Fixes the behavior of the range doc set, especially on multivalued

cardinality.

We stop scanning from the beginning of the column which is a great
improvement.
This is done by introducing a stateful `SelectCursor`.

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 }
@@ -55,13 +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" }
-fastfield_codecs = 			{ version= "0.3", path="./fastfield_codecs", default-features = false }
+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]
@@ -108,7 +107,7 @@ unstable = [] # useful for benches.
 quickwit = ["sstable"]
 
 [workspace]
-members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
+members = ["query-grammar", "bitpacker", "common", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
 
 # Following the "fail" crate best practises, we isolate
 # tests that define specific behavior in fail check points

bitpacker/src/bitpacker.rs

@@ -19,7 +19,7 @@ impl BitPacker {
     }
 
     #[inline]
-    pub fn write<TWrite: io::Write>(
+    pub fn write<TWrite: io::Write + ?Sized>(
         &mut self,
         val: u64,
         num_bits: u8,
@@ -43,7 +43,7 @@ impl BitPacker {
         Ok(())
     }
 
-    pub fn flush<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn flush<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
         if self.mini_buffer_written > 0 {
             let num_bytes = (self.mini_buffer_written + 7) / 8;
             let bytes = self.mini_buffer.to_le_bytes();
@@ -54,7 +54,7 @@ impl BitPacker {
         Ok(())
     }
 
-    pub fn close<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn close<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
         self.flush(output)?;
         Ok(())
     }

columnar/Cargo.toml

@@ -5,24 +5,28 @@ 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"
+criterion = "0.4"
 
 [features]
 unstable = []
+
+[[bench]]
+name = "bench_index"
+harness = false

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.rs

@@ -1,311 +0,0 @@
-#![feature(test)]
-
-extern crate test;
-
-#[cfg(test)]
-mod tests {
-    use std::ops::RangeInclusive;
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-    use rand::prelude::*;
-    use tantivy_columnar::*;
-    use test::Bencher;
-
-    use super::*;
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    fn generate_random() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64)
-            .map(|el| el + random::<u16>() as u64)
-            .collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation_gcd() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
-        column: &[T],
-    ) -> Arc<dyn Column<T>> {
-        let mut buffer = Vec::new();
-        serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
-        open(OwnedBytes::new(buffer)).unwrap()
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = permutation[a as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = column.get_val(a as u32);
-            }
-            a
-        });
-    }
-
-    const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
-    const SINGLE_ITEM: u64 = 90;
-    const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
-    const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
-    fn get_data_50percent_item() -> Vec<u128> {
-        let mut rng = StdRng::from_seed([1u8; 32]);
-
-        let mut data = vec![];
-        for _ in 0..300_000 {
-            let val = rng.gen_range(1..=100);
-            data.push(val);
-        }
-        data.push(SINGLE_ITEM);
-
-        data.shuffle(&mut rng);
-        let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
-        data
-    }
-    fn get_u128_column_random() -> Arc<dyn Column<u128>> {
-        let permutation = generate_random();
-        let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
-        get_u128_column_from_data(&permutation)
-    }
-
-    fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
-        let mut out = vec![];
-        let iter_gen = || data.iter().cloned();
-        serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
-        let out = OwnedBytes::new(out);
-        open_u128::<u128>(out).unwrap()
-    }
-
-    // U64 RANGE START
-    #[bench]
-    fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                FIFTY_PERCENT_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                ONE_PERCENT_ITEM_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                SINGLE_ITEM_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
-            positions
-        });
-    }
-    // U64 RANGE END
-
-    // U128 RANGE START
-    #[bench]
-    fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                *FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                *SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
-            positions
-        });
-    }
-    // U128 RANGE END
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let mut a = 0u128;
-            for i in 0u64..column.num_vals() as u64 {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let n = column.num_vals();
-            let mut a = 0u128;
-            for i in (0..n / 5).map(|val| val * 5) {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += permutation[i as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0u32..n as u32 {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
-        let permutation = generate_permutation_gcd();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..n {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..permutation.len() {
-                a += permutation[i as usize] as u64;
-            }
-            a
-        });
-    }
-}

columnar/benches/bench_index.rs

@@ -0,0 +1,91 @@
+use std::ops::Range;
+
+use criterion::*;
+use rand::prelude::*;
+use tantivy_columnar::column_index::MultiValueIndex;
+use tantivy_columnar::RowId;
+
+const WINDOW: usize = 40;
+
+fn bench_multi_value_index_util(
+    len_range: Range<u32>,
+    num_rows: RowId,
+    select_value_ratio: f64,
+    b: &mut criterion::Bencher,
+) {
+    let mut start_index: Vec<RowId> = vec![0u32];
+    let mut cursor: u32 = 0u32;
+    let mut rng = StdRng::from_seed([16u8; 32]);
+    for i in 0..num_rows {
+        let num_vals = rng.gen_range(len_range.clone());
+        cursor += num_vals;
+        start_index.push(cursor);
+    }
+    let select_rows: Vec<RowId> = (0u32..cursor)
+        .filter(|i| rng.gen_bool(select_value_ratio))
+        .collect();
+    let mv_index = MultiValueIndex::for_test(&start_index);
+
+    // mv_index.select_batch_in_place(0, &mut select_rows[..]);
+    let mut buffer = Vec::new();
+    b.iter(|| {
+        let mut start_row = 0u32;
+        let mut len = 0;
+        for chunk in select_rows.chunks(WINDOW) {
+            buffer.clear();
+            buffer.extend_from_slice(chunk);
+            mv_index.select_batch_in_place(start_row, &mut buffer);
+            start_row = buffer.last().copied().unwrap();
+            len += buffer.len()
+        }
+        assert_eq!(len, 4303);
+        len
+    });
+}
+
+fn bench_multi_value_index_util2(
+    len_range: Range<u32>,
+    num_rows: RowId,
+    select_value_ratio: f64,
+    b: &mut criterion::Bencher,
+) {
+    let mut start_index: Vec<RowId> = vec![0u32];
+    let mut cursor: u32 = 0u32;
+    let mut rng = StdRng::from_seed([16u8; 32]);
+    for i in 0..num_rows {
+        let num_vals = rng.gen_range(len_range.clone());
+        cursor += num_vals;
+        start_index.push(cursor);
+    }
+    let select_rows: Vec<RowId> = (0u32..cursor)
+        .filter(|i| rng.gen_bool(select_value_ratio))
+        .collect();
+    let mv_index = MultiValueIndex::for_test(&start_index);
+
+    // mv_index.select_batch_in_place(0, &mut select_rows[..]);
+    let mut buffer = Vec::new();
+    b.iter(|| {
+        let mut mv_index_cursor = mv_index.select_cursor();
+        let mut len = 0;
+        for chunk in select_rows.chunks(WINDOW) {
+            buffer.clear();
+            buffer.extend_from_slice(chunk);
+            mv_index_cursor.select_batch_in_place(&mut buffer);
+            len += buffer.len();
+        }
+        assert_eq!(len, 4303);
+        len
+    });
+}
+
+fn select_benchmark(c: &mut criterion::Criterion) {
+    c.bench_function("bench_multi_value_index_10_100", |b| {
+        bench_multi_value_index_util(0..10, 100_000, 0.01f64, b)
+    });
+    c.bench_function("bench_multi_value_cursor_index_10_100", |b| {
+        bench_multi_value_index_util2(0..10, 100_000, 0.01f64, b)
+    });
+}
+
+criterion_group!(benches, select_benchmark);
+criterion_main!(benches);

columnar/benches/bench_u128.rs

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

columnar/benches/bench_u64.rs

@@ -0,0 +1,211 @@
+#![feature(test)]
+extern crate test;
+
+use std::ops::RangeInclusive;
+use std::sync::Arc;
+
+use rand::prelude::*;
+use tantivy_columnar::column_values::{serialize_and_load_u64_based_column_values, CodecType};
+use tantivy_columnar::*;
+use test::Bencher;
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+fn generate_random() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (0u64..100_000u64)
+        .map(|el| el + random::<u16>() as u64)
+        .collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+// Warning: this generates the same permutation at each call
+fn generate_permutation_gcd() -> Vec<u64> {
+    let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
+    permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
+    permutation
+}
+
+pub fn serialize_and_load(column: &[u64], codec_type: CodecType) -> Arc<dyn ColumnValues<u64>> {
+    serialize_and_load_u64_based_column_values(&column, &[codec_type])
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = permutation[a as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_jumpy_fflookup_bitpacked(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for _ in 0..n {
+            a = column.get_val(a as u32);
+        }
+        a
+    });
+}
+
+const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
+const SINGLE_ITEM: u64 = 90;
+const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
+const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
+fn get_data_50percent_item() -> Vec<u128> {
+    let mut rng = StdRng::from_seed([1u8; 32]);
+
+    let mut data = vec![];
+    for _ in 0..300_000 {
+        let val = rng.gen_range(1..=100);
+        data.push(val);
+    }
+    data.push(SINGLE_ITEM);
+
+    data.shuffle(&mut rng);
+    let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
+    data
+}
+
+// U64 RANGE START
+#[bench]
+fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_docids_for_value_range(
+            FIFTY_PERCENT_RANGE,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_docids_for_value_range(
+            ONE_PERCENT_ITEM_RANGE,
+            0..data.len() as u32,
+            &mut positions,
+        );
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_docids_for_value_range(SINGLE_ITEM_RANGE, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+
+#[bench]
+fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
+    let data = get_data_50percent_item();
+    let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&data, CodecType::Bitpacked);
+
+    b.iter(|| {
+        let mut positions = Vec::new();
+        column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
+        positions
+    });
+}
+// U64 RANGE END
+
+#[bench]
+fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += permutation[i as usize];
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0;
+        for i in (0..n / 7).map(|val| val * 7) {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    let column_ref = column.as_ref();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0u32..n as u32 {
+            a += column_ref.get_val(i);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
+    let permutation = generate_permutation_gcd();
+    let n = permutation.len();
+    let column: Arc<dyn ColumnValues<u64>> = serialize_and_load(&permutation, CodecType::Bitpacked);
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..n {
+            a += column.get_val(i as u32);
+        }
+        a
+    });
+}
+
+#[bench]
+fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
+    let permutation = generate_permutation();
+    b.iter(|| {
+        let mut a = 0u64;
+        for i in 0..permutation.len() {
+            a += permutation[i as usize] as u64;
+        }
+        a
+    });
+}

columnar/columnar-cli/Cargo.toml

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

columnar/columnar-cli/src/main.rs

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

columnar/src/TODO.md

@@ -1,25 +1,21 @@
 # zero to one
-* merges
-* full still needs a num_values
-* replug u128
-* add dictionary encoded stuff
-* fix multivalued
-* find a way to make columnar work with strict types
-* plug to tantivy
-    - indexing
-    - aggregations
-    - merge
-* replug facets
-* replug range queries
-+ mutlivaued 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
+* move datetime to quickwit common
+* switch to nanos
+* reintroduce the gcd map.
 
 # Perf and Size
+* remove alloc in `ord_to_term`
 * re-add ZSTD compression for dictionaries
 no systematic monotonic mapping
 consider removing multilinear
 f32?
 adhoc solution for bool?
-
 add metrics helper for aggregate. sum(row_id)
 review inline absence/presence
 improv perf of select using PDEP
@@ -29,7 +25,6 @@ Add alignment?
 Consider another codec to bridge the gap between few and 5k elements
 
 # Cleanup and rationalization
-remove the 6 bit limitation of columntype. use 4 + 4 bits instead.
 in benchmark, unify percent vs ratio, f32 vs f64.
 investigate if should have better errors? io::Error is overused at the moment.
 rename rank/select in unit tests
@@ -40,16 +35,13 @@ use the rank & select naming in unit tests branch.
 multi-linear -> blockwise
 linear codec -> simply a multiplication for the index column
 rename columnar to something more explicit, like column_dictionary or columnar_table
-remove old column from the fast field API.
-remove the Column traits alias.
 rename fastfield -> column
 document changes
 rationalization FastFieldValue, HasColumnType
-
+isolate u128_based and uniform naming
 
 # Other
 fix enhance column-cli
 
 # Santa claus
-
 autodetect datetime ipaddr, plug customizable tokenizer.

columnar/src/column/dictionary_encoded.rs

@@ -56,18 +56,28 @@ 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
     }
 }
 
 impl StrColumn {
+    pub(crate) fn wrap(bytes_column: BytesColumn) -> StrColumn {
+        StrColumn(bytes_column)
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.0.dictionary.as_ref()
+    }
+
     /// Fills the buffer
     pub fn ord_to_str(&self, term_ord: u64, output: &mut String) -> io::Result<bool> {
         unsafe {
             let buf = output.as_mut_vec();
-            self.0.dictionary.ord_to_term(term_ord, buf)?;
+            if !self.0.dictionary.ord_to_term(term_ord, buf)? {
+                return Ok(false);
+            }
             // TODO consider remove checks if it hurts performance.
             if std::str::from_utf8(buf.as_slice()).is_err() {
                 buf.clear();

columnar/src/column/mod.rs

@@ -2,27 +2,46 @@ mod dictionary_encoded;
 mod serialize;
 
 use std::fmt::Debug;
+use std::io::Write;
 use std::ops::Deref;
 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;
-use crate::column_values::ColumnValues;
-use crate::{Cardinality, RowId};
+use crate::column_values::monotonic_mapping::StrictlyMonotonicMappingToInternal;
+use crate::column_values::{monotonic_map_column, ColumnValues};
+use crate::{Cardinality, MonotonicallyMappableToU64, RowId};
 
 #[derive(Clone)]
-pub struct Column<T> {
+pub struct Column<T = u64> {
     pub idx: ColumnIndex,
     pub values: Arc<dyn ColumnValues<T>>,
 }
 
+impl<T: MonotonicallyMappableToU64> Column<T> {
+    pub fn to_u64_monotonic(self) -> Column<u64> {
+        let values = Arc::new(monotonic_map_column(
+            self.values,
+            StrictlyMonotonicMappingToInternal::<T>::new(),
+        ));
+        Column {
+            idx: self.idx,
+            values,
+        }
+    }
+}
+
 impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
+    pub fn get_cardinality(&self) -> Cardinality {
+        self.idx.get_cardinality()
+    }
+
     pub fn num_rows(&self) -> RowId {
         match &self.idx {
             ColumnIndex::Full => self.values.num_vals() as u32,
@@ -52,6 +71,15 @@ impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
             .map(|value_row_id: RowId| self.values.get_val(value_row_id))
     }
 
+    /// 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,
@@ -69,7 +97,7 @@ impl<T> Deref for Column<T> {
 }
 
 impl BinarySerializable for Cardinality {
-    fn serialize<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> std::io::Result<()> {
         self.to_code().serialize(writer)
     }
 
@@ -105,7 +133,7 @@ impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
         match &self.column.idx {
             ColumnIndex::Full => self.column.values.num_vals(),
             ColumnIndex::Optional(optional_idx) => optional_idx.num_rows(),
-            ColumnIndex::Multivalued(_) => todo!(),
+            ColumnIndex::Multivalued(multivalue_idx) => multivalue_idx.num_rows(),
         }
     }
 }

columnar/src/column/serialize.rs

@@ -1,4 +1,3 @@
-use std::fmt::Debug;
 use std::io;
 use std::io::Write;
 use std::sync::Arc;
@@ -9,43 +8,31 @@ use sstable::Dictionary;
 use crate::column::{BytesColumn, Column};
 use crate::column_index::{serialize_column_index, SerializableColumnIndex};
 use crate::column_values::serialize::serialize_column_values_u128;
-use crate::column_values::{
-    serialize_column_values, ColumnValues, FastFieldCodecType, MonotonicallyMappableToU128,
-    MonotonicallyMappableToU64,
-};
+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<
-    F: Fn() -> I,
-    I: Iterator<Item = T>,
-    T: MonotonicallyMappableToU128,
->(
+pub fn serialize_column_mappable_to_u128<T: MonotonicallyMappableToU128>(
     column_index: SerializableColumnIndex<'_>,
-    column_values: F,
-    num_vals: u32,
+    iterable: &dyn Iterable<T>,
     output: &mut impl Write,
 ) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_column_values_u128(
-        || column_values().map(|val| val.to_u128()),
-        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 ColumnValues<T>,
+    column_values: &impl Iterable<T>,
     output: &mut impl Write,
 ) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_column_values(
+    serialize_u64_based_column_values(
         column_values,
-        &[
-            FastFieldCodecType::Bitpacked,
-            FastFieldCodecType::BlockwiseLinear,
-        ],
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
         output,
     )?;
     output.write_all(&column_index_num_bytes.to_le_bytes())?;
@@ -62,7 +49,8 @@ pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::
     );
     let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
     let column_index = crate::column_index::open_column_index(column_index_data)?;
-    let column_values = crate::column_values::open_u64_mapped(column_values_data)?;
+    let column_values =
+        crate::column_values::u64_based::load_u64_based_column_values(column_values_data)?;
     Ok(Column {
         idx: column_index,
         values: column_values,
@@ -88,15 +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 +1,171 @@
+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) => {
+                if optional_index.contains(row_addr.row_id) {
+                    1u32
+                } else {
+                    0u32
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_addr.row_id).len() as u32
+            }
+        }
+    })
+}
+
+/// Transforms an iterator containing the number of vals per row (with `num_rows` elements)
+/// into a `start_offset` iterator starting at 0 and (with `num_rows + 1` element)
+fn integrate_num_vals(num_vals: impl Iterator<Item = u32>) -> impl Iterator<Item = RowId> {
+    iter::once(0u32).chain(num_vals.scan(0, |state, num_vals| {
+        *state += num_vals;
+        Some(*state)
+    }))
+}
+
+impl<'a> Iterable<u32> for ShuffledMultivaluedIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        let num_vals_per_row = iter_num_values(self.column_indexes, self.merge_order);
+        Box::new(integrate_num_vals(num_vals_per_row))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_index::OptionalIndex;
+    use crate::RowAddr;
+
+    #[test]
+    fn test_integrate_num_vals_empty() {
+        assert!(integrate_num_vals(iter::empty()).eq(iter::once(0)));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_one_el() {
+        assert!(integrate_num_vals(iter::once(10)).eq([0, 10].into_iter()));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_several() {
+        assert!(integrate_num_vals([3, 0, 10, 20].into_iter()).eq([0, 3, 3, 13, 33].into_iter()));
+    }
+
+    #[test]
+    fn test_merge_column_index_optional_shuffle() {
+        let optional_index: ColumnIndex = OptionalIndex::for_test(2, &[0]).into();
+        let column_indexes = vec![Some(optional_index), Some(ColumnIndex::Full)];
+        let row_addrs = vec![
+            RowAddr {
+                segment_ord: 0u32,
+                row_id: 1u32,
+            },
+            RowAddr {
+                segment_ord: 1u32,
+                row_id: 0u32,
+            },
+        ];
+        let shuffle_merge_order = ShuffleMergeOrder::for_test(&[2, 1], row_addrs);
+        let serializable_index = merge_column_index_shuffled(
+            &column_indexes[..],
+            Cardinality::Optional,
+            &shuffle_merge_order,
+        );
+        let SerializableColumnIndex::Optional { non_null_row_ids, num_rows } = serializable_index else { panic!() };
+        assert_eq!(num_rows, 2);
+        let non_null_rows: Vec<RowId> = non_null_row_ids.boxed_iter().collect();
+        assert_eq!(&non_null_rows, &[1]);
+    }
+}

columnar/src/column_index/merge/stacked.rs

@@ -0,0 +1,154 @@
+use std::iter;
+
+use crate::column_index::{SerializableColumnIndex, Set};
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, StackMergeOrder};
+
+/// Simple case:
+/// The new mapping just consists in stacking the different column indexes.
+///
+/// There are no sort nor deletes involved.
+pub fn merge_column_index_stacked<'a>(
+    columns: &'a [Option<ColumnIndex>],
+    cardinality_after_merge: Cardinality,
+    stack_merge_order: &'a StackMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => SerializableColumnIndex::Optional {
+            non_null_row_ids: Box::new(StackedOptionalIndex {
+                columns,
+                stack_merge_order,
+            }),
+            num_rows: stack_merge_order.num_rows(),
+        },
+        Cardinality::Multivalued => {
+            let stacked_multivalued_index = StackedMultivaluedIndex {
+                columns,
+                stack_merge_order,
+            };
+            SerializableColumnIndex::Multivalued(Box::new(stacked_multivalued_index))
+        }
+    }
+}
+
+struct StackedOptionalIndex<'a> {
+    columns: &'a [Option<ColumnIndex>],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+impl<'a> Iterable<RowId> for StackedOptionalIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(
+            self.columns
+                .iter()
+                .enumerate()
+                .flat_map(|(columnar_id, column_index_opt)| {
+                    let columnar_row_range = self.stack_merge_order.columnar_range(columnar_id);
+                    let rows_it: Box<dyn Iterator<Item = RowId>> = match column_index_opt {
+                        Some(ColumnIndex::Full) => Box::new(columnar_row_range),
+                        Some(ColumnIndex::Optional(optional_index)) => Box::new(
+                            optional_index
+                                .iter_rows()
+                                .map(move |row_id: RowId| columnar_row_range.start + row_id),
+                        ),
+                        Some(ColumnIndex::Multivalued(_)) => {
+                            panic!("No multivalued index is allowed when stacking column index");
+                        }
+                        None => Box::new(std::iter::empty()),
+                    };
+                    rows_it
+                }),
+        )
+    }
+}
+
+#[derive(Clone, Copy)]
+struct StackedMultivaluedIndex<'a> {
+    columns: &'a [Option<ColumnIndex>],
+    stack_merge_order: &'a StackMergeOrder,
+}
+
+fn convert_column_opt_to_multivalued_index<'a>(
+    column_index_opt: Option<&'a ColumnIndex>,
+    num_rows: RowId,
+) -> Box<dyn Iterator<Item = RowId> + 'a> {
+    match column_index_opt {
+        None => Box::new(iter::repeat(0u32).take(num_rows as usize + 1)),
+        Some(ColumnIndex::Full) => Box::new(0..num_rows + 1),
+        Some(ColumnIndex::Optional(optional_index)) => {
+            Box::new(
+                (0..num_rows)
+                    // TODO optimize
+                    .map(|row_id| optional_index.rank(row_id))
+                    .chain(std::iter::once(optional_index.num_non_nulls())),
+            )
+        }
+        Some(ColumnIndex::Multivalued(multivalued_index)) => {
+            multivalued_index.start_index_column.iter()
+        }
+    }
+}
+
+impl<'a> Iterable<RowId> for StackedMultivaluedIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + '_> {
+        let multivalued_indexes =
+            self.columns
+                .iter()
+                .map(Option::as_ref)
+                .enumerate()
+                .map(|(columnar_id, column_opt)| {
+                    let num_rows =
+                        self.stack_merge_order.columnar_range(columnar_id).len() as RowId;
+                    convert_column_opt_to_multivalued_index(column_opt, num_rows)
+                });
+        stack_multivalued_indexes(multivalued_indexes)
+    }
+}
+
+// Refactor me
+fn stack_multivalued_indexes<'a>(
+    mut multivalued_indexes: impl Iterator<Item = Box<dyn Iterator<Item = RowId> + 'a>> + 'a,
+) -> Box<dyn Iterator<Item = RowId> + 'a> {
+    let mut offset = 0;
+    let mut last_row_id = 0;
+    let mut current_it = multivalued_indexes.next();
+    Box::new(std::iter::from_fn(move || loop {
+        let Some(multivalued_index) = current_it.as_mut() else {
+            return None;
+        };
+        if let Some(row_id) = multivalued_index.next() {
+            last_row_id = offset + row_id;
+            return Some(last_row_id);
+        }
+        offset = last_row_id;
+        loop {
+            current_it = multivalued_indexes.next();
+            if current_it.as_mut()?.next().is_some() {
+                break;
+            }
+        }
+    }))
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::RowId;
+
+    fn it<'a>(row_ids: &'a [RowId]) -> Box<dyn Iterator<Item = RowId> + 'a> {
+        Box::new(row_ids.iter().copied())
+    }
+
+    #[test]
+    fn test_stack() {
+        let columns = [
+            it(&[0u32, 0u32]),
+            it(&[0u32, 1u32, 1u32, 4u32]),
+            it(&[0u32, 3u32, 5u32]),
+            it(&[0u32, 4u32]),
+        ]
+        .into_iter();
+        let start_offsets: Vec<RowId> = super::stack_multivalued_indexes(columns).collect();
+        assert_eq!(start_offsets, &[0, 0, 1, 1, 4, 7, 9, 13]);
+    }
+}

columnar/src/column_index/mod.rs

@@ -1,15 +1,98 @@
+mod merge;
 mod multivalued_index;
 mod optional_index;
 mod serialize;
 
 use std::ops::Range;
 
-pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
+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;
+pub use crate::column_index::multivalued_index::{MultiValueIndex, MultiValueIndexCursor};
+use crate::column_index::optional_index::OptionalIndexSelectCursor;
 use crate::{Cardinality, RowId};
 
+pub struct ColumnIndexSelectCursor {
+    last_rank: Option<RowId>,
+    cardinality_specific_impl: CardinalitySpecificSelectCursor,
+}
+
+impl From<CardinalitySpecificSelectCursor> for ColumnIndexSelectCursor {
+    fn from(cardinality_specific_impl: CardinalitySpecificSelectCursor) -> Self {
+        ColumnIndexSelectCursor {
+            last_rank: None,
+            cardinality_specific_impl,
+        }
+    }
+}
+
+enum CardinalitySpecificSelectCursor {
+    Full,
+    Optional(OptionalIndexSelectCursor),
+    Multivalued(MultiValueIndexCursor),
+}
+
+/// This cursor object point is to compute batches of `select` operations.
+///
+/// Regardless of cardinality, a column index can always be seen as a mapping
+/// from row_id -> start_value_row_id. By definition, it is increasing.
+/// If `left <= right, column_index[left] <= column_index[right]`.
+///
+/// The select operation then identifies, given a value row id, which row it
+/// belong to: it is the inverse mapping.
+///
+/// As a more formal definition, `select(rank)` is defined as the only `i` such that
+/// mapping[i] <= rank and mapping[i+1] < rank.
+/// Another way to define it is to say that it is the last i such that
+/// mapping[i] <= rank.
+/// Finally it can be defined as the number of `row_id` such that
+/// mapping[i] <= rank.
+///
+/// `select_batch_in_place` is a complex function that copmutes
+/// select operation in batches and in place.
+///
+/// For optimization reasons, it only supports supplying ever striclty increasing
+/// values of `rank_ids`, even cross calls.
+///
+/// It is also required from the caller, to only supply rank_ids lower than max(mapping).
+/// Within those condition, the returned `row_ids` are guaranteed to be unique.
+///
+/// # Panics
+///
+/// Panics if the supplied rank_ids are not increasing from one call to another.
+/// We only check that the `rank_ids` Vec is increasing in debug mode for
+/// performance reason.
+impl ColumnIndexSelectCursor {
+    /// Returns a list of
+    pub fn select_batch_in_place(&mut self, rank_ids: &mut Vec<RowId>) {
+        // `rank_ids` has to be sorted.
+        debug_assert!(rank_ids.windows(2).all(|window| window[0] < window[1]));
+        // Two consecutive calls must pass strictly increasing `rank_ids`.
+        let (Some(first_rank), Some(new_last_rank)) = (rank_ids.first().copied(), rank_ids.last().copied()) else {
+            // rank_ids is empty, there is nothing to do.
+            return;
+        };
+        if let Some(last_rank) = self.last_rank {
+            assert!(last_rank < first_rank);
+        }
+        self.last_rank = Some(new_last_rank);
+        match &mut self.cardinality_specific_impl {
+            CardinalitySpecificSelectCursor::Full => {
+                // No need to do anything:
+                // `value_idx` and `row_idx` are the same.
+            }
+            CardinalitySpecificSelectCursor::Optional(optional_index) => {
+                optional_index.select_batch_in_place(&mut rank_ids[..]);
+            }
+            CardinalitySpecificSelectCursor::Multivalued(multivalued_index) => {
+                // TODO important: avoid using 0u32, and restart from the beginning all of the time.
+                multivalued_index.select_batch_in_place(rank_ids)
+            }
+        }
+    }
+}
+
 #[derive(Clone)]
 pub enum ColumnIndex {
     Full,
@@ -19,6 +102,18 @@ 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 {
@@ -28,6 +123,17 @@ impl ColumnIndex {
         }
     }
 
+    /// Returns true if and only if there are at least one value associated to the row.
+    pub fn has_value(&self, row_id: RowId) -> bool {
+        match self {
+            ColumnIndex::Full => true,
+            ColumnIndex::Optional(optional_index) => optional_index.contains(row_id),
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_id).len() > 0
+            }
+        }
+    }
+
     pub fn value_row_ids(&self, row_id: RowId) -> Range<RowId> {
         match self {
             ColumnIndex::Full => row_id..row_id + 1,
@@ -42,18 +148,15 @@ impl ColumnIndex {
         }
     }
 
-    pub fn select_batch_in_place(&self, rank_ids: &mut Vec<RowId>) {
+    pub fn select_cursor(&self) -> ColumnIndexSelectCursor {
         match self {
-            ColumnIndex::Full => {
-                // No need to do anything:
-                // value_idx and row_idx are the same.
-            }
+            ColumnIndex::Full => CardinalitySpecificSelectCursor::Full.into(),
             ColumnIndex::Optional(optional_index) => {
-                optional_index.select_batch(&mut rank_ids[..]);
+                CardinalitySpecificSelectCursor::Optional(optional_index.select_cursor()).into()
             }
             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)
+                CardinalitySpecificSelectCursor::Multivalued(multivalued_index.select_cursor())
+                    .into()
             }
         }
     }

columnar/src/column_index/multivalued_index.rs

@@ -5,16 +5,18 @@ use std::sync::Arc;
 
 use common::OwnedBytes;
 
