Added a columnar cli

* fix doc about term endianness

* rustfmt

Cargo.toml

@@ -59,9 +59,8 @@ columnar = { version="0.1", path="./columnar", package ="tantivy-columnar" }
 sstable = { version="0.1", path="./sstable", package ="tantivy-sstable", optional = true }
 stacker = { version="0.1", path="./stacker", package ="tantivy-stacker" }
 tantivy-query-grammar = { version= "0.19.0", path="./query-grammar" }
-tantivy-bitpacker = 		{ version= "0.3", path="./bitpacker" }
-common = 								{ version= "0.5", path = "./common/", package = "tantivy-common" }
-fastfield_codecs = 			{ version= "0.3", path="./fastfield_codecs", default-features = false }
+tantivy-bitpacker = { version= "0.3", path="./bitpacker" }
+common = { version= "0.5", path = "./common/", package = "tantivy-common" }
 tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
 
 [target.'cfg(windows)'.dependencies]
@@ -108,7 +107,7 @@ unstable = [] # useful for benches.
 quickwit = ["sstable"]
 
 [workspace]
-members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
+members = ["query-grammar", "bitpacker", "common", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
 
 # Following the "fail" crate best practises, we isolate
 # tests that define specific behavior in fail check points

TODO.txt

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

bitpacker/src/bitpacker.rs

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

columnar/Cargo.toml

@@ -19,7 +19,7 @@ common = { path = "../common", package = "tantivy-common" }
 tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
 
 [dev-dependencies]
-proptest = "1.0.0"
+proptest = "1"
 more-asserts = "0.3.1"
 rand = "0.8.5"
 

columnar/benches/bench_u128.rs

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

columnar/benches/bench_u64.rs

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

columnar/columnar-cli/Cargo.toml

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

columnar/columnar-cli/src/main.rs

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

columnar/src/TODO.md

@@ -1,22 +1,22 @@
 # zero to one
-* merges
-* full still needs a num_values
-* replug u128
-* add dictionary encoded stuff
-* fix multivalued
-* find a way to make columnar work with strict types
-* plug to tantivy
-    - indexing
-    - aggregations
-    - merge
+
+* revisit line codec
+* removal of all rows of a column in the schema due to deletes
+* add columns from schema on merge
+* Plugging JSON
+* replug examples
+* move datetime to quickwit common
+* switch to nanos
+* reintroduce the gcd map.
 
 # Perf and Size
+* remove alloc in `ord_to_term`
++ multivaued range queries restrat frm the beginning all of the time.
 * re-add ZSTD compression for dictionaries
 no systematic monotonic mapping
 consider removing multilinear
 f32?
 adhoc solution for bool?
-
 add metrics helper for aggregate. sum(row_id)
 review inline absence/presence
 improv perf of select using PDEP
@@ -36,11 +36,13 @@ use the rank & select naming in unit tests branch.
 multi-linear -> blockwise
 linear codec -> simply a multiplication for the index column
 rename columnar to something more explicit, like column_dictionary or columnar_table
+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

@@ -35,27 +35,49 @@ impl BytesColumn {
         self.term_ord_column.num_rows()
     }
 
+    pub fn term_ords(&self, row_id: RowId) -> impl Iterator<Item = u64> + '_ {
+        self.term_ord_column.values(row_id)
+    }
+
     /// Returns the column of ordinals
     pub fn ords(&self) -> &Column<u64> {
         &self.term_ord_column
     }
+
+    pub fn num_terms(&self) -> usize {
+        self.dictionary.num_terms()
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.dictionary.as_ref()
+    }
 }
 
 #[derive(Clone)]
 pub struct StrColumn(BytesColumn);
 
-impl From<BytesColumn> for StrColumn {
-    fn from(bytes_col: BytesColumn) -> Self {
-        StrColumn(bytes_col)
+impl From<StrColumn> for BytesColumn {
+    fn from(str_column: StrColumn) -> BytesColumn {
+        str_column.0
     }
 }
 
 impl StrColumn {
+    pub(crate) fn wrap(bytes_column: BytesColumn) -> StrColumn {
+        StrColumn(bytes_column)
+    }
+
+    pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
+        self.0.dictionary.as_ref()
+    }
+
     /// Fills the buffer
     pub fn ord_to_str(&self, term_ord: u64, output: &mut String) -> io::Result<bool> {
         unsafe {
             let buf = output.as_mut_vec();
-            self.0.dictionary.ord_to_term(term_ord, buf)?;
+            if !self.0.dictionary.ord_to_term(term_ord, buf)? {
+                return Ok(false);
+            }
             // TODO consider remove checks if it hurts performance.
             if std::str::from_utf8(buf.as_slice()).is_err() {
                 buf.clear();

columnar/src/column/mod.rs

@@ -1,27 +1,47 @@
 mod dictionary_encoded;
 mod serialize;
 
+use std::fmt::Debug;
+use std::io::Write;
 use std::ops::Deref;
 use std::sync::Arc;
 
 use common::BinarySerializable;
 pub use dictionary_encoded::{BytesColumn, StrColumn};
 pub use serialize::{
-    open_column_bytes, open_column_u128, open_column_u64, serialize_column_mappable_to_u128,
-    serialize_column_mappable_to_u64,
+    open_column_bytes, open_column_str, open_column_u128, open_column_u64,
+    serialize_column_mappable_to_u128, serialize_column_mappable_to_u64,
 };
 
 use crate::column_index::ColumnIndex;
-use crate::column_values::ColumnValues;
-use crate::{Cardinality, RowId};
+use crate::column_values::monotonic_mapping::StrictlyMonotonicMappingToInternal;
+use crate::column_values::{monotonic_map_column, ColumnValues};
+use crate::{Cardinality, MonotonicallyMappableToU64, RowId};
 
 #[derive(Clone)]
-pub struct Column<T> {
-    pub idx: ColumnIndex<'static>,
+pub struct Column<T = u64> {
+    pub idx: ColumnIndex,
     pub values: Arc<dyn ColumnValues<T>>,
 }
 
-impl<T: PartialOrd> Column<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,
@@ -29,7 +49,7 @@ impl<T: PartialOrd> Column<T> {
             ColumnIndex::Multivalued(col_index) => {
                 // The multivalued index contains all value start row_id,
                 // and one extra value at the end with the overall number of rows.
-                col_index.num_vals() - 1
+                col_index.num_rows()
             }
         }
     }
@@ -37,12 +57,11 @@ impl<T: PartialOrd> Column<T> {
     pub fn min_value(&self) -> T {
         self.values.min_value()
     }
+
     pub fn max_value(&self) -> T {
         self.values.max_value()
     }
-}
 
-impl<T: PartialOrd + Copy + Send + Sync + 'static> Column<T> {
     pub fn first(&self, row_id: RowId) -> Option<T> {
         self.values(row_id).next()
     }
@@ -52,6 +71,15 @@ impl<T: PartialOrd + Copy + Send + Sync + 'static> Column<T> {
             .map(|value_row_id: RowId| self.values.get_val(value_row_id))
     }
 
+    /// Fils the output vector with the (possibly multiple values that are associated_with
+    /// `row_id`.
+    ///
+    /// This method clears the `output` vector.
+    pub fn fill_vals(&self, row_id: RowId, output: &mut Vec<T>) {
+        output.clear();
+        output.extend(self.values(row_id));
+    }
+
     pub fn first_or_default_col(self, default_value: T) -> Arc<dyn ColumnValues<T>> {
         Arc::new(FirstValueWithDefault {
             column: self,
@@ -61,7 +89,7 @@ impl<T: PartialOrd + Copy + Send + Sync + 'static> Column<T> {
 }
 
 impl<T> Deref for Column<T> {
-    type Target = ColumnIndex<'static>;
+    type Target = ColumnIndex;
 
     fn deref(&self) -> &Self::Target {
         &self.idx
@@ -69,7 +97,7 @@ impl<T> Deref for Column<T> {
 }
 
 impl BinarySerializable for Cardinality {
-    fn serialize<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> std::io::Result<()> {
         self.to_code().serialize(writer)
     }
 
@@ -86,7 +114,9 @@ struct FirstValueWithDefault<T: Copy> {
     default_value: T,
 }
 
-impl<T: PartialOrd + Send + Sync + Copy + 'static> ColumnValues<T> for FirstValueWithDefault<T> {
+impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
+    for FirstValueWithDefault<T>
+{
     fn get_val(&self, idx: u32) -> T {
         self.column.first(idx).unwrap_or(self.default_value)
     }
@@ -103,7 +133,7 @@ impl<T: PartialOrd + Send + Sync + Copy + 'static> ColumnValues<T> for FirstValu
         match &self.column.idx {
             ColumnIndex::Full => self.column.values.num_vals(),
             ColumnIndex::Optional(optional_idx) => optional_idx.num_rows(),
-            ColumnIndex::Multivalued(_) => todo!(),
+            ColumnIndex::Multivalued(multivalue_idx) => multivalue_idx.num_rows(),
         }
     }
 }

columnar/src/column/serialize.rs

@@ -8,43 +8,31 @@ use sstable::Dictionary;
 use crate::column::{BytesColumn, Column};
 use crate::column_index::{serialize_column_index, SerializableColumnIndex};
 use crate::column_values::serialize::serialize_column_values_u128;
-use crate::column_values::{
-    serialize_column_values, ColumnValues, FastFieldCodecType, MonotonicallyMappableToU128,
-    MonotonicallyMappableToU64,
-};
+use crate::column_values::u64_based::{serialize_u64_based_column_values, CodecType};
+use crate::column_values::{MonotonicallyMappableToU128, MonotonicallyMappableToU64};
+use crate::iterable::Iterable;
+use crate::StrColumn;
 
-pub fn serialize_column_mappable_to_u128<
-    F: Fn() -> I,
-    I: Iterator<Item = T>,
-    T: MonotonicallyMappableToU128,
->(
+pub fn serialize_column_mappable_to_u128<T: MonotonicallyMappableToU128>(
     column_index: SerializableColumnIndex<'_>,
-    column_values: F,
-    num_vals: u32,
+    iterable: &dyn Iterable<T>,
     output: &mut impl Write,
 ) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_column_values_u128(
-        || column_values().map(|val| val.to_u128()),
-        num_vals,
-        output,
-    )?;
+    serialize_column_values_u128(iterable, output)?;
     output.write_all(&column_index_num_bytes.to_le_bytes())?;
     Ok(())
 }
 
 pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64>(
     column_index: SerializableColumnIndex<'_>,
-    column_values: &impl ColumnValues<T>,
+    column_values: &impl Iterable<T>,
     output: &mut impl Write,
 ) -> io::Result<()> {
     let column_index_num_bytes = serialize_column_index(column_index, output)?;
-    serialize_column_values(
+    serialize_u64_based_column_values(
         column_values,
-        &[
-            FastFieldCodecType::Bitpacked,
-            FastFieldCodecType::BlockwiseLinear,
-        ],
+        &[CodecType::Bitpacked, CodecType::BlockwiseLinear],
         output,
     )?;
     output.write_all(&column_index_num_bytes.to_le_bytes())?;
@@ -61,7 +49,8 @@ pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::
     );
     let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
     let column_index = crate::column_index::open_column_index(column_index_data)?;
-    let column_values = crate::column_values::open_u64_mapped(column_values_data)?;
+    let column_values =
+        crate::column_values::u64_based::load_u64_based_column_values(column_values_data)?;
     Ok(Column {
         idx: column_index,
         values: column_values,
@@ -87,15 +76,19 @@ pub fn open_column_u128<T: MonotonicallyMappableToU128>(
     })
 }
 
-pub fn open_column_bytes<T: From<BytesColumn>>(data: OwnedBytes) -> io::Result<T> {
+pub fn open_column_bytes(data: OwnedBytes) -> io::Result<BytesColumn> {
     let (body, dictionary_len_bytes) = data.rsplit(4);
     let dictionary_len = u32::from_le_bytes(dictionary_len_bytes.as_slice().try_into().unwrap());
     let (dictionary_bytes, column_bytes) = body.split(dictionary_len as usize);
     let dictionary = Arc::new(Dictionary::from_bytes(dictionary_bytes)?);
     let term_ord_column = crate::column::open_column_u64::<u64>(column_bytes)?;
-    let bytes_column = BytesColumn {
+    Ok(BytesColumn {
         dictionary,
         term_ord_column,
-    };
-    Ok(bytes_column.into())
+    })
+}
+
+pub fn open_column_str(data: OwnedBytes) -> io::Result<StrColumn> {
+    let bytes_column = open_column_bytes(data)?;
+    Ok(StrColumn::wrap(bytes_column))
 }

columnar/src/column_index/merge/mod.rs

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

columnar/src/column_index/merge/shuffled.rs

@@ -0,0 +1,171 @@
+use std::iter;
+
+use crate::column_index::{SerializableColumnIndex, Set};
+use crate::iterable::Iterable;
+use crate::{Cardinality, ColumnIndex, RowId, ShuffleMergeOrder};
+
+pub fn merge_column_index_shuffled<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    cardinality_after_merge: Cardinality,
+    shuffle_merge_order: &'a ShuffleMergeOrder,
+) -> SerializableColumnIndex<'a> {
+    match cardinality_after_merge {
+        Cardinality::Full => SerializableColumnIndex::Full,
+        Cardinality::Optional => {
+            let non_null_row_ids =
+                merge_column_index_shuffled_optional(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Optional {
+                non_null_row_ids,
+                num_rows: shuffle_merge_order.num_rows(),
+            }
+        }
+        Cardinality::Multivalued => {
+            let multivalue_start_index =
+                merge_column_index_shuffled_multivalued(column_indexes, shuffle_merge_order);
+            SerializableColumnIndex::Multivalued(multivalue_start_index)
+        }
+    }
+}
+
+/// Merge several column indexes into one, ordering rows according to the merge_order passed as
+/// argument. While it is true that the `merge_order` may imply deletes and hence could in theory a
+/// multivalued index into an optional one, this is not supported today for simplification.
+///
+/// In other words the column_indexes passed as argument may NOT be multivalued.
+fn merge_column_index_shuffled_optional<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledOptionalIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledOptionalIndex<'a> {
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+impl<'a> Iterable<u32> for ShuffledOptionalIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        Box::new(self.merge_order
+        .iter_new_to_old_row_addrs()
+        .enumerate()
+        .filter_map(|(new_row_id, old_row_addr)| {
+            let Some(column_index) = &self.column_indexes[old_row_addr.segment_ord as usize] else {
+                return None;
+            };
+            let row_id = new_row_id as u32;
+            if column_index.has_value(old_row_addr.row_id) {
+                Some(row_id)
+            } else {
+                None
+            }
+        }))
+    }
+}
+
+fn merge_column_index_shuffled_multivalued<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> Box<dyn Iterable<RowId> + 'a> {
+    Box::new(ShuffledMultivaluedIndex {
+        column_indexes,
+        merge_order,
+    })
+}
+
+struct ShuffledMultivaluedIndex<'a> {
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+}
+
+fn iter_num_values<'a>(
+    column_indexes: &'a [Option<ColumnIndex>],
+    merge_order: &'a ShuffleMergeOrder,
+) -> impl Iterator<Item = u32> + 'a {
+    merge_order.iter_new_to_old_row_addrs().map(|row_addr| {
+        let Some(column_index) = &column_indexes[row_addr.segment_ord as usize] else {
+                    // No values in the entire column. It surely means there are 0 values associated to this row.
+                    return 0u32;
+                };
+        match column_index {
+            ColumnIndex::Full => 1,
+            ColumnIndex::Optional(optional_index) => {
+                if optional_index.contains(row_addr.row_id) {
+                    1u32
+                } else {
+                    0u32
+                }
+            }
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_addr.row_id).len() as u32
+            }
+        }
+    })
+}
+
+/// Transforms an iterator containing the number of vals per row (with `num_rows` elements)
+/// into a `start_offset` iterator starting at 0 and (with `num_rows + 1` element)
+fn integrate_num_vals(num_vals: impl Iterator<Item = u32>) -> impl Iterator<Item = RowId> {
+    iter::once(0u32).chain(num_vals.scan(0, |state, num_vals| {
+        *state += num_vals;
+        Some(*state)
+    }))
+}
+
+impl<'a> Iterable<u32> for ShuffledMultivaluedIndex<'a> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
+        let num_vals_per_row = iter_num_values(self.column_indexes, self.merge_order);
+        Box::new(integrate_num_vals(num_vals_per_row))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_index::OptionalIndex;
+    use crate::RowAddr;
+
+    #[test]
+    fn test_integrate_num_vals_empty() {
+        assert!(integrate_num_vals(iter::empty()).eq(iter::once(0)));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_one_el() {
+        assert!(integrate_num_vals(iter::once(10)).eq([0, 10].into_iter()));
+    }
+
+    #[test]
+    fn test_integrate_num_vals_several() {
+        assert!(integrate_num_vals([3, 0, 10, 20].into_iter()).eq([0, 3, 3, 13, 33].into_iter()));
+    }
+
+    #[test]
+    fn test_merge_column_index_optional_shuffle() {
+        let optional_index: ColumnIndex = OptionalIndex::for_test(2, &[0]).into();
+        let column_indexes = vec![Some(optional_index), Some(ColumnIndex::Full)];
+        let row_addrs = vec![
+            RowAddr {
+                segment_ord: 0u32,
+                row_id: 1u32,
+            },
+            RowAddr {
+                segment_ord: 1u32,
+                row_id: 0u32,
+            },
+        ];
+        let shuffle_merge_order = ShuffleMergeOrder::for_test(&[2, 1], row_addrs);
+        let serializable_index = merge_column_index_shuffled(
+            &column_indexes[..],
+            Cardinality::Optional,
+            &shuffle_merge_order,
+        );
+        let SerializableColumnIndex::Optional { non_null_row_ids, num_rows } = serializable_index else { panic!() };
+        assert_eq!(num_rows, 2);
+        let non_null_rows: Vec<RowId> = non_null_row_ids.boxed_iter().collect();
+        assert_eq!(&non_null_rows, &[1]);
+    }
+}

columnar/src/column_index/merge/stacked.rs

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

columnar/src/column_index/mod.rs

@@ -1,30 +1,39 @@
+mod merge;
 mod multivalued_index;
 mod optional_index;
 mod serialize;
 
 use std::ops::Range;
-use std::sync::Arc;
 
-pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
+pub use merge::merge_column_index;
+pub use optional_index::{OptionalIndex, Set};
 pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
 
-use crate::column_values::ColumnValues;
+use crate::column_index::multivalued_index::MultiValueIndex;
 use crate::{Cardinality, RowId};
 
 #[derive(Clone)]
-pub enum ColumnIndex<'a> {
+pub enum ColumnIndex {
     Full,
     Optional(OptionalIndex),
-    // TODO Remove the static by fixing the codec if possible.
-    /// The column values enclosed contains for all row_id,
-    /// the value start_index.
-    ///
     /// In addition, at index num_rows, an extra value is added
     /// containing the overal number of values.
-    Multivalued(Arc<dyn ColumnValues<RowId> + 'a>),
+    Multivalued(MultiValueIndex),
 }
 
-impl<'a> ColumnIndex<'a> {
+impl From<OptionalIndex> for ColumnIndex {
+    fn from(optional_index: OptionalIndex) -> ColumnIndex {
+        ColumnIndex::Optional(optional_index)
+    }
+}
+
+impl From<MultiValueIndex> for ColumnIndex {
+    fn from(multi_value_index: MultiValueIndex) -> ColumnIndex {
+        ColumnIndex::Multivalued(multi_value_index)
+    }
+}
+
+impl ColumnIndex {
     pub fn get_cardinality(&self) -> Cardinality {
         match self {
             ColumnIndex::Full => Cardinality::Full,
@@ -33,6 +42,17 @@ impl<'a> ColumnIndex<'a> {
         }
     }
 
+    /// Returns true if and only if there are at least one value associated to the row.
+    pub fn has_value(&self, row_id: RowId) -> bool {
+        match self {
+            ColumnIndex::Full => true,
+            ColumnIndex::Optional(optional_index) => optional_index.contains(row_id),
+            ColumnIndex::Multivalued(multivalued_index) => {
+                multivalued_index.range(row_id).len() > 0
+            }
+        }
+    }
+
     pub fn value_row_ids(&self, row_id: RowId) -> Range<RowId> {
         match self {
             ColumnIndex::Full => row_id..row_id + 1,
@@ -43,11 +63,22 @@ impl<'a> ColumnIndex<'a> {
                     0..0
                 }
             }
+            ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(row_id),
+        }
+    }
+
+    pub fn select_batch_in_place(&self, rank_ids: &mut Vec<RowId>) {
+        match self {
+            ColumnIndex::Full => {
+                // No need to do anything:
+                // value_idx and row_idx are the same.
+            }
+            ColumnIndex::Optional(optional_index) => {
+                optional_index.select_batch(&mut rank_ids[..]);
+            }
             ColumnIndex::Multivalued(multivalued_index) => {
-                let multivalued_index_ref = &**multivalued_index;
-                let start: u32 = multivalued_index_ref.get_val(row_id);
-                let end: u32 = multivalued_index_ref.get_val(row_id + 1);
-                start..end
+                // TODO important: avoid using 0u32, and restart from the beginning all of the time.
+                multivalued_index.select_batch_in_place(0u32, rank_ids)
             }
         }
     }

columnar/src/column_index/multivalued_index.rs

@@ -1,29 +1,141 @@
 use std::io;
 use std::io::Write;
+use std::ops::Range;
 use std::sync::Arc;
 
 use common::OwnedBytes;
 
-use crate::column_values::{ColumnValues, FastFieldCodecType};
+use crate::column_values::u64_based::CodecType;
+use crate::column_values::ColumnValues;
+use crate::iterable::Iterable;
 use crate::RowId;
 
-#[derive(Clone)]
-pub struct MultivaluedIndex(Arc<dyn ColumnValues<RowId>>);
-
 pub fn serialize_multivalued_index(
-    multivalued_index: &dyn ColumnValues<RowId>,
+    multivalued_index: &dyn Iterable<RowId>,
     output: &mut impl Write,
 ) -> io::Result<()> {
-    crate::column_values::serialize_column_values(
-        &*multivalued_index,
-        &[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
+    crate::column_values::u64_based::serialize_u64_based_column_values(
+        multivalued_index,
+        &[CodecType::Bitpacked, CodecType::Linear],
         output,
     )?;
     Ok(())
 }
 
-pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<Arc<dyn ColumnValues<RowId>>> {
+pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex> {
     let start_index_column: Arc<dyn ColumnValues<RowId>> =
-        crate::column_values::open_u64_mapped(bytes)?;
-    Ok(start_index_column)
+        crate::column_values::u64_based::load_u64_based_column_values(bytes)?;
+    Ok(MultiValueIndex { start_index_column })
+}
+
+#[derive(Clone)]
+/// Index to resolve value range for given doc_id.
+/// Starts at 0.
+pub struct MultiValueIndex {
+    pub start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
+}
+
+impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {
+    fn from(start_index_column: Arc<dyn ColumnValues<RowId>>) -> Self {
+        MultiValueIndex { start_index_column }
+    }
+}
+
+impl MultiValueIndex {
+    pub fn for_test(start_offsets: &[RowId]) -> MultiValueIndex {
+        let mut buffer = Vec::new();
+        serialize_multivalued_index(&start_offsets, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_multivalued_index(bytes).unwrap()
+    }
+
+    /// Returns `[start, end)`, such that the values associated with
+    /// the given document are `start..end`.
+    #[inline]
+    pub(crate) fn range(&self, row_id: RowId) -> Range<RowId> {
+        let start = self.start_index_column.get_val(row_id);
+        let end = self.start_index_column.get_val(row_id + 1);
+        start..end
+    }
+
+    /// Returns the number of documents in the index.
+    #[inline]
+    pub fn num_rows(&self) -> u32 {
+        self.start_index_column.num_vals() - 1
+    }
+
+    /// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
+    /// row_ids. Positions are converted inplace to docids.
+    ///
+    /// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
+    /// index.
+    ///
+    /// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
+    /// increasing positions.
+    ///
+    /// TODO: Instead of a linear scan we can employ a exponential search into binary search to
+    /// match a docid to its value position.
+    #[allow(clippy::bool_to_int_with_if)]
+    pub(crate) fn select_batch_in_place(&self, row_start: RowId, ranks: &mut Vec<u32>) {
+        if ranks.is_empty() {
+            return;
+        }
+        let mut cur_doc = row_start;
+        let mut last_doc = None;
+
+        assert!(self.start_index_column.get_val(row_start) as u32 <= ranks[0]);
+
+        let mut write_doc_pos = 0;
+        for i in 0..ranks.len() {
+            let pos = ranks[i];
+            loop {
+                let end = self.start_index_column.get_val(cur_doc + 1) as u32;
+                if end > pos {
+                    ranks[write_doc_pos] = cur_doc;
+                    write_doc_pos += if last_doc == Some(cur_doc) { 0 } else { 1 };
+                    last_doc = Some(cur_doc);
+                    break;
+                }
+                cur_doc += 1;
+            }
+        }
+        ranks.truncate(write_doc_pos);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ops::Range;
+    use std::sync::Arc;
+
+    use super::MultiValueIndex;
+    use crate::column_values::IterColumn;
+    use crate::{ColumnValues, RowId};
+
+    fn index_to_pos_helper(
+        index: &MultiValueIndex,
+        doc_id_range: Range<u32>,
+        positions: &[u32],
+    ) -> Vec<u32> {
+        let mut positions = positions.to_vec();
+        index.select_batch_in_place(doc_id_range.start, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_positions_to_docid() {
+        let offsets: Vec<RowId> = vec![0, 10, 12, 15, 22, 23]; // docid values are [0..10, 10..12, 12..15, etc.]
+        let column: Arc<dyn ColumnValues<RowId>> = Arc::new(IterColumn::from(offsets.into_iter()));
+        let index = MultiValueIndex::from(column);
+        assert_eq!(index.num_rows(), 5);
+        let positions = &[10u32, 11, 15, 20, 21, 22];
+        assert_eq!(index_to_pos_helper(&index, 0..5, positions), vec![1, 3, 4]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, positions), vec![1, 3, 4]);
+        assert_eq!(index_to_pos_helper(&index, 0..5, &[9]), vec![0]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[10]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 1..5, &[11]), vec![1]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14]), vec![2]);
+        assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14, 15]), vec![2, 3]);
+    }
 }

columnar/src/column_index/optional_index/mod.rs

@@ -1,16 +1,16 @@
 use std::io::{self, Write};
-use std::ops::Range;
 use std::sync::Arc;
 
 mod set;
 mod set_block;
 
-use common::{BinarySerializable, GroupByIteratorExtended, OwnedBytes, VInt};
-pub use set::{Set, SetCodec};
+use common::{BinarySerializable, OwnedBytes, VInt};
+pub use set::{SelectCursor, Set, SetCodec};
 use set_block::{
     DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
 };
 
