unwrap(),
                 min_value: 0,
                 max_value: 0,

columnar/src/column_values/u64_based/bitpacked.rs

@@ -4,7 +4,7 @@ use common::{BinarySerializable, OwnedBytes};
 use fastdivide::DividerU64;
 use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
 
-use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
 use crate::{ColumnValues, RowId};
 
 /// Depending on the field type, a different
@@ -13,7 +13,7 @@ use crate::{ColumnValues, RowId};
 pub struct BitpackedReader {
     data: OwnedBytes,
     bit_unpacker: BitUnpacker,
-    stats: Stats,
+    stats: ColumnStats,
 }
 
 impl ColumnValues for BitpackedReader {
@@ -36,7 +36,7 @@ impl ColumnValues for BitpackedReader {
     }
 }
 
-fn num_bits(stats: &Stats) -> u8 {
+fn num_bits(stats: &ColumnStats) -> u8 {
     compute_num_bits(stats.amplitude() / stats.gcd)
 }
 
@@ -46,14 +46,14 @@ pub struct BitpackedCodecEstimator;
 impl ColumnCodecEstimator for BitpackedCodecEstimator {
     fn collect(&mut self, _value: u64) {}
 
-    fn estimate(&self, stats: &Stats) -> Option<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: &Stats,
+        stats: &ColumnStats,
         vals: &mut dyn Iterator<Item = u64>,
         wrt: &mut dyn Write,
     ) -> io::Result<()> {
@@ -72,12 +72,12 @@ impl ColumnCodecEstimator for BitpackedCodecEstimator {
 pub struct BitpackedCodec;
 
 impl ColumnCodec for BitpackedCodec {
-    type Reader = BitpackedReader;
+    type ColumnValues = BitpackedReader;
     type Estimator = BitpackedCodecEstimator;
 
     /// Opens a fast field given a file.
-    fn load(mut data: OwnedBytes) -> io::Result<Self::Reader> {
-        let stats = Stats::deserialize(&mut data)?;
+    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 {

columnar/src/column_values/u64_based/blockwise_linear.rs

@@ -7,7 +7,7 @@ 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, Stats};
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
 use crate::column_values::{ColumnValues, VecColumn};
 use crate::MonotonicallyMappableToU64;
 
@@ -84,7 +84,7 @@ impl ColumnCodecEstimator for BlockwiseLinearEstimator {
             self.block.clear();
         }
     }
-    fn estimate(&self, stats: &Stats) -> Option<u64> {
+    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 =
@@ -100,7 +100,7 @@ impl ColumnCodecEstimator for BlockwiseLinearEstimator {
 
     fn serialize(
         &self,
-        stats: &Stats,
+        stats: &ColumnStats,
         mut vals: &mut dyn Iterator<Item = u64>,
         wrt: &mut dyn Write,
     ) -> io::Result<()> {
@@ -165,12 +165,12 @@ impl ColumnCodecEstimator for BlockwiseLinearEstimator {
 pub struct BlockwiseLinearCodec;
 
 impl ColumnCodec<u64> for BlockwiseLinearCodec {
-    type Reader = BlockwiseLinearReader;
+    type ColumnValues = BlockwiseLinearReader;
 
     type Estimator = BlockwiseLinearEstimator;
 
-    fn load(mut bytes: OwnedBytes) -> io::Result<Self::Reader> {
-        let stats = Stats::deserialize(&mut bytes)?;
+    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);
@@ -195,7 +195,7 @@ impl ColumnCodec<u64> for BlockwiseLinearCodec {
 pub struct BlockwiseLinearReader {
     blocks: Arc<[Block]>,
     data: OwnedBytes,
-    stats: Stats,
+    stats: ColumnStats,
 }
 
 impl ColumnValues for BlockwiseLinearReader {

columnar/src/column_values/u64_based/line.rs

@@ -67,19 +67,6 @@ 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
     pub fn train_from(
         first_val: u64,

columnar/src/column_values/u64_based/linear.rs

@@ -5,9 +5,9 @@ use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
 
 use super::line::Line;
 use super::ColumnValues;
-use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, ColumnStats};
 use crate::column_values::VecColumn;
-use crate::{MonotonicallyMappableToU64, RowId};
+use crate::RowId;
 
 const HALF_SPACE: u64 = u64::MAX / 2;
 const LINE_ESTIMATION_BLOCK_LEN: usize = 512;
@@ -18,7 +18,7 @@ const LINE_ESTIMATION_BLOCK_LEN: usize = 512;
 pub struct LinearReader {
     data: OwnedBytes,
     linear_params: LinearParams,
-    stats: Stats,
+    stats: ColumnStats,
 }
 
 impl ColumnValues for LinearReader {
@@ -106,7 +106,7 @@ impl ColumnCodecEstimator for LinearCodecEstimator {
         }
     }
 
-    fn estimate(&self, stats: &Stats) -> Option<u64> {
+    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);
@@ -123,7 +123,7 @@ impl ColumnCodecEstimator for LinearCodecEstimator {
 
     fn serialize(
         &self,
-        stats: &Stats,
+        stats: &ColumnStats,
         vals: &mut dyn Iterator<Item = u64>,
         wrt: &mut dyn io::Write,
     ) -> io::Result<()> {
@@ -184,12 +184,12 @@ impl LinearCodecEstimator {
 }
 
 impl ColumnCodec for LinearCodec {
-    type Reader = LinearReader;
+    type ColumnValues = LinearReader;
 
     type Estimator = LinearCodecEstimator;
 
-    fn load(mut data: OwnedBytes) -> io::Result<Self::Reader> {
-        let stats = Stats::deserialize(&mut data)?;
+    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,

columnar/src/column_values/u64_based/mod.rs

@@ -13,35 +13,61 @@ use common::{BinarySerializable, OwnedBytes};
 use crate::column_values::monotonic_mapping::{
     StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
 };
-use crate::column_values::u64_based::bitpacked::BitpackedCodec;
-use crate::column_values::u64_based::blockwise_linear::BlockwiseLinearCodec;
-use crate::column_values::u64_based::linear::LinearCodec;
-use crate::column_values::u64_based::stats_collector::StatsCollector;
-use crate::column_values::{monotonic_map_column, Stats};
+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);
-    fn estimate(&self, stats: &Stats) -> Option<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: &Stats,
+        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> {
-    type Reader: ColumnValues<T> + 'static;
+    /// Specialized `ColumnValues` type.
+    type ColumnValues: ColumnValues<T> + 'static;
+    /// `Estimator` for the given codec.
     type Estimator: ColumnCodecEstimator + Default;
 
-    fn load(bytes: OwnedBytes) -> io::Result<Self::Reader>;
+    /// 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())
     }
@@ -62,6 +88,7 @@ pub enum CodecType {
     BlockwiseLinear = 2u8,
 }
 
+/// List of all available u64-base codecs.
 pub const ALL_U64_CODEC_TYPES: [CodecType; 3] = [
     CodecType::Bitpacked,
     CodecType::Linear,
@@ -106,6 +133,7 @@ fn load_specific_codec<C: ColumnCodec, T: MonotonicallyMappableToU64>(
 }
 
 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(),
@@ -115,7 +143,8 @@ impl CodecType {
     }
 }
 
-pub fn serialize_u64_based_column_values<'a, T: MonotonicallyMappableToU64>(
+/// 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,
@@ -156,11 +185,14 @@ pub fn serialize_u64_based_column_values<'a, T: MonotonicallyMappableToU64>(
     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
-        .get(0)
+        .first()
         .copied()
         .and_then(CodecType::try_from_code)
         .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Failed to read codec type"))?;

columnar/src/column_values/u64_based/stats_collector.rs

@@ -2,7 +2,7 @@ use std::num::NonZeroU64;
 
 use fastdivide::DividerU64;
 
-use crate::column_values::Stats;
+use crate::column_values::ColumnStats;
 use crate::RowId;
 
 /// Compute the gcd of two non null numbers.
@@ -33,14 +33,14 @@ pub struct StatsCollector {
 }
 
 impl StatsCollector {
-    pub fn stats(&self) -> Stats {
+    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()
         };
-        Stats {
+        ColumnStats {
             min_value,
             max_value,
             num_rows: self.num_rows,
@@ -97,9 +97,9 @@ mod tests {
     use std::num::NonZeroU64;
 
     use crate::column_values::u64_based::stats_collector::{compute_gcd, StatsCollector};
-    use crate::column_values::u64_based::Stats;
+    use crate::column_values::u64_based::ColumnStats;
 
-    fn compute_stats(vals: impl Iterator<Item = u64>) -> Stats {
+    fn compute_stats(vals: impl Iterator<Item = u64>) -> ColumnStats {
         let mut stats_collector = StatsCollector::default();
         for val in vals {
             stats_collector.collect(val);
@@ -144,7 +144,7 @@ mod tests {
     fn test_stats() {
         assert_eq!(
             compute_stats([].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(1).unwrap(),
                 min_value: 0,
                 max_value: 0,
@@ -153,7 +153,7 @@ mod tests {
         );
         assert_eq!(
             compute_stats([0, 1].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(1).unwrap(),
                 min_value: 0,
                 max_value: 1,
@@ -162,7 +162,7 @@ mod tests {
         );
         assert_eq!(
             compute_stats([0, 1].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(1).unwrap(),
                 min_value: 0,
                 max_value: 1,
@@ -171,7 +171,7 @@ mod tests {
         );
         assert_eq!(
             compute_stats([10, 20, 30].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(10).unwrap(),
                 min_value: 10,
                 max_value: 30,
@@ -180,7 +180,7 @@ mod tests {
         );
         assert_eq!(
             compute_stats([10, 50, 10, 30].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(20).unwrap(),
                 min_value: 10,
                 max_value: 50,
@@ -189,7 +189,7 @@ mod tests {
         );
         assert_eq!(
             compute_stats([10, 0, 30].into_iter()),
-            Stats {
+            ColumnStats {
                 gcd: NonZeroU64::new(10).unwrap(),
                 min_value: 0,
                 max_value: 30,

columnar/src/column_values/u64_based/tests.rs

@@ -60,7 +60,7 @@ pub(crate) fn create_and_validate<TColumnCodec: ColumnCodec>(
             .map(|(pos, _)| pos as u32)
             .collect();
         let mut positions = Vec::new();
-        reader.get_docids_for_value_range(
+        reader.get_row_ids_for_value_range(
             vals[test_rand_idx]..=vals[test_rand_idx],
             0..vals.len() as u32,
             &mut positions,

columnar/src/columnar/column_type.rs

@@ -4,24 +4,22 @@ use std::net::Ipv6Addr;
 use crate::value::NumericalType;
 use crate::InvalidData;
 
-/// The column type represents the column type and can fit on 6-bits.
-///
-/// - bits[0..3]: Column category type.
-/// - bits[3..6]: Numerical type if necessary.
+/// 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)]
 #[repr(u8)]
 pub enum ColumnType {
     I64 = 0u8,
     U64 = 1u8,
     F64 = 2u8,
-    Bytes = 10u8,
-    Str = 14u8,
-    Bool = 18u8,
-    IpAddr = 22u8,
-    DateTime = 26u8,
+    Bytes = 3u8,
+    Str = 4u8,
+    Bool = 5u8,
+    IpAddr = 6u8,
+    DateTime = 7u8,
 }
 
-#[cfg(test)]
+// The order needs to match _exactly_ the order in the enum
 const COLUMN_TYPES: [ColumnType; 8] = [
     ColumnType::I64,
     ColumnType::U64,
@@ -39,18 +37,7 @@ impl ColumnType {
     }
 
     pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
-        use ColumnType::*;
-        match code {
-            0u8 => Ok(I64),
-            1u8 => Ok(U64),
-            2u8 => Ok(F64),
-            10u8 => Ok(Bytes),
-            14u8 => Ok(Str),
-            18u8 => Ok(Bool),
-            22u8 => Ok(IpAddr),
-            26u8 => Ok(Self::DateTime),
-            _ => Err(InvalidData),
-        }
+        COLUMN_TYPES.get(code as usize).copied().ok_or(InvalidData)
     }
 }
 
@@ -124,7 +111,7 @@ impl HasAssociatedColumnType for bool {
     }
 }
 
-impl HasAssociatedColumnType for crate::DateTime {
+impl HasAssociatedColumnType for common::DateTime {
     fn column_type() -> ColumnType {
         ColumnType::DateTime
     }
@@ -143,70 +130,20 @@ impl HasAssociatedColumnType for Ipv6Addr {
     }
 }
 
-/// Column types are grouped into different categories that
-/// corresponds to the different types of `JsonValue` types.
-///
-/// The columnar writer will apply coercion rules to make sure that
-/// at most one column exist per `ColumnTypeCategory`.
-///
-/// See also [README.md].
-#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq, Hash, Debug)]
-#[repr(u8)]
-pub 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,
-        }
-    }
-}
-
 #[cfg(test)]
 mod tests {
-    use std::collections::HashSet;
-
     use super::*;
     use crate::Cardinality;
 
     #[test]
     fn test_column_type_to_code() {
-        let mut column_type_set: HashSet<ColumnType> = HashSet::new();
-        for code in u8::MIN..=u8::MAX {
-            if let Ok(column_type) = ColumnType::try_from_code(code) {
-                assert_eq!(column_type.to_code(), code);
-                assert!(column_type_set.insert(column_type));
+        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);
             }
         }
-        assert_eq!(column_type_set.len(), super::COLUMN_TYPES.len());
-    }
-
-    #[test]
-    fn test_column_category_sort_consistent_with_column_type_sort() {
-        // This is a very important property because we
-        // we need to serialize colunmn in the right order.
-        let mut column_types: Vec<ColumnType> = super::COLUMN_TYPES.iter().copied().collect();
-        column_types.sort_by_key(|col| col.to_code());
-        let column_categories: Vec<ColumnTypeCategory> = column_types
-            .into_iter()
-            .map(ColumnTypeCategory::from)
-            .collect();
-        for (prev, next) in column_categories.iter().zip(column_categories.iter()) {
-            assert!(prev <= next);
+        for code in COLUMN_TYPES.len() as u8..=u8::MAX {
+            assert!(ColumnType::try_from_code(code as u8).is_err());
         }
     }
 

columnar/src/columnar/format_version.rs

@@ -4,7 +4,7 @@ pub const VERSION_FOOTER_NUM_BYTES: usize = MAGIC_BYTES.len() + std::mem::size_o
 
 /// We end the file by these 4 bytes just to somewhat identify that
 /// this is indeed a columnar file.
-const MAGIC_BYTES: [u8; 4] = [2, 113, 119, 066];
+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];
@@ -27,8 +27,8 @@ pub enum Version {
 }
 
 impl Version {
-    fn to_bytes(&self) -> [u8; 4] {
-        (*self as u32).to_le_bytes()
+    fn to_bytes(self) -> [u8; 4] {
+        (self as u32).to_le_bytes()
     }
 
     fn try_from_bytes(bytes: [u8; 4]) -> Result<Version, InvalidData> {

columnar/src/columnar/merge/merge_dict_column.rs

@@ -1,68 +1,130 @@
 use std::io::{self, Write};
 
-use common::CountingWriter;
+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;
+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: &[BytesColumn],
+    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);
-    let term_ord_mapping = serialize_merged_dict(bytes_columns, &mut 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)?;
-    // serialize_bytes_or_str_column(column_index, bytes_columns, &term_ord_mapping, output)?;
     output.write_all(&dictionary_num_bytes.to_le_bytes())?;
     Ok(())
 }
 
 struct RemappedTermOrdinalsValues<'a> {
-    bytes_columns: &'a [BytesColumn],
+    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);
-                byte_column
-                    .ords()
-                    .values
-                    .iter()
-                    .map(move |term_ord| segment_ord[term_ord as usize])
+                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(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: &[BytesColumn],
+    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 {
+    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);
@@ -71,21 +133,57 @@ fn serialize_merged_dict(
     let mut merged_terms = TermMerger::new(field_term_streams);
     let mut sstable_builder = sstable::VoidSSTable::writer(output);
 
-    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);
+    // 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()?;
         }
-        current_term_ord += 1;
     }
-    sstable_builder.finish()?;
     Ok(term_ord_mapping)
 }
 
-#[derive(Default)]
+#[derive(Default, Debug)]
 struct TermOrdinalMapping {
     per_segment_new_term_ordinals: Vec<Vec<TermOrdinal>>,
 }
@@ -100,7 +198,7 @@ impl TermOrdinalMapping {
         self.per_segment_new_term_ordinals[segment_ord][from_ord as usize] = to_ord;
     }
 
-    fn get_segment(&self, segment_ord: usize) -> &[TermOrdinal] {
-        &(self.per_segment_new_term_ordinals[segment_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

@@ -1,6 +1,8 @@
 use std::ops::Range;
 
-use crate::{column, ColumnarReader, RowId};
+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
@@ -42,19 +44,75 @@ pub enum MergeRowOrder {
     /// 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 can interleaves rows from the different readers and
-    /// possibly drop rows.
-    Complex(()),
+    /// 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::Complex(_) => {
-                todo!()
-            }
+            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

@@ -9,24 +9,54 @@ use std::io;
 use std::net::Ipv6Addr;
 use std::sync::Arc;
 
-pub use merge_mapping::{MergeRowOrder, StackMergeOrder};
+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_index::stack_column_index;
-use crate::columnar::column_type::ColumnTypeCategory;
+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, MonotonicallyMappableToU128,
-    NumericalType, NumericalValue,
+    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,
+        }
+    }
+}
+
 pub fn merge_columnar(
     columnar_readers: &[&ColumnarReader],
-    mapping: MergeRowOrder,
+    merge_row_order: MergeRowOrder,
     output: &mut impl io::Write,
 ) -> io::Result<()> {
     let mut serializer = ColumnarSerializer::new(output);
@@ -35,9 +65,14 @@ pub fn merge_columnar(
     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, &mapping, &mut column_serializer)?;
+        merge_column(
+            column_type,
+            columns,
+            &merge_row_order,
+            &mut column_serializer,
+        )?;
     }
-    serializer.finalize(mapping.num_rows())?;
+    serializer.finalize(merge_row_order.num_rows())?;
 
     Ok(())
 }
@@ -53,7 +88,7 @@ fn dynamic_column_to_u64_monotonic(dynamic_column: DynamicColumn) -> Option<Colu
     }
 }
 
-pub fn merge_column(
+fn merge_column(
     column_type: ColumnType,
     columns: Vec<Option<DynamicColumn>>,
     merge_row_order: &MergeRowOrder,
@@ -66,68 +101,80 @@ pub fn merge_column(
         | ColumnType::DateTime
         | ColumnType::Bool => {
             let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
-            let mut column_values: Vec<Arc<dyn ColumnValues>> = 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(values);
+                    column_values.push(Some(values));
                 } else {
                     column_indexes.push(None);
+                    column_values.push(None);
                 }
             }
             let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            serialize_column_mappable_to_u64(merged_column_index, &&column_values[..], wrt)?;
+                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<Arc<dyn ColumnValues<Ipv6Addr>>> =
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues<Ipv6Addr>>>> =
                 Vec::with_capacity(columns.len());
-            let mut num_values = 0;
             for dynamic_column_opt in columns {
                 if let Some(DynamicColumn::IpAddr(Column { idx, values })) = dynamic_column_opt {
-                    num_values += values.num_vals();
                     column_indexes.push(Some(idx));
-                    column_values.push(values);
+                    column_values.push(Some(values));
                 } else {
                     column_indexes.push(None);
+                    column_values.push(None);
                 }
             }
+
             let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            serialize_column_mappable_to_u128(
-                merged_column_index,
-                &&column_values[..],
-                num_values,
-                wrt,
-            )?;
+                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<BytesColumn> = 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(str_column.into());
+                        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(bytes_column);
+                        bytes_columns.push(Some(bytes_column));
+                    }
+                    _ => {
+                        column_indexes.push(None);
+                        bytes_columns.push(None);
                     }
-                    _ => column_indexes.push(None),
                 }
             }
             let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            merge_bytes_or_str_column(merged_column_index, &bytes_columns, wrt)?;
+                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(())
 }
 
