Make several indexing optimizations #4350
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
/benchmark indexing |
Kerollmops
requested changes
Jan 23, 2024
Comment on lines
93
to
188
| .inspect(|result| { | ||
| if proximity_precision == ProximityPrecision::ByWord { | ||
| if let Ok((docid_word_positions_chunk, _)) = result { | ||
| run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>( | ||
| docid_word_positions_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| extract_word_pair_proximity_docids, | ||
| TypedChunk::WordPairProximityDocids, | ||
| "word-pair-proximity-docids", | ||
| ); | ||
| } | ||
| } | ||
| }) | ||
| .inspect(|result| { | ||
| if let Ok((docid_word_positions_chunk, _)) = result { | ||
| run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>( | ||
| docid_word_positions_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| extract_fid_word_count_docids, | ||
| TypedChunk::FieldIdWordCountDocids, | ||
| "field-id-wordcount-docids", | ||
| ); | ||
| } | ||
| }) | ||
| .inspect(|result| { | ||
| if let Ok((docid_word_positions_chunk, _)) = result { | ||
| let exact_attributes = exact_attributes.clone(); | ||
| run_extraction_task::< | ||
| _, | ||
| _, | ||
| ( | ||
| grenad::Reader<BufReader<File>>, | ||
| grenad::Reader<BufReader<File>>, | ||
| grenad::Reader<BufReader<File>>, | ||
| ), | ||
| >( | ||
| docid_word_positions_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| move |doc_word_pos, indexer| { | ||
| extract_word_docids(doc_word_pos, indexer, &exact_attributes) | ||
| }, | ||
| |( | ||
| word_docids_reader, | ||
| exact_word_docids_reader, | ||
| word_fid_docids_reader, | ||
| )| { | ||
| TypedChunk::WordDocids { | ||
| word_docids_reader, | ||
| exact_word_docids_reader, | ||
| word_fid_docids_reader, | ||
| } | ||
| }, | ||
| "word-docids", | ||
| ); | ||
| } | ||
| }) | ||
| .inspect(|result| { | ||
| if let Ok((docid_word_positions_chunk, _)) = result { | ||
| run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>( | ||
| docid_word_positions_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| extract_word_position_docids, | ||
| TypedChunk::WordPositionDocids, | ||
| "word-position-docids", | ||
| ); | ||
| } | ||
| }) | ||
| .inspect(|result| { | ||
| if let Ok((_, (_, fid_docid_facet_strings_chunk))) = result { | ||
| run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>( | ||
| fid_docid_facet_strings_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| extract_facet_string_docids, | ||
| TypedChunk::FieldIdFacetStringDocids, | ||
| "field-id-facet-string-docids", | ||
| ); | ||
| } | ||
| }) | ||
| .inspect(|result| { | ||
| if let Ok((_, (fid_docid_facet_numbers_chunk, _))) = result { | ||
| run_extraction_task::<_, _, grenad::Reader<BufReader<File>>>( | ||
| fid_docid_facet_numbers_chunk.clone(), | ||
| indexer, | ||
| lmdb_writer_sx.clone(), | ||
| extract_facet_number_docids, | ||
| TypedChunk::FieldIdFacetNumberDocids, | ||
| "field-id-facet-number-docids", | ||
| ); | ||
| } | ||
| }) | ||
| .map(|r| r.map(|_| ())) |
There was a problem hiding this comment.
Could you replace all those inspect and this map with a single map instead. inspect is more meant to be used to debug stuff than traversing for actual work. Also, can't these inspect tasks run in parallel? It seems that they are run sequentially here.
There was a problem hiding this comment.
Maybe you can use the for_each rayon trait method on various functions. Looping on the list of functions to run with the run_extraction_task could do the trick to run those extractions in parallel!