+use crate::iterable::Iterable;
 use crate::{InvalidData, RowId};
 
 /// The threshold for for number of elements after which we switch to dense block encoding.
@@ -88,16 +88,6 @@ pub struct OptionalIndex {
     block_metas: Arc<[BlockMeta]>,
 }
 
-impl OptionalIndex {
-    pub fn num_rows(&self) -> RowId {
-        self.num_rows
-    }
-
-    pub fn num_non_nulls(&self) -> RowId {
-        self.num_non_null_rows
-    }
-}
-
 /// Splits a value address into lower and upper 16bits.
 /// The lower 16 bits are the value in the block
 /// The upper 16 bits are the block index
@@ -115,7 +105,63 @@ fn row_addr_from_row_id(row_id: RowId) -> RowAddr {
     }
 }
 
+enum BlockSelectCursor<'a> {
+    Dense(<DenseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+    Sparse(<SparseBlock<'a> as Set<u16>>::SelectCursor<'a>),
+}
+
+impl<'a> BlockSelectCursor<'a> {
+    fn select(&mut self, rank: u16) -> u16 {
+        match self {
+            BlockSelectCursor::Dense(dense_select_cursor) => dense_select_cursor.select(rank),
+            BlockSelectCursor::Sparse(sparse_select_cursor) => sparse_select_cursor.select(rank),
+        }
+    }
+}
+pub struct OptionalIndexSelectCursor<'a> {
+    current_block_cursor: BlockSelectCursor<'a>,
+    current_block_id: u16,
+    // The current block is guaranteed to contain ranks < end_rank.
+    current_block_end_rank: RowId,
+    optional_index: &'a OptionalIndex,
+    block_doc_idx_start: RowId,
+    num_null_rows_before_block: RowId,
+}
+
+impl<'a> OptionalIndexSelectCursor<'a> {
+    fn search_and_load_block(&mut self, rank: RowId) {
+        if rank < self.current_block_end_rank {
+            // we are already in the right block
+            return;
+        }
+        self.current_block_id = self.optional_index.find_block(rank, self.current_block_id);
+        self.current_block_end_rank = self
+            .optional_index
+            .block_metas
+            .get(self.current_block_id as usize + 1)
+            .map(|block_meta| block_meta.non_null_rows_before_block)
+            .unwrap_or(u32::MAX);
+        self.block_doc_idx_start = (self.current_block_id as u32) * ELEMENTS_PER_BLOCK;
+        let block_meta = self.optional_index.block_metas[self.current_block_id as usize];
+        self.num_null_rows_before_block = block_meta.non_null_rows_before_block;
+        let block: Block<'_> = self.optional_index.block(block_meta);
+        self.current_block_cursor = match block {
+            Block::Dense(dense_block) => BlockSelectCursor::Dense(dense_block.select_cursor()),
+            Block::Sparse(sparse_block) => BlockSelectCursor::Sparse(sparse_block.select_cursor()),
+        };
+    }
+}
+
+impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor<'a> {
+    fn select(&mut self, rank: RowId) -> RowId {
+        self.search_and_load_block(rank);
+        let index_in_block = (rank - self.num_null_rows_before_block) as u16;
+        self.current_block_cursor.select(index_in_block) as RowId + self.block_doc_idx_start
+    }
+}
+
 impl Set<RowId> for OptionalIndex {
+    type SelectCursor<'b> = OptionalIndexSelectCursor<'b> where Self: 'b;
     // Check if value at position is not null.
     #[inline]
     fn contains(&self, row_id: RowId) -> bool {
@@ -130,6 +176,21 @@ impl Set<RowId> for OptionalIndex {
         }
     }
 
+    #[inline]
+    fn rank(&self, row_id: RowId) -> RowId {
+        let RowAddr {
+            block_id,
+            in_block_row_id,
+        } = row_addr_from_row_id(row_id);
+        let block_meta = self.block_metas[block_id as usize];
+        let block = self.block(block_meta);
+        let block_offset_row_id = match block {
+            Block::Dense(dense_block) => dense_block.rank(in_block_row_id),
+            Block::Sparse(sparse_block) => sparse_block.rank(in_block_row_id),
+        } as u32;
+        block_meta.non_null_rows_before_block + block_offset_row_id
+    }
+
     #[inline]
     fn rank_if_exists(&self, row_id: RowId) -> Option<RowId> {
         let RowAddr {
@@ -148,7 +209,7 @@ impl Set<RowId> for OptionalIndex {
     #[inline]
     fn select(&self, rank: RowId) -> RowId {
         let block_pos = self.find_block(rank, 0);
-        let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
+        let block_doc_idx_start = (block_pos as u32) * ELEMENTS_PER_BLOCK;
         let block_meta = self.block_metas[block_pos as usize];
         let block: Block<'_> = self.block(block_meta);
         let index_in_block = (rank - block_meta.non_null_rows_before_block) as u16;
@@ -159,39 +220,53 @@ impl Set<RowId> for OptionalIndex {
         block_doc_idx_start + in_block_rank as u32
     }
 
-    fn select_batch(&self, ranks: &[u32], output_idxs: &mut [u32]) {
-        let mut block_pos = 0u32;
-        let mut start = 0;
-        let group_by_it = ranks.iter().copied().group_by(move |codec_idx| {
-            block_pos = self.find_block(*codec_idx, block_pos);
-            block_pos
-        });
-        for (block_pos, block_iter) in group_by_it {
-            let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
-            let block_meta = self.block_metas[block_pos as usize];
-            let block: Block<'_> = self.block(block_meta);
-            let offset = block_meta.non_null_rows_before_block;
-            let indexes_in_block_iter =
-                block_iter.map(move |codec_idx| (codec_idx - offset) as u16);
-            match block {
-                Block::Dense(dense_block) => {
-                    for in_offset in dense_block.select_iter(indexes_in_block_iter) {
-                        output_idxs[start] = in_offset as u32 + block_doc_idx_start;
-                        start += 1;
-                    }
-                }
-                Block::Sparse(sparse_block) => {
-                    for in_offset in sparse_block.select_iter(indexes_in_block_iter) {
-                        output_idxs[start] = in_offset as u32 + block_doc_idx_start;
-                        start += 1;
-                    }
-                }
-            };
+    fn select_cursor<'b>(&'b self) -> OptionalIndexSelectCursor<'b> {
+        OptionalIndexSelectCursor {
+            current_block_cursor: BlockSelectCursor::Sparse(
+                SparseBlockCodec::open(b"").select_cursor(),
+            ),
+            current_block_id: 0u16,
+            current_block_end_rank: 0u32, //< this is sufficient to force the first load
+            optional_index: self,
+            block_doc_idx_start: 0u32,
+            num_null_rows_before_block: 0u32,
         }
     }
 }
 
 impl OptionalIndex {
+    pub fn for_test(num_rows: RowId, row_ids: &[RowId]) -> OptionalIndex {
+        assert!(row_ids
+            .last()
+            .copied()
+            .map(|last_row_id| last_row_id < num_rows)
+            .unwrap_or(true));
+        let mut buffer = Vec::new();
+        serialize_optional_index(&row_ids, num_rows, &mut buffer).unwrap();
+        let bytes = OwnedBytes::new(buffer);
+        open_optional_index(bytes).unwrap()
+    }
+
+    pub fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
+    pub fn num_non_nulls(&self) -> RowId {
+        self.num_non_null_rows
+    }
+
+    pub fn iter_rows<'a>(&'a self) -> impl Iterator<Item = RowId> + 'a {
+        // TODO optimize
+        let mut select_batch = self.select_cursor();
+        (0..self.num_non_null_rows).map(move |rank| select_batch.select(rank))
+    }
+    pub fn select_batch(&self, ranks: &mut [RowId]) {
+        let mut select_cursor = self.select_cursor();
+        for rank in ranks.iter_mut() {
+            *rank = select_cursor.select(*rank);
+        }
+    }
+
     #[inline]
     fn block<'a>(&'a self, block_meta: BlockMeta) -> Block<'a> {
         let BlockMeta {
@@ -214,14 +289,14 @@ impl OptionalIndex {
     }
 
     #[inline]
-    fn find_block(&self, dense_idx: u32, start_block_pos: u32) -> u32 {
-        for block_pos in start_block_pos..self.block_metas.len() as u32 {
+    fn find_block(&self, dense_idx: u32, start_block_pos: u16) -> u16 {
+        for block_pos in start_block_pos..self.block_metas.len() as u16 {
             let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
             if offset > dense_idx {
-                return block_pos - 1;
+                return block_pos - 1u16;
             }
         }
-        self.block_metas.len() as u32 - 1u32
+        self.block_metas.len() as u16 - 1u16
     }
 
     // TODO Add a good API for the codec_idx to original_idx translation.
@@ -255,7 +330,7 @@ impl OptionalIndexCodec {
 }
 
 impl BinarySerializable for OptionalIndexCodec {
-    fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         writer.write_all(&[self.to_code()])
     }
 
@@ -277,12 +352,13 @@ fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -
 }
 
 pub fn serialize_optional_index<'a, W: io::Write>(
-    serializable_optional_index: &dyn SerializableOptionalIndex<'a>,
+    non_null_rows: &dyn Iterable<RowId>,
+    num_rows: RowId,
     output: &mut W,
 ) -> io::Result<()> {
-    VInt(serializable_optional_index.num_rows() as u64).serialize(output)?;
+    VInt(num_rows as u64).serialize(output)?;
 
-    let mut rows_it = serializable_optional_index.non_null_rows();
+    let mut rows_it = non_null_rows.boxed_iter();
     let mut block_metadata: Vec<SerializedBlockMeta> = Vec::new();
     let mut current_block = Vec::new();
 
@@ -435,19 +511,5 @@ pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
     Ok(optional_index)
 }
 
-pub trait SerializableOptionalIndex<'a> {
-    fn num_rows(&self) -> RowId;
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a>;
-}
-
-impl SerializableOptionalIndex<'static> for Range<u32> {
-    fn num_rows(&self) -> RowId {
-        self.end
-    }
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'static> {
-        Box::new(self.clone())
-    }
-}
-
 #[cfg(test)]
 mod tests;

columnar/src/column_index/optional_index/set.rs

@@ -13,11 +13,25 @@ pub trait SetCodec {
     fn open<'a>(data: &'a [u8]) -> Self::Reader<'a>;
 }
 
+/// Stateful object that makes it possible to compute several select in a row,
+/// provided the rank passed as argument are increasing.
+pub trait SelectCursor<T> {
+    // May panic if rank is greater than the number of elements in the Set,
+    // or if rank is < than value provided in the previous call.
+    fn select(&mut self, rank: T) -> T;
+}
+
 pub trait Set<T> {
+    type SelectCursor<'b>: SelectCursor<T>
+    where Self: 'b;
+
     /// Returns true if the elements is contained in the Set
     fn contains(&self, el: T) -> bool;
 
-    /// If the set contains `el` returns its position in the sortd set of elements.
+    /// Returns the number of rows in the set that are < `el`
+    fn rank(&self, el: T) -> T;
+
+    /// If the set contains `el` returns the element rank.
     /// If the set does not contain the element, it returns `None`.
     fn rank_if_exists(&self, el: T) -> Option<T>;
 
@@ -28,11 +42,6 @@ pub trait Set<T> {
     /// May panic if rank is greater than the number of elements in the Set.
     fn select(&self, rank: T) -> T;
 
-    /// Batch version of select.
-    /// `ranks` is assumed to be sorted.
-    ///
-    /// # Panics
-    ///
-    /// May panic if rank is greater than the number of elements in the Set.
-    fn select_batch(&self, ranks: &[T], outputs: &mut [T]);
+    /// Creates a brand new select cursor.
+    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b>;
 }

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

@@ -3,7 +3,7 @@ use std::io::{self, Write};
 
 use common::BinarySerializable;
 
-use crate::column_index::optional_index::{Set, SetCodec, ELEMENTS_PER_BLOCK};
+use crate::column_index::optional_index::{SelectCursor, Set, SetCodec, ELEMENTS_PER_BLOCK};
 
 #[inline(always)]
 fn get_bit_at(input: u64, n: u16) -> bool {
@@ -105,7 +105,27 @@ impl DenseMiniBlock {
 #[derive(Copy, Clone)]
 pub struct DenseBlock<'a>(&'a [u8]);
 
+pub struct DenseBlockSelectCursor<'a> {
+    block_id: u16,
+    dense_block: DenseBlock<'a>,
+}
+
+impl<'a> SelectCursor<u16> for DenseBlockSelectCursor<'a> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        self.block_id = self
+            .dense_block
+            .find_miniblock_containing_rank(rank, self.block_id)
+            .unwrap();
+        let index_block = self.dense_block.mini_block(self.block_id);
+        let in_block_rank = rank - index_block.rank;
+        self.block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
+    }
+}
+
 impl<'a> Set<u16> for DenseBlock<'a> {
+    type SelectCursor<'b> = DenseBlockSelectCursor<'a> where Self: 'b;
+
     #[inline(always)]
     fn contains(&self, el: u16) -> bool {
         let mini_block_id = el / ELEMENTS_PER_MINI_BLOCK;
@@ -128,6 +148,15 @@ impl<'a> Set<u16> for DenseBlock<'a> {
         }
     }
 
+    #[inline(always)]
+    fn rank(&self, el: u16) -> u16 {
+        let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
+        let index_block = self.mini_block(block_pos);
+        let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
+        let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
+        index_block.rank + ones_in_block
+    }
+
     #[inline(always)]
     fn select(&self, rank: u16) -> u16 {
         let block_id = self.find_miniblock_containing_rank(rank, 0).unwrap();
@@ -136,37 +165,15 @@ impl<'a> Set<u16> for DenseBlock<'a> {
         block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
     }
 
-    fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
-        let orig_ids = self.select_iter(ranks.iter().copied());
-        for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
-            *output = original_id;
+    #[inline(always)]
+    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
+        DenseBlockSelectCursor {
+            block_id: 0,
+            dense_block: *self,
         }
     }
 }
 
-impl<'a> DenseBlock<'a> {
-    /// Iterator verison of select.
-    ///
-    /// # Panics
-    /// Panics if one of the rank is higher than the number of elements in the set.
-    pub fn select_iter<'b>(
-        &self,
-        rank_it: impl Iterator<Item = u16> + 'b,
-    ) -> impl Iterator<Item = u16> + 'b
-    where
-        Self: 'b,
-    {
-        let mut block_id = 0u16;
-        let me = *self;
-        rank_it.map(move |rank| {
-            block_id = me.find_miniblock_containing_rank(rank, block_id).unwrap();
-            let index_block = me.mini_block(block_id);
-            let in_block_rank = rank - index_block.rank;
-            block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
-        })
-    }
-}
-
 impl<'a> DenseBlock<'a> {
     #[inline]
     fn mini_block(&self, mini_block_id: u16) -> DenseMiniBlock {

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

@@ -1,4 +1,4 @@
-use crate::column_index::optional_index::{Set, SetCodec};
+use crate::column_index::optional_index::{SelectCursor, Set, SetCodec};
 
 pub struct SparseBlockCodec;
 
@@ -24,7 +24,16 @@ impl SetCodec for SparseBlockCodec {
 #[derive(Copy, Clone)]
 pub struct SparseBlock<'a>(&'a [u8]);
 
+impl<'a> SelectCursor<u16> for SparseBlock<'a> {
+    #[inline]
+    fn select(&mut self, rank: u16) -> u16 {
+        <SparseBlock<'a> as Set<u16>>::select(self, rank)
+    }
+}
+
 impl<'a> Set<u16> for SparseBlock<'a> {
+    type SelectCursor<'b> = Self where Self: 'b;
+
     #[inline(always)]
     fn contains(&self, el: u16) -> bool {
         self.binary_search(el).is_ok()
@@ -35,17 +44,20 @@ 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;
         u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
     }
 
-    fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
-        let orig_ids = self.select_iter(ranks.iter().copied());
-        for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
-            *output = original_id;
-        }
+    #[inline(always)]
+    fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
+        *self
     }
 }
 
@@ -96,17 +108,4 @@ impl<'a> SparseBlock<'a> {
         }
         Err(left)
     }
-
-    pub fn select_iter<'b>(
-        &self,
-        iter: impl Iterator<Item = u16> + 'b,
-    ) -> impl Iterator<Item = u16> + 'b
-    where
-        Self: 'b,
-    {
-        iter.map(|codec_id| {
-            let offset = codec_id as usize * 2;
-            u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
-        })
-    }
 }

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

@@ -1,9 +1,8 @@
 use std::collections::HashMap;
 
-use crate::column_index::optional_index::set_block::{
-    DenseBlockCodec, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
-};
-use crate::column_index::optional_index::{Set, SetCodec};
+use crate::column_index::optional_index::set_block::dense::DENSE_BLOCK_NUM_BYTES;
+use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
+use crate::column_index::optional_index::{SelectCursor, Set, SetCodec};
 
 fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
     let mut buffer = Vec::new();
@@ -18,6 +17,10 @@ fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
     for val in 0u16..=u16::MAX {
         assert_eq!(tested_set.contains(val), hash_set.contains_key(&val));
         assert_eq!(tested_set.rank_if_exists(val), hash_set.get(&val).copied());
+        assert_eq!(
+            tested_set.rank(val),
+            vals.iter().cloned().take_while(|v| *v < val).count() as u16
+        );
     }
     for rank in 0..vals.len() {
         assert_eq!(tested_set.select(rank as u16), vals[rank]);
@@ -75,12 +78,10 @@ fn test_simple_translate_codec_codec_idx_to_original_idx_dense() {
         .unwrap();
     let tested_set = DenseBlockCodec::open(buffer.as_slice());
     assert!(tested_set.contains(1));
-    assert_eq!(
-        &tested_set
-            .select_iter([0, 1, 2, 5].iter().copied())
-            .collect::<Vec<u16>>(),
-        &[1, 3, 17, 30_001]
-    );
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(select_cursor.select(0), 1);
+    assert_eq!(select_cursor.select(1), 3);
+    assert_eq!(select_cursor.select(2), 17);
 }
 
 #[test]
@@ -89,12 +90,10 @@ fn test_simple_translate_codec_idx_to_original_idx_sparse() {
     SparseBlockCodec::serialize([1, 3, 17].iter().copied(), &mut buffer).unwrap();
     let tested_set = SparseBlockCodec::open(buffer.as_slice());
     assert!(tested_set.contains(1));
-    assert_eq!(
-        &tested_set
-            .select_iter([0, 1, 2].iter().copied())
-            .collect::<Vec<u16>>(),
-        &[1, 3, 17]
-    );
+    let mut select_cursor = tested_set.select_cursor();
+    assert_eq!(SelectCursor::select(&mut select_cursor, 0), 1);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 1), 3);
+    assert_eq!(SelectCursor::select(&mut select_cursor, 2), 17);
 }
 
 #[test]
@@ -103,10 +102,8 @@ fn test_simple_translate_codec_idx_to_original_idx_dense() {
     DenseBlockCodec::serialize(0u16..150u16, &mut buffer).unwrap();
     let tested_set = DenseBlockCodec::open(buffer.as_slice());
     assert!(tested_set.contains(1));
-    let rg = 0u16..150u16;
-    let els: Vec<u16> = rg.clone().collect();
-    assert_eq!(
-        &tested_set.select_iter(rg.clone()).collect::<Vec<u16>>(),
-        &els
-    );
+    let mut select_cursor = tested_set.select_cursor();
+    for i in 0..150 {
+        assert_eq!(i, select_cursor.select(i));
+    }
 }

columnar/src/column_index/optional_index/tests.rs

@@ -37,13 +37,14 @@ proptest! {
 fn test_with_random_sets_simple() {
     let vals = 10..BLOCK_SIZE * 2;
     let mut out: Vec<u8> = Vec::new();
-    serialize_optional_index(&vals.clone(), &mut out).unwrap();
+    serialize_optional_index(&vals.clone(), 100, &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
     let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
-    let mut output = vec![0u32; ranks.len()];
-    null_index.select_batch(&ranks[..], &mut output[..]);
-    assert_eq!(&output, &els);
+    let mut select_cursor = null_index.select_cursor();
+    for (rank, el) in ranks.iter().copied().zip(els.iter().copied()) {
+        assert_eq!(select_cursor.select(rank), el);
+    }
 }
 
 #[test]
@@ -65,12 +66,8 @@ fn test_optional_index_one_block_true() {
     test_null_index(&iter[..]);
 }
 
-impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
-    fn num_rows(&self) -> RowId {
-        self.len() as u32
-    }
-
-    fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
+impl<'a> Iterable<RowId> for &'a [bool] {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
         Box::new(
             self.iter()
                 .cloned()
@@ -83,7 +80,7 @@ impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
 
 fn test_null_index(data: &[bool]) {
     let mut out: Vec<u8> = Vec::new();
-    serialize_optional_index(&data, &mut out).unwrap();
+    serialize_optional_index(&data, data.len() as RowId, &mut out).unwrap();
     let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
     let orig_idx_with_value: Vec<u32> = data
         .iter()
@@ -91,11 +88,10 @@ fn test_null_index(data: &[bool]) {
         .filter(|(_pos, val)| **val)
         .map(|(pos, _val)| pos as u32)
         .collect();
-    let ids: Vec<u32> = (0..orig_idx_with_value.len() as u32).collect();
-    let mut output = vec![0u32; ids.len()];
-    null_index.select_batch(&ids[..], &mut output);
-    // assert_eq!(&output[0..100], &orig_idx_with_value[0..100]);
-    assert_eq!(output, orig_idx_with_value);
+    let mut select_iter = null_index.select_cursor();
+    for i in 0..orig_idx_with_value.len() {
+        assert_eq!(select_iter.select(i as u32), orig_idx_with_value[i]);
+    }
 
     let step_size = (orig_idx_with_value.len() / 100).max(1);
     for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate().step_by(step_size) {
@@ -111,51 +107,96 @@ fn test_null_index(data: &[bool]) {
 
 #[test]
 fn test_optional_index_test_translation() {
-    let mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    let mut output = vec![0u32; 2];
-    null_index.select_batch(&[0, 1], &mut output);
-    assert_eq!(output, &[0, 2]);
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    let mut select_cursor = optional_index.select_cursor();
+    assert_eq!(select_cursor.select(0), 0);
+    assert_eq!(select_cursor.select(1), 2);
 }
 
 #[test]
 fn test_optional_index_translate() {
-    let mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert_eq!(null_index.rank_if_exists(0), Some(0));
-    assert_eq!(null_index.rank_if_exists(2), Some(1));
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert_eq!(optional_index.rank_if_exists(0), Some(0));
+    assert_eq!(optional_index.rank_if_exists(2), Some(1));
 }
 
 #[test]
 fn test_optional_index_small() {
-    let mut out = vec![];
-    let iter = &[true, false, true, false];
-    serialize_optional_index(&&iter[..], &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert!(null_index.contains(0));
-    assert!(!null_index.contains(1));
-    assert!(null_index.contains(2));
-    assert!(!null_index.contains(3));
+    let optional_index = OptionalIndex::for_test(4, &[0, 2]);
+    assert!(optional_index.contains(0));
+    assert!(!optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
 }
 
 #[test]
 fn test_optional_index_large() {
-    let mut docs = vec![];
-    docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
-    docs.extend((0..=1).map(|_idx| true));
+    let row_ids = &[ELEMENTS_PER_BLOCK, ELEMENTS_PER_BLOCK + 1];
+    let optional_index = OptionalIndex::for_test(ELEMENTS_PER_BLOCK + 2, row_ids);
+    assert!(!optional_index.contains(0));
+    assert!(!optional_index.contains(100));
+    assert!(!optional_index.contains(ELEMENTS_PER_BLOCK - 1));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK));
+    assert!(optional_index.contains(ELEMENTS_PER_BLOCK + 1));
+}
 
-    let mut out = vec![];
-    serialize_optional_index(&&docs[..], &mut out).unwrap();
-    let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
-    assert!(!null_index.contains(0));
-    assert!(!null_index.contains(100));
-    assert!(!null_index.contains(ELEMENTS_PER_BLOCK - 1));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK));
-    assert!(null_index.contains(ELEMENTS_PER_BLOCK + 1));
+fn test_optional_index_iter_aux(row_ids: &[RowId], num_rows: RowId) {
+    let optional_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(optional_index.num_rows(), num_rows);
+    assert!(optional_index.iter_rows().eq(row_ids.iter().copied()));
+}
+
+#[test]
+fn test_optional_index_iter_empty() {
+    test_optional_index_iter_aux(&[], 0u32);
+}
+
+fn test_optional_index_rank_aux(row_ids: &[RowId]) {
+    let num_rows = row_ids.last().copied().unwrap_or(0u32) + 1;
+    let null_index = OptionalIndex::for_test(num_rows, row_ids);
+    assert_eq!(null_index.num_rows(), num_rows);
+    for (row_id, row_val) in row_ids.iter().copied().enumerate() {
+        assert_eq!(null_index.rank(row_val), row_id as u32);
+        assert_eq!(null_index.rank_if_exists(row_val), Some(row_id as u32));
+        if row_val > 0 && !null_index.contains(&row_val - 1) {
+            assert_eq!(null_index.rank(row_val - 1), row_id as u32);
+        }
+        assert_eq!(null_index.rank(row_val + 1), row_id as u32 + 1);
+    }
+}
+
+#[test]
+fn test_optional_index_rank() {
+    test_optional_index_rank_aux(&[1u32]);
+    test_optional_index_rank_aux(&[0u32, 1u32]);
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_rank_aux(&block);
+}
+
+#[test]
+fn test_optional_index_iter_empty_one() {
+    test_optional_index_iter_aux(&[1], 2u32);
+    test_optional_index_iter_aux(&[100_000], 200_000u32);
+}
+
+#[test]
+fn test_optional_index_iter_dense_block() {
+    let mut block = Vec::new();
+    block.push(3u32);
+    block.extend((0..BLOCK_SIZE).map(|i| i + BLOCK_SIZE + 1));
+    test_optional_index_iter_aux(&block, 3 * BLOCK_SIZE);
+}
+
+#[test]
+fn test_optional_index_for_tests() {
+    let optional_index = OptionalIndex::for_test(4, &[1, 2]);
+    assert!(!optional_index.contains(0));
+    assert!(optional_index.contains(1));
+    assert!(optional_index.contains(2));
+    assert!(!optional_index.contains(3));
+    assert_eq!(optional_index.num_rows(), 4);
 }
 
