perf: use term ordinal order when sorting by keys

term ordinals are sorted lexicographically. we can use than to sort instead of reading the terms from the dictionary.

Cargo.toml

@@ -55,11 +55,11 @@ 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" }
 tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
 

bitpacker/src/bitpacker.rs

@@ -19,7 +19,7 @@ impl BitPacker {
     }
 
     #[inline]
-    pub fn write<TWrite: io::Write + ?Sized>(
+    pub fn write<TWrite: io::Write>(
         &mut self,
         val: u64,
         num_bits: u8,
@@ -43,7 +43,7 @@ impl BitPacker {
         Ok(())
     }
 
-    pub fn flush<TWrite: io::Write + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn flush<TWrite: io::Write>(&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 + ?Sized>(&mut self, output: &mut TWrite) -> io::Result<()> {
+    pub fn close<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
         self.flush(output)?;
         Ok(())
     }

columnar/Cargo.toml

@@ -5,19 +5,18 @@ 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"

columnar/benches/bench.rs

@@ -0,0 +1,311 @@
+#![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_u128.rs

@@ -1,129 +0,0 @@
-#![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.iter().copied()),
-        data.len() as u32,
-        &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;
-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
-}
-
-#[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

@@ -1,213 +0,0 @@
-#![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, ALL_U64_CODEC_TYPES,
-};
-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.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/src/TODO.md

@@ -1,12 +1,17 @@
 # zero to one
-* merges with non trivial mapping (deletes / sort)
-* emission of the sort mapping.
-+ muttivaued range queries restrat frm the beginning all of the time.
-* revisit line codec
-* removal of all rows of a column in the schema due to deletes
-* Plugging JSON
-replug examples
-
+* 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.
 
 # Perf and Size
 * re-add ZSTD compression for dictionaries
@@ -14,6 +19,7 @@ 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
@@ -21,8 +27,11 @@ compare with roaring bitmap/elias fano etc etc.
 SIMD range? (see blog post)
 Add alignment?
 Consider another codec to bridge the gap between few and 5k elements
+replug examples
+replug fast_field_codecs bench
 
 # 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
@@ -33,13 +42,16 @@ 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

@@ -62,12 +62,6 @@ impl From<BytesColumn> for StrColumn {
     }
 }
 
-impl From<StrColumn> for BytesColumn {
-    fn from(str_column: StrColumn) -> BytesColumn {
-        str_column.0
-    }
-}
-
 impl StrColumn {
     pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
         self.0.dictionary.as_ref()

columnar/src/column/mod.rs

@@ -2,7 +2,6 @@ mod dictionary_encoded;
 mod serialize;
 
 use std::fmt::Debug;
-use std::io::Write;
 use std::ops::Deref;
 use std::sync::Arc;
 
@@ -14,33 +13,16 @@ pub use serialize::{
 };
 
 use crate::column_index::ColumnIndex;
-use crate::column_values::monotonic_mapping::StrictlyMonotonicMappingToInternal;
-use crate::column_values::{monotonic_map_column, ColumnValues};
-use crate::{Cardinality, MonotonicallyMappableToU64, RowId};
+use crate::column_values::ColumnValues;
+use crate::{Cardinality, RowId};
 
 #[derive(Clone)]
-pub struct Column<T = u64> {
+pub struct Column<T> {
     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,
@@ -87,7 +69,7 @@ impl<T> Deref for Column<T> {
 }
 
 impl BinarySerializable for Cardinality {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> std::io::Result<()> {
+    fn serialize<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
         self.to_code().serialize(writer)
     }
 

columnar/src/column/serialize.rs

@@ -9,36 +9,43 @@ 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::u64_based::{serialize_u64_based_column_values, CodecType};
-use crate::column_values::{MonotonicallyMappableToU128, MonotonicallyMappableToU64};
-use crate::iterable::{map_iterable, Iterable};
+use crate::column_values::{
+    serialize_column_values, ColumnValues, FastFieldCodecType, MonotonicallyMappableToU128,
+    MonotonicallyMappableToU64,
+};
 
-pub fn serialize_column_mappable_to_u128<I, T: MonotonicallyMappableToU128>(
+pub fn serialize_column_mappable_to_u128<
+    F: Fn() -> I,
+    I: Iterator<Item = T>,
+    T: MonotonicallyMappableToU128,
+>(
     column_index: SerializableColumnIndex<'_>,
-    iterable: &dyn Fn() -> I,
+    column_values: F,
     num_vals: u32,
     output: &mut impl Write,
-) -> io::Result<()>
-where
-    I: Iterator<Item = T>,
-{
+) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    let u128_iterable = map_iterable(iterable, MonotonicallyMappableToU128::to_u128);
-    serialize_column_values_u128(&u128_iterable, num_vals, output)?;
+    serialize_column_values_u128(
+        || column_values().map(|val| val.to_u128()),
+        num_vals,
+        output,
+    )?;
     output.write_all(&column_index_num_bytes.to_le_bytes())?;
     Ok(())
 }
 
-pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64 + Debug, I>(
+pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64 + Debug>(
     column_index: SerializableColumnIndex<'_>,
-    column_values: &dyn Fn() -> I,
+    column_values: &impl ColumnValues<T>,
     output: &mut impl Write,
-) -> io::Result<()>
-where I: Iterator<Item=T> {
+) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_u64_based_column_values(
+    serialize_column_values(
         column_values,
-        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
+        &[
+            FastFieldCodecType::Bitpacked,
+            FastFieldCodecType::BlockwiseLinear,
+        ],
         output,
     )?;
     output.write_all(&column_index_num_bytes.to_le_bytes())?;
@@ -55,8 +62,7 @@ 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::u64_based::load_u64_based_column_values(column_values_data)?;
+    let column_values = crate::column_values::open_u64_mapped(column_values_data)?;
     Ok(Column {
         idx: column_index,
         values: column_values,

columnar/src/column_index/merge.rs

@@ -1,174 +0,0 @@
-use std::iter;
-
-use crate::column_index::{
-    multivalued_index, serialize_column_index, SerializableColumnIndex, Set,
-};
-use crate::iterable::Iterable;
-use crate::{Cardinality, ColumnIndex, MergeRowOrder, RowId, StackMergeOrder};
-
-fn detect_cardinality(columns: &[Option<ColumnIndex>]) -> Cardinality {
-    columns
-        .iter()
-        .flatten()
-        .map(ColumnIndex::get_cardinality)
-        .max()
-        .unwrap_or(Cardinality::Full)
-}
-
-pub fn stack_column_index<'a>(
-    columns: &'a [Option<ColumnIndex>],
-    merge_row_order: &'a MergeRowOrder,
-) -> SerializableColumnIndex<'a> {
-    let MergeRowOrder::Stack(stack_merge_order) = merge_row_order else {
-        panic!("only supporting stacking at the moment.");
-    };
-    let cardinality = detect_cardinality(columns);
-    match cardinality {
-        Cardinality::Full => SerializableColumnIndex::Full,
-        Cardinality::Optional =>  {
-            let stacked_optional_index: StackedOptionalIndex<'a> = StackedOptionalIndex {
-                columns,
-                stack_merge_order,
-            };
-            SerializableColumnIndex::Optional {
-                non_null_row_ids: Box::new(move || Box::new(stacked_optional_index.iter())),
-                num_rows: stack_merge_order.num_rows(),
-            }
-        },
-        Cardinality::Multivalued => {
-            let stacked_multivalued_index = StackedMultivaluedIndex {
-                columns,
-                stack_merge_order,
-            };
-            SerializableColumnIndex::Multivalued(Box::new(move || stacked_multivalued_index.boxed_iter()))
-        }
-    }
-}
-
-struct StackedOptionalIndex<'a> {
-    columns: &'a [Option<ColumnIndex>],
-    stack_merge_order: &'a StackMergeOrder,
-}
-
-impl<'a> StackedOptionalIndex<'a> {
-    fn iter(&self) -> impl 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| row_id + columnar_row_range.start),
-                        ),
-                        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> StackedMultivaluedIndex<'a> {
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
-        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;
-            }
-        }
-    }))
-}
-
-fn stack_multivalued_index<'a>(
-    columns: &'a [Option<ColumnIndex>],
-    stack_merge_order: &StackMergeOrder,
-) -> Box<dyn Iterable<RowId> + 'a> {
-    todo!()
-}
-
-#[cfg(test)]
-mod tests {
-    use crate::RowId;
-
-    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

@@ -1,12 +1,10 @@
-mod merge;
 mod multivalued_index;
 mod optional_index;
 mod serialize;
 
 use std::ops::Range;
 
-pub use merge::stack_column_index;
-pub use optional_index::{OptionalIndex, Set};
+pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
 pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
 
 use crate::column_index::multivalued_index::MultiValueIndex;

columnar/src/column_index/multivalued_index.rs

@@ -5,17 +5,16 @@ use std::sync::Arc;
 
 use common::OwnedBytes;
 
-use crate::column_values::u64_based::CodecType;
-use crate::column_values::ColumnValues;
+use crate::column_values::{ColumnValues, FastFieldCodecType};
 use crate::RowId;
 
-pub fn serialize_multivalued_index<'a>(
-    multivalued_index: &'a dyn Fn() -> Box<dyn Iterator<Item=RowId> + 'a>,
+pub fn serialize_multivalued_index(
+    multivalued_index: &dyn ColumnValues<RowId>,
     output: &mut impl Write,
 ) -> io::Result<()> {
-    crate::column_values::u64_based::serialize_u64_based_column_values(
-        multivalued_index,
-        &[CodecType::Bitpacked, CodecType::Linear],
+    crate::column_values::serialize_column_values(
+        &*multivalued_index,
+        &[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
         output,
     )?;
     Ok(())
@@ -23,7 +22,7 @@ pub fn serialize_multivalued_index<'a>(
 
 pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex> {
     let start_index_column: Arc<dyn ColumnValues<RowId>> =
-        crate::column_values::u64_based::load_u64_based_column_values(bytes)?;
+        crate::column_values::open_u64_mapped(bytes)?;
     Ok(MultiValueIndex { start_index_column })
 }
 
