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>
608: Fix soft deleted documents r=ManyTheFish a=ManyTheFish
When we replaced or updated some documents, the indexing was skipping the replaced documents.
Related to https://github.com/meilisearch/meilisearch/issues/2672
Co-authored-by: ManyTheFish <many@meilisearch.com>
594: Fix(Search): Fix phrase search candidates computation r=Kerollmops a=ManyTheFish
This bug is an old bug but was hidden by the proximity criterion,
Phrase searches were always returning an empty candidates list when the proximity criterion is deactivated.
Before the fix, we were trying to find any words[n] near words[n]
instead of finding any words[n] near words[n+1], for example:
for a phrase search '"Hello world"' we were searching for "hello" near "hello" first, instead of "hello" near "world".
Co-authored-by: ManyTheFish <many@meilisearch.com>
NOTE: The token_at_depth is method is a bit useless now, as the only
cases where there would be a toke at depth 1000 are the cases where
the parser already stack-overflowed earlier.
Example: (((((... (x=1) ...)))))
602: Use mimalloc as the default allocator r=Kerollmops a=loiclec
## What does this PR do?
Use mimalloc as the global allocator for milli's benchmarks on macOS.
## Why?
On Linux, we use jemalloc, which is a very fast allocator. But on macOS, we currently use the system allocator, which is very slow. In practice, this difference in allocator speed means that it is difficult to gain insight into milli's performance by running benchmarks locally on the Mac.
By using mimalloc, which is another excellent allocator, we reduce the speed difference between the two platforms.
Co-authored-by: Loïc Lecrenier <loic@meilisearch.com>
New full snapshot:
---
source: milli/src/update/word_prefix_pair_proximity_docids.rs
---
5 a 1 [101, ]
5 a 2 [101, ]
5 am 1 [101, ]
5 b 4 [101, ]
5 be 4 [101, ]
am a 3 [101, ]
amazing a 1 [100, ]
amazing a 2 [100, ]
amazing a 3 [100, ]
amazing an 1 [100, ]
amazing an 2 [100, ]
amazing b 2 [100, ]
amazing be 2 [100, ]
an a 1 [100, ]
an a 2 [100, 202, ]
an am 1 [100, ]
an an 2 [100, ]
an b 3 [100, ]
an be 3 [100, ]
and a 2 [100, ]
and a 3 [100, ]
and a 4 [100, ]
and am 2 [100, ]
and an 3 [100, ]
and b 1 [100, ]
and be 1 [100, ]
at a 1 [100, 202, ]
at a 2 [100, 101, ]
at a 3 [100, ]
at am 2 [100, 101, ]
at an 1 [100, 202, ]
at an 3 [100, ]
at b 3 [101, ]
at b 4 [100, ]
at be 3 [101, ]
at be 4 [100, ]
beautiful a 2 [100, ]
beautiful a 3 [100, ]
beautiful a 4 [100, ]
beautiful am 3 [100, ]
beautiful an 2 [100, ]
beautiful an 4 [100, ]
bell a 2 [101, ]
bell a 4 [101, ]
bell am 4 [101, ]
extraordinary a 2 [202, ]
extraordinary a 3 [202, ]
extraordinary an 2 [202, ]
house a 3 [100, 202, ]
house a 4 [100, 202, ]
house am 4 [100, ]
house an 3 [100, 202, ]
house b 2 [100, ]
house be 2 [100, ]
rings a 1 [101, ]
rings a 3 [101, ]
rings am 3 [101, ]
rings b 2 [101, ]
rings be 2 [101, ]
the a 3 [101, ]
the b 1 [101, ]
the be 1 [101, ]
New snapshot (yes, it's wrong as well, it will get fixed later):
---
source: milli/src/update/word_prefix_pair_proximity_docids.rs
---
5 a 1 [101, ]
5 a 2 [101, ]
5 am 1 [101, ]
5 b 4 [101, ]
5 be 4 [101, ]
am a 3 [101, ]
amazing a 1 [100, ]
amazing a 2 [100, ]
amazing a 3 [100, ]
amazing an 1 [100, ]
amazing an 2 [100, ]