+#[allow(clippy::type_complexity)]
 fn group_columns_for_merge(
     columnar_readers: &[&ColumnarReader],
 ) -> io::Result<BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>>> {

columnar/src/columnar/merge/sorted_doc_id_multivalue_column.rs

@@ -1,169 +0,0 @@
-use std::cmp;
-
-use fastfield_codecs::Column;
-
-use super::flat_map_with_buffer::FlatMapWithBufferIter;
-use crate::fastfield::{MultiValueIndex, MultiValuedFastFieldReader};
-use crate::indexer::doc_id_mapping::SegmentDocIdMapping;
-use crate::{DocAddress, SegmentReader};
-
-pub(crate) struct RemappedDocIdMultiValueColumn<'a> {
-    doc_id_mapping: &'a SegmentDocIdMapping,
-    fast_field_readers: Vec<MultiValuedFastFieldReader<u64>>,
-    min_value: u64,
-    max_value: u64,
-    num_vals: u32,
-}
-
-impl<'a> RemappedDocIdMultiValueColumn<'a> {
-    pub(crate) fn new(
-        readers: &'a [SegmentReader],
-        doc_id_mapping: &'a SegmentDocIdMapping,
-        field: &str,
-    ) -> Self {
-        // Our values are bitpacked and we need to know what should be
-        // our bitwidth and our minimum value before serializing any values.
-        //
-        // Computing those is non-trivial if some documents are deleted.
-        // We go through a complete first pass to compute the minimum and the
-        // maximum value and initialize our Serializer.
-        let mut num_vals = 0;
-        let mut min_value = u64::MAX;
-        let mut max_value = u64::MIN;
-        let mut vals = Vec::new();
-        let mut fast_field_readers = Vec::with_capacity(readers.len());
-        for reader in readers {
-            let ff_reader: MultiValuedFastFieldReader<u64> = reader
-                .fast_fields()
-                .typed_fast_field_multi_reader::<u64>(field)
-                .expect(
-                    "Failed to find multivalued fast field reader. This is a bug in tantivy. \
-                     Please report.",
-                );
-            for doc in reader.doc_ids_alive() {
-                ff_reader.get_vals(doc, &mut vals);
-                for &val in &vals {
-                    min_value = cmp::min(val, min_value);
-                    max_value = cmp::max(val, max_value);
-                }
-                num_vals += vals.len();
-            }
-            fast_field_readers.push(ff_reader);
-            // TODO optimize when no deletes
-        }
-        if min_value > max_value {
-            min_value = 0;
-            max_value = 0;
-        }
-        RemappedDocIdMultiValueColumn {
-            doc_id_mapping,
-            fast_field_readers,
-            min_value,
-            max_value,
-            num_vals: num_vals as u32,
-        }
-    }
-}
-
-impl<'a> Column for RemappedDocIdMultiValueColumn<'a> {
-    fn get_val(&self, _pos: u32) -> u64 {
-        unimplemented!()
-    }
-
-    fn iter(&self) -> Box<dyn Iterator<Item = u64> + '_> {
-        Box::new(
-            self.doc_id_mapping
-                .iter_old_doc_addrs()
-                .flat_map_with_buffer(|old_doc_addr: DocAddress, buffer| {
-                    let ff_reader = &self.fast_field_readers[old_doc_addr.segment_ord as usize];
-                    ff_reader.get_vals(old_doc_addr.doc_id, buffer);
-                }),
-        )
-    }
-    fn min_value(&self) -> u64 {
-        self.min_value
-    }
-
-    fn max_value(&self) -> u64 {
-        self.max_value
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.num_vals
-    }
-}
-
-pub(crate) struct RemappedDocIdMultiValueIndexColumn<'a> {
-    doc_id_mapping: &'a SegmentDocIdMapping,
-    multi_value_length_readers: Vec<&'a MultiValueIndex>,
-    min_value: u64,
-    max_value: u64,
-    num_vals: u32,
-}
-
-impl<'a> RemappedDocIdMultiValueIndexColumn<'a> {
-    pub(crate) fn new(
-        segment_and_ff_readers: &'a [(&'a SegmentReader, &'a MultiValueIndex)],
-        doc_id_mapping: &'a SegmentDocIdMapping,
-    ) -> Self {
-        // We go through a complete first pass to compute the minimum and the
-        // maximum value and initialize our Column.
-        let mut num_vals = 0;
-        let min_value = 0;
-        let mut max_value = 0;
-        let mut multi_value_length_readers = Vec::with_capacity(segment_and_ff_readers.len());
-        for segment_and_ff_reader in segment_and_ff_readers {
-            let segment_reader = segment_and_ff_reader.0;
-            let multi_value_length_reader = segment_and_ff_reader.1;
-            if !segment_reader.has_deletes() {
-                max_value += multi_value_length_reader.total_num_vals() as u64;
-            } else {
-                for doc in segment_reader.doc_ids_alive() {
-                    max_value += multi_value_length_reader.num_vals_for_doc(doc) as u64;
-                }
-            }
-            num_vals += segment_reader.num_docs();
-            multi_value_length_readers.push(multi_value_length_reader);
-        }
-        // The value range is always get_val(doc)..get_val(doc + 1)
-        num_vals += 1;
-        Self {
-            doc_id_mapping,
-            multi_value_length_readers,
-            min_value,
-            max_value,
-            num_vals,
-        }
-    }
-}
-
-impl<'a> Column for RemappedDocIdMultiValueIndexColumn<'a> {
-    fn get_val(&self, _pos: u32) -> u64 {
-        unimplemented!()
-    }
-
-    fn iter(&self) -> Box<dyn Iterator<Item = u64> + '_> {
-        let mut offset = 0;
-        Box::new(
-            std::iter::once(0).chain(self.doc_id_mapping.iter_old_doc_addrs().map(
-                move |old_doc_addr| {
-                    let ff_reader =
-                        &self.multi_value_length_readers[old_doc_addr.segment_ord as usize];
-                    offset += ff_reader.num_vals_for_doc(old_doc_addr.doc_id);
-                    offset as u64
-                },
-            )),
-        )
-    }
-    fn min_value(&self) -> u64 {
-        self.min_value
-    }
-
-    fn max_value(&self) -> u64 {
-        self.max_value
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.num_vals
-    }
-}

columnar/src/columnar/merge/tests.rs

@@ -12,7 +12,7 @@ fn make_columnar<T: Into<NumericalValue> + HasAssociatedColumnType + Copy>(
     }
     let mut buffer: Vec<u8> = Vec::new();
     dataframe_writer
-        .serialize(vals.len() as RowId, &mut buffer)
+        .serialize(vals.len() as RowId, None, &mut buffer)
         .unwrap();
     ColumnarReader::open(buffer).unwrap()
 }
@@ -90,7 +90,9 @@ fn make_numerical_columnar_multiple_columns(
         .max()
         .unwrap_or(0u32);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
     ColumnarReader::open(buffer).unwrap()
 }
 
@@ -109,7 +111,9 @@ fn make_byte_columnar_multiple_columns(columns: &[(&str, &[&[&[u8]]])]) -> Colum
         .max()
         .unwrap_or(0u32);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
     ColumnarReader::open(buffer).unwrap()
 }
 
@@ -128,7 +132,9 @@ fn make_text_columnar_multiple_columns(columns: &[(&str, &[&[&str]])]) -> Column
         .max()
         .unwrap_or(0u32);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(num_rows, &mut buffer).unwrap();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
     ColumnarReader::open(buffer).unwrap()
 }
 

columnar/src/columnar/mod.rs

@@ -6,6 +6,6 @@ mod reader;
 mod writer;
 
 pub use column_type::{ColumnType, HasAssociatedColumnType};
-pub use merge::{merge_columnar, MergeRowOrder, StackMergeOrder};
+pub use merge::{merge_columnar, MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
 pub use reader::ColumnarReader;
 pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -137,7 +137,7 @@ mod tests {
         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, &mut buffer).unwrap();
+        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);
@@ -153,7 +153,7 @@ mod tests {
         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, &mut buffer).unwrap();
+        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);
@@ -162,7 +162,7 @@ mod tests {
     }
 
     #[test]
-    #[should_panic(expect = "Input type forbidden")]
+    #[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/src/columnar/writer/column_writers.rs

@@ -41,10 +41,31 @@ impl ColumnWriter {
     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))
     }
@@ -189,10 +210,12 @@ impl CompatibleNumericalTypes {
 }
 
 impl NumericalColumnWriter {
-    pub fn column_type_and_cardinality(&self, num_docs: RowId) -> (NumericalType, Cardinality) {
-        let numerical_type = self.compatible_numerical_types.to_numerical_type();
-        let cardinality = self.column_writer.get_cardinality(num_docs);
-        (numerical_type, cardinality)
+    pub fn 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(
@@ -208,9 +231,11 @@ impl NumericalColumnWriter {
     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, buffer)
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, buffer)
     }
 }
 
@@ -251,9 +276,11 @@ impl StrOrBytesColumnWriter {
     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, byte_buffer)
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, byte_buffer)
     }
 }
 

columnar/src/columnar/writer/mod.rs

@@ -16,7 +16,7 @@ use crate::column_index::SerializableColumnIndex;
 use crate::column_values::{
     ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64, VecColumn,
 };
-use crate::columnar::column_type::{ColumnType, ColumnTypeCategory};
+use crate::columnar::column_type::ColumnType;
 use crate::columnar::writer::column_writers::{
     ColumnWriter, NumericalColumnWriter, StrOrBytesColumnWriter,
 };
@@ -45,8 +45,9 @@ struct SpareBuffers {
 /// columnar_writer.record_str(1u32 /* doc id */, "product_name", "Apple");
 /// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10.5f64); //< uh oh we ended up mixing integer and floats.
 /// let mut wrt: Vec<u8> =  Vec::new();
-/// columnar_writer.serialize(2u32, &mut wrt).unwrap();
+/// columnar_writer.serialize(2u32, None, &mut wrt).unwrap();
 /// ```
+#[derive(Default)]
 pub struct ColumnarWriter {
     numerical_field_hash_map: ArenaHashMap,
     datetime_field_hash_map: ArenaHashMap,
@@ -60,22 +61,6 @@ pub struct ColumnarWriter {
     buffers: SpareBuffers,
 }
 
-impl Default for ColumnarWriter {
-    fn default() -> Self {
-        ColumnarWriter {
-            numerical_field_hash_map: ArenaHashMap::new(10_000),
-            bool_field_hash_map: ArenaHashMap::new(10_000),
-            ip_addr_field_hash_map: ArenaHashMap::new(10_000),
-            bytes_field_hash_map: ArenaHashMap::new(10_000),
-            str_field_hash_map: ArenaHashMap::new(10_000),
-            datetime_field_hash_map: ArenaHashMap::new(10_000),
-            dictionaries: Vec::new(),
-            arena: MemoryArena::default(),
-            buffers: SpareBuffers::default(),
-        }
-    }
-}
-
 #[inline]
 fn mutate_or_create_column<V, TMutator>(
     arena_hash_map: &mut ArenaHashMap,
@@ -104,6 +89,48 @@ impl ColumnarWriter {
             + 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`.
@@ -224,11 +251,15 @@ impl ColumnarWriter {
         });
     }
 
-    pub fn record_datetime(&mut self, doc: RowId, column_name: &str, datetime: crate::DateTime) {
+    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.timestamp_micros), arena);
+            column.record(
+                doc,
+                NumericalValue::I64(datetime.into_timestamp_micros()),
+                arena,
+            );
             column
         });
     }
@@ -278,37 +309,47 @@ impl ColumnarWriter {
             },
         );
     }
-    pub fn serialize(&mut self, num_docs: RowId, wrt: &mut dyn io::Write) -> io::Result<()> {
+    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], ColumnTypeCategory, Addr)> = self
+        let mut columns: Vec<(&[u8], ColumnType, Addr)> = self
             .numerical_field_hash_map
             .iter()
-            .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Numerical, addr))
+            .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, ColumnTypeCategory::Bytes, addr)),
+                .map(|(term, addr, _)| (term, ColumnType::Bytes, addr)),
         );
         columns.extend(
             self.str_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Str, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Str, addr)),
         );
         columns.extend(
             self.bool_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Bool, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Bool, addr)),
         );
         columns.extend(
             self.ip_addr_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::IpAddr, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::IpAddr, addr)),
         );
         columns.extend(
             self.datetime_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::DateTime, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::DateTime, addr)),
         );
         columns.sort_unstable_by_key(|(column_name, col_type, _)| (*column_name, *col_type));
 
@@ -316,20 +357,24 @@ impl ColumnarWriter {
         let mut symbol_byte_buffer: Vec<u8> = Vec::new();
         for (column_name, column_type, addr) in columns {
             match column_type {
-                ColumnTypeCategory::Bool => {
+                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, ColumnType::Bool);
+                        serializer.serialize_column(column_name, column_type);
                     serialize_bool_column(
                         cardinality,
                         num_docs,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::IpAddr => {
+                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 =
@@ -337,50 +382,64 @@ impl ColumnarWriter {
                     serialize_ip_addr_column(
                         cardinality,
                         num_docs,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::Bytes | ColumnTypeCategory::Str => {
-                    let (column_type, str_column_writer): (ColumnType, StrOrBytesColumnWriter) =
-                        if column_type == ColumnTypeCategory::Bytes {
-                            (ColumnType::Bytes, self.bytes_field_hash_map.read(addr))
+                ColumnType::Bytes | ColumnType::Str => {
+                    let str_or_bytes_column_writer: StrOrBytesColumnWriter =
+                        if column_type == ColumnType::Bytes {
+                            self.bytes_field_hash_map.read(addr)
                         } else {
-                            (ColumnType::Str, self.str_field_hash_map.read(addr))
+                            self.str_field_hash_map.read(addr)
                         };
                     let dictionary_builder =
-                        &dictionaries[str_column_writer.dictionary_id as usize];
-                    let cardinality = str_column_writer.column_writer.get_cardinality(num_docs);
+                        &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_column_writer.sort_values_within_row,
+                        str_or_bytes_column_writer.sort_values_within_row,
                         dictionary_builder,
-                        str_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        str_or_bytes_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::Numerical => {
+                ColumnType::F64 | ColumnType::I64 | ColumnType::U64 => {
                     let numerical_column_writer: NumericalColumnWriter =
                         self.numerical_field_hash_map.read(addr);
-                    let (numerical_type, cardinality) =
-                        numerical_column_writer.column_type_and_cardinality(num_docs);
+                    let cardinality = numerical_column_writer.cardinality(num_docs);
                     let mut column_serializer =
-                        serializer.serialize_column(column_name, ColumnType::from(numerical_type));
+                        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, &mut symbol_byte_buffer),
+                        numerical_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::DateTime => {
+                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 =
@@ -389,7 +448,11 @@ impl ColumnarWriter {
                         cardinality,
                         num_docs,
                         NumericalType::I64,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
@@ -588,13 +651,12 @@ where
     crate::column::serialize_column_mappable_to_u128(
         serializable_column_index,
         &&values[..],
-        values.len() as u32,
         &mut wrt,
     )?;
     Ok(())
 }
 
-fn sort_values_within_row_in_place(multivalued_index: &[RowId], values: &mut Vec<u64>) {
+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;
@@ -699,7 +761,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(3), Cardinality::Full);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 6);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
@@ -728,7 +790,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(3), Cardinality::Optional);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 4);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(1u32)));
@@ -751,7 +813,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(2), Cardinality::Optional);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 2);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
@@ -770,7 +832,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(1), Cardinality::Multivalued);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 3);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));

columnar/src/columnar/writer/value_index.rs

@@ -29,7 +29,7 @@ pub struct OptionalIndexBuilder {
 }
 
 impl OptionalIndexBuilder {
-    pub fn finish<'a>(&'a mut self, num_rows: RowId) -> impl Iterable<RowId> + 'a {
+    pub fn finish(&mut self, num_rows: RowId) -> impl Iterable<RowId> + '_ {
         debug_assert!(self
             .docs
             .last()

columnar/src/dynamic_column.rs

@@ -3,12 +3,12 @@ use std::net::Ipv6Addr;
 use std::sync::Arc;
 
 use common::file_slice::FileSlice;
-use common::{HasLen, OwnedBytes};
+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, DateTime, NumericalType};
+use crate::{Cardinality, NumericalType};
 
 #[derive(Clone)]
 pub enum DynamicColumn {
@@ -166,9 +166,9 @@ impl StrictlyMonotonicFn<i64, u64> for MapI64ToU64 {
 
 macro_rules! static_dynamic_conversions {
     ($typ:ty, $enum_name:ident) => {
-        impl Into<Option<$typ>> for DynamicColumn {
-            fn into(self) -> Option<$typ> {
-                if let DynamicColumn::$enum_name(col) = self {
+        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
@@ -188,7 +188,7 @@ 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<crate::DateTime>, DateTime);
+static_dynamic_conversions!(Column<DateTime>, DateTime);
 static_dynamic_conversions!(StrColumn, Str);
 static_dynamic_conversions!(BytesColumn, Bytes);
 static_dynamic_conversions!(Column<Ipv6Addr>, IpAddr);
@@ -235,17 +235,15 @@ impl DynamicColumnHandle {
 
     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::<BytesColumn>(column_bytes)?.into()
-            }
-            ColumnType::Str => crate::column::open_column_bytes::<StrColumn>(column_bytes)?.into(),
+            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::<crate::DateTime>(column_bytes)?.into()
+                crate::column::open_column_u64::<DateTime>(column_bytes)?.into()
             }
         };
         Ok(dynamic_column)

columnar/src/lib.rs

@@ -24,7 +24,7 @@ pub use column_index::ColumnIndex;
 pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
 pub use columnar::{
     merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
-    MergeRowOrder, StackMergeOrder,
+    MergeRowOrder, ShuffleMergeOrder, StackMergeOrder,
 };
 use sstable::VoidSSTable;
 pub use value::{NumericalType, NumericalValue};
@@ -32,13 +32,18 @@ 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>;
 
-#[derive(Clone, Copy, PartialOrd, PartialEq, Default, Debug)]
-pub struct DateTime {
-    pub timestamp_micros: i64,
-}
+pub use common::DateTime;
 
 #[derive(Copy, Clone, Debug)]
 pub struct InvalidData;

columnar/src/tests.rs

@@ -12,7 +12,7 @@ fn test_dataframe_writer_str() {
     dataframe_writer.record_str(1u32, "my_string", "hello");
     dataframe_writer.record_str(3u32, "my_string", "helloeee");
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    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();
@@ -26,7 +26,7 @@ fn test_dataframe_writer_bytes() {
     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, &mut buffer).unwrap();
+    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();
@@ -40,7 +40,7 @@ fn test_dataframe_writer_bool() {
     dataframe_writer.record_bool(1u32, "bool.value", false);
     dataframe_writer.record_bool(3u32, "bool.value", true);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    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();
@@ -63,7 +63,7 @@ fn test_dataframe_writer_u64_multivalued() {
     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, &mut buffer).unwrap();
+    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();
@@ -75,7 +75,7 @@ fn test_dataframe_writer_u64_multivalued() {
         divisor_col.get_cardinality(),
         crate::Cardinality::Multivalued
     );
-    assert_eq!(divisor_col.num_rows(), 7);
+    assert_eq!(divisor_col.num_docs(), 7);
 }
 
 #[test]
@@ -84,7 +84,7 @@ fn test_dataframe_writer_ip_addr() {
     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, &mut buffer).unwrap();
+    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();
@@ -113,7 +113,7 @@ fn test_dataframe_writer_numerical() {
     dataframe_writer.record_numerical(2u32, "srical.value", NumericalValue::U64(13u64));
     dataframe_writer.record_numerical(4u32, "srical.value", NumericalValue::U64(15u64));
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(6, &mut buffer).unwrap();
+    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();
@@ -144,7 +144,7 @@ fn test_dictionary_encoded_str() {
     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, &mut buffer).unwrap();
+    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();
@@ -176,7 +176,7 @@ fn test_dictionary_encoded_bytes() {
     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, &mut buffer).unwrap();
+    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();

columnar/src/value.rs

@@ -1,3 +1,5 @@
+use common::DateTime;
+
 use crate::InvalidData;
 
 #[derive(Copy, Clone, PartialEq, Debug)]
@@ -104,10 +106,10 @@ impl Coerce for f64 {
     }
 }
 
-impl Coerce for crate::DateTime {
+impl Coerce for DateTime {
     fn coerce(value: NumericalValue) -> Self {
         let timestamp_micros = i64::coerce(value);
-        crate::DateTime { timestamp_micros }
+        DateTime::from_timestamp_micros(timestamp_micros)
     }
 }
 

common/Cargo.toml

@@ -16,6 +16,8 @@ repository = "https://github.com/quickwit-oss/tantivy"
 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)]
+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/lib.rs

@@ -5,12 +5,14 @@ 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};

examples/aggregation.rs

@@ -24,8 +24,7 @@ fn main() -> tantivy::Result<()> {
         )
         .set_stored();
     let text_field = schema_builder.add_text_field("text", text_fieldtype);
-    let score_fieldtype =
-        crate::schema::NumericOptions::default().set_fast();
+    let score_fieldtype = crate::schema::NumericOptions::default().set_fast();
     let highscore_field = schema_builder.add_f64_field("highscore", score_fieldtype.clone());
     let price_field = schema_builder.add_f64_field("price", score_fieldtype);
 

examples/custom_collector.rs

@@ -7,9 +7,7 @@
 // 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};
@@ -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(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 {

examples/faceted_search.rs

@@ -71,7 +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".
@@ -97,7 +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

