This database can easily contain millions of entries. Thus, iterating
over it can be very expensive.
For regular `documentAdditionOrUpdate` tasks, `del_prefix_fst_words`
will always be empty. Thus, we can save a significant amount of time
by adding this `if !del_prefix_fst_words.is_empty()` condition.
The code's behaviour remains completely unchanged.
733: Avoid a prefix-related worst-case scenario in the proximity criterion r=loiclec a=loiclec
# Pull Request
## Related issue
Somewhat fixes (until merged into meilisearch) https://github.com/meilisearch/meilisearch/issues/3118
## What does this PR do?
When a query ends with a word and a prefix, such as:
```
word pr
```
Then we first determine whether `pre` *could possibly* be in the proximity prefix database before querying it. There are then three possibilities:
1. `pr` is not in any prefix cache because it is not the prefix of many words. We don't query the proximity prefix database. Instead, we list all the word derivations of `pre` through the FST and query the regular proximity databases.
2. `pr` is in the prefix cache but cannot be found in the proximity prefix databases. **In this case, we partially disable the proximity ranking rule for the pair `word pre`.** This is done as follows:
1. Only find the documents where `word` is in proximity to `pre` **exactly** (no derivations)
2. Otherwise, assume that their proximity in all the documents in which they coexist is >= 8
3. `pr` is in the prefix cache and can be found in the proximity prefix databases. In this case we simply query the proximity prefix databases.
Note that if a prefix is longer than 2 bytes, then it cannot be in the proximity prefix databases. Also, proximities larger than 4 are not present in these databases either. Therefore, the impact on relevancy is:
1. For common prefixes of one or two letters: we no longer distinguish between proximities from 4 to 8
2. For common prefixes of more than two letters: we no longer distinguish between any proximities
3. For uncommon prefixes: nothing changes
Regarding (1), it means that these two documents would be considered equally relevant according to the proximity rule for the query `heard pr` (IF `pr` is the prefix of more than 200 words in the dataset):
```json
[
{ "text": "I heard there is a faster proximity criterion" },
{ "text": "I heard there is a faster but less relevant proximity criterion" }
]
```
Regarding (2), it means that two documents would be considered equally relevant according to the proximity rule for the query "faster pro":
```json
[
{ "text": "I heard there is a faster but less relevant proximity criterion" }
{ "text": "I heard there is a faster proximity criterion" },
]
```
But the following document would be considered more relevant than the two documents above:
```json
{ "text": "I heard there is a faster swimmer who is competing in the pro section of the competition " }
```
Note, however, that this change of behaviour only occurs when using the set-based version of the proximity criterion. In cases where there are fewer than 1000 candidate documents when the proximity criterion is called, this PR does not change anything.
---
## Performance
I couldn't use the existing search benchmarks to measure the impact of the PR, but I did some manual tests with the `songs` benchmark dataset.
```
1. 10x 'a':
- 640ms ⟹ 630ms = no significant difference
2. 10x 'b':
- set-based: 4.47s ⟹ 7.42 = bad, ~2x regression
- dynamic: 1s ⟹ 870 ms = no significant difference
3. 'Someone I l':
- set-based: 250ms ⟹ 12 ms = very good, x20 speedup
- dynamic: 21ms ⟹ 11 ms = good, x2 speedup
4. 'billie e':
- set-based: 623ms ⟹ 2ms = very good, x300 speedup
- dynamic: ~4ms ⟹ 4ms = no difference
5. 'billie ei':
- set-based: 57ms ⟹ 20ms = good, ~2x speedup
- dynamic: ~4ms ⟹ ~2ms. = no significant difference
6. 'i am getting o'
- set-based: 300ms ⟹ 60ms = very good, 5x speedup
- dynamic: 30ms ⟹ 6ms = very good, 5x speedup
7. 'prologue 1 a 1:
- set-based: 3.36s ⟹ 120ms = very good, 30x speedup
- dynamic: 200ms ⟹ 30ms = very good, 6x speedup
8. 'prologue 1 a 10':
- set-based: 590ms ⟹ 18ms = very good, 30x speedup
- dynamic: 82ms ⟹ 35ms = good, ~2x speedup
```
Performance is often significantly better, but there is also one regression in the set-based implementation with the query `b b b b b b b b b b`.
Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
736: Update charabia r=curquiza a=ManyTheFish
Update Charabia to the last version.
> We are now Romanizing Chinese characters into Pinyin.
> Note that we keep the accent because they are in fact never typed directly by the end-user, moreover, changing an accent leads to a different Chinese character, and I don't have sufficient knowledge to forecast the impact of removing accents in this context.
Co-authored-by: ManyTheFish <many@meilisearch.com>
Displays log message in the form:
```
[2022-12-21T09:19:42Z INFO milli::update::index_documents::enrich] Primary key was not specified in index. Inferred to 'id'
```
By creating snapshots and updating the format of the existing
snapshots. The next commit will apply the fix, which will show
its effects cleanly on the old and new snapshot tests
706: Limit the reindexing caused by updating settings when not needed r=curquiza a=GregoryConrad
## What does this PR do?
When updating index settings using `update::Settings`, sometimes a `reindex` of `update::Settings` is triggered when it doesn't need to be. This PR aims to prevent those unnecessary `reindex` calls.
For reference, here is a snippet from the current `execute` method in `update::Settings`:
```rust
// ...
if stop_words_updated
|| faceted_updated
|| synonyms_updated
|| searchable_updated
|| exact_attributes_updated
{
self.reindex(&progress_callback, &should_abort, old_fields_ids_map)?;
}
```
- [x] `faceted_updated` - looks good as-is ✅
- [x] `stop_words_updated` - looks good as-is ✅
- [x] `synonyms_updated` - looks good as-is ✅
- [x] `searchable_updated` - fixed in this PR
- [x] `exact_attributes_updated` - fixed in this PR
## PR checklist
Please check if your PR fulfills the following requirements:
- [x] Does this PR fix an existing issue, or have you listed the changes applied in the PR description (and why they are needed)?
- [x] Have you read the contributing guidelines?
- [x] Have you made sure that the title is accurate and descriptive of the changes?
Thank you so much for contributing to Meilisearch!
Co-authored-by: Gregory Conrad <gregorysconrad@gmail.com>
708: Reduce memory usage of the MatchingWords structure r=ManyTheFish a=loiclec
# Pull Request
## Related issue
Fixes (partially) https://github.com/meilisearch/meilisearch/issues/3115
## What does this PR do?
1. Reduces the memory usage caused by the creation of a 10-word query tree by 20x.
This is done by deduplicating the `MatchingWord` values, which are heavy because of their inner DFA. The deduplication works by wrapping each `MatchingWord` in a reference-counted box and using a hash map to determine whether a `MatchingWord` DFA already exists for a certain signature, or whether a new one needs to be built.
2. Avoid the worst-case scenario of creating a `MatchingWord` for extremely long words that cannot be indexed by milli.
Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
697: Fix bug in prefix DB indexing r=loiclec a=loiclec
Where the batch's information was not properly updated in cases where only the proximity changed between two consecutive word pair proximities.
Closes partially https://github.com/meilisearch/meilisearch/issues/3043
Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
1. Handle keys with variable length correctly
This fixes https://github.com/meilisearch/meilisearch/issues/3042 and
is easily reproducible with the updated fuzz tests, which now generate
keys with variable lengths.
2. Prevent adding facets to the database if their encoded value does
not satisfy `valid_lmdb_key`.
This fixes an indexing failure when a document had a filterable
attribute containing a value whose length is higher than ~500 bytes.
668: Fix many Clippy errors part 2 r=ManyTheFish a=ehiggs
This brings us a step closer to enforcing clippy on each build.