amazing b 2 [100, ]
amazing be 2 [100, ]
an a 1 [100, ]
an a 2 [100, 202, ]
an am 1 [100, ]
an b 3 [100, ]
an be 3 [100, ]
and a 2 [100, ]
and a 3 [100, ]
and a 4 [100, ]
and b 1 [100, ]
and be 1 [100, ]
d\0 0 [100, 202, ]
an an 2 [100, ]
and am 2 [100, ]
and an 3 [100, ]
at a 2 [100, 101, ]
at a 3 [100, ]
at am 2 [100, 101, ]
at an 1 [100, 202, ]
at an 3 [100, ]
at b 3 [101, ]
at b 4 [100, ]
at be 3 [101, ]
at be 4 [100, ]
beautiful a 2 [100, ]
beautiful a 3 [100, ]
beautiful a 4 [100, ]
beautiful am 3 [100, ]
beautiful an 2 [100, ]
beautiful an 4 [100, ]
bell a 2 [101, ]
bell a 4 [101, ]
bell am 4 [101, ]
extraordinary a 2 [202, ]
extraordinary a 3 [202, ]
extraordinary an 2 [202, ]
house a 4 [100, 202, ]
house a 4 [100, ]
house am 4 [100, ]
house an 3 [100, 202, ]
house b 2 [100, ]
house be 2 [100, ]
rings a 1 [101, ]
rings a 3 [101, ]
rings am 3 [101, ]
rings b 2 [101, ]
rings be 2 [101, ]
the a 3 [101, ]
the b 1 [101, ]
the be 1 [101, ]
556: Add EXISTS filter r=loiclec a=loiclec
## What does this PR do?
Fixes issue [#2484](https://github.com/meilisearch/meilisearch/issues/2484) in the meilisearch repo.
It creates a `field EXISTS` filter which selects all documents containing the `field` key.
For example, with the following documents:
```json
[{
"id": 0,
"colour": []
},
{
"id": 1,
"colour": ["blue", "green"]
},
{
"id": 2,
"colour": 145238
},
{
"id": 3,
"colour": null
},
{
"id": 4,
"colour": {
"green": []
}
},
{
"id": 5,
"colour": {}
},
{
"id": 6
}]
```
Then the filter `colour EXISTS` selects the ids `[0, 1, 2, 3, 4, 5]`. The filter `colour NOT EXISTS` selects `[6]`.
## Details
There is a new database named `facet-id-exists-docids`. Its keys are field ids and its values are bitmaps of all the document ids where the corresponding field exists.
To create this database, the indexing part of milli had to be adapted. The implementation there is basically copy/pasted from the code handling the `facet-id-f64-docids` database, with appropriate modifications in place.
There was an issue involving the flattening of documents during (re)indexing. Previously, the following JSON:
```json
{
"id": 0,
"colour": [],
"size": {}
}
```
would be flattened to:
```json
{
"id": 0
}
```
prior to being given to the extraction pipeline.
This transformation would lose the information that is needed to populate the `facet-id-exists-docids` database. Therefore, I have also changed the implementation of the `flatten-serde-json` crate. Now, as it traverses the Json, it keeps track of which key was encountered. Then, at the end, if a previously encountered key is not present in the flattened object, it adds that key to the object with an empty array as value. For example:
```json
{
"id": 0,
"colour": {
"green": [],
"blue": 1
},
"size": {}
}
```
becomes
```json
{
"id": 0,
"colour": [],
"colour.green": [],
"colour.blue": 1,
"size": []
}
```
Co-authored-by: Kerollmops <clement@meilisearch.com>
This bug is an old bug but was hidden by the proximity criterion,
Phrase search were always returning an empty candidates list.
Before the fix, we were trying to find any words[n] near words[n]
instead of finding any words[n] near words[n+1], for example:
for a phrase search '"Hello world"' we were searching for "hello" near "hello" first, instead of "hello" near "world".