 #[cfg(all(test, feature = "unstable"))]
@@ -175,7 +216,6 @@ mod bench {
             .map(|_| rng.gen_bool(fill_ratio))
             .collect();
         serialize_optional_index(&&vals[..], &mut out).unwrap();
-
         let codec = open_optional_index(OwnedBytes::new(out)).unwrap();
         codec
     }
@@ -311,7 +351,8 @@ mod bench {
         };
         let mut output = vec![0u32; idxs.len()];
         bench.iter(|| {
-            codec.select_batch(&idxs[..], &mut output);
+            output.copy_from_slice(&idxs[..]);
+            codec.select_batch(&mut output);
         });
     }
 

columnar/src/column_index/serialize.rs

@@ -5,23 +5,26 @@ use common::{CountingWriter, OwnedBytes};
 
 use crate::column_index::multivalued_index::serialize_multivalued_index;
 use crate::column_index::optional_index::serialize_optional_index;
-use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
-use crate::column_values::ColumnValues;
+use crate::column_index::ColumnIndex;
+use crate::iterable::Iterable;
 use crate::{Cardinality, RowId};
 
 pub enum SerializableColumnIndex<'a> {
     Full,
-    Optional(Box<dyn SerializableOptionalIndex<'a> + 'a>),
+    Optional {
+        non_null_row_ids: Box<dyn Iterable<RowId> + 'a>,
+        num_rows: RowId,
+    },
     // TODO remove the Arc<dyn> apart from serialization this is not
     // dynamic at all.
-    Multivalued(Box<dyn ColumnValues<RowId> + 'a>),
+    Multivalued(Box<dyn Iterable<RowId> + 'a>),
 }
 
 impl<'a> SerializableColumnIndex<'a> {
     pub fn get_cardinality(&self) -> Cardinality {
         match self {
             SerializableColumnIndex::Full => Cardinality::Full,
-            SerializableColumnIndex::Optional(_) => Cardinality::Optional,
+            SerializableColumnIndex::Optional { .. } => Cardinality::Optional,
             SerializableColumnIndex::Multivalued(_) => Cardinality::Multivalued,
         }
     }
@@ -36,9 +39,10 @@ pub fn serialize_column_index(
     output.write_all(&[cardinality])?;
     match column_index {
         SerializableColumnIndex::Full => {}
-        SerializableColumnIndex::Optional(optional_index) => {
-            serialize_optional_index(&*optional_index, &mut output)?
-        }
+        SerializableColumnIndex::Optional {
+            non_null_row_ids,
+            num_rows,
+        } => serialize_optional_index(non_null_row_ids.as_ref(), num_rows, &mut output)?,
         SerializableColumnIndex::Multivalued(multivalued_index) => {
             serialize_multivalued_index(&*multivalued_index, &mut output)?
         }
@@ -47,7 +51,7 @@ pub fn serialize_column_index(
     Ok(column_index_num_bytes)
 }
 
-pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex<'static>> {
+pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex> {
     if bytes.is_empty() {
         return Err(io::Error::new(
             io::ErrorKind::UnexpectedEof,
@@ -64,8 +68,8 @@ pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex<'stati
             Ok(ColumnIndex::Optional(optional_index))
         }
         Cardinality::Multivalued => {
-            let multivalued_index = super::multivalued_index::open_multivalued_index(bytes)?;
-            Ok(ColumnIndex::Multivalued(multivalued_index))
+            let multivalue_index = super::multivalued_index::open_multivalued_index(bytes)?;
+            Ok(ColumnIndex::Multivalued(multivalue_index))
         }
     }
 }

columnar/src/column_values/bitpacked.rs

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

columnar/src/column_values/blockwise_linear.rs

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

columnar/src/column_values/column.rs

@@ -1,5 +1,7 @@
+use std::fmt::Debug;
 use std::marker::PhantomData;
 use std::ops::{Range, RangeInclusive};
+use std::sync::Arc;
 
 use tantivy_bitpacker::minmax;
 
@@ -26,7 +28,7 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     ///
     /// Must panic if `start + output.len()` is greater than
     /// the segment's `maxdoc`.
-    #[inline]
+    #[inline(always)]
     fn get_range(&self, start: u64, output: &mut [T]) {
         for (out, idx) in output.iter_mut().zip(start..) {
             *out = self.get_val(idx as u32);
@@ -36,7 +38,7 @@ 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]
+    #[inline(always)]
     fn get_docids_for_value_range(
         &self,
         value_range: RangeInclusive<T>,
@@ -44,7 +46,6 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
         positions: &mut Vec<u32>,
     ) {
         let doc_id_range = doc_id_range.start..doc_id_range.end.min(self.num_vals());
-
         for idx in doc_id_range.start..doc_id_range.end {
             let val = self.get_val(idx);
             if value_range.contains(&val) {
@@ -78,33 +79,39 @@ pub trait ColumnValues<T: PartialOrd = u64>: Send + Sync {
     }
 }
 
-impl<T: Copy + PartialOrd> ColumnValues<T> for std::sync::Arc<dyn ColumnValues<T>> {
+impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for Arc<dyn ColumnValues<T>> {
+    #[inline(always)]
     fn get_val(&self, idx: u32) -> T {
         self.as_ref().get_val(idx)
     }
 
+    #[inline(always)]
     fn min_value(&self) -> T {
         self.as_ref().min_value()
     }
 
+    #[inline(always)]
     fn max_value(&self) -> T {
         self.as_ref().max_value()
     }
 
+    #[inline(always)]
     fn num_vals(&self) -> u32 {
         self.as_ref().num_vals()
     }
 
+    #[inline(always)]
     fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
         self.as_ref().iter()
     }
 
+    #[inline(always)]
     fn get_range(&self, start: u64, output: &mut [T]) {
         self.as_ref().get_range(start, output)
     }
 }
 
-impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd> ColumnValues<T> for &'a C {
+impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd + Debug> ColumnValues<T> for &'a C {
     fn get_val(&self, idx: u32) -> T {
         (*self).get_val(idx)
     }
@@ -137,7 +144,7 @@ pub struct VecColumn<'a, T = u64> {
     pub(crate) max_value: T,
 }
 
-impl<'a, T: Copy + PartialOrd + Send + Sync> ColumnValues<T> for VecColumn<'a, T> {
+impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> ColumnValues<T> for VecColumn<'a, T> {
     fn get_val(&self, position: u32) -> T {
         self.values[position as usize]
     }
@@ -205,8 +212,8 @@ pub fn monotonic_map_column<C, T, Input, Output>(
 where
     C: ColumnValues<Input>,
     T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Clone,
-    Output: PartialOrd + Send + Sync + Clone,
+    Input: PartialOrd + Debug + Send + Sync + Clone,
+    Output: PartialOrd + Debug + Send + Sync + Clone,
 {
     MonotonicMappingColumn {
         from_column,
@@ -219,8 +226,8 @@ impl<C, T, Input, Output> ColumnValues<Output> for MonotonicMappingColumn<C, T,
 where
     C: ColumnValues<Input>,
     T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Clone,
-    Output: PartialOrd + Send + Sync + Clone,
+    Input: PartialOrd + Send + Debug + Sync + Clone,
+    Output: PartialOrd + Send + Debug + Sync + Clone,
 {
     #[inline]
     fn get_val(&self, idx: u32) -> Output {
@@ -282,7 +289,7 @@ where T: Iterator + Clone + ExactSizeIterator
 impl<T> ColumnValues<T::Item> for IterColumn<T>
 where
     T: Iterator + Clone + ExactSizeIterator + Send + Sync,
-    T::Item: PartialOrd,
+    T::Item: PartialOrd + Debug,
 {
     fn get_val(&self, idx: u32) -> T::Item {
         self.0.clone().nth(idx as usize).unwrap()

columnar/src/column_values/column_with_cardinality.rs

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

columnar/src/column_values/compact_space/mod.rs

@@ -22,6 +22,7 @@ use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
 
 use crate::column_values::compact_space::build_compact_space::get_compact_space;
 use crate::column_values::ColumnValues;
+use crate::RowId;
 
 mod blank_range;
 mod build_compact_space;
@@ -55,7 +56,7 @@ impl RangeMapping {
 }
 
 impl BinarySerializable for CompactSpace {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
 
         let mut prev_value = 0;
@@ -158,23 +159,30 @@ impl CompactSpace {
 pub struct CompactSpaceCompressor {
     params: IPCodecParams,
 }
+
 #[derive(Debug, Clone)]
 pub struct IPCodecParams {
     compact_space: CompactSpace,
     bit_unpacker: BitUnpacker,
     min_value: u128,
     max_value: u128,
-    num_vals: u32,
+    num_vals: RowId,
     num_bits: u8,
 }
 
 impl CompactSpaceCompressor {
-    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
-    pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
-        let mut values_sorted = BTreeSet::new();
-        values_sorted.extend(iter);
-        let total_num_values = num_vals;
+    pub fn num_vals(&self) -> RowId {
+        self.params.num_vals
+    }
 
+    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
+    pub fn train_from(iter: impl Iterator<Item = u128>) -> Self {
+        let mut values_sorted = BTreeSet::new();
+        let mut total_num_values = 0u32;
+        for val in iter {
+            total_num_values += 1u32;
+            values_sorted.insert(val);
+        }
         let compact_space =
             get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
         let amplitude_compact_space = compact_space.amplitude_compact_space();
@@ -247,7 +255,7 @@ pub struct CompactSpaceDecompressor {
 }
 
 impl BinarySerializable for IPCodecParams {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         // header flags for future optional dictionary encoding
         let footer_flags = 0u64;
         footer_flags.serialize(writer)?;
@@ -450,364 +458,352 @@ impl CompactSpaceDecompressor {
     }
 }
 
-// TODO reenable what can be reenabled.
-// #[cfg(test)]
-// mod tests {
-//
-// use super::*;
-// use crate::column::format_version::read_format_version;
-// use crate::column::column_footer::read_null_index_footer;
-// use crate::column::serialize::U128Header;
-// use crate::column::{open_u128, serialize_u128};
-//
-// #[test]
-// fn compact_space_test() {
-// let ips = &[
-// 2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
-// ]
-// .into_iter()
-// .collect();
-// let compact_space = get_compact_space(ips, ips.len() as u32, 11);
-// let amplitude = compact_space.amplitude_compact_space();
-// assert_eq!(amplitude, 17);
-// assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
-// assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
-// assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
-//
-// for (num1, num2) in (0..3).tuple_windows() {
-// assert_eq!(
-// compact_space.get_range_mapping(num1).compact_end() + 1,
-// compact_space.get_range_mapping(num2).compact_start
-// );
-// }
-//
-// let mut output: Vec<u8> = Vec::new();
-// compact_space.serialize(&mut output).unwrap();
-//
-// assert_eq!(
-// compact_space,
-// CompactSpace::deserialize(&mut &output[..]).unwrap()
-// );
-//
-// for ip in ips {
-// let compact = compact_space.u128_to_compact(*ip).unwrap();
-// assert_eq!(compact_space.compact_to_u128(compact), *ip);
-// }
-// }
-//
-// #[test]
-// fn compact_space_amplitude_test() {
-// let ips = &[100000u128, 1000000].into_iter().collect();
-// let compact_space = get_compact_space(ips, ips.len() as u32, 1);
-// let amplitude = compact_space.amplitude_compact_space();
-// assert_eq!(amplitude, 2);
-// }
-//
-// fn test_all(mut data: OwnedBytes, expected: &[u128]) {
-// let _header = U128Header::deserialize(&mut data);
-// let decompressor = CompactSpaceDecompressor::open(data).unwrap();
-// for (idx, expected_val) in expected.iter().cloned().enumerate() {
-// let val = decompressor.get(idx as u32);
-// assert_eq!(val, expected_val);
-//
-// let test_range = |range: RangeInclusive<u128>| {
-// let expected_positions = expected
-// .iter()
-// .positions(|val| range.contains(val))
-// .map(|pos| pos as u32)
-// .collect::<Vec<_>>();
-// let mut positions = Vec::new();
-// decompressor.get_positions_for_value_range(
-// range,
-// 0..decompressor.num_vals(),
-// &mut positions,
-// );
-// assert_eq!(positions, expected_positions);
-// };
-//
-// test_range(expected_val.saturating_sub(1)..=expected_val);
-// test_range(expected_val..=expected_val);
-// test_range(expected_val..=expected_val.saturating_add(1));
-// test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
-// }
-// }
-//
-// fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
-// let mut out = Vec::new();
-// serialize_u128(
-// || u128_vals.iter().cloned(),
-// u128_vals.len() as u32,
-// &mut out,
-// )
-// .unwrap();
-//
-// let data = OwnedBytes::new(out);
-// let (data, _format_version) = read_format_version(data).unwrap();
-// let (data, _null_index_footer) = read_null_index_footer(data).unwrap();
-// test_all(data.clone(), u128_vals);
-//
-// data
-// }
-//
-// #[test]
-// fn test_range_1() {
-// let vals = &[
-// 1u128,
-// 100u128,
-// 3u128,
-// 99999u128,
-// 100000u128,
-// 100001u128,
-// 4_000_211_221u128,
-// 4_000_211_222u128,
-// 333u128,
-// ];
-// let mut data = test_aux_vals(vals);
-//
-// let _header = U128Header::deserialize(&mut data);
-// let decomp = CompactSpaceDecompressor::open(data).unwrap();
-// let complete_range = 0..vals.len() as u32;
-// for (pos, val) in vals.iter().enumerate() {
-// let val = *val;
-// let pos = pos as u32;
-// let mut positions = Vec::new();
-// decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
-// assert_eq!(positions, vec![pos]);
-// }
-//
-// handle docid range out of bounds
-// let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
-// assert!(positions.is_empty());
-//
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
-// assert_eq!(positions, vec![0]);
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
-// assert_eq!(positions, vec![0]);
-// let positions =
-// get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
-// assert_eq!(positions, vec![0, 2]);
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99999u128..=99999u128,
-// complete_range.clone()
-// ),
-// vec![3]
-// );
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99999u128..=100000u128,
-// complete_range.clone()
-// ),
-// vec![3, 4]
-// );
-// assert_eq!(
-// get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=100000u128,
-// complete_range.clone()
-// ),
-// vec![3, 4]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=99999u128,
-// complete_range.clone()
-// ),
-// &[3]
-// );
-// assert!(get_positions_for_value_range_helper(
-// &decomp,
-// 99998u128..=99998u128,
-// complete_range.clone()
-// )
-// .is_empty());
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 333u128..=333u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 332u128..=333u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 332u128..=334u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 333u128..=334u128,
-// complete_range.clone()
-// ),
-// &[8]
-// );
-//
-// assert_eq!(
-// &get_positions_for_value_range_helper(
-// &decomp,
-// 4_000_211_221u128..=5_000_000_000u128,
-// complete_range
-// ),
-// &[6, 7]
-// );
-// }
-//
-// #[test]
-// fn test_empty() {
-// let vals = &[];
-// let data = test_aux_vals(vals);
-// let _decomp = CompactSpaceDecompressor::open(data).unwrap();
-// }
-//
-// #[test]
-// fn test_range_2() {
-// let vals = &[
-// 100u128,
-// 99999u128,
-// 100000u128,
-// 100001u128,
-// 4_000_211_221u128,
-// 4_000_211_222u128,
-// 333u128,
-// ];
-// let mut data = test_aux_vals(vals);
-// let _header = U128Header::deserialize(&mut data);
-// let decomp = CompactSpaceDecompressor::open(data).unwrap();
-// let complete_range = 0..vals.len() as u32;
-// assert!(
-// &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
-// .is_empty(),
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
-// &[0]
-// );
-// assert_eq!(
-// &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
-// &[0]
-// );
-// }
-//
-// fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd>(
-// column: &C,
-// value_range: RangeInclusive<T>,
-// doc_id_range: Range<u32>,
-// ) -> Vec<u32> {
-// let mut positions = Vec::new();
-// column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
-// positions
-// }
-//
-// #[test]
-// fn test_range_3() {
-// let vals = &[
-// 200u128,
-// 201,
-// 202,
-// 203,
-// 204,
-// 204,
-// 206,
-// 207,
-// 208,
-// 209,
-// 210,
-// 1_000_000,
-// 5_000_000_000,
-// ];
-// let mut out = Vec::new();
-// serialize_u128(|| vals.iter().cloned(), vals.len() as u32, &mut out).unwrap();
-// let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
-// let complete_range = 0..vals.len() as u32;
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
-// vec![0]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
-// vec![0, 1]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
-// vec![0]
-// );
-//
-// assert_eq!(
-// get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
-// vec![11]
-// );
-// }
-//
-// #[test]
-// fn test_bug1() {
-// let vals = &[9223372036854775806];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug2() {
-// let vals = &[340282366920938463463374607431768211455u128];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug3() {
-// let vals = &[340282366920938463463374607431768211454];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_bug4() {
-// let vals = &[340282366920938463463374607431768211455, 0];
-// let _data = test_aux_vals(vals);
-// }
-//
-// #[test]
-// fn test_first_large_gaps() {
-// let vals = &[1_000_000_000u128; 100];
-// let _data = test_aux_vals(vals);
-// }
-// use itertools::Itertools;
-// use proptest::prelude::*;
-//
-// fn num_strategy() -> impl Strategy<Value = u128> {
-// prop_oneof![
-// 1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
-// 1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
-// 20 => prop::num::u128::ANY,
-// ]
-// }
-//
-// proptest! {
-// #![proptest_config(ProptestConfig::with_cases(10))]
-//
-// #[test]
-// fn compress_decompress_random(vals in proptest::collection::vec(num_strategy()
-// , 1..1000)) {
-// let _data = test_aux_vals(&vals);
-// }
-// }
-// }
-//
+#[cfg(test)]
+mod tests {
+
+    use itertools::Itertools;
+
+    use super::*;
+    use crate::column_values::serialize::U128Header;
+    use crate::column_values::{open_u128_mapped, serialize_column_values_u128};
+
+    #[test]
+    fn compact_space_test() {
+        let ips = &[
+            2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
+        ]
+        .into_iter()
+        .collect();
+        let compact_space = get_compact_space(ips, ips.len() as u32, 11);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 17);
+        assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
+        assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
+        assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
+
+        for (num1, num2) in (0..3).tuple_windows() {
+            assert_eq!(
+                compact_space.get_range_mapping(num1).compact_end() + 1,
+                compact_space.get_range_mapping(num2).compact_start
+            );
+        }
+
+        let mut output: Vec<u8> = Vec::new();
+        compact_space.serialize(&mut output).unwrap();
+
+        assert_eq!(
+            compact_space,
+            CompactSpace::deserialize(&mut &output[..]).unwrap()
+        );
+
+        for ip in ips {
+            let compact = compact_space.u128_to_compact(*ip).unwrap();
+            assert_eq!(compact_space.compact_to_u128(compact), *ip);
+        }
+    }
+
+    #[test]
+    fn compact_space_amplitude_test() {
+        let ips = &[100000u128, 1000000].into_iter().collect();
+        let compact_space = get_compact_space(ips, ips.len() as u32, 1);
+        let amplitude = compact_space.amplitude_compact_space();
+        assert_eq!(amplitude, 2);
+    }
+
+    fn test_all(mut data: OwnedBytes, expected: &[u128]) {
+        let _header = U128Header::deserialize(&mut data);
+        let decompressor = CompactSpaceDecompressor::open(data).unwrap();
+        for (idx, expected_val) in expected.iter().cloned().enumerate() {
+            let val = decompressor.get(idx as u32);
+            assert_eq!(val, expected_val);
+
+            let test_range = |range: RangeInclusive<u128>| {
+                let expected_positions = expected
+                    .iter()
+                    .positions(|val| range.contains(val))
+                    .map(|pos| pos as u32)
+                    .collect::<Vec<_>>();
+                let mut positions = Vec::new();
+                decompressor.get_positions_for_value_range(
+                    range,
+                    0..decompressor.num_vals(),
+                    &mut positions,
+                );
+                assert_eq!(positions, expected_positions);
+            };
+
+            test_range(expected_val.saturating_sub(1)..=expected_val);
+            test_range(expected_val..=expected_val);
+            test_range(expected_val..=expected_val.saturating_add(1));
+            test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
+        }
+    }
+
+    fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
+        let mut out = Vec::new();
+        serialize_column_values_u128(&u128_vals, &mut out).unwrap();
+        let data = OwnedBytes::new(out);
+        test_all(data.clone(), u128_vals);
+        data
+    }
+
+    #[test]
+    fn test_range_1() {
+        let vals = &[
+            1u128,
+            100u128,
+            3u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        for (pos, val) in vals.iter().enumerate() {
+            let val = *val;
+            let pos = pos as u32;
+            let mut positions = Vec::new();
+            decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
+            assert_eq!(positions, vec![pos]);
+        }
+
+        // handle docid range out of bounds
+        let positions: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
+        assert!(positions.is_empty());
+
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
+        assert_eq!(positions, vec![0]);
+        let positions =
+            get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
+        assert_eq!(positions, vec![0, 2]);
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=99999u128,
+                complete_range.clone()
+            ),
+            vec![3]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99999u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=100000u128,
+                complete_range.clone()
+            ),
+            vec![3, 4]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                99998u128..=99999u128,
+                complete_range.clone()
+            ),
+            &[3]
+        );
+        assert!(get_positions_for_value_range_helper(
+            &decomp,
+            99998u128..=99998u128,
+            complete_range.clone()
+        )
+        .is_empty());
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=333u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                332u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                333u128..=334u128,
+                complete_range.clone()
+            ),
+            &[8]
+        );
+
+        assert_eq!(
+            &get_positions_for_value_range_helper(
+                &decomp,
+                4_000_211_221u128..=5_000_000_000u128,
+                complete_range
+            ),
+            &[6, 7]
+        );
+    }
+
+    #[test]
+    fn test_empty() {
+        let vals = &[];
+        let data = test_aux_vals(vals);
+        let _decomp = CompactSpaceDecompressor::open(data).unwrap();
+    }
+
+    #[test]
+    fn test_range_2() {
+        let vals = &[
+            100u128,
+            99999u128,
+            100000u128,
+            100001u128,
+            4_000_211_221u128,
+            4_000_211_222u128,
+            333u128,
+        ];
+        let mut data = test_aux_vals(vals);
+        let _header = U128Header::deserialize(&mut data);
+        let decomp = CompactSpaceDecompressor::open(data).unwrap();
+        let complete_range = 0..vals.len() as u32;
+        assert!(
+            &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
+                .is_empty(),
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
+            &[0]
+        );
+        assert_eq!(
+            &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
+            &[0]
+        );
+    }
+
+    fn get_positions_for_value_range_helper<C: ColumnValues<T> + ?Sized, T: PartialOrd>(
+        column: &C,
+        value_range: RangeInclusive<T>,
+        doc_id_range: Range<u32>,
+    ) -> Vec<u32> {
+        let mut positions = Vec::new();
+        column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
+        positions
+    }
+
+    #[test]
+    fn test_range_3() {
+        let vals = &[
+            200u128,
+            201,
+            202,
+            203,
+            204,
+            204,
+            206,
+            207,
+            208,
+            209,
+            210,
+            1_000_000,
+            5_000_000_000,
+        ];
+        let mut out = Vec::new();
+        serialize_column_values_u128(&&vals[..], &mut out).unwrap();
+        let decomp = open_u128_mapped(OwnedBytes::new(out)).unwrap();
+        let complete_range = 0..vals.len() as u32;
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
+            vec![0, 1]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
+            vec![0]
+        );
+
+        assert_eq!(
+            get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
+            vec![11]
+        );
+    }
+
+    #[test]
+    fn test_bug1() {
+        let vals = &[9223372036854775806];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug2() {
+        let vals = &[340282366920938463463374607431768211455u128];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug3() {
+        let vals = &[340282366920938463463374607431768211454];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_bug4() {
+        let vals = &[340282366920938463463374607431768211455, 0];
+        let _data = test_aux_vals(vals);
+    }
+
+    #[test]
+    fn test_first_large_gaps() {
+        let vals = &[1_000_000_000u128; 100];
+        let _data = test_aux_vals(vals);
+    }
+
+    use proptest::prelude::*;
+
+    fn num_strategy() -> impl Strategy<Value = u128> {
+        prop_oneof![
+            1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
+            1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
+            20 => prop::num::u128::ANY,
+        ]
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig::with_cases(10))]
+
+        #[test]
+        fn compress_decompress_random(vals in proptest::collection::vec(num_strategy() , 1..1000)) {
+            let _data = test_aux_vals(&vals);
+        }
+    }
+}

columnar/src/column_values/gcd.rs

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

columnar/src/column_values/linear.rs

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

columnar/src/column_values/main.rs

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

columnar/src/column_values/mod.rs

@@ -7,83 +7,74 @@
 //! - 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;
 use std::sync::Arc;
 
 use common::{BinarySerializable, OwnedBytes};
 use compact_space::CompactSpaceDecompressor;
-use monotonic_mapping::{
-    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
-    StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
-};
-use serialize::{Header, U128Header};
+pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
+use monotonic_mapping::{StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal};
+pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
+use serialize::U128Header;
 
-mod bitpacked;
-mod blockwise_linear;
 mod compact_space;
-mod line;
-mod linear;
 pub(crate) mod monotonic_mapping;
 pub(crate) mod monotonic_mapping_u128;
+mod stats;
+pub(crate) mod u64_based;
 
 mod column;
-mod column_with_cardinality;
-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};
-pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
-pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
-#[cfg(test)]
-pub use self::serialize::tests::serialize_and_load;
-pub use self::serialize::{serialize_column_values, NormalizedHeader};
-use crate::column_values::bitpacked::BitpackedCodec;
-use crate::column_values::blockwise_linear::BlockwiseLinearCodec;
-use crate::column_values::linear::LinearCodec;
+use crate::iterable::Iterable;
+use crate::{ColumnIndex, MergeRowOrder};
 
-#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
-#[repr(u8)]
-/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
-pub enum FastFieldCodecType {
-    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
-    /// `column.max_value() - column.min_value()`
-    Bitpacked = 1,
-    /// Linear interpolation puts a line between the first and last value and then bitpacks the
-    /// values by the offset from the line. The number of bits is defined by the max deviation from
-    /// the line.
-    Linear = 2,
-    /// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
-    BlockwiseLinear = 3,
+pub(crate) struct MergedColumnValues<'a, T> {
+    pub(crate) column_indexes: &'a [Option<ColumnIndex>],
+    pub(crate) column_values: &'a [Option<Arc<dyn ColumnValues<T>>>],
+    pub(crate) merge_row_order: &'a MergeRowOrder,
 }
 