-use crate::column_values::{ColumnValues, FastFieldCodecType};
+use crate::column_values::u64_based::CodecType;
+use crate::column_values::ColumnValues;
+use crate::iterable::Iterable;
 use crate::RowId;
 
 pub fn serialize_multivalued_index(
-    multivalued_index: &dyn ColumnValues<RowId>,
+    multivalued_index: &dyn Iterable<RowId>,
     output: &mut impl Write,
 ) -> io::Result<()> {
-    crate::column_values::serialize_column_values(
-        &*multivalued_index,
-        &[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
+    crate::column_values::u64_based::serialize_u64_based_column_values(
+        multivalued_index,
+        &[CodecType::Bitpacked, CodecType::Linear],
         output,
     )?;
     Ok(())
@@ -22,7 +24,7 @@ pub fn serialize_multivalued_index(
 
 pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex> {
     let start_index_column: Arc<dyn ColumnValues<RowId>> =
-        crate::column_values::open_u64_mapped(bytes)?;
+        crate::column_values::u64_based::load_u64_based_column_values(bytes)?;
     Ok(MultiValueIndex { start_index_column })
 }
 
@@ -30,16 +32,17 @@ pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex>
 /// Index to resolve value range for given doc_id.
 /// Starts at 0.
 pub struct MultiValueIndex {
-    start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
-}
-
-impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {
-    fn from(start_index_column: Arc<dyn ColumnValues<RowId>>) -> Self {
-        MultiValueIndex { start_index_column }
-    }
+    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
 }
 
 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]
@@ -55,78 +58,302 @@ impl MultiValueIndex {
         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.
+    pub fn select_cursor(&self) -> MultiValueIndexCursor {
+        MultiValueIndexCursor {
+            multivalued_index: self.clone(),
+            row_cursor: 0u32,
+        }
+    }
+}
+
+pub struct MultiValueIndexCursor {
+    multivalued_index: MultiValueIndex,
+    row_cursor: RowId,
+}
+
+impl MultiValueIndexCursor {
+    /// See contract in `ColumnIndexSelectCursor`.
     ///
-    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
-    /// index.
+    /// Multi valued cardinality is special for two different
+    /// ranks `rank_left` and `rank_right`, we can end up with
+    /// the same `select(rank_left)` and `select(rank_right)`.
     ///
-    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
-    /// increasing positions.
-    ///
-    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
-    /// match a docid to its value position.
-    #[allow(clippy::bool_to_int_with_if)]
-    pub(crate) fn select_batch_in_place(&self, row_start: RowId, ranks: &mut Vec<u32>) {
+    /// For this reason, this function includes extra complexity
+    /// to prevent the cursor from emitting the same row_id.
+    /// - From a last call, by skipping ranks mapping to
+    /// the same row_id
+    /// - With the batch, by simply deduplicating the output.
+    pub fn select_batch_in_place(&mut self, ranks: &mut Vec<RowId>) {
         if ranks.is_empty() {
             return;
         }
-        let mut cur_doc = row_start;
-        let mut last_doc = None;
+        let mut row_cursor = self.row_cursor;
 
-        assert!(self.start_index_column.get_val(row_start) as u32 <= ranks[0]);
+        let mut write_cursor_id = usize::MAX;
+        let mut last_written_row_id = u32::MAX;
 
-        let mut write_doc_pos = 0;
-        for i in 0..ranks.len() {
-            let pos = ranks[i];
-            loop {
-                let end = self.start_index_column.get_val(cur_doc + 1) as u32;
-                if end > pos {
-                    ranks[write_doc_pos] = cur_doc;
-                    write_doc_pos += if last_doc == Some(cur_doc) { 0 } else { 1 };
-                    last_doc = Some(cur_doc);
-                    break;
-                }
-                cur_doc += 1;
+        // We skip all of the ranks that we already passed.
+        //
+        // It is possible in the case of multivalued, for a the first
+        // few rank to belong to the same row_id as the last rank
+        // of the previous call.
+        let start_bound = self
+            .multivalued_index
+            .start_index_column
+            .get_val(row_cursor);
+
+        let mut skip = 0;
+        while ranks[skip] < start_bound {
+            skip += 1;
+            if skip == ranks.len() {
+                ranks.clear();
+                return;
             }
         }
-        ranks.truncate(write_doc_pos);
+
+        for i in skip..ranks.len() {
+            let rank = ranks[i];
+            let row_id = loop {
+                // TODO See if we can find a way to introduce a function in
+                // ColumnValue to remove dynamic dispatch.
+                // This is tricky however... because it only applies to T=u32.
+                //
+                // TODO consider using exponential search.
+                let end = self
+                    .multivalued_index
+                    .start_index_column
+                    .get_val(row_cursor + 1) as u32;
+                if end > rank {
+                    break row_cursor;
+                }
+                row_cursor += 1;
+            };
+            // We remove duplicates in a branchless fashion: we only advance
+            // the write cursor when we are writing a value different from
+            // the last written value.
+            write_cursor_id =
+                write_cursor_id.wrapping_add(if row_id == last_written_row_id { 0 } else { 1 });
+            ranks[write_cursor_id] = row_id;
+            last_written_row_id = row_id;
+        }
+
+        self.row_cursor = row_cursor + 1;
+        ranks.truncate(write_cursor_id + 1);
     }
 }
 
 #[cfg(test)]
 mod tests {
-    use std::ops::Range;
     use std::sync::Arc;
 
     use super::MultiValueIndex;
     use crate::column_values::IterColumn;
     use crate::{ColumnValues, RowId};
+    use proptest::prelude::*;
 
-    fn index_to_pos_helper(
-        index: &MultiValueIndex,
-        doc_id_range: Range<u32>,
-        positions: &[u32],
-    ) -> Vec<u32> {
+    fn index_to_pos_helper(index: &MultiValueIndex, positions: &[u32]) -> Vec<u32> {
         let mut positions = positions.to_vec();
-        index.select_batch_in_place(doc_id_range.start, &mut positions);
+        let mut cursor = index.select_cursor();
+        cursor.select_batch_in_place(&mut positions);
         positions
     }
 
+    // Value row id ranges are [0..10, 10..12, 12..15, etc.]
+    const START_OFFSETS: &[RowId] = &[0, 10, 12, 15, 22, 23];
+
+    #[track_caller]
+    fn test_multivalue_select_cursor_aux(
+        start_offsets: &'static [RowId],
+        ranks: &[RowId],
+        expected: &[RowId],
+    ) {
+        let column: Arc<dyn ColumnValues<RowId>> =
+            Arc::new(IterColumn::from(start_offsets.iter().copied()));
+        let index = MultiValueIndex {
+            start_index_column: column,
+        };
+        assert_eq!(&index_to_pos_helper(&index, &ranks), expected);
+    }
+
     #[test]
-    fn test_positions_to_docid() {
-        let offsets: Vec<RowId> = vec![0, 10, 12, 15, 22, 23]; // docid values are [0..10, 10..12, 12..15, etc.]
-        let column: Arc<dyn ColumnValues<RowId>> = Arc::new(IterColumn::from(offsets.into_iter()));
-        let index = MultiValueIndex::from(column);
-        assert_eq!(index.num_rows(), 5);
-        let positions = &[10u32, 11, 15, 20, 21, 22];
-        assert_eq!(index_to_pos_helper(&index, 0..5, positions), vec![1, 3, 4]);
-        assert_eq!(index_to_pos_helper(&index, 1..5, positions), vec![1, 3, 4]);
-        assert_eq!(index_to_pos_helper(&index, 0..5, &[9]), vec![0]);
-        assert_eq!(index_to_pos_helper(&index, 1..5, &[10]), vec![1]);
-        assert_eq!(index_to_pos_helper(&index, 1..5, &[11]), vec![1]);
-        assert_eq!(index_to_pos_helper(&index, 2..5, &[12]), vec![2]);
-        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14]), vec![2]);
-        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14, 15]), vec![2, 3]);
+    fn test_multivalue_select_cursor_empty() {
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[], &[]);
+    }
+
+    #[test]
+    fn test_multivalue_select_cursor_single() {
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[9], &[0]);
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[10], &[1]);
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[11], &[1]);
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[11], &[1]);
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[12], &[2]);
+    }
+
+    #[test]
+    fn test_multivalue_select_cursor_duplicates() {
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[12, 14], &[2]);
+    }
+
+    #[test]
+    fn test_multivalue_select_cursor_complex() {
+        test_multivalue_select_cursor_aux(START_OFFSETS, &[10, 11, 15, 20, 21, 22], &[1, 3, 4])
+    }
+
+
+    #[test]
+    fn test_multivalue_select_corner_case_skip_all() {
+        let column: Arc<dyn ColumnValues<RowId>> =
+            Arc::new(IterColumn::from([0, 10].into_iter()));
+        let index = MultiValueIndex {
+            start_index_column: column,
+        };
+        let mut cursor = index.select_cursor();
+        {
+            let mut ranks = vec![0];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[0]);
+        }
+        {
+            let mut ranks = vec![5];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[]);
+        }
+    }
+
+    #[test]
+    fn test_multi_value_index_cursor_bug() {
+        let column: Arc<dyn ColumnValues<RowId>> =
+            Arc::new(IterColumn::from([0, 10].into_iter()));
+        let index = MultiValueIndex {
+            start_index_column: column,
+        };
+        let mut cursor = index.select_cursor();
+        {
+            let mut ranks = vec![0];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[0]);
+        }
+        {
+            let mut ranks = vec![4];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[]);
+        }
+        {
+            let mut ranks = vec![9];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[]);
+        }
+    }
+
+    #[test]
+    fn test_multivalue_select_cursor_skip_already_emitted() {
+        let column: Arc<dyn ColumnValues<RowId>> =
+            Arc::new(IterColumn::from(START_OFFSETS.iter().copied()));
+        let index = MultiValueIndex {
+            start_index_column: column,
+        };
+        let mut cursor = index.select_cursor();
+        {
+            let mut ranks = vec![1, 10];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[0, 1]);
+        }
+        {
+            // Here we skip row_id = 1.
+            let mut ranks = vec![11, 12];
+            cursor.select_batch_in_place(&mut ranks);
+            assert_eq!(ranks, &[2]);
+        }
+    }
+
+    fn start_index_strategy() -> impl Strategy<Value = Vec<RowId>> {
+        proptest::collection::vec(0u32..3u32, 1..6)
+            .prop_map(|deltas: Vec<u32>| {
+                let mut start_offsets: Vec<RowId> = Vec::with_capacity(deltas.len() + 1);
+                let mut cumul = 0u32;
+                start_offsets.push(cumul);
+                for delta in deltas {
+                    cumul += delta;
+                    if cumul >= 10 {
+                        break;
+                    }
+                    start_offsets.push(cumul);
+                }
+                start_offsets.push(10);
+                start_offsets
+            })
+   }
+
+    fn query_strategy() -> impl Strategy<Value = Vec<Vec<RowId>> > {
+        proptest::collection::btree_set(0u32..10u32, 1..=10)
+            .prop_flat_map(|els| {
+                let els: Vec<RowId> = els.into_iter().collect();
+                proptest::collection::btree_set(0..els.len(), 0..els.len())
+                    .prop_map(move |mut split_positions| {
+                        split_positions.insert(els.len());
+                        let mut queries: Vec<Vec<RowId>> = Vec::with_capacity(split_positions.len() + 1);
+                        let mut cursor = 0;
+                        for split_position in split_positions {
+                            queries.push(els[cursor..split_position].to_vec());
+                            cursor = split_position;
+                        }
+                        queries
+                    })
+            })
+    }
+
+
+    /// Simple inefficient implementation used for reference.
+    struct SimpleSelectCursor {
+        start_indexes: Vec<RowId>,
+        last_emitted_row_id: Option<RowId>,
+    }
+
+    impl SimpleSelectCursor {
+        fn select(&self, rank: u32) -> RowId {
+            for i in 0..self.start_indexes.len() - 1 {
+                if self.start_indexes[i] <= rank && self.start_indexes[i + 1] > rank{
+                    return i as u32;
+                }
+            }
+            panic!();
+        }
+
+        fn select_batch_in_place(&mut self, ranks: &mut Vec<RowId>) {
+            if ranks.is_empty() {
+                return;
+            }
+            for rank in ranks.iter_mut() {
+                *rank = self.select(*rank);
+            }
+            ranks.dedup();
+            if ranks.first().copied() == self.last_emitted_row_id {
+                ranks.remove(0);
+            }
+            if let Some(last_emitted) = ranks.last().copied() {
+                self.last_emitted_row_id = Some(last_emitted);
+            }
+        }
+    }
+
+
+    proptest! {
+        #[test]
+        fn test_multi_value_index_cursor_proptest(start_indexes in start_index_strategy(), mut queries in query_strategy()) {
+            let mut simple_select_cursor = SimpleSelectCursor {
+                start_indexes: start_indexes.clone(),
+                last_emitted_row_id: None
+            };
+            let column: Arc<dyn ColumnValues<RowId>> =
+                Arc::new(IterColumn::from(start_indexes.into_iter()));
+            let index = MultiValueIndex { start_index_column: column };
+            let mut select_cursor = index.select_cursor();
+            for query in queries.iter_mut() {
+                let mut query_clone = query.clone();
+                select_cursor.select_batch_in_place(query);
+                simple_select_cursor.select_batch_in_place(&mut query_clone);
+                assert_eq!(&query[..], &query_clone[..]);
+            }
+        }
     }
 }

columnar/src/column_index/optional_index/mod.rs

@@ -1,5 +1,4 @@
 use std::io::{self, Write};
-use std::ops::Range;
 use std::sync::Arc;
 
 mod set;
@@ -11,6 +10,7 @@ use set_block::{
     DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
 };
 
+use crate::iterable::Iterable;
 use crate::{InvalidData, RowId};
 
 /// The threshold for for number of elements after which we switch to dense block encoding.
@@ -88,16 +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
-    }
-}
-
 /// 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
@@ -128,17 +118,17 @@ impl<'a> BlockSelectCursor<'a> {
         }
     }
 }
-pub struct OptionalIndexSelectCursor<'a> {
-    current_block_cursor: BlockSelectCursor<'a>,
+pub struct OptionalIndexSelectCursor {
+    current_block_cursor: BlockSelectCursor<'static>,
     current_block_id: u16,
     // The current block is guaranteed to contain ranks < end_rank.
     current_block_end_rank: RowId,
-    optional_index: &'a OptionalIndex,
+    optional_index: OptionalIndex,
     block_doc_idx_start: RowId,
     num_null_rows_before_block: RowId,
 }
 
-impl<'a> OptionalIndexSelectCursor<'a> {
+impl OptionalIndexSelectCursor {
     fn search_and_load_block(&mut self, rank: RowId) {
         if rank < self.current_block_end_rank {
             // we are already in the right block
@@ -155,14 +145,23 @@ impl<'a> OptionalIndexSelectCursor<'a> {
         let block_meta = self.optional_index.block_metas[self.current_block_id as usize];
         self.num_null_rows_before_block = block_meta.non_null_rows_before_block;
         let block: Block<'_> = self.optional_index.block(block_meta);
-        self.current_block_cursor = match block {
+        let current_block_cursor = match block {
             Block::Dense(dense_block) => BlockSelectCursor::Dense(dense_block.select_cursor()),
             Block::Sparse(sparse_block) => BlockSelectCursor::Sparse(sparse_block.select_cursor()),
         };
+        // We are building a self-owned `OptionalIndexSelectCursor`.
+        self.current_block_cursor = unsafe { std::mem::transmute(current_block_cursor) };
+    }
+
+    pub fn select_batch_in_place(&mut self, ranks: &mut [RowId]) {
+        // TODO see if we can batch at the block level as well for optimization purposes.
+        for rank in ranks {
+            *rank = self.select(*rank);
+        }
     }
 }
 
-impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor<'a> {
+impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor {
     fn select(&mut self, rank: RowId) -> RowId {
         self.search_and_load_block(rank);
         let index_in_block = (rank - self.num_null_rows_before_block) as u16;
@@ -171,7 +170,7 @@ impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor<'a> {
 }
 
 impl Set<RowId> for OptionalIndex {
-    type SelectCursor<'b> = OptionalIndexSelectCursor<'b> where Self: 'b;
+    type SelectCursor<'a> = OptionalIndexSelectCursor;
     // Check if value at position is not null.
     #[inline]
     fn contains(&self, row_id: RowId) -> bool {
@@ -186,6 +185,21 @@ impl Set<RowId> for OptionalIndex {
         }
     }
 
+    #[inline]
+    fn rank(&self, row_id: RowId) -> RowId {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(row_id);
+        let block_meta = self.block_metas[block_id as usize];
+        let block = self.block(block_meta);
+        let block_offset_row_id = match block {
+            Block::Dense(dense_block) => dense_block.rank(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank(in_block_row_id),
+        } as u32;
+        block_meta.non_null_rows_before_block + block_offset_row_id
+    }
+
     #[inline]
     fn rank_if_exists(&self, row_id: RowId) -> Option<RowId> {
         let RowAddr {
@@ -215,14 +229,14 @@ 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(),
             ),
             current_block_id: 0u16,
             current_block_end_rank: 0u32, //< this is sufficient to force the first load
-            optional_index: self,
+            optional_index: self.clone(),
             block_doc_idx_start: 0u32,
             num_null_rows_before_block: 0u32,
         }
@@ -230,6 +244,31 @@ 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_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))
+    }
     pub fn select_batch(&self, ranks: &mut [RowId]) {
         let mut select_cursor = self.select_cursor();
         for rank in ranks.iter_mut() {
@@ -300,7 +339,7 @@ impl OptionalIndexCodec {
 }
 
 impl BinarySerializable for OptionalIndexCodec {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_all(&[self.to_code()])
     }
 
@@ -322,12 +361,13 @@ fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -
 }
 
 pub fn serialize_optional_index<'a, W: io::Write>(
-    serializable_optional_index: &dyn SerializableOptionalIndex<'a>,
+    non_null_rows: &dyn Iterable<RowId>,
+    num_rows: RowId,
     output: &mut W,
 ) -> io::Result<()> {
-    VInt(serializable_optional_index.num_rows() as u64).serialize(output)?;
+    VInt(num_rows as u64).serialize(output)?;
 
-    let mut rows_it = serializable_optional_index.non_null_rows();
+    let mut rows_it = non_null_rows.boxed_iter();
     let mut block_metadata: Vec<SerializedBlockMeta> = Vec::new();
     let mut current_block = Vec::new();
 
@@ -480,19 +520,5 @@ pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
     Ok(optional_index)
 }
 
-pub trait SerializableOptionalIndex<'a> {
-    fn num_rows(&self) -> RowId;
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a>;
-}
-
-impl SerializableOptionalIndex<'static> for Range<u32> {
-    fn num_rows(&self) -> RowId {
-        self.end
-    }
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'static> {
-        Box::new(self.clone())
-    }
-}
-
 #[cfg(test)]
 mod tests;

columnar/src/column_index/optional_index/set.rs

@@ -28,7 +28,10 @@ pub trait Set<T> {
     /// Returns true if the elements is contained in the Set
     fn contains(&self, el: T) -> bool;
 
-    /// If the set contains `el` returns its position in the sortd set of elements.
+    /// Returns the number of rows in the set that are < `el`
+    fn rank(&self, el: T) -> T;
+
+    /// If the set contains `el` returns the element rank.
     /// If the set does not contain the element, it returns `None`.
     fn rank_if_exists(&self, el: T) -> Option<T>;
 

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

@@ -148,6 +148,15 @@ impl<'a> Set<u16> for DenseBlock<'a> {
         }
     }
 
+    #[inline(always)]
+    fn rank(&self, el: u16) -> u16 {
+        let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
+        let index_block = self.mini_block(block_pos);
+        let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
+        let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
+        index_block.rank + ones_in_block
+    }
+
     #[inline(always)]
     fn select(&self, rank: u16) -> u16 {
         let block_id = self.find_miniblock_containing_rank(rank, 0).unwrap();

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

@@ -44,6 +44,11 @@ impl<'a> Set<u16> for SparseBlock<'a> {
         self.binary_search(el).ok()
     }
 
+    #[inline(always)]
+    fn rank(&self, el: u16) -> u16 {
+        self.binary_search(el).unwrap_or_else(|el| el)
+    }
+
     #[inline(always)]
     fn select(&self, rank: u16) -> u16 {
         let offset = rank as usize * 2;

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

@@ -17,6 +17,10 @@ fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
     for val in 0u16..=u16::MAX {
         assert_eq!(tested_set.contains(val), hash_set.contains_key(&val));
         assert_eq!(tested_set.rank_if_exists(val), hash_set.get(&val).copied());
+        assert_eq!(
+            tested_set.rank(val),
+            vals.iter().cloned().take_while(|v| *v < val).count() as u16
+        );
     }
     for rank in 0..vals.len() {
         assert_eq!(tested_set.select(rank as u16), vals[rank]);

columnar/src/column_index/optional_index/tests.rs

@@ -37,7 +37,7 @@ proptest! {
 fn test_with_random_sets_simple() {
     let vals = 10..BLOCK_SIZE * 2;
     let mut out: Vec<u8> = Vec::new();
-    serialize_optional_index(&vals.clone(), &mut out).unwrap();
+    serialize_optional_index(&vals.clone(), 100, &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
     let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
@@ -66,12 +66,8 @@ fn test_optional_index_one_block_true() {
     test_null_index(&iter[..]);
 }
 
-impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
-    fn num_rows(&self) -> RowId {
-        self.len() as u32
-    }
-
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+impl<'a> Iterable<RowId> for &'a [bool] {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
         Box::new(
             self.iter()
                 .cloned()
@@ -84,7 +80,7 @@ impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
 
 fn test_null_index(data: &[bool]) {
     let mut out: Vec<u8> = Vec::new();
-    serialize_optional_index(&data, &mut out).unwrap();
+    serialize_optional_index(&data, data.len() as RowId, &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let orig_idx_with_value: Vec<u32> = data
         .iter()
@@ -111,51 +107,96 @@ 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[..], &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[..], &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[..], &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 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));
+}
 
-    let mut out = vec![];
-    serialize_optional_index(&&docs[..], &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert!(!null_index.contains(0));
-    assert!(!null_index.contains(100));
-    assert!(!null_index.contains(ELEMENTS_PER_BLOCK - 1));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK + 1));
+fn test_optional_index_iter_aux(row_ids: &[RowId], num_rows: RowId) {
+    let optional_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(optional_index.num_rows(), num_rows);
+    assert!(optional_index.iter_rows().eq(row_ids.iter().copied()));
+}
+
+#[test]
+fn test_optional_index_iter_empty() {
+    test_optional_index_iter_aux(&[], 0u32);
+}
+
+fn test_optional_index_rank_aux(row_ids: &[RowId]) {
+    let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
+    let null_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(null_index.num_rows(), num_rows);
+    for (row_id, row_val) in row_ids.iter().copied().enumerate() {
+        assert_eq!(null_index.rank(row_val), row_id as u32);
+        assert_eq!(null_index.rank_if_exists(row_val), Some(row_id as u32));
+        if row_val > 0 && !null_index.contains(&row_val - 1) {
+            assert_eq!(null_index.rank(row_val - 1), row_id as u32);
+        }
+        assert_eq!(null_index.rank(row_val + 1), row_id as u32 + 1);
+    }
+}
+
+#[test]
+fn test_optional_index_rank() {
+    test_optional_index_rank_aux(&[1u32]);
+    test_optional_index_rank_aux(&[0u32, 1u32]);
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_rank_aux(&block);
+}
+
+#[test]
+fn test_optional_index_iter_empty_one() {
+    test_optional_index_iter_aux(&[1], 2u32);
+    test_optional_index_iter_aux(&[100_000], 200_000u32);
+}
+
+#[test]
+fn test_optional_index_iter_dense_block() {
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_iter_aux(&block, 3 * BLOCK_SIZE);
+}
+
+#[test]
+fn test_optional_index_for_tests() {
+    let optional_index = OptionalIndex::for_test(4, &[1, 2]);
+    assert!(!optional_index.contains(0));
+    assert!(optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+    assert_eq!(optional_index.num_rows(), 4);
 }
 
 #[cfg(all(test, feature = "unstable"))]

columnar/src/column_index/serialize.rs

@@ -5,23 +5,26 @@ use common::{CountingWriter, OwnedBytes};
 
 use crate::column_index::multivalued_index::serialize_multivalued_index;
 use crate::column_index::optional_index::serialize_optional_index;
-use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
-use crate::column_values::ColumnValues;
+use crate::column_index::ColumnIndex;
+use crate::iterable::Iterable;
 use crate::{Cardinality, RowId};
 
 pub enum SerializableColumnIndex<'a> {
     Full,
-    Optional(Box<dyn SerializableOptionalIndex<'a> + 'a>),
+    Optional {
+        non_null_row_ids: Box<dyn Iterable<RowId> + 'a>,
+        num_rows: RowId,
+    },
     // TODO remove the Arc<dyn> apart from serialization this is not
     // dynamic at all.
-    Multivalued(Box<dyn ColumnValues<RowId> + 'a>),
+    Multivalued(Box<dyn Iterable<RowId> + 'a>),
 }
 
 impl<'a> SerializableColumnIndex<'a> {
     pub fn get_cardinality(&self) -> Cardinality {
         match self {
             SerializableColumnIndex::Full => Cardinality::Full,
-            SerializableColumnIndex::Optional(_) => Cardinality::Optional,
+            SerializableColumnIndex::Optional { .. } => Cardinality::Optional,
             SerializableColumnIndex::Multivalued(_) => Cardinality::Multivalued,
         }
     }
@@ -36,9 +39,10 @@ pub fn serialize_column_index(
     output.write_all(&[cardinality])?;
     match column_index {
         SerializableColumnIndex::Full => {}
-        SerializableColumnIndex::Optional(optional_index) => {
-            serialize_optional_index(&*optional_index, &mut output)?
-        }
+        SerializableColumnIndex::Optional {
+            non_null_row_ids,
+            num_rows,
+        } => serialize_optional_index(non_null_row_ids.as_ref(), num_rows, &mut output)?,
         SerializableColumnIndex::Multivalued(multivalued_index) => {
             serialize_multivalued_index(&*multivalued_index, &mut output)?
         }

columnar/src/column_values/bitpacked.rs

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

columnar/src/column_values/blockwise_linear.rs

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

columnar/src/column_values/column.rs

@@ -1,6 +1,7 @@
 use std::fmt::Debug;
 use std::marker::PhantomData;
 use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
 
 use tantivy_bitpacker::minmax;
 
@@ -9,7 +10,7 @@ use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
 /// `ColumnValues` provides access to a dense field column.
 ///
 /// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
-pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
+pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     /// Return the value associated with the given idx.
     ///
     /// This accessor should return as fast as possible.
@@ -27,7 +28,7 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
     ///
     /// Must panic if `start + output.len()` is greater than
     /// the segment's `maxdoc`.
-    #[inline]
+    #[inline(always)]
     fn get_range(&self, start: u64, output: &mut [T]) {
         for (out, idx) in output.iter_mut().zip(start..) {
             *out = self.get_val(idx as u32);
@@ -37,7 +38,7 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
     /// Get the positions of values which are in the provided value range.
     ///
     /// Note that position == docid for single value fast fields
-    #[inline]
+    #[inline(always)]
     fn get_docids_for_value_range(
         &self,
         value_range: RangeInclusive<T>,
@@ -78,27 +79,33 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
     }
 }
 
-impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for std::sync::Arc<dyn ColumnValues<T>> {
+impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
+    #[inline(always)]
     fn get_val(&self, idx: u32) -> T {
         self.as_ref().get_val(idx)
     }
 
+    #[inline(always)]
     fn min_value(&self) -> T {
         self.as_ref().min_value()
     }
 
+    #[inline(always)]
     fn max_value(&self) -> T {
         self.as_ref().max_value()
     }
 
+    #[inline(always)]
     fn num_vals(&self) -> u32 {
         self.as_ref().num_vals()
     }
 
+    #[inline(always)]
     fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
         self.as_ref().iter()
     }
 
+    #[inline(always)]
     fn get_range(&self, start: u64, output: &mut [T]) {
         self.as_ref().get_range(start, output)
     }

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;
@@ -55,7 +56,7 @@ impl RangeMapping {
 }
 
 impl BinarySerializable for CompactSpace {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
 
         let mut prev_value = 0;
@@ -158,23 +159,30 @@ impl CompactSpace {
 pub struct CompactSpaceCompressor {
     params: IPCodecParams,
 }
+
 #[derive(Debug, Clone)]
 pub struct IPCodecParams {
     compact_space: CompactSpace,
     bit_unpacker: BitUnpacker,
     min_value: u128,
     max_value: u128,
-    num_vals: u32,
+    num_vals: RowId,
     num_bits: u8,
 }
 
 impl CompactSpaceCompressor {
-    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
-    pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
-        let mut values_sorted = BTreeSet::new();
-        values_sorted.extend(iter);
-        let total_num_values = num_vals;
+    pub fn num_vals(&self) -> RowId {
+        self.params.num_vals
+    }
 