# Pull Request
## Related issue
This does not fix any issue outright, but it is a second round of fixes for clippy after https://github.com/meilisearch/milli/pull/665. This should contribute to fixing https://github.com/meilisearch/milli/pull/659.
## What does this PR do?
Satisfies many issues for clippy. The complaints are mostly:
* Passing reference where a variable is already a reference.
* Using clone where a struct already implements `Copy`
* Using `ok_or_else` when it is a closure that returns a value instead of using the closure to call function (hence we use `ok_or`)
* Unambiguous lifetimes don't need names, so we can just use `'_`
* Using `return` when it is not needed as we are on the last expression of a function.
## PR checklist
Please check if your PR fulfills the following requirements:
- [x] Does this PR fix an existing issue, or have you listed the changes applied in the PR description (and why they are needed)?
- [x] Have you read the contributing guidelines?
- [x] Have you made sure that the title is accurate and descriptive of the changes?
Thank you so much for contributing to Meilisearch!
Co-authored-by: Ewan Higgs <ewan.higgs@gmail.com>
e.g. add one facet value incrementally with a group_size = X and then
add another one with group_size = Y
It is not actually possible to do so with the public API of milli,
but I wanted to make sure the algorithm worked well in those cases
anyway.
The bugs were found by fuzzing the code with fuzzcheck, which I've added
to milli as a conditional dev-dependency. But it can be removed later.
616: Introduce an indexation abortion function when indexing documents r=Kerollmops a=Kerollmops
Co-authored-by: Kerollmops <clement@meilisearch.com>
Co-authored-by: Clément Renault <clement@meilisearch.com>
Most of these are calling clone when the struct supports Copy.
Many are using & and &mut on `self` when the function they are called
from already has an immutable or mutable borrow so this isn't needed.
I tried to stay away from actual changes or places where I'd have to
name fresh variables.
635: Use an unstable algorithm for `grenad::Sorter` when possible r=Kerollmops a=loiclec
# Pull Request
## What does this PR do?
Use an unstable algorithm to sort the internal vector used by `grenad::Sorter` whenever possible to speed up indexing.
In practice, every time the merge function creates a `RoaringBitmap`, we use an unstable sort. For every other merge function, such as `keep_first`, `keep_last`, etc., a stable sort is used.
Co-authored-by: Loïc Lecrenier <loic@meilisearch.com>
587: Word prefix pair proximity docids indexation refactor r=Kerollmops a=loiclec
# Pull Request
## What does this PR do?
Refactor the code of `WordPrefixPairProximityDocIds` to make it much faster, fix a bug, and add a unit test.
## Why is it faster?
Because we avoid using a sorter to insert the (`word1`, `prefix`, `proximity`) keys and their associated bitmaps, and thus we don't have to sort a potentially very big set of data. I have also added a couple of other optimisations:
1. reusing allocations
2. using a prefix trie instead of an array of prefixes to get all the prefixes of a word
3. inserting directly into the database instead of putting the data in an intermediary grenad when possible. Also avoid checking for pre-existing values in the database when we know for certain that they do not exist.
## What bug was fixed?
When reindexing, the `new_prefix_fst_words` prefixes may look like:
```
["ant", "axo", "bor"]
```
which we group by first letter:
```
[["ant", "axo"], ["bor"]]
```
Later in the code, if we have the word2 "axolotl", we try to find which subarray of prefixes contains its prefixes. This check is done with `word2.starts_with(subarray_prefixes[0])`, but `"axolotl".starts_with("ant")` is false, and thus we wrongly think that there are no prefixes in `new_prefix_fst_words` that are prefixes of `axolotl`.
## StrStrU8Codec
I had to change the encoding of `StrStrU8Codec` to make the second string null-terminated as well. I don't think this should be a problem, but I may have missed some nuances about the impacts of this change.