-impl BinarySerializable for FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
-        self.to_code().serialize(wrt)
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let code = u8::deserialize(reader)?;
-        let codec_type: Self = Self::from_code(code)
-            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
-        Ok(codec_type)
-    }
-}
-
-impl FastFieldCodecType {
-    pub(crate) fn to_code(self) -> u8 {
-        self as u8
-    }
-
-    pub(crate) fn from_code(code: u8) -> Option<Self> {
-        match code {
-            1 => Some(Self::Bitpacked),
-            2 => Some(Self::Linear),
-            3 => Some(Self::BlockwiseLinear),
-            _ => None,
+impl<'a, T: Copy + PartialOrd + Debug> Iterable<T> for MergedColumnValues<'a, T> {
+    fn boxed_iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
+        match self.merge_row_order {
+            MergeRowOrder::Stack(_) => {
+                Box::new(self
+                    .column_values
+                    .iter()
+                    .flatten()
+                    .flat_map(|column_value| column_value.iter()))
+            },
+            MergeRowOrder::Shuffled(shuffle_merge_order) => {
+                Box::new(shuffle_merge_order
+                    .iter_new_to_old_row_addrs()
+                    .flat_map(|row_addr| {
+                        let Some(column_index) = self.column_indexes[row_addr.segment_ord as usize].as_ref() else {
+                            return None;
+                        };
+                        let Some(column_values) = self.column_values[row_addr.segment_ord as usize].as_ref() else {
+                            return None;
+                        };
+                        let value_range = column_index.value_row_ids(row_addr.row_id);
+                        Some((value_range, column_values))
+                    })
+                    .flat_map(|(value_range, column_values)| {
+                        value_range
+                            .into_iter()
+                            .map(|val| column_values.get_val(val))
+                    })
+                )
+            },
         }
     }
 }
@@ -98,7 +89,7 @@ pub enum U128FastFieldCodecType {
 }
 
 impl BinarySerializable for U128FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
+    fn serialize<W: Write + ?Sized>(&self, wrt: &mut W) -> io::Result<()> {
         self.to_code().serialize(wrt)
     }
 
@@ -124,7 +115,7 @@ impl U128FastFieldCodecType {
 }
 
 /// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open_u128_mapped<T: MonotonicallyMappableToU128>(
+pub fn open_u128_mapped<T: MonotonicallyMappableToU128 + Debug>(
     mut bytes: OwnedBytes,
 ) -> io::Result<Arc<dyn ColumnValues<T>>> {
     let header = U128Header::deserialize(&mut bytes)?;
@@ -136,68 +127,6 @@ pub fn open_u128_mapped<T: MonotonicallyMappableToU128>(
     Ok(Arc::new(monotonic_map_column(reader, inverted)))
 }
 
-/// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open_u64_mapped<T: MonotonicallyMappableToU64>(
-    mut bytes: OwnedBytes,
-) -> io::Result<Arc<dyn ColumnValues<T>>> {
-    let header = Header::deserialize(&mut bytes)?;
-    match header.codec_type {
-        FastFieldCodecType::Bitpacked => open_specific_codec::<BitpackedCodec, _>(bytes, &header),
-        FastFieldCodecType::Linear => open_specific_codec::<LinearCodec, _>(bytes, &header),
-        FastFieldCodecType::BlockwiseLinear => {
-            open_specific_codec::<BlockwiseLinearCodec, _>(bytes, &header)
-        }
-    }
-}
-
-fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
-    bytes: OwnedBytes,
-    header: &Header,
-) -> io::Result<Arc<dyn ColumnValues<Item>>> {
-    let normalized_header = header.normalized();
-    let reader = C::open_from_bytes(bytes, normalized_header)?;
-    let min_value = header.min_value;
-    if let Some(gcd) = header.gcd {
-        let mapping = StrictlyMonotonicMappingInverter::from(
-            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
-        );
-        Ok(Arc::new(monotonic_map_column(reader, mapping)))
-    } else {
-        let mapping = StrictlyMonotonicMappingInverter::from(
-            StrictlyMonotonicMappingToInternalBaseval::new(min_value),
-        );
-        Ok(Arc::new(monotonic_map_column(reader, mapping)))
-    }
-}
-
-/// The FastFieldSerializerEstimate trait is required on all variants
-/// of fast field compressions, to decide which one to choose.
-pub(crate) trait FastFieldCodec: 'static {
-    /// A codex needs to provide a unique name and id, which is
-    /// used for debugging and de/serialization.
-    const CODEC_TYPE: FastFieldCodecType;
-
-    type Reader: ColumnValues<u64> + 'static;
-
-    /// Reads the metadata and returns the CodecReader
-    fn open_from_bytes(bytes: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader>;
-
-    /// Serializes the data using the serializer into write.
-    ///
-    /// The column iterator should be preferred over using column `get_val` method for
-    /// performance reasons.
-    fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()>;
-
-    /// Returns an estimate of the compression ratio.
-    /// If the codec is not applicable, returns `None`.
-    ///
-    /// The baseline is uncompressed 64bit data.
-    ///
-    /// It could make sense to also return a value representing
-    /// computational complexity.
-    fn estimate(column: &dyn ColumnValues) -> Option<f32>;
-}
-
 #[cfg(all(test, feature = "unstable"))]
 mod bench {
     use std::sync::Arc;

columnar/src/column_values/monotonic_mapping.rs

@@ -1,3 +1,4 @@
+use std::fmt::Debug;
 use std::marker::PhantomData;
 
 use fastdivide::DividerU64;
@@ -7,7 +8,7 @@ use crate::RowId;
 
 /// Monotonic maps a value to u64 value space.
 /// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
-pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
+pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Debug + Copy + Send + Sync {
     /// Converts a value to u64.
     ///
     /// Internally all fast field values are encoded as u64.

columnar/src/column_values/monotonic_mapping_u128.rs

@@ -1,8 +1,9 @@
+use std::fmt::Debug;
 use std::net::Ipv6Addr;
 
 /// Montonic maps a value to u128 value space
 /// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
-pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Send + Sync {
+pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Debug + Send + Sync {
     /// Converts a value to u128.
     ///
     /// Internally all fast field values are encoded as u64.

columnar/src/column_values/serialize.rs

@@ -1,40 +1,12 @@
-// Copyright (C) 2022 Quickwit, Inc.
-//
-// Quickwit is offered under the AGPL v3.0 and as commercial software.
-// For commercial licensing, contact us at hello@quickwit.io.
-//
-// AGPL:
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU Affero General Public License as
-// published by the Free Software Foundation, either version 3 of the
-// License, or (at your option) any later version.
-//
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-// GNU Affero General Public License for more details.
-//
-// You should have received a copy of the GNU Affero General Public License
-// along with this program. If not, see <http://www.gnu.org/licenses/>.
-
+use std::fmt::Debug;
 use std::io;
-use std::num::NonZeroU64;
 
 use common::{BinarySerializable, VInt};
-use log::warn;
 
-use super::bitpacked::BitpackedCodec;
-use super::blockwise_linear::BlockwiseLinearCodec;
-use super::linear::LinearCodec;
-use super::monotonic_mapping::{
-    StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
-    StrictlyMonotonicMappingToInternalGCDBaseval,
-};
-use super::{
-    monotonic_map_column, ColumnValues, FastFieldCodec, FastFieldCodecType,
-    MonotonicallyMappableToU64, U128FastFieldCodecType,
-};
 use crate::column_values::compact_space::CompactSpaceCompressor;
+use crate::column_values::U128FastFieldCodecType;
+use crate::iterable::Iterable;
+use crate::MonotonicallyMappableToU128;
 
 /// The normalized header gives some parameters after applying the following
 /// normalization of the vector:
@@ -49,53 +21,6 @@ pub struct NormalizedHeader {
     pub max_value: u64,
 }
 
-#[derive(Debug, Copy, Clone)]
-pub(crate) struct Header {
-    pub num_vals: u32,
-    pub min_value: u64,
-    pub max_value: u64,
-    pub gcd: Option<NonZeroU64>,
-    pub codec_type: FastFieldCodecType,
-}
-
-impl Header {
-    pub fn normalized(self) -> NormalizedHeader {
-        let gcd = self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
-        let gcd_min_val_mapping =
-            StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, self.min_value);
-
-        let max_value = gcd_min_val_mapping.mapping(self.max_value);
-        NormalizedHeader {
-            num_vals: self.num_vals,
-            max_value,
-        }
-    }
-
-    pub(crate) fn normalize_column<C: ColumnValues>(&self, from_column: C) -> impl ColumnValues {
-        normalize_column(from_column, self.min_value, self.gcd)
-    }
-
-    pub fn compute_header(
-        column: impl ColumnValues<u64>,
-        codecs: &[FastFieldCodecType],
-    ) -> Option<Header> {
-        let num_vals = column.num_vals();
-        let min_value = column.min_value();
-        let max_value = column.max_value();
-        let gcd = super::gcd::find_gcd(column.iter().map(|val| val - min_value))
-            .filter(|gcd| gcd.get() > 1u64);
-        let normalized_column = normalize_column(column, min_value, gcd);
-        let codec_type = detect_codec(normalized_column, codecs)?;
-        Some(Header {
-            num_vals,
-            min_value,
-            max_value,
-            gcd,
-            codec_type,
-        })
-    }
-}
-
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
 pub(crate) struct U128Header {
     pub num_vals: u32,
@@ -103,7 +28,7 @@ pub(crate) struct U128Header {
 }
 
 impl BinarySerializable for U128Header {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
+    fn serialize<W: io::Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
         VInt(self.num_vals as u64).serialize(writer)?;
         self.codec_type.serialize(writer)?;
         Ok(())
@@ -119,157 +44,39 @@ impl BinarySerializable for U128Header {
     }
 }
 
-fn normalize_column<C: ColumnValues>(
-    from_column: C,
-    min_value: u64,
-    gcd: Option<NonZeroU64>,
-) -> impl ColumnValues {
-    let gcd = gcd.map(|gcd| gcd.get()).unwrap_or(1);
-    let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, min_value);
-    monotonic_map_column(from_column, mapping)
-}
-
-impl BinarySerializable for Header {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
-        VInt(self.num_vals as u64).serialize(writer)?;
-        VInt(self.min_value).serialize(writer)?;
-        VInt(self.max_value - self.min_value).serialize(writer)?;
-        if let Some(gcd) = self.gcd {
-            VInt(gcd.get()).serialize(writer)?;
-        } else {
-            VInt(0u64).serialize(writer)?;
-        }
-        self.codec_type.serialize(writer)?;
-        Ok(())
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let num_vals = VInt::deserialize(reader)?.0 as u32;
-        let min_value = VInt::deserialize(reader)?.0;
-        let amplitude = VInt::deserialize(reader)?.0;
-        let max_value = min_value + amplitude;
-        let gcd_u64 = VInt::deserialize(reader)?.0;
-        let codec_type = FastFieldCodecType::deserialize(reader)?;
-        Ok(Header {
-            num_vals,
-            min_value,
-            max_value,
-            gcd: NonZeroU64::new(gcd_u64),
-            codec_type,
-        })
-    }
-}
-
 /// Serializes u128 values with the compact space codec.
-pub fn serialize_column_values_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
-    iter_gen: F,
-    num_vals: u32,
+pub fn serialize_column_values_u128<T: MonotonicallyMappableToU128>(
+    iterable: &dyn Iterable<T>,
     output: &mut impl io::Write,
 ) -> io::Result<()> {
+    let compressor = CompactSpaceCompressor::train_from(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+    );
     let header = U128Header {
-        num_vals,
+        num_vals: compressor.num_vals(),
         codec_type: U128FastFieldCodecType::CompactSpace,
     };
     header.serialize(output)?;
-    let compressor = CompactSpaceCompressor::train_from(iter_gen(), num_vals);
-    compressor.compress_into(iter_gen(), output)?;
-
-    Ok(())
-}
-
-/// Serializes the column with the codec with the best estimate on the data.
-pub fn serialize_column_values<T: MonotonicallyMappableToU64>(
-    typed_column: impl ColumnValues<T>,
-    codecs: &[FastFieldCodecType],
-    output: &mut impl io::Write,
-) -> io::Result<()> {
-    let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
-    let header = Header::compute_header(&column, codecs).ok_or_else(|| {
-        io::Error::new(
-            io::ErrorKind::InvalidInput,
-            format!(
-                "Data cannot be serialized with this list of codec. {:?}",
-                codecs
-            ),
-        )
-    })?;
-    header.serialize(output)?;
-    let normalized_column = header.normalize_column(column);
-    assert_eq!(normalized_column.min_value(), 0u64);
-    serialize_given_codec(normalized_column, header.codec_type, output)?;
-    Ok(())
-}
-
-fn detect_codec(
-    column: impl ColumnValues<u64>,
-    codecs: &[FastFieldCodecType],
-) -> Option<FastFieldCodecType> {
-    let mut estimations = Vec::new();
-    for &codec in codecs {
-        let estimation_opt = match codec {
-            FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&column),
-            FastFieldCodecType::Linear => LinearCodec::estimate(&column),
-            FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&column),
-        };
-        if let Some(estimation) = estimation_opt {
-            estimations.push((estimation, codec));
-        }
-    }
-    if let Some(broken_estimation) = estimations.iter().find(|estimation| estimation.0.is_nan()) {
-        warn!(
-            "broken estimation for fast field codec {:?}",
-            broken_estimation.1
-        );
-    }
-    // removing nan values for codecs with broken calculations, and max values which disables
-    // codecs
-    estimations.retain(|estimation| !estimation.0.is_nan() && estimation.0 != f32::MAX);
-    estimations.sort_by(|(score_left, _), (score_right, _)| score_left.total_cmp(score_right));
-    Some(estimations.first()?.1)
-}
-
-pub(crate) fn serialize_given_codec(
-    column: impl ColumnValues<u64>,
-    codec_type: FastFieldCodecType,
-    output: &mut impl io::Write,
-) -> io::Result<()> {
-    match codec_type {
-        FastFieldCodecType::Bitpacked => {
-            BitpackedCodec::serialize(&column, output)?;
-        }
-        FastFieldCodecType::Linear => {
-            LinearCodec::serialize(&column, output)?;
-        }
-        FastFieldCodecType::BlockwiseLinear => {
-            BlockwiseLinearCodec::serialize(&column, output)?;
-        }
-    }
+    compressor.compress_into(
+        iterable
+            .boxed_iter()
+            .map(MonotonicallyMappableToU128::to_u128),
+        output,
+    )?;
     Ok(())
 }
 
 #[cfg(test)]
 pub mod tests {
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-
     use super::*;
-    use crate::column_values::{open_u64_mapped, VecColumn};
+    use crate::column_values::u64_based::{
+        serialize_and_load_u64_based_column_values, serialize_u64_based_column_values,
+        ALL_U64_CODEC_TYPES,
+    };
+    use crate::column_values::CodecType;
 
-    const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
-        FastFieldCodecType::Bitpacked,
-        FastFieldCodecType::Linear,
-        FastFieldCodecType::BlockwiseLinear,
-    ];
-
-    /// Helper function to serialize a column (autodetect from all codecs) and then open it
-    pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
-        column: &[T],
-    ) -> Arc<dyn ColumnValues<T>> {
-        let mut buffer = Vec::new();
-        serialize_column_values(&VecColumn::from(&column), &ALL_CODEC_TYPES, &mut buffer).unwrap();
-        open_u64_mapped(OwnedBytes::new(buffer)).unwrap()
-    }
     #[test]
     fn test_serialize_deserialize_u128_header() {
         let original = U128Header {
@@ -285,15 +92,22 @@ pub mod tests {
     #[test]
     fn test_serialize_deserialize() {
         let original = [1u64, 5u64, 10u64];
-        let restored: Vec<u64> = serialize_and_load(&original[..]).iter().collect();
+        let restored: Vec<u64> =
+            serialize_and_load_u64_based_column_values(&&original[..], &ALL_U64_CODEC_TYPES)
+                .iter()
+                .collect();
         assert_eq!(&restored, &original[..]);
     }
 
     #[test]
     fn test_fastfield_bool_size_bitwidth_1() {
         let mut buffer = Vec::new();
-        let col = VecColumn::from(&[false, true][..]);
-        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
         // TODO put the header as a footer so that it serves as a padding.
         // 5 bytes of header, 1 byte of value, 7 bytes of padding.
         assert_eq!(buffer.len(), 5 + 1);
@@ -302,19 +116,23 @@ pub mod tests {
     #[test]
     fn test_fastfield_bool_bit_size_bitwidth_0() {
         let mut buffer = Vec::new();
-        let col = VecColumn::from(&[true][..]);
-        serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
-        // 5 bytes of header, 0 bytes of value, 7 bytes of padding.
-        assert_eq!(buffer.len(), 5);
+        serialize_u64_based_column_values::<bool>(
+            &&[false, true][..],
+            &ALL_U64_CODEC_TYPES,
+            &mut buffer,
+        )
+        .unwrap();
+        // 6 bytes of header, 0 bytes of value, 7 bytes of padding.
+        assert_eq!(buffer.len(), 6);
     }
 
     #[test]
     fn test_fastfield_gcd() {
         let mut buffer = Vec::new();
         let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
-        let col = VecColumn::from(&vals[..]);
-        serialize_column_values(&col, &[FastFieldCodecType::Bitpacked], &mut buffer).unwrap();
+        serialize_u64_based_column_values(&&vals[..], &[CodecType::Bitpacked], &mut buffer)
+            .unwrap();
         // Values are stored over 3 bits.
-        assert_eq!(buffer.len(), 7 + (3 * 80 / 8));
+        assert_eq!(buffer.len(), 6 + (3 * 80 / 8));
     }
 }

columnar/src/column_values/stats.rs

@@ -0,0 +1,96 @@
+use std::io;
+use std::io::Write;
+use std::num::NonZeroU64;
+
+use common::{BinarySerializable, VInt};
+
+use crate::RowId;
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub struct Stats {
+    pub gcd: NonZeroU64,
+    pub min_value: u64,
+    pub max_value: u64,
+    pub num_rows: RowId,
+}
+
+impl Stats {
+    pub fn amplitude(&self) -> u64 {
+        self.max_value - self.min_value
+    }
+}
+
+impl BinarySerializable for Stats {
+    fn serialize<W: Write + ?Sized>(&self, writer: &mut W) -> io::Result<()> {
+        VInt(self.min_value).serialize(writer)?;
+        VInt(self.gcd.get()).serialize(writer)?;
+        VInt(self.amplitude() / self.gcd).serialize(writer)?;
+        VInt(self.num_rows as u64).serialize(writer)?;
+        Ok(())
+    }
+
+    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
+        let min_value = VInt::deserialize(reader)?.0;
+        let gcd = VInt::deserialize(reader)?.0;
+        let gcd = NonZeroU64::new(gcd)
+            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "GCD of 0 is forbidden"))?;
+        let amplitude = VInt::deserialize(reader)?.0 * gcd.get();
+        let max_value = min_value + amplitude;
+        let num_rows = VInt::deserialize(reader)?.0 as RowId;
+        Ok(Stats {
+            min_value,
+            max_value,
+            num_rows,
+            gcd,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU64;
+
+    use common::BinarySerializable;
+
+    use crate::column_values::Stats;
+
+    #[track_caller]
+    fn test_stats_ser_deser_aux(stats: &Stats, num_bytes: usize) {
+        let mut buffer: Vec<u8> = Vec::new();
+        stats.serialize(&mut buffer).unwrap();
+        assert_eq!(buffer.len(), num_bytes);
+        let deser_stats = Stats::deserialize(&mut &buffer[..]).unwrap();
+        assert_eq!(stats, &deser_stats);
+    }
+
+    #[test]
+    fn test_stats_serialization() {
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(3).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(1_000).unwrap(),
+                min_value: 1,
+                max_value: 3001,
+                num_rows: 10,
+            }),
+            5,
+        );
+        test_stats_ser_deser_aux(
+            &(Stats {
+                gcd: NonZeroU64::new(1).unwrap(),
+                min_value: 0,
+                max_value: 0,
+                num_rows: 0,
+            }),
+            4,
+        );
+    }
+}

columnar/src/column_values/u64_based/bitpacked.rs

@@ -0,0 +1,127 @@
+use std::io::{self, Write};
+
+use common::{BinarySerializable, OwnedBytes};
+use fastdivide::DividerU64;
+use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
+
+use crate::column_values::u64_based::{ColumnCodec, ColumnCodecEstimator, Stats};
+use crate::{ColumnValues, RowId};
+
+/// Depending on the field type, a different
+/// fast field is required.
+#[derive(Clone)]
+pub struct BitpackedReader {
+    data: OwnedBytes,
+    bit_unpacker: BitUnpacker,
+    stats: Stats,
+}
+
+impl ColumnValues for BitpackedReader {
+    #[inline(always)]
+    fn get_val(&self, doc: u32) -> u64 {
+        self.stats.min_value + self.stats.gcd.get() * self.bit_unpacker.get(doc, &self.data)
+    }
+
+    #[inline]
+    fn min_value(&self) -> u64 {
+        self.stats.min_value
+    }
+    #[inline]
+    fn max_value(&self) -> u64 {
+        self.stats.max_value
+    }
+    #[inline]
+    fn num_vals(&self) -> RowId {
+        self.stats.num_rows
+    }
+}
+
+fn num_bits(stats: &Stats) -> u8 {
+    compute_num_bits(stats.amplitude() / stats.gcd)
+}
+
+#[derive(Default)]
+pub struct BitpackedCodecEstimator;
+
+impl ColumnCodecEstimator for BitpackedCodecEstimator {
+    fn collect(&mut self, _value: u64) {}
+
+    fn estimate(&self, stats: &Stats) -> Option<u64> {
+        let num_bits_per_value = num_bits(stats);
+        Some(stats.num_bytes() + (stats.num_rows as u64 * (num_bits_per_value as u64) + 7) / 8)
+    }
+
+    fn serialize(
+        &self,
+        stats: &Stats,
+        vals: &mut dyn Iterator<Item = u64>,
+        wrt: &mut dyn Write,
+    ) -> io::Result<()> {
+        stats.serialize(wrt)?;
+        let num_bits = num_bits(stats);
+        let mut bit_packer = BitPacker::new();
+        let divider = DividerU64::divide_by(stats.gcd.get());
+        for val in vals {
+            bit_packer.write(divider.divide(val - stats.min_value), num_bits, wrt)?;
+        }
+        bit_packer.close(wrt)?;
+        Ok(())
+    }
+}
+
+pub struct BitpackedCodec;
+
+impl ColumnCodec for BitpackedCodec {
+    type Reader = BitpackedReader;
+    type Estimator = BitpackedCodecEstimator;
+
+    /// Opens a fast field given a file.
+    fn load(mut data: OwnedBytes) -> io::Result<Self::Reader> {
+        let stats = Stats::deserialize(&mut data)?;
+        let num_bits = num_bits(&stats);
+        let bit_unpacker = BitUnpacker::new(num_bits);
+        Ok(BitpackedReader {
+            data,
+            bit_unpacker,
+            stats,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::column_values::u64_based::tests::create_and_validate;
+
+    #[test]
+    fn test_with_codec_data_sets_simple() {
+        create_and_validate::<BitpackedCodec>(&[4, 3, 12], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets_simple_gcd() {
+        create_and_validate::<BitpackedCodec>(&[1000, 2000, 3000], "name");
+    }
+
+    #[test]
+    fn test_with_codec_data_sets() {
+        let data_sets = crate::column_values::u64_based::tests::get_codec_test_datasets();
+        for (mut data, name) in data_sets {
+            create_and_validate::<BitpackedCodec>(&data, name);
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, name);
+        }
+    }
+
+    #[test]
+    fn bitpacked_fast_field_rand() {
+        for _ in 0..500 {
+            let mut data = (0..1 + rand::random::<u8>() as usize)
+                .map(|_| rand::random::<i64>() as u64 / 2)
+                .collect::<Vec<_>>();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+            data.reverse();
+            create_and_validate::<BitpackedCodec>(&data, "rand");
+        }
+    }
+}

columnar/src/column_values/u64_based/blockwise_linear.rs

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

columnar/src/column_values/u64_based/line.rs

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

columnar/src/column_values/u64_based/linear.rs

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

columnar/src/column_values/u64_based/mod.rs

@@ -0,0 +1,182 @@
+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<'a, T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable<T>,
+    codec_types: &[CodecType],
+    wrt: &mut dyn Write,
+) -> io::Result<()> {
+    let mut stats_collector = StatsCollector::default();
+    let mut estimators: Vec<(CodecType, Box<dyn ColumnCodecEstimator>)> =
+        Vec::with_capacity(codec_types.len());
+    for &codec_type in codec_types {
+        estimators.push((codec_type, codec_type.estimator()));
+    }
+    for val in vals.boxed_iter() {
+        let val_u64 = val.to_u64();
+        stats_collector.collect(val_u64);
+        for (_, estimator) in &mut estimators {
+            estimator.collect(val_u64);
+        }
+    }
+    for (_, estimator) in &mut estimators {
+        estimator.finalize();
+    }
+    let stats = stats_collector.stats();
+    let (_, best_codec, best_codec_estimator) = estimators
+        .into_iter()
+        .flat_map(|(codec_type, estimator)| {
+            let num_bytes = estimator.estimate(&stats)?;
+            Some((num_bytes, codec_type, estimator))
+        })
+        .min_by_key(|(num_bytes, _, _)| *num_bytes)
+        .ok_or_else(|| {
+            io::Error::new(io::ErrorKind::InvalidData, "No available applicable codec.")
+        })?;
+    best_codec.to_code().serialize(wrt)?;
+    best_codec_estimator.serialize(
+        &stats,
+        &mut vals.boxed_iter().map(MonotonicallyMappableToU64::to_u64),
+        wrt,
+    )?;
+    Ok(())
+}
+
+pub fn load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    mut bytes: OwnedBytes,
+) -> io::Result<Arc<dyn ColumnValues<T>>> {
+    let codec_type: CodecType = bytes
+        .get(0)
+        .copied()
+        .and_then(CodecType::try_from_code)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Failed to read codec type"))?;
+    bytes.advance(1);
+    codec_type.load(bytes)
+}
+
+/// Helper function to serialize a column (autodetect from all codecs) and then open it
+pub fn serialize_and_load_u64_based_column_values<T: MonotonicallyMappableToU64>(
+    vals: &dyn Iterable,
+    codec_types: &[CodecType],
+) -> Arc<dyn ColumnValues<T>> {
+    let mut buffer = Vec::new();
+    serialize_u64_based_column_values(vals, codec_types, &mut buffer).unwrap();
+    load_u64_based_column_values::<T>(OwnedBytes::new(buffer)).unwrap()
+}
+
+#[cfg(test)]
+mod tests;

columnar/src/column_values/u64_based/stats_collector.rs

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

columnar/src/column_values/u64_based/tests.rs

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

columnar/src/columnar/column_type.rs

@@ -1,3 +1,4 @@
+use std::fmt::Debug;
 use std::net::Ipv6Addr;
 
 use crate::value::NumericalType;
@@ -66,7 +67,7 @@ impl ColumnType {
 }
 
 // TODO remove if possible
-pub trait HasAssociatedColumnType: 'static + Send + Sync + Copy + PartialOrd {
+pub trait HasAssociatedColumnType: 'static + Debug + Send + Sync + Copy + PartialOrd {
     fn column_type() -> ColumnType;
     fn default_value() -> Self;
 }

columnar/src/columnar/merge.rs

@@ -1,208 +0,0 @@
-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

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

columnar/src/columnar/merge/merge_mapping.rs

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

columnar/src/columnar/merge/mod.rs

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

columnar/src/columnar/merge/term_merger.rs

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

columnar/src/columnar/merge/tests.rs

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

columnar/src/columnar/merge_index.rs

@@ -0,0 +1 @@
+

columnar/src/columnar/mod.rs

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

columnar/src/columnar/reader/mod.rs

@@ -6,6 +6,7 @@ use sstable::{Dictionary, RangeSSTable};
 
 use crate::columnar::{format_version, ColumnType};
 use crate::dynamic_column::DynamicColumnHandle;
+use crate::RowId;
 
 fn io_invalid_data(msg: String) -> io::Error {
     io::Error::new(io::ErrorKind::InvalidData, msg)
@@ -13,9 +14,11 @@ fn io_invalid_data(msg: String) -> io::Error {
 