+    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
+    pub fn train_from(iter: impl Iterator<Item = u128>) -> Self {
+        let mut values_sorted = BTreeSet::new();
+        let mut total_num_values = 0u32;
+        for val in iter {
+            total_num_values += 1u32;
+            values_sorted.insert(val);
+        }
         let compact_space =
             get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
         let amplitude_compact_space = compact_space.amplitude_compact_space();
@@ -247,7 +255,7 @@ pub struct CompactSpaceDecompressor {
 }
 
 impl BinarySerializable for IPCodecParams {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         // header flags for future optional dictionary encoding
         let footer_flags = 0u64;
         footer_flags.serialize(writer)?;
@@ -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_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_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/gcd.rs

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

columnar/src/column_values/linear.rs

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

columnar/src/column_values/main.rs

@@ -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

@@ -7,9 +7,6 @@
 //! - Encode data in different codecs.
 //! - Monotonically map values to u64/u128
 
-#[cfg(test)]
-mod tests;
-
 use std::fmt::Debug;
 use std::io;
 use std::io::Write;
@@ -18,72 +15,66 @@ use std::sync::Arc;
 use common::{BinarySerializable, OwnedBytes};
 use compact_space::CompactSpaceDecompressor;
 pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
-use monotonic_mapping::{
-    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
-    StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
-};
+use monotonic_mapping::{StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal};
 pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
-use serialize::{Header, U128Header};
+use serialize::U128Header;
 
-mod bitpacked;
-mod blockwise_linear;
 mod compact_space;
-mod line;
-mod linear;
 pub(crate) mod monotonic_mapping;
 pub(crate) mod monotonic_mapping_u128;
+mod stats;
+pub(crate) mod u64_based;
 
 mod column;
-mod gcd;
 pub mod serialize;
 
+pub use serialize::serialize_column_values_u128;
+pub use stats::Stats;
+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};
-#[cfg(test)]
-pub use self::serialize::tests::serialize_and_load;
-pub use self::serialize::{serialize_column_values, NormalizedHeader};
-use crate::column_values::bitpacked::BitpackedCodec;
-use crate::column_values::blockwise_linear::BlockwiseLinearCodec;
-use crate::column_values::linear::LinearCodec;
+use crate::iterable::Iterable;
+use crate::{ColumnIndex, MergeRowOrder};
 
-#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
-#[repr(u8)]
-/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
-pub enum FastFieldCodecType {
-    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
-    /// `column.max_value() - column.min_value()`
-    Bitpacked = 1,
-    /// Linear interpolation puts a line between the first and last value and then bitpacks the
-    /// values by the offset from the line. The number of bits is defined by the max deviation from
-    /// the line.
-    Linear = 2,
-    /// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
-    BlockwiseLinear = 3,
+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 BinarySerializable for FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
-        self.to_code().serialize(wrt)
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let code = u8::deserialize(reader)?;
-        let codec_type: Self = Self::from_code(code)
-            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
-        Ok(codec_type)
-    }
-}
-
-impl FastFieldCodecType {
-    pub(crate) fn to_code(self) -> u8 {
-        self as u8
-    }
-
-    pub(crate) fn from_code(code: u8) -> Option<Self> {
-        match code {
-            1 => Some(Self::Bitpacked),
-            2 => Some(Self::Linear),
-            3 => Some(Self::BlockwiseLinear),
-            _ => None,
+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))
+                    })
+                )
+            },
         }
     }
 }
@@ -98,7 +89,7 @@ pub enum U128FastFieldCodecType {
 }
 
 impl BinarySerializable for U128FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, wrt: &mut W) -> io::Result<()> {
         self.to_code().serialize(wrt)
     }
 
@@ -136,68 +127,6 @@ pub fn open_u128_mapped<T: MonotonicallyMappableToU128 + Debug>(
     Ok(Arc::new(monotonic_map_column(reader, inverted)))
 }
 
-/// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open_u64_mapped<T: MonotonicallyMappableToU64 + Debug>(
-    mut bytes: OwnedBytes,
-) -> io::Result<Arc<dyn ColumnValues<T>>> {
-    let header = Header::deserialize(&mut bytes)?;
-    match header.codec_type {
-        FastFieldCodecType::Bitpacked => open_specific_codec::<BitpackedCodec, _>(bytes, &header),
-        FastFieldCodecType::Linear => open_specific_codec::<LinearCodec, _>(bytes, &header),
-        FastFieldCodecType::BlockwiseLinear => {
-            open_specific_codec::<BlockwiseLinearCodec, _>(bytes, &header)
-        }
-    }
-}
-
-fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64 + Debug>(
-    bytes: OwnedBytes,
-    header: &Header,
-) -> io::Result<Arc<dyn ColumnValues<Item>>> {
-    let normalized_header = header.normalized();
-    let reader = C::open_from_bytes(bytes, normalized_header)?;
-    let min_value = header.min_value;
-    if let Some(gcd) = header.gcd {
-        let mapping = StrictlyMonotonicMappingInverter::from(
-            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
-        );
-        Ok(Arc::new(monotonic_map_column(reader, mapping)))
-    } else {
-        let mapping = StrictlyMonotonicMappingInverter::from(
-            StrictlyMonotonicMappingToInternalBaseval::new(min_value),
-        );
-        Ok(Arc::new(monotonic_map_column(reader, mapping)))
-    }
-}
-
-/// The FastFieldSerializerEstimate trait is required on all variants
-/// of fast field compressions, to decide which one to choose.
-pub(crate) trait FastFieldCodec: 'static {
-    /// A codex needs to provide a unique name and id, which is
-    /// used for debugging and de/serialization.
-    const CODEC_TYPE: FastFieldCodecType;
-
-    type Reader: ColumnValues<u64> + 'static;
-
-    /// Reads the metadata and returns the CodecReader
-    fn open_from_bytes(bytes: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader>;
-
-    /// Serializes the data using the serializer into write.
-    ///
-    /// The column iterator should be preferred over using column `get_val` method for
-    /// performance reasons.
-    fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()>;
-
-    /// Returns an estimate of the compression ratio.
-    /// If the codec is not applicable, returns `None`.
-    ///
-    /// The baseline is uncompressed 64bit data.
-    ///
-    /// It could make sense to also return a value representing
-    /// computational complexity.
-    fn estimate(column: &dyn ColumnValues) -> Option<f32>;
-}
-
 #[cfg(all(test, feature = "unstable"))]
 mod bench {
     use std::sync::Arc;

columnar/src/column_values/serialize.rs

@@ -1,22 +1,12 @@
 use std::fmt::Debug;
 use std::io;
-use std::num::NonZeroU64;
 
 use common::{BinarySerializable, VInt};
-use log::warn;
 
-use super::bitpacked::BitpackedCodec;
-use super::blockwise_linear::BlockwiseLinearCodec;
-use super::linear::LinearCodec;
-use super::monotonic_mapping::{
-    StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
-    StrictlyMonotonicMappingToInternalGCDBaseval,
-};
-use super::{
-    monotonic_map_column, ColumnValues, FastFieldCodec, FastFieldCodecType,
-    MonotonicallyMappableToU64, U128FastFieldCodecType,
-};
 use crate::column_values::compact_space::CompactSpaceCompressor;
+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:
@@ -31,53 +21,6 @@ pub struct NormalizedHeader {
     pub max_value: u64,
 }
 
-#[derive(Debug, Copy, Clone)]
-pub(crate) struct Header {
-    pub num_vals: u32,
-    pub min_value: u64,
-    pub max_value: u64,
-    pub gcd: Option<NonZeroU64>,
-    pub codec_type: FastFieldCodecType,
-}
-
-impl Header {
-    pub fn normalized(self) -> NormalizedHeader {
-        let gcd = self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
-        let gcd_min_val_mapping =
-            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, self.min_value);
-
-        let max_value = gcd_min_val_mapping.mapping(self.max_value);
-        NormalizedHeader {
-            num_vals: self.num_vals,
-            max_value,
-        }
-    }
-
-    pub(crate) fn normalize_column<C: ColumnValues>(&self, from_column: C) -> impl ColumnValues {
-        normalize_column(from_column, self.min_value, self.gcd)
-    }
-
-    pub fn compute_header(
-        column: impl ColumnValues<u64>,
-        codecs: &[FastFieldCodecType],
-    ) -> Option<Header> {
-        let num_vals = column.num_vals();
-        let min_value = column.min_value();
-        let max_value = column.max_value();
-        let gcd = super::gcd::find_gcd(column.iter().map(|val| val - min_value))
-            .filter(|gcd| gcd.get() > 1u64);
-        let normalized_column = normalize_column(column, min_value, gcd);
-        let codec_type = detect_codec(normalized_column, codecs)?;
-        Some(Header {
-            num_vals,
-            min_value,
-            max_value,
-            gcd,
-            codec_type,
-        })
-    }
-}
-
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
 pub(crate) struct U128Header {
     pub num_vals: u32,
@@ -85,7 +28,7 @@ pub(crate) struct U128Header {
 }
 
 impl BinarySerializable for U128Header {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.num_vals as u64).serialize(writer)?;
         self.codec_type.serialize(writer)?;
         Ok(())
@@ -101,157 +44,39 @@ impl BinarySerializable for U128Header {
     }
 }
 
-fn normalize_column<C: ColumnValues>(
-    from_column: C,
-    min_value: u64,
-    gcd: Option<NonZeroU64>,
-) -> impl ColumnValues {
-    let gcd = gcd.map(|gcd| gcd.get()).unwrap_or(1);
-    let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, min_value);
-    monotonic_map_column(from_column, mapping)
-}
-
-impl BinarySerializable for Header {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
-        VInt(self.num_vals as u64).serialize(writer)?;
-        VInt(self.min_value).serialize(writer)?;
-        VInt(self.max_value - self.min_value).serialize(writer)?;
-        if let Some(gcd) = self.gcd {
-            VInt(gcd.get()).serialize(writer)?;
-        } else {
-            VInt(0u64).serialize(writer)?;
-        }
-        self.codec_type.serialize(writer)?;
-        Ok(())
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let num_vals = VInt::deserialize(reader)?.0 as u32;
-        let min_value = VInt::deserialize(reader)?.0;
-        let amplitude = VInt::deserialize(reader)?.0;
-        let max_value = min_value + amplitude;
-        let gcd_u64 = VInt::deserialize(reader)?.0;
-        let codec_type = FastFieldCodecType::deserialize(reader)?;
-        Ok(Header {
-            num_vals,
-            min_value,
-            max_value,
-            gcd: NonZeroU64::new(gcd_u64),
-            codec_type,
-        })
-    }
-}
-
 /// Serializes u128 values with the compact space codec.
-pub fn serialize_column_values_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
-    iter_gen: F,
-    num_vals: u32,
+pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
+    iterable: &dyn Iterable<T>,
     output: &mut impl io::Write,
 ) -> io::Result<()> {
+    let compressor = CompactSpaceCompressor::train_from(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+    );
     let header = U128Header {
-        num_vals,
+        num_vals: compressor.num_vals(),
         codec_type: U128FastFieldCodecType::CompactSpace,
     };
     header.serialize(output)?;
-    let compressor = CompactSpaceCompressor::train_from(iter_gen(), num_vals);
-    compressor.compress_into(iter_gen(), output)?;
-
-    Ok(())
-}
-
-/// Serializes the column with the codec with the best estimate on the data.
-pub fn serialize_column_values<T: MonotonicallyMappableToU64 + Debug>(
-    typed_column: impl ColumnValues<T>,
-    codecs: &[FastFieldCodecType],
-    output: &mut impl io::Write,
-) -> io::Result<()> {
-    let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
-    let header = Header::compute_header(&column, codecs).ok_or_else(|| {
-        io::Error::new(
-            io::ErrorKind::InvalidInput,
-            format!(
-                "Data cannot be serialized with this list of codec. {:?}",
-                codecs
-            ),
-        )
-    })?;
-    header.serialize(output)?;
-    let normalized_column = header.normalize_column(column);
-    assert_eq!(normalized_column.min_value(), 0u64);
-    serialize_given_codec(normalized_column, header.codec_type, output)?;
-    Ok(())
-}
-
-fn detect_codec(
-    column: impl ColumnValues<u64>,
-    codecs: &[FastFieldCodecType],
-) -> Option<FastFieldCodecType> {
-    let mut estimations = Vec::new();
-    for &codec in codecs {
-        let estimation_opt = match codec {
-            FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&column),
-            FastFieldCodecType::Linear => LinearCodec::estimate(&column),
-            FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&column),
-        };
-        if let Some(estimation) = estimation_opt {
-            estimations.push((estimation, codec));
-        }
-    }
-    if let Some(broken_estimation) = estimations.iter().find(|estimation| estimation.0.is_nan()) {
-        warn!(
-            "broken estimation for fast field codec {:?}",
-            broken_estimation.1
-        );
-    }
-    // removing nan values for codecs with broken calculations, and max values which disables
-    // codecs
-    estimations.retain(|estimation| !estimation.0.is_nan() && estimation.0 != f32::MAX);
-    estimations.sort_by(|(score_left, _), (score_right, _)| score_left.total_cmp(score_right));
-    Some(estimations.first()?.1)
-}
-
-pub(crate) fn serialize_given_codec(
-    column: impl ColumnValues<u64>,
-    codec_type: FastFieldCodecType,
-    output: &mut impl io::Write,
-) -> io::Result<()> {
-    match codec_type {
-        FastFieldCodecType::Bitpacked => {
-            BitpackedCodec::serialize(&column, output)?;
-        }
-        FastFieldCodecType::Linear => {
-            LinearCodec::serialize(&column, output)?;
-        }
-        FastFieldCodecType::BlockwiseLinear => {
-            BlockwiseLinearCodec::serialize(&column, output)?;
-        }
-    }
+    compressor.compress_into(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+        output,
+    )?;
     Ok(())
 }
 
 #[cfg(test)]
 pub mod tests {
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-
     use super::*;
-    use crate::column_values::{open_u64_mapped, VecColumn};
+    use crate::column_values::u64_based::{
+        serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
+        ALL_U64_CODEC_TYPES,
+    };
+    use crate::column_values::CodecType;
 
-    const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
-        FastFieldCodecType::Bitpacked,
-        FastFieldCodecType::Linear,
-        FastFieldCodecType::BlockwiseLinear,
-    ];
-
-    /// Helper function to serialize a column (autodetect from all codecs) and then open it
-    pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
-        column: &[T],
-    ) -> Arc<dyn ColumnValues<T>> {
-        let mut buffer = Vec::new();
-        serialize_column_values(&VecColumn::from(&column), &ALL_CODEC_TYPES, &mut buffer).unwrap();
-        open_u64_mapped(OwnedBytes::new(buffer)).unwrap()
-    }
     #[test]
     fn test_serialize_deserialize_u128_header() {
         let original = U128Header {
@@ -267,15 +92,22 @@ pub mod tests {
     #[test]
     fn test_serialize_deserialize() {
         let original = [1u64, 5u64, 10u64];
-        let restored: Vec<u64> = serialize_and_load(&original[..]).iter().collect();
+        let restored: Vec<u64> =
+            serialize_and_load_u64_based_column_values(&&original[..], &ALL_U64_CODEC_TYPES)
+                .iter()
+                .collect();
         assert_eq!(&restored, &original[..]);
     }
 
     #[test]
     fn test_fastfield_bool_size_bitwidth_1() {
         let mut buffer = Vec::new();
-        let col = VecColumn::from(&[false, true][..]);
-        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
         // TODO put the header as a footer so that it serves as a padding.
         // 5 bytes of header, 1 byte of value, 7 bytes of padding.
         assert_eq!(buffer.len(), 5 + 1);
@@ -284,19 +116,23 @@ pub mod tests {
     #[test]
     fn test_fastfield_bool_bit_size_bitwidth_0() {
         let mut buffer = Vec::new();
-        let col = VecColumn::from(&[true][..]);
-        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
-        // 5 bytes of header, 0 bytes of value, 7 bytes of padding.
-        assert_eq!(buffer.len(), 5);
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
+        // 6 bytes of header, 0 bytes of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 6);
     }
 
     #[test]
     fn test_fastfield_gcd() {
         let mut buffer = Vec::new();
         let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
-        let col = VecColumn::from(&vals[..]);
-        serialize_column_values(&col, &[FastFieldCodecType::Bitpacked], &mut buffer).unwrap();
+        serialize_u64_based_column_values(&&vals[..], &[CodecType::Bitpacked], &mut buffer)
+            .unwrap();
         // Values are stored over 3 bits.
-        assert_eq!(buffer.len(), 7 + (3 * 80 / 8));
+        assert_eq!(buffer.len(), 6 + (3 * 80 / 8));
     }
 }

columnar/src/column_values/stats.rs

@@ -0,0 +1,96 @@
+use std::io;
+use std::io::Write;
+use std::num::NonZeroU64;
+
+use common::{BinarySerializable, VInt};
+
+use crate::RowId;
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct Stats {
+    pub gcd: NonZeroU64,
+    pub min_value: u64,
+    pub max_value: u64,
+    pub num_rows: RowId,
+}
+
+impl Stats {
+    pub fn amplitude(&self) -> u64 {
+        self.max_value - self.min_value
+    }
+}
+
+impl BinarySerializable for Stats {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.min_value).serialize(writer)?;
+        VInt(self.gcd.get()).serialize(writer)?;
+        VInt(self.amplitude() / self.gcd).serialize(writer)?;
+        VInt(self.num_rows as u64).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let min_value = VInt::deserialize(reader)?.0;
+        let gcd = VInt::deserialize(reader)?.0;
+        let gcd = NonZeroU64::new(gcd)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "GCD of 0 is forbidden"))?;
+        let amplitude = VInt::deserialize(reader)?.0 * gcd.get();
+        let max_value = min_value + amplitude;
+        let num_rows = VInt::deserialize(reader)?.0 as RowId;
+        Ok(Stats {
+            min_value,
+            max_value,
+            num_rows,
+            gcd,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use common::BinarySerializable;
+
+    use crate::column_values::Stats;
+
+    #[track_caller]
+    fn test_stats_ser_deser_aux(stats: &Stats, 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();
+        assert_eq!(stats, &deser_stats);
+    }
+
+    #[test]
+    fn test_stats_serialization() {
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(3).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(1_000).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0,
+            }),
+            4,
+        );
+    }
+}

columnar/src/column_values/u64_based/bitpacked.rs

@@ -0,0 +1,127 @@
+use std::io::{self, Write};
+
+use common::{BinarySerializable, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::{ColumnValues, RowId};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct BitpackedReader {
+    data: OwnedBytes,
+    bit_unpacker: BitUnpacker,
+    stats: Stats,
+}
+
+impl ColumnValues for BitpackedReader {
+    #[inline(always)]
+    fn get_val(&self, doc: u32) -> u64 {
+        self.stats.min_value + self.stats.gcd.get() * self.bit_unpacker.get(doc, &self.data)
+    }
+
+    #[inline]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+    #[inline]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+    #[inline]
+    fn num_vals(&self) -> RowId {
+        self.stats.num_rows
+    }
+}
+
+fn num_bits(stats: &Stats) -> u8 {
+    compute_num_bits(stats.amplitude() / stats.gcd)
+}
+
+#[derive(Default)]
+pub struct BitpackedCodecEstimator;
+
+impl ColumnCodecEstimator for BitpackedCodecEstimator {
+    fn collect(&mut self, _value: u64) {}
+
+    fn estimate(&self, stats: &Stats) -> 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,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let num_bits = num_bits(stats);
+        let mut bit_packer = BitPacker::new();
+        let divider = DividerU64::divide_by(stats.gcd.get());
+        for val in vals {
+            bit_packer.write(divider.divide(val - stats.min_value), num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+}
+
+pub struct BitpackedCodec;
+
+impl ColumnCodec for BitpackedCodec {
+    type Reader = 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)?;
+        let num_bits = num_bits(&stats);
+        let bit_unpacker = BitUnpacker::new(num_bits);
+        Ok(BitpackedReader {
+            data,
+            bit_unpacker,
+            stats,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BitpackedCodec>(&[4, 3, 12], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        create_and_validate::<BitpackedCodec>(&[1000, 2000, 3000], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BitpackedCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn bitpacked_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/blockwise_linear.rs

@@ -0,0 +1,281 @@
+use std::io::Write;
+use std::sync::Arc;
+use std::{io, iter};
+
+use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::u64_based::line::Line;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::column_values::{ColumnValues, VecColumn};
+use crate::MonotonicallyMappableToU64;
+
+const BLOCK_SIZE: u32 = 512u32;
+
+#[derive(Debug, Default)]
+struct Block {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+    data_start_offset: usize,
+}
+
+impl BinarySerializable for Block {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Block {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+            data_start_offset: 0,
+        })
+    }
+}
+
+fn compute_num_blocks(num_vals: u32) -> u32 {
+    (num_vals + BLOCK_SIZE - 1) / BLOCK_SIZE
+}
+
+pub struct BlockwiseLinearEstimator {
+    block: Vec<u64>,
+    values_num_bytes: u64,
+    meta_num_bytes: u64,
+}
+
+impl Default for BlockwiseLinearEstimator {
+    fn default() -> Self {
+        Self {
+            block: Vec::with_capacity(BLOCK_SIZE as usize),
+            values_num_bytes: 0u64,
+            meta_num_bytes: 0u64,
+        }
+    }
+}
+
+impl BlockwiseLinearEstimator {
+    fn flush_block_estimate(&mut self) {
+        if self.block.is_empty() {
+            return;
+        }
+        let line = Line::train(&VecColumn::from(&self.block));
+        let mut max_value = 0u64;
+        for (i, buffer_val) in self.block.iter().enumerate() {
+            let interpolated_val = line.eval(i as u32);
+            let val = buffer_val.wrapping_sub(interpolated_val);
+            max_value = val.max(max_value);
+        }
+        let bit_width = compute_num_bits(max_value) as usize;
+        self.values_num_bytes += (bit_width * self.block.len() + 7) as u64 / 8;
+        self.meta_num_bytes += 1 + line.num_bytes();
+    }
+}
+
+impl ColumnCodecEstimator for BlockwiseLinearEstimator {
+    fn collect(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == BLOCK_SIZE as usize {
+            self.flush_block_estimate();
+            self.block.clear();
+        }
+    }
+    fn estimate(&self, stats: &Stats) -> Option<u64> {
+        let mut estimate = 4 + stats.num_bytes() + self.meta_num_bytes + self.values_num_bytes;
+        if stats.gcd.get() > 1 {
+            let estimate_gain_from_gcd =
+                (stats.gcd.get() as f32).log2().floor() * stats.num_rows as f32 / 8.0f32;
+            estimate = estimate.saturating_sub(estimate_gain_from_gcd as u64);
+        }
+        Some(estimate)
+    }
+
+    fn finalize(&mut self) {
+        self.flush_block_estimate();
+    }
+
+    fn serialize(
+        &self,
+        stats: &Stats,
+        mut vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let mut buffer = Vec::with_capacity(BLOCK_SIZE as usize);
+        let num_blocks = compute_num_blocks(stats.num_rows) as usize;
+        let mut blocks = Vec::with_capacity(num_blocks);
+
+        let mut bit_packer = BitPacker::new();
+
+        let gcd_divider = DividerU64::divide_by(stats.gcd.get());
+
+        for _ in 0..num_blocks {
+            buffer.clear();
+            buffer.extend(
+                (&mut vals)
+                    .map(MonotonicallyMappableToU64::to_u64)
+                    .take(BLOCK_SIZE as usize),
+            );
+
+            for buffer_val in buffer.iter_mut() {
+                *buffer_val = gcd_divider.divide(*buffer_val - stats.min_value);
+            }
+
+            let line = Line::train(&VecColumn::from(&buffer));
+
+            assert!(!buffer.is_empty());
+
+            for (i, buffer_val) in buffer.iter_mut().enumerate() {
+                let interpolated_val = line.eval(i as u32);
+                *buffer_val = buffer_val.wrapping_sub(interpolated_val);
+            }
+
+            let bit_width = buffer.iter().copied().map(compute_num_bits).max().unwrap();
+
+            for &buffer_val in &buffer {
+                bit_packer.write(buffer_val, bit_width, wrt)?;
+            }
+
+            blocks.push(Block {
+                line,
+                bit_unpacker: BitUnpacker::new(bit_width),
+                data_start_offset: 0,
+            });
+        }
+
+        bit_packer.close(wrt)?;
+
+        assert_eq!(blocks.len(), num_blocks);
+
+        let mut counting_wrt = CountingWriter::wrap(wrt);
+        for block in &blocks {
+            block.serialize(&mut counting_wrt)?;
+        }
+        let footer_len = counting_wrt.written_bytes();
+        (footer_len as u32).serialize(&mut counting_wrt)?;
+
+        Ok(())
+    }
+}
+
+pub struct BlockwiseLinearCodec;
+
+impl ColumnCodec<u64> for BlockwiseLinearCodec {
+    type Reader = BlockwiseLinearReader;
+
+    type Estimator = BlockwiseLinearEstimator;
+
+    fn load(mut bytes: OwnedBytes) -> io::Result<Self::Reader> {
+        let stats = Stats::deserialize(&mut bytes)?;
+        let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
+        let footer_offset = bytes.len() - 4 - footer_len as usize;
+        let (data, mut footer) = bytes.split(footer_offset);
+        let num_blocks = compute_num_blocks(stats.num_rows);
+        let mut blocks: Vec<Block> = iter::repeat_with(|| Block::deserialize(&mut footer))
+            .take(num_blocks as usize)
+            .collect::<io::Result<_>>()?;
+        let mut start_offset = 0;
+        for block in &mut blocks {
+            block.data_start_offset = start_offset;
+            start_offset += (block.bit_unpacker.bit_width() as usize) * BLOCK_SIZE as usize / 8;
+        }
+        Ok(BlockwiseLinearReader {
+            blocks: blocks.into_boxed_slice().into(),
+            data,
+            stats,
+        })
+    }
+}
+
+#[derive(Clone)]
+pub struct BlockwiseLinearReader {
+    blocks: Arc<[Block]>,
+    data: OwnedBytes,
+    stats: Stats,
+}
+
+impl ColumnValues for BlockwiseLinearReader {
+    #[inline(always)]
+    fn get_val(&self, idx: u32) -> u64 {
+        let block_id = (idx / BLOCK_SIZE as u32) as usize;
+        let idx_within_block = idx % (BLOCK_SIZE as u32);
+        let block = &self.blocks[block_id];
+        let interpoled_val: u64 = block.line.eval(idx_within_block);
+        let block_bytes = &self.data[block.data_start_offset..];
+        let bitpacked_diff = block.bit_unpacker.get(idx_within_block, block_bytes);
+        // TODO optimize me! the line parameters could be tweaked to include the multiplication and
+        // remove the dependency.
+        self.stats.min_value
+            + self
+                .stats
+                .gcd
+                .get()
+                .wrapping_mul(interpoled_val.wrapping_add(bitpacked_diff))
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline(always)]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BlockwiseLinearCodec>(
+            &[11, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "simple test",
+        )
+        .unwrap();
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        let (_, actual_compression_rate) = create_and_validate::<BlockwiseLinearCodec>(
+            &[10, 20, 40, 20, 10, 10, 10, 10, 10, 10],
+            "name",
+        )
+        .unwrap();
+        assert_eq!(actual_compression_rate, 0.175);
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn test_blockwise_linear_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate::<BlockwiseLinearCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BlockwiseLinearCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/line.rs

@@ -17,8 +17,8 @@ const MID_POINT: u64 = (1u64 << 32) - 1u64;
 /// `y = m * x >> 32 + b`
 #[derive(Debug, Clone, Copy, Default)]
 pub struct Line {
-    slope: u64,
-    intercept: u64,
+    pub(crate) slope: u64,
+    pub(crate) intercept: u64,
 }
 
 /// Compute the line slope.
@@ -67,21 +67,8 @@ impl Line {
         self.intercept.wrapping_add(linear_part)
     }
 
