MeiliSearch/milli/src/update/facets.rs
bors[bot] 25123af3b8
Merge #436
436: Speed up the word prefix databases computation time r=Kerollmops a=Kerollmops

This PR depends on the fixes done in #431 and must be merged after it.

In this PR we will bring the `WordPrefixPairProximityDocids`, `WordPrefixDocids` and, `WordPrefixPositionDocids` update structures to a new era, a better era, where computing the word prefix pair proximities costs much fewer CPU cycles, an era where this update structure can use the, previously computed, set of new word docids from the newly indexed batch of documents.

---

The `WordPrefixPairProximityDocids` is an update structure, which means that it is an object that we feed with some parameters and which modifies the LMDB database of an index when asked for. This structure specifically computes the list of word prefix pair proximities, which correspond to a list of pairs of words associated with a proximity (the distance between both words) where the second word is not a word but a prefix e.g. `s`, `se`, `a`. This word prefix pair proximity is associated with the list of documents ids which contains the pair of words and prefix at the given proximity.

The origin of the performances issue that this struct brings is related to the fact that it starts its job from the beginning, it clears the LMDB database before rewriting everything from scratch, using the other LMDB databases to achieve that. I hope you understand that this is absolutely not an optimized way of doing things.

Co-authored-by: Clément Renault <clement@meilisearch.com>
Co-authored-by: Kerollmops <clement@meilisearch.com>
2022-02-16 15:41:14 +00:00