 /// The ColumnarReader makes it possible to access a set of columns
 /// associated to field names.
+#[derive(Clone)]
 pub struct ColumnarReader {
     column_dictionary: Dictionary<RangeSSTable>,
     column_data: FileSlice,
+    num_rows: RowId,
 }
 
 impl ColumnarReader {
@@ -27,23 +30,27 @@ impl ColumnarReader {
 
     fn open_inner(file_slice: FileSlice) -> io::Result<ColumnarReader> {
         let (file_slice_without_sstable_len, footer_slice) = file_slice
-            .split_from_end(mem::size_of::<u64>() + format_version::VERSION_FOOTER_NUM_BYTES);
+            .split_from_end(mem::size_of::<u64>() + 4 + format_version::VERSION_FOOTER_NUM_BYTES);
         let footer_bytes = footer_slice.read_bytes()?;
-        let (mut sstable_len_bytes, version_footer_bytes) =
-            footer_bytes.rsplit(format_version::VERSION_FOOTER_NUM_BYTES);
+        let sstable_len = u64::deserialize(&mut &footer_bytes[0..8])?;
+        let num_rows = u32::deserialize(&mut &footer_bytes[8..12])?;
         let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
-            version_footer_bytes.as_slice().try_into().unwrap();
+            footer_bytes[12..].try_into().unwrap();
         let _version = format_version::parse_footer(version_footer_bytes)?;
-        let sstable_len = u64::deserialize(&mut sstable_len_bytes)?;
         let (column_data, sstable) =
             file_slice_without_sstable_len.split_from_end(sstable_len as usize);
         let column_dictionary = Dictionary::open(sstable)?;
         Ok(ColumnarReader {
             column_dictionary,
             column_data,
+            num_rows,
         })
     }
 
+    pub fn num_rows(&self) -> RowId {
+        self.num_rows
+    }
+
     // TODO Add unit tests
     pub fn list_columns(&self) -> io::Result<Vec<(String, DynamicColumnHandle)>> {
         let mut stream = self.column_dictionary.stream()?;
@@ -73,7 +80,6 @@ impl ColumnarReader {
     ///
     /// There can be more than one column associated to a given column name, provided they have
     /// different types.
-    // TODO fix ugly API
     pub fn read_columns(&self, column_name: &str) -> io::Result<Vec<DynamicColumnHandle>> {
         // Each column is a associated to a given `column_key`,
         // that starts by `column_name\0column_header`.
@@ -120,3 +126,46 @@ impl ColumnarReader {
         self.column_dictionary.num_terms()
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::{ColumnType, ColumnarReader, ColumnarWriter};
+
+    #[test]
+    fn test_list_columns() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("col1", ColumnType::Str, false);
+        columnar_writer.record_column_type("col2", ColumnType::U64, false);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(1, None, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 2);
+        assert_eq!(&columns[0].0, "col1");
+        assert_eq!(columns[0].1.column_type(), ColumnType::Str);
+        assert_eq!(&columns[1].0, "col2");
+        assert_eq!(columns[1].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    fn test_list_columns_strict_typing_prevents_coercion() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1u64);
+        let mut buffer = Vec::new();
+        columnar_writer.serialize(2, None, &mut buffer).unwrap();
+        let columnar = ColumnarReader::open(buffer).unwrap();
+        let columns = columnar.list_columns().unwrap();
+        assert_eq!(columns.len(), 1);
+        assert_eq!(&columns[0].0, "count");
+        assert_eq!(columns[0].1.column_type(), ColumnType::U64);
+    }
+
+    #[test]
+    #[should_panic(expect = "Input type forbidden")]
+    fn test_list_columns_strict_typing_panics_on_wrong_types() {
+        let mut columnar_writer = ColumnarWriter::default();
+        columnar_writer.record_column_type("count", ColumnType::U64, false);
+        columnar_writer.record_numerical(1, "count", 1i64);
+    }
+}

columnar/src/columnar/writer/column_writers.rs

@@ -41,10 +41,31 @@ impl ColumnWriter {
     pub(super) fn operation_iterator<'a, V: SymbolValue>(
         &self,
         arena: &MemoryArena,
+        old_to_new_ids_opt: Option<&[RowId]>,
         buffer: &'a mut Vec<u8>,
     ) -> impl Iterator<Item = ColumnOperation<V>> + 'a {
         buffer.clear();
         self.values.read_to_end(arena, buffer);
+        if let Some(old_to_new_ids) = old_to_new_ids_opt {
+            // TODO avoid the extra deserialization / serialization.
+            let mut sorted_ops: Vec<(RowId, ColumnOperation<V>)> = Vec::new();
+            let mut new_doc = 0u32;
+            let mut cursor = &buffer[..];
+            for op in std::iter::from_fn(|| ColumnOperation::<V>::deserialize(&mut cursor)) {
+                if let ColumnOperation::NewDoc(doc) = &op {
+                    new_doc = old_to_new_ids[*doc as usize];
+                    sorted_ops.push((new_doc, ColumnOperation::NewDoc(new_doc)));
+                } else {
+                    sorted_ops.push((new_doc, op));
+                }
+            }
+            // stable sort is crucial here.
+            sorted_ops.sort_by_key(|(new_doc_id, _)| *new_doc_id);
+            buffer.clear();
+            for (_, op) in sorted_ops {
+                buffer.extend_from_slice(op.serialize().as_ref());
+            }
+        }
         let mut cursor: &[u8] = &buffer[..];
         std::iter::from_fn(move || ColumnOperation::deserialize(&mut cursor))
     }
@@ -114,7 +135,7 @@ impl NumericalColumnWriter {
 /// State used to store what types are still acceptable
 /// after having seen a set of numerical values.
 #[derive(Clone, Copy)]
-enum CompatibleNumericalTypes {
+pub(crate) enum CompatibleNumericalTypes {
     Dynamic {
         all_values_within_i64_range: bool,
         all_values_within_u64_range: bool,
@@ -132,7 +153,7 @@ impl Default for CompatibleNumericalTypes {
 }
 
 impl CompatibleNumericalTypes {
-    fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
+    pub fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -148,7 +169,7 @@ impl CompatibleNumericalTypes {
         }
     }
 
-    fn accept_value(&mut self, numerical_value: NumericalValue) {
+    pub fn accept_value(&mut self, numerical_value: NumericalValue) {
         match self {
             CompatibleNumericalTypes::Dynamic {
                 all_values_within_i64_range,
@@ -168,7 +189,12 @@ impl CompatibleNumericalTypes {
                 }
             },
             CompatibleNumericalTypes::StaticType(typ) => {
-                assert_eq!(numerical_value.numerical_type(), *typ);
+                assert_eq!(
+                    numerical_value.numerical_type(),
+                    *typ,
+                    "Input type forbidden. This column has been forced to type {typ:?}, received \
+                     {numerical_value:?}"
+                );
             }
         }
     }
@@ -205,9 +231,11 @@ impl NumericalColumnWriter {
     pub(super) fn operation_iterator<'a>(
         self,
         arena: &MemoryArena,
+        old_to_new_ids: Option<&[RowId]>,
         buffer: &'a mut Vec<u8>,
     ) -> impl Iterator<Item = ColumnOperation<NumericalValue>> + 'a {
-        self.column_writer.operation_iterator(arena, buffer)
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, buffer)
     }
 }
 
@@ -215,6 +243,14 @@ impl NumericalColumnWriter {
 pub(crate) struct StrOrBytesColumnWriter {
     pub(crate) dictionary_id: u32,
     pub(crate) column_writer: ColumnWriter,
+    // If true, when facing a multivalued cardinality,
+    // values associated to a given document will be sorted.
+    //
+    // This is useful for facets.
+    //
+    // If false, the order of appearance in the document will be
+    // observed.
+    pub(crate) sort_values_within_row: bool,
 }
 
 impl StrOrBytesColumnWriter {
@@ -222,6 +258,7 @@ impl StrOrBytesColumnWriter {
         StrOrBytesColumnWriter {
             dictionary_id,
             column_writer: Default::default(),
+            sort_values_within_row: false,
         }
     }
 
@@ -239,9 +276,11 @@ impl StrOrBytesColumnWriter {
     pub(super) fn operation_iterator<'a>(
         &self,
         arena: &MemoryArena,
+        old_to_new_ids: Option<&[RowId]>,
         byte_buffer: &'a mut Vec<u8>,
     ) -> impl Iterator<Item = ColumnOperation<UnorderedId>> + 'a {
-        self.column_writer.operation_iterator(arena, byte_buffer)
+        self.column_writer
+            .operation_iterator(arena, old_to_new_ids, byte_buffer)
     }
 }
 

columnar/src/columnar/writer/mod.rs

@@ -7,8 +7,9 @@ use std::io;
 use std::net::Ipv6Addr;
 
 use column_operation::ColumnOperation;
+pub(crate) use column_writers::CompatibleNumericalTypes;
 use common::CountingWriter;
-use serializer::ColumnarSerializer;
+pub(crate) use serializer::ColumnarSerializer;
 use stacker::{Addr, ArenaHashMap, MemoryArena};
 
 use crate::column_index::SerializableColumnIndex;
@@ -29,10 +30,7 @@ use crate::{Cardinality, RowId};
 #[derive(Default)]
 struct SpareBuffers {
     value_index_builders: PreallocatedIndexBuilders,
-    i64_values: Vec<i64>,
     u64_values: Vec<u64>,
-    f64_values: Vec<f64>,
-    bool_values: Vec<bool>,
     ip_addr_values: Vec<Ipv6Addr>,
 }
 
@@ -47,7 +45,7 @@ struct SpareBuffers {
 /// columnar_writer.record_str(1u32 /* doc id */, "product_name", "Apple");
 /// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10.5f64); //< uh oh we ended up mixing integer and floats.
 /// let mut wrt: Vec<u8> =  Vec::new();
-/// columnar_writer.serialize(2u32, &mut wrt).unwrap();
+/// columnar_writer.serialize(2u32, None, &mut wrt).unwrap();
 /// ```
 pub struct ColumnarWriter {
     numerical_field_hash_map: ArenaHashMap,
@@ -106,7 +104,65 @@ impl ColumnarWriter {
             + self.datetime_field_hash_map.mem_usage()
     }
 
-    pub fn record_column_type(&mut self, column_name: &str, column_type: ColumnType) {
+    /// Returns the list of doc ids from 0..num_docs sorted by the `sort_field`
+    /// column.
+    ///
+    /// If the column is multivalued, use the first value for scoring.
+    /// If no value is associated to a specific row, the document is assigned
+    /// the lowest possible score.
+    ///
+    /// The sort applied is stable.
+    pub fn sort_order(&self, sort_field: &str, num_docs: RowId, reversed: bool) -> Vec<u32> {
+        let Some(numerical_col_writer) =
+            self.numerical_field_hash_map.get::<NumericalColumnWriter>(sort_field.as_bytes()) else {
+                return Vec::new();
+        };
+        let mut symbols_buffer = Vec::new();
+        let mut values = Vec::new();
+        let mut last_doc_opt: Option<RowId> = None;
+        for op in numerical_col_writer.operation_iterator(&self.arena, None, &mut symbols_buffer) {
+            match op {
+                ColumnOperation::NewDoc(doc) => {
+                    last_doc_opt = Some(doc);
+                }
+                ColumnOperation::Value(numerical_value) => {
+                    if let Some(last_doc) = last_doc_opt {
+                        let score: f32 = f64::coerce(numerical_value) as f32;
+                        values.push((score, last_doc));
+                    }
+                }
+            }
+        }
+        for doc in values.len() as u32..num_docs {
+            values.push((0.0f32, doc));
+        }
+        values.sort_by(|(left_score, _), (right_score, _)| {
+            if reversed {
+                right_score.partial_cmp(left_score).unwrap()
+            } else {
+                left_score.partial_cmp(right_score).unwrap()
+            }
+        });
+        values.into_iter().map(|(_score, doc)| doc).collect()
+    }
+
+    /// Records a column type. This is useful to bypass the coercion process,
+    /// makes sure the empty is present in the resulting columnar, or set
+    /// the `sort_values_within_row`.
+    ///
+    /// `sort_values_within_row` is only allowed for `Bytes` or `Str` columns.
+    pub fn record_column_type(
+        &mut self,
+        column_name: &str,
+        column_type: ColumnType,
+        sort_values_within_row: bool,
+    ) {
+        if sort_values_within_row {
+            assert!(
+                column_type == ColumnType::Bytes || column_type == ColumnType::Str,
+                "sort_values_within_row is only allowed for Bytes and Str columns",
+            );
+        }
         match column_type {
             ColumnType::Str | ColumnType::Bytes => {
                 let (hash_map, dictionaries) = (
@@ -121,13 +177,15 @@ impl ColumnarWriter {
                     hash_map,
                     column_name,
                     |column_opt: Option<StrOrBytesColumnWriter>| {
-                        if let Some(column_writer) = column_opt {
+                        let mut column_writer = if let Some(column_writer) = column_opt {
                             column_writer
                         } else {
                             let dictionary_id = dictionaries.len() as u32;
                             dictionaries.push(DictionaryBuilder::default());
                             StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
-                        }
+                        };
+                        column_writer.sort_values_within_row = sort_values_within_row;
+                        column_writer
                     },
                 );
             }
@@ -165,18 +223,6 @@ impl ColumnarWriter {
         }
     }
 
-    pub fn force_numerical_type(&mut self, column_name: &str, numerical_type: NumericalType) {
-        mutate_or_create_column(
-            &mut self.numerical_field_hash_map,
-            column_name,
-            |column_opt: Option<NumericalColumnWriter>| {
-                let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
-                column.force_numerical_type(numerical_type);
-                column
-            },
-        );
-    }
-
     pub fn record_numerical<T: Into<NumericalValue> + Copy>(
         &mut self,
         doc: RowId,
@@ -274,7 +320,12 @@ impl ColumnarWriter {
             },
         );
     }
-    pub fn serialize(&mut self, num_docs: RowId, wrt: &mut dyn io::Write) -> io::Result<()> {
+    pub fn serialize(
+        &mut self,
+        num_docs: RowId,
+        old_to_new_row_ids: Option<&[RowId]>,
+        wrt: &mut dyn io::Write,
+    ) -> io::Result<()> {
         let mut serializer = ColumnarSerializer::new(wrt);
         let mut columns: Vec<(&[u8], ColumnType, Addr)> = self
             .numerical_field_hash_map
@@ -317,19 +368,19 @@ impl ColumnarWriter {
         let mut symbol_byte_buffer: Vec<u8> = Vec::new();
         for (column_name, column_type, addr) in columns {
             match column_type {
-                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)
-                    };
+                ColumnType::Bool => {
+                    let column_writer: ColumnWriter = self.bool_field_hash_map.read(addr);
                     let cardinality = column_writer.get_cardinality(num_docs);
                     let mut column_serializer =
-                        serializer.serialize_column(column_name, ColumnType::Bool);
+                        serializer.serialize_column(column_name, column_type);
                     serialize_bool_column(
                         cardinality,
                         num_docs,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
@@ -342,7 +393,11 @@ impl ColumnarWriter {
                     serialize_ip_addr_column(
                         cardinality,
                         num_docs,
-                        column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
@@ -364,39 +419,68 @@ impl ColumnarWriter {
                     serialize_bytes_or_str_column(
                         cardinality,
                         num_docs,
+                        str_or_bytes_column_writer.sort_values_within_row,
                         dictionary_builder,
-                        str_or_bytes_column_writer
-                            .operation_iterator(arena, &mut symbol_byte_buffer),
+                        str_or_bytes_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
-                ColumnType::I64 | ColumnType::F64 | ColumnType::U64 => {
+                ColumnType::F64 | ColumnType::I64 | ColumnType::U64 => {
                     let numerical_column_writer: NumericalColumnWriter =
                         self.numerical_field_hash_map.read(addr);
-                    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));
+                        serializer.serialize_column(column_name, column_type);
+                    let numerical_type = column_type.numerical_type().unwrap();
                     serialize_numerical_column(
                         cardinality,
                         num_docs,
                         numerical_type,
-                        numerical_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
+                        numerical_column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
+                        buffers,
+                        &mut column_serializer,
+                    )?;
+                }
+                ColumnType::DateTime => {
+                    let column_writer: ColumnWriter = self.datetime_field_hash_map.read(addr);
+                    let cardinality = column_writer.get_cardinality(num_docs);
+                    let mut column_serializer =
+                        serializer.serialize_column(column_name, ColumnType::DateTime);
+                    serialize_numerical_column(
+                        cardinality,
+                        num_docs,
+                        NumericalType::I64,
+                        column_writer.operation_iterator(
+                            arena,
+                            old_to_new_row_ids,
+                            &mut symbol_byte_buffer,
+                        ),
                         buffers,
                         &mut column_serializer,
                     )?;
                 }
             };
         }
-        serializer.finalize()?;
+        serializer.finalize(num_docs)?;
         Ok(())
     }
 }
 
+// Serialize [Dictionary, Column, dictionary num bytes U32::LE]
+// Column: [Column Index, Column Values, column index num bytes U32::LE]
 fn serialize_bytes_or_str_column(
     cardinality: Cardinality,
     num_docs: RowId,
+    sort_values_within_row: bool,
     dictionary_builder: &DictionaryBuilder,
     operation_it: impl Iterator<Item = ColumnOperation<UnorderedId>>,
     buffers: &mut SpareBuffers,
@@ -425,6 +509,7 @@ fn serialize_bytes_or_str_column(
         operation_iterator,
         cardinality,
         num_docs,
+        sort_values_within_row,
         value_index_builders,
         u64_values,
         &mut wrt,
@@ -444,8 +529,6 @@ fn serialize_numerical_column(
     let SpareBuffers {
         value_index_builders,
         u64_values,
-        i64_values,
-        f64_values,
         ..
     } = buffers;
     match numerical_type {
@@ -454,8 +537,9 @@ fn serialize_numerical_column(
                 coerce_numerical_symbol::<i64>(op_iterator),
                 cardinality,
                 num_docs,
+                false,
                 value_index_builders,
-                i64_values,
+                u64_values,
                 wrt,
             )?;
         }
@@ -464,6 +548,7 @@ fn serialize_numerical_column(
                 coerce_numerical_symbol::<u64>(op_iterator),
                 cardinality,
                 num_docs,
+                false,
                 value_index_builders,
                 u64_values,
                 wrt,
@@ -474,8 +559,9 @@ fn serialize_numerical_column(
                 coerce_numerical_symbol::<f64>(op_iterator),
                 cardinality,
                 num_docs,
+                false,
                 value_index_builders,
-                f64_values,
+                u64_values,
                 wrt,
             )?;
         }
@@ -492,15 +578,19 @@ fn serialize_bool_column(
 ) -> io::Result<()> {
     let SpareBuffers {
         value_index_builders,
-        bool_values,
+        u64_values,
         ..
     } = buffers;
     send_to_serialize_column_mappable_to_u64(
-        column_operations_it,
+        column_operations_it.map(|bool_column_operation| match bool_column_operation {
+            ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
+            ColumnOperation::Value(bool_val) => ColumnOperation::Value(bool_val.to_u64()),
+        }),
         cardinality,
         num_docs,
+        false,
         value_index_builders,
-        bool_values,
+        u64_values,
         wrt,
     )?;
     Ok(())
@@ -530,11 +620,11 @@ fn serialize_ip_addr_column(
 }
 
 fn send_to_serialize_column_mappable_to_u128<
-    T: Copy + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU128 + PartialOrd,
+    T: Copy + Ord + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU128 + PartialOrd,
 >(
     op_iterator: impl Iterator<Item = ColumnOperation<T>>,
     cardinality: Cardinality,
-    num_docs: RowId,
+    num_rows: RowId,
     value_index_builders: &mut PreallocatedIndexBuilders,
     values: &mut Vec<T>,
     mut wrt: impl io::Write,
@@ -556,37 +646,47 @@ where
         Cardinality::Optional => {
             let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let optional_index = optional_index_builder.finish(num_docs);
-            SerializableColumnIndex::Optional(Box::new(optional_index))
+            let optional_index = optional_index_builder.finish(num_rows);
+            SerializableColumnIndex::Optional {
+                num_rows,
+                non_null_row_ids: Box::new(optional_index),
+            }
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            let multivalued_index = multivalued_index_builder.finish(num_rows);
             SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u128(
         serializable_column_index,
-        || values.iter().cloned(),
-        values.len() as u32,
+        &&values[..],
         &mut wrt,
     )?;
     Ok(())
 }
 
-fn send_to_serialize_column_mappable_to_u64<
-    T: Copy + Default + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU64 + PartialOrd,
->(
-    op_iterator: impl Iterator<Item = ColumnOperation<T>>,
+fn sort_values_within_row_in_place(multivalued_index: &[RowId], values: &mut Vec<u64>) {
+    let mut start_index: usize = 0;
+    for end_index in multivalued_index.iter().copied() {
+        let end_index = end_index as usize;
+        values[start_index..end_index].sort_unstable();
+        start_index = end_index;
+    }
+}
+
+fn send_to_serialize_column_mappable_to_u64(
+    op_iterator: impl Iterator<Item = ColumnOperation<u64>>,
     cardinality: Cardinality,
-    num_docs: RowId,
+    num_rows: RowId,
+    sort_values_within_row: bool,
     value_index_builders: &mut PreallocatedIndexBuilders,
-    values: &mut Vec<T>,
+    values: &mut Vec<u64>,
     mut wrt: impl io::Write,
 ) -> io::Result<()>
 where
-    for<'a> VecColumn<'a, T>: ColumnValues<T>,
+    for<'a> VecColumn<'a, u64>: ColumnValues<u64>,
 {
     values.clear();
     let serializable_column_index = match cardinality {
@@ -601,19 +701,25 @@ where
         Cardinality::Optional => {
             let optional_index_builder = value_index_builders.borrow_optional_index_builder();
             consume_operation_iterator(op_iterator, optional_index_builder, values);
-            let optional_index = optional_index_builder.finish(num_docs);
-            SerializableColumnIndex::Optional(Box::new(optional_index))
+            let optional_index = optional_index_builder.finish(num_rows);
+            SerializableColumnIndex::Optional {
+                non_null_row_ids: Box::new(optional_index),
+                num_rows,
+            }
         }
         Cardinality::Multivalued => {
             let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
             consume_operation_iterator(op_iterator, multivalued_index_builder, values);
-            let multivalued_index = multivalued_index_builder.finish(num_docs);
+            let multivalued_index = multivalued_index_builder.finish(num_rows);
+            if sort_values_within_row {
+                sort_values_within_row_in_place(multivalued_index, values);
+            }
             SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
         }
     };
     crate::column::serialize_column_mappable_to_u64(
         serializable_column_index,
-        &VecColumn::from(&values[..]),
+        &&values[..],
         &mut wrt,
     )?;
     Ok(())
@@ -621,17 +727,17 @@ where
 
 fn coerce_numerical_symbol<T>(
     operation_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
-) -> impl Iterator<Item = ColumnOperation<T>>
-where T: Coerce {
+) -> impl Iterator<Item = ColumnOperation<u64>>
+where T: Coerce + MonotonicallyMappableToU64 {
     operation_iterator.map(|symbol| match symbol {
         ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
         ColumnOperation::Value(numerical_value) => {
-            ColumnOperation::Value(Coerce::coerce(numerical_value))
+            ColumnOperation::Value(T::coerce(numerical_value).to_u64())
         }
     })
 }
 
-fn consume_operation_iterator<T: std::fmt::Debug, TIndexBuilder: IndexBuilder>(
+fn consume_operation_iterator<T: Ord, TIndexBuilder: IndexBuilder>(
     operation_iterator: impl Iterator<Item = ColumnOperation<T>>,
     index_builder: &mut TIndexBuilder,
     values: &mut Vec<T>,
@@ -666,7 +772,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(3), Cardinality::Full);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 6);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
@@ -695,7 +801,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(3), Cardinality::Optional);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 4);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(1u32)));
@@ -718,7 +824,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(2), Cardinality::Optional);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 2);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
@@ -737,7 +843,7 @@ mod tests {
         assert_eq!(column_writer.get_cardinality(1), Cardinality::Multivalued);
         let mut buffer = Vec::new();
         let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
-            .operation_iterator(&mut arena, &mut buffer)
+            .operation_iterator(&mut arena, None, &mut buffer)
             .collect();
         assert_eq!(symbols.len(), 3);
         assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));

columnar/src/columnar/writer/serializer.rs

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

columnar/src/columnar/writer/value_index.rs

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

columnar/src/dynamic_column.rs

@@ -8,7 +8,7 @@ use common::{HasLen, OwnedBytes};
 use crate::column::{BytesColumn, Column, StrColumn};
 use crate::column_values::{monotonic_map_column, StrictlyMonotonicFn};
 use crate::columnar::ColumnType;
-use crate::{DateTime, NumericalType};
+use crate::{Cardinality, DateTime, NumericalType};
 
 #[derive(Clone)]
 pub enum DynamicColumn {
@@ -23,6 +23,18 @@ pub enum DynamicColumn {
 }
 
 impl DynamicColumn {
+    pub fn get_cardinality(&self) -> Cardinality {
+        match self {
+            DynamicColumn::Bool(c) => c.get_cardinality(),
+            DynamicColumn::I64(c) => c.get_cardinality(),
+            DynamicColumn::U64(c) => c.get_cardinality(),
+            DynamicColumn::F64(c) => c.get_cardinality(),
+            DynamicColumn::IpAddr(c) => c.get_cardinality(),
+            DynamicColumn::DateTime(c) => c.get_cardinality(),
+            DynamicColumn::Bytes(c) => c.ords().get_cardinality(),
+            DynamicColumn::Str(c) => c.ords().get_cardinality(),
+        }
+    }
     pub fn column_type(&self) -> ColumnType {
         match self {
             DynamicColumn::Bool(_) => ColumnType::Bool,
@@ -36,6 +48,14 @@ impl DynamicColumn {
         }
     }
 