-    // Same as train, but the intercept is only estimated from provided sample positions
-    pub fn estimate(sample_positions_and_values: &[(u64, u64)]) -> Self {
-        let first_val = sample_positions_and_values[0].1;
-        let last_val = sample_positions_and_values[sample_positions_and_values.len() - 1].1;
-        let num_vals = sample_positions_and_values[sample_positions_and_values.len() - 1].0 + 1;
-        Self::train_from(
-            first_val,
-            last_val,
-            num_vals as u32,
-            sample_positions_and_values.iter().cloned(),
-        )
-    }
-
     // Intercept is only computed from provided positions
-    fn train_from(
+    pub fn train_from(
         first_val: u64,
         last_val: u64,
         num_vals: u32,
@@ -145,6 +132,7 @@ impl Line {
     ///
     /// This function is only invariable by translation if all of the
     /// `ys` are packaged into half of the space. (See heuristic below)
+    /// TODO USE array
     pub fn train(ys: &dyn ColumnValues) -> Self {
         let first_val = ys.iter().next().unwrap();
         let last_val = ys.iter().nth(ys.num_vals() as usize - 1).unwrap();
@@ -158,7 +146,7 @@ impl Line {
 }
 
 impl BinarySerializable for Line {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.slope).serialize(writer)?;
         VInt(self.intercept).serialize(writer)?;
         Ok(())

columnar/src/column_values/u64_based/linear.rs

@@ -0,0 +1,277 @@
+use std::io;
+
+use common::{BinarySerializable, OwnedBytes};
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use super::line::Line;
+use super::ColumnValues;
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::column_values::VecColumn;
+use crate::RowId;
+
+const HALF_SPACE: u64 = u64::MAX / 2;
+const LINE_ESTIMATION_BLOCK_LEN: usize = 512;
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct LinearReader {
+    data: OwnedBytes,
+    linear_params: LinearParams,
+    stats: Stats,
+}
+
+impl ColumnValues for LinearReader {
+    #[inline]
+    fn get_val(&self, doc: u32) -> u64 {
+        let interpoled_val: u64 = self.linear_params.line.eval(doc);
+        let bitpacked_diff = self.linear_params.bit_unpacker.get(doc, &self.data);
+        interpoled_val.wrapping_add(bitpacked_diff)
+    }
+
+    #[inline(always)]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+
+    #[inline(always)]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+
+    #[inline]
+    fn num_vals(&self) -> u32 {
+        self.stats.num_rows
+    }
+}
+
+/// Fastfield serializer, which tries to guess values by linear interpolation
+/// and stores the difference bitpacked.
+pub struct LinearCodec;
+
+#[derive(Debug, Clone)]
+struct LinearParams {
+    line: Line,
+    bit_unpacker: BitUnpacker,
+}
+
+impl BinarySerializable for LinearParams {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        self.line.serialize(writer)?;
+        self.bit_unpacker.bit_width().serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let line = Line::deserialize(reader)?;
+        let bit_width = u8::deserialize(reader)?;
+        Ok(Self {
+            line,
+            bit_unpacker: BitUnpacker::new(bit_width),
+        })
+    }
+}
+
+pub struct LinearCodecEstimator {
+    block: Vec<u64>,
+    line: Option<Line>,
+    row_id: RowId,
+    min_deviation: u64,
+    max_deviation: u64,
+    first_val: u64,
+    last_val: u64,
+}
+
+impl Default for LinearCodecEstimator {
+    fn default() -> LinearCodecEstimator {
+        LinearCodecEstimator {
+            block: Vec::with_capacity(LINE_ESTIMATION_BLOCK_LEN),
+            line: None,
+            row_id: 0,
+            min_deviation: u64::MAX,
+            max_deviation: u64::MIN,
+            first_val: 0u64,
+            last_val: 0u64,
+        }
+    }
+}
+
+impl ColumnCodecEstimator for LinearCodecEstimator {
+    fn finalize(&mut self) {
+        if let Some(line) = self.line.as_mut() {
+            line.intercept = line
+                .intercept
+                .wrapping_add(self.min_deviation)
+                .wrapping_sub(HALF_SPACE);
+        }
+    }
+
+    fn estimate(&self, stats: &Stats) -> Option<u64> {
+        let line = self.line?;
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        Some(
+            stats.num_bytes()
+                + linear_params.num_bytes()
+                + (num_bits as u64 * stats.num_rows as u64 + 7) / 8,
+        )
+    }
+
+    fn serialize(
+        &self,
+        stats: &Stats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let line = self.line.unwrap();
+        let amplitude = self.max_deviation - self.min_deviation;
+        let num_bits = compute_num_bits(amplitude);
+        let linear_params = LinearParams {
+            line,
+            bit_unpacker: BitUnpacker::new(num_bits),
+        };
+        linear_params.serialize(wrt)?;
+        let mut bit_packer = BitPacker::new();
+        for (pos, value) in vals.enumerate() {
+            let calculated_value = line.eval(pos as u32);
+            let offset = value.wrapping_sub(calculated_value);
+            bit_packer.write(offset, num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+
+    fn collect(&mut self, value: u64) {
+        if let Some(line) = self.line {
+            self.collect_after_line_estimation(&line, value);
+        } else {
+            self.collect_before_line_estimation(value);
+        }
+    }
+}
+
+impl LinearCodecEstimator {
+    #[inline]
+    fn collect_after_line_estimation(&mut self, line: &Line, value: u64) {
+        let interpoled_val: u64 = line.eval(self.row_id);
+        let deviation = value.wrapping_add(HALF_SPACE).wrapping_sub(interpoled_val);
+        self.min_deviation = self.min_deviation.min(deviation);
+        self.max_deviation = self.max_deviation.max(deviation);
+        if self.row_id == 0 {
+            self.first_val = value;
+        }
+        self.last_val = value;
+        self.row_id += 1u32;
+    }
+
+    #[inline]
+    fn collect_before_line_estimation(&mut self, value: u64) {
+        self.block.push(value);
+        if self.block.len() == LINE_ESTIMATION_BLOCK_LEN {
+            let line = Line::train(&VecColumn::from(&self.block));
+            let block = std::mem::take(&mut self.block);
+            for val in block {
+                self.collect_after_line_estimation(&line, val);
+            }
+            self.line = Some(line);
+        }
+    }
+}
+
+impl ColumnCodec for LinearCodec {
+    type Reader = LinearReader;
+
+    type Estimator = LinearCodecEstimator;
+
+    fn load(mut data: OwnedBytes) -> io::Result<Self::Reader> {
+        let stats = Stats::deserialize(&mut data)?;
+        let linear_params = LinearParams::deserialize(&mut data)?;
+        Ok(LinearReader {
+            stats,
+            linear_params,
+            data,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use rand::RngCore;
+
+    use super::*;
+    use crate::column_values::u64_based::tests::{create_and_validate, get_codec_test_datasets};
+
+    #[test]
+    fn test_compression_simple() {
+        let vals = (100u64..)
+            .take(super::LINE_ESTIMATION_BLOCK_LEN)
+            .collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&vals, "simple monotonically large").unwrap();
+    }
+
+    #[test]
+    fn test_compression() {
+        let data = (10..=6_000_u64).collect::<Vec<_>>();
+        let (estimate, actual_compression) =
+            create_and_validate::<LinearCodec>(&data, "simple monotonically large").unwrap();
+        assert_le!(actual_compression, 0.001);
+        assert_le!(estimate, 0.02);
+    }
+
+    #[test]
+    fn test_with_codec_datasets() {
+        let data_sets = get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<LinearCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<LinearCodec>(&data, name);
+        }
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_large_amplitude() {
+        let data = vec![
+            i64::MAX as u64 / 2,
+            i64::MAX as u64 / 3,
+            i64::MAX as u64 / 2,
+        ];
+        create_and_validate::<LinearCodec>(&data, "large amplitude");
+    }
+
+    #[test]
+    fn overflow_error_test() {
+        let data = vec![1572656989877777, 1170935903116329, 720575940379279, 0];
+        create_and_validate::<LinearCodec>(&data, "overflow test");
+    }
+
+    #[test]
+    fn linear_interpol_fast_concave_data() {
+        let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
+        create_and_validate::<LinearCodec>(&data, "concave data");
+    }
+    #[test]
+    fn linear_interpol_fast_convex_data() {
+        let data = vec![0, 40, 60, 70, 75, 77];
+        create_and_validate::<LinearCodec>(&data, "convex data");
+    }
+    #[test]
+    fn linear_interpol_fast_field_test_simple() {
+        let data = (10..=20_u64).collect::<Vec<_>>();
+        create_and_validate::<LinearCodec>(&data, "simple monotonically");
+    }
+
+    #[test]
+    fn linear_interpol_fast_field_rand() {
+        let mut rng = rand::thread_rng();
+        for _ in 0..50 {
+            let mut data = (0..10_000).map(|_| rng.next_u64()).collect::<Vec<_>>();
+            create_and_validate::<LinearCodec>(&data, "random");
+            data.reverse();
+            create_and_validate::<LinearCodec>(&data, "random");
+        }
+    }
+}

columnar/src/column_values/u64_based/mod.rs

@@ -0,0 +1,183 @@
+mod bitpacked;
+mod blockwise_linear;
+mod line;
+mod linear;
+mod stats_collector;
+
+use std::io;
+use std::io::Write;
+use std::sync::Arc;
+
+use common::{BinarySerializable, OwnedBytes};
+
+use crate::column_index::MultiValueIndex;
+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};
+use crate::iterable::Iterable;
+use crate::{ColumnValues, MonotonicallyMappableToU64};
+
+pub trait ColumnCodecEstimator<T = u64>: 'static {
+    fn collect(&mut self, value: u64);
+    fn estimate(&self, stats: &Stats) -> Option<u64>;
+    fn finalize(&mut self) {}
+    fn serialize(
+        &self,
+        stats: &Stats,
+        vals: &mut dyn Iterator<Item = T>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()>;
+}
+
+pub trait ColumnCodec<T: PartialOrd = u64> {
+    type Reader: ColumnValues<T> + 'static;
+    type Estimator: ColumnCodecEstimator + Default;
+
+    fn load(bytes: OwnedBytes) -> io::Result<Self::Reader>;
+
+    fn estimator() -> Self::Estimator {
+        Self::Estimator::default()
+    }
+    fn boxed_estimator() -> Box<dyn ColumnCodecEstimator> {
+        Box::new(Self::estimator())
+    }
+}
+
+/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+pub enum CodecType {
+    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
+    /// `column.max_value() - column.min_value()`
+    Bitpacked = 0u8,
+    /// Linear interpolation puts a line between the first and last value and then bitpacks the
+    /// values by the offset from the line. The number of bits is defined by the max deviation from
+    /// the line.
+    Linear = 1u8,
+    /// Same as [`CodecType::Linear`], but encodes in blocks of 512 elements.
+    BlockwiseLinear = 2u8,
+}
+
+pub const ALL_U64_CODEC_TYPES: [CodecType; 3] = [
+    CodecType::Bitpacked,
+    CodecType::Linear,
+    CodecType::BlockwiseLinear,
+];
+
+impl CodecType {
+    fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    fn try_from_code(code: u8) -> Option<CodecType> {
+        match code {
+            0u8 => Some(CodecType::Bitpacked),
+            1u8 => Some(CodecType::Linear),
+            2u8 => Some(CodecType::BlockwiseLinear),
+            _ => None,
+        }
+    }
+
+    fn load<T: MonotonicallyMappableToU64>(
+        &self,
+        bytes: OwnedBytes,
+    ) -> io::Result<Arc<dyn ColumnValues<T>>> {
+        match self {
+            CodecType::Bitpacked => load_specific_codec::<BitpackedCodec, T>(bytes),
+            CodecType::Linear => load_specific_codec::<LinearCodec, T>(bytes),
+            CodecType::BlockwiseLinear => load_specific_codec::<BlockwiseLinearCodec, T>(bytes),
+        }
+    }
+}
+
+fn load_specific_codec<C: ColumnCodec, T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let reader = C::load(bytes)?;
+    let reader_typed = monotonic_map_column(
+        reader,
+        StrictlyMonotonicMappingInverter::from(StrictlyMonotonicMappingToInternal::<T>::new()),
+    );
+    Ok(Arc::new(reader_typed))
+}
+
+impl CodecType {
+    pub fn estimator(&self) -> Box<dyn ColumnCodecEstimator> {
+        match self {
+            CodecType::Bitpacked => BitpackedCodec::boxed_estimator(),
+            CodecType::Linear => LinearCodec::boxed_estimator(),
+            CodecType::BlockwiseLinear => BlockwiseLinearCodec::boxed_estimator(),
+        }
+    }
+}
+
+pub fn serialize_u64_based_column_values<'a, T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable<T>,
+    codec_types: &[CodecType],
+    wrt: &mut dyn Write,
+) -> io::Result<()> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimators: Vec<(CodecType, Box<dyn ColumnCodecEstimator>)> =
+        Vec::with_capacity(codec_types.len());
+    for &codec_type in codec_types {
+        estimators.push((codec_type, codec_type.estimator()));
+    }
+    for val in vals.boxed_iter() {
+        let val_u64 = val.to_u64();
+        stats_collector.collect(val_u64);
+        for (_, estimator) in &mut estimators {
+            estimator.collect(val_u64);
+        }
+    }
+    for (_, estimator) in &mut estimators {
+        estimator.finalize();
+    }
+    let stats = stats_collector.stats();
+    let (_, best_codec, best_codec_estimator) = estimators
+        .into_iter()
+        .flat_map(|(codec_type, estimator)| {
+            let num_bytes = estimator.estimate(&stats)?;
+            Some((num_bytes, codec_type, estimator))
+        })
+        .min_by_key(|(num_bytes, _, _)| *num_bytes)
+        .ok_or_else(|| {
+            io::Error::new(io::ErrorKind::InvalidData, "No available applicable codec.")
+        })?;
+    best_codec.to_code().serialize(wrt)?;
+    best_codec_estimator.serialize(
+        &stats,
+        &mut vals.boxed_iter().map(MonotonicallyMappableToU64::to_u64),
+        wrt,
+    )?;
+    Ok(())
+}
+
+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)
+        .copied()
+        .and_then(CodecType::try_from_code)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Failed to read codec type"))?;
+    bytes.advance(1);
+    codec_type.load(bytes)
+}
+
+/// Helper function to serialize a column (autodetect from all codecs) and then open it
+pub fn serialize_and_load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable,
+    codec_types: &[CodecType],
+) -> Arc<dyn ColumnValues<T>> {
+    let mut buffer = Vec::new();
+    serialize_u64_based_column_values(vals, codec_types, &mut buffer).unwrap();
+    load_u64_based_column_values::<T>(OwnedBytes::new(buffer)).unwrap()
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_values/u64_based/stats_collector.rs

@@ -0,0 +1,200 @@
+use std::num::NonZeroU64;
+
+use fastdivide::DividerU64;
+
+use crate::column_values::Stats;
+use crate::RowId;
+
+/// Compute the gcd of two non null numbers.
+///
+/// It is recommended, but not required, to feed values such that `large >= small`.
+fn compute_gcd(mut large: NonZeroU64, mut small: NonZeroU64) -> NonZeroU64 {
+    loop {
+        let rem: u64 = large.get() % small;
+        if let Some(new_small) = NonZeroU64::new(rem) {
+            (large, small) = (small, new_small);
+        } else {
+            return small;
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct StatsCollector {
+    min_max_opt: Option<(u64, u64)>,
+    num_rows: RowId,
+    // We measure the GCD of the difference between the values and the minimal value.
+    // This is the same as computing the difference between the values and the first value.
+    //
+    // This way, we can compress i64-converted-to-u64 (e.g. timestamp that were supplied in
+    // seconds, only to be converted in microseconds).
+    increment_gcd_opt: Option<(NonZeroU64, DividerU64)>,
+    first_value_opt: Option<u64>,
+}
+
+impl StatsCollector {
+    pub fn stats(&self) -> Stats {
+        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 {
+            min_value,
+            max_value,
+            num_rows: self.num_rows,
+            gcd: increment_gcd,
+        }
+    }
+
+    #[inline]
+    fn update_increment_gcd(&mut self, value: u64) {
+        let Some(first_value) = self.first_value_opt else {
+            // We set the first value and just quit.
+            self.first_value_opt = Some(value);
+            return;
+        };
+        let Some(non_zero_value) = NonZeroU64::new(value.abs_diff(first_value)) else {
+            // We can simply skip 0 values.
+            return;
+        };
+        let Some((gcd, gcd_divider)) = self.increment_gcd_opt else {
+            self.set_increment_gcd(non_zero_value);
+            return;
+        };
+        if gcd.get() == 1 {
+            // It won't see any update now.
+            return;
+        }
+        let remainder =
+            non_zero_value.get() - (gcd_divider.divide(non_zero_value.get())) * gcd.get();
+        if remainder == 0 {
+            return;
+        }
+        let new_gcd = compute_gcd(non_zero_value, gcd);
+        self.set_increment_gcd(new_gcd);
+    }
+
+    fn set_increment_gcd(&mut self, gcd: NonZeroU64) {
+        let new_divider = DividerU64::divide_by(gcd.get());
+        self.increment_gcd_opt = Some((gcd, new_divider));
+    }
+
+    pub fn collect(&mut self, value: u64) {
+        self.min_max_opt = Some(if let Some((min, max)) = self.min_max_opt {
+            (min.min(value), max.max(value))
+        } else {
+            (value, value)
+        });
+        self.num_rows += 1;
+        self.update_increment_gcd(value);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use crate::column_values::u64_based::stats_collector::{compute_gcd, StatsCollector};
+    use crate::column_values::u64_based::Stats;
+
+    fn compute_stats(vals: impl Iterator<Item = u64>) -> Stats {
+        let mut stats_collector = StatsCollector::default();
+        for val in vals {
+            stats_collector.collect(val);
+        }
+        stats_collector.stats()
+    }
+
+    fn find_gcd(vals: impl Iterator<Item = u64>) -> u64 {
+        compute_stats(vals).gcd.get()
+    }
+
+    #[test]
+    fn test_compute_gcd() {
+        let test_compute_gcd_aux = |large, small, expected| {
+            let large = NonZeroU64::new(large).unwrap();
+            let small = NonZeroU64::new(small).unwrap();
+            let expected = NonZeroU64::new(expected).unwrap();
+            assert_eq!(compute_gcd(small, large), expected);
+            assert_eq!(compute_gcd(large, small), expected);
+        };
+        test_compute_gcd_aux(1, 4, 1);
+        test_compute_gcd_aux(2, 4, 2);
+        test_compute_gcd_aux(10, 25, 5);
+        test_compute_gcd_aux(25, 25, 25);
+    }
+
+    #[test]
+    fn test_gcd() {
+        assert_eq!(find_gcd([0].into_iter()), 1);
+        assert_eq!(find_gcd([0, 10].into_iter()), 10);
+        assert_eq!(find_gcd([10, 0].into_iter()), 10);
+        assert_eq!(find_gcd([].into_iter()), 1);
+        assert_eq!(find_gcd([15, 30, 5, 10].into_iter()), 5);
+        assert_eq!(find_gcd([15, 16, 10].into_iter()), 1);
+        assert_eq!(find_gcd([0, 5, 5, 5].into_iter()), 5);
+        assert_eq!(find_gcd([0, 0].into_iter()), 1);
+        assert_eq!(find_gcd([1, 10, 4, 1, 7, 10].into_iter()), 3);
+        assert_eq!(find_gcd([1, 10, 0, 4, 1, 7, 10].into_iter()), 1);
+    }
+
+    #[test]
+    fn test_stats() {
+        assert_eq!(
+            compute_stats([].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([0, 1].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 1,
+                num_rows: 2
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 20, 30].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 10,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 50, 10, 30].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(20).unwrap(),
+                min_value: 10,
+                max_value: 50,
+                num_rows: 4
+            }
+        );
+        assert_eq!(
+            compute_stats([10, 0, 30].into_iter()),
+            Stats {
+                gcd: NonZeroU64::new(10).unwrap(),
+                min_value: 0,
+                max_value: 30,
+                num_rows: 3
+            }
+        );
+    }
+}

columnar/src/column_values/u64_based/tests.rs

@@ -2,53 +2,88 @@ use proptest::prelude::*;
 use proptest::strategy::Strategy;
 use proptest::{prop_oneof, proptest};
 
-use super::bitpacked::BitpackedCodec;
-use super::blockwise_linear::BlockwiseLinearCodec;
-use super::linear::LinearCodec;
-use super::serialize::Header;
-
-pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
-    data: &[u64],
+#[test]
+fn test_serialize_and_load_simple() {
+    let mut buffer = Vec::new();
+    let vals = &[1u64, 2u64, 5u64];
+    serialize_u64_based_column_values(
+        &&vals[..],
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        &mut buffer,
+    )
+    .unwrap();
+    assert_eq!(buffer.len(), 7);
+    let col = load_u64_based_column_values::<u64>(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(col.num_vals(), 3);
+    assert_eq!(col.get_val(0), 1);
+    assert_eq!(col.get_val(1), 2);
+    assert_eq!(col.get_val(2), 5);
+}
+pub(crate) fn create_and_validate<TColumnCodec: ColumnCodec>(
+    vals: &[u64],
     name: &str,
 ) -> Option<(f32, f32)> {
-    let col = &VecColumn::from(data);
-    let header = Header::compute_header(col, &[Codec::CODEC_TYPE])?;
-    let normalized_col = header.normalize_column(col);
-    let estimation = Codec::estimate(&normalized_col)?;
+    let mut stats_collector = StatsCollector::default();
+    let mut codec_estimator: TColumnCodec::Estimator = Default::default();
 
-    let mut out = Vec::new();
-    let col = VecColumn::from(data);
-    serialize_column_values(&col, &[Codec::CODEC_TYPE], &mut out).unwrap();
+    for val in vals.boxed_iter() {
+        stats_collector.collect(val);
+        codec_estimator.collect(val);
+    }
+    codec_estimator.finalize();
+    let stats = stats_collector.stats();
+    let estimation = codec_estimator.estimate(&stats)?;
 
-    let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
+    let mut buffer = Vec::new();
+    codec_estimator
+        .serialize(&stats, vals.boxed_iter().as_mut(), &mut buffer)
+        .unwrap();
 
-    let reader = super::open_u64_mapped::<u64>(OwnedBytes::new(out)).unwrap();
-    assert_eq!(reader.num_vals(), data.len() as u32);
-    for (doc, orig_val) in data.iter().copied().enumerate() {
+    let actual_compression = buffer.len() as u64;
+
+    let reader = TColumnCodec::load(OwnedBytes::new(buffer)).unwrap();
+    assert_eq!(reader.num_vals(), vals.len() as u32);
+    for (doc, orig_val) in vals.iter().copied().enumerate() {
         let val = reader.get_val(doc as u32);
         assert_eq!(
             val, orig_val,
-            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{data:?}`",
+            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{vals:?}`",
         );
     }
 
-    if !data.is_empty() {
-        let test_rand_idx = rand::thread_rng().gen_range(0..=data.len() - 1);
-        let expected_positions: Vec<u32> = data
+    if !vals.is_empty() {
+        let test_rand_idx = rand::thread_rng().gen_range(0..=vals.len() - 1);
+        let expected_positions: Vec<u32> = vals
             .iter()
             .enumerate()
-            .filter(|(_, el)| **el == data[test_rand_idx])
+            .filter(|(_, el)| **el == vals[test_rand_idx])
             .map(|(pos, _)| pos as u32)
             .collect();
         let mut positions = Vec::new();
         reader.get_docids_for_value_range(
-            data[test_rand_idx]..=data[test_rand_idx],
-            0..data.len() as u32,
+            vals[test_rand_idx]..=vals[test_rand_idx],
+            0..vals.len() as u32,
             &mut positions,
         );
         assert_eq!(expected_positions, positions);
     }
-    Some((estimation, actual_compression))
+    if actual_compression > 1000 {
+        assert!(relative_difference(estimation, actual_compression) < 0.10f32);
+    }
+    Some((
+        compression_rate(estimation, stats.num_rows),
+        compression_rate(actual_compression, stats.num_rows),
+    ))
+}
+
+fn compression_rate(num_bytes: u64, num_values: u32) -> f32 {
+    num_bytes as f32 / (num_values as f32 * 8.0)
+}
+
+fn relative_difference(left: u64, right: u64) -> f32 {
+    let left = left as f32;
+    let right = right as f32;
+    2.0f32 * (left - right).abs() / (left + right)
 }
 
 proptest! {
@@ -64,12 +99,21 @@ proptest! {
         create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
     }
 
+
     #[test]
     fn test_proptest_small_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
         create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
     }
 }
 
+#[test]
+fn test_small_blockwise_linear_example() {
+    create_and_validate::<BlockwiseLinearCodec>(
+        &[9223372036854775808, 9223370937344622593],
+        "proptest multilinearinterpol",
+    );
+}
+
 proptest! {
     #![proptest_config(ProptestConfig::with_cases(10))]
 
@@ -118,8 +162,8 @@ pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
     data_and_names
 }
 
-fn test_codec<C: FastFieldCodec>() {
-    let codec_name = format!("{:?}", C::CODEC_TYPE);
+fn test_codec<C: ColumnCodec>() {
+    let codec_name = std::any::type_name::<C>();
     for (data, dataset_name) in get_codec_test_datasets() {
         let estimate_actual_opt: Option<(f32, f32)> =
             tests::create_and_validate::<C>(&data, dataset_name);
@@ -146,53 +190,48 @@ fn test_codec_multi_interpolation() {
 
 use super::*;
 
+fn estimate<C: ColumnCodec>(vals: &[u64]) -> Option<f32> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimator = C::Estimator::default();
+    for &val in vals {
+        stats_collector.collect(val);
+        estimator.collect(val);
+    }
+    estimator.finalize();
+    let stats = stats_collector.stats();
+    let num_bytes = estimator.estimate(&stats)?;
+    if stats.num_rows == 0 {
+        return None;
+    }
+    Some(num_bytes as f32 / (8.0 * stats.num_rows as f32))
+}
+
 #[test]
 fn estimation_good_interpolation_case() {
     let data = (10..=20000_u64).collect::<Vec<_>>();
-    let data: VecColumn = data.as_slice().into();
 
-    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    let linear_interpol_estimation = estimate::<LinearCodec>(&data).unwrap();
     assert_le!(linear_interpol_estimation, 0.01);
 
-    let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
+    let multi_linear_interpol_estimation = estimate::<BlockwiseLinearCodec>(&data).unwrap();
     assert_le!(multi_linear_interpol_estimation, 0.2);
     assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
 
-    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
     assert_lt!(linear_interpol_estimation, bitpacked_estimation);
 }
-#[test]
-fn estimation_test_bad_interpolation_case() {
-    let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
-
-    let data: VecColumn = data.into();
-    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-    assert_le!(linear_interpol_estimation, 0.34);
-
-    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
-}
-
-#[test]
-fn estimation_prefer_bitpacked() {
-    let data = VecColumn::from(&[10, 10, 10, 10]);
-    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
-}
 
 #[test]
 fn estimation_test_bad_interpolation_case_monotonically_increasing() {
     let mut data: Vec<u64> = (201..=20000_u64).collect();
     data.push(1_000_000);
-    let data: VecColumn = data.as_slice().into();
 
     // in this case the linear interpolation can't in fact not be worse than bitpacking,
     // but the estimator adds some threshold, which leads to estimated worse behavior
-    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    let linear_interpol_estimation = estimate::<LinearCodec>(&data[..]).unwrap();
     assert_le!(linear_interpol_estimation, 0.35);
 
-    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
     assert_le!(bitpacked_estimation, 0.32);
     assert_le!(bitpacked_estimation, linear_interpol_estimation);
 }
@@ -201,7 +240,7 @@ fn estimation_test_bad_interpolation_case_monotonically_increasing() {
 fn test_fast_field_codec_type_to_code() {
     let mut count_codec = 0;
     for code in 0..=255 {
-        if let Some(codec_type) = FastFieldCodecType::from_code(code) {
+        if let Some(codec_type) = CodecType::try_from_code(code) {
             assert_eq!(codec_type.to_code(), code);
             count_codec += 1;
         }
@@ -209,19 +248,16 @@ fn test_fast_field_codec_type_to_code() {
     assert_eq!(count_codec, 3);
 }
 
-fn test_fastfield_gcd_i64_with_codec(
-    codec_type: FastFieldCodecType,
-    num_vals: usize,
-) -> io::Result<()> {
+fn test_fastfield_gcd_i64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
     let mut vals: Vec<i64> = (-4..=(num_vals as i64) - 5).map(|val| val * 1000).collect();
     let mut buffer: Vec<u8> = Vec::new();
-    crate::column_values::serialize_column_values(
-        &VecColumn::from(&vals),
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
         &[codec_type],
         &mut buffer,
     )?;
     let buffer = OwnedBytes::new(buffer);
-    let column = crate::column_values::open_u64_mapped::<i64>(buffer.clone())?;
+    let column = crate::column_values::load_u64_based_column_values::<i64>(buffer.clone())?;
     assert_eq!(column.get_val(0), -4000i64);
     assert_eq!(column.get_val(1), -3000i64);
     assert_eq!(column.get_val(2), -2000i64);
@@ -232,8 +268,8 @@ fn test_fastfield_gcd_i64_with_codec(
     let mut buffer_without_gcd = Vec::new();
     vals.pop();
     vals.push(1001i64);
-    crate::column_values::serialize_column_values(
-        &VecColumn::from(&vals),
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
         &[codec_type],
         &mut buffer_without_gcd,
     )?;
@@ -246,28 +282,25 @@ fn test_fastfield_gcd_i64_with_codec(
 #[test]
 fn test_fastfield_gcd_i64() -> io::Result<()> {
     for &codec_type in &[
-        FastFieldCodecType::Bitpacked,
-        FastFieldCodecType::BlockwiseLinear,
-        FastFieldCodecType::Linear,
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::Linear,
     ] {
         test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
     }
     Ok(())
 }
 
-fn test_fastfield_gcd_u64_with_codec(
-    codec_type: FastFieldCodecType,
-    num_vals: usize,
-) -> io::Result<()> {
+fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
     let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
     let mut buffer: Vec<u8> = Vec::new();
-    crate::column_values::serialize_column_values(
-        &VecColumn::from(&vals),
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
         &[codec_type],
         &mut buffer,
     )?;
     let buffer = OwnedBytes::new(buffer);
-    let column = crate::column_values::open_u64_mapped::<u64>(buffer.clone())?;
+    let column = crate::column_values::load_u64_based_column_values::<u64>(buffer.clone())?;
     assert_eq!(column.get_val(0), 1000u64);
     assert_eq!(column.get_val(1), 2000u64);
     assert_eq!(column.get_val(2), 3000u64);
@@ -278,8 +311,8 @@ fn test_fastfield_gcd_u64_with_codec(
     let mut buffer_without_gcd = Vec::new();
     vals.pop();
     vals.push(1001u64);
-    crate::column_values::serialize_column_values(
-        &VecColumn::from(&vals),
+    crate::column_values::serialize_u64_based_column_values(
+        &&vals[..],
         &[codec_type],
         &mut buffer_without_gcd,
     )?;
@@ -291,9 +324,9 @@ fn test_fastfield_gcd_u64_with_codec(
 #[test]
 fn test_fastfield_gcd_u64() -> io::Result<()> {
     for &codec_type in &[
-        FastFieldCodecType::Bitpacked,
-        FastFieldCodecType::BlockwiseLinear,
-        FastFieldCodecType::Linear,
+        CodecType::Bitpacked,
+        CodecType::BlockwiseLinear,
+        CodecType::Linear,
     ] {
         test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
     }
@@ -302,7 +335,10 @@ fn test_fastfield_gcd_u64() -> io::Result<()> {
 