561: Enriched documents batch reader r=curquiza a=Kerollmops
~This PR is based on #555 and must be rebased on main after it has been merged to ease the review.~
This PR contains the work in #555 and can be merged on main as soon as reviewed and approved.
- [x] Create an `EnrichedDocumentsBatchReader` that contains the external documents id.
- [x] Extract the primary key name and make it accessible in the `EnrichedDocumentsBatchReader`.
- [x] Use the external id from the `EnrichedDocumentsBatchReader` in the `Transform::read_documents`.
- [x] Remove the `update_primary_key` from the _transform.rs_ file.
- [x] Really generate the auto-generated documents ids.
- [x] Insert the (auto-generated) document ids in the document while processing it in `Transform::read_documents`.
Co-authored-by: Kerollmops <clement@meilisearch.com>
The idea is to directly create a sorted and merged list of bitmaps
in the form of a BTreeMap<FieldId, RoaringBitmap> instead of creating
a grenad::Reader where the keys are field_id and the values are docids.
Then we send that BTreeMap to the thing that handles TypedChunks, which
inserts its content into the database.
When a document deletion occurs, instead of deleting the document we mark it as deleted
in the new “soft deleted” bitmap. It is then removed from the search, and all the other
endpoints.
564: Rename the limitedTo parameter into maxTotalHits r=curquiza a=Kerollmops
This PR is related to https://github.com/meilisearch/meilisearch/issues/2542, it renames the `limitedTo` parameter into `maxTotalHits`.
Co-authored-by: Kerollmops <clement@meilisearch.com>
563: Improve the `estimatedNbHits` when a `distinctAttribute` is specified r=irevoire a=Kerollmops
This PR is related to https://github.com/meilisearch/meilisearch/issues/2532 but it doesn't fix it entirely. It improves it by computing the excluded documents (the ones with an already-seen distinct value) before stopping the loop, I think it was a mistake and should always have been this way.
The reason it doesn't fix the issue is that Meilisearch is lazy, just to be sure not to compute too many things and answer by taking too much time. When we deduplicate the documents by their distinct value we must do it along the water, everytime we see a new document we check that its distinct value of it doesn't collide with an already returned document.
The reason we can see the correct result when enough documents are fetched is that we were lucky to see all of the different distinct values possible in the dataset and all of the deduplication was done, no document can be returned.
If we wanted to implement that to have a correct `extimatedNbHits` every time we should have done a pass on the whole set of possible distinct values for the distinct attribute and do a big intersection, this could cost a lot of CPU cycles.
Co-authored-by: Kerollmops <clement@meilisearch.com>
552: Fix escaped quotes in filter r=Kerollmops a=irevoire
Will fix https://github.com/meilisearch/meilisearch/issues/2380
The issue was that in the evaluation of the filter, I was using the deref implementation instead of calling the `value` method of my token.
To avoid the problem happening again, I removed the deref implementation; now, you need to either call the `lexeme` or the `value` methods but can't rely on a « default » implementation to get a string out of a token.
Co-authored-by: Tamo <tamo@meilisearch.com>
547: Update version for next release (v0.29.1) r=Kerollmops a=curquiza
A new milli version will be released once this PR is merged https://github.com/meilisearch/milli/pull/543
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
541: Update version for next release (v0.29.0) r=ManyTheFish a=curquiza
Need to update the version since #540 was merged and breaking
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
535: Reintroduce the max values by facet limit r=ManyTheFish a=Kerollmops
This PR reintroduces the max values by facet limit this is related to https://github.com/meilisearch/meilisearch/issues/2349.
~I would like some help in deciding on whether I keep the default 100 max values in milli and set up the `FacetDistribution` settings in Meilisearch to use 1000 as the new value, I expose the `max_values_by_facet` for this purpose.~
I changed the default value to 1000 and the max to 10000, thank you `@ManyTheFish` for the help!
