use optional index in multivalued index (#2439)

* use optional index in multivalued index For mostly empty multivalued indices there was a large overhead during creation when iterating all docids. This is alleviated by placing an optional index in the multivalued index to mark documents that have values. There's some performance overhead when accessing values in a multivalued index. The accessing cost is now optional index + multivalue index. The sparse codec performs relatively bad with the binary_search when accessing data. This is reflected in the benchmarks below. This changes the format of columnar to v2, but code is added to handle the v1 formats. ``` Running benches/bench_access.rs (/home/pascal/Development/tantivy/optional_multivalues/target/release/deps/bench_access-ea323c028db88db4) multi sparse 1/13 access_values_for_doc Avg: 42.8946ms (+241.80%) Median: 42.8869ms (+244.10%) [42.7484ms .. 43.1074ms] access_first_vals Avg: 42.8022ms (+421.93%) Median: 42.7553ms (+439.84%) [42.6794ms .. 43.7404ms] multi 2x access_values_for_doc Avg: 31.1244ms (+24.17%) Median: 30.8339ms (+23.46%) [30.7192ms .. 33.6059ms] access_first_vals Avg: 24.3070ms (+70.92%) Median: 24.0966ms (+70.18%) [23.9328ms .. 26.4851ms] sparse 1/13 access_values_for_doc Avg: 42.2490ms (+0.61%) Median: 42.2346ms (+2.28%) [41.8988ms .. 43.7821ms] access_first_vals Avg: 43.6272ms (+0.23%) Median: 43.6197ms (+1.78%) [43.4920ms .. 43.9009ms] dense 1/12 access_values_for_doc Avg: 8.6184ms (+23.18%) Median: 8.6126ms (+23.78%) [8.5843ms .. 8.7527ms] access_first_vals Avg: 6.8112ms (+4.47%) Median: 6.8002ms (+4.55%) [6.7887ms .. 6.8991ms] full access_values_for_doc Avg: 9.4073ms (-5.09%) Median: 9.4023ms (-2.23%) [9.3694ms .. 9.4568ms] access_first_vals Avg: 4.9531ms (+6.24%) Median: 4.9502ms (+7.85%) [4.9423ms .. 4.9718ms] ``` ``` Running benches/bench_merge.rs (/home/pascal/Development/tantivy/optional_multivalues/target/release/deps/bench_merge-475697dfceb3639f) merge_multi 2x_and_multi 2x Avg: 20.2280ms (+34.33%) Median: 20.1829ms (+35.33%) [19.9933ms .. 20.8806ms] merge_multi sparse 1/13_and_multi sparse 1/13 Avg: 0.8961ms (-78.04%) Median: 0.8943ms (-77.61%) [0.8899ms .. 0.9272ms] merge_dense 1/12_and_dense 1/12 Avg: 0.6619ms (-1.26%) Median: 0.6616ms (+2.20%) [0.6473ms .. 0.6837ms] merge_sparse 1/13_and_sparse 1/13 Avg: 0.5508ms (-0.85%) Median: 0.5508ms (+2.80%) [0.5420ms .. 0.5634ms] merge_sparse 1/13_and_dense 1/12 Avg: 0.6046ms (-4.64%) Median: 0.6038ms (+2.80%) [0.5939ms .. 0.6296ms] merge_multi sparse 1/13_and_dense 1/12 Avg: 0.9111ms (-83.48%) Median: 0.9063ms (-83.50%) [0.9047ms .. 0.9663ms] merge_multi sparse 1/13_and_sparse 1/13 Avg: 0.8451ms (-89.49%) Median: 0.8428ms (-89.43%) [0.8411ms .. 0.8563ms] merge_multi 2x_and_dense 1/12 Avg: 10.6624ms (-4.82%) Median: 10.6568ms (-4.49%) [10.5738ms .. 10.8353ms] merge_multi 2x_and_sparse 1/13 Avg: 10.6336ms (-22.95%) Median: 10.5925ms (-22.33%) [10.5149ms .. 11.5657ms] ``` * Update columnar/src/columnar/format_version.rs Co-authored-by: Paul Masurel <[email protected]> * Update columnar/src/column_index/mod.rs Co-authored-by: Paul Masurel <[email protected]> --------- Co-authored-by: Paul Masurel <[email protected]>
quickwit-oss · Jun 19, 2024 · 5908414 · 5908414
1 parent 511b027
commit 5908414
Show file tree