 #[test]
 pub fn test_fastfield2() {
-    let test_fastfield = crate::column_values::serialize_and_load(&[100u64, 200u64, 300u64]);
+    let test_fastfield = crate::column_values::serialize_and_load_u64_based_column_values::<u64>(
+        &&[100u64, 200u64, 300u64][..],
+        &ALL_U64_CODEC_TYPES,
+    );
     assert_eq!(test_fastfield.get_val(0), 100);
     assert_eq!(test_fastfield.get_val(1), 200);
     assert_eq!(test_fastfield.get_val(2), 300);

columnar/src/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.
-#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy)]
+/// 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)
     }
 }
 
@@ -65,18 +52,6 @@ impl From<NumericalType> for ColumnType {
 }
 
 impl ColumnType {
-    /// get column type category
-    pub(crate) fn column_type_category(self) -> ColumnTypeCategory {
-        match self {
-            ColumnType::I64 | ColumnType::U64 | 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 numerical_type(&self) -> Option<NumericalType> {
         match self {
             ColumnType::I64 => Some(NumericalType::I64),
@@ -155,70 +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/merge.rs

@@ -1,176 +0,0 @@
-use std::collections::HashMap;
-use std::io;
-
-use super::column_type::ColumnTypeCategory;
-use crate::columnar::ColumnarReader;
-use crate::dynamic_column::DynamicColumn;
-
-pub enum MergeDocOrder {
-    /// Columnar tables are simply stacked one above the other.
-    /// If the i-th columnar_readers has n_rows_i rows, then
-    /// in the resulting columnar,
-    /// rows [r0..n_row_0) contains the row of columnar_readers[0], in ordder
-    /// rows [n_row_0..n_row_0 + n_row_1 contains the row of columnar_readers[1], in order.
-    /// ..
-    Stack,
-    /// Some more complex mapping, that can interleaves rows from the different readers and
-    /// possibly drop rows.
-    Complex(()),
-}
-
-pub fn merge_columnar(
-    _columnar_readers: &[ColumnarReader],
-    mapping: MergeDocOrder,
-    _output: &mut impl io::Write,
-) -> io::Result<()> {
-    match mapping {
-        MergeDocOrder::Stack => {
-            // implement me :)
-            todo!();
-        }
-        MergeDocOrder::Complex(_) => {
-            // for later
-            todo!();
-        }
-    }
-}
-
-pub fn collect_columns(
-    columnar_readers: &[&ColumnarReader],
-) -> io::Result<HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>>> {
-    // Each column name may have multiple types of column associated.
-    // For merging we are interested in the same column type category since they can be merged.
-    let mut field_name_to_group: HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>> =
-        HashMap::new();
-
-    for columnar_reader in columnar_readers {
-        let column_name_and_handle = columnar_reader.list_columns()?;
-        for (column_name, handle) in column_name_and_handle {
-            let column_type_to_handles = field_name_to_group
-                .entry(column_name.to_string())
-                .or_default();
-
-            let columns = column_type_to_handles
-                .entry(handle.column_type().column_type_category())
-                .or_default();
-            columns.push(handle.open()?);
-        }
-    }
-
-    normalize_columns(&mut field_name_to_group);
-
-    Ok(field_name_to_group)
-}
-
-/// Cast numerical type columns to the same type
-pub(crate) fn normalize_columns(
-    map: &mut HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>>,
-) {
-    for (_field_name, type_category_to_columns) in map.iter_mut() {
-        for (type_category, columns) in type_category_to_columns {
-            if type_category == &ColumnTypeCategory::Numerical {
-                let casted_columns = cast_to_common_numerical_column(&columns);
-                *columns = casted_columns;
-            }
-        }
-    }
-}
-
-/// Receives a list of columns of numerical types (u64, i64, f64)
-///
-/// Returns a list of `DynamicColumn` which are all of the same numerical type
-fn cast_to_common_numerical_column(columns: &[DynamicColumn]) -> Vec<DynamicColumn> {
-    assert!(columns
-        .iter()
-        .all(|column| column.column_type().numerical_type().is_some()));
-    let coerce_to_i64: Vec<_> = columns
-        .iter()
-        .map(|column| column.clone().coerce_to_i64())
-        .collect();
-
-    if coerce_to_i64.iter().all(|column| column.is_some()) {
-        return coerce_to_i64
-            .into_iter()
-            .map(|column| column.unwrap())
-            .collect();
-    }
-
-    let coerce_to_u64: Vec<_> = columns
-        .iter()
-        .map(|column| column.clone().coerce_to_u64())
-        .collect();
-
-    if coerce_to_u64.iter().all(|column| column.is_some()) {
-        return coerce_to_u64
-            .into_iter()
-            .map(|column| column.unwrap())
-            .collect();
-    }
-
-    columns
-        .iter()
-        .map(|column| {
-            column
-                .clone()
-                .coerce_to_f64()
-                .expect("couldn't cast column to f64")
-        })
-        .collect()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::ColumnarWriter;
-
-    #[test]
-    fn test_column_coercion() {
-        // i64 type
-        let columnar1 = {
-            let mut dataframe_writer = ColumnarWriter::default();
-            dataframe_writer.record_numerical(1u32, "numbers", 1i64);
-            let mut buffer: Vec<u8> = Vec::new();
-            dataframe_writer.serialize(2, &mut buffer).unwrap();
-            ColumnarReader::open(buffer).unwrap()
-        };
-        // u64 type
-        let columnar2 = {
-            let mut dataframe_writer = ColumnarWriter::default();
-            dataframe_writer.record_numerical(1u32, "numbers", u64::MAX - 100);
-            let mut buffer: Vec<u8> = Vec::new();
-            dataframe_writer.serialize(2, &mut buffer).unwrap();
-            ColumnarReader::open(buffer).unwrap()
-        };
-
-        // f64 type
-        let columnar3 = {
-            let mut dataframe_writer = ColumnarWriter::default();
-            dataframe_writer.record_numerical(1u32, "numbers", 30.5);
-            let mut buffer: Vec<u8> = Vec::new();
-            dataframe_writer.serialize(2, &mut buffer).unwrap();
-            ColumnarReader::open(buffer).unwrap()
-        };
-
-        let column_map = collect_columns(&[&columnar1, &columnar2, &columnar3]).unwrap();
-        assert_eq!(column_map.len(), 1);
-        let cat_to_columns = column_map.get("numbers").unwrap();
-        assert_eq!(cat_to_columns.len(), 1);
-
-        let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
-        assert!(numerical.iter().all(|column| column.is_f64()));
-
-        let column_map = collect_columns(&[&columnar1, &columnar1]).unwrap();
-        assert_eq!(column_map.len(), 1);
-        let cat_to_columns = column_map.get("numbers").unwrap();
-        assert_eq!(cat_to_columns.len(), 1);
-        let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
-        assert!(numerical.iter().all(|column| column.is_i64()));
-
-        let column_map = collect_columns(&[&columnar2, &columnar2]).unwrap();
-        assert_eq!(column_map.len(), 1);
-        let cat_to_columns = column_map.get("numbers").unwrap();
-        assert_eq!(cat_to_columns.len(), 1);
-        let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
-        assert!(numerical.iter().all(|column| column.is_u64()));
-    }
-}

columnar/src/columnar/merge/merge_dict_column.rs

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

columnar/src/columnar/merge/merge_mapping.rs

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

columnar/src/columnar/merge/mod.rs

@@ -0,0 +1,271 @@
+mod merge_dict_column;
+mod merge_mapping;
+mod term_merger;
+
+// mod sorted_doc_id_column;
+
+use std::collections::{BTreeMap, HashMap, HashSet};
+use std::io;
+use std::net::Ipv6Addr;
+use std::sync::Arc;
+
+pub use merge_mapping::{MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
+
+use super::writer::ColumnarSerializer;
+use crate::column::{serialize_column_mappable_to_u128, serialize_column_mappable_to_u64};
+use crate::column_values::MergedColumnValues;
+use crate::columnar::merge::merge_dict_column::merge_bytes_or_str_column;
+use crate::columnar::writer::CompatibleNumericalTypes;
+use crate::columnar::ColumnarReader;
+use crate::dynamic_column::DynamicColumn;
+use crate::{
+    BytesColumn, Column, ColumnIndex, ColumnType, ColumnValues, NumericalType, NumericalValue,
+};
+
+/// Column types are grouped into different categories.
+/// After merge, all columns belonging to the same category are coerced to
+/// the same column type.
+///
+/// In practise, today, only Numerical colummns are coerced into one type today.
+///
+/// See also [README.md].
+#[derive(Copy, Clone, Eq, PartialEq, Hash, Debug)]
+enum ColumnTypeCategory {
+    Bool,
+    Str,
+    Numerical,
+    DateTime,
+    Bytes,
+    IpAddr,
+}
+
+impl From<ColumnType> for ColumnTypeCategory {
+    fn from(column_type: ColumnType) -> Self {
+        match column_type {
+            ColumnType::I64 => ColumnTypeCategory::Numerical,
+            ColumnType::U64 => ColumnTypeCategory::Numerical,
+            ColumnType::F64 => ColumnTypeCategory::Numerical,
+            ColumnType::Bytes => ColumnTypeCategory::Bytes,
+            ColumnType::Str => ColumnTypeCategory::Str,
+            ColumnType::Bool => ColumnTypeCategory::Bool,
+            ColumnType::IpAddr => ColumnTypeCategory::IpAddr,
+            ColumnType::DateTime => ColumnTypeCategory::DateTime,
+        }
+    }
+}
+
+pub fn merge_columnar(
+    columnar_readers: &[&ColumnarReader],
+    merge_row_order: MergeRowOrder,
+    output: &mut impl io::Write,
+) -> io::Result<()> {
+    let mut serializer = ColumnarSerializer::new(output);
+
+    let columns_to_merge = group_columns_for_merge(columnar_readers)?;
+    for ((column_name, column_type), columns) in columns_to_merge {
+        let mut column_serializer =
+            serializer.serialize_column(column_name.as_bytes(), column_type);
+        merge_column(
+            column_type,
+            columns,
+            &merge_row_order,
+            &mut column_serializer,
+        )?;
+    }
+    serializer.finalize(merge_row_order.num_rows())?;
+
+    Ok(())
+}
+
+fn dynamic_column_to_u64_monotonic(dynamic_column: DynamicColumn) -> Option<Column<u64>> {
+    match dynamic_column {
+        DynamicColumn::Bool(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::I64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::U64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::F64(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::DateTime(column) => Some(column.to_u64_monotonic()),
+        DynamicColumn::IpAddr(_) | DynamicColumn::Bytes(_) | DynamicColumn::Str(_) => None,
+    }
+}
+
+fn merge_column(
+    column_type: ColumnType,
+    columns: Vec<Option<DynamicColumn>>,
+    merge_row_order: &MergeRowOrder,
+    wrt: &mut impl io::Write,
+) -> io::Result<()> {
+    match column_type {
+        ColumnType::I64
+        | ColumnType::U64
+        | ColumnType::F64
+        | ColumnType::DateTime
+        | ColumnType::Bool => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues>>> =
+                Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                if let Some(Column { idx, values }) =
+                    dynamic_column_opt.and_then(dynamic_column_to_u64_monotonic)
+                {
+                    column_indexes.push(Some(idx));
+                    column_values.push(Some(values));
+                } else {
+                    column_indexes.push(None);
+                    column_values.push(None);
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values[..],
+                merge_row_order,
+            };
+            serialize_column_mappable_to_u64(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::IpAddr => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut column_values: Vec<Option<Arc<dyn ColumnValues<Ipv6Addr>>>> =
+                Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                if let Some(DynamicColumn::IpAddr(Column { idx, values })) = dynamic_column_opt {
+                    column_indexes.push(Some(idx));
+                    column_values.push(Some(values));
+                } else {
+                    column_indexes.push(None);
+                    column_values.push(None);
+                }
+            }
+
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            let merge_column_values = MergedColumnValues {
+                column_indexes: &column_indexes[..],
+                column_values: &column_values,
+                merge_row_order,
+            };
+
+            serialize_column_mappable_to_u128(merged_column_index, &merge_column_values, wrt)?;
+        }
+        ColumnType::Bytes | ColumnType::Str => {
+            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
+            let mut bytes_columns: Vec<Option<BytesColumn>> = Vec::with_capacity(columns.len());
+            for dynamic_column_opt in columns {
+                match dynamic_column_opt {
+                    Some(DynamicColumn::Str(str_column)) => {
+                        column_indexes.push(Some(str_column.term_ord_column.idx.clone()));
+                        bytes_columns.push(Some(str_column.into()));
+                    }
+                    Some(DynamicColumn::Bytes(bytes_column)) => {
+                        column_indexes.push(Some(bytes_column.term_ord_column.idx.clone()));
+                        bytes_columns.push(Some(bytes_column));
+                    }
+                    _ => {
+                        column_indexes.push(None);
+                        bytes_columns.push(None);
+                    }
+                }
+            }
+            let merged_column_index =
+                crate::column_index::merge_column_index(&column_indexes[..], merge_row_order);
+            merge_bytes_or_str_column(merged_column_index, &bytes_columns, merge_row_order, wrt)?;
+        }
+    }
+    Ok(())
+}
+
+fn group_columns_for_merge(
+    columnar_readers: &[&ColumnarReader],
+) -> io::Result<BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>>> {
+    // Each column name may have multiple types of column associated.
+    // For merging we are interested in the same column type category since they can be merged.
+    let mut columns_grouped: HashMap<(String, ColumnTypeCategory), Vec<Option<DynamicColumn>>> =
+        HashMap::new();
+
+    let num_columnars = columnar_readers.len();
+
+    for (columnar_id, columnar_reader) in columnar_readers.iter().enumerate() {
+        let column_name_and_handle = columnar_reader.list_columns()?;
+        for (column_name, handle) in column_name_and_handle {
+            let column_type_category: ColumnTypeCategory = handle.column_type().into();
+            let columns = columns_grouped
+                .entry((column_name, column_type_category))
+                .or_insert_with(|| vec![None; num_columnars]);
+            let column = handle.open()?;
+            columns[columnar_id] = Some(column);
+        }
+    }
+
+    let mut merge_columns: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        BTreeMap::default();
+
+    for ((column_name, col_category), mut columns) in columns_grouped {
+        if col_category == ColumnTypeCategory::Numerical {
+            coerce_numerical_columns_to_same_type(&mut columns);
+        }
+        let column_type = columns
+            .iter()
+            .flatten()
+            .map(|col| col.column_type())
+            .next()
+            .unwrap();
+        merge_columns.insert((column_name, column_type), columns);
+    }
+
+    Ok(merge_columns)
+}
+
+/// Coerce a set of numerical columns to the same type.
+///
+/// If all columns are already from the same type, keep this type
+/// (even if they could all be coerced to i64).
+fn coerce_numerical_columns_to_same_type(columns: &mut [Option<DynamicColumn>]) {
+    let mut column_types: HashSet<NumericalType> = HashSet::default();
+    let mut compatible_numerical_types = CompatibleNumericalTypes::default();
+    for column in columns.iter().flatten() {
+        let min_value: NumericalValue;
+        let max_value: NumericalValue;
+        match column {
+            DynamicColumn::I64(column) => {
+                min_value = column.min_value().into();
+                max_value = column.max_value().into();
+            }
+            DynamicColumn::U64(column) => {
+                min_value = column.min_value().into();
+                max_value = column.min_value().into();
+            }
+            DynamicColumn::F64(column) => {
+                min_value = column.min_value().into();
+                max_value = column.min_value().into();
+            }
+            DynamicColumn::Bool(_)
+            | DynamicColumn::IpAddr(_)
+            | DynamicColumn::DateTime(_)
+            | DynamicColumn::Bytes(_)
+            | DynamicColumn::Str(_) => {
+                panic!("We expected only numerical columns.");
+            }
+        }
+        column_types.insert(column.column_type().numerical_type().unwrap());
+        compatible_numerical_types.accept_value(min_value);
+        compatible_numerical_types.accept_value(max_value);
+    }
+    if column_types.len() <= 1 {
+        // No need to do anything. The columns are already all from the same type.
+        // This is necessary to let use force a given type.
+
+        // TODO This works in a world where we do not allow a change of schema,
+        // but in the future, we will have to pass some kind of schema to enforce
+        // the logic.
+        return;
+    }
+    let coerce_type = compatible_numerical_types.to_numerical_type();
+    for column_opt in columns.iter_mut() {
+        if let Some(column) = column_opt.take() {
+            *column_opt = column.coerce_numerical(coerce_type);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/columnar/merge/term_merger.rs

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

columnar/src/columnar/merge/tests.rs

@@ -0,0 +1,258 @@
+use super::*;
+use crate::{Cardinality, ColumnarWriter, HasAssociatedColumnType, RowId};
+
+fn make_columnar<T: Into<NumericalValue> + HasAssociatedColumnType + Copy>(
+    column_name: &str,
+    vals: &[T],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    dataframe_writer.record_column_type(column_name, T::column_type(), false);
+    for (row_id, val) in vals.iter().copied().enumerate() {
+        dataframe_writer.record_numerical(row_id as RowId, column_name, val.into());
+    }
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(vals.len() as RowId, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_column_coercion_to_u64() {
+    // i64 type
+    let columnar1 = make_columnar("numbers", &[1i64]);
+    // u64 type
+    let columnar2 = make_columnar("numbers", &[u64::MAX]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_column_no_coercion_if_all_the_same() {
+    let columnar1 = make_columnar("numbers", &[1u64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::U64)));
+}
+
+#[test]
+fn test_column_coercion_to_i64() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2]).unwrap();
+    assert_eq!(column_map.len(), 1);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::I64)));
+}
+
+#[test]
+fn test_missing_column() {
+    let columnar1 = make_columnar("numbers", &[-1i64]);
+    let columnar2 = make_columnar("numbers2", &[2u64]);
+    let column_map: BTreeMap<(String, ColumnType), Vec<Option<DynamicColumn>>> =
+        group_columns_for_merge(&[&columnar1, &columnar2]).unwrap();
+    assert_eq!(column_map.len(), 2);
+    assert!(column_map.contains_key(&("numbers".to_string(), ColumnType::I64)));
+    {
+        let columns = column_map
+            .get(&("numbers".to_string(), ColumnType::I64))
+            .unwrap();
+        assert!(columns[0].is_some());
+        assert!(columns[1].is_none());
+    }
+    {
+        let columns = column_map
+            .get(&("numbers2".to_string(), ColumnType::U64))
+            .unwrap();
+        assert!(columns[0].is_none());
+        assert!(columns[1].is_some());
+    }
+}
+
+fn make_numerical_columnar_multiple_columns(
+    columns: &[(&str, &[&[NumericalValue]])],
+) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_numerical(row_id as u32, column_name, *val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+fn make_byte_columnar_multiple_columns(columns: &[(&str, &[&[&[u8]]])]) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_bytes(row_id as u32, column_name, *val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+fn make_text_columnar_multiple_columns(columns: &[(&str, &[&[&str]])]) -> ColumnarReader {
+    let mut dataframe_writer = ColumnarWriter::default();
+    for (column_name, column_values) in columns {
+        for (row_id, vals) in column_values.iter().enumerate() {
+            for val in vals.iter() {
+                dataframe_writer.record_str(row_id as u32, column_name, *val);
+            }
+        }
+    }
+    let num_rows = columns
+        .iter()
+        .map(|(_, val_rows)| val_rows.len() as RowId)
+        .max()
+        .unwrap_or(0u32);
+    let mut buffer: Vec<u8> = Vec::new();
+    dataframe_writer
+        .serialize(num_rows, None, &mut buffer)
+        .unwrap();
+    ColumnarReader::open(buffer).unwrap()
+}
+
+#[test]
+fn test_merge_columnar_numbers() {
+    let columnar1 =
+        make_numerical_columnar_multiple_columns(&[("numbers", &[&[NumericalValue::from(-1f64)]])]);
+    let columnar2 = make_numerical_columnar_multiple_columns(&[(
+        "numbers",
+        &[&[], &[NumericalValue::from(-3f64)]],
+    )]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("numbers").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::F64(vals) = dynamic_column else { panic!() };
+    assert_eq!(vals.get_cardinality(), Cardinality::Optional);
+    assert_eq!(vals.first(0u32), Some(-1f64));
+    assert_eq!(vals.first(1u32), None);
+    assert_eq!(vals.first(2u32), Some(-3f64));
+}
+
+#[test]
+fn test_merge_columnar_texts() {
+    let columnar1 = make_text_columnar_multiple_columns(&[("texts", &[&["a"]])]);
+    let columnar2 = make_text_columnar_multiple_columns(&[("texts", &[&[], &["b"]])]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 3);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("texts").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Str(vals) = dynamic_column else { panic!() };
+    let get_str_for_ord = |ord| {
+        let mut out = String::new();
+        vals.ord_to_str(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 2);
+    assert_eq!(get_str_for_ord(0), "a");
+    assert_eq!(get_str_for_ord(1), "b");
+
+    let get_str_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = String::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_str(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_str_for_row(0), "a");
+    assert_eq!(get_str_for_row(1), "");
+    assert_eq!(get_str_for_row(2), "b");
+}
+
+#[test]
+fn test_merge_columnar_byte() {
+    let columnar1 = make_byte_columnar_multiple_columns(&[("bytes", &[&[b"bbbb"], &[b"baaa"]])]);
+    let columnar2 = make_byte_columnar_multiple_columns(&[("bytes", &[&[], &[b"a"]])]);
+    let mut buffer = Vec::new();
+    let columnars = &[&columnar1, &columnar2];
+    let stack_merge_order = StackMergeOrder::stack(columnars);
+    crate::columnar::merge_columnar(
+        columnars,
+        MergeRowOrder::Stack(stack_merge_order),
+        &mut buffer,
+    )
+    .unwrap();
+    let columnar_reader = ColumnarReader::open(buffer).unwrap();
+    assert_eq!(columnar_reader.num_rows(), 4);
+    assert_eq!(columnar_reader.num_columns(), 1);
+    let cols = columnar_reader.read_columns("bytes").unwrap();
+    let dynamic_column = cols[0].open().unwrap();
+    let DynamicColumn::Bytes(vals) = dynamic_column else { panic!() };
+    let get_bytes_for_ord = |ord| {
+        let mut out = Vec::new();
+        vals.ord_to_bytes(ord, &mut out).unwrap();
+        out
+    };
+
+    assert_eq!(vals.dictionary.num_terms(), 3);
+    assert_eq!(get_bytes_for_ord(0), b"a");
+    assert_eq!(get_bytes_for_ord(1), b"baaa");
+    assert_eq!(get_bytes_for_ord(2), b"bbbb");
+
+    let get_bytes_for_row = |row_id| {
+        let term_ords: Vec<u64> = vals.term_ords(row_id).collect();
+        assert!(term_ords.len() <= 1);
+        let mut out = Vec::new();
+        if term_ords.len() == 1 {
+            vals.ord_to_bytes(term_ords[0], &mut out).unwrap();
+        }
+        out
+    };
+
+    assert_eq!(get_bytes_for_row(0), b"bbbb");
+    assert_eq!(get_bytes_for_row(1), b"baaa");
+    assert_eq!(get_bytes_for_row(2), b"");
+    assert_eq!(get_bytes_for_row(3), b"a");
+}

columnar/src/columnar/merge_index.rs

@@ -0,0 +1 @@
+

columnar/src/columnar/mod.rs

@@ -1,10 +1,11 @@
 mod column_type;
 mod format_version;
 mod merge;
+mod merge_index;
 mod reader;
 mod writer;
 
 pub use column_type::{ColumnType, HasAssociatedColumnType};
-pub use merge::{merge_columnar, MergeDocOrder};
+pub use merge::{merge_columnar, MergeRowOrder, ShuffleMergeOrder, StackMergeOrder};
 pub use reader::ColumnarReader;
 pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -6,6 +6,7 @@ use sstable::{Dictionary, RangeSSTable};
 
 use crate::columnar::{format_version, ColumnType};
 use crate::dynamic_column::DynamicColumnHandle;
+use crate::RowId;
 
 fn io_invalid_data(msg: String) -> io::Error {
     io::Error::new(io::ErrorKind::InvalidData, msg)
@@ -13,9 +14,11 @@ fn io_invalid_data(msg: String) -> io::Error {
 
 /// The ColumnarReader makes it possible to access a set of columns
 /// associated to field names.
+#[derive(Clone)]
 pub struct ColumnarReader {
     column_dictionary: Dictionary<RangeSSTable>,
     column_data: FileSlice,
+    num_rows: RowId,
 }
 
 impl ColumnarReader {
@@ -27,23 +30,27 @@ impl ColumnarReader {
 
     fn open_inner(file_slice: FileSlice) -> io::Result<ColumnarReader> {
         let (file_slice_without_sstable_len, footer_slice) = file_slice
-            .split_from_end(mem::size_of::<u64>() + format_version::VERSION_FOOTER_NUM_BYTES);
+            .split_from_end(mem::size_of::<u64>() + 4 + format_version::VERSION_FOOTER_NUM_BYTES);
         let footer_bytes = footer_slice.read_bytes()?;
-        let (mut sstable_len_bytes, version_footer_bytes) =
-            footer_bytes.rsplit(format_version::VERSION_FOOTER_NUM_BYTES);
+        let sstable_len = u64::deserialize(&mut &footer_bytes[0..8])?;
+        let num_rows = u32::deserialize(&mut &footer_bytes[8..12])?;
         let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
-            version_footer_bytes.as_slice().try_into().unwrap();
+            footer_bytes[12..].try_into().unwrap();
         let _version = format_version::parse_footer(version_footer_bytes)?;
-        let sstable_len = u64::deserialize(&mut sstable_len_bytes)?;
         let (column_data, sstable) =
             file_slice_without_sstable_len.split_from_end(sstable_len as usize);
         let column_dictionary = Dictionary::open(sstable)?;
         Ok(ColumnarReader {
             column_dictionary,
             column_data,
+            num_rows,
         })
     }
 
+    pub fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
     // TODO Add unit tests
     pub fn list_columns(&self) -> io::Result<Vec<(String, DynamicColumnHandle)>> {
         let mut stream = self.column_dictionary.stream()?;
@@ -130,7 +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);
@@ -146,7 +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);

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))
     }
@@ -114,7 +135,7 @@ impl NumericalColumnWriter {
 /// State used to store what types are still acceptable
 /// after having seen a set of numerical values.
 #[derive(Clone, Copy)]
-enum CompatibleNumericalTypes {
+pub(crate) enum CompatibleNumericalTypes {
     Dynamic {
         all_values_within_i64_range: bool,
         all_values_within_u64_range: bool,
@@ -132,7 +153,7 @@ impl Default for CompatibleNumericalTypes {
 }
 
 impl CompatibleNumericalTypes {
-    fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
+    pub fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -148,7 +169,7 @@ impl CompatibleNumericalTypes {
         }
     }
 
-    fn accept_value(&mut self, numerical_value: NumericalValue) {
+    pub fn accept_value(&mut self, numerical_value: NumericalValue) {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -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

@@ -7,15 +7,16 @@ use std::io;
 use std::net::Ipv6Addr;
 
 use column_operation::ColumnOperation;
+pub(crate) use column_writers::CompatibleNumericalTypes;
 use common::CountingWriter;
-use serializer::ColumnarSerializer;
+pub(crate) use serializer::ColumnarSerializer;
 use stacker::{Addr, ArenaHashMap, MemoryArena};
 
 use crate::column_index::SerializableColumnIndex;
 use crate::column_values::{
     ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64, VecColumn,
 };
-use crate::columnar::column_type::{ColumnType, ColumnTypeCategory};
+use crate::columnar::column_type::ColumnType;
 use crate::columnar::writer::column_writers::{
     ColumnWriter, NumericalColumnWriter, StrOrBytesColumnWriter,
 };
@@ -44,7 +45,7 @@ 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();
 /// ```
 pub struct ColumnarWriter {
     numerical_field_hash_map: ArenaHashMap,
@@ -103,6 +104,48 @@ 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`.
@@ -277,37 +320,47 @@ 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));
 
@@ -315,20 +368,24 @@ 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 =
@@ -336,50 +393,64 @@ 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 =
@@ -388,18 +459,24 @@ 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,
                     )?;
                 }
             };
         }
-        serializer.finalize()?;
+        serializer.finalize(num_docs)?;
         Ok(())
     }
 }
 
+// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
+// Column: [Column Index, Column Values, column index num bytes U32::LE]
 fn serialize_bytes_or_str_column(
     cardinality: Cardinality,
     num_docs: RowId,
@@ -547,7 +624,7 @@ fn send_to_serialize_column_mappable_to_u128<
 >(
     op_iterator: impl Iterator<Item = ColumnOperation<T>>,
     cardinality: Cardinality,
-    num_docs: RowId,
+    num_rows: RowId,
     value_index_builders: &mut PreallocatedIndexBuilders,
     values: &mut Vec<T>,
     mut wrt: impl io::Write,
@@ -569,31 +646,30 @@ where
         Cardinality::Optional => {
             let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let optional_index = optional_index_builder.finish(num_docs);
-            SerializableColumnIndex::Optional(Box::new(optional_index))
+            let optional_index = optional_index_builder.finish(num_rows);
+            SerializableColumnIndex::Optional {
+                num_rows,
+                non_null_row_ids: Box::new(optional_index),
+            }
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            let multivalued_index = multivalued_index_builder.finish(num_rows);
             SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u128(
         serializable_column_index,
-        || values.iter().cloned(),
-        values.len() as u32,
+        &&values[..],
         &mut wrt,
     )?;
     Ok(())
 }
 