@@ -56,7 +56,7 @@ fn main() -> tantivy::Result<()> {
         );
         let top_docs_by_custom_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 +64,9 @@ 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()
+                    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/warmer.rs

@@ -48,7 +48,10 @@ 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))

src/aggregation/agg_req_with_accessor.rs

@@ -2,9 +2,8 @@
 
 use std::rc::Rc;
 use std::sync::atomic::AtomicU32;
-use std::sync::Arc;
 
-use fastfield_codecs::Column;
+use columnar::{Column, StrColumn};
 
 use super::agg_req::{Aggregation, Aggregations, BucketAggregationType, MetricAggregation};
 use super::bucket::{HistogramAggregation, RangeAggregation, TermsAggregation};
@@ -14,9 +13,8 @@ use super::metric::{
 };
 use super::segment_agg_result::BucketCount;
 use super::VecWithNames;
-use crate::fastfield::{type_and_cardinality, MultiValuedFastFieldReader};
 use crate::schema::Type;
-use crate::{InvertedIndexReader, SegmentReader, TantivyError};
+use crate::{SegmentReader, TantivyError};
 
 #[derive(Clone, Default)]
 pub(crate) struct AggregationsWithAccessor {
@@ -37,38 +35,12 @@ impl AggregationsWithAccessor {
     }
 }
 
-#[derive(Clone)]
-pub(crate) enum FastFieldAccessor {
-    Multi(MultiValuedFastFieldReader<u64>),
-    Single(Arc<dyn Column<u64>>),
-}
-impl FastFieldAccessor {
-    pub fn as_single(&self) -> Option<&dyn Column<u64>> {
-        match self {
-            FastFieldAccessor::Multi(_) => None,
-            FastFieldAccessor::Single(reader) => Some(&**reader),
-        }
-    }
-    pub fn into_single(self) -> Option<Arc<dyn Column<u64>>> {
-        match self {
-            FastFieldAccessor::Multi(_) => None,
-            FastFieldAccessor::Single(reader) => Some(reader),
-        }
-    }
-    pub fn as_multi(&self) -> Option<&MultiValuedFastFieldReader<u64>> {
-        match self {
-            FastFieldAccessor::Multi(reader) => Some(reader),
-            FastFieldAccessor::Single(_) => None,
-        }
-    }
-}
-
 #[derive(Clone)]
 pub struct BucketAggregationWithAccessor {
     /// In general there can be buckets without fast field access, e.g. buckets that are created
     /// based on search terms. So eventually this needs to be Option or moved.
-    pub(crate) accessor: FastFieldAccessor,
-    pub(crate) inverted_index: Option<Arc<InvertedIndexReader>>,
+    pub(crate) accessor: Column<u64>,
+    pub(crate) str_dict_column: Option<StrColumn>,
     pub(crate) field_type: Type,
     pub(crate) bucket_agg: BucketAggregationType,
     pub(crate) sub_aggregation: AggregationsWithAccessor,
@@ -83,20 +55,19 @@ impl BucketAggregationWithAccessor {
         bucket_count: Rc<AtomicU32>,
         max_bucket_count: u32,
     ) -> crate::Result<BucketAggregationWithAccessor> {
-        let mut inverted_index = None;
+        let mut str_dict_column = None;
         let (accessor, field_type) = match &bucket {
             BucketAggregationType::Range(RangeAggregation {
                 field: field_name, ..
-            }) => get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?,
+            }) => get_ff_reader_and_validate(reader, field_name)?,
             BucketAggregationType::Histogram(HistogramAggregation {
                 field: field_name, ..
-            }) => get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?,
+            }) => get_ff_reader_and_validate(reader, field_name)?,
             BucketAggregationType::Terms(TermsAggregation {
                 field: field_name, ..
             }) => {
-                let field = reader.schema().get_field(field_name)?;
-                inverted_index = Some(reader.inverted_index(field)?);
-                get_ff_reader_and_validate(reader, field_name, Cardinality::MultiValues)?
+                str_dict_column = reader.fast_fields().str(field_name)?;
+                get_ff_reader_and_validate(reader, field_name)?
             }
         };
         let sub_aggregation = sub_aggregation.clone();
@@ -110,7 +81,7 @@ impl BucketAggregationWithAccessor {
                 max_bucket_count,
             )?,
             bucket_agg: bucket.clone(),
-            inverted_index,
+            str_dict_column,
             bucket_count: BucketCount {
                 bucket_count,
                 max_bucket_count,
@@ -124,7 +95,7 @@ impl BucketAggregationWithAccessor {
 pub struct MetricAggregationWithAccessor {
     pub metric: MetricAggregation,
     pub field_type: Type,
-    pub accessor: Arc<dyn Column>,
+    pub accessor: Column<u64>,
 }
 
 impl MetricAggregationWithAccessor {
@@ -139,13 +110,10 @@ impl MetricAggregationWithAccessor {
             | MetricAggregation::Min(MinAggregation { field: field_name })
             | MetricAggregation::Stats(StatsAggregation { field: field_name })
             | MetricAggregation::Sum(SumAggregation { field: field_name }) => {
-                let (accessor, field_type) =
-                    get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?;
+                let (accessor, field_type) = get_ff_reader_and_validate(reader, field_name)?;
 
                 Ok(MetricAggregationWithAccessor {
-                    accessor: accessor
-                        .into_single()
-                        .expect("unexpected fast field cardinality"),
+                    accessor,
                     field_type,
                     metric: metric.clone(),
                 })
@@ -190,32 +158,22 @@ pub(crate) fn get_aggs_with_accessor_and_validate(
 fn get_ff_reader_and_validate(
     reader: &SegmentReader,
     field_name: &str,
-    cardinality: Cardinality,
-) -> crate::Result<(FastFieldAccessor, Type)> {
+) -> crate::Result<(columnar::Column<u64>, Type)> {
     let field = reader.schema().get_field(field_name)?;
-    let field_type = reader.schema().get_field_entry(field).field_type();
-
-    if let Some((_ff_type, field_cardinality)) = type_and_cardinality(field_type) {
-        if cardinality != field_cardinality {
-            return Err(TantivyError::InvalidArgument(format!(
-                "Invalid field cardinality on field {} expected {:?}, but got {:?}",
-                field_name, cardinality, field_cardinality
-            )));
-        }
-    } else {
-        return Err(TantivyError::InvalidArgument(format!(
-            "Only fast fields of type f64, u64, i64 are supported, but got {:?} ",
-            field_type.value_type()
-        )));
-    };
+    // TODO we should get type metadata from columnar
+    let field_type = reader
+        .schema()
+        .get_field_entry(field)
+        .field_type()
+        .value_type();
+    // TODO Do validation
 
     let ff_fields = reader.fast_fields();
-    match cardinality {
-        Cardinality::SingleValue => ff_fields
-            .u64_lenient(field_name)
-            .map(|field| (FastFieldAccessor::Single(field), field_type.value_type())),
-        Cardinality::MultiValues => ff_fields
-            .u64s_lenient(field_name)
-            .map(|field| (FastFieldAccessor::Multi(field), field_type.value_type())),
-    }
+    let ff_field = ff_fields.u64_lenient(field_name)?.ok_or_else(|| {
+        TantivyError::InvalidArgument(format!(
+            "No numerical fast field found for field: {}",
+            field_name
+        ))
+    })?;
+    Ok((ff_field, field_type))
 }

src/aggregation/bucket/histogram/date_histogram.rs

@@ -0,0 +1,137 @@
+use serde::{Deserialize, Serialize};
+
+/// DateHistogramAggregation is similar to `HistogramAggregation`, but it can only be used with date
+/// type.
+///
+/// Currently only **fixed time** intervals are supported. Calendar-aware time intervals are not
+/// supported.
+///
+/// Like the histogram, values are rounded down into the closest bucket.
+///
+/// For this calculation all fastfield values are converted to f64.
+///
+/// # Limitations/Compatibility
+/// Only fixed time intervals are supported.
+///
+/// # JSON Format
+/// ```json
+/// {
+///     "prices": {
+///         "date_histogram": {
+///             "field": "price",
+///             "fixed_interval": "30d"
+///         }
+///     }
+/// }
+/// ```
+///
+/// Response
+/// See [`BucketEntry`](crate::aggregation::agg_result::BucketEntry)
+#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
+pub struct DateHistogramAggregationReq {
+    /// The field to aggregate on.
+    pub field: String,
+    /// The interval to chunk your data range. Each bucket spans a value range of
+    /// [0..fixed_interval). Accepted values
+    ///
+    /// Fixed intervals are configured with the `fixed_interval` parameter.
+    /// In contrast to calendar-aware intervals, fixed intervals are a fixed number of SI units and
+    /// never deviate, regardless of where they fall on the calendar. One second is always
+    /// composed of 1000ms. This allows fixed intervals to be specified in any multiple of the
+    /// supported units. However, it means fixed intervals cannot express other units such as
+    /// months, since the duration of a month is not a fixed quantity. Attempting to specify a
+    /// calendar interval like month or quarter will return an Error.
+    ///
+    /// The accepted units for fixed intervals are:
+    /// * `ms`: milliseconds
+    /// * `s`: seconds. Defined as 1000 milliseconds each.
+    /// * `m`: minutes. Defined as 60 seconds each (60_000 milliseconds).
+    /// * `h`: hours. Defined as 60 minutes each (3_600_000 milliseconds).
+    /// * `d`: days. Defined as 24 hours (86_400_000 milliseconds).
+    ///
+    /// Fractional time values are not supported, but you can address this by shifting to another
+    /// time unit (e.g., `1.5h` could instead be specified as `90m`).
+    pub fixed_interval: String,
+    /// Intervals implicitly defines an absolute grid of buckets `[interval * k, interval * (k +
+    /// 1))`.
+    pub offset: Option<String>,
+    /// Whether to return the buckets as a hash map
+    #[serde(default)]
+    pub keyed: bool,
+}
+
+impl DateHistogramAggregationReq {
+    fn validate(&self) -> crate::Result<()> {
+        Ok(())
+    }
+}
+
+#[derive(Debug, PartialEq, Eq)]
+/// Errors when parsing the fixed interval for `DateHistogramAggregationReq`.
+pub enum DateHistogramParseError {
+    /// Unit not recognized in passed String
+    UnitNotRecognized(String),
+    /// Number not found in passed String
+    NumberMissing(String),
+    /// Unit not found in passed String
+    UnitMissing(String),
+}
+
+fn parse_into_milliseconds(input: &str) -> Result<u64, DateHistogramParseError> {
+    let split_boundary = input
+        .as_bytes()
+        .iter()
+        .take_while(|byte| byte.is_ascii_digit())
+        .count();
+    let (number, unit) = input.split_at(split_boundary);
+    if number.is_empty() {
+        return Err(DateHistogramParseError::NumberMissing(input.to_string()));
+    }
+    if unit.is_empty() {
+        return Err(DateHistogramParseError::UnitMissing(input.to_string()));
+    }
+    let number: u64 = number
+        .parse()
+        // Technically this should never happen, but there was a bug
+        // here and being defensive does not hurt.
+        .map_err(|_err| DateHistogramParseError::NumberMissing(input.to_string()))?;
+
+    let multiplier_from_unit = match unit {
+        "ms" => 1,
+        "s" => 1000,
+        "m" => 60 * 1000,
+        "h" => 60 * 60 * 1000,
+        "d" => 24 * 60 * 60 * 1000,
+        _ => return Err(DateHistogramParseError::UnitNotRecognized(unit.to_string())),
+    };
+
+    Ok(number * multiplier_from_unit)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_parse_into_milliseconds() {
+        assert_eq!(parse_into_milliseconds("1m").unwrap(), 60_000);
+        assert_eq!(parse_into_milliseconds("2m").unwrap(), 120_000);
+        assert_eq!(
+            parse_into_milliseconds("2y").unwrap_err(),
+            DateHistogramParseError::UnitNotRecognized("y".to_string())
+        );
+        assert_eq!(
+            parse_into_milliseconds("2000").unwrap_err(),
+            DateHistogramParseError::UnitMissing("2000".to_string())
+        );
+        assert_eq!(
+            parse_into_milliseconds("ms").unwrap_err(),
+            DateHistogramParseError::NumberMissing("ms".to_string())
+        );
+    }
+
+    #[test]
+    fn test_parse_into_milliseconds_do_not_accept_non_ascii() {
+        assert!(parse_into_milliseconds("１m").is_err());
+    }
+}

src/aggregation/bucket/histogram/histogram.rs

@@ -1,7 +1,7 @@
 use std::cmp::Ordering;
 use std::fmt::Display;
 
-use fastfield_codecs::Column;
+use columnar::Column;
 use itertools::Itertools;
 use serde::{Deserialize, Serialize};
 
@@ -13,7 +13,9 @@ use crate::aggregation::agg_result::BucketEntry;
 use crate::aggregation::intermediate_agg_result::{
     IntermediateAggregationResults, IntermediateBucketResult, IntermediateHistogramBucketEntry,
 };
-use crate::aggregation::segment_agg_result::SegmentAggregationResultsCollector;
+use crate::aggregation::segment_agg_result::{
+    GenericSegmentAggregationResultsCollector, SegmentAggregationCollector,
+};
 use crate::aggregation::{f64_from_fastfield_u64, format_date};
 use crate::schema::{Schema, Type};
 use crate::{DocId, TantivyError};
@@ -62,7 +64,6 @@ use crate::{DocId, TantivyError};
 ///
 /// Response
 /// See [`BucketEntry`](crate::aggregation::agg_result::BucketEntry)
-
 #[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
 pub struct HistogramAggregation {
     /// The field to aggregate on.
@@ -184,7 +185,7 @@ pub(crate) struct SegmentHistogramBucketEntry {
 impl SegmentHistogramBucketEntry {
     pub(crate) fn into_intermediate_bucket_entry(
         self,
-        sub_aggregation: SegmentAggregationResultsCollector,
+        sub_aggregation: GenericSegmentAggregationResultsCollector,
         agg_with_accessor: &AggregationsWithAccessor,
     ) -> crate::Result<IntermediateHistogramBucketEntry> {
         Ok(IntermediateHistogramBucketEntry {
@@ -198,11 +199,11 @@ impl SegmentHistogramBucketEntry {
 
 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug)]
 pub struct SegmentHistogramCollector {
     /// The buckets containing the aggregation data.
     buckets: Vec<SegmentHistogramBucketEntry>,
-    sub_aggregations: Option<Vec<SegmentAggregationResultsCollector>>,
+    sub_aggregations: Option<Vec<GenericSegmentAggregationResultsCollector>>,
     field_type: Type,
     interval: f64,
     offset: f64,
@@ -283,7 +284,7 @@ impl SegmentHistogramCollector {
         req: &HistogramAggregation,
         sub_aggregation: &AggregationsWithAccessor,
         field_type: Type,
-        accessor: &dyn Column<u64>,
+        accessor: &Column<u64>,
     ) -> crate::Result<Self> {
         req.validate()?;
         let min = f64_from_fastfield_u64(accessor.min_value(), &field_type);
@@ -300,7 +301,7 @@ impl SegmentHistogramCollector {
             None
         } else {
             let sub_aggregation =
-                SegmentAggregationResultsCollector::from_req_and_validate(sub_aggregation)?;
+                GenericSegmentAggregationResultsCollector::from_req_and_validate(sub_aggregation)?;
             Some(buckets.iter().map(|_| sub_aggregation.clone()).collect())
         };
 
@@ -335,7 +336,7 @@ impl SegmentHistogramCollector {
     #[inline]
     pub(crate) fn collect_block(
         &mut self,
-        doc: &[DocId],
+        docs: &[DocId],
         bucket_with_accessor: &BucketAggregationWithAccessor,
         force_flush: bool,
     ) -> crate::Result<()> {
@@ -346,64 +347,20 @@ impl SegmentHistogramCollector {
         let get_bucket_num =
             |val| (get_bucket_num_f64(val, interval, offset) as i64 - first_bucket_num) as usize;
 
-        let accessor = bucket_with_accessor
-            .accessor
-            .as_single()
-            .expect("unexpected fast field cardinatility");
-        let mut iter = doc.chunks_exact(4);
-        for docs in iter.by_ref() {
-            let val0 = self.f64_from_fastfield_u64(accessor.get_val(docs[0]));
-            let val1 = self.f64_from_fastfield_u64(accessor.get_val(docs[1]));
-            let val2 = self.f64_from_fastfield_u64(accessor.get_val(docs[2]));
-            let val3 = self.f64_from_fastfield_u64(accessor.get_val(docs[3]));
+        let accessor = &bucket_with_accessor.accessor;
+        for doc in docs {
+            for val in accessor.values(*doc) {
+                let val = self.f64_from_fastfield_u64(val);
 
-            let bucket_pos0 = get_bucket_num(val0);
-            let bucket_pos1 = get_bucket_num(val1);
-            let bucket_pos2 = get_bucket_num(val2);
-            let bucket_pos3 = get_bucket_num(val3);
-
-            self.increment_bucket_if_in_bounds(
-                val0,
-                &bounds,
-                bucket_pos0,
-                docs[0],
-                &bucket_with_accessor.sub_aggregation,
-            )?;
-            self.increment_bucket_if_in_bounds(
-                val1,
-                &bounds,
-                bucket_pos1,
-                docs[1],
-                &bucket_with_accessor.sub_aggregation,
-            )?;
-            self.increment_bucket_if_in_bounds(
-                val2,
-                &bounds,
-                bucket_pos2,
-                docs[2],
-                &bucket_with_accessor.sub_aggregation,
-            )?;
-            self.increment_bucket_if_in_bounds(
-                val3,
-                &bounds,
-                bucket_pos3,
-                docs[3],
-                &bucket_with_accessor.sub_aggregation,
-            )?;
-        }
-        for &doc in iter.remainder() {
-            let val = f64_from_fastfield_u64(accessor.get_val(doc), &self.field_type);
-            if !bounds.contains(val) {
-                continue;
+                let bucket_pos = get_bucket_num(val);
+                self.increment_bucket_if_in_bounds(
+                    val,
+                    &bounds,
+                    bucket_pos,
+                    *doc,
+                    &bucket_with_accessor.sub_aggregation,
+                )?;
             }
-            let bucket_pos = (get_bucket_num_f64(val, self.interval, self.offset) as i64
-                - self.first_bucket_num) as usize;
-
-            debug_assert_eq!(
-                self.buckets[bucket_pos].key,
-                get_bucket_val(val, self.interval, self.offset)
-            );
-            self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
         }
         if force_flush {
             if let Some(sub_aggregations) = self.sub_aggregations.as_mut() {

src/aggregation/bucket/histogram/mod.rs

@@ -1,2 +1,4 @@
+// mod date_histogram;
 mod histogram;
+// pub use date_histogram::*;
 pub use histogram::*;

src/aggregation/bucket/range.rs

@@ -1,7 +1,7 @@
 use std::fmt::Debug;
 use std::ops::Range;
 
-use fastfield_codecs::MonotonicallyMappableToU64;
+use columnar::MonotonicallyMappableToU64;
 use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
 
@@ -11,7 +11,9 @@ use crate::aggregation::agg_req_with_accessor::{
 use crate::aggregation::intermediate_agg_result::{
     IntermediateBucketResult, IntermediateRangeBucketEntry, IntermediateRangeBucketResult,
 };
-use crate::aggregation::segment_agg_result::{BucketCount, SegmentAggregationResultsCollector};
+use crate::aggregation::segment_agg_result::{
+    BucketCount, GenericSegmentAggregationResultsCollector, SegmentAggregationCollector,
+};
 use crate::aggregation::{
     f64_from_fastfield_u64, f64_to_fastfield_u64, format_date, Key, SerializedKey,
 };
@@ -114,7 +116,7 @@ impl From<Range<u64>> for InternalRangeAggregationRange {
     }
 }
 
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug)]
 pub(crate) struct SegmentRangeAndBucketEntry {
     range: Range<u64>,
     bucket: SegmentRangeBucketEntry,
@@ -122,18 +124,18 @@ pub(crate) struct SegmentRangeAndBucketEntry {
 
 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug)]
 pub struct SegmentRangeCollector {
     /// The buckets containing the aggregation data.
     buckets: Vec<SegmentRangeAndBucketEntry>,
     field_type: Type,
 }
 