@@ -31,7 +30,7 @@ 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 {
-    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
+    start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
 }
 
 impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {

columnar/src/column_index/optional_index/mod.rs

@@ -1,4 +1,5 @@
 use std::io::{self, Write};
+use std::ops::Range;
 use std::sync::Arc;
 
 mod set;
@@ -10,7 +11,6 @@ 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.
@@ -96,12 +96,6 @@ impl OptionalIndex {
     pub fn num_non_nulls(&self) -> RowId {
         self.num_non_null_rows
     }
-
-    pub fn iter_rows<'a>(&'a self) -> impl Iterator<Item = RowId> + 'a {
-        // TODO optimize
-        let mut select_batch = self.select_cursor();
-        (0..self.num_non_null_rows).map(move |rank| select_batch.select(rank))
-    }
 }
 
 /// Splits a value address into lower and upper 16bits.
@@ -192,21 +186,6 @@ 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 {
@@ -321,7 +300,7 @@ impl OptionalIndexCodec {
 }
 
 impl BinarySerializable for OptionalIndexCodec {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_all(&[self.to_code()])
     }
 
@@ -343,13 +322,12 @@ fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -
 }
 
 pub fn serialize_optional_index<'a, W: io::Write>(
-    non_null_rows: &dyn Fn() -> Box<dyn Iterator<Item=RowId> + 'a>,
-    num_rows: RowId,
+    serializable_optional_index: &dyn SerializableOptionalIndex<'a>,
     output: &mut W,
 ) -> io::Result<()> {
-    VInt(num_rows as u64).serialize(output)?;
+    VInt(serializable_optional_index.num_rows() as u64).serialize(output)?;
 
-    let mut rows_it = non_null_rows();
+    let mut rows_it = serializable_optional_index.non_null_rows();
     let mut block_metadata: Vec<SerializedBlockMeta> = Vec::new();
     let mut current_block = Vec::new();
 
@@ -502,5 +480,19 @@ 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

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

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

@@ -148,15 +148,6 @@ 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

@@ -44,11 +44,6 @@ 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

@@ -1,7 +1,8 @@
 use std::collections::HashMap;
 
-use crate::column_index::optional_index::set_block::dense::DENSE_BLOCK_NUM_BYTES;
-use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
+use crate::column_index::optional_index::set_block::{
+    DenseBlockCodec, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
+};
 use crate::column_index::optional_index::{SelectCursor, Set, SetCodec};
 
 fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
@@ -17,10 +18,6 @@ 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(), 100, &mut out).unwrap();
+    serialize_optional_index(&vals.clone(), &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
     let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
@@ -66,8 +66,12 @@ fn test_optional_index_one_block_true() {
     test_null_index(&iter[..]);
 }
 
-impl<'a> Iterable<RowId> for &'a [bool] {
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
+    fn num_rows(&self) -> RowId {
+        self.len() as u32
+    }
+
+    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
         Box::new(
             self.iter()
                 .cloned()
@@ -80,7 +84,7 @@ impl<'a> Iterable<RowId> for &'a [bool] {
 
 fn test_null_index(data: &[bool]) {
     let mut out: Vec<u8> = Vec::new();
-    serialize_optional_index(&data, data.len() as RowId, &mut out).unwrap();
+    serialize_optional_index(&data, &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let orig_idx_with_value: Vec<u32> = data
         .iter()
@@ -109,7 +113,7 @@ fn test_null_index(data: &[bool]) {
 fn test_optional_index_test_translation() {
     let mut out = vec![];
     let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as u32, &mut out).unwrap();
+    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();
     assert_eq!(select_cursor.select(0), 0);
@@ -120,7 +124,7 @@ fn test_optional_index_test_translation() {
 fn test_optional_index_translate() {
     let mut out = vec![];
     let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as RowId, &mut out).unwrap();
+    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));
@@ -130,7 +134,7 @@ fn test_optional_index_translate() {
 fn test_optional_index_small() {
     let mut out = vec![];
     let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], iter.len() as RowId, &mut out).unwrap();
+    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));
@@ -145,7 +149,7 @@ fn test_optional_index_large() {
     docs.extend((0..=1).map(|_idx| true));
 
     let mut out = vec![];
-    serialize_optional_index(&&docs[..], docs.len() as RowId, &mut out).unwrap();
+    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));
@@ -154,59 +158,6 @@ fn test_optional_index_large() {
     assert!(null_index.contains(ELEMENTS_PER_BLOCK + 1));
 }
 
-fn test_optional_index_iter_aux(row_ids: &[RowId], num_rows: RowId) {
-    let mut buffer: Vec<u8> = Vec::new();
-    serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(buffer)).unwrap();
-    assert_eq!(null_index.num_rows(), num_rows);
-    assert!(null_index.iter_rows().eq(row_ids.iter().copied()));
-}
-
-#[test]
-fn test_optional_index_iter_empty() {
-    test_optional_index_iter_aux(&[], 0u32);
-}
-
-fn test_optional_index_rank_aux(row_ids: &[RowId]) {
-    let mut buffer: Vec<u8> = Vec::new();
-    let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
-    serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(buffer)).unwrap();
-    assert_eq!(null_index.num_rows(), num_rows);
-    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);
-}
-
 #[cfg(all(test, feature = "unstable"))]
 mod bench {
 

columnar/src/column_index/serialize.rs

@@ -3,35 +3,32 @@ use std::io::Write;
 
 use common::{CountingWriter, OwnedBytes};
 
-use crate::column_index::multivalued_index::{serialize_multivalued_index, self};
+use crate::column_index::multivalued_index::serialize_multivalued_index;
 use crate::column_index::optional_index::serialize_optional_index;
-use crate::column_index::ColumnIndex;
-use crate::iterable::Iterable;
+use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
+use crate::column_values::ColumnValues;
 use crate::{Cardinality, RowId};
 
 pub enum SerializableColumnIndex<'a> {
     Full,
-    Optional {
-        non_null_row_ids: Box<dyn Fn() -> Box<dyn Iterator<Item=RowId> + 'a> + 'a>,
-        num_rows: RowId,
-    },
+    Optional(Box<dyn SerializableOptionalIndex<'a> + 'a>),
     // TODO remove the Arc<dyn> apart from serialization this is not
     // dynamic at all.
-    Multivalued(&'a dyn Fn() -> Box<dyn Iterator<Item=RowId> + 'a>),
+    Multivalued(Box<dyn ColumnValues<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,
         }
     }
 }
 
-pub fn serialize_column_index<'a>(
-    column_index: SerializableColumnIndex<'a>,
+pub fn serialize_column_index(
+    column_index: SerializableColumnIndex,
     output: &mut impl Write,
 ) -> io::Result<u32> {
     let mut output = CountingWriter::wrap(output);
@@ -39,13 +36,11 @@ pub fn serialize_column_index<'a>(
     output.write_all(&[cardinality])?;
     match column_index {
         SerializableColumnIndex::Full => {}
-        SerializableColumnIndex::Optional {
-            non_null_row_ids,
-            num_rows,
-        } => serialize_optional_index(non_null_row_ids.as_ref(), num_rows, &mut output)?,
+        SerializableColumnIndex::Optional(optional_index) => {
+            serialize_optional_index(&*optional_index, &mut output)?
+        }
         SerializableColumnIndex::Multivalued(multivalued_index) => {
-            let multivalued_index_ref: &'a dyn Fn() -> Box<dyn Iterator<Item=RowId> + 'a> = multivalued_index.as_ref();
-            serialize_multivalued_index(multivalued_index_ref, &mut output)?
+            serialize_multivalued_index(&*multivalued_index, &mut output)?
         }
     }
     let column_index_num_bytes = output.written_bytes() as u32;

columnar/src/column_values/bitpacked.rs

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

columnar/src/column_values/blockwise_linear.rs

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

columnar/src/column_values/column.rs

@@ -1,17 +1,15 @@
 use std::fmt::Debug;
 use std::marker::PhantomData;
 use std::ops::{Range, RangeInclusive};
-use std::sync::Arc;
 
 use tantivy_bitpacker::minmax;
 
 use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
-use crate::iterable::Iterable;
 
 /// `ColumnValues` provides access to a dense field column.
 ///
 /// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
-pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
+pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
     /// Return the value associated with the given idx.
     ///
     /// This accessor should return as fast as possible.
@@ -29,7 +27,7 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     ///
     /// Must panic if `start + output.len()` is greater than
     /// the segment's `maxdoc`.
-    #[inline(always)]
+    #[inline]
     fn get_range(&self, start: u64, output: &mut [T]) {
         for (out, idx) in output.iter_mut().zip(start..) {
             *out = self.get_val(idx as u32);
@@ -39,7 +37,7 @@ pub trait ColumnValues<T: PartialOrd = 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(always)]
+    #[inline]
     fn get_docids_for_value_range(
         &self,
         value_range: RangeInclusive<T>,
@@ -80,33 +78,27 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     }
 }
 
-impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
-    #[inline(always)]
+impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for std::sync::Arc<dyn ColumnValues<T>> {
     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

@@ -55,7 +55,7 @@ impl RangeMapping {
 }
 
 impl BinarySerializable for CompactSpace {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
 
         let mut prev_value = 0;
@@ -247,7 +247,7 @@ pub struct CompactSpaceDecompressor {
 }
 
 impl BinarySerializable for IPCodecParams {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         // header flags for future optional dictionary encoding
         let footer_flags = 0u64;
         footer_flags.serialize(writer)?;

columnar/src/column_values/gcd.rs

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

columnar/src/column_values/line.rs

@@ -17,8 +17,8 @@ const MID_POINT: u64 = (1u64 << 32) - 1u64;
 /// `y = m * x >> 32 + b`
 #[derive(Debug, Clone, Copy, Default)]
 pub struct Line {
-    pub(crate) slope: u64,
-    pub(crate) intercept: u64,
+    slope: u64,
+    intercept: u64,
 }
 