345 lines
12 KiB
Rust

use std::fs::File;
use std::num::{NonZeroU8, NonZeroUsize};
use std::{cmp, mem};
use grenad::{CompressionType, Reader, Writer};
use heed::types::{ByteSlice, DecodeIgnore};
use heed::{BytesEncode, Error};
use log::debug;
use roaring::RoaringBitmap;
use time::OffsetDateTime;
use crate::error::InternalError;
use crate::heed_codec::facet::{
FacetLevelValueF64Codec, FacetLevelValueU32Codec, FacetStringLevelZeroCodec,
FacetStringLevelZeroValueCodec, FacetStringZeroBoundsValueCodec,
};
use crate::heed_codec::CboRoaringBitmapCodec;
use crate::update::index_documents::{create_writer, write_into_lmdb_database, writer_into_reader};
use crate::{FieldId, Index, Result};
pub struct Facets<'t, 'u, 'i> {
wtxn: &'t mut heed::RwTxn<'i, 'u>,
index: &'i Index,
pub(crate) chunk_compression_type: CompressionType,
pub(crate) chunk_compression_level: Option<u32>,
level_group_size: NonZeroUsize,
min_level_size: NonZeroUsize,
}
impl<'t, 'u, 'i> Facets<'t, 'u, 'i> {
pub fn new(wtxn: &'t mut heed::RwTxn<'i, 'u>, index: &'i Index) -> Facets<'t, 'u, 'i> {
Facets {
wtxn,
index,
chunk_compression_type: CompressionType::None,
chunk_compression_level: None,
level_group_size: NonZeroUsize::new(4).unwrap(),
min_level_size: NonZeroUsize::new(5).unwrap(),
}
}
pub fn level_group_size(&mut self, value: NonZeroUsize) -> &mut Self {
self.level_group_size = NonZeroUsize::new(cmp::max(value.get(), 2)).unwrap();
self
}
pub fn min_level_size(&mut self, value: NonZeroUsize) -> &mut Self {
self.min_level_size = value;
self
}
#[logging_timer::time("Facets::{}")]
pub fn execute(self) -> Result<()> {
self.index.set_updated_at(self.wtxn, &OffsetDateTime::now_utc())?;
// We get the faceted fields to be able to create the facet levels.
let faceted_fields = self.index.faceted_fields_ids(self.wtxn)?;
debug!("Computing and writing the facet values levels docids into LMDB on disk...");
for field_id in faceted_fields {
// Clear the facet string levels.
clear_field_string_levels(
self.wtxn,
self.index.facet_id_string_docids.remap_types::<ByteSlice, DecodeIgnore>(),
field_id,
)?;
// Compute and store the faceted strings documents ids.
let string_documents_ids = compute_faceted_strings_documents_ids(
self.wtxn,
self.index.facet_id_string_docids.remap_key_type::<ByteSlice>(),
field_id,
)?;
let facet_string_levels = compute_facet_string_levels(
self.wtxn,
self.index.facet_id_string_docids,
self.chunk_compression_type,
self.chunk_compression_level,
self.level_group_size,
self.min_level_size,
field_id,
)?;
// Clear the facet number levels.
clear_field_number_levels(self.wtxn, self.index.facet_id_f64_docids, field_id)?;
// Compute and store the faceted numbers documents ids.
let number_documents_ids = compute_faceted_numbers_documents_ids(
self.wtxn,
self.index.facet_id_f64_docids.remap_key_type::<ByteSlice>(),
field_id,
)?;
let facet_number_levels = compute_facet_number_levels(
self.wtxn,
self.index.facet_id_f64_docids,
self.chunk_compression_type,
self.chunk_compression_level,
self.level_group_size,
self.min_level_size,
field_id,
)?;
self.index.put_string_faceted_documents_ids(
self.wtxn,
field_id,
&string_documents_ids,
)?;
self.index.put_number_faceted_documents_ids(
self.wtxn,
field_id,
&number_documents_ids,
)?;
write_into_lmdb_database(
self.wtxn,
*self.index.facet_id_f64_docids.as_polymorph(),
facet_number_levels,
|_, _| Err(InternalError::IndexingMergingKeys { process: "facet number levels" })?,
)?;
write_into_lmdb_database(
self.wtxn,
*self.index.facet_id_string_docids.as_polymorph(),
facet_string_levels,
|_, _| Err(InternalError::IndexingMergingKeys { process: "facet string levels" })?,
)?;
}
Ok(())
}
}
fn clear_field_number_levels<'t>(
wtxn: &'t mut heed::RwTxn,
db: heed::Database<FacetLevelValueF64Codec, CboRoaringBitmapCodec>,
field_id: FieldId,
) -> heed::Result<()> {
let left = (field_id, 1, f64::MIN, f64::MIN);
let right = (field_id, u8::MAX, f64::MAX, f64::MAX);
let range = left..=right;
db.delete_range(wtxn, &range).map(drop)
}
fn compute_facet_number_levels<'t>(
rtxn: &'t heed::RoTxn,
db: heed::Database<FacetLevelValueF64Codec, CboRoaringBitmapCodec>,
compression_type: CompressionType,
compression_level: Option<u32>,
level_group_size: NonZeroUsize,
min_level_size: NonZeroUsize,
field_id: FieldId,
) -> Result<Reader<File>> {
let first_level_size = db
.remap_key_type::<ByteSlice>()
.prefix_iter(rtxn, &field_id.to_be_bytes())?
.remap_types::<DecodeIgnore, DecodeIgnore>()
.fold(Ok(0usize), |count, result| result.and(count).map(|c| c + 1))?;
// It is forbidden to keep a cursor and write in a database at the same time with LMDB
// therefore we write the facet levels entries into a grenad file before transfering them.
let mut writer = create_writer(compression_type, compression_level, tempfile::tempfile()?);
let level_0_range = {
let left = (field_id, 0, f64::MIN, f64::MIN);
let right = (field_id, 0, f64::MAX, f64::MAX);
left..=right
};
// Groups sizes are always a power of the original level_group_size and therefore a group
// always maps groups of the previous level and never splits previous levels groups in half.
let group_size_iter = (1u8..)
.map(|l| (l, level_group_size.get().pow(l as u32)))
.take_while(|(_, s)| first_level_size / *s >= min_level_size.get());
for (level, group_size) in group_size_iter {