+    pub fn coerce_numerical(self, target_numerical_type: NumericalType) -> Option<Self> {
+        match target_numerical_type {
+            NumericalType::I64 => self.coerce_to_i64(),
+            NumericalType::U64 => self.coerce_to_u64(),
+            NumericalType::F64 => self.coerce_to_f64(),
+        }
+    }
+
     pub fn is_numerical(&self) -> bool {
         self.column_type().numerical_type().is_some()
     }
@@ -50,7 +70,7 @@ impl DynamicColumn {
         self.column_type().numerical_type() == Some(NumericalType::U64)
     }
 
-    pub fn coerce_to_f64(self) -> Option<DynamicColumn> {
+    fn coerce_to_f64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
                 idx: column.idx,
@@ -64,7 +84,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    pub fn coerce_to_i64(self) -> Option<DynamicColumn> {
+    fn coerce_to_i64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::U64(column) => {
                 if column.max_value() > i64::MAX as u64 {
@@ -79,7 +99,7 @@ impl DynamicColumn {
             _ => None,
         }
     }
-    pub fn coerce_to_u64(self) -> Option<DynamicColumn> {
+    fn coerce_to_u64(self) -> Option<DynamicColumn> {
         match self {
             DynamicColumn::I64(column) => {
                 if column.min_value() < 0 {
@@ -215,10 +235,8 @@ impl DynamicColumnHandle {
 
     fn open_internal(&self, column_bytes: OwnedBytes) -> io::Result<DynamicColumn> {
         let dynamic_column: DynamicColumn = match self.column_type {
-            ColumnType::Bytes => {
-                crate::column::open_column_bytes::<BytesColumn>(column_bytes)?.into()
-            }
-            ColumnType::Str => crate::column::open_column_bytes::<StrColumn>(column_bytes)?.into(),
+            ColumnType::Bytes => crate::column::open_column_bytes(column_bytes)?.into(),
+            ColumnType::Str => crate::column::open_column_str(column_bytes)?.into(),
             ColumnType::I64 => crate::column::open_column_u64::<i64>(column_bytes)?.into(),
             ColumnType::U64 => crate::column::open_column_u64::<u64>(column_bytes)?.into(),
             ColumnType::F64 => crate::column::open_column_u64::<f64>(column_bytes)?.into(),

columnar/src/iterable.rs

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

columnar/src/lib.rs

@@ -11,30 +11,48 @@ use std::io;
 
 mod column;
 mod column_index;
-mod column_values;
+pub mod column_values;
 mod columnar;
 mod dictionary;
 mod dynamic_column;
+mod iterable;
 pub(crate) mod utils;
 mod value;
 
 pub use column::{BytesColumn, Column, StrColumn};
-pub use column_values::ColumnValues;
+pub use column_index::ColumnIndex;
+pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
 pub use columnar::{
     merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
-    MergeDocOrder,
+    MergeRowOrder, ShuffleMergeOrder, StackMergeOrder,
 };
+use sstable::VoidSSTable;
 pub use value::{NumericalType, NumericalValue};
 
 pub use self::dynamic_column::{DynamicColumn, DynamicColumnHandle};
 
 pub type RowId = u32;
 
+#[derive(Clone, Copy)]
+pub struct RowAddr {
+    pub segment_ord: u32,
+    pub row_id: RowId,
+}
+
+pub use sstable::Dictionary;
+pub type Streamer<'a> = sstable::Streamer<'a, VoidSSTable>;
+
 #[derive(Clone, Copy, PartialOrd, PartialEq, Default, Debug)]
 pub struct DateTime {
     pub timestamp_micros: i64,
 }
 
+impl DateTime {
+    pub fn into_timestamp_micros(self) -> i64 {
+        self.timestamp_micros
+    }
+}
+
 #[derive(Copy, Clone, Debug)]
 pub struct InvalidData;
 
@@ -62,6 +80,12 @@ pub enum Cardinality {
 }
 
 impl Cardinality {
+    pub fn is_optional(&self) -> bool {
+        matches!(self, Cardinality::Optional)
+    }
+    pub fn is_multivalue(&self) -> bool {
+        matches!(self, Cardinality::Multivalued)
+    }
     pub(crate) fn to_code(self) -> u8 {
         self as u8
     }

columnar/src/tests.rs

@@ -12,7 +12,7 @@ fn test_dataframe_writer_str() {
     dataframe_writer.record_str(1u32, "my_string", "hello");
     dataframe_writer.record_str(3u32, "my_string", "helloeee");
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    dataframe_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
@@ -26,7 +26,7 @@ fn test_dataframe_writer_bytes() {
     dataframe_writer.record_bytes(1u32, "my_string", b"hello");
     dataframe_writer.record_bytes(3u32, "my_string", b"helloeee");
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    dataframe_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
@@ -40,7 +40,7 @@ fn test_dataframe_writer_bool() {
     dataframe_writer.record_bool(1u32, "bool.value", false);
     dataframe_writer.record_bool(3u32, "bool.value", true);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    dataframe_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("bool.value").unwrap();
@@ -63,7 +63,7 @@ fn test_dataframe_writer_u64_multivalued() {
     dataframe_writer.record_numerical(6u32, "divisor", 2u64);
     dataframe_writer.record_numerical(6u32, "divisor", 3u64);
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(7, &mut buffer).unwrap();
+    dataframe_writer.serialize(7, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("divisor").unwrap();
@@ -84,7 +84,7 @@ fn test_dataframe_writer_ip_addr() {
     dataframe_writer.record_ip_addr(1, "ip_addr", Ipv6Addr::from_u128(1001));
     dataframe_writer.record_ip_addr(3, "ip_addr", Ipv6Addr::from_u128(1050));
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(5, &mut buffer).unwrap();
+    dataframe_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("ip_addr").unwrap();
@@ -113,7 +113,7 @@ fn test_dataframe_writer_numerical() {
     dataframe_writer.record_numerical(2u32, "srical.value", NumericalValue::U64(13u64));
     dataframe_writer.record_numerical(4u32, "srical.value", NumericalValue::U64(15u64));
     let mut buffer: Vec<u8> = Vec::new();
-    dataframe_writer.serialize(6, &mut buffer).unwrap();
+    dataframe_writer.serialize(6, None, &mut buffer).unwrap();
     let columnar = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar.num_columns(), 1);
     let cols: Vec<DynamicColumnHandle> = columnar.read_columns("srical.value").unwrap();
@@ -144,7 +144,7 @@ fn test_dictionary_encoded_str() {
     columnar_writer.record_str(3, "my.column", "c");
     columnar_writer.record_str(3, "my.column2", "different_column!");
     columnar_writer.record_str(4, "my.column", "b");
-    columnar_writer.serialize(5, &mut buffer).unwrap();
+    columnar_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar_reader = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar_reader.num_columns(), 2);
     let col_handles = columnar_reader.read_columns("my.column").unwrap();
@@ -176,7 +176,7 @@ fn test_dictionary_encoded_bytes() {
     columnar_writer.record_bytes(3, "my.column", b"c");
     columnar_writer.record_bytes(3, "my.column2", b"different_column!");
     columnar_writer.record_bytes(4, "my.column", b"b");
-    columnar_writer.serialize(5, &mut buffer).unwrap();
+    columnar_writer.serialize(5, None, &mut buffer).unwrap();
     let columnar_reader = ColumnarReader::open(buffer).unwrap();
     assert_eq!(columnar_reader.num_columns(), 2);
     let col_handles = columnar_reader.read_columns("my.column").unwrap();

common/src/serialize.rs

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

common/src/vint.rs

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

examples-disabled/aggregation.rs

@@ -13,7 +13,7 @@ use tantivy::aggregation::agg_result::AggregationResults;
 use tantivy::aggregation::metric::AverageAggregation;
 use tantivy::aggregation::AggregationCollector;
 use tantivy::query::TermQuery;
-use tantivy::schema::{self, Cardinality, IndexRecordOption, Schema, TextFieldIndexing};
+use tantivy::schema::{self, IndexRecordOption, Schema, TextFieldIndexing};
 use tantivy::{doc, Index, Term};
 
 fn main() -> tantivy::Result<()> {
@@ -25,7 +25,7 @@ fn main() -> tantivy::Result<()> {
         .set_stored();
     let text_field = schema_builder.add_text_field("text", text_fieldtype);
     let score_fieldtype =
-        crate::schema::NumericOptions::default().set_fast(Cardinality::SingleValue);
+        crate::schema::NumericOptions::default().set_fast();
     let highscore_field = schema_builder.add_f64_field("highscore", score_fieldtype.clone());
     let price_field = schema_builder.add_f64_field("price", score_fieldtype);
 

examples-disabled/date_time_field.rs

@@ -4,7 +4,7 @@
 
 use tantivy::collector::TopDocs;
 use tantivy::query::QueryParser;
-use tantivy::schema::{Cardinality, DateOptions, Schema, Value, INDEXED, STORED, STRING};
+use tantivy::schema::{DateOptions, Schema, Value, INDEXED, STORED, STRING};
 use tantivy::Index;
 
 fn main() -> tantivy::Result<()> {
@@ -12,7 +12,7 @@ fn main() -> tantivy::Result<()> {
     let mut schema_builder = Schema::builder();
     let opts = DateOptions::from(INDEXED)
         .set_stored()
-        .set_fast(Cardinality::SingleValue)
+        .set_fast()
         .set_precision(tantivy::DatePrecision::Seconds);
     let occurred_at = schema_builder.add_date_field("occurred_at", opts);
     let event_type = schema_builder.add_text_field("event", STRING | STORED);

fastfield_codecs/Cargo.toml

@@ -1,33 +0,0 @@
-[package]
-name = "fastfield_codecs"
-version = "0.3.0"
-authors = ["Pascal Seitz <pascal@quickwit.io>"]
-license = "MIT"
-edition = "2021"
-description = "Fast field codecs used by tantivy"
-documentation = "https://docs.rs/fastfield_codecs/"
-homepage = "https://github.com/quickwit-oss/tantivy"
-repository = "https://github.com/quickwit-oss/tantivy"
-
-# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[dependencies]
-common = { version = "0.5", path = "../common/", package = "tantivy-common" }
-tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
-prettytable-rs = {version="0.10.0", optional= true}
-rand = {version="0.8.3", optional= true}
-fastdivide = "0.4"
-log = "0.4"
-itertools = { version = "0.10.3" }
-measure_time = { version="0.8.2", optional=true}
-
-[dev-dependencies]
-more-asserts = "0.3.0"
-proptest = "1.0.0"
-rand = "0.8.3"
-
-[features]
-bin = ["prettytable-rs", "rand", "measure_time"]
-default = ["bin"]
-unstable = []
-

fastfield_codecs/README.md

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

fastfield_codecs/benches/bench.rs

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

fastfield_codecs/src/bitpacked.rs

@@ -1,116 +0,0 @@
-use std::io::{self, Write};
-
-use common::OwnedBytes;
-use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
-
-use crate::serialize::NormalizedHeader;
-use crate::{Column, FastFieldCodec, FastFieldCodecType};
-
-/// Depending on the field type, a different
-/// fast field is required.
-#[derive(Clone)]
-pub struct BitpackedReader {
-    data: OwnedBytes,
-    bit_unpacker: BitUnpacker,
-    normalized_header: NormalizedHeader,
-}
-
-impl Column for BitpackedReader {
-    #[inline]
-    fn get_val(&self, doc: u32) -> u64 {
-        self.bit_unpacker.get(doc, &self.data)
-    }
-    #[inline]
-    fn min_value(&self) -> u64 {
-        // The BitpackedReader assumes a normalized vector.
-        0
-    }
-    #[inline]
-    fn max_value(&self) -> u64 {
-        self.normalized_header.max_value
-    }
-    #[inline]
-    fn num_vals(&self) -> u32 {
-        self.normalized_header.num_vals
-    }
-}
-
-pub struct BitpackedCodec;
-
-impl FastFieldCodec for BitpackedCodec {
-    /// The CODEC_TYPE is an enum value used for serialization.
-    const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::Bitpacked;
-
-    type Reader = BitpackedReader;
-
-    /// Opens a fast field given a file.
-    fn open_from_bytes(
-        data: OwnedBytes,
-        normalized_header: NormalizedHeader,
-    ) -> io::Result<Self::Reader> {
-        let num_bits = compute_num_bits(normalized_header.max_value);
-        let bit_unpacker = BitUnpacker::new(num_bits);
-        Ok(BitpackedReader {
-            data,
-            bit_unpacker,
-            normalized_header,
-        })
-    }
-
-    /// Serializes data with the BitpackedFastFieldSerializer.
-    ///
-    /// The bitpacker assumes that the column has been normalized.
-    /// i.e. It has already been shifted by its minimum value, so that its
-    /// current minimum value is 0.
-    ///
-    /// Ideally, we made a shift upstream on the column so that `col.min_value() == 0`.
-    fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()> {
-        assert_eq!(column.min_value(), 0u64);
-        let num_bits = compute_num_bits(column.max_value());
-        let mut bit_packer = BitPacker::new();
-        for val in column.iter() {
-            bit_packer.write(val, num_bits, write)?;
-        }
-        bit_packer.close(write)?;
-        Ok(())
-    }
-
-    fn estimate(column: &dyn Column) -> Option<f32> {
-        let num_bits = compute_num_bits(column.max_value());
-        let num_bits_uncompressed = 64;
-        Some(num_bits as f32 / num_bits_uncompressed as f32)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::tests::get_codec_test_datasets;
-
-    fn create_and_validate(data: &[u64], name: &str) {
-        crate::tests::create_and_validate::<BitpackedCodec>(data, name);
-    }
-
-    #[test]
-    fn test_with_codec_data_sets() {
-        let data_sets = get_codec_test_datasets();
-        for (mut data, name) in data_sets {
-            create_and_validate(&data, name);
-            data.reverse();
-            create_and_validate(&data, name);
-        }
-    }
-
-    #[test]
-    fn bitpacked_fast_field_rand() {
-        for _ in 0..500 {
-            let mut data = (0..1 + rand::random::<u8>() as usize)
-                .map(|_| rand::random::<i64>() as u64 / 2)
-                .collect::<Vec<_>>();
-            create_and_validate(&data, "rand");
-
-            data.reverse();
-            create_and_validate(&data, "rand");
-        }
-    }
-}

fastfield_codecs/src/blockwise_linear.rs

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

fastfield_codecs/src/column.rs

@@ -1,352 +0,0 @@
-use std::fmt::{self, Debug};
-use std::marker::PhantomData;
-use std::ops::{Range, RangeInclusive};
-
-use tantivy_bitpacker::minmax;
-
-use crate::monotonic_mapping::StrictlyMonotonicFn;
-
-/// `Column` provides columnar access on a field.
-pub trait Column<T: PartialOrd + Debug = u64>: Send + Sync {
-    /// Return the value associated with the given idx.
-    ///
-    /// This accessor should return as fast as possible.
-    ///
-    /// # Panics
-    ///
-    /// May panic if `idx` is greater than the column length.
-    fn get_val(&self, idx: u32) -> T;
-
-    /// Fills an output buffer with the fast field values
-    /// associated with the `DocId` going from
-    /// `start` to `start + output.len()`.
-    ///
-    /// # Panics
-    ///
-    /// Must panic if `start + output.len()` is greater than
-    /// the segment's `maxdoc`.
-    #[inline]
-    fn get_range(&self, start: u64, output: &mut [T]) {
-        for (out, idx) in output.iter_mut().zip(start..) {
-            *out = self.get_val(idx as u32);
-        }
-    }
-
-    /// Get the positions of values which are in the provided value range.
-    ///
-    /// Note that position == docid for single value fast fields
-    #[inline]
-    fn get_docids_for_value_range(
-        &self,
-        value_range: RangeInclusive<T>,
-        doc_id_range: Range<u32>,
-        positions: &mut Vec<u32>,
-    ) {
-        let doc_id_range = doc_id_range.start..doc_id_range.end.min(self.num_vals());
-
-        for idx in doc_id_range.start..doc_id_range.end {
-            let val = self.get_val(idx);
-            if value_range.contains(&val) {
-                positions.push(idx);
-            }
-        }
-    }
-
-    /// Returns the minimum value for this fast field.
-    ///
-    /// This min_value may not be exact.
-    /// For instance, the min value does not take in account of possible
-    /// deleted document. All values are however guaranteed to be higher than
-    /// `.min_value()`.
-    fn min_value(&self) -> T;
-
-    /// Returns the maximum value for this fast field.
-    ///
-    /// This max_value may not be exact.
-    /// For instance, the max value does not take in account of possible
-    /// deleted document. All values are however guaranteed to be higher than
-    /// `.max_value()`.
-    fn max_value(&self) -> T;
-
-    /// The number of values in the column.
-    fn num_vals(&self) -> u32;
-
-    /// Returns a iterator over the data
-    fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = T> + 'a> {
-        Box::new((0..self.num_vals()).map(|idx| self.get_val(idx)))
-    }
-}
-
-/// VecColumn provides `Column` over a slice.
-pub struct VecColumn<'a, T = u64> {
-    values: &'a [T],
-    min_value: T,
-    max_value: T,
-}
-
-impl<'a, C: Column<T>, T: Copy + PartialOrd + fmt::Debug> Column<T> for &'a C {
-    fn get_val(&self, idx: u32) -> T {
-        (*self).get_val(idx)
-    }
-
-    fn min_value(&self) -> T {
-        (*self).min_value()
-    }
-
-    fn max_value(&self) -> T {
-        (*self).max_value()
-    }
-
-    fn num_vals(&self) -> u32 {
-        (*self).num_vals()
-    }
-
-    fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
-        (*self).iter()
-    }
-
-    fn get_range(&self, start: u64, output: &mut [T]) {
-        (*self).get_range(start, output)
-    }
-}
-
-impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> Column<T> for VecColumn<'a, T> {
-    fn get_val(&self, position: u32) -> T {
-        self.values[position as usize]
-    }
-
-    fn iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
-        Box::new(self.values.iter().copied())
-    }
-
-    fn min_value(&self) -> T {
-        self.min_value
-    }
-
-    fn max_value(&self) -> T {
-        self.max_value
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.values.len() as u32
-    }
-
-    fn get_range(&self, start: u64, output: &mut [T]) {
-        output.copy_from_slice(&self.values[start as usize..][..output.len()])
-    }
-}
-
-impl<'a, T: Copy + PartialOrd + Default, V> From<&'a V> for VecColumn<'a, T>
-where V: AsRef<[T]> + ?Sized
-{
-    fn from(values: &'a V) -> Self {
-        let values = values.as_ref();
-        let (min_value, max_value) = minmax(values.iter().copied()).unwrap_or_default();
-        Self {
-            values,
-            min_value,
-            max_value,
-        }
-    }
-}
-
-struct MonotonicMappingColumn<C, T, Input> {
-    from_column: C,
-    monotonic_mapping: T,
-    _phantom: PhantomData<Input>,
-}
-
-/// Creates a view of a column transformed by a strictly monotonic mapping. See
-/// [`StrictlyMonotonicFn`].
-///
-/// E.g. apply a gcd monotonic_mapping([100, 200, 300]) == [1, 2, 3]
-/// monotonic_mapping.mapping() is expected to be injective, and we should always have
-/// monotonic_mapping.inverse(monotonic_mapping.mapping(el)) == el
-///
-/// The inverse of the mapping is required for:
-/// `fn get_positions_for_value_range(&self, range: RangeInclusive<T>) -> Vec<u64> `
-/// The user provides the original value range and we need to monotonic map them in the same way the
-/// serialization does before calling the underlying column.
-///
-/// Note that when opening a codec, the monotonic_mapping should be the inverse of the mapping
-/// during serialization. And therefore the monotonic_mapping_inv when opening is the same as
-/// monotonic_mapping during serialization.
-pub fn monotonic_map_column<C, T, Input, Output>(
-    from_column: C,
-    monotonic_mapping: T,
-) -> impl Column<Output>
-where
-    C: Column<Input>,
-    T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Copy + Debug,
-    Output: PartialOrd + Send + Sync + Copy + Debug,
-{
-    MonotonicMappingColumn {
-        from_column,
-        monotonic_mapping,
-        _phantom: PhantomData,
-    }
-}
-
-impl<C, T, Input, Output> Column<Output> for MonotonicMappingColumn<C, T, Input>
-where
-    C: Column<Input>,
-    T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
-    Input: PartialOrd + Send + Sync + Copy + Debug,
-    Output: PartialOrd + Send + Sync + Copy + Debug,
-{
-    #[inline]
-    fn get_val(&self, idx: u32) -> Output {
-        let from_val = self.from_column.get_val(idx);
-        self.monotonic_mapping.mapping(from_val)
-    }
-
-    fn min_value(&self) -> Output {
-        let from_min_value = self.from_column.min_value();
-        self.monotonic_mapping.mapping(from_min_value)
-    }
-
-    fn max_value(&self) -> Output {
-        let from_max_value = self.from_column.max_value();
-        self.monotonic_mapping.mapping(from_max_value)
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.from_column.num_vals()
-    }
-
-    fn iter(&self) -> Box<dyn Iterator<Item = Output> + '_> {
-        Box::new(
-            self.from_column
-                .iter()
-                .map(|el| self.monotonic_mapping.mapping(el)),
-        )
-    }
-
-    fn get_docids_for_value_range(
-        &self,
-        range: RangeInclusive<Output>,
-        doc_id_range: Range<u32>,
-        positions: &mut Vec<u32>,
-    ) {
-        if range.start() > &self.max_value() || range.end() < &self.min_value() {
-            return;
-        }
-        let range = self.monotonic_mapping.inverse_coerce(range);
-        if range.start() > range.end() {
-            return;
-        }
-        self.from_column
-            .get_docids_for_value_range(range, doc_id_range, positions)
-    }
-
-    // We voluntarily do not implement get_range as it yields a regression,
-    // and we do not have any specialized implementation anyway.
-}
-
-/// Wraps an iterator into a `Column`.
-pub struct IterColumn<T>(T);
-
-impl<T> From<T> for IterColumn<T>
-where T: Iterator + Clone + ExactSizeIterator
-{
-    fn from(iter: T) -> Self {
-        IterColumn(iter)
-    }
-}
-
-impl<T> Column<T::Item> for IterColumn<T>
-where
-    T: Iterator + Clone + ExactSizeIterator + Send + Sync,
-    T::Item: PartialOrd + fmt::Debug,
-{
-    fn get_val(&self, idx: u32) -> T::Item {
-        self.0.clone().nth(idx as usize).unwrap()
-    }
-
-    fn min_value(&self) -> T::Item {
-        self.0.clone().next().unwrap()
-    }
-
-    fn max_value(&self) -> T::Item {
-        self.0.clone().last().unwrap()
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.0.len() as u32
-    }
-
-    fn iter(&self) -> Box<dyn Iterator<Item = T::Item> + '_> {
-        Box::new(self.0.clone())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::monotonic_mapping::{
-        StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternalBaseval,
-        StrictlyMonotonicMappingToInternalGCDBaseval,
-    };
-
-    #[test]
-    fn test_monotonic_mapping() {
-        let vals = &[3u64, 5u64][..];
-        let col = VecColumn::from(vals);
-        let mapped = monotonic_map_column(col, StrictlyMonotonicMappingToInternalBaseval::new(2));
-        assert_eq!(mapped.min_value(), 1u64);
-        assert_eq!(mapped.max_value(), 3u64);
-        assert_eq!(mapped.num_vals(), 2);
-        assert_eq!(mapped.num_vals(), 2);
-        assert_eq!(mapped.get_val(0), 1);
-        assert_eq!(mapped.get_val(1), 3);
-    }
-
-    #[test]
-    fn test_range_as_col() {
-        let col = IterColumn::from(10..100);
-        assert_eq!(col.num_vals(), 90);
-        assert_eq!(col.max_value(), 99);
-    }
-
-    #[test]
-    fn test_monotonic_mapping_iter() {
-        let vals: Vec<u64> = (10..110u64).map(|el| el * 10).collect();
-        let col = VecColumn::from(&vals);
-        let mapped = monotonic_map_column(
-            col,
-            StrictlyMonotonicMappingInverter::from(
-                StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100),
-            ),
-        );
-        let val_i64s: Vec<u64> = mapped.iter().collect();
-        for i in 0..100 {
-            assert_eq!(val_i64s[i as usize], mapped.get_val(i));
-        }
-    }
-
-    #[test]
-    fn test_monotonic_mapping_get_range() {
-        let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
-        let col = VecColumn::from(&vals);
-        let mapped = monotonic_map_column(
-            col,
-            StrictlyMonotonicMappingInverter::from(
-                StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 0),
-            ),
-        );
-
-        assert_eq!(mapped.min_value(), 0u64);
-        assert_eq!(mapped.max_value(), 9900u64);
-        assert_eq!(mapped.num_vals(), 100);
-        let val_u64s: Vec<u64> = mapped.iter().collect();
-        assert_eq!(val_u64s.len(), 100);
-        for i in 0..100 {
-            assert_eq!(val_u64s[i as usize], mapped.get_val(i));
-            assert_eq!(val_u64s[i as usize], vals[i as usize] * 10);
-        }
-        let mut buf = [0u64; 20];
-        mapped.get_range(7, &mut buf[..]);
-        assert_eq!(&val_u64s[7..][..20], &buf);
-    }
-}