-fn sort_values_within_row_in_place(
-    multivalued_index: &impl ColumnValues<RowId>,
-    values: &mut Vec<u64>,
-) {
+fn sort_values_within_row_in_place(multivalued_index: &[RowId], values: &mut Vec<u64>) {
     let mut start_index: usize = 0;
-    for end_index in multivalued_index.iter() {
+    for end_index in multivalued_index.iter().copied() {
         let end_index = end_index as usize;
         values[start_index..end_index].sort_unstable();
         start_index = end_index;
@@ -603,7 +679,7 @@ fn sort_values_within_row_in_place(
 fn send_to_serialize_column_mappable_to_u64(
     op_iterator: impl Iterator<Item = ColumnOperation<u64>>,
     cardinality: Cardinality,
-    num_docs: RowId,
+    num_rows: RowId,
     sort_values_within_row: bool,
     value_index_builders: &mut PreallocatedIndexBuilders,
     values: &mut Vec<u64>,
@@ -625,22 +701,25 @@ where
         Cardinality::Optional => {
             let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let optional_index = optional_index_builder.finish(num_docs);
-            SerializableColumnIndex::Optional(Box::new(optional_index))
+            let optional_index = optional_index_builder.finish(num_rows);
+            SerializableColumnIndex::Optional {
+                non_null_row_ids: Box::new(optional_index),
+                num_rows,
+            }
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            let multivalued_index = multivalued_index_builder.finish(num_rows);
             if sort_values_within_row {
-                sort_values_within_row_in_place(&multivalued_index, values);
+                sort_values_within_row_in_place(multivalued_index, values);
             }
             SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u64(
         serializable_column_index,
-        &VecColumn::from(&values[..]),
+        &&values[..],
         &mut wrt,
     )?;
     Ok(())
@@ -693,7 +772,7 @@ 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)));
@@ -722,7 +801,7 @@ 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)));
@@ -745,7 +824,7 @@ 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)));
@@ -764,7 +843,7 @@ 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/serializer.rs

@@ -1,11 +1,12 @@
 use std::io;
 use std::io::Write;
 
-use common::CountingWriter;
+use common::{BinarySerializable, CountingWriter};
 use sstable::value::RangeValueWriter;
 use sstable::RangeSSTable;
 
 use crate::columnar::ColumnType;
+use crate::RowId;
 
 pub struct ColumnarSerializer<W: io::Write> {
     wrt: CountingWriter<W>,
@@ -46,11 +47,12 @@ impl<W: io::Write> ColumnarSerializer<W> {
         }
     }
 
-    pub(crate) fn finalize(mut self) -> io::Result<()> {
+    pub(crate) fn finalize(mut self, num_rows: RowId) -> io::Result<()> {
         let sstable_bytes: Vec<u8> = self.sstable_range.finish()?;
         let sstable_num_bytes: u64 = sstable_bytes.len() as u64;
         self.wrt.write_all(&sstable_bytes)?;
         self.wrt.write_all(&sstable_num_bytes.to_le_bytes()[..])?;
+        num_rows.serialize(&mut self.wrt)?;
         self.wrt
             .write_all(&super::super::format_version::footer())?;
         self.wrt.flush()?;

columnar/src/columnar/writer/value_index.rs

@@ -1,5 +1,4 @@
-use crate::column_index::SerializableOptionalIndex;
-use crate::column_values::{ColumnValues, VecColumn};
+use crate::iterable::Iterable;
 use crate::RowId;
 
 /// The `IndexBuilder` interprets a sequence of
@@ -29,34 +28,15 @@ pub struct OptionalIndexBuilder {
     docs: Vec<RowId>,
 }
 
-struct SingleValueArrayIndex<'a> {
-    // RowIds with a value, in a strictly increasing order
-    row_ids: &'a [RowId],
-    num_rows: RowId,
-}
-
-impl<'a> SerializableOptionalIndex<'a> for SingleValueArrayIndex<'a> {
-    fn num_rows(&self) -> RowId {
-        self.num_rows
-    }
-
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
-        Box::new(self.row_ids.iter().copied())
-    }
-}
-
 impl OptionalIndexBuilder {
-    pub fn finish<'a>(&'a mut self, num_rows: RowId) -> impl SerializableOptionalIndex + 'a {
+    pub fn finish<'a>(&'a mut self, num_rows: RowId) -> impl Iterable<RowId> + 'a {
         debug_assert!(self
             .docs
             .last()
             .copied()
             .map(|last_doc| last_doc < num_rows)
             .unwrap_or(true));
-        SingleValueArrayIndex {
-            row_ids: &self.docs[..],
-            num_rows,
-        }
+        &self.docs[..]
     }
 
     fn reset(&mut self) {
@@ -84,14 +64,10 @@ pub struct MultivaluedIndexBuilder {
 }
 
 impl MultivaluedIndexBuilder {
-    pub fn finish(&mut self, num_docs: RowId) -> impl ColumnValues<u32> + '_ {
+    pub fn finish(&mut self, num_docs: RowId) -> &[u32] {
         self.start_offsets
             .resize(num_docs as usize + 1, self.total_num_vals_seen);
-        VecColumn {
-            values: &&self.start_offsets[..],
-            min_value: 0,
-            max_value: self.start_offsets.last().copied().unwrap_or(0),
-        }
+        &self.start_offsets[..]
     }
 
     fn reset(&mut self) {
@@ -149,7 +125,7 @@ mod tests {
         assert_eq!(
             &opt_value_index_builder
                 .finish(1u32)
-                .non_null_rows()
+                .boxed_iter()
                 .collect::<Vec<u32>>(),
             &[0]
         );
@@ -159,7 +135,7 @@ mod tests {
         assert_eq!(
             &opt_value_index_builder
                 .finish(2u32)
-                .non_null_rows()
+                .boxed_iter()
                 .collect::<Vec<u32>>(),
             &[1]
         );
@@ -177,6 +153,7 @@ mod tests {
             multivalued_value_index_builder
                 .finish(4u32)
                 .iter()
+                .copied()
                 .collect::<Vec<u32>>(),
             vec![0, 0, 2, 3, 3]
         );
@@ -188,6 +165,7 @@ mod tests {
             multivalued_value_index_builder
                 .finish(4u32)
                 .iter()
+                .copied()
                 .collect::<Vec<u32>>(),
             vec![0, 0, 0, 2, 2]
         );

columnar/src/dynamic_column.rs

@@ -8,7 +8,7 @@ use common::{HasLen, OwnedBytes};
 use crate::column::{BytesColumn, Column, StrColumn};
 use crate::column_values::{monotonic_map_column, StrictlyMonotonicFn};
 use crate::columnar::ColumnType;
-use crate::{DateTime, NumericalType};
+use crate::{Cardinality, DateTime, NumericalType};
 
 #[derive(Clone)]
 pub enum DynamicColumn {
@@ -23,6 +23,18 @@ pub enum DynamicColumn {
 }
 
 impl DynamicColumn {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            DynamicColumn::Bool(c) => c.get_cardinality(),
+            DynamicColumn::I64(c) => c.get_cardinality(),
+            DynamicColumn::U64(c) => c.get_cardinality(),
+            DynamicColumn::F64(c) => c.get_cardinality(),
+            DynamicColumn::IpAddr(c) => c.get_cardinality(),
+            DynamicColumn::DateTime(c) => c.get_cardinality(),
+            DynamicColumn::Bytes(c) => c.ords().get_cardinality(),
+            DynamicColumn::Str(c) => c.ords().get_cardinality(),
+        }
+    }
     pub fn column_type(&self) -> ColumnType {
         match self {
             DynamicColumn::Bool(_) => ColumnType::Bool,
@@ -36,6 +48,14 @@ impl DynamicColumn {
         }
     }
 
+    pub fn coerce_numerical(self, target_numerical_type: NumericalType) -> Option<Self> {
+        match target_numerical_type {
+            NumericalType::I64 => self.coerce_to_i64(),
+            NumericalType::U64 => self.coerce_to_u64(),
+            NumericalType::F64 => self.coerce_to_f64(),
+        }
+    }
+
     pub fn is_numerical(&self) -> bool {
         self.column_type().numerical_type().is_some()
     }
@@ -50,7 +70,7 @@ impl DynamicColumn {
         self.column_type().numerical_type() == Some(NumericalType::U64)
     }
 
-    pub fn coerce_to_f64(self) -> Option<DynamicColumn> {
+    fn coerce_to_f64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
                 idx: column.idx,
@@ -64,7 +84,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    pub fn coerce_to_i64(self) -> Option<DynamicColumn> {
+    fn coerce_to_i64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::U64(column) => {
                 if column.max_value() > i64::MAX as u64 {
@@ -79,7 +99,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    pub fn coerce_to_u64(self) -> Option<DynamicColumn> {
+    fn coerce_to_u64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => {
                 if column.min_value() < 0 {
@@ -215,10 +235,8 @@ 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(),

columnar/src/iterable.rs

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

columnar/src/lib.rs

@@ -10,11 +10,12 @@ extern crate test;
 use std::io;
 
 mod column;
-mod column_index;
-mod column_values;
+pub mod column_index;
+pub mod column_values;
 mod columnar;
 mod dictionary;
 mod dynamic_column;
+mod iterable;
 pub(crate) mod utils;
 mod value;
 
@@ -23,7 +24,7 @@ pub use column_index::ColumnIndex;
 pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
 pub use columnar::{
     merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
-    MergeDocOrder,
+    MergeRowOrder, ShuffleMergeOrder, StackMergeOrder,
 };
 use sstable::VoidSSTable;
 pub use value::{NumericalType, NumericalValue};
@@ -31,6 +32,13 @@ pub use value::{NumericalType, NumericalValue};
 pub use self::dynamic_column::{DynamicColumn, DynamicColumnHandle};
 
 pub type RowId = 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>;
 
@@ -39,6 +47,12 @@ pub struct DateTime {
     pub timestamp_micros: i64,
 }
 
+impl DateTime {
+    pub fn into_timestamp_micros(self) -> i64 {
+        self.timestamp_micros
+    }
+}
+
 #[derive(Copy, Clone, Debug)]
 pub struct InvalidData;
 
@@ -66,6 +80,12 @@ pub enum Cardinality {
 }
 
 impl Cardinality {
+    pub fn is_optional(&self) -> bool {
+        matches!(self, Cardinality::Optional)
+    }
+    pub fn is_multivalue(&self) -> bool {
+        matches!(self, Cardinality::Multivalued)
+    }
     pub(crate) fn to_code(self) -> u8 {
         self as u8
     }

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();
@@ -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();

common/src/serialize.rs

@@ -5,12 +5,37 @@ use byteorder::{ReadBytesExt, WriteBytesExt};
 
 use crate::{Endianness, VInt};
 
+#[derive(Default)]
+struct Counter(u64);
+
+impl io::Write for Counter {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        self.0 += buf.len() as u64;
+        Ok(buf.len())
+    }
+
+    fn write_all(&mut self, buf: &[u8]) -> io::Result<()> {
+        self.0 += buf.len() as u64;
+        Ok(())
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
+
 /// Trait for a simple binary serialization.
 pub trait BinarySerializable: fmt::Debug + Sized {
     /// Serialize
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()>;
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()>;
     /// Deserialize
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self>;
+
+    fn num_bytes(&self) -> u64 {
+        let mut counter = Counter::default();
+        self.serialize(&mut counter).unwrap();
+        counter.0
+    }
 }
 
 pub trait DeserializeFrom<T: BinarySerializable> {
@@ -34,7 +59,7 @@ pub trait FixedSize: BinarySerializable {
 }
 
 impl BinarySerializable for () {
-    fn serialize<W: Write>(&self, _: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, _: &mut W) -> io::Result<()> {
         Ok(())
     }
     fn deserialize<R: Read>(_: &mut R) -> io::Result<Self> {
@@ -47,7 +72,7 @@ impl FixedSize for () {
 }
 
 impl<T: BinarySerializable> BinarySerializable for Vec<T> {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.len() as u64).serialize(writer)?;
         for it in self {
             it.serialize(writer)?;
@@ -66,7 +91,7 @@ impl<T: BinarySerializable> BinarySerializable for Vec<T> {
 }
 
 impl<Left: BinarySerializable, Right: BinarySerializable> BinarySerializable for (Left, Right) {
-    fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, write: &mut W) -> io::Result<()> {
         self.0.serialize(write)?;
         self.1.serialize(write)
     }
@@ -81,7 +106,7 @@ impl<Left: BinarySerializable + FixedSize, Right: BinarySerializable + FixedSize
 }
 
 impl BinarySerializable for u32 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u32::<Endianness>(*self)
     }
 
@@ -95,7 +120,7 @@ impl FixedSize for u32 {
 }
 
 impl BinarySerializable for u16 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u16::<Endianness>(*self)
     }
 
@@ -109,7 +134,7 @@ impl FixedSize for u16 {
 }
 
 impl BinarySerializable for u64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -122,7 +147,7 @@ impl FixedSize for u64 {
 }
 
 impl BinarySerializable for u128 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u128::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -135,7 +160,7 @@ impl FixedSize for u128 {
 }
 
 impl BinarySerializable for f32 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_f32::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -148,7 +173,7 @@ impl FixedSize for f32 {
 }
 
 impl BinarySerializable for i64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_i64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -161,7 +186,7 @@ impl FixedSize for i64 {
 }
 
 impl BinarySerializable for f64 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_f64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -174,7 +199,7 @@ impl FixedSize for f64 {
 }
 
 impl BinarySerializable for u8 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u8(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<u8> {
@@ -187,7 +212,7 @@ impl FixedSize for u8 {
 }
 
 impl BinarySerializable for bool {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u8(u8::from(*self))
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<bool> {
@@ -208,7 +233,7 @@ impl FixedSize for bool {
 }
 
 impl BinarySerializable for String {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let data: &[u8] = self.as_bytes();
         VInt(data.len() as u64).serialize(writer)?;
         writer.write_all(data)

common/src/vint.rs

@@ -44,7 +44,7 @@ pub fn deserialize_vint_u128(data: &[u8]) -> io::Result<(u128, &[u8])> {
 pub struct VIntU128(pub u128);
 
 impl BinarySerializable for VIntU128 {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let mut buffer = vec![];
         serialize_vint_u128(self.0, &mut buffer);
         writer.write_all(&buffer)
@@ -211,7 +211,7 @@ impl VInt {
 }
 
 impl BinarySerializable for VInt {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         let mut buffer = [0u8; 10];
         let num_bytes = self.serialize_into(&mut buffer);
         writer.write_all(&buffer[0..num_bytes])

fastfield_codecs/Cargo.toml

@@ -1,34 +0,0 @@
-[package]
-name = "fastfield_codecs"
-version = "0.3.0"
-authors = ["Pascal Seitz <pascal@quickwit.io>"]
-license = "MIT"
-edition = "2021"
-description = "Fast field codecs used by tantivy"
-documentation = "https://docs.rs/fastfield_codecs/"
-homepage = "https://github.com/quickwit-oss/tantivy"
-repository = "https://github.com/quickwit-oss/tantivy"
-
-# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[dependencies]
-common = { version = "0.5", path = "../common/", package = "tantivy-common" }
-tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-columnar = { version= "0.1", path="../columnar", package="tantivy-columnar" }
-prettytable-rs = {version="0.10.0", optional= true}
-rand = {version="0.8.3", optional= true}
-fastdivide = "0.4"
-log = "0.4"
-itertools = { version = "0.10.3" }
-measure_time = { version="0.8.2", optional=true}
-
-[dev-dependencies]
-more-asserts = "0.3.0"
-proptest = "1.0.0"
-rand = "0.8.3"
-
-[features]
-bin = ["prettytable-rs", "rand", "measure_time"]
-default = ["bin"]
-unstable = []
-

fastfield_codecs/README.md

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

fastfield_codecs/benches/bench.rs

@@ -1,311 +0,0 @@
-#![feature(test)]
-
-extern crate test;
-
-#[cfg(test)]
-mod tests {
-    use std::ops::RangeInclusive;
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-    use fastfield_codecs::*;
-    use rand::prelude::*;
-    use test::Bencher;
-
-    use super::*;
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    fn generate_random() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (0u64..100_000u64)
-            .map(|el| el + random::<u16>() as u64)
-            .collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    // Warning: this generates the same permutation at each call
-    fn generate_permutation_gcd() -> Vec<u64> {
-        let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
-        permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
-        permutation
-    }
-
-    pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
-        column: &[T],
-    ) -> Arc<dyn Column<T>> {
-        let mut buffer = Vec::new();
-        serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
-        open(OwnedBytes::new(buffer)).unwrap()
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = permutation[a as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for _ in 0..n {
-                a = column.get_val(a as u32);
-            }
-            a
-        });
-    }
-
-    const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
-    const SINGLE_ITEM: u64 = 90;
-    const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
-    const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
-    fn get_data_50percent_item() -> Vec<u128> {
-        let mut rng = StdRng::from_seed([1u8; 32]);
-
-        let mut data = vec![];
-        for _ in 0..300_000 {
-            let val = rng.gen_range(1..=100);
-            data.push(val);
-        }
-        data.push(SINGLE_ITEM);
-
-        data.shuffle(&mut rng);
-        let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
-        data
-    }
-    fn get_u128_column_random() -> Arc<dyn Column<u128>> {
-        let permutation = generate_random();
-        let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
-        get_u128_column_from_data(&permutation)
-    }
-
-    fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
-        let mut out = vec![];
-        let iter_gen = || data.iter().cloned();
-        serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
-        let out = OwnedBytes::new(out);
-        open_u128::<u128>(out).unwrap()
-    }
-
-    // U64 RANGE START
-    #[bench]
-    fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                FIFTY_PERCENT_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                ONE_PERCENT_ITEM_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                SINGLE_ITEM_RANGE,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
-            positions
-        });
-    }
-    // U64 RANGE END
-
-    // U128 RANGE START
-    #[bench]
-    fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                *FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(
-                *SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
-                0..data.len() as u32,
-                &mut positions,
-            );
-            positions
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
-        let data = get_data_50percent_item();
-        let column = get_u128_column_from_data(&data);
-
-        b.iter(|| {
-            let mut positions = Vec::new();
-            column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
-            positions
-        });
-    }
-    // U128 RANGE END
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let mut a = 0u128;
-            for i in 0u64..column.num_vals() as u64 {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
-        let column = get_u128_column_random();
-
-        b.iter(|| {
-            let n = column.num_vals();
-            let mut a = 0u128;
-            for i in (0..n / 5).map(|val| val * 5) {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += permutation[i as usize];
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0;
-            for i in (0..n / 7).map(|val| val * 7) {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0u32..n as u32 {
-                a += column.get_val(i);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
-        let permutation = generate_permutation_gcd();
-        let n = permutation.len();
-        let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..n {
-                a += column.get_val(i as u32);
-            }
-            a
-        });
-    }
-
-    #[bench]
-    fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
-        let permutation = generate_permutation();
-        b.iter(|| {
-            let mut a = 0u64;
-            for i in 0..permutation.len() {
-                a += permutation[i as usize] as u64;
-            }
-            a
-        });
-    }
-}

fastfield_codecs/src/lib.rs

@@ -1,10 +0,0 @@
-#![warn(missing_docs)]
-#![cfg_attr(all(feature = "unstable", test), feature(test))]
-
-//! # `fastfield_codecs`
-//!
-//! - Columnar storage of data for tantivy [`Column`].
-//! - Encode data in different codecs.
-//! - Monotonically map values to u64/u128
-
-pub use columnar::ColumnValues as Column;

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 +1,126 @@
+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
+        .char_indices()
+        .take_while(|(pos, el)| el.is_numeric())
+        .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().unwrap();
+    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 parser_test() {
+        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())
+        );
+    }
+}

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

@@ -1,5 +1,6 @@
 use std::fmt::Debug;
 
+use columnar::Column;
 use itertools::Itertools;
 use rustc_hash::FxHashMap;
 use serde::{Deserialize, Serialize};
@@ -11,9 +12,11 @@ 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, GenericSegmentAggregationResultsCollector,
+    SegmentAggregationCollector,
+};
 use crate::error::DataCorruption;
-use crate::fastfield::MultiValuedFastFieldReader;
 use crate::schema::Type;
 use crate::{DocId, TantivyError};
 
@@ -196,17 +199,16 @@ 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 +220,7 @@ 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(),
@@ -247,46 +249,11 @@ impl TermBuckets {
         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 +266,12 @@ 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 +283,8 @@ 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 +302,7 @@ 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 +311,6 @@ impl SegmentTermCollector {
         Ok(SegmentTermCollector {
             req: TermsAggregationInternal::from_req(req),
             term_buckets,
-            field_type,
             blueprint,
         })
     }
@@ -362,13 +322,19 @@ 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 +350,40 @@ 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 +392,6 @@ 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 +404,26 @@ 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 +842,14 @@ 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 +876,7 @@ 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 +900,14 @@ 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 +930,14 @@ 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 +957,14 @@ 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 +987,14 @@ 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 +1126,37 @@ 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 +1176,13 @@ 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 +1273,3 @@ 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 +4,10 @@ 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, GenericSegmentAggregationResultsCollector,
+    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 +140,7 @@ 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 +154,7 @@ 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

@@ -12,8 +12,7 @@
 use std::marker::PhantomData;
 use std::sync::Arc;
 
-use columnar::{DynamicColumn, HasAssociatedColumnType};
-use fastfield_codecs::Column;
+use columnar::{ColumnValues, DynamicColumn, HasAssociatedColumnType};
 
 use crate::collector::{Collector, SegmentCollector};
 use crate::schema::Field;
@@ -121,7 +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
@@ -152,7 +151,7 @@ where
     TPredicate: 'static,
     DynamicColumn: Into<Option<columnar::Column<TPredicateValue>>>,
 {
-    fast_field_reader: Arc<dyn Column<TPredicateValue>>,
+    fast_field_reader: Arc<dyn ColumnValues<TPredicateValue>>,
     segment_collector: TSegmentCollector,
     predicate: TPredicate,
     t_predicate_value: PhantomData<TPredicateValue>,

src/collector/histogram_collector.rs

@@ -1,7 +1,7 @@
 use std::sync::Arc;
 
+use columnar::ColumnValues;
 use fastdivide::DividerU64;
-use fastfield_codecs::Column;
 
 use crate::collector::{Collector, SegmentCollector};
 use crate::fastfield::{FastFieldNotAvailableError, FastValue};
@@ -87,7 +87,7 @@ impl HistogramComputer {
 }
 pub struct SegmentHistogramCollector {
     histogram_computer: HistogramComputer,
-    column_u64: Arc<dyn Column<u64>>,
+    column_u64: Arc<dyn ColumnValues<u64>>,
 }
 
 impl SegmentCollector for SegmentHistogramCollector {

src/collector/tests.rs

@@ -1,12 +1,12 @@
 use std::sync::Arc;
 
-use fastfield_codecs::Column;
+use columnar::{BytesColumn, ColumnValues};
 
 use super::*;
 use crate::collector::{Count, FilterCollector, TopDocs};
 use crate::core::SegmentReader;
 use crate::query::{AllQuery, QueryParser};
-use crate::schema::{Field, Schema, FAST, TEXT};
+use crate::schema::{Schema, FAST, TEXT};
 use crate::time::format_description::well_known::Rfc3339;
 use crate::time::OffsetDateTime;
 use crate::{doc, DateTime, DocAddress, DocId, Document, Index, Score, Searcher, SegmentOrdinal};
@@ -160,7 +160,7 @@ pub struct FastFieldTestCollector {
 
 pub struct FastFieldSegmentCollector {
     vals: Vec<u64>,
-    reader: Arc<dyn Column<u64>>,
+    reader: Arc<dyn columnar::ColumnValues>,
 }
 
 impl FastFieldTestCollector {
@@ -212,62 +212,73 @@ 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

@@ -3,7 +3,7 @@ use std::fmt;
 use std::marker::PhantomData;
 use std::sync::Arc;
 
-use fastfield_codecs::Column;
+use columnar::ColumnValues;
 
 use super::Collector;
 use crate::collector::custom_score_top_collector::CustomScoreTopCollector;
@@ -14,7 +14,6 @@ use crate::collector::{
 };
 use crate::fastfield::{FastFieldNotAvailableError, FastValue};
 use crate::query::Weight;
-use crate::schema::Field;
 use crate::{DocAddress, DocId, Score, SegmentOrdinal, SegmentReader, TantivyError};
 
 struct FastFieldConvertCollector<
@@ -133,7 +132,7 @@ impl fmt::Debug for TopDocs {
 }
 
 struct ScorerByFastFieldReader {
-    sort_column: Arc<dyn Column<u64>>,
+    sort_column: Arc<dyn ColumnValues<u64>>,
 }
 
 impl CustomSegmentScorer<u64> for ScorerByFastFieldReader {
@@ -236,7 +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();
@@ -253,7 +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))]);
@@ -263,8 +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
@@ -272,7 +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
@@ -324,22 +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))]);
@@ -350,7 +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
@@ -359,7 +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

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/composite_file.rs

@@ -22,7 +22,7 @@ impl FileAddr {
 }
 
 impl BinarySerializable for FileAddr {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         self.field.serialize(writer)?;
         VInt(self.idx as u64).serialize(writer)?;
         Ok(())

src/fastfield/mod.rs

@@ -21,15 +21,15 @@
 
 use std::net::Ipv6Addr;
 
+pub use columnar::Column;
 use columnar::MonotonicallyMappableToU64;
-pub use fastfield_codecs::Column;
 
 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;
@@ -115,33 +115,13 @@ impl columnar::MonotonicallyMappableToU64 for DateTime {
     }
 }
 
-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::collections::HashMap;
     use std::ops::{Range, RangeInclusive};
     use std::path::Path;
-    use std::sync::Arc;
 
+    use columnar::Column;
     use common::{HasLen, TerminatingWrite};
     use once_cell::sync::Lazy;
     use rand::prelude::SliceRandom;
@@ -149,7 +129,7 @@ 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 +171,7 @@ mod tests {
         }
         let file = directory.open_read(path).unwrap();
 
-        assert_eq!(file.len(), 157);
+        assert_eq!(file.len(), 161);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let column = fast_field_readers.u64("field").unwrap();
         assert_eq!(column.get_val(0), 13u64);
@@ -238,7 +218,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 185);
+        assert_eq!(file.len(), 189);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let col = fast_field_readers.u64("field").unwrap();
         assert_eq!(col.get_val(0), 4u64);
@@ -268,7 +248,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 158);
+        assert_eq!(file.len(), 162);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let fast_field_reader = fast_field_readers.u64("field").unwrap();
         for doc in 0..10_000 {
@@ -297,7 +277,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 80166);
+        assert_eq!(file.len(), 4557);
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
             let col = fast_field_readers.u64("field").unwrap();
@@ -327,7 +307,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 329_usize);
+        assert_eq!(file.len(), 333_usize);
 
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
@@ -438,39 +418,38 @@ 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() {
@@ -526,57 +505,61 @@ 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]
@@ -590,11 +573,7 @@ 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]);
@@ -620,155 +599,162 @@ 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::<columnar::DateTime>("date")
+            .unwrap()
+            .unwrap()
+            .first_or_default_col(Default::default());
+        let dates_fast_field = fast_fields
+            .column_opt::<columnar::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() {
@@ -795,7 +781,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 171);
+        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);
@@ -827,7 +813,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 183);
+        assert_eq!(file.len(), 187);
         let readers = FastFieldReaders::open(file).unwrap();
         let bool_col = readers.bool("field_bool").unwrap();
         for i in 0..25 {
@@ -852,7 +838,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 173);
+        assert_eq!(file.len(), 177);
         let fastfield_readers = FastFieldReaders::open(file).unwrap();
         let col = fastfield_readers.bool("field_bool").unwrap();
         assert_eq!(col.get_val(0), false);
@@ -978,8 +964,7 @@ 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();
+        let column: Column<Ipv6Addr> = fastfields.column_opt("ip").unwrap().unwrap();
         assert_eq!(column.num_rows(), 3);
         assert_eq!(column.first(0), None);
         assert_eq!(column.first(1), Some(ip_addr));

src/fastfield/readers.rs

@@ -1,16 +1,14 @@
-use std::collections::HashMap;
 use std::io;
 use std::net::Ipv6Addr;
 use std::sync::Arc;
 
 use columnar::{
-    BytesColumn, ColumnType, ColumnValues, ColumnarReader, DynamicColumn, DynamicColumnHandle,
-    HasAssociatedColumnType, StrColumn,
+    BytesColumn, Column, ColumnType, ColumnValues, ColumnarReader, DynamicColumn,
+    DynamicColumnHandle, HasAssociatedColumnType, StrColumn,
 };
-use fastfield_codecs::Column;
 
 use crate::directory::FileSlice;
-use crate::schema::{Field, Schema};
+use crate::schema::Schema;
 use crate::space_usage::{FieldUsage, PerFieldSpaceUsage};
 
 /// Provides access to all of the BitpackedFastFieldReader.
@@ -28,6 +26,10 @@ impl FastFieldReaders {
         Ok(FastFieldReaders { columnar })
     }
 