 /// Compute the line slope.
@@ -81,7 +81,7 @@ impl Line {
     }
 
     // Intercept is only computed from provided positions
-    pub fn train_from(
+    fn train_from(
         first_val: u64,
         last_val: u64,
         num_vals: u32,
@@ -145,7 +145,6 @@ 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();
@@ -159,7 +158,7 @@ impl Line {
 }
 
 impl BinarySerializable for Line {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.slope).serialize(writer)?;
         VInt(self.intercept).serialize(writer)?;
         Ok(())

columnar/src/column_values/linear.rs

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

columnar/src/column_values/mod.rs

@@ -7,6 +7,9 @@
 //! - 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;
@@ -15,27 +18,75 @@ use std::sync::Arc;
 use common::{BinarySerializable, OwnedBytes};
 use compact_space::CompactSpaceDecompressor;
 pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
-use monotonic_mapping::{StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal};
+use monotonic_mapping::{
+    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
+    StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
+};
 pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
-use serialize::U128Header;
+use serialize::{Header, 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;
+
+#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
+#[repr(u8)]
+/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
+pub enum FastFieldCodecType {
+    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
+    /// `column.max_value() - column.min_value()`
+    Bitpacked = 1,
+    /// Linear interpolation puts a line between the first and last value and then bitpacks the
+    /// values by the offset from the line. The number of bits is defined by the max deviation from
+    /// the line.
+    Linear = 2,
+    /// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
+    BlockwiseLinear = 3,
+}
+
+impl BinarySerializable for FastFieldCodecType {
+    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
+        self.to_code().serialize(wrt)
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let code = u8::deserialize(reader)?;
+        let codec_type: Self = Self::from_code(code)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
+        Ok(codec_type)
+    }
+}
+
+impl FastFieldCodecType {
+    pub(crate) fn to_code(self) -> u8 {
+        self as u8
+    }
+
+    pub(crate) fn from_code(code: u8) -> Option<Self> {
+        match code {
+            1 => Some(Self::Bitpacked),
+            2 => Some(Self::Linear),
+            3 => Some(Self::BlockwiseLinear),
+            _ => None,
+        }
+    }
+}
 
 #[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
 #[repr(u8)]
@@ -47,7 +98,7 @@ pub enum U128FastFieldCodecType {
 }
 
 impl BinarySerializable for U128FastFieldCodecType {
-    fn serialize<W: Write + ?Sized>(&self, wrt: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
         self.to_code().serialize(wrt)
     }
 
@@ -85,6 +136,68 @@ 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

@@ -5,17 +5,18 @@ 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, u64_based, ColumnValues, MonotonicallyMappableToU64,
-    U128FastFieldCodecType,
+    monotonic_map_column, ColumnValues, FastFieldCodec, FastFieldCodecType,
+    MonotonicallyMappableToU64, U128FastFieldCodecType,
 };
 use crate::column_values::compact_space::CompactSpaceCompressor;
-use crate::column_values::u64_based::CodecType;
-use crate::iterable::Iterable;
 
 /// The normalized header gives some parameters after applying the following
 /// normalization of the vector:
@@ -30,6 +31,53 @@ 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,
@@ -37,7 +85,7 @@ pub(crate) struct U128Header {
 }
 
 impl BinarySerializable for U128Header {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.num_vals as u64).serialize(writer)?;
         self.codec_type.serialize(writer)?;
         Ok(())
@@ -63,9 +111,40 @@ fn normalize_column<C: ColumnValues>(
     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<I: Iterator<Item = u128>>(
-    iterable: &dyn Fn() -> I,
+pub fn serialize_column_values_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
+    iter_gen: F,
     num_vals: u32,
     output: &mut impl io::Write,
 ) -> io::Result<()> {
@@ -74,20 +153,105 @@ pub fn serialize_column_values_u128<I: Iterator<Item = u128>>(
         codec_type: U128FastFieldCodecType::CompactSpace,
     };
     header.serialize(output)?;
-    let compressor = CompactSpaceCompressor::train_from(iterable(), num_vals);
-    compressor.compress_into(iterable(), 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)?;
+        }
+    }
+    Ok(())
+}
+
 #[cfg(test)]
 pub mod tests {
-    use super::*;
-    use crate::column_values::u64_based::{
-        self, serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
-        ALL_U64_CODEC_TYPES,
-    };
+    use std::sync::Arc;
 
+    use common::OwnedBytes;
+
+    use super::*;
+    use crate::column_values::{open_u64_mapped, VecColumn};
+
+    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 {
@@ -103,22 +267,15 @@ pub mod tests {
     #[test]
     fn test_serialize_deserialize() {
         let original = [1u64, 5u64, 10u64];
-        let restored: Vec<u64> =
-            serialize_and_load_u64_based_column_values(&&original[..], &ALL_U64_CODEC_TYPES)
-                .iter()
-                .collect();
+        let restored: Vec<u64> = serialize_and_load(&original[..]).iter().collect();
         assert_eq!(&restored, &original[..]);
     }
 
     #[test]
     fn test_fastfield_bool_size_bitwidth_1() {
         let mut buffer = Vec::new();
-        serialize_u64_based_column_values(
-            || [false, true].into_iter(),
-            &ALL_U64_CODEC_TYPES,
-            &mut buffer,
-        )
-        .unwrap();
+        let col = VecColumn::from(&[false, true][..]);
+        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
         // TODO put the header as a footer so that it serves as a padding.
         // 5 bytes of header, 1 byte of value, 7 bytes of padding.
         assert_eq!(buffer.len(), 5 + 1);
@@ -127,27 +284,19 @@ pub mod tests {
     #[test]
     fn test_fastfield_bool_bit_size_bitwidth_0() {
         let mut buffer = Vec::new();
-        serialize_u64_based_column_values(
-            || [false, true].into_iter(),
-            &ALL_U64_CODEC_TYPES,
-            &mut buffer,
-        )
-        .unwrap();
-        // 6 bytes of header, 0 bytes of value, 7 bytes of padding.
-        assert_eq!(buffer.len(), 6);
+        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);
     }
 
     #[test]
     fn test_fastfield_gcd() {
         let mut buffer = Vec::new();
         let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
-        serialize_u64_based_column_values(
-            || vals.iter().cloned(),
-            &[CodecType::Bitpacked],
-            &mut buffer,
-        )
-        .unwrap();
+        let col = VecColumn::from(&vals[..]);
+        serialize_column_values(&col, &[FastFieldCodecType::Bitpacked], &mut buffer).unwrap();
         // Values are stored over 3 bits.
-        assert_eq!(buffer.len(), 6 + (3 * 80 / 8));
+        assert_eq!(buffer.len(), 7 + (3 * 80 / 8));
     }
 }

columnar/src/column_values/stats.rs

@@ -1,96 +0,0 @@
-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/tests.rs

@@ -2,90 +2,53 @@ use proptest::prelude::*;
 use proptest::strategy::Strategy;
 use proptest::{prop_oneof, proptest};
 
-#[test]
-fn test_serialize_and_load_simple() {
-    let mut buffer = Vec::new();
-    let vals = &[1u64, 2u64, 5u64];
-    serialize_u64_based_column_values(
-        || vals.iter().cloned(),
-        &[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],
+use super::bitpacked::BitpackedCodec;
+use super::blockwise_linear::BlockwiseLinearCodec;
+use super::linear::LinearCodec;
+use super::serialize::Header;
+
+pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
+    data: &[u64],
     name: &str,
 ) -> Option<(f32, f32)> {
-    let mut stats_collector = StatsCollector::default();
-    let mut codec_estimator: TColumnCodec::Estimator = Default::default();
+    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)?;
 
-    for val in vals.boxed_iter() {
-        stats_collector.collect(val);
-        codec_estimator.collect(val);
-    }
-    codec_estimator.finalize();
-    let stats = stats_collector.stats();
-    let estimation = codec_estimator.estimate(&stats)?;
+    let mut out = Vec::new();
+    let col = VecColumn::from(data);
+    serialize_column_values(&col, &[Codec::CODEC_TYPE], &mut out).unwrap();
 
-    let mut buffer = Vec::new();
-    codec_estimator
-        .serialize(&stats, vals.boxed_iter().as_mut(), &mut buffer)
-        .unwrap();
+    let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
 
-    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 reader = super::open_u64_mapped::<u64>(OwnedBytes::new(out)).unwrap();
+    assert_eq!(reader.num_vals(), data.len() as u32);
+    for (doc, orig_val) in data.iter().copied().enumerate() {
         let val = reader.get_val(doc as u32);
         assert_eq!(
             val, orig_val,
-            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{vals:?}`",
+            "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data `{data:?}`",
         );
     }
 
-    if !vals.is_empty() {
-        let test_rand_idx = rand::thread_rng().gen_range(0..=vals.len() - 1);
-        let expected_positions: Vec<u32> = vals
+    if !data.is_empty() {
+        let test_rand_idx = rand::thread_rng().gen_range(0..=data.len() - 1);
+        let expected_positions: Vec<u32> = data
             .iter()
             .enumerate()
-            .filter(|(_, el)| **el == vals[test_rand_idx])
+            .filter(|(_, el)| **el == data[test_rand_idx])
             .map(|(pos, _)| pos as u32)
             .collect();
         let mut positions = Vec::new();
         reader.get_docids_for_value_range(
-            vals[test_rand_idx]..=vals[test_rand_idx],
-            0..vals.len() as u32,
+            data[test_rand_idx]..=data[test_rand_idx],
+            0..data.len() as u32,
             &mut positions,
         );
         assert_eq!(expected_positions, positions);
     }
-    dbg!(estimation);
-    dbg!(actual_compression);
-    if actual_compression > 20 {
-        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)
+    Some((estimation, actual_compression))
 }
 
 proptest! {
@@ -155,8 +118,8 @@ pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
     data_and_names
 }
 
-fn test_codec<C: ColumnCodec>() {
-    let codec_name = std::any::type_name::<C>();
+fn test_codec<C: FastFieldCodec>() {
+    let codec_name = format!("{:?}", C::CODEC_TYPE);
     for (data, dataset_name) in get_codec_test_datasets() {
         let estimate_actual_opt: Option<(f32, f32)> =
             tests::create_and_validate::<C>(&data, dataset_name);
@@ -183,48 +146,53 @@ 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 = estimate::<LinearCodec>(&data).unwrap();
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
     assert_le!(linear_interpol_estimation, 0.01);
 
-    let multi_linear_interpol_estimation = estimate::<BlockwiseLinearCodec>(&data).unwrap();
+    let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
     assert_le!(multi_linear_interpol_estimation, 0.2);
     assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
 
-    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
     assert_lt!(linear_interpol_estimation, bitpacked_estimation);
 }
+#[test]
+fn estimation_test_bad_interpolation_case() {
+    let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
+
+    let data: VecColumn = data.into();
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    assert_le!(linear_interpol_estimation, 0.34);
+
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
+}
+
+#[test]
+fn estimation_prefer_bitpacked() {
+    let data = VecColumn::from(&[10, 10, 10, 10]);
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
+    assert_lt!(bitpacked_estimation, linear_interpol_estimation);
+}
 
 #[test]
 fn estimation_test_bad_interpolation_case_monotonically_increasing() {
     let mut data: Vec<u64> = (201..=20000_u64).collect();
     data.push(1_000_000);
+    let data: VecColumn = data.as_slice().into();
 
     // in this case the linear interpolation can't in fact not be worse than bitpacking,
     // but the estimator adds some threshold, which leads to estimated worse behavior
-    let linear_interpol_estimation = estimate::<LinearCodec>(&data[..]).unwrap();
+    let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
     assert_le!(linear_interpol_estimation, 0.35);
 
-    let bitpacked_estimation = estimate::<BitpackedCodec>(&data).unwrap();
+    let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
     assert_le!(bitpacked_estimation, 0.32);
     assert_le!(bitpacked_estimation, linear_interpol_estimation);
 }
@@ -233,7 +201,7 @@ 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) = CodecType::try_from_code(code) {
+        if let Some(codec_type) = FastFieldCodecType::from_code(code) {
             assert_eq!(codec_type.to_code(), code);
             count_codec += 1;
         }
@@ -241,16 +209,19 @@ fn test_fast_field_codec_type_to_code() {
     assert_eq!(count_codec, 3);
 }
 
-fn test_fastfield_gcd_i64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
+fn test_fastfield_gcd_i64_with_codec(
+    codec_type: FastFieldCodecType,
+    num_vals: usize,
+) -> io::Result<()> {
     let mut vals: Vec<i64> = (-4..=(num_vals as i64) - 5).map(|val| val * 1000).collect();
     let mut buffer: Vec<u8> = Vec::new();
-    crate::column_values::serialize_u64_based_column_values(
-        || vals.iter().cloned(),
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
         &[codec_type],
         &mut buffer,
     )?;
     let buffer = OwnedBytes::new(buffer);
-    let column = crate::column_values::load_u64_based_column_values::<i64>(buffer.clone())?;
+    let column = crate::column_values::open_u64_mapped::<i64>(buffer.clone())?;
     assert_eq!(column.get_val(0), -4000i64);
     assert_eq!(column.get_val(1), -3000i64);
     assert_eq!(column.get_val(2), -2000i64);
@@ -261,8 +232,8 @@ fn test_fastfield_gcd_i64_with_codec(codec_type: CodecType, num_vals: usize) ->
     let mut buffer_without_gcd = Vec::new();
     vals.pop();
     vals.push(1001i64);
-    crate::column_values::serialize_u64_based_column_values(
-        || vals.iter().cloned(),
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
         &[codec_type],
         &mut buffer_without_gcd,
     )?;
@@ -275,25 +246,28 @@ fn test_fastfield_gcd_i64_with_codec(codec_type: CodecType, num_vals: usize) ->
 #[test]
 fn test_fastfield_gcd_i64() -> io::Result<()> {
     for &codec_type in &[
-        CodecType::Bitpacked,
-        CodecType::BlockwiseLinear,
-        CodecType::Linear,
+        FastFieldCodecType::Bitpacked,
+        FastFieldCodecType::BlockwiseLinear,
+        FastFieldCodecType::Linear,
     ] {
         test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
     }
     Ok(())
 }
 
-fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, num_vals: usize) -> io::Result<()> {
+fn test_fastfield_gcd_u64_with_codec(
+    codec_type: FastFieldCodecType,
+    num_vals: usize,
+) -> io::Result<()> {
     let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
     let mut buffer: Vec<u8> = Vec::new();
-    crate::column_values::serialize_u64_based_column_values(
-        || vals.iter().cloned(),
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
         &[codec_type],
         &mut buffer,
     )?;
     let buffer = OwnedBytes::new(buffer);
-    let column = crate::column_values::load_u64_based_column_values::<u64>(buffer.clone())?;
+    let column = crate::column_values::open_u64_mapped::<u64>(buffer.clone())?;
     assert_eq!(column.get_val(0), 1000u64);
     assert_eq!(column.get_val(1), 2000u64);
     assert_eq!(column.get_val(2), 3000u64);
@@ -304,8 +278,8 @@ fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, num_vals: usize) ->
     let mut buffer_without_gcd = Vec::new();
     vals.pop();
     vals.push(1001u64);
-    crate::column_values::serialize_u64_based_column_values(
-        || vals.iter().cloned(),
+    crate::column_values::serialize_column_values(
+        &VecColumn::from(&vals),
         &[codec_type],
         &mut buffer_without_gcd,
     )?;
@@ -317,9 +291,9 @@ fn test_fastfield_gcd_u64_with_codec(codec_type: CodecType, num_vals: usize) ->
 #[test]
 fn test_fastfield_gcd_u64() -> io::Result<()> {
     for &codec_type in &[
-        CodecType::Bitpacked,
-        CodecType::BlockwiseLinear,
-        CodecType::Linear,
+        FastFieldCodecType::Bitpacked,
+        FastFieldCodecType::BlockwiseLinear,
+        FastFieldCodecType::Linear,
     ] {
         test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
     }
@@ -328,10 +302,7 @@ fn test_fastfield_gcd_u64() -> io::Result<()> {
 
 #[test]
 pub fn test_fastfield2() {
-    let test_fastfield = crate::column_values::serialize_and_load_u64_based_column_values::<u64>(
-        &&[100u64, 200u64, 300u64][..],
-        &ALL_U64_CODEC_TYPES,
-    );
+    let test_fastfield = crate::column_values::serialize_and_load(&[100u64, 200u64, 300u64]);
     assert_eq!(test_fastfield.get_val(0), 100);
     assert_eq!(test_fastfield.get_val(1), 200);
     assert_eq!(test_fastfield.get_val(2), 300);

columnar/src/column_values/u64_based/bitpacked.rs

@@ -1,127 +0,0 @@
-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

@@ -1,281 +0,0 @@
-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 mut 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/linear.rs

@@ -1,277 +0,0 @@
-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::{MonotonicallyMappableToU64, 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

@@ -1,186 +0,0 @@
-mod bitpacked;
-mod blockwise_linear;
-mod line;
-mod linear;
-mod stats_collector;
-
-use std::io;
-use std::io::Write;
-use std::sync::Arc;
-
-use common::{BinarySerializable, OwnedBytes};
-
-use crate::column_values::monotonic_mapping::{
-    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
-};
-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<T: MonotonicallyMappableToU64, F, I>(
-    vals: F,
-    codec_types: &[CodecType],
-    wrt: &mut dyn Write,
-) -> io::Result<()>
-where
-    I: Iterator<Item = T>,
-    F: Fn() -> I,
-{
-    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() {
-        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().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.boxed_iter(), 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

@@ -1,200 +0,0 @@
-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/columnar/column_type.rs

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

columnar/src/columnar/merge.rs

@@ -0,0 +1,208 @@
+use std::collections::HashMap;
+use std::io;
+
+use crate::columnar::ColumnarReader;
+use crate::dynamic_column::DynamicColumn;
+use crate::ColumnType;
+
+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!();
+        }
+    }
+}
+
+/// 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)]
+#[repr(u8)]
+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,
+        }
+    }
+}
+
+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().into())
+                .or_default();
+            columns.push(handle.open()?);
+        }
+    }
+
+    normalize_columns(&mut field_name_to_group);
+
+    Ok(field_name_to_group)
+}
+
+/// Coerce numerical type columns to the same type
+/// TODO rename to `coerce_columns`
+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

@@ -1,114 +0,0 @@
-use std::io::{self, Write};
-
-use common::CountingWriter;
-use itertools::Itertools;
-use sstable::{SSTable, TermOrdinal};
-
-use super::term_merger::TermMerger;
-use crate::column_index::{serialize_column_index, SerializableColumnIndex};
-use crate::column_values::{serialize_u64_based_column_values, CodecType};
-use crate::BytesColumn;
-
-// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
-// Column: [Column Index, Column Values, column index num bytes U32::LE]
-pub fn merge_bytes_or_str_column(
-    column_index: SerializableColumnIndex<'_>,
-    bytes_columns: &[BytesColumn],
-    output: &mut impl Write,
-) -> io::Result<()> {
-    // Serialize dict and generate mapping for values
-    let mut output = CountingWriter::wrap(output);
-    let term_ord_mapping = serialize_merged_dict(bytes_columns, &mut output)?;
-    let dictionary_num_bytes: u32 = output.written_bytes() as u32;
-    let output = output.finish();
-
-    serialize_bytes_or_str_column(column_index, bytes_columns, &term_ord_mapping, output)?;
-
-    output.write_all(&dictionary_num_bytes.to_le_bytes())?;
-    Ok(())
-}
-
-fn serialize_bytes_or_str_column(
-    column_index: SerializableColumnIndex<'_>,
-    bytes_columns: &[BytesColumn],
-    term_ord_mapping: &TermOrdinalMapping,
-    output: &mut impl Write,
-) -> io::Result<()> {
-    let column_index_num_bytes = serialize_column_index(column_index, output)?;
-
-    let column_values = move || {
-        let iter = bytes_columns
-            .iter()
-            .enumerate()
-            .flat_map(|(segment_ord, byte_column)| {
-                let segment_ord = term_ord_mapping.get_segment(segment_ord);
-                byte_column
-                    .ords()
-                    .values
-                    .iter()
-                    .map(move |term_ord| segment_ord[term_ord as usize])
-            });
-        iter
-    };
-
-    serialize_u64_based_column_values(
-        column_values,
-        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
-        output,
-    )?;
-
-    output.write_all(&column_index_num_bytes.to_le_bytes())?;
-
-    Ok(())
-}
-
-fn serialize_merged_dict(
-    bytes_columns: &[BytesColumn],
-    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 {
-        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);
-
-    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()?;
-
-    Ok(term_ord_mapping)
-}
-
-#[derive(Default)]
-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: usize) -> &[TermOrdinal] {
-        &(self.per_segment_new_term_ordinals[segment_ord])[..]
-    }
-}