| @@ -82,90 +82,6 @@ pub unsafe fn as_cloneable_grenad( | |||
| Ok(reader) | |||
| } | |||
|
|
|||
| pub trait MergeableReader | |||
|
Here are your indexing benchmarks diff 👊 |
|
/benchmark indexing |
|
Here are your indexing benchmarks diff 👊 |
|
I measured this PR with a dump loading of a big index of about 150M documents. Writing along the way without merging at the end of the extraction was not interesting and was slowing down the whole process. However, keeping the chunk size formula this way was very good, reducing the number of files open in parallel and the number of files to be merged before writing them. This screenshot shows:
|
|
/benchmark indexing |
This was referenced Feb 8, 2024
ManyTheFish
force-pushed
the
remove-merging-phase-from-indexing
branch
from
e8ed27b
to
169f27f
Compare
ManyTheFish
changed the title
|
Here are your indexing benchmarks diff 👊 |
ManyTheFish
force-pushed
the
remove-merging-phase-from-indexing
branch
2 times, most recently
from
f8632f4
to
c659cc2
Compare
dureuill
force-pushed
the
remove-merging-phase-from-indexing
branch
from
c659cc2
to
747d4fb
Compare
ManyTheFish
force-pushed
the
remove-merging-phase-from-indexing
branch
from
d65368a
to
39c83cb
Compare
Kerollmops
requested changes
Feb 12, 2024
milli/src/update/facet/mod.rs
Outdated
| .map(|(_, c)| c) | ||
| .collect(); | ||
| normalized_facet = normalized_truncated_facet.into(); | ||
| if let Some(normalized_delta_data) = self.normalized_delta_data { |
There was a problem hiding this comment.
Can you extract this into another function, please?
Comment on lines
226
to
245
| pub fn merge_btreeset_string<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> { | ||
| if values.len() == 1 { | ||
| Ok(values[0].clone()) | ||
| } else { | ||
| // TODO improve the perf by using a `#[borrow] Cow<str>`. | ||
| let strings: BTreeSet<String> = values | ||
| .iter() | ||
| .map(AsRef::as_ref) | ||
| .map(serde_json::from_slice::<BTreeSet<String>>) | ||
| .map(StdResult::unwrap) | ||
| .reduce(|mut current, new| { | ||
| for x in new { | ||
| current.insert(x); | ||
| } | ||
| current | ||
| }) | ||
| .unwrap(); | ||
| Ok(Cow::Owned(serde_json::to_vec(&strings).unwrap())) | ||
| } | ||
| } |
ManyTheFish
commented
Feb 13, 2024
Comment on lines
90
to
116
| impl PartialEq for TypedChunk { | ||
| fn eq(&self, other: &Self) -> bool { | ||
| use TypedChunk::*; | ||
| match (self, other) { | ||
| (FieldIdDocidFacetStrings(_), FieldIdDocidFacetStrings(_)) | ||
| | (FieldIdDocidFacetNumbers(_), FieldIdDocidFacetNumbers(_)) | ||
| | (Documents(_), Documents(_)) | ||
| | (FieldIdWordCountDocids(_), FieldIdWordCountDocids(_)) | ||
| | (WordDocids { .. }, WordDocids { .. }) | ||
| | (WordPositionDocids(_), WordPositionDocids(_)) | ||
| | (WordPairProximityDocids(_), WordPairProximityDocids(_)) | ||
| | (FieldIdFacetStringDocids(_), FieldIdFacetStringDocids(_)) | ||
| | (FieldIdFacetNumberDocids(_), FieldIdFacetNumberDocids(_)) | ||
| | (FieldIdFacetExistsDocids(_), FieldIdFacetExistsDocids(_)) | ||
| | (FieldIdFacetIsNullDocids(_), FieldIdFacetIsNullDocids(_)) | ||
| | (FieldIdFacetIsEmptyDocids(_), FieldIdFacetIsEmptyDocids(_)) | ||
| | (GeoPoints(_), GeoPoints(_)) | ||
| | (ScriptLanguageDocids(_), ScriptLanguageDocids(_)) => true, | ||
| ( | ||
| VectorPoints { embedder_name: left, expected_dimension: left_dim, .. }, | ||
| VectorPoints { embedder_name: right, expected_dimension: right_dim, .. }, | ||
| ) => left == right && left_dim == right_dim, | ||
| _ => false, | ||
| } | ||
| } | ||
| } | ||
| impl Eq for TypedChunk {} |
There was a problem hiding this comment.
can_accumulate_with / is_batchable_with
Co-authored-by: Clément Renault <clement@meilisearch.com>
Co-authored-by: Clément Renault <clement@meilisearch.com>
ManyTheFish
force-pushed
the
remove-merging-phase-from-indexing
branch
from
a14bf6f
to
48026aa
Compare
This was referenced Feb 14, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Implement several enhancements to reduce the indexing time.
Steps
Running Indexing process
mainEach type of data is written after a merging phase:
remove-merging-phase-from-indexingWhen the extraction of a chunk is finished, the data is written:
Related
This PR removes the appending writes on several indexing parts, which may fix #4300. However, all of the appending writes are not removed. There are 2 remaining calls that could trigger this bug: