4090: Diff indexing r=ManyTheFish a=ManyTheFish
This pull request aims to reduce the indexing time by computing a difference between the data added to the index and the data removed from the index before writing in LMDB.
## Why focus on reducing the writings in LMDB?
The indexing in Meilisearch is split into 3 main phases:
1) The computing or the extraction of the data (Multi-threaded)
2) The writing of the data in LMDB (Mono-threaded)
3) The processing of the prefix databases (Mono-threaded)
see below:
Because the writing is mono-threaded, it represents a bottleneck in the indexing, reducing the number of writes in LMDB will reduce the pressure on the main thread and should reduce the global time spent on the indexing.
## Give Feedback
We created [a dedicated discussion](https://github.com/meilisearch/meilisearch/discussions/4196) for users to try this new feature and to give feedback on bugs or performance issues.
## Technical approach
### Part 1: merge the addition and the deletion process
This part:
a) Aims to reduce the time spent on indexing only the filterable/sortable fields of documents, for example:
- Updating the number of "likes" or "stars" of a song or a movie
- Updating the "stock count" or the "price" of a product
b) Aims to reduce the time spent on writing in LMDB which should reduce the global indexing time for the highly multi-threaded machines by reducing the writing bottleneck.
c) Aims to reduce the average time spent to delete documents without having to keep the soft-deleted documents implementation
- [x] Create a preprocessing function that creates the diff-based documents chuck (`OBKV<fid, OBKV<AddDel, value>>`)
- [x] and clearly separate the faceted fields and the searchable fields in two different chunks
- Change the parameters of the input extractor by taking an `OBKV<fid, OBKV<AddDel, value>>` instead of `OBKV<fid, value>`.
- [x] extract_docid_word_positions
- [x] extract_geo_points
- [x] extract_vector_points
- [x] extract_fid_docid_facet_values
- Adapt the searchable extractors to the new diff-chucks
- [x] extract_fid_word_count_docids
- [x] extract_word_pair_proximity_docids
- [x] extract_word_position_docids
- [x] extract_word_docids
- Adapt the facet extractors to the new diff-chucks
- [x] extract_facet_number_docids
- [x] extract_facet_string_docids
- [x] extract_fid_docid_facet_values
- [x] FacetsUpdate
- [x] Adapt the prefix database extractors ⚠️⚠️
- [x] Make the LMDB writer remove the document_ids to delete at the same time the new document_ids are added
- [x] Remove document deletion pipeline
- [x] remove `new_documents_ids` entirely and `replaced_documents_ids`
- [x] reuse extracted external id from transform instead of re-extracting in `TypedChunks::Documents`
- [x] Remove deletion pipeline after autobatcher
- [x] remove autobatcher deletion pipeline
- [x] everything uses `IndexOperation::DocumentOperation`
- [x] repair deletion by internal id for filter by delete
- [x] Improve the deletion via internal ids by avoiding iterating over the whole set of external document ids.
- [x] Remove soft-deleted documents
#### FIXME
- [x] field distribution is not correctly updated after deletion
- [x] missing documents in the tests of tokenizer_customization
### Part 2: Only compute the documents field by field
This part aims to reduce the global indexing time for any kind of partial document modification on any size of machine from the mono-threaded one to the highly multi-threaded one.
- [ ] Make the preprocessing function only send the fields that changed to the extractors
- [ ] remove the `word_docids` and `exact_word_docids` database and adapt the search (⚠️ could impact the search performances)
- [ ] replace the `word_pair_proximity_docids` database with a `word_pair_proximity_fid_docids` database and adapt the search (⚠️ could impact the search performances)
- [ ] Adapt the prefix database extractors ⚠️⚠️
## Technical Concerns
- The part 1 implementation could increase the indexing time for the smallest machines (with few threads) by increasing the extracting time (multi-threaded) more than the writing time (mono-threaded)
- The part 2 implementation needs to change the databases which could have a significant impact on the search performances
- The prefix databases are a bit special to process and may be a pain to adapt to the difference-based indexing
Co-authored-by: ManyTheFish <many@meilisearch.com>
Co-authored-by: Clément Renault <clement@meilisearch.com>
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
4208: Makes the dump cancellable r=Kerollmops a=irevoire
# Pull Request
Make the dump tasks cancellable even when they have already started processing.
## Related issue
Fixes https://github.com/meilisearch/meilisearch/issues/4157
Co-authored-by: Tamo <tamo@meilisearch.com>
4203: Extract external document docids from docs on deletion by filter r=Kerollmops a=dureuill
This fixes some of the performance regression observed on `diff-indexing` when doing delete-by-filter with a filter matching many documents.
To delete 19 768 771 documents (hackernews dataset, all documents matching `type = comment`), here are the observed time:
|branch (commit sha1sum)|time|speed-down factor (lower is better)|
|--|--|--|
|`main` (48865470d7)|1212.885536s (~20min)|x1.0 (baseline)|
|`diff-indexing` (523519fdbf)|5385.550543s (90min)|x4.44|
|**`diff-indexing-extract-primary-key`**(f8289cd974)|2582.323324s (43min) | x2.13|
So we're still suffering a speed-down of x2.13, but that's much better than x4.44.
---
Changes:
- Refactor the logic of PrimaryKey extraction to a struct
- Add a trait to abstract the extraction of field id from a name between `DocumentBatch` and `FieldIdMap`.
- Add `Index::external_id_of` to get the external ids of a bitmap of internal ids.
- Use this new method to add new Transform and Batch methods to remove documents that are known to be from the DB.
- Modify delete-by-filter to use the new method
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
4205: Prevent search hang on the processing index r=Kerollmops a=dureuill
Fixes#4206, an issue originally [reported on Discord](https://discord.com/channels/1006923006964154428/1148983671026618579/1148983671026618579) where having parallel search requests on more indexes than the index cache capacity would cause search requests on the currently updating index to hang until the index is done updating.
## Test setup
- Create 20 empty indexes by sending settings to them
- repeatedly send placeholder search requests to each of the indexes in a loop
- Create another index and send a significant batch of documents to index.
- Attempt to perform a search request on that last index.
- Before this PR, the search request hangs while the index update task is processing
- After this PR, the search request respond immediately even while the index update task is processing
## Changes
- When getting the handle to an index for some potentially long running batches of tasks, save it in the index scheduler.
- Drop the handle from the index-scheduler when the task is done so that we don't leak indexes.
- When getting an index from outside the task queue processor, check if there is such an handle matching the requested index. If so, skip the cache entirely and clone the handle.
Co-authored-by: Louis Dureuil <louis.dureuil@xinra.net>
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
4204: Throw error when the vector search is sent with the wrong size r=Kerollmops a=dureuill
# Pull Request
## Related issue
Fixes#4201
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