columnar/src/columnar/merge/merge_mapping.rs

@@ -1,60 +0,0 @@
-use std::ops::Range;
-
-use crate::{column, ColumnarReader, 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 from_columnars(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.
-    /// ..
-    Stack(StackMergeOrder),
-    /// Some more complex mapping, that can interleaves rows from the different readers and
-    /// possibly drop rows.
-    Complex(()),
-}
-
-impl MergeRowOrder {
-    pub fn num_rows(&self) -> RowId {
-        match self {
-            MergeRowOrder::Stack(stack_row_order) => stack_row_order.num_rows(),
-            MergeRowOrder::Complex(_) => {
-                todo!()
-            }
-        }
-    }
-}

columnar/src/columnar/merge/mod.rs

@@ -1,231 +0,0 @@
-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, StackMergeOrder};
-
-use super::writer::ColumnarSerializer;
-use crate::column::{serialize_column_mappable_to_u128, serialize_column_mappable_to_u64};
-use crate::columnar::column_type::ColumnTypeCategory;
-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,
-};
-
-pub fn merge_columnar(
-    columnar_readers: &[&ColumnarReader],
-    mapping: 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, &mapping, &mut column_serializer)?;
-    }
-    serializer.finalize(mapping.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,
-    }
-}
-
-pub 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<Arc<dyn ColumnValues>> = Vec::with_capacity(columns.len());
-            for dynamic_column_opt in columns {
-                if let Some(Column { idx, values }) =
-                    dynamic_column_opt.and_then(dynamic_column_to_u64_monotonic)
-                {
-                    column_indexes.push(Some(idx));
-                    column_values.push(values);
-                } else {
-                    column_indexes.push(None);
-                }
-            }
-            let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            let stacked_columns_iterable = || column_values
-                .iter()
-                .flat_map(|column| column.iter());
-            serialize_column_mappable_to_u64(merged_column_index, &stacked_columns_iterable, wrt)?;
-        }
-        ColumnType::IpAddr => {
-            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
-            let mut column_values: Vec<Arc<dyn ColumnValues<Ipv6Addr>>> =
-                Vec::with_capacity(columns.len());
-            let mut num_values = 0;
-            for dynamic_column_opt in columns {
-                if let Some(DynamicColumn::IpAddr(Column { idx, values })) = dynamic_column_opt {
-                    num_values += values.num_vals();
-                    column_indexes.push(Some(idx));
-                    column_values.push(values);
-                } else {
-                    column_indexes.push(None);
-                }
-            }
-            let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            serialize_column_mappable_to_u128(
-                merged_column_index,
-                &|| {
-                    column_values
-                        .iter()
-                        .flat_map(|column_value| column_value.iter())
-                },
-                num_values,
-                wrt,
-            )?;
-        }
-        ColumnType::Bytes | ColumnType::Str => {
-            let mut column_indexes: Vec<Option<ColumnIndex>> = Vec::with_capacity(columns.len());
-            let mut bytes_columns: Vec<BytesColumn> = Vec::with_capacity(columns.len());
-            for dynamic_column_opt in columns {
-                match dynamic_column_opt {
-                    Some(DynamicColumn::Str(str_column)) => {
-                        column_indexes.push(Some(str_column.term_ord_column.idx.clone()));
-                        bytes_columns.push(str_column.into());
-                    }
-                    Some(DynamicColumn::Bytes(bytes_column)) => {
-                        column_indexes.push(Some(bytes_column.term_ord_column.idx.clone()));
-                        bytes_columns.push(bytes_column);
-                    }
-                    _ => column_indexes.push(None),
-                }
-            }
-            let merged_column_index =
-                crate::column_index::stack_column_index(&column_indexes[..], merge_row_order);
-            merge_bytes_or_str_column(merged_column_index, &bytes_columns, wrt)?;
-        }
-    }
-    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

@@ -1,107 +0,0 @@
-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

@@ -1,252 +0,0 @@
-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, &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, &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, &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, &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::from_columnars(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::from_columnars(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::from_columnars(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

@@ -1 +0,0 @@
-

columnar/src/columnar/mod.rs

@@ -1,11 +1,10 @@
 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, MergeRowOrder, StackMergeOrder};
+pub use merge::{merge_columnar, MergeDocOrder};
 pub use reader::ColumnarReader;
 pub use writer::ColumnarWriter;

columnar/src/columnar/reader/mod.rs

@@ -6,7 +6,6 @@ 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)
@@ -14,11 +13,9 @@ 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 {
@@ -30,27 +27,23 @@ 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>() + 4 + format_version::VERSION_FOOTER_NUM_BYTES);
+            .split_from_end(mem::size_of::<u64>() + format_version::VERSION_FOOTER_NUM_BYTES);
         let footer_bytes = footer_slice.read_bytes()?;
-        let sstable_len = u64::deserialize(&mut &footer_bytes[0..8])?;
-        let num_rows = u32::deserialize(&mut &footer_bytes[8..12])?;
+        let (mut sstable_len_bytes, version_footer_bytes) =
+            footer_bytes.rsplit(format_version::VERSION_FOOTER_NUM_BYTES);
         let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
-            footer_bytes[12..].try_into().unwrap();
+            version_footer_bytes.as_slice().try_into().unwrap();
         let _version = format_version::parse_footer(version_footer_bytes)?;
+        let sstable_len = u64::deserialize(&mut sstable_len_bytes)?;
         let (column_data, sstable) =
             file_slice_without_sstable_len.split_from_end(sstable_len as usize);
         let column_dictionary = Dictionary::open(sstable)?;
         Ok(ColumnarReader {
             column_dictionary,
             column_data,
-            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()?;

columnar/src/columnar/writer/column_writers.rs

@@ -114,7 +114,7 @@ impl NumericalColumnWriter {
 /// State used to store what types are still acceptable
 /// after having seen a set of numerical values.
 #[derive(Clone, Copy)]
-pub(crate) enum CompatibleNumericalTypes {
+enum CompatibleNumericalTypes {
     Dynamic {
         all_values_within_i64_range: bool,
         all_values_within_u64_range: bool,
@@ -132,7 +132,7 @@ impl Default for CompatibleNumericalTypes {
 }
 
 impl CompatibleNumericalTypes {
-    pub fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
+    fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -148,7 +148,7 @@ impl CompatibleNumericalTypes {
         }
     }
 
-    pub fn accept_value(&mut self, numerical_value: NumericalValue) {
+    fn accept_value(&mut self, numerical_value: NumericalValue) {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -189,10 +189,12 @@ 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(

columnar/src/columnar/writer/mod.rs

@@ -7,20 +7,19 @@ use std::io;
 use std::net::Ipv6Addr;
 
 use column_operation::ColumnOperation;
-pub(crate) use column_writers::CompatibleNumericalTypes;
 use common::CountingWriter;
-pub(crate) use serializer::ColumnarSerializer;
+use 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,
 };
-use crate::columnar::writer::value_index::{IndexBuilder, PreallocatedIndexBuilders, OptionalIndexBuilder};
+use crate::columnar::writer::value_index::{IndexBuilder, PreallocatedIndexBuilders};
 use crate::dictionary::{DictionaryBuilder, TermIdMapping, UnorderedId};
 use crate::value::{Coerce, NumericalType, NumericalValue};
 use crate::{Cardinality, RowId};
@@ -280,35 +279,40 @@ impl ColumnarWriter {
     }
     pub fn serialize(&mut self, num_docs: RowId, wrt: &mut dyn io::Write) -> io::Result<()> {
         let mut serializer = ColumnarSerializer::new(wrt);
-        let mut columns: Vec<(&[u8], ColumnTypeCategory, Addr)> = self
+        let mut columns: Vec<(&[u8], ColumnType, Addr)> = self
             .numerical_field_hash_map
             .iter()
-            .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Numerical, addr))
+            .map(|(column_name, addr, _)| {
+                let numerical_column_writer: NumericalColumnWriter =
+                    self.numerical_field_hash_map.read(addr);
+                let column_type = numerical_column_writer.numerical_type().into();
+                (column_name, column_type, addr)
+            })
             .collect();
         columns.extend(
             self.bytes_field_hash_map
                 .iter()
-                .map(|(term, addr, _)| (term, ColumnTypeCategory::Bytes, addr)),
+                .map(|(term, addr, _)| (term, ColumnType::Bytes, addr)),
         );
         columns.extend(
             self.str_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Str, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Str, addr)),
         );
         columns.extend(
             self.bool_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::Bool, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::Bool, addr)),
         );
         columns.extend(
             self.ip_addr_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::IpAddr, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::IpAddr, addr)),
         );
         columns.extend(
             self.datetime_field_hash_map
                 .iter()
-                .map(|(column_name, addr, _)| (column_name, ColumnTypeCategory::DateTime, addr)),
+                .map(|(column_name, addr, _)| (column_name, ColumnType::DateTime, addr)),
         );
         columns.sort_unstable_by_key(|(column_name, col_type, _)| (*column_name, *col_type));
 