Hide file tree

Showing 28 changed files with 1,007 additions and 366 deletions.
diff --git a/columnar/benches/bench_access.rs b/columnar/benches/bench_access.rs
@@ -1,73 +1,13 @@
-use core::fmt;
-use std::fmt::{Display, Formatter};
-
 use binggan::{black_box, InputGroup};
-use tantivy_columnar::*;
+use common::*;
+use tantivy_columnar::Column;
 
-pub enum Card {
-    MultiSparse,
-    Multi,
-    Sparse,
-    Dense,
-    Full,
-}
-impl Display for Card {
-    fn fmt(&self, f: &mut Formatter) -> fmt::Result {
-        match self {
-            Card::MultiSparse => write!(f, "multi sparse 1/13"),
-            Card::Multi => write!(f, "multi 2x"),
-            Card::Sparse => write!(f, "sparse 1/13"),
-            Card::Dense => write!(f, "dense 1/12"),
-            Card::Full => write!(f, "full"),
-        }
-    }
-}
+pub mod common;
 
 const NUM_DOCS: u32 = 2_000_000;
 
-pub fn generate_columnar(card: Card, num_docs: u32) -> ColumnarReader {
-    use tantivy_columnar::ColumnarWriter;
-
-    let mut columnar_writer = ColumnarWriter::default();
-
-    match card {
-        Card::MultiSparse => {
-            columnar_writer.record_numerical(0, "price", 10u64);
-            columnar_writer.record_numerical(0, "price", 10u64);
-        }
-        _ => {}
-    }
-
-    for i in 0..num_docs {
-        match card {
-            Card::MultiSparse | Card::Sparse => {
-                if i % 13 == 0 {
-                    columnar_writer.record_numerical(i, "price", i as u64);
-                }
-            }
-            Card::Dense => {
-                if i % 12 == 0 {
-                    columnar_writer.record_numerical(i, "price", i as u64);
-                }
-            }
-            Card::Full => {
-                columnar_writer.record_numerical(i, "price", i as u64);
-            }
-            Card::Multi => {
-                columnar_writer.record_numerical(i, "price", i as u64);
-                columnar_writer.record_numerical(i, "price", i as u64);
-            }
-        }
-    }
-
-    let mut wrt: Vec<u8> = Vec::new();
-    columnar_writer.serialize(num_docs, &mut wrt).unwrap();
-    let reader = ColumnarReader::open(wrt).unwrap();
-    reader
-}
-
 pub fn generate_columnar_and_open(card: Card, num_docs: u32) -> Column {
-    let reader = generate_columnar(card, num_docs);
+    let reader = generate_columnar_with_name(card, num_docs, "price");
     reader.read_columns("price").unwrap()[0]
         .open_u64_lenient()
         .unwrap()
@@ -116,9 +56,8 @@ fn bench_group(mut runner: InputGroup<Column>) {
 
             column.first_vals(&docs, &mut buffer);
             for val in buffer.iter() {
-                if let Some(val) = val {
-                    sum += *val;
-                }
+                let Some(val) = val else { continue };
+                sum += *val;
             }
         }
 

diff --git a/columnar/benches/bench_merge.rs b/columnar/benches/bench_merge.rs
@@ -1,7 +1,7 @@
-mod bench_access;
+pub mod common;
 
-use bench_access::{generate_columnar, Card};
 use binggan::{black_box, BenchRunner};
+use common::{generate_columnar_with_name, Card};
 use tantivy_columnar::*;
 
 const NUM_DOCS: u32 = 100_000;
@@ -13,8 +13,8 @@ fn main() {
         inputs.push((
             format!("merge_{card1}_and_{card2}"),
             vec![
-                generate_columnar(card1, NUM_DOCS),
-                generate_columnar(card2, NUM_DOCS),
+                generate_columnar_with_name(card1, NUM_DOCS, "price"),
+                generate_columnar_with_name(card2, NUM_DOCS, "price"),
             ],
         ));
     };

diff --git a/columnar/benches/common.rs b/columnar/benches/common.rs
@@ -0,0 +1,59 @@
+extern crate tantivy_columnar;
+
+use core::fmt;
+use std::fmt::{Display, Formatter};
+
+use tantivy_columnar::{ColumnarReader, ColumnarWriter};
+
+pub enum Card {
+    MultiSparse,
+    Multi,
+    Sparse,
+    Dense,
+    Full,
+}
+impl Display for Card {
+    fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+        match self {
+            Card::MultiSparse => write!(f, "multi sparse 1/13"),
+            Card::Multi => write!(f, "multi 2x"),
+            Card::Sparse => write!(f, "sparse 1/13"),
+            Card::Dense => write!(f, "dense 1/12"),
+            Card::Full => write!(f, "full"),
+        }
+    }
+}
+pub fn generate_columnar_with_name(card: Card, num_docs: u32, column_name: &str) -> ColumnarReader {
+    let mut columnar_writer = ColumnarWriter::default();
+
+    if let Card::MultiSparse = card {
+        columnar_writer.record_numerical(0, column_name, 10u64);
+        columnar_writer.record_numerical(0, column_name, 10u64);
+    }
+
+    for i in 0..num_docs {
+        match card {
+            Card::MultiSparse | Card::Sparse => {
+                if i % 13 == 0 {
+                    columnar_writer.record_numerical(i, column_name, i as u64);
+                }
+            }
+            Card::Dense => {
+                if i % 12 == 0 {
+                    columnar_writer.record_numerical(i, column_name, i as u64);
+                }
+            }
+            Card::Full => {
+                columnar_writer.record_numerical(i, column_name, i as u64);
+            }
+            Card::Multi => {
+                columnar_writer.record_numerical(i, column_name, i as u64);
+                columnar_writer.record_numerical(i, column_name, i as u64);
+            }
+        }
+    }
+
+    let mut wrt: Vec<u8> = Vec::new();
+    columnar_writer.serialize(num_docs, &mut wrt).unwrap();
+    ColumnarReader::open(wrt).unwrap()
+}
diff --git a/columnar/compat_tests_data/v1.columnar b/columnar/compat_tests_data/v1.columnar
diff --git a/columnar/compat_tests_data/v2.columnar b/columnar/compat_tests_data/v2.columnar
diff --git a/columnar/src/column/mod.rs b/columnar/src/column/mod.rs
@@ -136,7 +136,7 @@ impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
             .map(|value_row_id: RowId| self.values.get_val(value_row_id))
     }
 
-    /// Get the docids of values which are in the provided value range.
+    /// Get the docids of values which are in the provided value and docid range.
     #[inline]
     pub fn get_docids_for_value_range(
         &self,

diff --git a/columnar/src/column/serialize.rs b/columnar/src/column/serialize.rs
@@ -12,7 +12,7 @@ use crate::column_values::{
     CodecType, MonotonicallyMappableToU128, MonotonicallyMappableToU64,
 };
 use crate::iterable::Iterable;
-use crate::StrColumn;
+use crate::{StrColumn, Version};
 
 pub fn serialize_column_mappable_to_u128<T: MonotonicallyMappableToU128>(
     column_index: SerializableColumnIndex<'_>,
@@ -40,7 +40,10 @@ pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64>(
     Ok(())
 }
 
-pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Column<T>> {
+pub fn open_column_u64<T: MonotonicallyMappableToU64>(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<Column<T>> {
     let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
     let column_index_num_bytes = u32::from_le_bytes(
         column_index_num_bytes_payload
@@ -49,7 +52,7 @@ pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::
             .unwrap(),
     );
     let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
-    let column_index = crate::column_index::open_column_index(column_index_data)?;
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
     let column_values = load_u64_based_column_values(column_values_data)?;
     Ok(Column {
         index: column_index,
@@ -59,6 +62,7 @@ pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::
 
 pub fn open_column_u128<T: MonotonicallyMappableToU128>(
     bytes: OwnedBytes,
+    format_version: Version,
 ) -> io::Result<Column<T>> {
     let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
     let column_index_num_bytes = u32::from_le_bytes(
@@ -68,7 +72,7 @@ pub fn open_column_u128<T: MonotonicallyMappableToU128>(
             .unwrap(),
     );
     let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
-    let column_index = crate::column_index::open_column_index(column_index_data)?;
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
     let column_values = crate::column_values::open_u128_mapped(column_values_data)?;
     Ok(Column {
         index: column_index,
@@ -79,7 +83,10 @@ pub fn open_column_u128<T: MonotonicallyMappableToU128>(
 /// Open the column as u64.
 ///
 /// See [`open_u128_as_compact_u64`] for more details.
-pub fn open_column_u128_as_compact_u64(bytes: OwnedBytes) -> io::Result<Column<u64>> {
+pub fn open_column_u128_as_compact_u64(
+    bytes: OwnedBytes,
+    format_version: Version,
+) -> io::Result<Column<u64>> {
     let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
     let column_index_num_bytes = u32::from_le_bytes(
         column_index_num_bytes_payload
@@ -88,27 +95,27 @@ pub fn open_column_u128_as_compact_u64(bytes: OwnedBytes) -> io::Result<Column<u
             .unwrap(),
     );
     let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
-    let column_index = crate::column_index::open_column_index(column_index_data)?;
+    let column_index = crate::column_index::open_column_index(column_index_data, format_version)?;
     let column_values = crate::column_values::open_u128_as_compact_u64(column_values_data)?;
     Ok(Column {
         index: column_index,
         values: column_values,
     })
 }
 
-pub fn open_column_bytes(data: OwnedBytes) -> io::Result<BytesColumn> {
+pub fn open_column_bytes(data: OwnedBytes, format_version: Version) -> 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 term_ord_column = crate::column::open_column_u64::<u64>(column_bytes, format_version)?;
     Ok(BytesColumn {
         dictionary,
         term_ord_column,
     })
 }
 
-pub fn open_column_str(data: OwnedBytes) -> io::Result<StrColumn> {
-    let bytes_column = open_column_bytes(data)?;
+pub fn open_column_str(data: OwnedBytes, format_version: Version) -> io::Result<StrColumn> {
+    let bytes_column = open_column_bytes(data, format_version)?;
     Ok(StrColumn::wrap(bytes_column))
 }
diff --git a/columnar/src/column_index/merge/mod.rs b/columnar/src/column_index/merge/mod.rs
@@ -95,8 +95,12 @@ pub fn merge_column_index<'a>(
 
 #[cfg(test)]
 mod tests {
+    use common::OwnedBytes;
+
     use crate::column_index::merge::detect_cardinality;
-    use crate::column_index::multivalued_index::MultiValueIndex;
+    use crate::column_index::multivalued_index::{
+        open_multivalued_index, serialize_multivalued_index, MultiValueIndex,
+    };
     use crate::column_index::{merge_column_index, OptionalIndex, SerializableColumnIndex};
     use crate::{
         Cardinality, ColumnIndex, MergeRowOrder, RowAddr, RowId, ShuffleMergeOrder, StackMergeOrder,
@@ -171,7 +175,11 @@ mod tests {
         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();
+        let mut output = Vec::new();
+        serialize_multivalued_index(&start_index_iterable, &mut output).unwrap();
+        let multivalue =
+            open_multivalued_index(OwnedBytes::new(output), crate::Version::V2).unwrap();
+        let start_indexes: Vec<RowId> = multivalue.get_start_index_column().iter().collect();
         assert_eq!(&start_indexes, &[0, 3, 5]);
     }
 
@@ -200,11 +208,16 @@ mod tests {
             ],
         )
         .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();
+        let mut output = Vec::new();
+        serialize_multivalued_index(&start_index_iterable, &mut output).unwrap();
+        let multivalue =
+            open_multivalued_index(OwnedBytes::new(output), crate::Version::V2).unwrap();
+        let start_indexes: Vec<RowId> = multivalue.get_start_index_column().iter().collect();
         assert_eq!(&start_indexes, &[0, 3, 5, 6]);
     }
 }