+    pub(crate) fn columnar(&self) -> &ColumnarReader {
+        self.columnar.as_ref()
+    }
+
     pub(crate) fn space_usage(&self, schema: &Schema) -> io::Result<PerFieldSpaceUsage> {
         let mut per_field_usages: Vec<FieldUsage> = Default::default();
         for (field, field_entry) in schema.fields() {
@@ -44,33 +46,18 @@ impl FastFieldReaders {
         Ok(PerFieldSpaceUsage::new(per_field_usages))
     }
 
-    pub fn typed_column_opt<T>(
-        &self,
-        field_name: &str,
-    ) -> crate::Result<Option<columnar::Column<T>>>
+    /// Returns a typed column associated to a given field name.
+    ///
+    /// If no column associated with that field_name exists,
+    /// or existing columns do not have the required type,
+    /// returns `None`.
+    pub fn column_opt<T>(&self, field_name: &str) -> crate::Result<Option<Column<T>>>
     where
         T: PartialOrd + Copy + HasAssociatedColumnType + Send + Sync + 'static,
-        DynamicColumn: Into<Option<columnar::Column<T>>>,
+        DynamicColumn: Into<Option<Column<T>>>,
     {
         let column_type = T::column_type();
-        let Some(dynamic_column_handle) = self.column_handle(field_name, column_type)?
-        else {
-            return Ok(None);
-        };
-        let dynamic_column = dynamic_column_handle.open()?;
-        Ok(dynamic_column.into())
-    }
-
-    pub fn bytes_column_opt(&self, field_name: &str) -> crate::Result<Option<BytesColumn>> {
-        let Some(dynamic_column_handle) = self.column_handle(field_name, ColumnType::Bytes)?
-        else {
-            return Ok(None);
-        };
-        let dynamic_column = dynamic_column_handle.open()?;
-        Ok(dynamic_column.into())
-    }
-    pub fn str_column_opt(&self, field_name: &str) -> crate::Result<Option<StrColumn>> {
-        let Some(dynamic_column_handle) = self.column_handle(field_name, ColumnType::Str)?
+        let Some(dynamic_column_handle) = self.dynamic_column_handle(field_name, column_type)?
         else {
             return Ok(None);
         };
@@ -78,6 +65,7 @@ impl FastFieldReaders {
         Ok(dynamic_column.into())
     }
 
+    /// Returns the number of `bytes` associated with a column.
     pub fn column_num_bytes(&self, field: &str) -> crate::Result<usize> {
         Ok(self
             .columnar
@@ -87,12 +75,17 @@ impl FastFieldReaders {
             .sum())
     }
 
-    pub fn typed_column_first_or_default<T>(&self, field: &str) -> crate::Result<Arc<dyn Column<T>>>
+    /// Returns a typed column value object.
+    ///
+    /// In that column value:
+    /// - Rows with no value are associated with the default value.
+    /// - Rows with several values are associated with the first value.
+    pub fn column_first_or_default<T>(&self, field: &str) -> crate::Result<Arc<dyn ColumnValues<T>>>
     where
         T: PartialOrd + Copy + HasAssociatedColumnType + Send + Sync + 'static,
-        DynamicColumn: Into<Option<columnar::Column<T>>>,
+        DynamicColumn: Into<Option<Column<T>>>,
     {
-        let col_opt: Option<columnar::Column<T>> = self.typed_column_opt(field)?;
+        let col_opt: Option<Column<T>> = self.column_opt(field)?;
         if let Some(col) = col_opt {
             Ok(col.first_or_default_col(T::default_value()))
         } else {
@@ -106,32 +99,45 @@ impl FastFieldReaders {
     ///
     /// If `field` is not a u64 fast field, this method returns an Error.
     pub fn u64(&self, field: &str) -> crate::Result<Arc<dyn ColumnValues<u64>>> {
-        self.typed_column_first_or_default(field)
+        self.column_first_or_default(field)
     }
 
     /// Returns the `date` fast field reader reader associated with `field`.
     ///
     /// If `field` is not a date fast field, this method returns an Error.
     pub fn date(&self, field: &str) -> crate::Result<Arc<dyn ColumnValues<columnar::DateTime>>> {
-        self.typed_column_first_or_default(field)
+        self.column_first_or_default(field)
     }
 
     /// Returns the `ip` fast field reader reader associated to `field`.
     ///
     /// If `field` is not a u128 fast field, this method returns an Error.
-    pub fn ip_addr(&self, field: &str) -> crate::Result<Arc<dyn Column<Ipv6Addr>>> {
-        self.typed_column_first_or_default(field)
+    pub fn ip_addr(&self, field: &str) -> crate::Result<Arc<dyn ColumnValues<Ipv6Addr>>> {
+        self.column_first_or_default(field)
     }
 
-    pub fn str(&self, field: &str) -> crate::Result<Option<columnar::StrColumn>> {
-        self.str_column_opt(field)
+    /// Returns a `str` column.
+    pub fn str(&self, field_name: &str) -> crate::Result<Option<StrColumn>> {
+        let Some(dynamic_column_handle) = self.dynamic_column_handle(field_name, ColumnType::Str)?
+        else {
+            return Ok(None);
+        };
+        let dynamic_column = dynamic_column_handle.open()?;
+        Ok(dynamic_column.into())
     }
 
-    pub fn bytes(&self, field: &str) -> crate::Result<Option<columnar::BytesColumn>> {
-        self.bytes_column_opt(field)
+    /// Returns a `bytes` column.
+    pub fn bytes(&self, field_name: &str) -> crate::Result<Option<BytesColumn>> {
+        let Some(dynamic_column_handle) = self.dynamic_column_handle(field_name, ColumnType::Bytes)?
+        else {
+            return Ok(None);
+        };
+        let dynamic_column = dynamic_column_handle.open()?;
+        Ok(dynamic_column.into())
     }
 
-    pub fn column_handle(
+    /// Returning a `dynamic_column_handle`.
+    pub fn dynamic_column_handle(
         &self,
         field_name: &str,
         column_type: ColumnType,
@@ -145,7 +151,9 @@ impl FastFieldReaders {
         Ok(dynamic_column_handle_opt)
     }
 
-    pub fn u64_lenient(&self, field_name: &str) -> crate::Result<Option<columnar::Column<u64>>> {
+    /// Returns the `u64` column used to represent any `u64`-mapped typed (i64, u64, f64, DateTime).
+    #[doc(hidden)]
+    pub fn u64_lenient(&self, field_name: &str) -> crate::Result<Option<Column<u64>>> {
         for col in self.columnar.read_columns(field_name)? {
             if let Some(col_u64) = col.open_u64_lenient()? {
                 return Ok(Some(col_u64));
@@ -157,21 +165,21 @@ impl FastFieldReaders {
     /// Returns the `i64` fast field reader reader associated with `field`.
     ///
     /// If `field` is not a i64 fast field, this method returns an Error.
-    pub fn i64(&self, field_name: &str) -> crate::Result<Arc<dyn Column<i64>>> {
-        self.typed_column_first_or_default(field_name)
+    pub fn i64(&self, field_name: &str) -> crate::Result<Arc<dyn ColumnValues<i64>>> {
+        self.column_first_or_default(field_name)
     }
 
     /// Returns the `f64` fast field reader reader associated with `field`.
     ///
     /// If `field` is not a f64 fast field, this method returns an Error.
-    pub fn f64(&self, field_name: &str) -> crate::Result<Arc<dyn Column<f64>>> {
-        self.typed_column_first_or_default(field_name)
+    pub fn f64(&self, field_name: &str) -> crate::Result<Arc<dyn ColumnValues<f64>>> {
+        self.column_first_or_default(field_name)
     }
 
     /// Returns the `bool` fast field reader reader associated with `field`.
     ///
     /// If `field` is not a bool fast field, this method returns an Error.
-    pub fn bool(&self, field_name: &str) -> crate::Result<Arc<dyn Column<bool>>> {
-        self.typed_column_first_or_default(field_name)
+    pub fn bool(&self, field_name: &str) -> crate::Result<Arc<dyn ColumnValues<bool>>> {
+        self.column_first_or_default(field_name)
     }
 }

src/fastfield/writer.rs

@@ -67,6 +67,16 @@ impl FastFieldsWriter {
         self.columnar_writer.mem_usage()
     }
 
+    pub(crate) fn sort_order(
+        &self,
+        sort_field: &str,
+        num_docs: DocId,
+        reversed: bool,
+    ) -> Vec<DocId> {
+        self.columnar_writer
+            .sort_order(sort_field, num_docs, reversed)
+    }
+
     /// Indexes all of the fastfields of a new document.
     pub fn add_document(&mut self, doc: &Document) -> crate::Result<()> {
         let doc_id = self.num_docs;
@@ -143,11 +153,13 @@ impl FastFieldsWriter {
     pub fn serialize(
         mut self,
         wrt: &mut dyn io::Write,
-        doc_id_map: Option<&DocIdMapping>,
+        doc_id_map_opt: Option<&DocIdMapping>,
     ) -> io::Result<()> {
-        assert!(doc_id_map.is_none()); // TODO handle doc id map
         let num_docs = self.num_docs;
-        self.columnar_writer.serialize(num_docs, wrt)?;
+        let old_to_new_row_ids =
+            doc_id_map_opt.map(|doc_id_mapping| doc_id_mapping.old_to_new_ids());
+        self.columnar_writer
+            .serialize(num_docs, old_to_new_row_ids, wrt)?;
         Ok(())
     }
 }

src/indexer/doc_id_mapping.rs

@@ -1,27 +1,44 @@
 //! This module is used when sorting the index by a property, e.g.
 //! to get mappings from old doc_id to new doc_id and vice versa, after sorting
 
-use std::cmp::Reverse;
+use common::ReadOnlyBitSet;
 
 use super::SegmentWriter;
 use crate::schema::{Field, Schema};
-use crate::{DocAddress, DocId, IndexSortByField, Order, TantivyError};
+use crate::{DocAddress, DocId, IndexSortByField, TantivyError};
+
+#[derive(Copy, Clone, Eq, PartialEq)]
+pub enum MappingType {
+    Stacked,
+    StackedWithDeletes,
+    Shuffled,
+}
 
 /// Struct to provide mapping from new doc_id to old doc_id and segment.
 #[derive(Clone)]
 pub(crate) struct SegmentDocIdMapping {
-    new_doc_id_to_old_doc_addr: Vec<DocAddress>,
-    is_trivial: bool,
+    pub(crate) new_doc_id_to_old_doc_addr: Vec<DocAddress>,
+    pub(crate) alive_bitsets: Vec<Option<ReadOnlyBitSet>>,
+    mapping_type: MappingType,
 }
 
 impl SegmentDocIdMapping {
-    pub(crate) fn new(new_doc_id_to_old_and_segment: Vec<DocAddress>, is_trivial: bool) -> Self {
+    pub(crate) fn new(
+        new_doc_id_to_old_doc_addr: Vec<DocAddress>,
+        mapping_type: MappingType,
+        alive_bitsets: Vec<Option<ReadOnlyBitSet>>,
+    ) -> Self {
         Self {
-            new_doc_id_to_old_doc_addr: new_doc_id_to_old_and_segment,
-            is_trivial,
+            new_doc_id_to_old_doc_addr,
+            mapping_type,
+            alive_bitsets,
         }
     }
 
+    pub fn mapping_type(&self) -> MappingType {
+        self.mapping_type
+    }
+
     /// Returns an iterator over the old document addresses, ordered by the new document ids.
     ///
     /// In the returned `DocAddress`, the `segment_ord` is the ordinal of targeted segment
@@ -30,16 +47,20 @@ impl SegmentDocIdMapping {
         self.new_doc_id_to_old_doc_addr.iter().copied()
     }
 
-    pub(crate) fn len(&self) -> usize {
-        self.new_doc_id_to_old_doc_addr.len()
-    }
-
     /// This flags means the segments are simply stacked in the order of their ordinal.
     /// e.g. [(0, 1), .. (n, 1), (0, 2)..., (m, 2)]
     ///
+    /// The different segment may present some deletes, in which case it is expressed by skipping a
+    /// `DocId`. [(0, 1), (0, 3)] <--- here doc_id=0 and doc_id=1 have been deleted
+    ///
+    /// Being trivial is equivalent to having the `new_doc_id_to_old_doc_addr` array sorted.
+    ///
     /// This allows for some optimization.
     pub(crate) fn is_trivial(&self) -> bool {
-        self.is_trivial
+        match self.mapping_type {
+            MappingType::Stacked | MappingType::StackedWithDeletes => true,
+            MappingType::Shuffled => false,
+        }
     }
 }
 
@@ -80,6 +101,11 @@ impl DocIdMapping {
     pub fn iter_old_doc_ids(&self) -> impl Iterator<Item = DocId> + Clone + '_ {
         self.new_doc_id_to_old.iter().cloned()
     }
+
+    pub fn old_to_new_ids(&self) -> &[DocId] {
+        &self.old_doc_id_to_new[..]
+    }
+
     /// Remaps a given array to the new doc ids.
     pub fn remap<T: Copy>(&self, els: &[T]) -> Vec<T> {
         self.new_doc_id_to_old
@@ -113,36 +139,15 @@ pub(crate) fn get_doc_id_mapping_from_field(
     sort_by_field: IndexSortByField,
     segment_writer: &SegmentWriter,
 ) -> crate::Result<DocIdMapping> {
-    todo!()
-    // let schema = segment_writer.segment_serializer.segment().schema();
-    // let field_id = expect_field_id_for_sort_field(&schema, &sort_by_field)?; // for now expect
-    // fastfield, but not strictly required
-    // let fast_field = segment_writer
-    //     .fast_field_writers
-    //     .get_field_writer(field_id)
-    //     .ok_or_else(|| {
-    //         TantivyError::InvalidArgument(format!(
-    //             "sort index by field is required to be a fast field {:?}",
-    //             sort_by_field.field
-    //         ))
-    //     })?;
-
-    // // create new doc_id to old doc_id index (used in fast_field_writers)
-    // let mut doc_id_and_data = fast_field
-    //     .iter()
-    //     .enumerate()
-    //     .map(|el| (el.0 as DocId, el.1))
-    //     .collect::<Vec<_>>();
-    // if sort_by_field.order == Order::Desc {
-    //     doc_id_and_data.sort_by_key(|k| Reverse(k.1));
-    // } else {
-    //     doc_id_and_data.sort_by_key(|k| k.1);
-    // }
-    // let new_doc_id_to_old = doc_id_and_data
-    //     .into_iter()
-    //     .map(|el| el.0)
-    //     .collect::<Vec<_>>();
-    // Ok(DocIdMapping::from_new_id_to_old_id(new_doc_id_to_old))
+    let schema = segment_writer.segment_serializer.segment().schema();
+    expect_field_id_for_sort_field(&schema, &sort_by_field)?; // for now expect
+    let new_doc_id_to_old = segment_writer.fast_field_writers.sort_order(
+        sort_by_field.field.as_str(),
+        segment_writer.max_doc(),
+        sort_by_field.order.is_desc(),
+    );
+    // create new doc_id to old doc_id index (used in fast_field_writers)
+    Ok(DocIdMapping::from_new_id_to_old_id(new_doc_id_to_old))
 }
 
 #[cfg(test)]

src/indexer/merger.rs

@@ -1,18 +1,19 @@
-use std::collections::HashMap;
 use std::sync::Arc;
 
+use columnar::{
+    ColumnValues, ColumnarReader, MergeRowOrder, RowAddr, ShuffleMergeOrder, StackMergeOrder,
+};
+use common::ReadOnlyBitSet;
 use itertools::Itertools;
 use measure_time::debug_time;
 
-// use super::sorted_doc_id_multivalue_column::RemappedDocIdMultiValueIndexColumn;
 use crate::core::{Segment, SegmentReader};
 use crate::directory::WritePtr;
 use crate::docset::{DocSet, TERMINATED};
 use crate::error::DataCorruption;
-use crate::fastfield::{AliveBitSet, Column, FastFieldNotAvailableError};
+use crate::fastfield::{AliveBitSet, FastFieldNotAvailableError};
 use crate::fieldnorm::{FieldNormReader, FieldNormReaders, FieldNormsSerializer, FieldNormsWriter};
-use crate::indexer::doc_id_mapping::SegmentDocIdMapping;
-// use crate::indexer::sorted_doc_id_multivalue_column::RemappedDocIdMultiValueColumn;
+use crate::indexer::doc_id_mapping::{MappingType, SegmentDocIdMapping};
 use crate::indexer::SegmentSerializer;
 use crate::postings::{InvertedIndexSerializer, Postings, SegmentPostings};
 use crate::schema::{Field, FieldType, Schema};
@@ -85,29 +86,6 @@ pub struct IndexMerger {
     max_doc: u32,
 }
 
-struct TermOrdinalMapping {
-    per_segment_new_term_ordinals: Vec<Vec<TermOrdinal>>,
-}
-
-impl TermOrdinalMapping {
-    fn new(max_term_ords: Vec<TermOrdinal>) -> TermOrdinalMapping {
-        TermOrdinalMapping {
-            per_segment_new_term_ordinals: max_term_ords
-                .into_iter()
-                .map(|max_term_ord| vec![TermOrdinal::default(); max_term_ord as usize])
-                .collect(),
-        }
-    }
-
-    fn register_from_to(&mut self, segment_ord: usize, from_ord: TermOrdinal, to_ord: TermOrdinal) {
-        self.per_segment_new_term_ordinals[segment_ord][from_ord as usize] = to_ord;
-    }
-
-    fn get_segment(&self, segment_ord: usize) -> &[TermOrdinal] {
-        &(self.per_segment_new_term_ordinals[segment_ord])[..]
-    }
-}
-
 struct DeltaComputer {
     buffer: Vec<u32>,
 }
@@ -132,6 +110,31 @@ impl DeltaComputer {
     }
 }
 
+fn convert_to_merge_order<'a>(
+    columnars: &[&'a ColumnarReader],
+    doc_id_mapping: SegmentDocIdMapping,
+) -> MergeRowOrder {
+    match doc_id_mapping.mapping_type() {
+        MappingType::Stacked => MergeRowOrder::Stack(StackMergeOrder::stack(columnars)),
+        MappingType::StackedWithDeletes | MappingType::Shuffled => {
+            // RUST/LLVM is amazing. The following conversion is actually a no-op:
+            // no allocation, no copy.
+            let new_row_id_to_old_row_id: Vec<RowAddr> = doc_id_mapping
+                .new_doc_id_to_old_doc_addr
+                .into_iter()
+                .map(|doc_addr| RowAddr {
+                    segment_ord: doc_addr.segment_ord,
+                    row_id: doc_addr.doc_id,
+                })
+                .collect();
+            MergeRowOrder::Shuffled(ShuffleMergeOrder {
+                new_row_id_to_old_row_id,
+                alive_bitsets: doc_id_mapping.alive_bitsets,
+            })
+        }
+    }
+}
+
 impl IndexMerger {
     pub fn open(
         schema: Schema,
@@ -223,7 +226,6 @@ impl IndexMerger {
         let mut fieldnorms_data = Vec::with_capacity(self.max_doc as usize);
         for field in fields {
             fieldnorms_data.clear();
-
             let fieldnorms_readers: Vec<FieldNormReader> = self
                 .readers
                 .iter()
@@ -234,7 +236,6 @@ impl IndexMerger {
                 let fieldnorm_id = fieldnorms_reader.fieldnorm_id(old_doc_addr.doc_id);
                 fieldnorms_data.push(fieldnorm_id);
             }
-
             fieldnorms_serializer.serialize_field(field, &fieldnorms_data[..])?;
         }
         fieldnorms_serializer.close()?;
@@ -244,67 +245,16 @@ impl IndexMerger {
     fn write_fast_fields(
         &self,
         fast_field_wrt: &mut WritePtr,
-        mut term_ord_mappings: HashMap<Field, TermOrdinalMapping>,
-        doc_id_mapping: &SegmentDocIdMapping,
+        doc_id_mapping: SegmentDocIdMapping,
     ) -> crate::Result<()> {
-        debug_time!("wrie-fast-fields");
-        for (_field, field_entry) in self.schema.fields() {
-            if field_entry.is_fast() {
-                todo!();
-            }
-        }
-
-        // for (field, field_entry) in self.schema.fields() {
-        // let field_type = field_entry.field_type();
-        // match field_type {
-        // FieldType::Facet(_) | FieldType::Str(_) if field_type.is_fast() => {
-        // let term_ordinal_mapping = term_ord_mappings.remove(&field).expect(
-        // "Logic Error in Tantivy (Please report). Facet field should have required \
-        // a`term_ordinal_mapping`.",
-        // );
-        // self.write_term_id_fast_field(
-        // field,
-        // &term_ordinal_mapping,
-        // fast_field_serializer,
-        // doc_id_mapping,
-        // )?;
-        // }
-        // FieldType::U64(ref options)
-        // | FieldType::I64(ref options)
-        // | FieldType::F64(ref options)
-        // | FieldType::Bool(ref options) => {
-        // todo!()
-        // }
-        // FieldType::Date(ref options) => {
-        // if options.is_fast() {
-        // todo!();
-        // }
-        // Some(Cardinality::SingleValue) => {
-        //     self.write_single_fast_field(field, fast_field_serializer, doc_id_mapping)?;
-        // }
-        // Some(Cardinality::MultiValues) => {
-        //     self.write_multi_fast_field(field, fast_field_serializer, doc_id_mapping)?;
-        // }
-        // None => {}
-        // },
-        // FieldType::Bytes(byte_options) => {
-        // if byte_options.is_fast() {
-        // self.write_bytes_fast_field(field, fast_field_serializer, doc_id_mapping)?;
-        // }
-        // }
-        // FieldType::IpAddr(options) =>  {
-        // if options.is_fast() {
-        // todo!();
-        // }
-        // },
-        //
-        // FieldType::JsonObject(_) | FieldType::Facet(_) | FieldType::Str(_) => {
-        // We don't handle json fast field for the moment
-        // They can be implemented using what is done
-        // for facets in the future
-        // }
-        // }
-        // }
+        debug_time!("write-fast-fields");
+        let columnars: Vec<&ColumnarReader> = self
+            .readers
+            .iter()
+            .map(|reader| reader.fast_fields().columnar())
+            .collect();
+        let merge_row_order = convert_to_merge_order(&columnars[..], doc_id_mapping);
+        columnar::merge_columnar(&columnars[..], merge_row_order, fast_field_wrt)?;
         Ok(())
     }
 
@@ -334,7 +284,7 @@ impl IndexMerger {
     pub(crate) fn get_sort_field_accessor(
         reader: &SegmentReader,
         sort_by_field: &IndexSortByField,
-    ) -> crate::Result<Arc<dyn Column>> {
+    ) -> crate::Result<Arc<dyn ColumnValues>> {
         reader.schema().get_field(&sort_by_field.field)?;
         let value_accessor = reader
             .fast_fields()
@@ -348,7 +298,7 @@ impl IndexMerger {
     pub(crate) fn get_reader_with_sort_field_accessor(
         &self,
         sort_by_field: &IndexSortByField,
-    ) -> crate::Result<Vec<(SegmentOrdinal, Arc<dyn Column>)>> {
+    ) -> crate::Result<Vec<(SegmentOrdinal, Arc<dyn ColumnValues>)>> {
         let reader_ordinal_and_field_accessors = self
             .readers
             .iter()
@@ -369,7 +319,7 @@ impl IndexMerger {
     /// doc_id.
     /// ReaderWithOrdinal will include the ordinal position of the
     /// reader in self.readers.
-    pub(crate) fn generate_doc_id_mapping(
+    pub(crate) fn generate_doc_id_mapping_with_sort_by_field(
         &self,
         sort_by_field: &IndexSortByField,
     ) -> crate::Result<SegmentDocIdMapping> {
@@ -414,7 +364,23 @@ impl IndexMerger {
                     segment_ord,
                 }),
         );
-        Ok(SegmentDocIdMapping::new(sorted_doc_ids, false))
+
+        let alive_bitsets: Vec<Option<ReadOnlyBitSet>> = self
+            .readers
+            .iter()
+            .map(|segment_reader| {
+                if let Some(alive_bitset) = segment_reader.alive_bitset() {
+                    Some(alive_bitset.bitset().clone())
+                } else {
+                    None
+                }
+            })
+            .collect();
+        Ok(SegmentDocIdMapping::new(
+            sorted_doc_ids,
+            MappingType::Shuffled,
+            alive_bitsets,
+        ))
     }
 
     /// Creates a mapping if the segments are stacked. this is helpful to merge codelines between
@@ -439,17 +405,39 @@ impl IndexMerger {
                     })
                 }),
         );
-        Ok(SegmentDocIdMapping::new(mapping, true))
+
+        let has_deletes: bool = self.readers.iter().any(SegmentReader::has_deletes);
+        let mapping_type = if has_deletes {
+            MappingType::StackedWithDeletes
+        } else {
+            MappingType::Stacked
+        };
+        let alive_bitsets: Vec<Option<ReadOnlyBitSet>> = self
+            .readers
+            .iter()
+            .map(|reader| {
+                if let Some(bitset) = reader.alive_bitset() {
+                    Some(bitset.bitset().clone())
+                } else {
+                    None
+                }
+            })
+            .collect();
+        Ok(SegmentDocIdMapping::new(
+            mapping,
+            mapping_type,
+            alive_bitsets,
+        ))
     }
 
     fn write_postings_for_field(
         &self,
         indexed_field: Field,
-        field_type: &FieldType,
+        _field_type: &FieldType,
         serializer: &mut InvertedIndexSerializer,
         fieldnorm_reader: Option<FieldNormReader>,
         doc_id_mapping: &SegmentDocIdMapping,
-    ) -> crate::Result<Option<TermOrdinalMapping>> {
+    ) -> crate::Result<()> {
         debug_time!("write-postings-for-field");
         let mut positions_buffer: Vec<u32> = Vec::with_capacity(1_000);
         let mut delta_computer = DeltaComputer::new();
@@ -469,14 +457,6 @@ impl IndexMerger {
             max_term_ords.push(terms.num_terms() as u64);
         }
 
-        let mut term_ord_mapping_opt = match field_type {
-            FieldType::Facet(_) => Some(TermOrdinalMapping::new(max_term_ords)),
-            FieldType::Str(options) if options.is_fast() => {
-                Some(TermOrdinalMapping::new(max_term_ords))
-            }
-            _ => None,
-        };
-
         let mut merged_terms = TermMerger::new(field_term_streams);
 
         // map from segment doc ids to the resulting merged segment doc id.
@@ -559,13 +539,7 @@ impl IndexMerger {
                 continue;
             }
 
-            let to_term_ord = field_serializer.new_term(term_bytes, total_doc_freq)?;
-
-            if let Some(ref mut term_ord_mapping) = term_ord_mapping_opt {
-                for (segment_ord, from_term_ord) in merged_terms.matching_segments() {
-                    term_ord_mapping.register_from_to(segment_ord, from_term_ord, to_term_ord);
-                }
-            }
+            field_serializer.new_term(term_bytes, total_doc_freq)?;
 
             // We can now serialize this postings, by pushing each document to the
             // postings serializer.
@@ -617,7 +591,7 @@ impl IndexMerger {
             field_serializer.close_term()?;
         }
         field_serializer.close()?;
-        Ok(term_ord_mapping_opt)
+        Ok(())
     }
 
     fn write_postings(
@@ -625,23 +599,20 @@ impl IndexMerger {
         serializer: &mut InvertedIndexSerializer,
         fieldnorm_readers: FieldNormReaders,
         doc_id_mapping: &SegmentDocIdMapping,
-    ) -> crate::Result<HashMap<Field, TermOrdinalMapping>> {
-        let mut term_ordinal_mappings = HashMap::new();
+    ) -> crate::Result<()> {
         for (field, field_entry) in self.schema.fields() {
             let fieldnorm_reader = fieldnorm_readers.get_field(field)?;
             if field_entry.is_indexed() {
-                if let Some(term_ordinal_mapping) = self.write_postings_for_field(
+                self.write_postings_for_field(
                     field,
                     field_entry.field_type(),
                     serializer,
                     fieldnorm_reader,
                     doc_id_mapping,
-                )? {
-                    term_ordinal_mappings.insert(field, term_ordinal_mapping);
-                }
+                )?;
             }
         }
-        Ok(term_ordinal_mappings)
+        Ok(())
     }
 
     fn write_storable_fields(
@@ -726,7 +697,7 @@ impl IndexMerger {
             if self.is_disjunct_and_sorted_on_sort_property(sort_by_field)? {
                 self.get_doc_id_from_concatenated_data()?
             } else {
-                self.generate_doc_id_mapping(sort_by_field)?
+                self.generate_doc_id_mapping_with_sort_by_field(sort_by_field)?
             }
         } else {
             self.get_doc_id_from_concatenated_data()?
@@ -740,19 +711,17 @@ impl IndexMerger {
             .segment()
             .open_read(SegmentComponent::FieldNorms)?;
         let fieldnorm_readers = FieldNormReaders::open(fieldnorm_data)?;
-        let term_ord_mappings = self.write_postings(
+        self.write_postings(
             serializer.get_postings_serializer(),
             fieldnorm_readers,
             &doc_id_mapping,
         )?;
-        debug!("write-fastfields");
-        self.write_fast_fields(
-            serializer.get_fast_field_write(),
-            term_ord_mappings,
-            &doc_id_mapping,
-        )?;
+
         debug!("write-storagefields");
         self.write_storable_fields(serializer.get_store_writer(), &doc_id_mapping)?;
+        debug!("write-fastfields");
+        self.write_fast_fields(serializer.get_fast_field_write(), doc_id_mapping)?;
+
         debug!("close-serializer");
         serializer.close()?;
         Ok(self.max_doc)
@@ -761,11 +730,14 @@ impl IndexMerger {
 