@@ -316,8 +320,12 @@ impl ColumnarWriter {
         let mut symbol_byte_buffer: Vec<u8> = Vec::new();
         for (column_name, column_type, addr) in columns {
             match column_type {
-                ColumnTypeCategory::Bool => {
-                    let column_writer: ColumnWriter = self.bool_field_hash_map.read(addr);
+                ColumnType::Bool | ColumnType::DateTime => {
+                    let column_writer: ColumnWriter = if column_type == ColumnType::Bool {
+                        self.bool_field_hash_map.read(addr)
+                    } else {
+                        self.datetime_field_hash_map.read(addr)
+                    };
                     let cardinality = column_writer.get_cardinality(num_docs);
                     let mut column_serializer =
                         serializer.serialize_column(column_name, ColumnType::Bool);
@@ -329,7 +337,7 @@ impl ColumnarWriter {
                         &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 =
@@ -342,33 +350,36 @@ impl ColumnarWriter {
                         &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, &mut symbol_byte_buffer),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::Numerical => {
+                ColumnType::I64 | ColumnType::F64 | 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 numerical_type = column_type.numerical_type().unwrap();
+                    let cardinality = numerical_column_writer.cardinality(num_docs);
                     let mut column_serializer =
                         serializer.serialize_column(column_name, ColumnType::from(numerical_type));
                     serialize_numerical_column(
@@ -380,29 +391,13 @@ impl ColumnarWriter {
                         &mut column_serializer,
                     )?;
                 }
-                ColumnTypeCategory::DateTime => {
-                    let column_writer: ColumnWriter = self.datetime_field_hash_map.read(addr);
-                    let cardinality = column_writer.get_cardinality(num_docs);
-                    let mut column_serializer =
-                        serializer.serialize_column(column_name, ColumnType::DateTime);
-                    serialize_numerical_column(
-                        cardinality,
-                        num_docs,
-                        NumericalType::I64,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
-                        buffers,
-                        &mut column_serializer,
-                    )?;
-                }
             };
         }
-        serializer.finalize(num_docs)?;
+        serializer.finalize()?;
         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,
@@ -550,7 +545,7 @@ fn send_to_serialize_column_mappable_to_u128<
 >(
     op_iterator: impl Iterator<Item = ColumnOperation<T>>,
     cardinality: Cardinality,
-    num_rows: RowId,
+    num_docs: RowId,
     value_index_builders: &mut PreallocatedIndexBuilders,
     values: &mut Vec<T>,
     mut wrt: impl io::Write,
@@ -572,48 +567,48 @@ where
         Cardinality::Optional => {
             let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let non_null_rows: &[u32] = optional_index_builder.finish(num_rows);
-            SerializableColumnIndex::Optional {
-                num_rows,
-                non_null_row_ids: Box::new(|| Box::new(non_null_rows.iter().copied())),
-            }
+            let optional_index = optional_index_builder.finish(num_docs);
+            SerializableColumnIndex::Optional(Box::new(optional_index))
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_rows);
-            SerializableColumnIndex::Multivalued(Box::new(|| Box::new(multivalued_index.iter().copied())))
+            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u128(
         serializable_column_index,
-        &|| values.iter().copied(),
+        || values.iter().cloned(),
         values.len() as u32,
         &mut wrt,
     )?;
     Ok(())
 }
 
-fn sort_values_within_row_in_place(multivalued_index: &[RowId], values: &mut Vec<u64>) {
+fn sort_values_within_row_in_place(
+    multivalued_index: &impl ColumnValues<RowId>,
+    values: &mut Vec<u64>,
+) {
     let mut start_index: usize = 0;
-    for end_index in multivalued_index.iter().copied() {
+    for end_index in multivalued_index.iter() {
         let end_index = end_index as usize;
         values[start_index..end_index].sort_unstable();
         start_index = end_index;
     }
 }
 
-fn send_to_serialize_column_mappable_to_u64<'a>(
+fn send_to_serialize_column_mappable_to_u64(
     op_iterator: impl Iterator<Item = ColumnOperation<u64>>,
     cardinality: Cardinality,
-    num_rows: RowId,
+    num_docs: RowId,
     sort_values_within_row: bool,
-    value_index_builders: &'a mut PreallocatedIndexBuilders,
+    value_index_builders: &mut PreallocatedIndexBuilders,
     values: &mut Vec<u64>,
     mut wrt: impl io::Write,
 ) -> io::Result<()>
 where
-    for<'b> VecColumn<'b, u64>: ColumnValues<u64>,
+    for<'a> VecColumn<'a, u64>: ColumnValues<u64>,
 {
     values.clear();
     let serializable_column_index = match cardinality {
@@ -626,27 +621,24 @@ where
             SerializableColumnIndex::Full
         }
         Cardinality::Optional => {
-            let optional_index_builder: &'a mut OptionalIndexBuilder = value_index_builders.borrow_optional_index_builder();
+            let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let optional_index: &'a [u32] = optional_index_builder.finish(num_rows);
-            SerializableColumnIndex::Optional {
-                non_null_row_ids: Box::new(move || Box::new(optional_index.iter().copied())),
-                num_rows,
-            }
+            let optional_index = optional_index_builder.finish(num_docs);
+            SerializableColumnIndex::Optional(Box::new(optional_index))
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_rows);
+            let multivalued_index = multivalued_index_builder.finish(num_docs);
             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(|| Box::new(multivalued_index.iter().copied())))
+            SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u64(
         serializable_column_index,
-        &|| values.iter().copied(),
+        &VecColumn::from(&values[..]),
         &mut wrt,
     )?;
     Ok(())

columnar/src/columnar/writer/serializer.rs

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

@@ -1,3 +1,5 @@
+use crate::column_index::SerializableOptionalIndex;
+use crate::column_values::{ColumnValues, VecColumn};
 use crate::RowId;
 
 /// The `IndexBuilder` interprets a sequence of
@@ -27,15 +29,34 @@ 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) -> &'a [RowId] {
+    pub fn finish<'a>(&'a mut self, num_rows: RowId) -> impl SerializableOptionalIndex + 'a {
         debug_assert!(self
             .docs
             .last()
             .copied()
             .map(|last_doc| last_doc < num_rows)
             .unwrap_or(true));
-        &self.docs[..]
+        SingleValueArrayIndex {
+            row_ids: &self.docs[..],
+            num_rows,
+        }
     }
 
     fn reset(&mut self) {
@@ -63,10 +84,14 @@ pub struct MultivaluedIndexBuilder {
 }
 
 impl MultivaluedIndexBuilder {
-    pub fn finish(&mut self, num_docs: RowId) -> &[u32] {
+    pub fn finish(&mut self, num_docs: RowId) -> impl ColumnValues<u32> + '_ {
         self.start_offsets
             .resize(num_docs as usize + 1, self.total_num_vals_seen);
-        &self.start_offsets[..]
+        VecColumn {
+            values: &&self.start_offsets[..],
+            min_value: 0,
+            max_value: self.start_offsets.last().copied().unwrap_or(0),
+        }
     }
 
     fn reset(&mut self) {
@@ -122,14 +147,20 @@ mod tests {
         opt_value_index_builder.record_row(0u32);
         opt_value_index_builder.record_value();
         assert_eq!(
-            &opt_value_index_builder.finish(1u32),
+            &opt_value_index_builder
+                .finish(1u32)
+                .non_null_rows()
+                .collect::<Vec<u32>>(),
             &[0]
         );
         opt_value_index_builder.reset();
         opt_value_index_builder.record_row(1u32);
         opt_value_index_builder.record_value();
         assert_eq!(
-            &opt_value_index_builder.finish(2u32),
+            &opt_value_index_builder
+                .finish(2u32)
+                .non_null_rows()
+                .collect::<Vec<u32>>(),
             &[1]
         );
     }
@@ -146,7 +177,6 @@ mod tests {
             multivalued_value_index_builder
                 .finish(4u32)
                 .iter()
-                .copied()
                 .collect::<Vec<u32>>(),
             vec![0, 0, 2, 3, 3]
         );
@@ -158,7 +188,6 @@ 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::{Cardinality, DateTime, NumericalType};
+use crate::{DateTime, NumericalType};
 
 #[derive(Clone)]
 pub enum DynamicColumn {
@@ -23,18 +23,6 @@ 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,
@@ -48,14 +36,6 @@ 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()
     }
@@ -70,7 +50,7 @@ impl DynamicColumn {
         self.column_type().numerical_type() == Some(NumericalType::U64)
     }
 
-    fn coerce_to_f64(self) -> Option<DynamicColumn> {
+    pub fn coerce_to_f64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
                 idx: column.idx,
@@ -84,7 +64,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    fn coerce_to_i64(self) -> Option<DynamicColumn> {
+    pub fn coerce_to_i64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::U64(column) => {
                 if column.max_value() > i64::MAX as u64 {
@@ -99,7 +79,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    fn coerce_to_u64(self) -> Option<DynamicColumn> {
+    pub fn coerce_to_u64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => {
                 if column.min_value() < 0 {

columnar/src/iterable.rs

@@ -1,61 +0,0 @@
-use std::marker::PhantomData;
-use std::ops::Range;
-
-pub trait Iterable<T = u64> {
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_>;
-}
-
-struct Mapped<U, Original, Transform> {
-    original_iterable: Original,
-    transform: Transform,
-    input_type: PhantomData<U>,
-}
-
-impl<U, V, Original, Transform> Iterable<V> for Mapped<U, Original, Transform>
-where
-    Original: Iterable<U>,
-    Transform: Fn(U) -> V,
-{
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = V> + '_> {
-        Box::new(self.original_iterable.boxed_iter().map(&self.transform))
-    }
-}
-
-impl<U> Iterable<U> for &dyn Iterable<U> {
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = U> + '_> {
-        (*self).boxed_iter()
-    }
-}
-
-impl<F, T> Iterable<T> for F
-where F: Fn() -> Box<dyn Iterator<Item = T>>
-{
-    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
-        self()
-    }
-}
-
-pub fn map_iterable<U, V, F, I>(
-    original_iterable: impl Fn() -> I,
-    transform: F,
-) -> impl Fn() -> std::iter::Map<I, F>
-where
-    F: Fn(U) -> V + Clone,
-    I: Iterator<Item = U>,
-{
-    move || original_iterable().map(transform.clone())
-}
-
-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

@@ -11,11 +11,10 @@ use std::io;
 
 mod column;
 mod column_index;
-pub mod column_values;
+mod column_values;
 mod columnar;
 mod dictionary;
 mod dynamic_column;
-mod iterable;
 pub(crate) mod utils;
 mod value;
 
@@ -24,9 +23,8 @@ pub use column_index::ColumnIndex;
 pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
 pub use columnar::{
     merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
-    MergeRowOrder, StackMergeOrder,
+    MergeDocOrder,
 };
-pub(crate) use iterable::{map_iterable, Iterable};
 use sstable::VoidSSTable;
 pub use value::{NumericalType, NumericalValue};
 