-#[derive(Clone, PartialEq)]
+#[derive(Clone)]
 pub(crate) struct SegmentRangeBucketEntry {
     pub key: Key,
     pub doc_count: u64,
-    pub sub_aggregation: Option<SegmentAggregationResultsCollector>,
+    pub sub_aggregation: Option<GenericSegmentAggregationResultsCollector>,
     /// The from range of the bucket. Equals `f64::MIN` when `None`.
     pub from: Option<f64>,
     /// The to range of the bucket. Equals `f64::MAX` when `None`. Open interval, `to` is not
@@ -227,9 +229,11 @@ impl SegmentRangeCollector {
                 let sub_aggregation = if sub_aggregation.is_empty() {
                     None
                 } else {
-                    Some(SegmentAggregationResultsCollector::from_req_and_validate(
-                        sub_aggregation,
-                    )?)
+                    Some(
+                        GenericSegmentAggregationResultsCollector::from_req_and_validate(
+                            sub_aggregation,
+                        )?,
+                    )
                 };
 
                 Ok(SegmentRangeAndBucketEntry {
@@ -257,35 +261,18 @@ impl SegmentRangeCollector {
     #[inline]
     pub(crate) fn collect_block(
         &mut self,
-        doc: &[DocId],
+        docs: &[DocId],
         bucket_with_accessor: &BucketAggregationWithAccessor,
         force_flush: bool,
     ) -> crate::Result<()> {
-        let mut iter = doc.chunks_exact(4);
-        let accessor = bucket_with_accessor
-            .accessor
-            .as_single()
-            .expect("unexpected fast field cardinality");
-        for docs in iter.by_ref() {
-            let val1 = accessor.get_val(docs[0]);
-            let val2 = accessor.get_val(docs[1]);
-            let val3 = accessor.get_val(docs[2]);
-            let val4 = accessor.get_val(docs[3]);
-            let bucket_pos1 = self.get_bucket_pos(val1);
-            let bucket_pos2 = self.get_bucket_pos(val2);
-            let bucket_pos3 = self.get_bucket_pos(val3);
-            let bucket_pos4 = self.get_bucket_pos(val4);
+        let accessor = &bucket_with_accessor.accessor;
+        for doc in docs {
+            for val in accessor.values(*doc) {
+                let bucket_pos = self.get_bucket_pos(val);
+                self.increment_bucket(bucket_pos, *doc, &bucket_with_accessor.sub_aggregation)?;
+            }
+        }
 
-            self.increment_bucket(bucket_pos1, docs[0], &bucket_with_accessor.sub_aggregation)?;
-            self.increment_bucket(bucket_pos2, docs[1], &bucket_with_accessor.sub_aggregation)?;
-            self.increment_bucket(bucket_pos3, docs[2], &bucket_with_accessor.sub_aggregation)?;
-            self.increment_bucket(bucket_pos4, docs[3], &bucket_with_accessor.sub_aggregation)?;
-        }
-        for &doc in iter.remainder() {
-            let val = accessor.get_val(doc);
-            let bucket_pos = self.get_bucket_pos(val);
-            self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
-        }
         if force_flush {
             for bucket in &mut self.buckets {
                 if let Some(sub_aggregation) = &mut bucket.bucket.sub_aggregation {
@@ -434,7 +421,7 @@ pub(crate) fn range_to_key(range: &Range<u64>, field_type: &Type) -> crate::Resu
 #[cfg(test)]
 mod tests {
 
-    use fastfield_codecs::MonotonicallyMappableToU64;
+    use columnar::MonotonicallyMappableToU64;
     use serde_json::Value;
 
     use super::*;

src/aggregation/bucket/term_agg.rs

@@ -1,6 +1,5 @@
 use std::fmt::Debug;
 
-use itertools::Itertools;
 use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
 
@@ -11,10 +10,10 @@ use crate::aggregation::agg_req_with_accessor::{
 use crate::aggregation::intermediate_agg_result::{
     IntermediateBucketResult, IntermediateTermBucketEntry, IntermediateTermBucketResult,
 };
-use crate::aggregation::segment_agg_result::{BucketCount, SegmentAggregationResultsCollector};
+use crate::aggregation::segment_agg_result::{
+    build_segment_agg_collector, SegmentAggregationCollector,
+};
 use crate::error::DataCorruption;
-use crate::fastfield::MultiValuedFastFieldReader;
-use crate::schema::Type;
 use crate::{DocId, TantivyError};
 
 /// Creates a bucket for every unique term and counts the number of occurences.
@@ -75,9 +74,9 @@ use crate::{DocId, TantivyError};
 ///     ...
 ///     "aggregations": {
 ///         "genres": {
-///             "doc_count_error_upper_bound": 0,   
-///             "sum_other_doc_count": 0,           
-///             "buckets": [                        
+///             "doc_count_error_upper_bound": 0,
+///             "sum_other_doc_count": 0,
+///             "buckets": [
 ///                 { "key": "drumnbass", "doc_count": 6 },
 ///                 { "key": "raggae", "doc_count": 4 },
 ///                 { "key": "jazz", "doc_count": 2 }
@@ -196,17 +195,16 @@ impl TermsAggregationInternal {
     }
 }
 
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug, Default)]
 /// Container to store term_ids and their buckets.
 struct TermBuckets {
     pub(crate) entries: FxHashMap<u32, TermBucketEntry>,
-    blueprint: Option<SegmentAggregationResultsCollector>,
 }
 
-#[derive(Clone, PartialEq, Default)]
+#[derive(Clone, Default)]
 struct TermBucketEntry {
     doc_count: u64,
-    sub_aggregations: Option<SegmentAggregationResultsCollector>,
+    sub_aggregations: Option<Box<dyn SegmentAggregationCollector>>,
 }
 
 impl Debug for TermBucketEntry {
@@ -218,7 +216,7 @@ impl Debug for TermBucketEntry {
 }
 
 impl TermBucketEntry {
-    fn from_blueprint(blueprint: &Option<SegmentAggregationResultsCollector>) -> Self {
+    fn from_blueprint(blueprint: &Option<Box<dyn SegmentAggregationCollector>>) -> Self {
         Self {
             doc_count: 0,
             sub_aggregations: blueprint.clone(),
@@ -243,50 +241,6 @@ impl TermBucketEntry {
 }
 
 impl TermBuckets {
-    pub(crate) fn from_req_and_validate(
-        sub_aggregation: &AggregationsWithAccessor,
-        _max_term_id: usize,
-    ) -> crate::Result<Self> {
-        let has_sub_aggregations = sub_aggregation.is_empty();
-
-        let blueprint = if has_sub_aggregations {
-            let sub_aggregation =
-                SegmentAggregationResultsCollector::from_req_and_validate(sub_aggregation)?;
-            Some(sub_aggregation)
-        } else {
-            None
-        };
-
-        Ok(TermBuckets {
-            blueprint,
-            entries: Default::default(),
-        })
-    }
-
-    fn increment_bucket(
-        &mut self,
-        term_ids: &[u64],
-        doc: DocId,
-        sub_aggregation: &AggregationsWithAccessor,
-        bucket_count: &BucketCount,
-        blueprint: &Option<SegmentAggregationResultsCollector>,
-    ) -> crate::Result<()> {
-        for &term_id in term_ids {
-            let entry = self.entries.entry(term_id as u32).or_insert_with(|| {
-                bucket_count.add_count(1);
-
-                TermBucketEntry::from_blueprint(blueprint)
-            });
-            entry.doc_count += 1;
-            if let Some(sub_aggregations) = entry.sub_aggregations.as_mut() {
-                sub_aggregations.collect(doc, sub_aggregation)?;
-            }
-        }
-        bucket_count.validate_bucket_count()?;
-
-        Ok(())
-    }
-
     fn force_flush(&mut self, agg_with_accessor: &AggregationsWithAccessor) -> crate::Result<()> {
         for entry in &mut self.entries.values_mut() {
             if let Some(sub_aggregations) = entry.sub_aggregations.as_mut() {
@@ -299,13 +253,12 @@ impl TermBuckets {
 
 /// The collector puts values from the fast field into the correct buckets and does a conversion to
 /// the correct datatype.
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug)]
 pub struct SegmentTermCollector {
     /// The buckets containing the aggregation data.
     term_buckets: TermBuckets,
     req: TermsAggregationInternal,
-    field_type: Type,
-    blueprint: Option<SegmentAggregationResultsCollector>,
+    blueprint: Option<Box<dyn SegmentAggregationCollector>>,
 }
 
 pub(crate) fn get_agg_name_and_property(name: &str) -> (&str, &str) {
@@ -317,12 +270,8 @@ impl SegmentTermCollector {
     pub(crate) fn from_req_and_validate(
         req: &TermsAggregation,
         sub_aggregations: &AggregationsWithAccessor,
-        field_type: Type,
-        accessor: &MultiValuedFastFieldReader<u64>,
     ) -> crate::Result<Self> {
-        let max_term_id = accessor.max_value();
-        let term_buckets =
-            TermBuckets::from_req_and_validate(sub_aggregations, max_term_id as usize)?;
+        let term_buckets = TermBuckets::default();
 
         if let Some(custom_order) = req.order.as_ref() {
             // Validate sub aggregtion exists
@@ -340,8 +289,7 @@ impl SegmentTermCollector {
 
         let has_sub_aggregations = !sub_aggregations.is_empty();
         let blueprint = if has_sub_aggregations {
-            let sub_aggregation =
-                SegmentAggregationResultsCollector::from_req_and_validate(sub_aggregations)?;
+            let sub_aggregation = build_segment_agg_collector(sub_aggregations)?;
             Some(sub_aggregation)
         } else {
             None
@@ -350,7 +298,6 @@ impl SegmentTermCollector {
         Ok(SegmentTermCollector {
             req: TermsAggregationInternal::from_req(req),
             term_buckets,
-            field_type,
             blueprint,
         })
     }
@@ -362,13 +309,19 @@ impl SegmentTermCollector {
         let mut entries: Vec<(u32, TermBucketEntry)> =
             self.term_buckets.entries.into_iter().collect();
 
-        let order_by_key = self.req.order.target == OrderTarget::Key;
         let order_by_sub_aggregation =
             matches!(self.req.order.target, OrderTarget::SubAggregation(_));
 
         match self.req.order.target {
             OrderTarget::Key => {
-                // defer order and cut_off after loading the texts from the dictionary
+                // We rely on the fact, that term ordinals match the order of the strings
+                // TODO: We could have a special collector, that keeps only TOP n results at any
+                // time.
+                if self.req.order.order == Order::Desc {
+                    entries.sort_unstable_by_key(|bucket| std::cmp::Reverse(bucket.0));
+                } else {
+                    entries.sort_unstable_by_key(|bucket| bucket.0);
+                }
             }
             OrderTarget::SubAggregation(_name) => {
                 // don't sort and cut off since it's hard to make assumptions on the quality of the
@@ -384,34 +337,40 @@ impl SegmentTermCollector {
             }
         }
 
-        let (term_doc_count_before_cutoff, mut sum_other_doc_count) =
-            if order_by_key || order_by_sub_aggregation {
-                (0, 0)
-            } else {
-                cut_off_buckets(&mut entries, self.req.segment_size as usize)
-            };
+        let (term_doc_count_before_cutoff, sum_other_doc_count) = if order_by_sub_aggregation {
+            (0, 0)
+        } else {
+            cut_off_buckets(&mut entries, self.req.segment_size as usize)
+        };
 
         let inverted_index = agg_with_accessor
-            .inverted_index
+            .str_dict_column
             .as_ref()
             .expect("internal error: inverted index not loaded for term aggregation");
-        let term_dict = inverted_index.terms();
+        let term_dict = inverted_index;
 
         let mut dict: FxHashMap<String, IntermediateTermBucketEntry> = Default::default();
-        let mut buffer = vec![];
+        let mut buffer = String::new();
         for (term_id, entry) in entries {
-            term_dict
-                .ord_to_term(term_id as u64, &mut buffer)
-                .expect("could not find term");
+            if !term_dict.ord_to_str(term_id as u64, &mut buffer)? {
+                return Err(TantivyError::InternalError(format!(
+                    "Couldn't find term_id {} in dict",
+                    term_id
+                )));
+            }
             dict.insert(
-                String::from_utf8(buffer.to_vec())
-                    .map_err(|utf8_err| DataCorruption::comment_only(utf8_err.to_string()))?,
+                buffer.to_string(),
                 entry.into_intermediate_bucket_entry(&agg_with_accessor.sub_aggregation)?,
             );
         }
         if self.req.min_doc_count == 0 {
-            let mut stream = term_dict.stream()?;
+            // TODO: Handle rev streaming for descending sorting by keys
+            let mut stream = term_dict.dictionary().stream()?;
             while let Some((key, _ord)) = stream.next() {
+                if dict.len() >= self.req.segment_size as usize {
+                    break;
+                }
+
                 let key = std::str::from_utf8(key)
                     .map_err(|utf8_err| DataCorruption::comment_only(utf8_err.to_string()))?;
                 if !dict.contains_key(key) {
@@ -420,20 +379,6 @@ impl SegmentTermCollector {
             }
         }
 
-        if order_by_key {
-            let mut dict_entries = dict.into_iter().collect_vec();
-            if self.req.order.order == Order::Desc {
-                dict_entries.sort_unstable_by(|(key1, _), (key2, _)| key1.cmp(key2));
-            } else {
-                dict_entries.sort_unstable_by(|(key1, _), (key2, _)| key2.cmp(key1));
-            }
-            let (_, sum_other_docs) =
-                cut_off_buckets(&mut dict_entries, self.req.segment_size as usize);
-
-            sum_other_doc_count += sum_other_docs;
-            dict = dict_entries.into_iter().collect();
-        }
-
         Ok(IntermediateBucketResult::Terms(
             IntermediateTermBucketResult {
                 entries: dict,
@@ -446,65 +391,26 @@ impl SegmentTermCollector {
     #[inline]
     pub(crate) fn collect_block(
         &mut self,
-        doc: &[DocId],
+        docs: &[DocId],
         bucket_with_accessor: &BucketAggregationWithAccessor,
         force_flush: bool,
     ) -> crate::Result<()> {
-        let accessor = bucket_with_accessor
-            .accessor
-            .as_multi()
-            .expect("unexpected fast field cardinatility");
-        let mut iter = doc.chunks_exact(4);
-        let mut vals1 = vec![];
-        let mut vals2 = vec![];
-        let mut vals3 = vec![];
-        let mut vals4 = vec![];
-        for docs in iter.by_ref() {
-            accessor.get_vals(docs[0], &mut vals1);
-            accessor.get_vals(docs[1], &mut vals2);
-            accessor.get_vals(docs[2], &mut vals3);
-            accessor.get_vals(docs[3], &mut vals4);
+        let accessor = &bucket_with_accessor.accessor;
 
-            self.term_buckets.increment_bucket(
-                &vals1,
-                docs[0],
-                &bucket_with_accessor.sub_aggregation,
-                &bucket_with_accessor.bucket_count,
-                &self.blueprint,
-            )?;
-            self.term_buckets.increment_bucket(
-                &vals2,
-                docs[1],
-                &bucket_with_accessor.sub_aggregation,
-                &bucket_with_accessor.bucket_count,
-                &self.blueprint,
-            )?;
-            self.term_buckets.increment_bucket(
-                &vals3,
-                docs[2],
-                &bucket_with_accessor.sub_aggregation,
-                &bucket_with_accessor.bucket_count,
-                &self.blueprint,
-            )?;
-            self.term_buckets.increment_bucket(
-                &vals4,
-                docs[3],
-                &bucket_with_accessor.sub_aggregation,
-                &bucket_with_accessor.bucket_count,
-                &self.blueprint,
-            )?;
+        for doc in docs {
+            for term_id in accessor.values(*doc) {
+                let entry = self
+                    .term_buckets
+                    .entries
+                    .entry(term_id as u32)
+                    .or_insert_with(|| TermBucketEntry::from_blueprint(&self.blueprint));
+                entry.doc_count += 1;
+                if let Some(sub_aggregations) = entry.sub_aggregations.as_mut() {
+                    sub_aggregations.collect(*doc, &bucket_with_accessor.sub_aggregation)?;
+                }
+            }
         }
-        for &doc in iter.remainder() {
-            accessor.get_vals(doc, &mut vals1);
 
-            self.term_buckets.increment_bucket(
-                &vals1,
-                doc,
-                &bucket_with_accessor.sub_aggregation,
-                &bucket_with_accessor.bucket_count,
-                &self.blueprint,
-            )?;
-        }
         if force_flush {
             self.term_buckets
                 .force_flush(&bucket_with_accessor.sub_aggregation)?;
@@ -923,14 +829,14 @@ mod tests {
         ];
         let index = get_test_index_from_values_and_terms(merge_segments, &segment_and_terms)?;
 
-        // key desc
+        // key asc
         let agg_req: Aggregations = vec![(
             "my_texts".to_string(),
             Aggregation::Bucket(BucketAggregation {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Desc,
+                        order: Order::Asc,
                         target: OrderTarget::Key,
                     }),
                     ..Default::default()
@@ -957,7 +863,7 @@ mod tests {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Desc,
+                        order: Order::Asc,
                         target: OrderTarget::Key,
                     }),
                     size: Some(2),
@@ -981,14 +887,14 @@ mod tests {
 
         assert_eq!(res["my_texts"]["sum_other_doc_count"], 3);
 
-        // key desc and segment_size cut_off
+        // key asc and segment_size cut_off
         let agg_req: Aggregations = vec![(
             "my_texts".to_string(),
             Aggregation::Bucket(BucketAggregation {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Desc,
+                        order: Order::Asc,
                         target: OrderTarget::Key,
                     }),
                     size: Some(2),
@@ -1011,14 +917,14 @@ mod tests {
             serde_json::Value::Null
         );
 
-        // key asc
+        // key desc
         let agg_req: Aggregations = vec![(
             "my_texts".to_string(),
             Aggregation::Bucket(BucketAggregation {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Asc,
+                        order: Order::Desc,
                         target: OrderTarget::Key,
                     }),
                     ..Default::default()
@@ -1038,14 +944,14 @@ mod tests {
         assert_eq!(res["my_texts"]["buckets"][2]["doc_count"], 5);
         assert_eq!(res["my_texts"]["sum_other_doc_count"], 0);
 
-        // key asc, size cut_off
+        // key desc, size cut_off
         let agg_req: Aggregations = vec![(
             "my_texts".to_string(),
             Aggregation::Bucket(BucketAggregation {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Asc,
+                        order: Order::Desc,
                         target: OrderTarget::Key,
                     }),
                     size: Some(2),
@@ -1068,14 +974,14 @@ mod tests {
         );
         assert_eq!(res["my_texts"]["sum_other_doc_count"], 5);
 
-        // key asc, segment_size cut_off
+        // key desc, segment_size cut_off
         let agg_req: Aggregations = vec![(
             "my_texts".to_string(),
             Aggregation::Bucket(BucketAggregation {
                 bucket_agg: BucketAggregationType::Terms(TermsAggregation {
                     field: "string_id".to_string(),
                     order: Some(CustomOrder {
-                        order: Order::Asc,
+                        order: Order::Desc,
                         target: OrderTarget::Key,
                     }),
                     size: Some(2),
@@ -1207,36 +1113,37 @@ mod tests {
         Ok(())
     }
 
-    #[test]
-    fn terms_aggregation_term_bucket_limit() -> crate::Result<()> {
-        let terms: Vec<String> = (0..100_000).map(|el| el.to_string()).collect();
-        let terms_per_segment = vec![terms.iter().map(|el| el.as_str()).collect()];
+    // TODO reenable with memory limit
+    //#[test]
+    // fn terms_aggregation_term_bucket_limit() -> crate::Result<()> {
+    // let terms: Vec<String> = (0..100_000).map(|el| el.to_string()).collect();
+    // let terms_per_segment = vec![terms.iter().map(|el| el.as_str()).collect()];
 
-        let index = get_test_index_from_terms(true, &terms_per_segment)?;
+    // let index = get_test_index_from_terms(true, &terms_per_segment)?;
 
-        let agg_req: Aggregations = vec![(
-            "my_texts".to_string(),
-            Aggregation::Bucket(BucketAggregation {
-                bucket_agg: BucketAggregationType::Terms(TermsAggregation {
-                    field: "string_id".to_string(),
-                    min_doc_count: Some(0),
-                    ..Default::default()
-                }),
-                sub_aggregation: Default::default(),
-            }),
-        )]
-        .into_iter()
-        .collect();
+    // let agg_req: Aggregations = vec![(
+    //"my_texts".to_string(),
+    // Aggregation::Bucket(BucketAggregation {
+    // bucket_agg: BucketAggregationType::Terms(TermsAggregation {
+    // field: "string_id".to_string(),
+    // min_doc_count: Some(0),
+    //..Default::default()
+    //}),
+    // sub_aggregation: Default::default(),
+    //}),
+    //)]
+    //.into_iter()
+    //.collect();
 
-        let res = exec_request_with_query(agg_req, &index, None);
+    // let res = exec_request_with_query(agg_req, &index, None);
 
-        assert!(res.is_err());
+    // assert!(res.is_err());
 