fastfield_codecs/src/compact_space/blank_range.rs

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

fastfield_codecs/src/compact_space/build_compact_space.rs

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

fastfield_codecs/src/compact_space/mod.rs

@@ -1,815 +0,0 @@
-/// This codec takes a large number space (u128) and reduces it to a compact number space.
-///
-/// It will find spaces in the number range. For example:
-///
-/// 100, 101, 102, 103, 104, 50000, 50001
-/// could be mapped to
-/// 100..104 -> 0..4
-/// 50000..50001 -> 5..6
-///
-/// Compact space 0..=6 requires much less bits than 100..=50001
-///
-/// The codec is created to compress ip addresses, but may be employed in other use cases.
-use std::{
-    cmp::Ordering,
-    collections::BTreeSet,
-    io::{self, Write},
-    ops::{Range, RangeInclusive},
-};
-
-use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
-use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
-
-use crate::compact_space::build_compact_space::get_compact_space;
-use crate::Column;
-
-mod blank_range;
-mod build_compact_space;
-
-/// The cost per blank is quite hard actually, since blanks are delta encoded, the actual cost of
-/// blanks depends on the number of blanks.
-///
-/// The number is taken by looking at a real dataset. It is optimized for larger datasets.
-const COST_PER_BLANK_IN_BITS: usize = 36;
-
-#[derive(Debug, Clone, Eq, PartialEq)]
-pub struct CompactSpace {
-    ranges_mapping: Vec<RangeMapping>,
-}
-
-/// Maps the range from the original space to compact_start + range.len()
-#[derive(Debug, Clone, Eq, PartialEq)]
-struct RangeMapping {
-    value_range: RangeInclusive<u128>,
-    compact_start: u64,
-}
-impl RangeMapping {
-    fn range_length(&self) -> u64 {
-        (self.value_range.end() - self.value_range.start()) as u64 + 1
-    }
-
-    // The last value of the compact space in this range
-    fn compact_end(&self) -> u64 {
-        self.compact_start + self.range_length() - 1
-    }
-}
-
-impl BinarySerializable for CompactSpace {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
-        VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
-
-        let mut prev_value = 0;
-        for value_range in self
-            .ranges_mapping
-            .iter()
-            .map(|range_mapping| &range_mapping.value_range)
-        {
-            let blank_delta_start = value_range.start() - prev_value;
-            VIntU128(blank_delta_start).serialize(writer)?;
-            prev_value = *value_range.start();
-
-            let blank_delta_end = value_range.end() - prev_value;
-            VIntU128(blank_delta_end).serialize(writer)?;
-            prev_value = *value_range.end();
-        }
-
-        Ok(())
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let num_ranges = VInt::deserialize(reader)?.0;
-        let mut ranges_mapping: Vec<RangeMapping> = vec![];
-        let mut value = 0u128;
-        let mut compact_start = 1u64; // 0 is reserved for `null`
-        for _ in 0..num_ranges {
-            let blank_delta_start = VIntU128::deserialize(reader)?.0;
-            value += blank_delta_start;
-            let blank_start = value;
-
-            let blank_delta_end = VIntU128::deserialize(reader)?.0;
-            value += blank_delta_end;
-            let blank_end = value;
-
-            let range_mapping = RangeMapping {
-                value_range: blank_start..=blank_end,
-                compact_start,
-            };
-            let range_length = range_mapping.range_length();
-            ranges_mapping.push(range_mapping);
-            compact_start += range_length;
-        }
-
-        Ok(Self { ranges_mapping })
-    }
-}
-
-impl CompactSpace {
-    /// Amplitude is the value range of the compact space including the sentinel value used to
-    /// identify null values. The compact space is 0..=amplitude .
-    ///
-    /// It's only used to verify we don't exceed u64 number space, which would indicate a bug.
-    fn amplitude_compact_space(&self) -> u128 {
-        self.ranges_mapping
-            .last()
-            .map(|last_range| last_range.compact_end() as u128)
-            .unwrap_or(1) // compact space starts at 1, 0 == null
-    }
-
-    fn get_range_mapping(&self, pos: usize) -> &RangeMapping {
-        &self.ranges_mapping[pos]
-    }
-
-    /// Returns either Ok(the value in the compact space) or if it is outside the compact space the
-    /// Err(position where it would be inserted)
-    fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
-        self.ranges_mapping
-            .binary_search_by(|probe| {
-                let value_range = &probe.value_range;
-                if value < *value_range.start() {
-                    Ordering::Greater
-                } else if value > *value_range.end() {
-                    Ordering::Less
-                } else {
-                    Ordering::Equal
-                }
-            })
-            .map(|pos| {
-                let range_mapping = &self.ranges_mapping[pos];
-                let pos_in_range = (value - range_mapping.value_range.start()) as u64;
-                range_mapping.compact_start + pos_in_range
-            })
-    }
-
-    /// Unpacks a value from compact space u64 to u128 space
-    fn compact_to_u128(&self, compact: u64) -> u128 {
-        let pos = self
-            .ranges_mapping
-            .binary_search_by_key(&compact, |range_mapping| range_mapping.compact_start)
-            // Correctness: Overflow. The first range starts at compact space 0, the error from
-            // binary search can never be 0
-            .map_or_else(|e| e - 1, |v| v);
-
-        let range_mapping = &self.ranges_mapping[pos];
-        let diff = compact - range_mapping.compact_start;
-        range_mapping.value_range.start() + diff as u128
-    }
-}
-
-pub struct CompactSpaceCompressor {
-    params: IPCodecParams,
-}
-#[derive(Debug, Clone)]
-pub struct IPCodecParams {
-    compact_space: CompactSpace,
-    bit_unpacker: BitUnpacker,
-    min_value: u128,
-    max_value: u128,
-    num_vals: u32,
-    num_bits: u8,
-}
-
-impl CompactSpaceCompressor {
-    /// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
-    pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
-        let mut values_sorted = BTreeSet::new();
-        values_sorted.extend(iter);
-        let total_num_values = num_vals;
-
-        let compact_space =
-            get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
-        let amplitude_compact_space = compact_space.amplitude_compact_space();
-
-        assert!(
-            amplitude_compact_space <= u64::MAX as u128,
-            "case unsupported."
-        );
-
-        let num_bits = tantivy_bitpacker::compute_num_bits(amplitude_compact_space as u64);
-        let min_value = *values_sorted.iter().next().unwrap_or(&0);
-        let max_value = *values_sorted.iter().last().unwrap_or(&0);
-        assert_eq!(
-            compact_space
-                .u128_to_compact(max_value)
-                .expect("could not convert max value to compact space"),
-            amplitude_compact_space as u64
-        );
-        CompactSpaceCompressor {
-            params: IPCodecParams {
-                compact_space,
-                bit_unpacker: BitUnpacker::new(num_bits),
-                min_value,
-                max_value,
-                num_vals: total_num_values,
-                num_bits,
-            },
-        }
-    }
-
-    fn write_footer(self, writer: &mut impl Write) -> io::Result<()> {
-        let writer = &mut CountingWriter::wrap(writer);
-        self.params.serialize(writer)?;
-
-        let footer_len = writer.written_bytes() as u32;
-        footer_len.serialize(writer)?;
-
-        Ok(())
-    }
-
-    pub fn compress_into(
-        self,
-        vals: impl Iterator<Item = u128>,
-        write: &mut impl Write,
-    ) -> io::Result<()> {
-        let mut bitpacker = BitPacker::default();
-        for val in vals {
-            let compact = self
-                .params
-                .compact_space
-                .u128_to_compact(val)
-                .map_err(|_| {
-                    io::Error::new(
-                        io::ErrorKind::InvalidData,
-                        "Could not convert value to compact_space. This is a bug.",
-                    )
-                })?;
-            bitpacker.write(compact, self.params.num_bits, write)?;
-        }
-        bitpacker.close(write)?;
-        self.write_footer(write)?;
-        Ok(())
-    }
-}
-
-#[derive(Debug, Clone)]
-pub struct CompactSpaceDecompressor {
-    data: OwnedBytes,
-    params: IPCodecParams,
-}
-
-impl BinarySerializable for IPCodecParams {
-    fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
-        // header flags for future optional dictionary encoding
-        let footer_flags = 0u64;
-        footer_flags.serialize(writer)?;
-
-        VIntU128(self.min_value).serialize(writer)?;
-        VIntU128(self.max_value).serialize(writer)?;
-        VIntU128(self.num_vals as u128).serialize(writer)?;
-        self.num_bits.serialize(writer)?;
-
-        self.compact_space.serialize(writer)?;
-
-        Ok(())
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let _header_flags = u64::deserialize(reader)?;
-        let min_value = VIntU128::deserialize(reader)?.0;
-        let max_value = VIntU128::deserialize(reader)?.0;
-        let num_vals = VIntU128::deserialize(reader)?.0 as u32;
-        let num_bits = u8::deserialize(reader)?;
-        let compact_space = CompactSpace::deserialize(reader)?;
-
-        Ok(Self {
-            compact_space,
-            bit_unpacker: BitUnpacker::new(num_bits),
-            min_value,
-            max_value,
-            num_vals,
-            num_bits,
-        })
-    }
-}
-
-impl Column<u128> for CompactSpaceDecompressor {
-    #[inline]
-    fn get_val(&self, doc: u32) -> u128 {
-        self.get(doc)
-    }
-
-    fn min_value(&self) -> u128 {
-        self.min_value()
-    }
-
-    fn max_value(&self) -> u128 {
-        self.max_value()
-    }
-
-    fn num_vals(&self) -> u32 {
-        self.params.num_vals
-    }
-
-    #[inline]
-    fn iter(&self) -> Box<dyn Iterator<Item = u128> + '_> {
-        Box::new(self.iter())
-    }
-
-    #[inline]
-    fn get_docids_for_value_range(
-        &self,
-        value_range: RangeInclusive<u128>,
-        positions_range: Range<u32>,
-        positions: &mut Vec<u32>,
-    ) {
-        self.get_positions_for_value_range(value_range, positions_range, positions)
-    }
-}
-
-impl CompactSpaceDecompressor {
-    pub fn open(data: OwnedBytes) -> io::Result<CompactSpaceDecompressor> {
-        let (data_slice, footer_len_bytes) = data.split_at(data.len() - 4);
-        let footer_len = u32::deserialize(&mut &footer_len_bytes[..])?;
-
-        let data_footer = &data_slice[data_slice.len() - footer_len as usize..];
-        let params = IPCodecParams::deserialize(&mut &data_footer[..])?;
-        let decompressor = CompactSpaceDecompressor { data, params };
-
-        Ok(decompressor)
-    }
-
-    /// Converting to compact space for the decompressor is more complex, since we may get values
-    /// which are outside the compact space. e.g. if we map
-    /// 1000 => 5
-    /// 2000 => 6
-    ///
-    /// and we want a mapping for 1005, there is no equivalent compact space. We instead return an
-    /// error with the index of the next range.
-    fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
-        self.params.compact_space.u128_to_compact(value)
-    }
-
-    fn compact_to_u128(&self, compact: u64) -> u128 {
-        self.params.compact_space.compact_to_u128(compact)
-    }
-
-    /// Comparing on compact space: Random dataset 0,24 (50% random hit) - 1.05 GElements/s
-    /// Comparing on compact space: Real dataset 1.08 GElements/s
-    ///
-    /// Comparing on original space: Real dataset .06 GElements/s (not completely optimized)
-    #[inline]
-    pub fn get_positions_for_value_range(
-        &self,
-        value_range: RangeInclusive<u128>,
-        position_range: Range<u32>,
-        positions: &mut Vec<u32>,
-    ) {
-        if value_range.start() > value_range.end() {
-            return;
-        }
-        let position_range = position_range.start..position_range.end.min(self.num_vals());
-        let from_value = *value_range.start();
-        let to_value = *value_range.end();
-        assert!(to_value >= from_value);
-        let compact_from = self.u128_to_compact(from_value);
-        let compact_to = self.u128_to_compact(to_value);
-
-        // Quick return, if both ranges fall into the same non-mapped space, the range can't cover
-        // any values, so we can early exit
-        match (compact_to, compact_from) {
-            (Err(pos1), Err(pos2)) if pos1 == pos2 => return,
-            _ => {}
-        }
-
-        let compact_from = compact_from.unwrap_or_else(|pos| {
-            // Correctness: Out of bounds, if this value is Err(last_index + 1), we early exit,
-            // since the to_value also mapps into the same non-mapped space
-            let range_mapping = self.params.compact_space.get_range_mapping(pos);
-            range_mapping.compact_start
-        });
-        // If there is no compact space, we go to the closest upperbound compact space
-        let compact_to = compact_to.unwrap_or_else(|pos| {
-            // Correctness: Overflow, if this value is Err(0), we early exit,
-            // since the from_value also mapps into the same non-mapped space
-
-            // Get end of previous range
-            let pos = pos - 1;
-            let range_mapping = self.params.compact_space.get_range_mapping(pos);
-            range_mapping.compact_end()
-        });
-
-        let range = compact_from..=compact_to;
-
-        let scan_num_docs = position_range.end - position_range.start;
-
-        let step_size = 4;
-        let cutoff = position_range.start + scan_num_docs - scan_num_docs % step_size;
-
-        let mut push_if_in_range = |idx, val| {
-            if range.contains(&val) {
-                positions.push(idx);
-            }
-        };
-        let get_val = |idx| self.params.bit_unpacker.get(idx, &self.data);
-        // unrolled loop
-        for idx in (position_range.start..cutoff).step_by(step_size as usize) {
-            let idx1 = idx;
-            let idx2 = idx + 1;
-            let idx3 = idx + 2;
-            let idx4 = idx + 3;
-            let val1 = get_val(idx1);
-            let val2 = get_val(idx2);
-            let val3 = get_val(idx3);
-            let val4 = get_val(idx4);
-            push_if_in_range(idx1, val1);
-            push_if_in_range(idx2, val2);
-            push_if_in_range(idx3, val3);
-            push_if_in_range(idx4, val4);
-        }
-
-        // handle rest
-        for idx in cutoff..position_range.end {
-            push_if_in_range(idx, get_val(idx));
-        }
-    }
-
-    #[inline]
-    fn iter_compact(&self) -> impl Iterator<Item = u64> + '_ {
-        (0..self.params.num_vals).map(move |idx| self.params.bit_unpacker.get(idx, &self.data))
-    }
-
-    #[inline]
-    fn iter(&self) -> impl Iterator<Item = u128> + '_ {
-        // TODO: Performance. It would be better to iterate on the ranges and check existence via
-        // the bit_unpacker.
-        self.iter_compact()
-            .map(|compact| self.compact_to_u128(compact))
-    }
-
-    #[inline]
-    pub fn get(&self, idx: u32) -> u128 {
-        let compact = self.params.bit_unpacker.get(idx, &self.data);
-        self.compact_to_u128(compact)
-    }
-
-    pub fn min_value(&self) -> u128 {
-        self.params.min_value
-    }
-
-    pub fn max_value(&self) -> u128 {
-        self.params.max_value
-    }
-}
-
-#[cfg(test)]
-mod tests {
-
-    use std::fmt;
-
-    use super::*;
-    use crate::format_version::read_format_version;
-    use crate::null_index_footer::read_null_index_footer;
-    use crate::serialize::U128Header;
-    use crate::{open_u128, serialize_u128};
-
-    #[test]
-    fn compact_space_test() {
-        let ips = &[
-            2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
-        ]
-        .into_iter()
-        .collect();
-        let compact_space = get_compact_space(ips, ips.len() as u32, 11);
-        let amplitude = compact_space.amplitude_compact_space();
-        assert_eq!(amplitude, 17);
-        assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
-        assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
-        assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
-
-        for (num1, num2) in (0..3).tuple_windows() {
-            assert_eq!(
-                compact_space.get_range_mapping(num1).compact_end() + 1,
-                compact_space.get_range_mapping(num2).compact_start
-            );
-        }
-
-        let mut output: Vec<u8> = Vec::new();
-        compact_space.serialize(&mut output).unwrap();
-
-        assert_eq!(
-            compact_space,
-            CompactSpace::deserialize(&mut &output[..]).unwrap()
-        );
-
-        for ip in ips {
-            let compact = compact_space.u128_to_compact(*ip).unwrap();
-            assert_eq!(compact_space.compact_to_u128(compact), *ip);
-        }
-    }
-
-    #[test]
-    fn compact_space_amplitude_test() {
-        let ips = &[100000u128, 1000000].into_iter().collect();
-        let compact_space = get_compact_space(ips, ips.len() as u32, 1);
-        let amplitude = compact_space.amplitude_compact_space();
-        assert_eq!(amplitude, 2);
-    }
-
-    fn test_all(mut data: OwnedBytes, expected: &[u128]) {
-        let _header = U128Header::deserialize(&mut data);
-        let decompressor = CompactSpaceDecompressor::open(data).unwrap();
-        for (idx, expected_val) in expected.iter().cloned().enumerate() {
-            let val = decompressor.get(idx as u32);
-            assert_eq!(val, expected_val);
-
-            let test_range = |range: RangeInclusive<u128>| {
-                let expected_positions = expected
-                    .iter()
-                    .positions(|val| range.contains(val))
-                    .map(|pos| pos as u32)
-                    .collect::<Vec<_>>();
-                let mut positions = Vec::new();
-                decompressor.get_positions_for_value_range(
-                    range,
-                    0..decompressor.num_vals(),
-                    &mut positions,
-                );
-                assert_eq!(positions, expected_positions);
-            };
-
-            test_range(expected_val.saturating_sub(1)..=expected_val);
-            test_range(expected_val..=expected_val);
-            test_range(expected_val..=expected_val.saturating_add(1));
-            test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
-        }
-    }
-
-    fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
-        let mut out = Vec::new();
-        serialize_u128(
-            || u128_vals.iter().cloned(),
-            u128_vals.len() as u32,
-            &mut out,
-        )
-        .unwrap();
-
-        let data = OwnedBytes::new(out);
-        let (data, _format_version) = read_format_version(data).unwrap();
-        let (data, _null_index_footer) = read_null_index_footer(data).unwrap();
-        test_all(data.clone(), u128_vals);
-
-        data
-    }
-
-    #[test]
-    fn test_range_1() {
-        let vals = &[
-            1u128,
-            100u128,
-            3u128,
-            99999u128,
-            100000u128,
-            100001u128,
-            4_000_211_221u128,
-            4_000_211_222u128,
-            333u128,
-        ];
-        let mut data = test_aux_vals(vals);
-
-        let _header = U128Header::deserialize(&mut data);
-        let decomp = CompactSpaceDecompressor::open(data).unwrap();
-        let complete_range = 0..vals.len() as u32;
-        for (pos, val) in vals.iter().enumerate() {
-            let val = *val;
-            let pos = pos as u32;
-            let mut positions = Vec::new();
-            decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
-            assert_eq!(positions, vec![pos]);
-        }
-
-        // handle docid range out of bounds
-        let positions = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
-        assert_eq!(positions, vec![]);
-
-        let positions =
-            get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
-        assert_eq!(positions, vec![0]);
-        let positions =
-            get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
-        assert_eq!(positions, vec![0]);
-        let positions =
-            get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
-        assert_eq!(positions, vec![0, 2]);
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                99999u128..=99999u128,
-                complete_range.clone()
-            ),
-            vec![3]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                99999u128..=100000u128,
-                complete_range.clone()
-            ),
-            vec![3, 4]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                99998u128..=100000u128,
-                complete_range.clone()
-            ),
-            vec![3, 4]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                99998u128..=99999u128,
-                complete_range.clone()
-            ),
-            vec![3]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                99998u128..=99998u128,
-                complete_range.clone()
-            ),
-            vec![]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                333u128..=333u128,
-                complete_range.clone()
-            ),
-            vec![8]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                332u128..=333u128,
-                complete_range.clone()
-            ),
-            vec![8]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                332u128..=334u128,
-                complete_range.clone()
-            ),
-            vec![8]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                333u128..=334u128,
-                complete_range.clone()
-            ),
-            vec![8]
-        );
-
-        assert_eq!(
-            get_positions_for_value_range_helper(
-                &decomp,
-                4_000_211_221u128..=5_000_000_000u128,
-                complete_range
-            ),
-            vec![6, 7]
-        );
-    }
-
-    #[test]
-    fn test_empty() {
-        let vals = &[];
-        let data = test_aux_vals(vals);
-        let _decomp = CompactSpaceDecompressor::open(data).unwrap();
-    }
-
-    #[test]
-    fn test_range_2() {
-        let vals = &[
-            100u128,
-            99999u128,
-            100000u128,
-            100001u128,
-            4_000_211_221u128,
-            4_000_211_222u128,
-            333u128,
-        ];
-        let mut data = test_aux_vals(vals);
-        let _header = U128Header::deserialize(&mut data);
-        let decomp = CompactSpaceDecompressor::open(data).unwrap();
-        let complete_range = 0..vals.len() as u32;
-        assert_eq!(
-            get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone()),
-            vec![]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
-            vec![0]
-        );
-        assert_eq!(
-            get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
-            vec![0]
-        );
-    }
-
-    fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd + fmt::Debug>(
-        column: &C,
-        value_range: RangeInclusive<T>,
-        doc_id_range: Range<u32>,
-    ) -> Vec<u32> {
-        let mut positions = Vec::new();
-        column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
-        positions
-    }
-
-    #[test]
-    fn test_range_3() {
-        let vals = &[
-            200u128,
-            201,
-            202,
-            203,
-            204,
-            204,
-            206,
-            207,
-            208,
-            209,
-            210,
-            1_000_000,
-            5_000_000_000,
-        ];
-        let mut out = Vec::new();
-        serialize_u128(|| vals.iter().cloned(), vals.len() as u32, &mut out).unwrap();
-        let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
-        let complete_range = 0..vals.len() as u32;
-
-        assert_eq!(
-            get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
-            vec![0]
-        );
-
-        assert_eq!(
-            get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
-            vec![0, 1]
-        );
-
-        assert_eq!(
-            get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
-            vec![0]
-        );
-
-        assert_eq!(
-            get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
-            vec![11]
-        );
-    }
-
-    #[test]
-    fn test_bug1() {
-        let vals = &[9223372036854775806];
-        let _data = test_aux_vals(vals);
-    }
-
-    #[test]
-    fn test_bug2() {
-        let vals = &[340282366920938463463374607431768211455u128];
-        let _data = test_aux_vals(vals);
-    }
-
-    #[test]
-    fn test_bug3() {
-        let vals = &[340282366920938463463374607431768211454];
-        let _data = test_aux_vals(vals);
-    }
-
-    #[test]
-    fn test_bug4() {
-        let vals = &[340282366920938463463374607431768211455, 0];
-        let _data = test_aux_vals(vals);
-    }
-
-    #[test]
-    fn test_first_large_gaps() {
-        let vals = &[1_000_000_000u128; 100];
-        let _data = test_aux_vals(vals);
-    }
-    use itertools::Itertools;
-    use proptest::prelude::*;
-
-    fn num_strategy() -> impl Strategy<Value = u128> {
-        prop_oneof![
-            1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
-            1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
-            1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
-            1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
-            20 => prop::num::u128::ANY,
-        ]
-    }
-
-    proptest! {
-        #![proptest_config(ProptestConfig::with_cases(10))]
-
-            #[test]
-            fn compress_decompress_random(vals in proptest::collection::vec(num_strategy()
-    , 1..1000)) {
-                let _data = test_aux_vals(&vals);
-            }
-        }
-}

fastfield_codecs/src/format_version.rs

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

fastfield_codecs/src/gcd.rs

@@ -1,170 +0,0 @@
-use std::num::NonZeroU64;
-
-use fastdivide::DividerU64;
-
-/// Compute the gcd of two non null numbers.
-///
-/// It is recommended, but not required, to feed values such that `large >= small`.
-fn compute_gcd(mut large: NonZeroU64, mut small: NonZeroU64) -> NonZeroU64 {
-    loop {
-        let rem: u64 = large.get() % small;
-        if let Some(new_small) = NonZeroU64::new(rem) {
-            (large, small) = (small, new_small);
-        } else {
-            return small;
-        }
-    }
-}
-
-// Find GCD for iterator of numbers
-pub fn find_gcd(numbers: impl Iterator<Item = u64>) -> Option<NonZeroU64> {
-    let mut numbers = numbers.flat_map(NonZeroU64::new);
-    let mut gcd: NonZeroU64 = numbers.next()?;
-    if gcd.get() == 1 {
-        return Some(gcd);
-    }
-
-    let mut gcd_divider = DividerU64::divide_by(gcd.get());
-    for val in numbers {
-        let remainder = val.get() - (gcd_divider.divide(val.get())) * gcd.get();
-        if remainder == 0 {
-            continue;
-        }
-        gcd = compute_gcd(val, gcd);
-        if gcd.get() == 1 {
-            return Some(gcd);
-        }
-
-        gcd_divider = DividerU64::divide_by(gcd.get());
-    }
-    Some(gcd)
-}
-
-#[cfg(test)]
-mod tests {
-    use std::io;
-    use std::num::NonZeroU64;
-
-    use common::OwnedBytes;
-
-    use crate::gcd::{compute_gcd, find_gcd};
-    use crate::{FastFieldCodecType, VecColumn};
-
-    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::serialize(VecColumn::from(&vals), &mut buffer, &[codec_type])?;
-        let buffer = OwnedBytes::new(buffer);
-        let column = crate::open::<i64>(buffer.clone())?;
-        assert_eq!(column.get_val(0), -4000i64);
-        assert_eq!(column.get_val(1), -3000i64);
-        assert_eq!(column.get_val(2), -2000i64);
-        assert_eq!(column.max_value(), (num_vals as i64 - 5) * 1000);
-        assert_eq!(column.min_value(), -4000i64);
-
-        // Can't apply gcd
-        let mut buffer_without_gcd = Vec::new();
-        vals.pop();
-        vals.push(1001i64);
-        crate::serialize(
-            VecColumn::from(&vals),
-            &mut buffer_without_gcd,
-            &[codec_type],
-        )?;
-        let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
-        assert!(buffer_without_gcd.len() > buffer.len());
-
-        Ok(())
-    }
-
-    #[test]
-    fn test_fastfield_gcd_i64() -> io::Result<()> {
-        for &codec_type in &[
-            FastFieldCodecType::Bitpacked,
-            FastFieldCodecType::BlockwiseLinear,
-            FastFieldCodecType::Linear,
-        ] {
-            test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
-        }
-        Ok(())
-    }
-
-    fn test_fastfield_gcd_u64_with_codec(
-        codec_type: FastFieldCodecType,
-        num_vals: usize,
-    ) -> io::Result<()> {
-        let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
-        let mut buffer: Vec<u8> = Vec::new();
-        crate::serialize(VecColumn::from(&vals), &mut buffer, &[codec_type])?;
-        let buffer = OwnedBytes::new(buffer);
-        let column = crate::open::<u64>(buffer.clone())?;
-        assert_eq!(column.get_val(0), 1000u64);
-        assert_eq!(column.get_val(1), 2000u64);
-        assert_eq!(column.get_val(2), 3000u64);
-        assert_eq!(column.max_value(), num_vals as u64 * 1000);
-        assert_eq!(column.min_value(), 1000u64);
-
-        // Can't apply gcd
-        let mut buffer_without_gcd = Vec::new();
-        vals.pop();
-        vals.push(1001u64);
-        crate::serialize(
-            VecColumn::from(&vals),
-            &mut buffer_without_gcd,
-            &[codec_type],
-        )?;
-        let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
-        assert!(buffer_without_gcd.len() > buffer.len());
-        Ok(())
-    }
-
-    #[test]
-    fn test_fastfield_gcd_u64() -> io::Result<()> {
-        for &codec_type in &[
-            FastFieldCodecType::Bitpacked,
-            FastFieldCodecType::BlockwiseLinear,
-            FastFieldCodecType::Linear,
-        ] {
-            test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
-        }
-        Ok(())
-    }
-
-    #[test]
-    pub fn test_fastfield2() {
-        let test_fastfield = crate::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);
-    }
-
-    #[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);
-    }
-}