@@ -68,12 +66,6 @@ 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
     }

common/src/serialize.rs

@@ -5,37 +5,12 @@ 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 + ?Sized>(&self, writer: &mut W) -> io::Result<()>;
+    fn serialize<W: Write>(&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> {
@@ -59,7 +34,7 @@ pub trait FixedSize: BinarySerializable {
 }
 
 impl BinarySerializable for () {
-    fn serialize<W: Write + ?Sized>(&self, _: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, _: &mut W) -> io::Result<()> {
         Ok(())
     }
     fn deserialize<R: Read>(_: &mut R) -> io::Result<Self> {
@@ -72,7 +47,7 @@ impl FixedSize for () {
 }
 
 impl<T: BinarySerializable> BinarySerializable for Vec<T> {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.len() as u64).serialize(writer)?;
         for it in self {
             it.serialize(writer)?;
@@ -91,7 +66,7 @@ impl<T: BinarySerializable> BinarySerializable for Vec<T> {
 }
 
 impl<Left: BinarySerializable, Right: BinarySerializable> BinarySerializable for (Left, Right) {
-    fn serialize<W: Write + ?Sized>(&self, write: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
         self.0.serialize(write)?;
         self.1.serialize(write)
     }
@@ -106,7 +81,7 @@ impl<Left: BinarySerializable + FixedSize, Right: BinarySerializable + FixedSize
 }
 
 impl BinarySerializable for u32 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u32::<Endianness>(*self)
     }
 
@@ -120,7 +95,7 @@ impl FixedSize for u32 {
 }
 
 impl BinarySerializable for u16 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u16::<Endianness>(*self)
     }
 
@@ -134,7 +109,7 @@ impl FixedSize for u16 {
 }
 
 impl BinarySerializable for u64 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -147,7 +122,7 @@ impl FixedSize for u64 {
 }
 
 impl BinarySerializable for u128 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u128::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -160,7 +135,7 @@ impl FixedSize for u128 {
 }
 
 impl BinarySerializable for f32 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_f32::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -173,7 +148,7 @@ impl FixedSize for f32 {
 }
 
 impl BinarySerializable for i64 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_i64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -186,7 +161,7 @@ impl FixedSize for i64 {
 }
 
 impl BinarySerializable for f64 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_f64::<Endianness>(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
@@ -199,7 +174,7 @@ impl FixedSize for f64 {
 }
 
 impl BinarySerializable for u8 {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u8(*self)
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<u8> {
@@ -212,7 +187,7 @@ impl FixedSize for u8 {
 }
 
 impl BinarySerializable for bool {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_u8(u8::from(*self))
     }
     fn deserialize<R: Read>(reader: &mut R) -> io::Result<bool> {
@@ -233,7 +208,7 @@ impl FixedSize for bool {
 }
 
 impl BinarySerializable for String {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&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 + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&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 + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&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

@@ -0,0 +1,34 @@
+[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

@@ -0,0 +1,68 @@
+
+
+# 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

@@ -0,0 +1,311 @@
+#![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

@@ -0,0 +1,10 @@
+#![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

@@ -4,7 +4,7 @@ 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};
@@ -14,8 +14,7 @@ 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::schema::{FieldType, Type};
 use crate::{InvertedIndexReader, SegmentReader, TantivyError};
 
 #[derive(Clone, Default)]
@@ -37,38 +36,12 @@ 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 +56,19 @@ 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 +82,7 @@ impl BucketAggregationWithAccessor {
                 max_bucket_count,
             )?,
             bucket_agg: bucket.clone(),
-            inverted_index,
+            str_dict_column,
             bucket_count: BucketCount {
                 bucket_count,
                 max_bucket_count,
@@ -124,7 +96,7 @@ 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 +111,10 @@ 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 +159,19 @@ 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 fast field found for field: {}", field_name))
+    })?;
+    Ok((ff_field, field_type))
 }

src/aggregation/bucket/histogram/date_histogram.rs

@@ -0,0 +1,123 @@
+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,
         })
     }
@@ -368,7 +328,14 @@ impl SegmentTermCollector {
 
         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 +351,40 @@ 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, mut 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) {
@@ -446,65 +419,26 @@ 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)?;
@@ -1207,36 +1141,37 @@ 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 +1191,13 @@ 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 +1288,3 @@ 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))
                 }
             },

src/aggregation/metric/stats.rs

@@ -1,10 +1,17 @@
-use fastfield_codecs::Column;
+use columnar::Column;
 use serde::{Deserialize, Serialize};
 
-use crate::aggregation::f64_from_fastfield_u64;
+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};
 
+use super::*;
+
 /// A multi-value metric aggregation that computes a collection of statistics on numeric values that
 /// are extracted from the aggregated documents.
 /// See [`Stats`] for returned statistics.
@@ -160,27 +167,74 @@ 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);
@@ -1168,6 +1164,9 @@ mod tests {
     #[cfg(all(test, feature = "unstable"))]
     mod bench {
 
+        use crate::aggregation::bucket::CustomOrder;
+        use crate::aggregation::bucket::Order;
+        use crate::aggregation::bucket::OrderTarget;
         use rand::prelude::SliceRandom;
         use rand::{thread_rng, Rng};
         use test::{self, Bencher};
@@ -1177,7 +1176,7 @@ mod tests {
         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 +1188,19 @@ 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 +1215,6 @@ 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 +1366,42 @@ 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 +1426,36 @@ 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 +25,90 @@ 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 +119,7 @@ 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 +130,9 @@ 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 +150,7 @@ 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 +163,7 @@ impl SegmentAggregationResultsCollector {
         Ok(())
     }
 
-    pub(crate) fn flush_staged_docs(
+    fn flush_staged_docs(
         &mut self,
         agg_with_accessor: &AggregationsWithAccessor,
         force_flush: bool,
@@ -162,6 +197,66 @@ 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 +300,7 @@ 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 +310,7 @@ 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 +338,7 @@ 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 +353,7 @@ 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

@@ -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/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

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

src/fastfield/mod.rs

@@ -189,7 +189,7 @@ mod tests {
         }
         let file = directory.open_read(path).unwrap();
 
-        assert_eq!(file.len(), 161);
+        assert_eq!(file.len(), 157);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let column = fast_field_readers.u64("field").unwrap();
         assert_eq!(column.get_val(0), 13u64);
@@ -236,7 +236,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 189);
+        assert_eq!(file.len(), 185);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let col = fast_field_readers.u64("field").unwrap();
         assert_eq!(col.get_val(0), 4u64);
@@ -266,7 +266,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 162);
+        assert_eq!(file.len(), 158);
         let fast_field_readers = FastFieldReaders::open(file).unwrap();
         let fast_field_reader = fast_field_readers.u64("field").unwrap();
         for doc in 0..10_000 {
@@ -295,7 +295,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 4557);
+        assert_eq!(file.len(), 80166);
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
             let col = fast_field_readers.u64("field").unwrap();
@@ -325,7 +325,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 333_usize);
+        assert_eq!(file.len(), 329_usize);
 
         {
             let fast_field_readers = FastFieldReaders::open(file).unwrap();
@@ -793,7 +793,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 175);
+        assert_eq!(file.len(), 171);
         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);
@@ -825,7 +825,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 187);
+        assert_eq!(file.len(), 183);
         let readers = FastFieldReaders::open(file).unwrap();
         let bool_col = readers.bool("field_bool").unwrap();
         for i in 0..25 {
@@ -850,7 +850,7 @@ mod tests {
             write.terminate().unwrap();
         }
         let file = directory.open_read(path).unwrap();
-        assert_eq!(file.len(), 177);
+        assert_eq!(file.len(), 173);
         let fastfield_readers = FastFieldReaders::open(file).unwrap();
         let col = fastfield_readers.bool("field_bool").unwrap();
         assert_eq!(col.get_val(0), false);

src/fastfield/readers.rs

@@ -26,10 +26,6 @@ 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() {

src/indexer/merger.rs

@@ -1,7 +1,7 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use columnar::{ColumnValues, ColumnarReader, MergeRowOrder, StackMergeOrder};
+use columnar::ColumnValues;
 use itertools::Itertools;
 use measure_time::debug_time;
 
@@ -248,17 +248,13 @@ impl IndexMerger {
         mut term_ord_mappings: HashMap<Field, TermOrdinalMapping>,
         doc_id_mapping: &SegmentDocIdMapping,
     ) -> crate::Result<()> {
-        debug_time!("write-fast-fields");
-        let columnars: Vec<&ColumnarReader> = self
-            .readers
-            .iter()
-            .map(|reader| reader.fast_fields().columnar())
-            .collect();
-        if !doc_id_mapping.is_trivial() {
-            todo!()
+        debug_time!("wrie-fast-fields");
+        for (_field, field_entry) in self.schema.fields() {
+            if field_entry.is_fast() {
+                todo!();
+            }
         }
-        let merge_row_order = MergeRowOrder::Stack(StackMergeOrder::from_columnars(&columnars[..]));
-        columnar::merge_columnar(&columnars[..], merge_row_order, fast_field_wrt)?;
+
         // for (field, field_entry) in self.schema.fields() {
         // let field_type = field_entry.field_type();
         // match field_type {

src/indexer/merger_sorted_index_test.rs

@@ -476,12 +476,11 @@ 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 +511,42 @@ 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

@@ -19,6 +19,8 @@ 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;

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;

src/postings/term_info.rs

@@ -39,7 +39,7 @@ impl FixedSize for TermInfo {
 }
 
 impl BinarySerializable for TermInfo {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         self.doc_freq.serialize(writer)?;
         (self.postings_range.start as u64).serialize(writer)?;
         self.posting_num_bytes().serialize(writer)?;

src/query/range_query/range_query_ip_fastfield.rs

@@ -88,7 +88,7 @@ fn bound_to_value_range(
 }
 
 #[cfg(test)]
-mod tests {
+pub mod tests {
     use proptest::prelude::ProptestConfig;
     use proptest::strategy::Strategy;
     use proptest::{prop_oneof, proptest};
@@ -188,7 +188,7 @@ mod tests {
         assert_eq!(count, 2);
     }
 
-    fn create_index_from_docs(docs: &[Doc]) -> Index {
+    pub fn create_index_from_docs(docs: &[Doc]) -> Index {
         let mut schema_builder = Schema::builder();
         let ip_field = schema_builder.add_ip_addr_field("ip", STORED | FAST);
         let ips_field = schema_builder.add_ip_addr_field("ips", FAST | INDEXED);

src/query/range_query/range_query_u64_fastfield.rs

@@ -86,7 +86,7 @@ fn bound_to_value_range<T: MonotonicallyMappableToU64>(
 }
 
 #[cfg(test)]
-mod tests {
+pub mod tests {
     use std::ops::{Bound, RangeInclusive};
 
     use proptest::prelude::ProptestConfig;
@@ -191,7 +191,7 @@ mod tests {
         assert!(test_id_range_for_docs(ops).is_ok());
     }
 
-    fn create_index_from_docs(docs: &[Doc]) -> Index {
+    pub fn create_index_from_docs(docs: &[Doc]) -> Index {
         let mut schema_builder = Schema::builder();
         let id_u64_field = schema_builder.add_u64_field("id", INDEXED | STORED | FAST);
         let ids_u64_field =

src/reader/mod.rs

@@ -44,7 +44,7 @@ pub struct IndexReaderBuilder {
     index: Index,
     warmers: Vec<Weak<dyn Warmer>>,
     num_warming_threads: usize,
-    doc_store_cache_size: usize,
+    doc_store_cache_num_blocks: usize,
 }
 
 impl IndexReaderBuilder {
@@ -55,7 +55,7 @@ impl IndexReaderBuilder {
             index,
             warmers: Vec::new(),
             num_warming_threads: 1,
-            doc_store_cache_size: DOCSTORE_CACHE_CAPACITY,
+            doc_store_cache_num_blocks: DOCSTORE_CACHE_CAPACITY,
         }
     }
 
@@ -72,7 +72,7 @@ impl IndexReaderBuilder {
             searcher_generation_inventory.clone(),
         )?;
         let inner_reader = InnerIndexReader::new(
-            self.doc_store_cache_size,
+            self.doc_store_cache_num_blocks,
             self.index,
             warming_state,
             searcher_generation_inventory,
@@ -119,8 +119,11 @@ impl IndexReaderBuilder {
     ///
     /// The doc store readers cache by default DOCSTORE_CACHE_CAPACITY(100) decompressed blocks.
     #[must_use]
-    pub fn doc_store_cache_size(mut self, doc_store_cache_size: usize) -> IndexReaderBuilder {
-        self.doc_store_cache_size = doc_store_cache_size;
+    pub fn doc_store_cache_num_blocks(
+        mut self,
+        doc_store_cache_num_blocks: usize,
+    ) -> IndexReaderBuilder {
+        self.doc_store_cache_num_blocks = doc_store_cache_num_blocks;
         self
     }
 
@@ -151,7 +154,7 @@ impl TryInto<IndexReader> for IndexReaderBuilder {
 }
 
 struct InnerIndexReader {
-    doc_store_cache_size: usize,
+    doc_store_cache_num_blocks: usize,
     index: Index,
     warming_state: WarmingState,
     searcher: arc_swap::ArcSwap<SearcherInner>,
@@ -161,7 +164,7 @@ struct InnerIndexReader {
 
 impl InnerIndexReader {
     fn new(
-        doc_store_cache_size: usize,
+        doc_store_cache_num_blocks: usize,
         index: Index,
         warming_state: WarmingState,
         // The searcher_generation_inventory is not used as source, but as target to track the
@@ -172,13 +175,13 @@ impl InnerIndexReader {
 
         let searcher = Self::create_searcher(
             &index,
-            doc_store_cache_size,
+            doc_store_cache_num_blocks,
             &warming_state,
             &searcher_generation_counter,
             &searcher_generation_inventory,
         )?;
         Ok(InnerIndexReader {
-            doc_store_cache_size,
+            doc_store_cache_num_blocks,
             index,
             warming_state,
             searcher: ArcSwap::from(searcher),
@@ -214,7 +217,7 @@ impl InnerIndexReader {
 
     fn create_searcher(
         index: &Index,
-        doc_store_cache_size: usize,
+        doc_store_cache_num_blocks: usize,
         warming_state: &WarmingState,
         searcher_generation_counter: &Arc<AtomicU64>,
         searcher_generation_inventory: &Inventory<SearcherGeneration>,
@@ -232,7 +235,7 @@ impl InnerIndexReader {
             index.clone(),
             segment_readers,
             searcher_generation,
-            doc_store_cache_size,
+            doc_store_cache_num_blocks,
         )?);
 
         warming_state.warm_new_searcher_generation(&searcher.clone().into())?;
@@ -242,7 +245,7 @@ impl InnerIndexReader {
     fn reload(&self) -> crate::Result<()> {
         let searcher = Self::create_searcher(
             &self.index,
-            self.doc_store_cache_size,
+            self.doc_store_cache_num_blocks,
             &self.warming_state,
             &self.searcher_generation_counter,
             &self.searcher_generation_inventory,

src/schema/document.rs

@@ -228,7 +228,7 @@ impl Document {
 }
 
 impl BinarySerializable for Document {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         let field_values = self.field_values();
         VInt(field_values.len() as u64).serialize(writer)?;
         for field_value in field_values {

src/schema/facet.rs

@@ -191,7 +191,7 @@ impl<'a, T: ?Sized + AsRef<str>> From<&'a T> for Facet {
 }
 
 impl BinarySerializable for Facet {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         <String as BinarySerializable>::serialize(&self.0, writer)
     }
 

src/schema/field.rs

@@ -23,7 +23,7 @@ impl Field {
 }
 
 impl BinarySerializable for Field {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         self.0.serialize(writer)
     }
 

src/schema/field_value.rs

@@ -36,7 +36,7 @@ impl From<FieldValue> for Value {
 }
 
 impl BinarySerializable for FieldValue {
-    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
         self.field.serialize(writer)?;
         self.value.serialize(writer)
     }

src/schema/value.rs

@@ -344,7 +344,7 @@ mod binary_serialize {
     const TOK_STR_CODE: u8 = 0;
 
     impl BinarySerializable for Value {
-        fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
             match *self {
                 Value::Str(ref text) => {
                     TEXT_CODE.serialize(writer)?;

src/store/footer.rs

@@ -16,7 +16,7 @@ pub struct DocStoreFooter {
 /// - compressor id: 1 byte
 /// - reserved for future use: 15 bytes
 impl BinarySerializable for DocStoreFooter {
-    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
         BinarySerializable::serialize(&DOC_STORE_VERSION, writer)?;
         BinarySerializable::serialize(&self.offset, writer)?;
         BinarySerializable::serialize(&self.decompressor.get_id(), writer)?;

src/store/mod.rs

@@ -4,8 +4,8 @@
 //! order to be handled in the `Store`.
 //!
 //! Internally, documents (or rather their stored fields) are serialized to a buffer.
-//! When the buffer exceeds 16K, the buffer is compressed using `brotli`, `LZ4` or `snappy`
-//! and the resulting block is written to disk.
+//! When the buffer exceeds `block_size` (defaults to 16K), the buffer is compressed using `brotli`,
+//! `LZ4` or `snappy` and the resulting block is written to disk.
 //!
 //! One can then request for a specific `DocId`.
 //! A skip list helps navigating to the right block,
@@ -28,8 +28,6 @@
 //! - at the segment level, the
 //! [`SegmentReader`'s `doc` method](../struct.SegmentReader.html#method.doc)
 //! - at the index level, the [`Searcher::doc()`](crate::Searcher::doc) method
-//!
-//! !
 
 mod compressors;
 mod decompressors;

src/store/reader.rs

@@ -114,7 +114,10 @@ impl Sum for CacheStats {
 
 impl StoreReader {
     /// Opens a store reader
-    pub fn open(store_file: FileSlice, cache_size: usize) -> io::Result<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 open(store_file: FileSlice, cache_num_blocks: usize) -> io::Result<StoreReader> {
         let (footer, data_and_offset) = DocStoreFooter::extract_footer(store_file)?;
 
         let (data_file, offset_index_file) = data_and_offset.split(footer.offset as usize);
@@ -125,8 +128,8 @@ impl StoreReader {
             decompressor: footer.decompressor,
             data: data_file,
             cache: BlockCache {
-                cache: NonZeroUsize::new(cache_size)
-                    .map(|cache_size| Mutex::new(LruCache::new(cache_size))),
+                cache: NonZeroUsize::new(cache_num_blocks)
+                    .map(|cache_num_blocks| Mutex::new(LruCache::new(cache_num_blocks))),
                 cache_hits: Default::default(),
                 cache_misses: Default::default(),
             },

src/termdict/fst_termdict/term_info_store.rs

@@ -21,7 +21,7 @@ struct TermInfoBlockMeta {
 }
 
 impl BinarySerializable for TermInfoBlockMeta {
-    fn serialize<W: Write + ?Sized>(&self, write: &mut W) -> io::Result<()> {
+    fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
         self.offset.serialize(write)?;
         self.ref_term_info.serialize(write)?;
         write.write_all(&[
@@ -272,7 +272,7 @@ impl TermInfoStoreWriter {
         Ok(())
     }
 
-    pub fn serialize<W: io::Write + ?Sized>(&mut self, write: &mut W) -> io::Result<()> {
+    pub fn serialize<W: io::Write>(&mut self, write: &mut W) -> io::Result<()> {
         if !self.term_infos.is_empty() {
             self.flush_block()?;
         }

sstable/src/dictionary.rs

@@ -30,7 +30,6 @@ use crate::{BlockAddr, DeltaReader, Reader, SSTable, SSTableIndex, TermOrdinal,
 /// block boundary.
 ///
 /// (See also README.md)
-#[derive(Debug, Clone)]
 pub struct Dictionary<TSSTable: SSTable = VoidSSTable> {
     pub sstable_slice: FileSlice,
     pub sstable_index: SSTableIndex,

sstable/src/lib.rs

@@ -117,7 +117,6 @@ impl SSTable for MonotonicU64SSTable {
 /// `range_sstable[k1].end == range_sstable[k2].start`.
 ///
 /// The first range is not required to start at `0`.
-#[derive(Clone, Copy, Debug)]
 pub struct RangeSSTable;
 
 impl SSTable for RangeSSTable {

sstable/src/sstable_index.rs

@@ -5,7 +5,7 @@ use serde::{Deserialize, Serialize};
 
 use crate::{common_prefix_len, SSTableDataCorruption, TermOrdinal};
 
-#[derive(Default, Debug, Clone, Serialize, Deserialize)]
+#[derive(Default, Debug, Serialize, Deserialize)]
 pub struct SSTableIndex {
     blocks: Vec<BlockMeta>,
 }
@@ -75,7 +75,7 @@ pub struct BlockAddr {
     pub first_ordinal: u64,
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize)]
 pub(crate) struct BlockMeta {
     /// Any byte string that is lexicographically greater or equal to
     /// the last key in the block,