-        Ok(())
-    }
+    // Ok(())
+    //}
 
     #[test]
-    fn terms_aggregation_multi_token_per_doc() -> crate::Result<()> {
+    fn terms_aggregation_different_tokenizer_on_ff_test() -> crate::Result<()> {
         let terms = vec!["Hello Hello", "Hallo Hallo"];
 
         let index = get_test_index_from_terms(true, &[terms])?;
@@ -1256,12 +1163,13 @@ mod tests {
         .collect();
 
         let res = exec_request_with_query(agg_req, &index, None).unwrap();
+        println!("{}", serde_json::to_string_pretty(&res).unwrap());
 
-        assert_eq!(res["my_texts"]["buckets"][0]["key"], "hello");
-        assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 2);
+        assert_eq!(res["my_texts"]["buckets"][0]["key"], "Hallo Hallo");
+        assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 1);
 
-        assert_eq!(res["my_texts"]["buckets"][1]["key"], "hallo");
-        assert_eq!(res["my_texts"]["buckets"][1]["doc_count"], 2);
+        assert_eq!(res["my_texts"]["buckets"][1]["key"], "Hello Hello");
+        assert_eq!(res["my_texts"]["buckets"][1]["doc_count"], 1);
 
         Ok(())
     }
@@ -1352,68 +1260,3 @@ mod tests {
         Ok(())
     }
 }
-
-#[cfg(all(test, feature = "unstable"))]
-mod bench {
-
-    use itertools::Itertools;
-    use rand::seq::SliceRandom;
-    use rand::thread_rng;
-
-    use super::*;
-
-    fn get_collector_with_buckets(num_docs: u64) -> TermBuckets {
-        TermBuckets::from_req_and_validate(&Default::default(), num_docs as usize).unwrap()
-    }
-
-    fn get_rand_terms(total_terms: u64, num_terms_returned: u64) -> Vec<u64> {
-        let mut rng = thread_rng();
-
-        let all_terms = (0..total_terms - 1).collect_vec();
-
-        let mut vals = vec![];
-        for _ in 0..num_terms_returned {
-            let val = all_terms.as_slice().choose(&mut rng).unwrap();
-            vals.push(*val);
-        }
-
-        vals
-    }
-
-    fn bench_term_buckets(b: &mut test::Bencher, num_terms: u64, total_terms: u64) {
-        let mut collector = get_collector_with_buckets(total_terms);
-        let vals = get_rand_terms(total_terms, num_terms);
-        let aggregations_with_accessor: AggregationsWithAccessor = Default::default();
-        let bucket_count: BucketCount = BucketCount {
-            bucket_count: Default::default(),
-            max_bucket_count: 1_000_001u32,
-        };
-        b.iter(|| {
-            for &val in &vals {
-                collector
-                    .increment_bucket(&[val], 0, &aggregations_with_accessor, &bucket_count, &None)
-                    .unwrap();
-            }
-        })
-    }
-
-    #[bench]
-    fn bench_term_buckets_500_of_1_000_000(b: &mut test::Bencher) {
-        bench_term_buckets(b, 500u64, 1_000_000u64)
-    }
-
-    #[bench]
-    fn bench_term_buckets_1_000_000_of_50_000(b: &mut test::Bencher) {
-        bench_term_buckets(b, 1_000_000u64, 50_000u64)
-    }
-
-    #[bench]
-    fn bench_term_buckets_1_000_000_of_50(b: &mut test::Bencher) {
-        bench_term_buckets(b, 1_000_000u64, 50u64)
-    }
-
-    #[bench]
-    fn bench_term_buckets_1_000_000_of_1_000_000(b: &mut test::Bencher) {
-        bench_term_buckets(b, 1_000_000u64, 1_000_000u64)
-    }
-}

src/aggregation/collector.rs

@@ -4,7 +4,7 @@ use super::agg_req::Aggregations;
 use super::agg_req_with_accessor::AggregationsWithAccessor;
 use super::agg_result::AggregationResults;
 use super::intermediate_agg_result::IntermediateAggregationResults;
-use super::segment_agg_result::SegmentAggregationResultsCollector;
+use super::segment_agg_result::{build_segment_agg_collector, SegmentAggregationCollector};
 use crate::aggregation::agg_req_with_accessor::get_aggs_with_accessor_and_validate;
 use crate::collector::{Collector, SegmentCollector};
 use crate::schema::Schema;
@@ -137,7 +137,7 @@ fn merge_fruits(
 /// `AggregationSegmentCollector` does the aggregation collection on a segment.
 pub struct AggregationSegmentCollector {
     aggs_with_accessor: AggregationsWithAccessor,
-    result: SegmentAggregationResultsCollector,
+    result: Box<dyn SegmentAggregationCollector>,
     error: Option<TantivyError>,
 }
 
@@ -151,8 +151,7 @@ impl AggregationSegmentCollector {
     ) -> crate::Result<Self> {
         let aggs_with_accessor =
             get_aggs_with_accessor_and_validate(agg, reader, Rc::default(), max_bucket_count)?;
-        let result =
-            SegmentAggregationResultsCollector::from_req_and_validate(&aggs_with_accessor)?;
+        let result = build_segment_agg_collector(&aggs_with_accessor)?;
         Ok(AggregationSegmentCollector {
             aggs_with_accessor,
             result,

src/aggregation/intermediate_agg_result.rs

@@ -222,24 +222,23 @@ pub enum IntermediateMetricResult {
 
 impl From<SegmentMetricResultCollector> for IntermediateMetricResult {
     fn from(tree: SegmentMetricResultCollector) -> Self {
+        use super::metric::SegmentStatsType;
         match tree {
             SegmentMetricResultCollector::Stats(collector) => match collector.collecting_for {
-                super::metric::SegmentStatsType::Average => IntermediateMetricResult::Average(
+                SegmentStatsType::Average => IntermediateMetricResult::Average(
                     IntermediateAverage::from_collector(collector),
                 ),
-                super::metric::SegmentStatsType::Count => {
+                SegmentStatsType::Count => {
                     IntermediateMetricResult::Count(IntermediateCount::from_collector(collector))
                 }
-                super::metric::SegmentStatsType::Max => {
+                SegmentStatsType::Max => {
                     IntermediateMetricResult::Max(IntermediateMax::from_collector(collector))
                 }
-                super::metric::SegmentStatsType::Min => {
+                SegmentStatsType::Min => {
                     IntermediateMetricResult::Min(IntermediateMin::from_collector(collector))
                 }
-                super::metric::SegmentStatsType::Stats => {
-                    IntermediateMetricResult::Stats(collector.stats)
-                }
-                super::metric::SegmentStatsType::Sum => {
+                SegmentStatsType::Stats => IntermediateMetricResult::Stats(collector.stats),
+                SegmentStatsType::Sum => {
                     IntermediateMetricResult::Sum(IntermediateSum::from_collector(collector))
                 }
             },
@@ -499,7 +498,7 @@ impl IntermediateTermBucketResult {
         match req.order.target {
             OrderTarget::Key => {
                 buckets.sort_by(|left, right| {
-                    if req.order.order == Order::Desc {
+                    if req.order.order == Order::Asc {
                         left.key.partial_cmp(&right.key)
                     } else {
                         right.key.partial_cmp(&left.key)

src/aggregation/metric/stats.rs

@@ -1,7 +1,13 @@
-use fastfield_codecs::Column;
+use columnar::Column;
 use serde::{Deserialize, Serialize};
 
-use crate::aggregation::f64_from_fastfield_u64;
+use super::*;
+use crate::aggregation::agg_req_with_accessor::AggregationsWithAccessor;
+use crate::aggregation::intermediate_agg_result::{
+    IntermediateAggregationResults, IntermediateMetricResult,
+};
+use crate::aggregation::segment_agg_result::SegmentAggregationCollector;
+use crate::aggregation::{f64_from_fastfield_u64, VecWithNames};
 use crate::schema::Type;
 use crate::{DocId, TantivyError};
 
@@ -160,27 +166,74 @@ impl SegmentStatsCollector {
             stats: IntermediateStats::default(),
         }
     }
-    pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &dyn Column<u64>) {
-        let mut iter = doc.chunks_exact(4);
-        for docs in iter.by_ref() {
-            let val1 = field.get_val(docs[0]);
-            let val2 = field.get_val(docs[1]);
-            let val3 = field.get_val(docs[2]);
-            let val4 = field.get_val(docs[3]);
-            let val1 = f64_from_fastfield_u64(val1, &self.field_type);
-            let val2 = f64_from_fastfield_u64(val2, &self.field_type);
-            let val3 = f64_from_fastfield_u64(val3, &self.field_type);
-            let val4 = f64_from_fastfield_u64(val4, &self.field_type);
+    pub(crate) fn collect_block(&mut self, docs: &[DocId], field: &Column<u64>) {
+        // TODO special case for Required, Optional column type
+        for doc in docs {
+            for val in field.values(*doc) {
+                let val1 = f64_from_fastfield_u64(val, &self.field_type);
+                self.stats.collect(val1);
+            }
+        }
+    }
+}
+
+impl SegmentAggregationCollector for SegmentStatsCollector {
+    fn into_intermediate_aggregations_result(
+        self: Box<Self>,
+        agg_with_accessor: &AggregationsWithAccessor,
+    ) -> crate::Result<IntermediateAggregationResults> {
+        let name = agg_with_accessor.metrics.keys[0].to_string();
+
+        let intermediate_metric_result = match self.collecting_for {
+            SegmentStatsType::Average => {
+                IntermediateMetricResult::Average(IntermediateAverage::from_collector(*self))
+            }
+            SegmentStatsType::Count => {
+                IntermediateMetricResult::Count(IntermediateCount::from_collector(*self))
+            }
+            SegmentStatsType::Max => {
+                IntermediateMetricResult::Max(IntermediateMax::from_collector(*self))
+            }
+            SegmentStatsType::Min => {
+                IntermediateMetricResult::Min(IntermediateMin::from_collector(*self))
+            }
+            SegmentStatsType::Stats => IntermediateMetricResult::Stats(self.stats),
+            SegmentStatsType::Sum => {
+                IntermediateMetricResult::Sum(IntermediateSum::from_collector(*self))
+            }
+        };
+
+        let metrics = Some(VecWithNames::from_entries(vec![(
+            name,
+            intermediate_metric_result,
+        )]));
+
+        Ok(IntermediateAggregationResults {
+            metrics,
+            buckets: None,
+        })
+    }
+
+    fn collect(
+        &mut self,
+        doc: crate::DocId,
+        agg_with_accessor: &AggregationsWithAccessor,
+    ) -> crate::Result<()> {
+        let accessor = &agg_with_accessor.metrics.values[0].accessor;
+        for val in accessor.values(doc) {
+            let val1 = f64_from_fastfield_u64(val, &self.field_type);
             self.stats.collect(val1);
-            self.stats.collect(val2);
-            self.stats.collect(val3);
-            self.stats.collect(val4);
-        }
-        for &doc in iter.remainder() {
-            let val = field.get_val(doc);
-            let val = f64_from_fastfield_u64(val, &self.field_type);
-            self.stats.collect(val);
         }
+
+        Ok(())
+    }
+
+    fn flush_staged_docs(
+        &mut self,
+        _agg_with_accessor: &AggregationsWithAccessor,
+        _force_flush: bool,
+    ) -> crate::Result<()> {
+        Ok(())
     }
 }
 

src/aggregation/mod.rs

@@ -172,8 +172,8 @@ pub use collector::{
     AggregationCollector, AggregationSegmentCollector, DistributedAggregationCollector,
     MAX_BUCKET_COUNT,
 };
+use columnar::MonotonicallyMappableToU64;
 pub(crate) use date::format_date;
-use fastfield_codecs::MonotonicallyMappableToU64;
 use itertools::Itertools;
 use serde::{Deserialize, Serialize};
 
@@ -182,7 +182,7 @@ use crate::schema::Type;
 /// Represents an associative array `(key => values)` in a very efficient manner.
 #[derive(Clone, PartialEq, Serialize, Deserialize)]
 pub(crate) struct VecWithNames<T: Clone> {
-    values: Vec<T>,
+    pub(crate) values: Vec<T>,
     keys: Vec<String>,
 }
 
@@ -248,9 +248,6 @@ impl<T: Clone> VecWithNames<T> {
     fn values_mut(&mut self) -> impl Iterator<Item = &mut T> + '_ {
         self.values.iter_mut()
     }
-    fn entries(&self) -> impl Iterator<Item = (&str, &T)> + '_ {
-        self.keys().zip(self.values.iter())
-    }
     fn is_empty(&self) -> bool {
         self.keys.is_empty()
     }
@@ -336,8 +333,9 @@ mod tests {
     use crate::aggregation::intermediate_agg_result::IntermediateAggregationResults;
     use crate::aggregation::segment_agg_result::DOC_BLOCK_SIZE;
     use crate::aggregation::DistributedAggregationCollector;
+    use crate::indexer::NoMergePolicy;
     use crate::query::{AllQuery, TermQuery};
-    use crate::schema::{Cardinality, IndexRecordOption, Schema, TextFieldIndexing, FAST, STRING};
+    use crate::schema::{IndexRecordOption, Schema, TextFieldIndexing, FAST, STRING};
     use crate::{DateTime, Index, Term};
 
     fn get_avg_req(field_name: &str) -> Aggregation {
@@ -432,8 +430,7 @@ mod tests {
         let text_field = schema_builder.add_text_field("text", text_fieldtype.clone());
         let text_field_id = schema_builder.add_text_field("text_id", text_fieldtype);
         let string_field_id = schema_builder.add_text_field("string_id", STRING | FAST);
-        let score_fieldtype =
-            crate::schema::NumericOptions::default().set_fast();
+        let score_fieldtype = crate::schema::NumericOptions::default().set_fast();
         let score_field = schema_builder.add_u64_field("score", score_fieldtype.clone());
         let score_field_f64 = schema_builder.add_f64_field("score_f64", score_fieldtype.clone());
         let score_field_i64 = schema_builder.add_i64_field("score_i64", score_fieldtype);
@@ -445,6 +442,7 @@ mod tests {
         {
             // let mut index_writer = index.writer_for_tests()?;
             let mut index_writer = index.writer_with_num_threads(1, 30_000_000)?;
+            index_writer.set_merge_policy(Box::new(NoMergePolicy));
             for values in segment_and_values {
                 for (i, term) in values {
                     let i = *i;
@@ -656,13 +654,11 @@ mod tests {
         let text_field = schema_builder.add_text_field("text", text_fieldtype);
         let date_field = schema_builder.add_date_field("date", FAST);
         schema_builder.add_text_field("dummy_text", STRING);
-        let score_fieldtype =
-            crate::schema::NumericOptions::default().set_fast();
+        let score_fieldtype = crate::schema::NumericOptions::default().set_fast();
         let score_field = schema_builder.add_u64_field("score", score_fieldtype.clone());
         let score_field_f64 = schema_builder.add_f64_field("score_f64", score_fieldtype.clone());
 
-        let multivalue =
-            crate::schema::NumericOptions::default().set_fast();
+        let multivalue = crate::schema::NumericOptions::default().set_fast();
         let scores_field_i64 = schema_builder.add_i64_field("scores_i64", multivalue);
 
         let score_field_i64 = schema_builder.add_i64_field("score_i64", score_fieldtype);
@@ -1147,7 +1143,7 @@ mod tests {
         let agg_res = avg_on_field("dummy_text");
         assert_eq!(
             format!("{:?}", agg_res),
-            r#"InvalidArgument("Only fast fields of type f64, u64, i64 are supported, but got Str ")"#
+            r#"InvalidArgument("No numerical fast field found for field: dummy_text")"#
         );
 
         let agg_res = avg_on_field("not_exist_field");
@@ -1156,12 +1152,6 @@ mod tests {
             r#"FieldNotFound("not_exist_field")"#
         );
 
-        let agg_res = avg_on_field("scores_i64");
-        assert_eq!(
-            format!("{:?}", agg_res),
-            r#"InvalidArgument("Invalid field cardinality on field scores_i64 expected SingleValue, but got MultiValues")"#
-        );
-
         Ok(())
     }
 
@@ -1173,11 +1163,14 @@ mod tests {
         use test::{self, Bencher};
 
         use super::*;
-        use crate::aggregation::bucket::{HistogramAggregation, HistogramBounds, TermsAggregation};
+        use crate::aggregation::bucket::{
+            CustomOrder, HistogramAggregation, HistogramBounds, Order, OrderTarget,
+            TermsAggregation,
+        };
         use crate::aggregation::metric::StatsAggregation;
         use crate::query::AllQuery;
 
-        fn get_test_index_bench(merge_segments: bool) -> crate::Result<Index> {
+        fn get_test_index_bench(_merge_segments: bool) -> crate::Result<Index> {
             let mut schema_builder = Schema::builder();
             let text_fieldtype = crate::schema::TextOptions::default()
                 .set_indexing_options(
@@ -1189,20 +1182,19 @@ mod tests {
                 schema_builder.add_text_field("text_many_terms", STRING | FAST);
             let text_field_few_terms =
                 schema_builder.add_text_field("text_few_terms", STRING | FAST);
-            let score_fieldtype =
-                crate::schema::NumericOptions::default().set_fast();
+            let score_fieldtype = crate::schema::NumericOptions::default().set_fast();
             let score_field = schema_builder.add_u64_field("score", score_fieldtype.clone());
             let score_field_f64 =
                 schema_builder.add_f64_field("score_f64", score_fieldtype.clone());
             let score_field_i64 = schema_builder.add_i64_field("score_i64", score_fieldtype);
             let index = Index::create_from_tempdir(schema_builder.build())?;
             let few_terms_data = vec!["INFO", "ERROR", "WARN", "DEBUG"];
-            let many_terms_data = (0..15_000)
+            let many_terms_data = (0..150_000)
                 .map(|num| format!("author{}", num))
                 .collect::<Vec<_>>();
             {
                 let mut rng = thread_rng();
-                let mut index_writer = index.writer_for_tests()?;
+                let mut index_writer = index.writer_with_num_threads(1, 100_000_000)?;
                 // writing the segment
                 for _ in 0..1_000_000 {
                     let val: f64 = rng.gen_range(0.0..1_000_000.0);
@@ -1217,14 +1209,6 @@ mod tests {
                 }
                 index_writer.commit()?;
             }
-            if merge_segments {
-                let segment_ids = index
-                    .searchable_segment_ids()
-                    .expect("Searchable segments failed.");
-                let mut index_writer = index.writer_for_tests()?;
-                index_writer.merge(&segment_ids).wait()?;
-                index_writer.wait_merging_threads()?;
-            }
 
             Ok(index)
         }
@@ -1376,7 +1360,42 @@ mod tests {
         }
 
         #[bench]
-        fn bench_aggregation_terms_many(b: &mut Bencher) {
+        fn bench_aggregation_terms_many_with_sub_agg(b: &mut Bencher) {
+            let index = get_test_index_bench(false).unwrap();
+            let reader = index.reader().unwrap();
+
+            b.iter(|| {
+                let sub_agg_req: Aggregations = vec![(
+                    "average_f64".to_string(),
+                    Aggregation::Metric(MetricAggregation::Average(
+                        AverageAggregation::from_field_name("score_f64".to_string()),
+                    )),
+                )]
+                .into_iter()
+                .collect();
+
+                let agg_req: Aggregations = vec![(
+                    "my_texts".to_string(),
+                    Aggregation::Bucket(BucketAggregation {
+                        bucket_agg: BucketAggregationType::Terms(TermsAggregation {
+                            field: "text_many_terms".to_string(),
+                            ..Default::default()
+                        }),
+                        sub_aggregation: sub_agg_req,
+                    }),
+                )]
+                .into_iter()
+                .collect();
+
+                let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
+
+                let searcher = reader.searcher();
+                searcher.search(&AllQuery, &collector).unwrap()
+            });
+        }
+
+        #[bench]
+        fn bench_aggregation_terms_many2(b: &mut Bencher) {
             let index = get_test_index_bench(false).unwrap();
             let reader = index.reader().unwrap();
 