fastfield_codecs/src/lib.rs

@@ -1,568 +0,0 @@
-#![warn(missing_docs)]
-#![cfg_attr(all(feature = "unstable", test), feature(test))]
-
-//! # `fastfield_codecs`
-//!
-//! - Columnar storage of data for tantivy [`Column`].
-//! - Encode data in different codecs.
-//! - Monotonically map values to u64/u128
-
-#[cfg(test)]
-#[macro_use]
-extern crate more_asserts;
-
-#[cfg(all(test, feature = "unstable"))]
-extern crate test;
-
-use std::io::Write;
-use std::sync::Arc;
-use std::{fmt, io};
-
-use common::{BinarySerializable, OwnedBytes};
-use compact_space::CompactSpaceDecompressor;
-use format_version::read_format_version;
-use monotonic_mapping::{
-    StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
-    StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
-};
-use null_index_footer::read_null_index_footer;
-use serialize::{Header, U128Header};
-
-mod bitpacked;
-mod blockwise_linear;
-mod compact_space;
-mod format_version;
-mod line;
-mod linear;
-mod monotonic_mapping;
-mod monotonic_mapping_u128;
-#[allow(dead_code)]
-mod null_index;
-mod null_index_footer;
-
-mod column;
-mod gcd;
-pub mod serialize;
-
-use self::bitpacked::BitpackedCodec;
-use self::blockwise_linear::BlockwiseLinearCodec;
-pub use self::column::{monotonic_map_column, Column, IterColumn, VecColumn};
-use self::linear::LinearCodec;
-pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
-pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
-pub use self::serialize::{
-    estimate, serialize, serialize_and_load, serialize_u128, NormalizedHeader,
-};
-
-#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
-#[repr(u8)]
-/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
-pub enum FastFieldCodecType {
-    /// Bitpack all values in the value range. The number of bits is defined by the amplitude
-    /// `column.max_value() - column.min_value()`
-    Bitpacked = 1,
-    /// Linear interpolation puts a line between the first and last value and then bitpacks the
-    /// values by the offset from the line. The number of bits is defined by the max deviation from
-    /// the line.
-    Linear = 2,
-    /// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
-    BlockwiseLinear = 3,
-}
-
-impl BinarySerializable for FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
-        self.to_code().serialize(wrt)
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let code = u8::deserialize(reader)?;
-        let codec_type: Self = Self::from_code(code)
-            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
-        Ok(codec_type)
-    }
-}
-
-impl FastFieldCodecType {
-    pub(crate) fn to_code(self) -> u8 {
-        self as u8
-    }
-
-    pub(crate) fn from_code(code: u8) -> Option<Self> {
-        match code {
-            1 => Some(Self::Bitpacked),
-            2 => Some(Self::Linear),
-            3 => Some(Self::BlockwiseLinear),
-            _ => None,
-        }
-    }
-}
-
-#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
-#[repr(u8)]
-/// Available codecs to use to encode the u128 (via [`MonotonicallyMappableToU128`]) converted data.
-pub enum U128FastFieldCodecType {
-    /// This codec takes a large number space (u128) and reduces it to a compact number space, by
-    /// removing the holes.
-    CompactSpace = 1,
-}
-
-impl BinarySerializable for U128FastFieldCodecType {
-    fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
-        self.to_code().serialize(wrt)
-    }
-
-    fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
-        let code = u8::deserialize(reader)?;
-        let codec_type: Self = Self::from_code(code)
-            .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
-        Ok(codec_type)
-    }
-}
-
-impl U128FastFieldCodecType {
-    pub(crate) fn to_code(self) -> u8 {
-        self as u8
-    }
-
-    pub(crate) fn from_code(code: u8) -> Option<Self> {
-        match code {
-            1 => Some(Self::CompactSpace),
-            _ => None,
-        }
-    }
-}
-
-/// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open_u128<Item: MonotonicallyMappableToU128 + fmt::Debug>(
-    bytes: OwnedBytes,
-) -> io::Result<Arc<dyn Column<Item>>> {
-    let (bytes, _format_version) = read_format_version(bytes)?;
-    let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
-    let header = U128Header::deserialize(&mut bytes)?;
-    assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
-    let reader = CompactSpaceDecompressor::open(bytes)?;
-    let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<Item>> =
-        StrictlyMonotonicMappingToInternal::<Item>::new().into();
-    Ok(Arc::new(monotonic_map_column(reader, inverted)))
-}
-
-/// Returns the correct codec reader wrapped in the `Arc` for the data.
-pub fn open<T: MonotonicallyMappableToU64 + fmt::Debug>(
-    bytes: OwnedBytes,
-) -> io::Result<Arc<dyn Column<T>>> {
-    let (bytes, _format_version) = read_format_version(bytes)?;
-    let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
-    let header = Header::deserialize(&mut bytes)?;
-    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 + fmt::Debug>(
-    bytes: OwnedBytes,
-    header: &Header,
-) -> io::Result<Arc<dyn Column<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.
-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: Column<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 Column, 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 Column) -> Option<f32>;
-}
-
-/// The list of all available codecs for u64 convertible data.
-pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
-    FastFieldCodecType::Bitpacked,
-    FastFieldCodecType::BlockwiseLinear,
-    FastFieldCodecType::Linear,
-];
-
-#[cfg(test)]
-mod tests {
-
-    use proptest::prelude::*;
-    use proptest::strategy::Strategy;
-    use proptest::{prop_oneof, proptest};
-
-    use crate::bitpacked::BitpackedCodec;
-    use crate::blockwise_linear::BlockwiseLinearCodec;
-    use crate::linear::LinearCodec;
-    use crate::serialize::Header;
-
-    pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
-        data: &[u64],
-        name: &str,
-    ) -> Option<(f32, f32)> {
-        let col = &VecColumn::from(data);
-        let header = Header::compute_header(col, &[Codec::CODEC_TYPE])?;
-        let normalized_col = header.normalize_column(col);
-        let estimation = Codec::estimate(&normalized_col)?;
-
-        let mut out = Vec::new();
-        let col = VecColumn::from(data);
-        serialize(col, &mut out, &[Codec::CODEC_TYPE]).unwrap();
-
-        let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
-
-        let reader = crate::open::<u64>(OwnedBytes::new(out)).unwrap();
-        assert_eq!(reader.num_vals(), data.len() as u32);
-        for (doc, orig_val) in data.iter().copied().enumerate() {
-            let val = reader.get_val(doc as u32);
-            assert_eq!(
-                val, orig_val,
-                "val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data \
-                 `{data:?}`",
-            );
-        }
-
-        if !data.is_empty() {
-            let test_rand_idx = rand::thread_rng().gen_range(0..=data.len() - 1);
-            let expected_positions: Vec<u32> = data
-                .iter()
-                .enumerate()
-                .filter(|(_, el)| **el == data[test_rand_idx])
-                .map(|(pos, _)| pos as u32)
-                .collect();
-            let mut positions = Vec::new();
-            reader.get_docids_for_value_range(
-                data[test_rand_idx]..=data[test_rand_idx],
-                0..data.len() as u32,
-                &mut positions,
-            );
-            assert_eq!(expected_positions, positions);
-        }
-        Some((estimation, actual_compression))
-    }
-
-    proptest! {
-        #![proptest_config(ProptestConfig::with_cases(100))]
-
-        #[test]
-        fn test_proptest_small_bitpacked(data in proptest::collection::vec(num_strategy(), 1..10)) {
-            create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
-        }
-
-        #[test]
-        fn test_proptest_small_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
-            create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
-        }
-
-        #[test]
-        fn test_proptest_small_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
-            create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
-        }
-    }
-
-    proptest! {
-        #![proptest_config(ProptestConfig::with_cases(10))]
-
-        #[test]
-        fn test_proptest_large_bitpacked(data in proptest::collection::vec(num_strategy(), 1..6000)) {
-            create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
-        }
-
-        #[test]
-        fn test_proptest_large_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
-            create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
-        }
-
-        #[test]
-        fn test_proptest_large_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
-            create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
-        }
-    }
-
-    fn num_strategy() -> impl Strategy<Value = u64> {
-        prop_oneof![
-            1 => prop::num::u64::ANY.prop_map(|num| u64::MAX - (num % 10) ),
-            1 => prop::num::u64::ANY.prop_map(|num| num % 10 ),
-            20 => prop::num::u64::ANY,
-        ]
-    }
-
-    pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
-        let mut data_and_names = vec![];
-
-        let data = vec![10];
-        data_and_names.push((data, "minimal test"));
-
-        let data = (10..=10_000_u64).collect::<Vec<_>>();
-        data_and_names.push((data, "simple monotonically increasing"));
-
-        data_and_names.push((
-            vec![5, 6, 7, 8, 9, 10, 99, 100],
-            "offset in linear interpol",
-        ));
-
-        data_and_names.push((vec![3, 18446744073709551613, 5], "docid range regression"));
-
-        data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
-        data_and_names.push((vec![10], "single value"));
-
-        data_and_names.push((
-            vec![1572656989877777, 1170935903116329, 720575940379279, 0],
-            "overflow error",
-        ));
-
-        data_and_names
-    }
-
-    fn test_codec<C: FastFieldCodec>() {
-        let codec_name = format!("{:?}", C::CODEC_TYPE);
-        for (data, dataset_name) in get_codec_test_datasets() {
-            let estimate_actual_opt: Option<(f32, f32)> =
-                crate::tests::create_and_validate::<C>(&data, dataset_name);
-            let result = if let Some((estimate, actual)) = estimate_actual_opt {
-                format!("Estimate `{estimate}` Actual `{actual}`")
-            } else {
-                "Disabled".to_string()
-            };
-            println!("Codec {codec_name}, DataSet {dataset_name}, {result}");
-        }
-    }
-    #[test]
-    fn test_codec_bitpacking() {
-        test_codec::<BitpackedCodec>();
-    }
-    #[test]
-    fn test_codec_interpolation() {
-        test_codec::<LinearCodec>();
-    }
-    #[test]
-    fn test_codec_multi_interpolation() {
-        test_codec::<BlockwiseLinearCodec>();
-    }
-
-    use super::*;
-
-    #[test]
-    fn estimation_good_interpolation_case() {
-        let data = (10..=20000_u64).collect::<Vec<_>>();
-        let data: VecColumn = data.as_slice().into();
-
-        let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-        assert_le!(linear_interpol_estimation, 0.01);
-
-        let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
-        assert_le!(multi_linear_interpol_estimation, 0.2);
-        assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
-
-        let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-        assert_lt!(linear_interpol_estimation, bitpacked_estimation);
-    }
-    #[test]
-    fn estimation_test_bad_interpolation_case() {
-        let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
-
-        let data: VecColumn = data.into();
-        let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-        assert_le!(linear_interpol_estimation, 0.34);
-
-        let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-        assert_lt!(bitpacked_estimation, linear_interpol_estimation);
-    }
-
-    #[test]
-    fn estimation_prefer_bitpacked() {
-        let data = VecColumn::from(&[10, 10, 10, 10]);
-        let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-        let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-        assert_lt!(bitpacked_estimation, linear_interpol_estimation);
-    }
-
-    #[test]
-    fn estimation_test_bad_interpolation_case_monotonically_increasing() {
-        let mut data: Vec<u64> = (201..=20000_u64).collect();
-        data.push(1_000_000);
-        let data: VecColumn = data.as_slice().into();
-
-        // in this case the linear interpolation can't in fact not be worse than bitpacking,
-        // but the estimator adds some threshold, which leads to estimated worse behavior
-        let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
-        assert_le!(linear_interpol_estimation, 0.35);
-
-        let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
-        assert_le!(bitpacked_estimation, 0.32);
-        assert_le!(bitpacked_estimation, linear_interpol_estimation);
-    }
-
-    #[test]
-    fn test_fast_field_codec_type_to_code() {
-        let mut count_codec = 0;
-        for code in 0..=255 {
-            if let Some(codec_type) = FastFieldCodecType::from_code(code) {
-                assert_eq!(codec_type.to_code(), code);
-                count_codec += 1;
-            }
-        }
-        assert_eq!(count_codec, 3);
-    }
-}
-
-#[cfg(all(test, feature = "unstable"))]
-mod bench {
-    use std::sync::Arc;
-
-    use common::OwnedBytes;
-    use rand::rngs::StdRng;
-    use rand::{Rng, SeedableRng};
-    use test::{self, Bencher};
-
-    use super::*;
-    use crate::Column;
-
-    fn get_data() -> Vec<u64> {
-        let mut rng = StdRng::seed_from_u64(2u64);
-        let mut data: Vec<_> = (100..55000_u64)
-            .map(|num| num + rng.gen::<u8>() as u64)
-            .collect();
-        data.push(99_000);
-        data.insert(1000, 2000);
-        data.insert(2000, 100);
-        data.insert(3000, 4100);
-        data.insert(4000, 100);
-        data.insert(5000, 800);
-        data
-    }
-
-    #[inline(never)]
-    fn value_iter() -> impl Iterator<Item = u64> {
-        0..20_000
-    }
-    fn get_reader_for_bench<Codec: FastFieldCodec>(data: &[u64]) -> Codec::Reader {
-        let mut bytes = Vec::new();
-        let min_value = *data.iter().min().unwrap();
-        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
-        let col = VecColumn::from(&data);
-        let normalized_header = crate::NormalizedHeader {
-            num_vals: col.num_vals(),
-            max_value: col.max_value(),
-        };
-        Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
-        Codec::open_from_bytes(OwnedBytes::new(bytes), normalized_header).unwrap()
-    }
-    fn bench_get<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
-        let col = get_reader_for_bench::<Codec>(data);
-        b.iter(|| {
-            let mut sum = 0u64;
-            for pos in value_iter() {
-                let val = col.get_val(pos as u32);
-                sum = sum.wrapping_add(val);
-            }
-            sum
-        });
-    }
-
-    #[inline(never)]
-    fn bench_get_dynamic_helper(b: &mut Bencher, col: Arc<dyn Column>) {
-        b.iter(|| {
-            let mut sum = 0u64;
-            for pos in value_iter() {
-                let val = col.get_val(pos as u32);
-                sum = sum.wrapping_add(val);
-            }
-            sum
-        });
-    }
-
-    fn bench_get_dynamic<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
-        let col = Arc::new(get_reader_for_bench::<Codec>(data));
-        bench_get_dynamic_helper(b, col);
-    }
-    fn bench_create<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
-        let min_value = *data.iter().min().unwrap();
-        let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
-
-        let mut bytes = Vec::new();
-        b.iter(|| {
-            bytes.clear();
-            Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
-        });
-    }
-
-    #[bench]
-    fn bench_fastfield_bitpack_create(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_create::<BitpackedCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_linearinterpol_create(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_create::<LinearCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_multilinearinterpol_create(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_create::<BlockwiseLinearCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_bitpack_get(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get::<BitpackedCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_bitpack_get_dynamic(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get_dynamic::<BitpackedCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_linearinterpol_get(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get::<LinearCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_linearinterpol_get_dynamic(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get_dynamic::<LinearCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_multilinearinterpol_get(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get::<BlockwiseLinearCodec>(b, &data);
-    }
-    #[bench]
-    fn bench_fastfield_multilinearinterpol_get_dynamic(b: &mut Bencher) {
-        let data: Vec<_> = get_data();
-        bench_get_dynamic::<BlockwiseLinearCodec>(b, &data);
-    }
-}

fastfield_codecs/src/line.rs

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

fastfield_codecs/src/linear.rs

@@ -1,230 +0,0 @@
-use std::io::{self, Write};
-
-use common::{BinarySerializable, OwnedBytes};
-use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
-
-use crate::line::Line;
-use crate::serialize::NormalizedHeader;
-use crate::{Column, 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 Column 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 Column, 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 Column) -> 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::tests::get_codec_test_datasets;
-
-    fn create_and_validate(data: &[u64], name: &str) -> Option<(f32, f32)> {
-        crate::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 = 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");
-        }
-    }
-}

fastfield_codecs/src/main.rs

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

fastfield_codecs/src/monotonic_mapping.rs

@@ -1,320 +0,0 @@
-use std::fmt;
-use std::marker::PhantomData;
-use std::ops::RangeInclusive;
-
-use fastdivide::DividerU64;
-
-use crate::MonotonicallyMappableToU128;
-
-/// Monotonic maps a value to u64 value space.
-/// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
-pub trait MonotonicallyMappableToU64:
-    'static + PartialOrd + Copy + Send + Sync + fmt::Debug
-{
-    /// Converts a value to u64.
-    ///
-    /// Internally all fast field values are encoded as u64.
-    fn to_u64(self) -> u64;
-
-    /// Converts a value from u64
-    ///
-    /// Internally all fast field values are encoded as u64.
-    /// **Note: To be used for converting encoded Term, Posting values.**
-    fn from_u64(val: u64) -> Self;
-}
-
-/// Values need to be strictly monotonic mapped to a `Internal` value (u64 or u128) that can be
-/// used in fast field codecs.
-///
-/// The monotonic mapping is required so that `PartialOrd` can be used on `Internal` without
-/// converting to `External`.
-///
-/// All strictly monotonic functions are invertible because they are guaranteed to have a one-to-one
-/// mapping from their range to their domain. The `inverse` method is required when opening a codec,
-/// so a value can be converted back to its original domain (e.g. ip address or f64) from its
-/// internal representation.
-pub trait StrictlyMonotonicFn<External: Copy, Internal: Copy> {
-    /// Strictly monotonically maps the value from External to Internal.
-    fn mapping(&self, inp: External) -> Internal;
-    /// Inverse of `mapping`. Maps the value from Internal to External.
-    fn inverse(&self, out: Internal) -> External;
-
-    /// Maps a user provded value from External to Internal.
-    /// It may be necessary to coerce the value if it is outside the value space.
-    /// In that case it tries to find the next greater value in the value space.
-    ///
-    /// Returns a bool to mark if a value was outside the value space and had to be coerced _up_.
-    /// With that information we can detect if two values in a range both map outside the same value
-    /// space.
-    ///
-    /// coerce_up means the next valid upper value in the value space will be chosen if the value
-    /// has to be coerced.
-    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<Internal> {
-        self.mapping(*inp.start())..=self.mapping(*inp.end())
-    }
-    /// Inverse of `mapping_coerce`.
-    fn inverse_coerce(&self, out: RangeInclusive<Internal>) -> RangeInclusive<External> {
-        self.inverse(*out.start())..=self.inverse(*out.end())
-    }
-}
-
-/// Inverts a strictly monotonic mapping from `StrictlyMonotonicFn<A, B>` to
-/// `StrictlyMonotonicFn<B, A>`.
-///
-/// # Warning
-///
-/// This type comes with a footgun. A type being strictly monotonic does not impose that the inverse
-/// mapping is strictly monotonic over the entire space External. e.g. a -> a * 2. Use at your own
-/// risks.
-pub(crate) struct StrictlyMonotonicMappingInverter<T> {
-    orig_mapping: T,
-}
-impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
-    fn from(orig_mapping: T) -> Self {
-        Self { orig_mapping }
-    }
-}
-
-impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
-where
-    T: StrictlyMonotonicFn<From, To>,
-    From: Copy,
-    To: Copy,
-{
-    #[inline(always)]
-    fn mapping(&self, val: To) -> From {
-        self.orig_mapping.inverse(val)
-    }
-
-    #[inline(always)]
-    fn inverse(&self, val: From) -> To {
-        self.orig_mapping.mapping(val)
-    }
-
-    #[inline]
-    fn mapping_coerce(&self, inp: RangeInclusive<To>) -> RangeInclusive<From> {
-        self.orig_mapping.inverse_coerce(inp)
-    }
-    #[inline]
-    fn inverse_coerce(&self, out: RangeInclusive<From>) -> RangeInclusive<To> {
-        self.orig_mapping.mapping_coerce(out)
-    }
-}
-
-/// Applies the strictly monotonic mapping from `T` without any additional changes.
-pub(crate) struct StrictlyMonotonicMappingToInternal<T> {
-    _phantom: PhantomData<T>,
-}
-
-impl<T> StrictlyMonotonicMappingToInternal<T> {
-    pub(crate) fn new() -> StrictlyMonotonicMappingToInternal<T> {
-        Self {
-            _phantom: PhantomData,
-        }
-    }
-}
-
-impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
-    StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
-where T: MonotonicallyMappableToU128
-{
-    #[inline(always)]
-    fn mapping(&self, inp: External) -> u128 {
-        External::to_u128(inp)
-    }
-
-    #[inline(always)]
-    fn inverse(&self, out: u128) -> External {
-        External::from_u128(out)
-    }
-}
-
-impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
-    StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
-where T: MonotonicallyMappableToU64
-{
-    #[inline(always)]
-    fn mapping(&self, inp: External) -> u64 {
-        External::to_u64(inp)
-    }
-
-    #[inline(always)]
-    fn inverse(&self, out: u64) -> External {
-        External::from_u64(out)
-    }
-}
-
-/// Mapping dividing by  gcd and a base value.
-///
-/// The function is assumed to be only called on values divided by passed
-/// gcd value. (It is necessary for the function to be monotonic.)
-pub(crate) struct StrictlyMonotonicMappingToInternalGCDBaseval {
-    gcd_divider: DividerU64,
-    gcd: u64,
-    min_value: u64,
-}
-impl StrictlyMonotonicMappingToInternalGCDBaseval {
-    pub(crate) fn new(gcd: u64, min_value: u64) -> Self {
-        let gcd_divider = DividerU64::divide_by(gcd);
-        Self {
-            gcd_divider,
-            gcd,
-            min_value,
-        }
-    }
-}
-impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
-    for StrictlyMonotonicMappingToInternalGCDBaseval
-{
-    #[inline(always)]
-    fn mapping(&self, inp: External) -> u64 {
-        self.gcd_divider
-            .divide(External::to_u64(inp) - self.min_value)
-    }
-
-    #[inline(always)]
-    fn inverse(&self, out: u64) -> External {
-        External::from_u64(self.min_value + out * self.gcd)
-    }
-
-    #[inline]
-    #[allow(clippy::reversed_empty_ranges)]
-    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
-        let end = External::to_u64(*inp.end());
-        if end < self.min_value || inp.end() < inp.start() {
-            return 1..=0;
-        }
-        let map_coerce = |mut inp, coerce_up| {
-            let inp_lower_bound = self.inverse(0);
-            if inp < inp_lower_bound {
-                inp = inp_lower_bound;
-            }
-            let val = External::to_u64(inp);
-            let need_coercion = coerce_up && (val - self.min_value) % self.gcd != 0;
-            let mut mapped_val = self.mapping(inp);
-            if need_coercion {
-                mapped_val += 1;
-            }
-            mapped_val
-        };
-        let start = map_coerce(*inp.start(), true);
-        let end = map_coerce(*inp.end(), false);
-        start..=end
-    }
-}
-
-/// Strictly monotonic mapping with a base value.
-pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
-    min_value: u64,
-}
-impl StrictlyMonotonicMappingToInternalBaseval {
-    #[inline(always)]
-    pub(crate) fn new(min_value: u64) -> Self {
-        Self { min_value }
-    }
-}
-
-impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
-    for StrictlyMonotonicMappingToInternalBaseval
-{
-    #[inline]
-    #[allow(clippy::reversed_empty_ranges)]
-    fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
-        if External::to_u64(*inp.end()) < self.min_value {
-            return 1..=0;
-        }
-        let start = self.mapping(External::to_u64(*inp.start()).max(self.min_value));
-        let end = self.mapping(External::to_u64(*inp.end()));
-        start..=end
-    }
-
-    #[inline(always)]
-    fn mapping(&self, val: External) -> u64 {
-        External::to_u64(val) - self.min_value
-    }
-
-    #[inline(always)]
-    fn inverse(&self, val: u64) -> External {
-        External::from_u64(self.min_value + val)
-    }
-}
-
-impl MonotonicallyMappableToU64 for u64 {
-    #[inline(always)]
-    fn to_u64(self) -> u64 {
-        self
-    }
-
-    #[inline(always)]
-    fn from_u64(val: u64) -> Self {
-        val
-    }
-}
-
-impl MonotonicallyMappableToU64 for i64 {
-    #[inline(always)]
-    fn to_u64(self) -> u64 {
-        common::i64_to_u64(self)
-    }
-
-    #[inline(always)]
-    fn from_u64(val: u64) -> Self {
-        common::u64_to_i64(val)
-    }
-}
-
-impl MonotonicallyMappableToU64 for bool {
-    #[inline(always)]
-    fn to_u64(self) -> u64 {
-        u64::from(self)
-    }
-
-    #[inline(always)]
-    fn from_u64(val: u64) -> Self {
-        val > 0
-    }
-}
-
-// TODO remove me.
-// Tantivy should refuse NaN values and work with NotNaN internally.
-impl MonotonicallyMappableToU64 for f64 {
-    #[inline(always)]
-    fn to_u64(self) -> u64 {
-        common::f64_to_u64(self)
-    }
-
-    #[inline(always)]
-    fn from_u64(val: u64) -> Self {
-        common::u64_to_f64(val)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-
-    use super::*;
-
-    #[test]
-    fn strictly_monotonic_test() {
-        // identity mapping
-        test_round_trip(&StrictlyMonotonicMappingToInternal::<u64>::new(), 100u64);
-        // round trip to i64
-        test_round_trip(&StrictlyMonotonicMappingToInternal::<i64>::new(), 100u64);
-        // identity mapping
-        test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
-
-        // base value to i64 round trip
-        let mapping = StrictlyMonotonicMappingToInternalBaseval::new(100);
-        test_round_trip::<_, _, u64>(&mapping, 100i64);
-        // base value and gcd to u64 round trip
-        let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100);
-        test_round_trip::<_, _, u64>(&mapping, 100u64);
-    }
-
-    fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L: Copy>(
-        mapping: &T,
-        test_val: K,
-    ) {
-        assert_eq!(mapping.inverse(mapping.mapping(test_val)), test_val);
-    }
-}

fastfield_codecs/src/monotonic_mapping_u128.rs

@@ -1,43 +0,0 @@
-use std::fmt;
-use std::net::Ipv6Addr;
-
-/// Montonic maps a value to u128 value space
-/// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
-pub trait MonotonicallyMappableToU128:
-    'static + PartialOrd + Copy + Send + Sync + fmt::Debug
-{
-    /// Converts a value to u128.
-    ///
-    /// Internally all fast field values are encoded as u64.
-    fn to_u128(self) -> u128;
-
-    /// Converts a value from u128
-    ///
-    /// Internally all fast field values are encoded as u64.
-    /// **Note: To be used for converting encoded Term, Posting values.**
-    fn from_u128(val: u128) -> Self;
-}
-
-impl MonotonicallyMappableToU128 for u128 {
-    fn to_u128(self) -> u128 {
-        self
-    }
-
-    fn from_u128(val: u128) -> Self {
-        val
-    }
-}
-
-impl MonotonicallyMappableToU128 for Ipv6Addr {
-    fn to_u128(self) -> u128 {
-        ip_to_u128(self)
-    }
-
-    fn from_u128(val: u128) -> Self {
-        Ipv6Addr::from(val.to_be_bytes())
-    }
-}
-
-fn ip_to_u128(ip_addr: Ipv6Addr) -> u128 {
-    u128::from_be_bytes(ip_addr.octets())
-}