## Requests when reviewing this PR
I have explained what the code does in the module documentation of `word_pair_proximity_prefix_docids`. It would be nice if someone could read it and give their opinion on whether it is a clear explanation or not.
I also have a couple questions regarding the code itself:
- Should we clean up and factor out the `PrefixTrieNode` code to try and make broader use of it outside this module? For now, the prefixes undergo a few transformations: from FST, to array, to prefix trie. It seems like it could be simplified.
- I wrote a function called `write_into_lmdb_database_without_merging`. (1) Are we okay with such a function existing? (2) Should it be in `grenad_helpers` instead?
## Benchmark Results
We reduce the time it takes to index about 8% in most cases, but it varies between -3% and -20%.
```
group indexing_main_ce90fc62 indexing_word-prefix-pair-proximity-docids-refactor_cbad2023
----- ---------------------- ------------------------------------------------------------
indexing/-geo-delete-facetedNumber-facetedGeo-searchable- 1.00 1893.0±233.03µs ? ?/sec 1.01 1921.2±260.79µs ? ?/sec
indexing/-movies-delete-facetedString-facetedNumber-searchable- 1.05 9.4±3.51ms ? ?/sec 1.00 9.0±2.14ms ? ?/sec
indexing/-movies-delete-facetedString-facetedNumber-searchable-nested- 1.22 18.3±11.42ms ? ?/sec 1.00 15.0±5.79ms ? ?/sec
indexing/-songs-delete-facetedString-facetedNumber-searchable- 1.00 41.4±4.20ms ? ?/sec 1.28 53.0±13.97ms ? ?/sec
indexing/-wiki-delete-searchable- 1.00 285.6±18.12ms ? ?/sec 1.03 293.1±16.09ms ? ?/sec
indexing/Indexing geo_point 1.03 60.8±0.45s ? ?/sec 1.00 58.8±0.68s ? ?/sec
indexing/Indexing movies in three batches 1.14 16.5±0.30s ? ?/sec 1.00 14.5±0.24s ? ?/sec
indexing/Indexing movies with default settings 1.11 13.7±0.07s ? ?/sec 1.00 12.3±0.28s ? ?/sec
indexing/Indexing nested movies with default settings 1.10 10.6±0.11s ? ?/sec 1.00 9.6±0.15s ? ?/sec
indexing/Indexing nested movies without any facets 1.11 9.4±0.15s ? ?/sec 1.00 8.5±0.10s ? ?/sec
indexing/Indexing songs in three batches with default settings 1.18 66.2±0.39s ? ?/sec 1.00 56.0±0.67s ? ?/sec
indexing/Indexing songs with default settings 1.07 58.7±1.26s ? ?/sec 1.00 54.7±1.71s ? ?/sec
indexing/Indexing songs without any facets 1.08 53.1±0.88s ? ?/sec 1.00 49.3±1.43s ? ?/sec
indexing/Indexing songs without faceted numbers 1.08 57.7±1.33s ? ?/sec 1.00 53.3±0.98s ? ?/sec
indexing/Indexing wiki 1.06 1051.1±21.46s ? ?/sec 1.00 989.6±24.55s ? ?/sec
indexing/Indexing wiki in three batches 1.20 1184.8±8.93s ? ?/sec 1.00 989.7±7.06s ? ?/sec
indexing/Reindexing geo_point 1.04 67.5±0.75s ? ?/sec 1.00 64.9±0.32s ? ?/sec
indexing/Reindexing movies with default settings 1.12 13.9±0.17s ? ?/sec 1.00 12.4±0.13s ? ?/sec
indexing/Reindexing songs with default settings 1.05 60.6±0.84s ? ?/sec 1.00 57.5±0.99s ? ?/sec
indexing/Reindexing wiki 1.07 1725.0±17.92s ? ?/sec 1.00 1611.4±9.90s ? ?/sec
```
Co-authored-by: Loïc Lecrenier <loic@meilisearch.com>