@@ -1401,6 +1420,36 @@ mod tests {
             });
         }
 
+        #[bench]
+        fn bench_aggregation_terms_many_order_by_term(b: &mut Bencher) {
+            let index = get_test_index_bench(false).unwrap();
+            let reader = index.reader().unwrap();
+
+            b.iter(|| {
+                let agg_req: Aggregations = vec![(
+                    "my_texts".to_string(),
+                    Aggregation::Bucket(BucketAggregation {
+                        bucket_agg: BucketAggregationType::Terms(TermsAggregation {
+                            field: "text_many_terms".to_string(),
+                            order: Some(CustomOrder {
+                                order: Order::Desc,
+                                target: OrderTarget::Key,
+                            }),
+                            ..Default::default()
+                        }),
+                        sub_aggregation: Default::default(),
+                    }),
+                )]
+                .into_iter()
+                .collect();
+
+                let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
+
+                let searcher = reader.searcher();
+                searcher.search(&AllQuery, &collector).unwrap()
+            });
+        }
+
         #[bench]
         fn bench_aggregation_range_only(b: &mut Bencher) {
             let index = get_test_index_bench(false).unwrap();

src/aggregation/segment_agg_result.rs

@@ -25,15 +25,89 @@ use crate::{DocId, TantivyError};
 pub(crate) const DOC_BLOCK_SIZE: usize = 64;
 pub(crate) type DocBlock = [DocId; DOC_BLOCK_SIZE];
 
-#[derive(Clone, PartialEq)]
-pub(crate) struct SegmentAggregationResultsCollector {
+pub(crate) trait SegmentAggregationCollector: CollectorClone + Debug {
+    fn into_intermediate_aggregations_result(
+        self: Box<Self>,
+        agg_with_accessor: &AggregationsWithAccessor,
+    ) -> crate::Result<IntermediateAggregationResults>;
+
+    fn collect(
+        &mut self,
+        doc: crate::DocId,
+        agg_with_accessor: &AggregationsWithAccessor,
+    ) -> crate::Result<()>;
+
+    fn flush_staged_docs(
+        &mut self,
+        agg_with_accessor: &AggregationsWithAccessor,
+        force_flush: bool,
+    ) -> crate::Result<()>;
+}
+
+pub(crate) trait CollectorClone {
+    fn clone_box(&self) -> Box<dyn SegmentAggregationCollector>;
+}
+
+impl<T> CollectorClone for T
+where T: 'static + SegmentAggregationCollector + Clone
+{
+    fn clone_box(&self) -> Box<dyn SegmentAggregationCollector> {
+        Box::new(self.clone())
+    }
+}
+
+impl Clone for Box<dyn SegmentAggregationCollector> {
+    fn clone(&self) -> Box<dyn SegmentAggregationCollector> {
+        self.clone_box()
+    }
+}
+
+pub(crate) fn build_segment_agg_collector(
+    req: &AggregationsWithAccessor,
+) -> crate::Result<Box<dyn SegmentAggregationCollector>> {
+    // Single metric special case
+    if req.buckets.is_empty() && req.metrics.len() == 1 {
+        let req = &req.metrics.values[0];
+        let stats_collector = match &req.metric {
+            MetricAggregation::Average(AverageAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Average)
+            }
+            MetricAggregation::Count(CountAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Count)
+            }
+            MetricAggregation::Max(MaxAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Max)
+            }
+            MetricAggregation::Min(MinAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Min)
+            }
+            MetricAggregation::Stats(StatsAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Stats)
+            }
+            MetricAggregation::Sum(SumAggregation { .. }) => {
+                SegmentStatsCollector::from_req(req.field_type, SegmentStatsType::Sum)
+            }
+        };
+
+        return Ok(Box::new(stats_collector));
+    }
+
+    let agg = GenericSegmentAggregationResultsCollector::from_req_and_validate(req)?;
+    Ok(Box::new(agg))
+}
+
+#[derive(Clone)]
+/// The GenericSegmentAggregationResultsCollector is the generic version of the collector, which
+/// can handle arbitrary complexity of  sub-aggregations. Ideally we never have to pick this one
+/// and can provide specialized versions instead, that remove some of its overhead.
+pub(crate) struct GenericSegmentAggregationResultsCollector {
     pub(crate) metrics: Option<VecWithNames<SegmentMetricResultCollector>>,
     pub(crate) buckets: Option<VecWithNames<SegmentBucketResultCollector>>,
     staged_docs: DocBlock,
     num_staged_docs: usize,
 }
 
-impl Default for SegmentAggregationResultsCollector {
+impl Default for GenericSegmentAggregationResultsCollector {
     fn default() -> Self {
         Self {
             metrics: Default::default(),
@@ -44,7 +118,7 @@ impl Default for SegmentAggregationResultsCollector {
     }
 }
 
-impl Debug for SegmentAggregationResultsCollector {
+impl Debug for GenericSegmentAggregationResultsCollector {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("SegmentAggregationResultsCollector")
             .field("metrics", &self.metrics)
@@ -55,9 +129,9 @@ impl Debug for SegmentAggregationResultsCollector {
     }
 }
 
-impl SegmentAggregationResultsCollector {
-    pub fn into_intermediate_aggregations_result(
-        self,
+impl SegmentAggregationCollector for GenericSegmentAggregationResultsCollector {
+    fn into_intermediate_aggregations_result(
+        self: Box<Self>,
         agg_with_accessor: &AggregationsWithAccessor,
     ) -> crate::Result<IntermediateAggregationResults> {
         let buckets = if let Some(buckets) = self.buckets {
@@ -75,47 +149,7 @@ impl SegmentAggregationResultsCollector {
         Ok(IntermediateAggregationResults { metrics, buckets })
     }
 
-    pub(crate) fn from_req_and_validate(req: &AggregationsWithAccessor) -> crate::Result<Self> {
-        let buckets = req
-            .buckets
-            .entries()
-            .map(|(key, req)| {
-                Ok((
-                    key.to_string(),
-                    SegmentBucketResultCollector::from_req_and_validate(req)?,
-                ))
-            })
-            .collect::<crate::Result<Vec<(String, _)>>>()?;
-        let metrics = req
-            .metrics
-            .entries()
-            .map(|(key, req)| {
-                Ok((
-                    key.to_string(),
-                    SegmentMetricResultCollector::from_req_and_validate(req)?,
-                ))
-            })
-            .collect::<crate::Result<Vec<(String, _)>>>()?;
-        let metrics = if metrics.is_empty() {
-            None
-        } else {
-            Some(VecWithNames::from_entries(metrics))
-        };
-        let buckets = if buckets.is_empty() {
-            None
-        } else {
-            Some(VecWithNames::from_entries(buckets))
-        };
-        Ok(SegmentAggregationResultsCollector {
-            metrics,
-            buckets,
-            staged_docs: [0; DOC_BLOCK_SIZE],
-            num_staged_docs: 0,
-        })
-    }
-
-    #[inline]
-    pub(crate) fn collect(
+    fn collect(
         &mut self,
         doc: crate::DocId,
         agg_with_accessor: &AggregationsWithAccessor,
@@ -128,7 +162,7 @@ impl SegmentAggregationResultsCollector {
         Ok(())
     }
 
-    pub(crate) fn flush_staged_docs(
+    fn flush_staged_docs(
         &mut self,
         agg_with_accessor: &AggregationsWithAccessor,
         force_flush: bool,
@@ -162,6 +196,66 @@ impl SegmentAggregationResultsCollector {
     }
 }
 
+impl GenericSegmentAggregationResultsCollector {
+    pub fn into_intermediate_aggregations_result(
+        self,
+        agg_with_accessor: &AggregationsWithAccessor,
+    ) -> crate::Result<IntermediateAggregationResults> {
+        let buckets = if let Some(buckets) = self.buckets {
+            let entries = buckets
+                .into_iter()
+                .zip(agg_with_accessor.buckets.values())
+                .map(|((key, bucket), acc)| Ok((key, bucket.into_intermediate_bucket_result(acc)?)))
+                .collect::<crate::Result<Vec<(String, _)>>>()?;
+            Some(VecWithNames::from_entries(entries))
+        } else {
+            None
+        };
+        let metrics = self.metrics.map(VecWithNames::from_other);
+
+        Ok(IntermediateAggregationResults { metrics, buckets })
+    }
+
+    pub(crate) fn from_req_and_validate(req: &AggregationsWithAccessor) -> crate::Result<Self> {
+        let buckets = req
+            .buckets
+            .iter()
+            .map(|(key, req)| {
+                Ok((
+                    key.to_string(),
+                    SegmentBucketResultCollector::from_req_and_validate(req)?,
+                ))
+            })
+            .collect::<crate::Result<Vec<(String, _)>>>()?;
+        let metrics = req
+            .metrics
+            .iter()
+            .map(|(key, req)| {
+                Ok((
+                    key.to_string(),
+                    SegmentMetricResultCollector::from_req_and_validate(req)?,
+                ))
+            })
+            .collect::<crate::Result<Vec<(String, _)>>>()?;
+        let metrics = if metrics.is_empty() {
+            None
+        } else {
+            Some(VecWithNames::from_entries(metrics))
+        };
+        let buckets = if buckets.is_empty() {
+            None
+        } else {
+            Some(VecWithNames::from_entries(buckets))
+        };
+        Ok(GenericSegmentAggregationResultsCollector {
+            metrics,
+            buckets,
+            staged_docs: [0; DOC_BLOCK_SIZE],
+            num_staged_docs: 0,
+        })
+    }
+}
+
 #[derive(Clone, Debug, PartialEq)]
 pub(crate) enum SegmentMetricResultCollector {
     Stats(SegmentStatsCollector),
@@ -205,7 +299,7 @@ impl SegmentMetricResultCollector {
     pub(crate) fn collect_block(&mut self, doc: &[DocId], metric: &MetricAggregationWithAccessor) {
         match self {
             SegmentMetricResultCollector::Stats(stats_collector) => {
-                stats_collector.collect_block(doc, &*metric.accessor);
+                stats_collector.collect_block(doc, &metric.accessor);
             }
         }
     }
@@ -215,7 +309,7 @@ impl SegmentMetricResultCollector {
 /// segments.
 /// The typical structure of Map<Key, Bucket> is not suitable during collection for performance
 /// reasons.
-#[derive(Clone, Debug, PartialEq)]
+#[derive(Clone, Debug)]
 pub(crate) enum SegmentBucketResultCollector {
     Range(SegmentRangeCollector),
     Histogram(Box<SegmentHistogramCollector>),
@@ -243,14 +337,7 @@ impl SegmentBucketResultCollector {
     pub fn from_req_and_validate(req: &BucketAggregationWithAccessor) -> crate::Result<Self> {
         match &req.bucket_agg {
             BucketAggregationType::Terms(terms_req) => Ok(Self::Terms(Box::new(
-                SegmentTermCollector::from_req_and_validate(
-                    terms_req,
-                    &req.sub_aggregation,
-                    req.field_type,
-                    req.accessor
-                        .as_multi()
-                        .expect("unexpected fast field cardinality"),
-                )?,
+                SegmentTermCollector::from_req_and_validate(terms_req, &req.sub_aggregation)?,
             ))),
             BucketAggregationType::Range(range_req) => {
                 Ok(Self::Range(SegmentRangeCollector::from_req_and_validate(
@@ -265,9 +352,7 @@ impl SegmentBucketResultCollector {
                     histogram,
                     &req.sub_aggregation,
                     req.field_type,
-                    req.accessor
-                        .as_single()
-                        .expect("unexpected fast field cardinality"),
+                    &req.accessor,
                 )?,
             ))),
         }

src/collector/facet_collector.rs

@@ -150,7 +150,7 @@ fn facet_depth(facet_bytes: &[u8]) -> usize {
 ///     }
 ///
 ///     {
-///         let mut facet_collector = FacetCollector::for_field(facet);
+///         let mut facet_collector = FacetCollector::for_field("facet");
 ///         facet_collector.add_facet("/category/fiction");
 ///         let facet_counts = searcher.search(&AllQuery, &facet_collector)?;
 ///
@@ -829,7 +829,7 @@ mod bench {
         let reader = index.reader().unwrap();
         b.iter(|| {
             let searcher = reader.searcher();
-            let facet_collector = FacetCollector::for_field(facet_field);
+            let facet_collector = FacetCollector::for_field("facet");
             searcher.search(&AllQuery, &facet_collector).unwrap();
         });
     }

src/collector/filter_collector_wrapper.rs

@@ -120,7 +120,7 @@ where
 
         let fast_field_reader = segment_reader
             .fast_fields()
-            .typed_column_first_or_default(schema.get_field_name(self.field))?;
+            .column_first_or_default(schema.get_field_name(self.field))?;
 
         let segment_collector = self
             .collector

src/collector/tests.rs

@@ -1,6 +1,4 @@
-use std::sync::Arc;
-
-use columnar::ColumnValues;
+use columnar::{BytesColumn, Column};
 
 use super::*;
 use crate::collector::{Count, FilterCollector, TopDocs};
@@ -57,7 +55,7 @@ pub fn test_filter_collector() -> crate::Result<()> {
 
     assert_eq!(filtered_top_docs.len(), 0);
 
-    fn date_filter(value: columnar::DateTime) -> bool {
+    fn date_filter(value: DateTime) -> bool {
         (crate::DateTime::from(value).into_utc()
             - OffsetDateTime::parse("2019-04-09T00:00:00+00:00", &Rfc3339).unwrap())
         .whole_weeks()
@@ -160,7 +158,7 @@ pub struct FastFieldTestCollector {
 
 pub struct FastFieldSegmentCollector {
     vals: Vec<u64>,
-    reader: Arc<dyn columnar::ColumnValues>,
+    reader: Column,
 }
 
 impl FastFieldTestCollector {
@@ -203,8 +201,7 @@ impl SegmentCollector for FastFieldSegmentCollector {
     type Fruit = Vec<u64>;
 
     fn collect(&mut self, doc: DocId, _score: Score) {
-        let val = self.reader.get_val(doc);
-        self.vals.push(val);
+        self.vals.extend(self.reader.values(doc));
     }
 
     fn harvest(self) -> Vec<u64> {
@@ -212,62 +209,73 @@ impl SegmentCollector for FastFieldSegmentCollector {
     }
 }
 
-// /// Collects in order all of the fast field bytes for all of the
-// /// docs in the `DocSet`
-// ///
-// /// This collector is mainly useful for tests.
-// pub struct BytesFastFieldTestCollector {
-//     field: Field,
-// }
+/// Collects in order all of the fast field bytes for all of the
+/// docs in the `DocSet`
+///
+/// This collector is mainly useful for tests.
+/// It is very slow.
+pub struct BytesFastFieldTestCollector {
+    field: String,
+}
 
-// pub struct BytesFastFieldSegmentCollector {
-//     vals: Vec<u8>,
-//     reader: BytesFastFieldReader,
-// }
+pub struct BytesFastFieldSegmentCollector {
+    vals: Vec<u8>,
+    column_opt: Option<BytesColumn>,
+    buffer: Vec<u8>,
+}
 
-// impl BytesFastFieldTestCollector {
-//     pub fn for_field(field: Field) -> BytesFastFieldTestCollector {
-//         BytesFastFieldTestCollector { field }
-//     }
-// }
+impl BytesFastFieldTestCollector {
+    pub fn for_field(field: impl ToString) -> BytesFastFieldTestCollector {
+        BytesFastFieldTestCollector {
+            field: field.to_string(),
+        }
+    }
+}
 
-// impl Collector for BytesFastFieldTestCollector {
-//     type Fruit = Vec<u8>;
-//     type Child = BytesFastFieldSegmentCollector;
+impl Collector for BytesFastFieldTestCollector {
+    type Fruit = Vec<u8>;
+    type Child = BytesFastFieldSegmentCollector;
 
-//     fn for_segment(
-//         &self,
-//         _segment_local_id: u32,
-//         segment_reader: &SegmentReader,
-//     ) -> crate::Result<BytesFastFieldSegmentCollector> {
-//         let reader = segment_reader.fast_fields().bytes(self.field)?;
-//         Ok(BytesFastFieldSegmentCollector {
-//             vals: Vec::new(),
-//             reader,
-//         })
-//     }
+    fn for_segment(
+        &self,
+        _segment_local_id: u32,
+        segment_reader: &SegmentReader,
+    ) -> crate::Result<BytesFastFieldSegmentCollector> {
+        let column_opt = segment_reader.fast_fields().bytes(&self.field)?;
+        Ok(BytesFastFieldSegmentCollector {
+            vals: Vec::new(),
+            column_opt,
+            buffer: Vec::new(),
+        })
+    }
 
-//     fn requires_scoring(&self) -> bool {
-//         false
-//     }
+    fn requires_scoring(&self) -> bool {
+        false
+    }
 
-//     fn merge_fruits(&self, children: Vec<Vec<u8>>) -> crate::Result<Vec<u8>> {
-//         Ok(children.into_iter().flat_map(|c| c.into_iter()).collect())
-//     }
-// }
+    fn merge_fruits(&self, children: Vec<Vec<u8>>) -> crate::Result<Vec<u8>> {
+        Ok(children.into_iter().flat_map(|c| c.into_iter()).collect())
+    }
+}
 
-// impl SegmentCollector for BytesFastFieldSegmentCollector {
-//     type Fruit = Vec<u8>;
+impl SegmentCollector for BytesFastFieldSegmentCollector {
+    type Fruit = Vec<u8>;
 
-//     fn collect(&mut self, doc: u32, _score: Score) {
-//         let data = self.reader.get_bytes(doc);
-//         self.vals.extend(data);
-//     }
+    fn collect(&mut self, doc: DocId, _score: Score) {
+        if let Some(column) = self.column_opt.as_ref() {
+            for term_ord in column.term_ords(doc) {
+                let (vals, buffer) = (&mut self.vals, &mut self.buffer);
+                if column.ord_to_bytes(term_ord, buffer).unwrap() {
+                    vals.extend(&buffer[..]);
+                }
+            }
+        }
+    }
 
-//     fn harvest(self) -> <Self as SegmentCollector>::Fruit {
-//         self.vals
-//     }
-// }
+    fn harvest(self) -> <Self as SegmentCollector>::Fruit {
+        self.vals
+    }
+}
 
 fn make_test_searcher() -> crate::Result<Searcher> {
     let schema = Schema::builder().build();

src/collector/top_score_collector.rs

@@ -235,7 +235,6 @@ impl TopDocs {
     /// # use tantivy::query::{Query, QueryParser};
     /// use tantivy::Searcher;
     /// use tantivy::collector::TopDocs;
-    /// use tantivy::schema::Field;
     ///
     /// # fn main() -> tantivy::Result<()> {
     /// #   let mut schema_builder = Schema::builder();
@@ -252,7 +251,7 @@ impl TopDocs {
     /// #   index_writer.commit()?;
     /// #   let reader = index.reader()?;
     /// #   let query = QueryParser::for_index(&index, vec![title]).parse_query("diary")?;
-    /// #   let top_docs = docs_sorted_by_rating(&reader.searcher(), &query, rating)?;
+    /// #   let top_docs = docs_sorted_by_rating(&reader.searcher(), &query)?;
     /// #   assert_eq!(top_docs,
     /// #            vec![(97u64, DocAddress::new(0u32, 1)),
     /// #                 (80u64, DocAddress::new(0u32, 3))]);
@@ -262,8 +261,7 @@ impl TopDocs {
     /// /// collects the top 10 documents, order by the u64-`field`
     /// /// given in argument.
     /// fn docs_sorted_by_rating(searcher: &Searcher,
-    ///                          query: &dyn Query,
-    ///                          rating_field: Field)
+    ///                          query: &dyn Query)
     ///     -> tantivy::Result<Vec<(u64, DocAddress)>> {
     ///
     ///     // This is where we build our topdocs collector
@@ -271,7 +269,7 @@ impl TopDocs {
     ///     // Note the `rating_field` needs to be a FAST field here.
     ///     let top_books_by_rating = TopDocs
     ///                 ::with_limit(10)
-    ///                  .order_by_u64_field(rating_field);
+    ///                  .order_by_u64_field("rating");
     ///
     ///     // ... and here are our documents. Note this is a simple vec.
     ///     // The `u64` in the pair is the value of our fast field for
@@ -323,22 +321,21 @@ impl TopDocs {
     /// # use tantivy::query::{Query, AllQuery};
     /// use tantivy::Searcher;
     /// use tantivy::collector::TopDocs;
-    /// use tantivy::schema::Field;
     ///
     /// # fn main() -> tantivy::Result<()> {
     /// #   let mut schema_builder = Schema::builder();
     /// #   let title = schema_builder.add_text_field("company", TEXT);
-    /// #   let rating = schema_builder.add_i64_field("revenue", FAST);
+    /// #   let revenue = schema_builder.add_i64_field("revenue", FAST);
     /// #   let schema = schema_builder.build();
     /// #
     /// #   let index = Index::create_in_ram(schema);
     /// #   let mut index_writer = index.writer_with_num_threads(1, 10_000_000)?;
-    /// #   index_writer.add_document(doc!(title => "MadCow Inc.", rating => 92_000_000i64))?;
-    /// #   index_writer.add_document(doc!(title => "Zozo Cow KKK", rating => 119_000_000i64))?;
-    /// #   index_writer.add_document(doc!(title => "Declining Cow", rating => -63_000_000i64))?;
+    /// #   index_writer.add_document(doc!(title => "MadCow Inc.", revenue => 92_000_000i64))?;
+    /// #   index_writer.add_document(doc!(title => "Zozo Cow KKK", revenue => 119_000_000i64))?;
+    /// #   index_writer.add_document(doc!(title => "Declining Cow", revenue => -63_000_000i64))?;
     /// #   assert!(index_writer.commit().is_ok());
     /// #   let reader = index.reader()?;
-    /// #   let top_docs = docs_sorted_by_revenue(&reader.searcher(), &AllQuery, rating)?;
+    /// #   let top_docs = docs_sorted_by_revenue(&reader.searcher(), &AllQuery, "revenue")?;
     /// #   assert_eq!(top_docs,
     /// #            vec![(119_000_000i64, DocAddress::new(0, 1)),
     /// #                 (92_000_000i64, DocAddress::new(0, 0))]);
@@ -349,7 +346,7 @@ impl TopDocs {
     /// /// given in argument.
     /// fn docs_sorted_by_revenue(searcher: &Searcher,
     ///                          query: &dyn Query,
-    ///                          revenue_field: Field)
+    ///                          revenue_field: &str)
     ///     -> tantivy::Result<Vec<(i64, DocAddress)>> {
     ///
     ///     // This is where we build our topdocs collector
@@ -358,7 +355,7 @@ impl TopDocs {
     ///     // type `sort_by_field`. revenue_field here is a FAST i64 field.
     ///     let top_company_by_revenue = TopDocs
     ///                 ::with_limit(2)
-    ///                  .order_by_fast_field(revenue_field);
+    ///                  .order_by_fast_field("revenue");
     ///
     ///     // ... and here are our documents. Note this is a simple vec.
     ///     // The `i64` in the pair is the value of our fast field for
@@ -460,9 +457,10 @@ impl TopDocs {
     ///             // Typically, fast_fields.
     ///             //
     ///             // In our case, we will get a reader for the popularity
-    ///             // fast field.
+    ///             // fast field. For simplicity we read the first or default value in the fast
+    ///             // field.
     ///             let popularity_reader =
-    ///                 segment_reader.fast_fields().u64("popularity").unwrap();
+    ///                 segment_reader.fast_fields().u64("popularity").unwrap().first_or_default_col(0);
     ///
     ///             // We can now define our actual scoring function
     ///             move |doc: DocId, original_score: Score| {
@@ -569,9 +567,9 @@ impl TopDocs {
     ///             // Note that this is implemented by using a `(u64, u64)`
     ///             // as a score.
     ///             let popularity_reader =
-    ///                 segment_reader.fast_fields().u64("popularity").unwrap();
+    ///                 segment_reader.fast_fields().u64("popularity").unwrap().first_or_default_col(0);
     ///             let boosted_reader =
-    ///                 segment_reader.fast_fields().u64("boosted").unwrap();
+    ///                 segment_reader.fast_fields().u64("boosted").unwrap().first_or_default_col(0);
     ///
     ///             // We can now define our actual scoring function
     ///             move |doc: DocId| {

src/core/index.rs

@@ -905,12 +905,14 @@ mod tests {
         let field = schema.get_field("num_likes").unwrap();
         let index = Index::create(directory.clone(), schema, IndexSettings::default())?;
 