let mut left = 0.0;
let mut right = 0.0;
let mut group_docids = RoaringBitmap::new();
for (i, result) in db.range(rtxn, &level_0_range)?.enumerate() {
let ((_field_id, _level, value, _right), docids) = result?;
if i == 0 {
left = value;
} else if i % group_size == 0 {
// we found the first bound of the next group, we must store the left
// and right bounds associated with the docids.
write_number_entry(&mut writer, field_id, level, left, right, &group_docids)?;
// We save the left bound for the new group and also reset the docids.
group_docids = RoaringBitmap::new();
left = value;
}
// The right bound is always the bound we run through.
group_docids |= docids;
right = value;
}
if !group_docids.is_empty() {
write_number_entry(&mut writer, field_id, level, left, right, &group_docids)?;
}
}
writer_into_reader(writer)
}
fn write_number_entry(
writer: &mut Writer<File>,
field_id: FieldId,
level: u8,
left: f64,
right: f64,
ids: &RoaringBitmap,
) -> Result<()> {
let key = (field_id, level, left, right);
let key = FacetLevelValueF64Codec::bytes_encode(&key).ok_or(Error::Encoding)?;
let data = CboRoaringBitmapCodec::bytes_encode(&ids).ok_or(Error::Encoding)?;
writer.insert(&key, &data)?;
Ok(())
}
fn compute_faceted_strings_documents_ids(
rtxn: &heed::RoTxn,
db: heed::Database<ByteSlice, FacetStringLevelZeroValueCodec>,
field_id: FieldId,
) -> Result<RoaringBitmap> {
let mut documents_ids = RoaringBitmap::new();
for result in db.prefix_iter(rtxn, &field_id.to_be_bytes())? {
let (_key, (_original_value, docids)) = result?;
documents_ids |= docids;
}
Ok(documents_ids)
}
fn compute_faceted_numbers_documents_ids(
rtxn: &heed::RoTxn,
db: heed::Database<ByteSlice, CboRoaringBitmapCodec>,
field_id: FieldId,
) -> Result<RoaringBitmap> {
let mut documents_ids = RoaringBitmap::new();
for result in db.prefix_iter(rtxn, &field_id.to_be_bytes())? {
let (_key, docids) = result?;
documents_ids |= docids;
}
Ok(documents_ids)
}
fn clear_field_string_levels<'t>(
wtxn: &'t mut heed::RwTxn,
db: heed::Database<ByteSlice, DecodeIgnore>,
field_id: FieldId,
) -> heed::Result<()> {
let left = (field_id, NonZeroU8::new(1).unwrap(), u32::MIN, u32::MIN);
let right = (field_id, NonZeroU8::new(u8::MAX).unwrap(), u32::MAX, u32::MAX);
let range = left..=right;
db.remap_key_type::<FacetLevelValueU32Codec>().delete_range(wtxn, &range).map(drop)
}
fn compute_facet_string_levels<'t>(
rtxn: &'t heed::RoTxn,
db: heed::Database<FacetStringLevelZeroCodec, FacetStringLevelZeroValueCodec>,
compression_type: CompressionType,
compression_level: Option<u32>,
level_group_size: NonZeroUsize,
min_level_size: NonZeroUsize,
field_id: FieldId,
) -> Result<Reader<File>> {
let first_level_size = db
.remap_key_type::<ByteSlice>()
.prefix_iter(rtxn, &field_id.to_be_bytes())?
.remap_types::<DecodeIgnore, DecodeIgnore>()
.fold(Ok(0usize), |count, result| result.and(count).map(|c| c + 1))?;
// It is forbidden to keep a cursor and write in a database at the same time with LMDB
// therefore we write the facet levels entries into a grenad file before transfering them.
let mut writer = create_writer(compression_type, compression_level, tempfile::tempfile()?);
// Groups sizes are always a power of the original level_group_size and therefore a group
// always maps groups of the previous level and never splits previous levels groups in half.
let group_size_iter = (1u8..)
.map(|l| (l, level_group_size.get().pow(l as u32)))
.take_while(|(_, s)| first_level_size / *s >= min_level_size.get());
for (level, group_size) in group_size_iter {
let level = NonZeroU8::new(level).unwrap();
let mut left = (0, "");
let mut right = (0, "");
let mut group_docids = RoaringBitmap::new();
// Because we know the size of the level 0 we can use a range iterator that starts
// at the first value of the level and goes to the last by simply counting.
for (i, result) in db.range(rtxn, &((field_id, "")..))?.take(first_level_size).enumerate() {
let ((_field_id, value), (_original_value, docids)) = result?;
if i == 0 {
left = (i as u32, value);
} else if i % group_size == 0 {
// we found the first bound of the next group, we must store the left
// and right bounds associated with the docids. We also reset the docids.
let docids = mem::take(&mut group_docids);
write_string_entry(&mut writer, field_id, level, left, right, docids)?;
// We save the left bound for the new group.
left = (i as u32, value);
}
// The right bound is always the bound we run through.
group_docids |= docids;
right = (i as u32, value);
}
if !group_docids.is_empty() {
let docids = mem::take(&mut group_docids);
write_string_entry(&mut writer, field_id, level, left, right, docids)?;
}
}
writer_into_reader(writer)
}
fn write_string_entry(
writer: &mut Writer<File>,
field_id: FieldId,
level: NonZeroU8,
(left_id, left_value): (u32, &str),
(right_id, right_value): (u32, &str),
docids: RoaringBitmap,
) -> Result<()> {
let key = (field_id, level, left_id, right_id);
let key = FacetLevelValueU32Codec::bytes_encode(&key).ok_or(Error::Encoding)?;
let data = match level.get() {
1 => (Some((left_value, right_value)), docids),
_ => (None, docids),
};
let data = FacetStringZeroBoundsValueCodec::<CboRoaringBitmapCodec>::bytes_encode(&data)
.ok_or(Error::Encoding)?;
writer.insert(&key, &data)?;
Ok(())
}