 #[cfg(test)]
 mod tests {
-    use byteorder::{BigEndian, ReadBytesExt};
+
+    use columnar::Column;
     use schema::FAST;
 
-    use crate::collector::tests::{FastFieldTestCollector, TEST_COLLECTOR_WITH_SCORE};
-    use crate::collector::Count;
+    use crate::collector::tests::{
+        BytesFastFieldTestCollector, FastFieldTestCollector, TEST_COLLECTOR_WITH_SCORE,
+    };
+    use crate::collector::{Count, FacetCollector};
     use crate::core::Index;
     use crate::query::{AllQuery, BooleanQuery, EnableScoring, Scorer, TermQuery};
     use crate::schema::{
@@ -774,7 +746,7 @@ mod tests {
     };
     use crate::time::OffsetDateTime;
     use crate::{
-        assert_nearly_equals, schema, DateTime, DocAddress, DocSet, IndexSettings,
+        assert_nearly_equals, schema, DateTime, DocAddress, DocId, DocSet, IndexSettings,
         IndexSortByField, IndexWriter, Order, Searcher, SegmentId,
     };
 
@@ -907,27 +879,27 @@ mod tests {
                 );
             }
 
-            // {
-            //     let get_fast_vals = |terms: Vec<Term>| {
-            //         let query = BooleanQuery::new_multiterms_query(terms);
-            //         searcher.search(&query, &FastFieldTestCollector::for_field(score_field))
-            //     };
-            //     let get_fast_vals_bytes = |terms: Vec<Term>| {
-            //         let query = BooleanQuery::new_multiterms_query(terms);
-            //         searcher.search(
-            //             &query,
-            //             &BytesFastFieldTestCollector::for_field(bytes_score_field),
-            //         )
-            //     };
-            //     assert_eq!(
-            //         get_fast_vals(vec![Term::from_field_text(text_field, "a")])?,
-            //         vec![5, 7, 13]
-            //     );
-            //     assert_eq!(
-            //         get_fast_vals_bytes(vec![Term::from_field_text(text_field, "a")])?,
-            //         vec![0, 0, 0, 5, 0, 0, 0, 7, 0, 0, 0, 13]
-            //     );
-            // }
+            {
+                let get_fast_vals = |terms: Vec<Term>| {
+                    let query = BooleanQuery::new_multiterms_query(terms);
+                    searcher.search(&query, &FastFieldTestCollector::for_field("score"))
+                };
+                let get_fast_vals_bytes = |terms: Vec<Term>| {
+                    let query = BooleanQuery::new_multiterms_query(terms);
+                    searcher.search(
+                        &query,
+                        &BytesFastFieldTestCollector::for_field("score_bytes"),
+                    )
+                };
+                assert_eq!(
+                    get_fast_vals(vec![Term::from_field_text(text_field, "a")])?,
+                    vec![5, 7, 13]
+                );
+                assert_eq!(
+                    get_fast_vals_bytes(vec![Term::from_field_text(text_field, "a")])?,
+                    vec![0, 0, 0, 5, 0, 0, 0, 7, 0, 0, 0, 13]
+                );
+            }
         }
         Ok(())
     }
@@ -1240,209 +1212,206 @@ mod tests {
         Ok(())
     }
 
-    // TODO re-enable
-    // #[test]
-    // fn test_merge_facets_sort_none() {
-    //     test_merge_facets(None, true)
-    // }
+    #[test]
+    fn test_merge_facets_sort_none() {
+        test_merge_facets(None, true)
+    }
 
-    // #[test]
-    // fn test_merge_facets_sort_asc() {
-    //     // In the merge case this will go through the doc_id mapping code
-    //     test_merge_facets(
-    //         Some(IndexSettings {
-    //             sort_by_field: Some(IndexSortByField {
-    //                 field: "intval".to_string(),
-    //                 order: Order::Desc,
-    //             }),
-    //             ..Default::default()
-    //         }),
-    //         true,
-    //     );
-    //     // In the merge case this will not go through the doc_id mapping code, because the data
-    // is     // sorted and disjunct
-    //     test_merge_facets(
-    //         Some(IndexSettings {
-    //             sort_by_field: Some(IndexSortByField {
-    //                 field: "intval".to_string(),
-    //                 order: Order::Desc,
-    //             }),
-    //             ..Default::default()
-    //         }),
-    //         false,
-    //     );
-    // }
+    #[test]
+    fn test_merge_facets_sort_asc() {
+        // In the merge case this will go through the doc_id mapping code
+        test_merge_facets(
+            Some(IndexSettings {
+                sort_by_field: Some(IndexSortByField {
+                    field: "intval".to_string(),
+                    order: Order::Desc,
+                }),
+                ..Default::default()
+            }),
+            true,
+        );
+        // In the merge case this will not go through the doc_id mapping code, because the data
+        // sorted and disjunct
+        test_merge_facets(
+            Some(IndexSettings {
+                sort_by_field: Some(IndexSortByField {
+                    field: "intval".to_string(),
+                    order: Order::Desc,
+                }),
+                ..Default::default()
+            }),
+            false,
+        );
+    }
 
-    // #[test]
-    // fn test_merge_facets_sort_desc() {
-    //     // In the merge case this will go through the doc_id mapping code
-    //     test_merge_facets(
-    //         Some(IndexSettings {
-    //             sort_by_field: Some(IndexSortByField {
-    //                 field: "intval".to_string(),
-    //                 order: Order::Desc,
-    //             }),
-    //             ..Default::default()
-    //         }),
-    //         true,
-    //     );
-    //     // In the merge case this will not go through the doc_id mapping code, because the data
-    // is     // sorted and disjunct
-    //     test_merge_facets(
-    //         Some(IndexSettings {
-    //             sort_by_field: Some(IndexSortByField {
-    //                 field: "intval".to_string(),
-    //                 order: Order::Desc,
-    //             }),
-    //             ..Default::default()
-    //         }),
-    //         false,
-    //     );
-    // }
+    #[test]
+    fn test_merge_facets_sort_desc() {
+        // In the merge case this will go through the doc_id mapping code
+        test_merge_facets(
+            Some(IndexSettings {
+                sort_by_field: Some(IndexSortByField {
+                    field: "intval".to_string(),
+                    order: Order::Desc,
+                }),
+                ..Default::default()
+            }),
+            true,
+        );
+        // In the merge case this will not go through the doc_id mapping code, because the data
+        // sorted and disjunct
+        test_merge_facets(
+            Some(IndexSettings {
+                sort_by_field: Some(IndexSortByField {
+                    field: "intval".to_string(),
+                    order: Order::Desc,
+                }),
+                ..Default::default()
+            }),
+            false,
+        );
+    }
 
     // force_segment_value_overlap forces the int value for sorting to have overlapping min and max
     // ranges between segments so that merge algorithm can't apply certain optimizations
-    // fn test_merge_facets(index_settings: Option<IndexSettings>, force_segment_value_overlap:
-    // bool) {     let mut schema_builder = schema::Schema::builder();
-    //     let facet_field = schema_builder.add_facet_field("facet", FacetOptions::default());
-    //     let int_options = NumericOptions::default()
-    //         .set_fast()
-    //         .set_indexed();
-    //     let int_field = schema_builder.add_u64_field("intval", int_options);
-    //     let mut index_builder = Index::builder().schema(schema_builder.build());
-    //     if let Some(settings) = index_settings {
-    //         index_builder = index_builder.settings(settings);
-    //     }
-    //     let index = index_builder.create_in_ram().unwrap();
-    //     // let index = Index::create_in_ram(schema_builder.build());
-    //     let reader = index.reader().unwrap();
-    //     let mut int_val = 0;
-    //     {
-    //         let mut index_writer = index.writer_for_tests().unwrap();
-    //         let index_doc =
-    //             |index_writer: &mut IndexWriter, doc_facets: &[&str], int_val: &mut u64| {
-    //                 let mut doc = Document::default();
-    //                 for facet in doc_facets {
-    //                     doc.add_facet(facet_field, Facet::from(facet));
-    //                 }
-    //                 doc.add_u64(int_field, *int_val);
-    //                 *int_val += 1;
-    //                 index_writer.add_document(doc).unwrap();
-    //             };
+    fn test_merge_facets(index_settings: Option<IndexSettings>, force_segment_value_overlap: bool) {
+        let mut schema_builder = schema::Schema::builder();
+        let facet_field = schema_builder.add_facet_field("facet", FacetOptions::default());
+        let int_options = NumericOptions::default().set_fast().set_indexed();
+        let int_field = schema_builder.add_u64_field("intval", int_options);
+        let mut index_builder = Index::builder().schema(schema_builder.build());
+        if let Some(settings) = index_settings {
+            index_builder = index_builder.settings(settings);
+        }
+        let index = index_builder.create_in_ram().unwrap();
+        // let index = Index::create_in_ram(schema_builder.build());
+        let reader = index.reader().unwrap();
+        let mut int_val = 0;
+        {
+            let mut index_writer = index.writer_for_tests().unwrap();
+            let index_doc =
+                |index_writer: &mut IndexWriter, doc_facets: &[&str], int_val: &mut u64| {
+                    let mut doc = Document::default();
+                    for facet in doc_facets {
+                        doc.add_facet(facet_field, Facet::from(facet));
+                    }
+                    doc.add_u64(int_field, *int_val);
+                    *int_val += 1;
+                    index_writer.add_document(doc).unwrap();
+                };
 
-    //         index_doc(
-    //             &mut index_writer,
-    //             &["/top/a/firstdoc", "/top/b"],
-    //             &mut int_val,
-    //         );
-    //         index_doc(
-    //             &mut index_writer,
-    //             &["/top/a/firstdoc", "/top/b", "/top/c"],
-    //             &mut int_val,
-    //         );
-    //         index_doc(&mut index_writer, &["/top/a", "/top/b"], &mut int_val);
-    //         index_doc(&mut index_writer, &["/top/a"], &mut int_val);
+            index_doc(
+                &mut index_writer,
+                &["/top/a/firstdoc", "/top/b"],
+                &mut int_val,
+            );
+            index_doc(
+                &mut index_writer,
+                &["/top/a/firstdoc", "/top/b", "/top/c"],
+                &mut int_val,
+            );
+            index_doc(&mut index_writer, &["/top/a", "/top/b"], &mut int_val);
+            index_doc(&mut index_writer, &["/top/a"], &mut int_val);
 
-    //         index_doc(&mut index_writer, &["/top/b", "/top/d"], &mut int_val);
-    //         if force_segment_value_overlap {
-    //             index_doc(&mut index_writer, &["/top/d"], &mut 0);
-    //             index_doc(&mut index_writer, &["/top/e"], &mut 10);
-    //             index_writer.commit().expect("committed");
-    //             index_doc(&mut index_writer, &["/top/a"], &mut 5); // 5 is between 0 - 10 so the
-    //                                                                // segments don' have disjunct
-    //                                                                // ranges
-    //         } else {
-    //             index_doc(&mut index_writer, &["/top/d"], &mut int_val);
-    //             index_doc(&mut index_writer, &["/top/e"], &mut int_val);
-    //             index_writer.commit().expect("committed");
-    //             index_doc(&mut index_writer, &["/top/a"], &mut int_val);
-    //         }
-    //         index_doc(&mut index_writer, &["/top/b"], &mut int_val);
-    //         index_doc(&mut index_writer, &["/top/c"], &mut int_val);
-    //         index_writer.commit().expect("committed");
+            index_doc(&mut index_writer, &["/top/b", "/top/d"], &mut int_val);
+            if force_segment_value_overlap {
+                index_doc(&mut index_writer, &["/top/d"], &mut 0);
+                index_doc(&mut index_writer, &["/top/e"], &mut 10);
+                index_writer.commit().expect("committed");
+                index_doc(&mut index_writer, &["/top/a"], &mut 5); // 5 is between 0 - 10 so the
+                                                                   // segments don' have disjunct
+                                                                   // ranges
+            } else {
+                index_doc(&mut index_writer, &["/top/d"], &mut int_val);
+                index_doc(&mut index_writer, &["/top/e"], &mut int_val);
+                index_writer.commit().expect("committed");
+                index_doc(&mut index_writer, &["/top/a"], &mut int_val);
+            }
+            index_doc(&mut index_writer, &["/top/b"], &mut int_val);
+            index_doc(&mut index_writer, &["/top/c"], &mut int_val);
+            index_writer.commit().expect("committed");
 
-    //         index_doc(&mut index_writer, &["/top/e", "/top/f"], &mut int_val);
-    //         index_writer.commit().expect("committed");
-    //     }
+            index_doc(&mut index_writer, &["/top/e", "/top/f"], &mut int_val);
+            index_writer.commit().expect("committed");
+        }
 
-    //     reader.reload().unwrap();
-    //     let test_searcher = |expected_num_docs: usize, expected: &[(&str, u64)]| {
-    //         let searcher = reader.searcher();
-    //         let mut facet_collector = FacetCollector::for_field(facet_field);
-    //         facet_collector.add_facet(Facet::from("/top"));
-    //         let (count, facet_counts) = searcher
-    //             .search(&AllQuery, &(Count, facet_collector))
-    //             .unwrap();
-    //         assert_eq!(count, expected_num_docs);
-    //         let facets: Vec<(String, u64)> = facet_counts
-    //             .get("/top")
-    //             .map(|(facet, count)| (facet.to_string(), count))
-    //             .collect();
-    //         assert_eq!(
-    //             facets,
-    //             expected
-    //                 .iter()
-    //                 .map(|&(facet_str, count)| (String::from(facet_str), count))
-    //                 .collect::<Vec<_>>()
-    //         );
-    //     };
-    //     test_searcher(
-    //         11,
-    //         &[
-    //             ("/top/a", 5),
-    //             ("/top/b", 5),
-    //             ("/top/c", 2),
-    //             ("/top/d", 2),
-    //             ("/top/e", 2),
-    //             ("/top/f", 1),
-    //         ],
-    //     );
-    //     // Merging the segments
-    //     {
-    //         let segment_ids = index
-    //             .searchable_segment_ids()
-    //             .expect("Searchable segments failed.");
-    //         let mut index_writer = index.writer_for_tests().unwrap();
-    //         index_writer
-    //             .merge(&segment_ids)
-    //             .wait()
-    //             .expect("Merging failed");
-    //         index_writer.wait_merging_threads().unwrap();
-    //         reader.reload().unwrap();
-    //         test_searcher(
-    //             11,
-    //             &[
-    //                 ("/top/a", 5),
-    //                 ("/top/b", 5),
-    //                 ("/top/c", 2),
-    //                 ("/top/d", 2),
-    //                 ("/top/e", 2),
-    //                 ("/top/f", 1),
-    //             ],
-    //         );
-    //     }
+        reader.reload().unwrap();
+        let test_searcher = |expected_num_docs: usize, expected: &[(&str, u64)]| {
+            let searcher = reader.searcher();
+            let mut facet_collector = FacetCollector::for_field("facet");
+            facet_collector.add_facet(Facet::from("/top"));
+            let (count, facet_counts) = searcher
+                .search(&AllQuery, &(Count, facet_collector))
+                .unwrap();
+            assert_eq!(count, expected_num_docs);
+            let facets: Vec<(String, u64)> = facet_counts
+                .get("/top")
+                .map(|(facet, count)| (facet.to_string(), count))
+                .collect();
+            assert_eq!(
+                facets,
+                expected
+                    .iter()
+                    .map(|&(facet_str, count)| (String::from(facet_str), count))
+                    .collect::<Vec<_>>()
+            );
+        };
+        test_searcher(
+            11,
+            &[
+                ("/top/a", 5),
+                ("/top/b", 5),
+                ("/top/c", 2),
+                ("/top/d", 2),
+                ("/top/e", 2),
+                ("/top/f", 1),
+            ],
+        );
+        // Merging the segments
+        {
+            let segment_ids = index
+                .searchable_segment_ids()
+                .expect("Searchable segments failed.");
+            let mut index_writer = index.writer_for_tests().unwrap();
+            index_writer
+                .merge(&segment_ids)
+                .wait()
+                .expect("Merging failed");
+            index_writer.wait_merging_threads().unwrap();
+            reader.reload().unwrap();
+            test_searcher(
+                11,
+                &[
+                    ("/top/a", 5),
+                    ("/top/b", 5),
+                    ("/top/c", 2),
+                    ("/top/d", 2),
+                    ("/top/e", 2),
+                    ("/top/f", 1),
+                ],
+            );
+        }
 
-    //     // Deleting one term
-    //     {
-    //         let mut index_writer = index.writer_for_tests().unwrap();
-    //         let facet = Facet::from_path(vec!["top", "a", "firstdoc"]);
-    //         let facet_term = Term::from_facet(facet_field, &facet);
-    //         index_writer.delete_term(facet_term);
-    //         index_writer.commit().unwrap();
-    //         reader.reload().unwrap();
-    //         test_searcher(
-    //             9,
-    //             &[
-    //                 ("/top/a", 3),
-    //                 ("/top/b", 3),
-    //                 ("/top/c", 1),
-    //                 ("/top/d", 2),
-    //                 ("/top/e", 2),
-    //                 ("/top/f", 1),
-    //             ],
-    //         );
-    //     }
-    // }
+        // Deleting one term
+        {
+            let mut index_writer = index.writer_for_tests().unwrap();
+            let facet = Facet::from_path(vec!["top", "a", "firstdoc"]);
+            let facet_term = Term::from_facet(facet_field, &facet);
+            index_writer.delete_term(facet_term);
+            index_writer.commit().unwrap();
+            reader.reload().unwrap();
+            test_searcher(
+                9,
+                &[
+                    ("/top/a", 3),
+                    ("/top/b", 3),
+                    ("/top/c", 1),
+                    ("/top/d", 2),
+                    ("/top/e", 2),
+                    ("/top/f", 1),
+                ],
+            );
+        }
+    }
 
     #[test]
     fn test_bug_merge() -> crate::Result<()> {
@@ -1511,6 +1480,13 @@ mod tests {
         let int_field = schema_builder.add_u64_field("intvals", int_options);
         let index = Index::create_in_ram(schema_builder.build());
 
+        let mut vals: Vec<u64> = Vec::new();
+        let mut test_vals = move |col: &Column<u64>, doc: DocId, expected: &[u64]| {
+            vals.clear();
+            vals.extend(col.values(doc));
+            assert_eq!(&vals[..], expected);
+        };
+
         {
             let mut index_writer = index.writer_for_tests()?;
             let index_doc = |index_writer: &mut IndexWriter, int_vals: &[u64]| {
@@ -1536,49 +1512,42 @@ mod tests {
         }
         let reader = index.reader()?;
         let searcher = reader.searcher();
-        let mut vals: Vec<u64> = Vec::new();
 
         {
             let segment = searcher.segment_reader(0u32);
-            // let ff_reader = segment.fast_fields().u64s(int_field).unwrap();
-
-            // ff_reader.get_vals(0, &mut vals);
-            // assert_eq!(&vals, &[1, 2]);
-
-            // ff_reader.get_vals(1, &mut vals);
-            // assert_eq!(&vals, &[1, 2, 3]);
-
-            // ff_reader.get_vals(2, &mut vals);
-            // assert_eq!(&vals, &[4, 5]);
-
-            // ff_reader.get_vals(3, &mut vals);
-            // assert_eq!(&vals, &[1, 2]);
-
-            // ff_reader.get_vals(4, &mut vals);
-            // assert_eq!(&vals, &[1, 5]);
-
-            // ff_reader.get_vals(5, &mut vals);
-            // assert_eq!(&vals, &[3]);
-
-            // ff_reader.get_vals(6, &mut vals);
-            // assert_eq!(&vals, &[17]);
+            let column = segment
+                .fast_fields()
+                .column_opt::<u64>("intvals")
+                .unwrap()
+                .unwrap();
+            test_vals(&column, 0, &[1, 2]);
+            test_vals(&column, 1, &[1, 2, 3]);
+            test_vals(&column, 2, &[4, 5]);
+            test_vals(&column, 3, &[1, 2]);
+            test_vals(&column, 4, &[1, 5]);
+            test_vals(&column, 5, &[3]);
+            test_vals(&column, 6, &[17]);
         }
 
         {
             let segment = searcher.segment_reader(1u32);
-            // let ff_reader = segment.fast_fields().u64s(int_field).unwrap();
-            // ff_reader.get_vals(0, &mut vals);
-            // assert_eq!(&vals, &[28, 27]);
-
-            // ff_reader.get_vals(1, &mut vals);
-            // assert_eq!(&vals, &[1_000]);
+            let col = segment
+                .fast_fields()
+                .column_opt::<u64>("intvals")
+                .unwrap()
+                .unwrap();
+            test_vals(&col, 0, &[28, 27]);
+            test_vals(&col, 1, &[1000]);
         }
 
         {
             let segment = searcher.segment_reader(2u32);
-            // let ff_reader = segment.fast_fields().u64s(int_field).unwrap();
-            // ff_reader.get_vals(0, &mut vals);
-            // assert_eq!(&vals, &[20]);
+            let col = segment
+                .fast_fields()
+                .column_opt::<u64>("intvals")
+                .unwrap()
+                .unwrap();
+            test_vals(&col, 0, &[20]);
         }
 
         // Merging the segments
@@ -1593,37 +1562,21 @@ mod tests {
         {
             let searcher = reader.searcher();
             let segment = searcher.segment_reader(0u32);
-            // let ff_reader = segment.fast_fields().u64s(int_field).unwrap();
-
-            // ff_reader.get_vals(0, &mut vals);
-            // assert_eq!(&vals, &[1, 2]);
-
-            // ff_reader.get_vals(1, &mut vals);
-            // assert_eq!(&vals, &[1, 2, 3]);
-
-            // ff_reader.get_vals(2, &mut vals);
-            // assert_eq!(&vals, &[4, 5]);
-
-            // ff_reader.get_vals(3, &mut vals);
-            // assert_eq!(&vals, &[1, 2]);
-
-            // ff_reader.get_vals(4, &mut vals);
-            // assert_eq!(&vals, &[1, 5]);
-
-            // ff_reader.get_vals(5, &mut vals);
-            // assert_eq!(&vals, &[3]);
-
-            // ff_reader.get_vals(6, &mut vals);
-            // assert_eq!(&vals, &[17]);
-
-            // ff_reader.get_vals(7, &mut vals);
-            // assert_eq!(&vals, &[28, 27]);
-
-            // ff_reader.get_vals(8, &mut vals);
-            // assert_eq!(&vals, &[1_000]);
-
-            // ff_reader.get_vals(9, &mut vals);
-            // assert_eq!(&vals, &[20]);
+            let col = segment
+                .fast_fields()
+                .column_opt::<u64>("intvals")
+                .unwrap()
+                .unwrap();
+            test_vals(&col, 0, &[1, 2]);
+            test_vals(&col, 1, &[1, 2, 3]);
+            test_vals(&col, 2, &[4, 5]);
+            test_vals(&col, 3, &[1, 2]);
+            test_vals(&col, 4, &[1, 5]);
+            test_vals(&col, 5, &[3]);
+            test_vals(&col, 6, &[17]);
+            test_vals(&col, 7, &[28, 27]);
+            test_vals(&col, 8, &[1000]);
+            test_vals(&col, 9, &[20]);
         }
         Ok(())
     }

src/indexer/merger_sorted_index_test.rs

@@ -155,6 +155,7 @@ mod tests {
     fn test_merge_sorted_index_desc_not_disjunct() {
         test_merge_sorted_index_desc_(false);
     }
+
     #[test]
     fn test_merge_sorted_index_desc_disjunct() {
         test_merge_sorted_index_desc_(true);
@@ -476,12 +477,11 @@ mod bench_sorted_index_merge {
 
     use std::sync::Arc;
 
-    use fastfield_codecs::Column;
     use test::{self, Bencher};
 
     use crate::core::Index;
     use crate::indexer::merger::IndexMerger;
-    use crate::schema::{Cardinality, NumericOptions, Schema};
+    use crate::schema::{NumericOptions, Schema};
     use crate::{IndexSettings, IndexSortByField, IndexWriter, Order};
     fn create_index(sort_by_field: Option<IndexSortByField>) -> Index {
         let mut schema_builder = Schema::builder();
@@ -512,42 +512,42 @@ mod bench_sorted_index_merge {
         index
     }
 
-    #[bench]
-    fn create_sorted_index_walk_overkmerge_on_merge_fastfield(
-        b: &mut Bencher,
-    ) -> crate::Result<()> {
-        let sort_by_field = IndexSortByField {
-            field: "intval".to_string(),
-            order: Order::Desc,
-        };
-        let index = create_index(Some(sort_by_field.clone()));
-        let segments = index.searchable_segments().unwrap();
-        let merger: IndexMerger =
-            IndexMerger::open(index.schema(), index.settings().clone(), &segments[..])?;
-        let doc_id_mapping = merger.generate_doc_id_mapping(&sort_by_field).unwrap();
-        b.iter(|| {
-            let sorted_doc_ids = doc_id_mapping.iter_old_doc_addrs().map(|doc_addr| {
-                let reader = &merger.readers[doc_addr.segment_ord as usize];
-                let u64_reader: Arc<dyn Column<u64>> = reader
-                    .fast_fields()
-                    .typed_fast_field_reader("intval")
-                    .expect(
-                        "Failed to find a reader for single fast field. This is a tantivy bug and \
-                         it should never happen.",
-                    );
-                (doc_addr.doc_id, reader, u64_reader)
-            });
-            // add values in order of the new doc_ids
-            let mut val = 0;
-            for (doc_id, _reader, field_reader) in sorted_doc_ids {
-                val = field_reader.get_val(doc_id);
-            }
+    //#[bench]
+    // fn create_sorted_index_walk_overkmerge_on_merge_fastfield(
+    // b: &mut Bencher,
+    //) -> crate::Result<()> {
+    // let sort_by_field = IndexSortByField {
+    // field: "intval".to_string(),
+    // order: Order::Desc,
+    //};
+    // let index = create_index(Some(sort_by_field.clone()));
+    // let segments = index.searchable_segments().unwrap();
+    // let merger: IndexMerger =
+    // IndexMerger::open(index.schema(), index.settings().clone(), &segments[..])?;
+    // let doc_id_mapping = merger.generate_doc_id_mapping(&sort_by_field).unwrap();
+    // b.iter(|| {
+    // let sorted_doc_ids = doc_id_mapping.iter_old_doc_addrs().map(|doc_addr| {
+    // let reader = &merger.readers[doc_addr.segment_ord as usize];
+    // let u64_reader: Arc<dyn Column<u64>> = reader
+    //.fast_fields()
+    //.typed_fast_field_reader("intval")
+    //.expect(
+    //"Failed to find a reader for single fast field. This is a tantivy bug and \
+    // it should never happen.",
+    //);
+    //(doc_addr.doc_id, reader, u64_reader)
+    //});
+    //// add values in order of the new doc_ids
+    // let mut val = 0;
+    // for (doc_id, _reader, field_reader) in sorted_doc_ids {
+    // val = field_reader.get_val(doc_id);
+    //}
 
-            val
-        });
+    // val
+    //});
 
-        Ok(())
-    }
+    // Ok(())
+    //}
     #[bench]
     fn create_sorted_index_create_doc_id_mapping(b: &mut Bencher) -> crate::Result<()> {
         let sort_by_field = IndexSortByField {

src/indexer/mod.rs

@@ -19,8 +19,6 @@ mod segment_register;
 pub mod segment_serializer;
 pub mod segment_updater;
 mod segment_writer;
-// mod sorted_doc_id_column;
-// mod sorted_doc_id_multivalue_column;
 mod stamper;
 
 use crossbeam_channel as channel;
@@ -57,9 +55,9 @@ type AddBatchReceiver = channel::Receiver<AddBatch>;
 #[cfg(feature = "mmap")]
 #[cfg(test)]
 mod tests_mmap {
-    use crate::collector::Count;
+
     // use crate::query::QueryParser;
-    use crate::schema::{JsonObjectOptions, Schema, TEXT};
+    use crate::schema::{Schema, TEXT};
     use crate::{Index, Term};
 
     #[test]

src/indexer/segment_writer.rs

@@ -819,20 +819,23 @@ mod tests {
         // This is a bit of a contrived example.
         let tokens = PreTokenizedString {
             text: "contrived-example".to_string(), //< I can't think of a use case where this corner case happens in real life.
-            tokens: vec![Token { // Not the last token, yet ends after the last token.
-                offset_from: 0,
-                offset_to: 14,
-                position: 0,
-                text: "long_token".to_string(),
-                position_length: 3,
-            },
-            Token {
-                offset_from: 0,
-                offset_to: 14,
-                position: 1,
-                text: "short".to_string(),
-                position_length: 1,
-            }],
+            tokens: vec![
+                Token {
+                    // Not the last token, yet ends after the last token.
+                    offset_from: 0,
+                    offset_to: 14,
+                    position: 0,
+                    text: "long_token".to_string(),
+                    position_length: 3,
+                },
+                Token {
+                    offset_from: 0,
+                    offset_to: 14,
+                    position: 1,
+                    text: "short".to_string(),
+                    position_length: 1,
+                },
+            ],
         };
         doc.add_pre_tokenized_text(text, tokens);
         doc.add_text(text, "hello");

src/indexer/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
-    }
-}

src/lib.rs

@@ -279,7 +279,7 @@ mod indexer;
 pub mod error;
 pub mod tokenizer;
 
-// pub mod aggregation;
+pub mod aggregation;
 pub mod collector;
 pub mod directory;
 pub mod fastfield;