-        let mut writer = index.writer_with_num_threads(8, 24_000_000).unwrap();
-        for i in 0u64..8_000u64 {
-            writer.add_document(doc!(field => i))?;
+        let mut writer = index.writer_with_num_threads(1, 32_000_000).unwrap();
+        for _seg in 0..8 {
+            for i in 0u64..1_000u64 {
+                writer.add_document(doc!(field => i))?;
+            }
+            writer.commit()?;
         }
 
-        writer.commit()?;
         let mem_right_after_commit = directory.total_mem_usage();
 
         let reader = index

src/core/inverted_index_reader.rs

@@ -135,6 +135,8 @@ impl InvertedIndexReader {
         term_info: &TermInfo,
         option: IndexRecordOption,
     ) -> io::Result<SegmentPostings> {
+        let option = option.downgrade(self.record_option);
+
         let block_postings = self.read_block_postings_from_terminfo(term_info, option)?;
         let position_reader = {
             if option.has_positions() {

src/core/searcher.rs

@@ -249,7 +249,7 @@ impl SearcherInner {
         index: Index,
         segment_readers: Vec<SegmentReader>,
         generation: TrackedObject<SearcherGeneration>,
-        doc_store_cache_size: usize,
+        doc_store_cache_num_blocks: usize,
     ) -> io::Result<SearcherInner> {
         assert_eq!(
             &segment_readers
@@ -261,7 +261,7 @@ impl SearcherInner {
         );
         let store_readers: Vec<StoreReader> = segment_readers
             .iter()
-            .map(|segment_reader| segment_reader.get_store_reader(doc_store_cache_size))
+            .map(|segment_reader| segment_reader.get_store_reader(doc_store_cache_num_blocks))
             .collect::<io::Result<Vec<_>>>()?;
 
         Ok(SearcherInner {

src/core/segment_reader.rs

@@ -9,7 +9,7 @@ use crate::directory::{CompositeFile, FileSlice};
 use crate::error::DataCorruption;
 use crate::fastfield::{intersect_alive_bitsets, AliveBitSet, FacetReader, FastFieldReaders};
 use crate::fieldnorm::{FieldNormReader, FieldNormReaders};
-use crate::schema::{Field, FieldType, IndexRecordOption, Schema, Type};
+use crate::schema::{Field, IndexRecordOption, Schema, Type};
 use crate::space_usage::SegmentSpaceUsage;
 use crate::store::StoreReader;
 use crate::termdict::TermDictionary;
@@ -99,7 +99,7 @@ impl SegmentReader {
                 "`{field_name}` is not a facet field.`"
             )));
         }
-        let Some(facet_column) = self.fast_fields().str_column_opt(field_name)? else {
+        let Some(facet_column) = self.fast_fields().str(field_name)? else {
             panic!("Facet Field `{field_name}` is missing. This should not happen");
         };
         Ok(FacetReader::new(facet_column))
@@ -128,9 +128,12 @@ impl SegmentReader {
         &self.fieldnorm_readers
     }
 
-    /// Accessor to the segment's `StoreReader`.
-    pub fn get_store_reader(&self, cache_size: usize) -> io::Result<StoreReader> {
-        StoreReader::open(self.store_file.clone(), cache_size)
+    /// Accessor to the segment's [`StoreReader`](crate::store::StoreReader).
+    ///
+    /// `cache_num_blocks` sets the number of decompressed blocks to be cached in an LRU.
+    /// The size of blocks is configurable, this should be reflexted in the
+    pub fn get_store_reader(&self, cache_num_blocks: usize) -> io::Result<StoreReader> {
+        StoreReader::open(self.store_file.clone(), cache_num_blocks)
     }
 
     /// Open a new segment for reading.

src/directory/mmap_directory.rs

@@ -196,8 +196,21 @@ impl MmapDirectory {
                 directory_path,
             )));
         }
-        let canonical_path: PathBuf = directory_path.canonicalize().map_err(|io_err| {
-            OpenDirectoryError::wrap_io_error(io_err, PathBuf::from(directory_path))
+        #[allow(clippy::bind_instead_of_map)]
+        let canonical_path: PathBuf = directory_path.canonicalize().or_else(|io_err| {
+            let directory_path = directory_path.to_owned();
+
+            #[cfg(windows)]
+            {
+                // `canonicalize` returns "Incorrect function" (error code 1)
+                // for virtual drives (network drives, ramdisk, etc.).
+                if io_err.raw_os_error() == Some(1) && directory_path.exists() {
+                    // Should call `std::path::absolute` when it is stabilised.
+                    return Ok(directory_path);
+                }
+            }
+
+            Err(OpenDirectoryError::wrap_io_error(io_err, directory_path))
         })?;
         if !canonical_path.is_dir() {
             return Err(OpenDirectoryError::NotADirectory(PathBuf::from(
@@ -443,6 +456,16 @@ impl Directory for MmapDirectory {
         Ok(self.inner.watch(watch_callback))
     }
 
+    #[cfg(windows)]
+    fn sync_directory(&self) -> Result<(), io::Error> {
+        // On Windows, it is not necessary to fsync the parent directory to
+        // ensure that the directory entry containing the file has also reached
+        // disk, and calling sync_data on a handle to directory is a no-op on
+        // local disks, but will return an error on virtual drives.
+        Ok(())
+    }
+
+    #[cfg(not(windows))]
     fn sync_directory(&self) -> Result<(), io::Error> {
         let mut open_opts = OpenOptions::new();
 
@@ -450,19 +473,6 @@ impl Directory for MmapDirectory {
         // write must not be set, or it fails with EISDIR
         open_opts.read(true);
 
-        // On Windows, opening a directory requires FILE_FLAG_BACKUP_SEMANTICS
-        // and calling sync_all() only works if write access is requested.
-        #[cfg(windows)]
-        {
-            use std::os::windows::fs::OpenOptionsExt;
-
-            use winapi::um::winbase;
-
-            open_opts
-                .write(true)
-                .custom_flags(winbase::FILE_FLAG_BACKUP_SEMANTICS);
-        }
-
         let fd = open_opts.open(&self.inner.root_path)?;
         fd.sync_data()?;
         Ok(())

src/fastfield/alive_bitset.rs

@@ -49,11 +49,6 @@ impl AliveBitSet {
         Self::open(alive_bitset_bytes)
     }
 
-    pub(crate) fn from_bitset(bitset: &BitSet) -> AliveBitSet {
-        let readonly_bitset = ReadOnlyBitSet::from(bitset);
-        AliveBitSet::from(readonly_bitset)
-    }
-
     /// Opens an alive bitset given its file.
     pub fn open(bytes: OwnedBytes) -> AliveBitSet {
         let bitset = ReadOnlyBitSet::open(bytes);

src/fastfield/facet_reader.rs

@@ -54,6 +54,7 @@ impl FacetReader {
         self.facet_column.ords().values(doc)
     }
 
+    /// Accessor to the facet dictionary.
     pub fn facet_dict(&self) -> &columnar::Dictionary {
         self.facet_column.dictionary()
     }

src/fastfield/mod.rs

@@ -22,14 +22,13 @@
 use std::net::Ipv6Addr;
 
 pub use columnar::Column;
-use columnar::MonotonicallyMappableToU64;
 
 pub use self::alive_bitset::{intersect_alive_bitsets, write_alive_bitset, AliveBitSet};
 pub use self::error::{FastFieldNotAvailableError, Result};
 pub use self::facet_reader::FacetReader;
 pub use self::readers::FastFieldReaders;
 pub use self::writer::FastFieldsWriter;
-use crate::schema::{Type, Value};
+use crate::schema::Type;
 use crate::DateTime;
 
 mod alive_bitset;
@@ -103,43 +102,13 @@ impl FastValue for DateTime {
     }
 }
 
-impl columnar::MonotonicallyMappableToU64 for DateTime {
-    fn to_u64(self) -> u64 {
-        self.timestamp_micros.to_u64()
-    }
-
-    fn from_u64(val: u64) -> Self {
-        DateTime {
-            timestamp_micros: MonotonicallyMappableToU64::from_u64(val),
-        }
-    }
-}
-
-fn unexpected_value(expected: &str, actual: &Value) -> crate::TantivyError {
-    crate::TantivyError::SchemaError(format!(
-        "Expected a {:?} in fast field, but got {:?}",
-        expected, actual
-    ))
-}
-
-fn value_to_u64(value: &Value) -> crate::Result<u64> {
-    let value = match value {
-        Value::U64(val) => val.to_u64(),
-        Value::I64(val) => val.to_u64(),
-        Value::F64(val) => val.to_u64(),
-        Value::Bool(val) => val.to_u64(),
-        Value::Date(val) => val.to_u64(),
-        _ => return Err(unexpected_value("u64/i64/f64/bool/date", value)),
-    };
-    Ok(value)
-}
-
 #[cfg(test)]
 mod tests {
 
-    use std::ops::RangeInclusive;
+    use std::ops::{Range, RangeInclusive};
     use std::path::Path;
 
+    use columnar::{Column, MonotonicallyMappableToU64};
     use common::{HasLen, TerminatingWrite};
     use once_cell::sync::Lazy;
     use rand::prelude::SliceRandom;
@@ -147,7 +116,7 @@ mod tests {
     use rand::{Rng, SeedableRng};
 
     use super::*;
-    use crate::directory::{CompositeFile, Directory, RamDirectory, WritePtr};
+    use crate::directory::{Directory, RamDirectory, WritePtr};
     use crate::merge_policy::NoMergePolicy;
     use crate::schema::{
         Document, Facet, FacetOptions, Field, Schema, SchemaBuilder, FAST, INDEXED, STRING, TEXT,
@@ -191,7 +160,10 @@ mod tests {
 
         assert_eq!(file.len(), 161);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
-        let column = fast_field_readers.u64("field").unwrap();
+        let column = fast_field_readers
+            .u64("field")
+            .unwrap()
+            .first_or_default_col(0);
         assert_eq!(column.get_val(0), 13u64);
         assert_eq!(column.get_val(1), 14u64);
         assert_eq!(column.get_val(2), 2u64);
@@ -238,7 +210,10 @@ mod tests {
         let file = directory.open_read(path).unwrap();
         assert_eq!(file.len(), 189);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
-        let col = fast_field_readers.u64("field").unwrap();
+        let col = fast_field_readers
+            .u64("field")
+            .unwrap()
+            .first_or_default_col(0);
         assert_eq!(col.get_val(0), 4u64);
         assert_eq!(col.get_val(1), 14_082_001u64);
         assert_eq!(col.get_val(2), 3_052u64);
@@ -268,7 +243,10 @@ mod tests {
         let file = directory.open_read(path).unwrap();
         assert_eq!(file.len(), 162);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
-        let fast_field_reader = fast_field_readers.u64("field").unwrap();
+        let fast_field_reader = fast_field_readers
+            .u64("field")
+            .unwrap()
+            .first_or_default_col(0);
         for doc in 0..10_000 {
             assert_eq!(fast_field_reader.get_val(doc), 100_000u64);
         }
@@ -298,7 +276,10 @@ mod tests {
         assert_eq!(file.len(), 4557);
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
-            let col = fast_field_readers.u64("field").unwrap();
+            let col = fast_field_readers
+                .u64("field")
+                .unwrap()
+                .first_or_default_col(0);
             for doc in 1..10_000 {
                 assert_eq!(col.get_val(doc), 5_000_000_000_000_000_000u64 + doc as u64);
             }
@@ -329,7 +310,10 @@ mod tests {
 
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
-            let col = fast_field_readers.i64("field").unwrap();
+            let col = fast_field_readers
+                .i64("field")
+                .unwrap()
+                .first_or_default_col(0);
             assert_eq!(col.min_value(), -100i64);
             assert_eq!(col.max_value(), 9_999i64);
             for (doc, i) in (-100i64..10_000i64).enumerate() {
@@ -364,7 +348,18 @@ mod tests {
         let file = directory.open_read(path).unwrap();
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let col = fast_field_readers.i64("field").unwrap();
-        assert_eq!(col.get_val(0), 0i64);
+        assert_eq!(col.first(0), None);
+
+        let col = fast_field_readers
+            .i64("field")
+            .unwrap()
+            .first_or_default_col(0);
+        assert_eq!(col.get_val(0), 0);
+        let col = fast_field_readers
+            .i64("field")
+            .unwrap()
+            .first_or_default_col(-100);
+        assert_eq!(col.get_val(0), -100);
     }
 
     #[test]
@@ -385,8 +380,11 @@ mod tests {
 
         let file = directory.open_read(path).unwrap();
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
-        let col = fast_field_readers.date("date").unwrap();
-        assert_eq!(col.get_val(0), columnar::DateTime::default());
+        let col = fast_field_readers
+            .date("date")
+            .unwrap()
+            .first_or_default_col(DateTime::default());
+        assert_eq!(col.get_val(0), DateTime::default());
     }
 
     // Warning: this generates the same permutation at each call
@@ -418,7 +416,10 @@ mod tests {
         }
         let file = directory.open_read(path).unwrap();
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
-        let col = fast_field_readers.u64("field").unwrap();
+        let col = fast_field_readers
+            .u64("field")
+            .unwrap()
+            .first_or_default_col(0);
         for a in 0..n {
             assert_eq!(col.get_val(a as u32), permutation[a]);
         }
@@ -436,39 +437,38 @@ mod tests {
         test_intfastfield_permutation_with_data(permutation);
     }
 
-    // TODO reenable when merge is here.
-    // #[test]
-    // fn test_merge_missing_date_fast_field() {
-    //     let mut schema_builder = Schema::builder();
-    //     let date_field = schema_builder.add_date_field("date", FAST);
-    //     let schema = schema_builder.build();
-    //     let index = Index::create_in_ram(schema);
-    //     let mut index_writer = index.writer_for_tests().unwrap();
-    //     index_writer.set_merge_policy(Box::new(NoMergePolicy));
-    //     index_writer
-    //         .add_document(doc!(date_field =>DateTime::from_utc(OffsetDateTime::now_utc())))
-    //         .unwrap();
-    //     index_writer.commit().unwrap();
-    //     index_writer.add_document(doc!()).unwrap();
-    //     index_writer.commit().unwrap();
-    //     let reader = index.reader().unwrap();
-    //     let segment_ids: Vec<SegmentId> = reader
-    //         .searcher()
-    //         .segment_readers()
-    //         .iter()
-    //         .map(SegmentReader::segment_id)
-    //         .collect();
-    //     assert_eq!(segment_ids.len(), 2);
-    //     index_writer.merge(&segment_ids[..]).wait().unwrap();
-    //     reader.reload().unwrap();
-    //     assert_eq!(reader.searcher().segment_readers().len(), 1);
-    // }
+    #[test]
+    fn test_merge_missing_date_fast_field() {
+        let mut schema_builder = Schema::builder();
+        let date_field = schema_builder.add_date_field("date", FAST);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_for_tests().unwrap();
+        index_writer.set_merge_policy(Box::new(NoMergePolicy));
+        index_writer
+            .add_document(doc!(date_field => DateTime::from_utc(OffsetDateTime::now_utc())))
+            .unwrap();
+        index_writer.commit().unwrap();
+        index_writer.add_document(doc!()).unwrap();
+        index_writer.commit().unwrap();
+        let reader = index.reader().unwrap();
+        let segment_ids: Vec<SegmentId> = reader
+            .searcher()
+            .segment_readers()
+            .iter()
+            .map(SegmentReader::segment_id)
+            .collect();
+        assert_eq!(segment_ids.len(), 2);
+        index_writer.merge(&segment_ids[..]).wait().unwrap();
+        reader.reload().unwrap();
+        assert_eq!(reader.searcher().segment_readers().len(), 1);
+    }
 
-    // fn get_vals_for_docs(column: &columnar::Column<u64>, docs: Range<u32>) -> Vec<u64> {
-    //     docs.into_iter()
-    //         .flat_map(|doc| column.values(doc))
-    //         .collect()
-    // }
+    fn get_vals_for_docs(column: &Column<u64>, docs: Range<u32>) -> Vec<u64> {
+        docs.into_iter()
+            .flat_map(|doc| column.values(doc))
+            .collect()
+    }
 
     #[test]
     fn test_text_fastfield() {
@@ -524,57 +524,61 @@ mod tests {
             assert!(str_column.ord_to_str(0, &mut str_term).unwrap());
             assert_eq!("AAAAA", &str_term);
 
-            //     let inverted_index = segment_reader.inverted_index(text_field)?;
-            //     assert_eq!(inverted_index.terms().num_terms(), 3);
-            //     let mut bytes = vec![];
-            //     assert!(inverted_index.terms().ord_to_term(0, &mut bytes)?);
-            //     assert_eq!(bytes, "aaaaa".as_bytes());
-            // }
+            let inverted_index = segment_reader.inverted_index(text_field).unwrap();
+            assert_eq!(inverted_index.terms().num_terms(), 3);
+            let mut bytes = vec![];
+            assert!(inverted_index.terms().ord_to_term(0, &mut bytes).unwrap());
+            assert_eq!(bytes, "aaaaa".as_bytes());
+        }
 
-            // {
-            //     // second segment
-            //     let mut index_writer = index.writer_for_tests()?;
+        {
+            // second segment
+            let mut index_writer = index.writer_for_tests().unwrap();
 
-            //     index_writer.add_document(doc!(
-            //     text_field => "AAAAA", // term_ord 0
-            //     ))?;
+            index_writer
+                .add_document(doc!(
+                text_field => "AAAAA", // term_ord 0
+                ))
+                .unwrap();
 
-            //     index_writer.add_document(doc!(
-            //     text_field => "CCCCC AAAAA", // term_ord 1, after merge 2
-            //     ))?;
+            index_writer
+                .add_document(doc!(
+                text_field => "CCCCC AAAAA", // term_ord 1, after merge 2
+                ))
+                .unwrap();
 
-            //     index_writer.add_document(doc!())?;
-            //     index_writer.commit()?;
+            index_writer.add_document(doc!()).unwrap();
+            index_writer.commit().unwrap();
 
-            //     let reader = index.reader()?;
-            //     let searcher = reader.searcher();
-            //     assert_eq!(searcher.segment_readers().len(), 2);
-            //     let segment_reader = searcher.segment_reader(1);
-            //     let fast_fields = segment_reader.fast_fields();
-            //     let text_fast_field = fast_fields.u64s("text").unwrap();
+            let reader = index.reader().unwrap();
+            let searcher = reader.searcher();
+            assert_eq!(searcher.segment_readers().len(), 2);
+            let segment_reader = searcher.segment_reader(1);
+            let fast_fields = segment_reader.fast_fields();
+            let text_fast_field = fast_fields.str("text").unwrap().unwrap();
 
-            //     assert_eq!(get_vals_for_docs(&text_fast_field, 0..3), vec![0, 1, 0]);
+            assert_eq!(&get_vals_for_docs(&text_fast_field.ords(), 0..2), &[0, 1]);
         }
 
         // TODO uncomment once merging is available
         // Merging the segments
-        // {
-        // let segment_ids = index.searchable_segment_ids()?;
-        // let mut index_writer = index.writer_for_tests()?;
-        // index_writer.merge(&segment_ids).wait()?;
-        // index_writer.wait_merging_threads()?;
-        // }
-        //
-        // let reader = index.reader()?;
-        // let searcher = reader.searcher();
-        // let segment_reader = searcher.segment_reader(0);
-        // let fast_fields = segment_reader.fast_fields();
-        // let text_fast_field = fast_fields.u64s("text").unwrap();
-        //
-        // assert_eq!(
-        // get_vals_for_docs(&text_fast_field, 0..8),
-        // vec![1, 0, 0, 0, 1, 3 /* next segment */, 0, 2, 0]
-        // );
+        {
+            let segment_ids = index.searchable_segment_ids().unwrap();
+            let mut index_writer = index.writer_for_tests().unwrap();
+            index_writer.merge(&segment_ids).wait().unwrap();
+            index_writer.wait_merging_threads().unwrap();
+        }
+
+        let reader = index.reader().unwrap();
+        let searcher = reader.searcher();
+        let segment_reader = searcher.segment_reader(0);
+        let fast_fields = segment_reader.fast_fields();
+        let text_column = fast_fields.str("text").unwrap().unwrap();
+
+        assert_eq!(
+            get_vals_for_docs(text_column.ords(), 0..8),
+            vec![1, 0, 0, 0, 1, 3 /* next segment */, 0, 2]
+        );
     }
 
     #[test]
@@ -588,11 +592,7 @@ mod tests {
         writer.commit().unwrap();
         let searcher = index.reader().unwrap().searcher();
         let segment_reader = searcher.segment_reader(0);
-        let str_column = segment_reader
-            .fast_fields()
-            .str_column_opt("text")
-            .unwrap()
-            .unwrap();
+        let str_column = segment_reader.fast_fields().str("text").unwrap().unwrap();
         // The string values are not sorted here.
         let term_ords: Vec<u64> = str_column.term_ords(0u32).collect();
         assert_eq!(&term_ords, &[1, 0]);
@@ -618,155 +618,162 @@ mod tests {
         assert_eq!(&facet_ords, &[0, 1]);
     }
 
-    // #[test]
-    // fn test_string_fastfield() -> crate::Result<()> {
-    //     let mut schema_builder = Schema::builder();
-    //     let text_field = schema_builder.add_text_field("text", STRING | FAST);
-    //     let schema = schema_builder.build();
-    //     let index = Index::create_in_ram(schema);
+    #[test]
+    fn test_string_fastfield() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let text_field = schema_builder.add_text_field("text", STRING | FAST);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
 
-    //     {
-    //         // first segment
-    //         let mut index_writer = index.writer_for_tests()?;
-    //         index_writer.set_merge_policy(Box::new(NoMergePolicy));
-    //         index_writer.add_document(doc!(
-    //             text_field => "BBBBB", // term_ord 1
-    //         ))?;
-    //         index_writer.add_document(doc!())?;
-    //         index_writer.add_document(doc!(
-    //             text_field => "AAAAA", // term_ord 0
-    //         ))?;
-    //         index_writer.add_document(doc!(
-    //             text_field => "AAAAA", // term_ord 0
-    //         ))?;
-    //         index_writer.add_document(doc!(
-    //             text_field => "zumberthree", // term_ord 2, after merge term_ord 3
-    //         ))?;
+        {
+            // first segment
+            let mut index_writer = index.writer_for_tests()?;
+            index_writer.set_merge_policy(Box::new(NoMergePolicy));
+            index_writer.add_document(doc!(
+                text_field => "BBBBB", // term_ord 1
+            ))?;
+            index_writer.add_document(doc!())?;
+            index_writer.add_document(doc!(
+                text_field => "AAAAA", // term_ord 0
+            ))?;
+            index_writer.add_document(doc!(
+                text_field => "AAAAA", // term_ord 0
+            ))?;
+            index_writer.add_document(doc!(
+                text_field => "zumberthree", // term_ord 2, after merge term_ord 3
+            ))?;
 
-    //         index_writer.add_document(doc!())?;
-    //         index_writer.commit()?;
+            index_writer.add_document(doc!())?;
+            index_writer.commit()?;
 
-    //         let reader = index.reader()?;
-    //         let searcher = reader.searcher();
-    //         assert_eq!(searcher.segment_readers().len(), 1);
-    //         let segment_reader = searcher.segment_reader(0);
-    //         let fast_fields = segment_reader.fast_fields();
-    //         let text_fast_field = fast_fields.u64s(text_field).unwrap();
+            let reader = index.reader()?;
+            let searcher = reader.searcher();
+            assert_eq!(searcher.segment_readers().len(), 1);
+            let segment_reader = searcher.segment_reader(0);
+            let fast_fields = segment_reader.fast_fields();
+            let text_col = fast_fields.str("text").unwrap().unwrap();
 
-    //         assert_eq!(get_vals_for_docs(&text_fast_field, 0..6), vec![1, 0, 0, 2]);
+            assert_eq!(get_vals_for_docs(&text_col.ords(), 0..6), vec![1, 0, 0, 2]);
 
-    //         let inverted_index = segment_reader.inverted_index(text_field)?;
-    //         assert_eq!(inverted_index.terms().num_terms(), 3);
-    //         let mut bytes = vec![];
-    //         assert!(inverted_index.terms().ord_to_term(0, &mut bytes)?);
-    //         assert_eq!(bytes, "AAAAA".as_bytes());
-    //     }
+            let inverted_index = segment_reader.inverted_index(text_field)?;
+            assert_eq!(inverted_index.terms().num_terms(), 3);
+            let mut bytes = vec![];
+            assert!(inverted_index.terms().ord_to_term(0, &mut bytes)?);
+            assert_eq!(bytes, "AAAAA".as_bytes());
+        }
 
-    //     {
-    //         // second segment
-    //         let mut index_writer = index.writer_for_tests()?;
+        {
+            // second segment
+            let mut index_writer = index.writer_for_tests()?;
 
-    //         index_writer.add_document(doc!(
-    //             text_field => "AAAAA", // term_ord 0
-    //         ))?;
+            index_writer.add_document(doc!(
+                text_field => "AAAAA", // term_ord 0
+            ))?;
 
-    //         index_writer.add_document(doc!(
-    //             text_field => "CCCCC", // term_ord 1, after merge 2
-    //         ))?;
+            index_writer.add_document(doc!(
+                text_field => "CCCCC", // term_ord 1, after merge 2
+            ))?;
 
-    //         index_writer.add_document(doc!())?;
-    //         index_writer.commit()?;
+            index_writer.add_document(doc!())?;
+            index_writer.commit()?;
 
-    //         let reader = index.reader()?;
-    //         let searcher = reader.searcher();
-    //         assert_eq!(searcher.segment_readers().len(), 2);
-    //         let segment_reader = searcher.segment_reader(1);
-    //         let fast_fields = segment_reader.fast_fields();
-    //         let text_fast_field = fast_fields.u64s(text_field).unwrap();
+            let reader = index.reader()?;
+            let searcher = reader.searcher();
+            assert_eq!(searcher.segment_readers().len(), 2);
+            let segment_reader = searcher.segment_reader(1);
+            let fast_fields = segment_reader.fast_fields();
+            let text_fast_field = fast_fields.str("text").unwrap().unwrap();
 
-    //         assert_eq!(get_vals_for_docs(&text_fast_field, 0..2), vec![0, 1]);
-    //     }
-    //     // Merging the segments
-    //     {
-    //         let segment_ids = index.searchable_segment_ids()?;
-    //         let mut index_writer = index.writer_for_tests()?;
-    //         index_writer.merge(&segment_ids).wait()?;
-    //         index_writer.wait_merging_threads()?;
-    //     }
+            assert_eq!(&get_vals_for_docs(text_fast_field.ords(), 0..2), &[0, 1]);
+        }
+        // Merging the segments
+        {
+            let segment_ids = index.searchable_segment_ids()?;
+            let mut index_writer = index.writer_for_tests()?;
+            index_writer.merge(&segment_ids).wait()?;
+            index_writer.wait_merging_threads()?;
+        }
 
-    //     let reader = index.reader()?;
-    //     let searcher = reader.searcher();
-    //     let segment_reader = searcher.segment_reader(0);
-    //     let fast_fields = segment_reader.fast_fields();
-    //     let text_fast_field = fast_fields.u64s(text_field).unwrap();
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+        let segment_reader = searcher.segment_reader(0);
+        let fast_fields = segment_reader.fast_fields();
+        let text_fast_field = fast_fields.str("text").unwrap().unwrap();
 
-    //     assert_eq!(
-    //         get_vals_for_docs(&text_fast_field, 0..9),
-    //         vec![1, 0, 0, 3 /* next segment */, 0, 2]
-    //     );
+        assert_eq!(
+            get_vals_for_docs(&text_fast_field.ords(), 0..9),
+            vec![1, 0, 0, 3 /* next segment */, 0, 2]
+        );
 
-    //     Ok(())
-    // }
+        Ok(())
+    }
 
-    // #[test]
-    // fn test_datefastfield() -> crate::Result<()> {
-    //     let mut schema_builder = Schema::builder();
-    //     let date_field = schema_builder.add_date_field(
-    //         "date",
-    //         DateOptions::from(FAST).set_precision(DatePrecision::Microseconds),
-    //     );
-    //     let multi_date_field = schema_builder.add_date_field(
-    //         "multi_date",
-    //         DateOptions::default()
-    //             .set_precision(DatePrecision::Microseconds)
-    //             .set_fast(),
-    //     );
-    //     let schema = schema_builder.build();
-    //     let index = Index::create_in_ram(schema);
-    //     let mut index_writer = index.writer_for_tests()?;
-    //     index_writer.set_merge_policy(Box::new(NoMergePolicy));
-    //     index_writer.add_document(doc!(
-    //         date_field => DateTime::from_u64(1i64.to_u64()),
-    //         multi_date_field => DateTime::from_u64(2i64.to_u64()),
-    //         multi_date_field => DateTime::from_u64(3i64.to_u64())
-    //     ))?;
-    //     index_writer.add_document(doc!(
-    //         date_field => DateTime::from_u64(4i64.to_u64())
-    //     ))?;
-    //     index_writer.add_document(doc!(
-    //         multi_date_field => DateTime::from_u64(5i64.to_u64()),
-    //         multi_date_field => DateTime::from_u64(6i64.to_u64())
-    //     ))?;
-    //     index_writer.commit()?;
-    //     let reader = index.reader()?;
-    //     let searcher = reader.searcher();
-    //     assert_eq!(searcher.segment_readers().len(), 1);
-    //     let segment_reader = searcher.segment_reader(0);
-    //     let fast_fields = segment_reader.fast_fields();
-    //     let date_fast_field = fast_fields.date(date_field).unwrap();
-    //     let dates_fast_field = fast_fields.dates(multi_date_field).unwrap();
-    //     let mut dates = vec![];
-    //     {
-    //         assert_eq!(date_fast_field.get_val(0).into_timestamp_micros(), 1i64);
-    //         dates_fast_field.get_vals(0u32, &mut dates);
-    //         assert_eq!(dates.len(), 2);
-    //         assert_eq!(dates[0].into_timestamp_micros(), 2i64);
-    //         assert_eq!(dates[1].into_timestamp_micros(), 3i64);
-    //     }
-    //     {
-    //         assert_eq!(date_fast_field.get_val(1).into_timestamp_micros(), 4i64);
-    //         dates_fast_field.get_vals(1u32, &mut dates);
-    //         assert!(dates.is_empty());
-    //     }
-    //     {
-    //         assert_eq!(date_fast_field.get_val(2).into_timestamp_micros(), 0i64);
-    //         dates_fast_field.get_vals(2u32, &mut dates);
-    //         assert_eq!(dates.len(), 2);
-    //         assert_eq!(dates[0].into_timestamp_micros(), 5i64);
-    //         assert_eq!(dates[1].into_timestamp_micros(), 6i64);
-    //     }
-    //     Ok(())
-    // }
+    #[test]
+    fn test_datefastfield() -> crate::Result<()> {
+        let mut schema_builder = Schema::builder();
+        let date_field = schema_builder.add_date_field(
+            "date",
+            DateOptions::from(FAST).set_precision(DatePrecision::Microseconds),
+        );
+        let multi_date_field = schema_builder.add_date_field(
+            "multi_date",
+            DateOptions::default()
+                .set_precision(DatePrecision::Microseconds)
+                .set_fast(),
+        );
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        let mut index_writer = index.writer_for_tests()?;
+        index_writer.set_merge_policy(Box::new(NoMergePolicy));
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_u64(1i64.to_u64()),
+            multi_date_field => DateTime::from_u64(2i64.to_u64()),
+            multi_date_field => DateTime::from_u64(3i64.to_u64())
+        ))?;
+        index_writer.add_document(doc!(
+            date_field => DateTime::from_u64(4i64.to_u64())
+        ))?;
+        index_writer.add_document(doc!(
+            multi_date_field => DateTime::from_u64(5i64.to_u64()),
+            multi_date_field => DateTime::from_u64(6i64.to_u64())
+        ))?;
+        index_writer.commit()?;
+        let reader = index.reader()?;
+        let searcher = reader.searcher();
+        assert_eq!(searcher.segment_readers().len(), 1);
+        let segment_reader = searcher.segment_reader(0);
+        let fast_fields = segment_reader.fast_fields();
+        let date_fast_field = fast_fields
+            .column_opt::<DateTime>("date")
+            .unwrap()
+            .unwrap()
+            .first_or_default_col(Default::default());
+        let dates_fast_field = fast_fields
+            .column_opt::<DateTime>("multi_date")
+            .unwrap()
+            .unwrap();
+        let mut dates = vec![];
+        {
+            assert_eq!(date_fast_field.get_val(0).into_timestamp_micros(), 1i64);
+            dates_fast_field.fill_vals(0u32, &mut dates);
+            assert_eq!(dates.len(), 2);
+            assert_eq!(dates[0].into_timestamp_micros(), 2i64);
+            assert_eq!(dates[1].into_timestamp_micros(), 3i64);
+        }
+        {
+            assert_eq!(date_fast_field.get_val(1).into_timestamp_micros(), 4i64);
+            dates_fast_field.fill_vals(1u32, &mut dates);
+            assert!(dates.is_empty());
+        }
+        {
+            assert_eq!(date_fast_field.get_val(2).into_timestamp_micros(), 0i64);
+            dates_fast_field.fill_vals(2u32, &mut dates);
+            assert_eq!(dates.len(), 2);
+            assert_eq!(dates[0].into_timestamp_micros(), 5i64);
+            assert_eq!(dates[1].into_timestamp_micros(), 6i64);
+        }
+        Ok(())
+    }
 
     #[test]
     pub fn test_fastfield_bool_small() {
@@ -796,10 +803,10 @@ mod tests {
         assert_eq!(file.len(), 175);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let bool_col = fast_field_readers.bool("field_bool").unwrap();
-        assert_eq!(bool_col.get_val(0), true);
-        assert_eq!(bool_col.get_val(1), false);
-        assert_eq!(bool_col.get_val(2), true);
-        assert_eq!(bool_col.get_val(3), false);
+        assert_eq!(bool_col.first(0), Some(true));
+        assert_eq!(bool_col.first(1), Some(false));
+        assert_eq!(bool_col.first(2), Some(true));
+        assert_eq!(bool_col.first(3), Some(false));
     }
 
     #[test]
@@ -829,8 +836,8 @@ mod tests {
         let readers = FastFieldReaders::open(file).unwrap();
         let bool_col = readers.bool("field_bool").unwrap();
         for i in 0..25 {
-            assert_eq!(bool_col.get_val(i * 2), true);
-            assert_eq!(bool_col.get_val(i * 2 + 1), false);
+            assert_eq!(bool_col.first(i * 2), Some(true));
+            assert_eq!(bool_col.first(i * 2 + 1), Some(false));
         }
     }
 
@@ -853,7 +860,17 @@ mod tests {
         assert_eq!(file.len(), 177);
         let fastfield_readers = FastFieldReaders::open(file).unwrap();
         let col = fastfield_readers.bool("field_bool").unwrap();
+        assert_eq!(col.first(0), None);
+        let col = fastfield_readers
+            .bool("field_bool")
+            .unwrap()
+            .first_or_default_col(false);
         assert_eq!(col.get_val(0), false);
+        let col = fastfield_readers
+            .bool("field_bool")
+            .unwrap()
+            .first_or_default_col(true);
+        assert_eq!(col.get_val(0), true);
     }
 
     fn get_index(docs: &[crate::Document], schema: &Schema) -> crate::Result<RamDirectory> {
@@ -907,7 +924,7 @@ mod tests {
         let col = readers.date("field").unwrap();
 
         for (i, time) in times.iter().enumerate() {
-            let dt: crate::DateTime = col.get_val(i as u32).into();
+            let dt: DateTime = col.first(i as u32).unwrap().into();
             assert_eq!(dt, time.truncate(precision));
         }
         readers.column_num_bytes("field").unwrap()
@@ -943,13 +960,17 @@ mod tests {
         let reader = index.reader().unwrap();
         let searcher = reader.searcher();
         let segment = &searcher.segment_readers()[0];
-        let field = segment.fast_fields().u64("url_norm_hash").unwrap();
+        let field = segment
+            .fast_fields()
+            .u64("url_norm_hash")
+            .unwrap()
+            .first_or_default_col(0);
 
         let numbers = vec![100, 200, 300];
         let test_range = |range: RangeInclusive<u64>| {
             let expexted_count = numbers.iter().filter(|num| range.contains(num)).count();
             let mut vec = vec![];
-            field.get_docids_for_value_range(range, 0..u32::MAX, &mut vec);
+            field.get_row_ids_for_value_range(range, 0..u32::MAX, &mut vec);
             assert_eq!(vec.len(), expexted_count);
         };
         test_range(50..=50);
@@ -976,9 +997,8 @@ mod tests {
         index_writer.commit().unwrap();
         let searcher = index.reader().unwrap().searcher();
         let fastfields = searcher.segment_reader(0u32).fast_fields();
-        let column: columnar::Column<Ipv6Addr> =
-            fastfields.typed_column_opt("ip").unwrap().unwrap();
-        assert_eq!(column.num_rows(), 3);
+        let column: Column<Ipv6Addr> = fastfields.column_opt("ip").unwrap().unwrap();
+        assert_eq!(column.num_docs(), 3);
         assert_eq!(column.first(0), None);
         assert_eq!(column.first(1), Some(ip_addr));
         assert_eq!(column.first(2), None);
@@ -1014,13 +1034,17 @@ mod tests {
         let reader = index.reader().unwrap();
         let searcher = reader.searcher();
         let segment = &searcher.segment_readers()[0];
-        let field = segment.fast_fields().u64("url_norm_hash").unwrap();
+        let field = segment
+            .fast_fields()
+            .u64("url_norm_hash")
+            .unwrap()
+            .first_or_default_col(0);
 
         let numbers = vec![1000, 1001, 1003];
         let test_range = |range: RangeInclusive<u64>| {
             let expexted_count = numbers.iter().filter(|num| range.contains(num)).count();
             let mut vec = vec![];
-            field.get_docids_for_value_range(range, 0..u32::MAX, &mut vec);
+            field.get_row_ids_for_value_range(range, 0..u32::MAX, &mut vec);
             assert_eq!(vec.len(), expexted_count);
         };
         let test_range_variant = |start, stop| {