Co-authored-by: Kerollmops <clement@meilisearch.com>
538: speedup exact words r=Kerollmops a=MarinPostma
This PR make `exact_words` return an `Option` instead of an empty set, since set creation is costly, as noticed by `@kerollmops.`
I was not convinces that this was the cause for all of the performance drop we measured, and then realized that methods that initialized it were called recursively which caused initialization times to add up. While the first fix solves the issue when not using exact words, using exact word remained way more expensive that it should be. To address this issue, the exact words are cached into the `Context`, so they are only initialized once.
Co-authored-by: ad hoc <postma.marin@protonmail.com>
525: Simplify the error creation with thiserror r=irevoire a=irevoire
I introduced [`thiserror`](https://docs.rs/thiserror/latest/thiserror/) to implements all the `Display` trait and most of the `impl From<xxx> for yyy` in way less lines.
And then I introduced a cute macro to implements the `impl<X, Y, Z> From<X> for Z where Y: From<X>, Z: From<X>` more easily.
Co-authored-by: Tamo <tamo@meilisearch.com>
523: Improve geosearch error messages r=irevoire a=irevoire
Improve the geosearch error messages (#488).
And try to parse the string as specified in https://github.com/meilisearch/meilisearch/issues/2354
Co-authored-by: Tamo <tamo@meilisearch.com>
520: fix mistake in Settings initialization r=irevoire a=MarinPostma
fix settings not being correctly initialized and add a test to make sure that they are in the future.
fix https://github.com/meilisearch/meilisearch/issues/2358
Co-authored-by: ad hoc <postma.marin@protonmail.com>
518: Return facets even when there is no value associated to it r=Kerollmops a=Kerollmops
This PR is related to https://github.com/meilisearch/meilisearch/issues/2352 and should fix the issue when Meilisearch is up-to-date with this PR.
Co-authored-by: Kerollmops <clement@meilisearch.com>
511: Update version in every workspace r=curquiza a=curquiza
Checked with `@Kerollmops`
- Update the version into every workspace (the current version is v0.27.0, but I forgot to update it for the previous release)
- add `publish = false` except in `milli` workspace.
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
514: Stop flattening every field r=Kerollmops a=irevoire
When we need to flatten a document:
* The primary key contains a `.`.
* Some fields need to be flattened
Instead of flattening the whole object and thus creating a lot of allocations with the `serde_json_flatten_crate`, we instead generate a minimal sub-object containing only the fields that need to be flattened.
That should create fewer allocations and thus index faster.
---------
```
group indexing_main_e1e362fa indexing_stop-flattening-every-field_40d1bd6b
----- ---------------------- ---------------------------------------------
indexing/Indexing geo_point 1.99 23.7±0.23s ? ?/sec 1.00 11.9±0.21s ? ?/sec
indexing/Indexing movies in three batches 1.00 18.2±0.24s ? ?/sec 1.01 18.3±0.29s ? ?/sec
indexing/Indexing movies with default settings 1.00 17.5±0.09s ? ?/sec 1.01 17.7±0.26s ? ?/sec
indexing/Indexing songs in three batches with default settings 1.00 64.8±0.47s ? ?/sec 1.00 65.1±0.49s ? ?/sec
indexing/Indexing songs with default settings 1.00 54.9±0.99s ? ?/sec 1.01 55.7±1.34s ? ?/sec
indexing/Indexing songs without any facets 1.00 50.6±0.62s ? ?/sec 1.01 50.9±1.05s ? ?/sec
indexing/Indexing songs without faceted numbers 1.00 54.0±1.14s ? ?/sec 1.01 54.7±1.13s ? ?/sec
indexing/Indexing wiki 1.00 996.2±8.54s ? ?/sec 1.02 1021.1±30.63s ? ?/sec
indexing/Indexing wiki in three batches 1.00 1136.8±9.72s ? ?/sec 1.00 1138.6±6.59s ? ?/sec
```
So basically everything slowed down a liiiiiittle bit except the dataset with a nested field which got twice faster
Co-authored-by: Tamo <tamo@meilisearch.com>
505: normalize exact words r=curquiza a=MarinPostma
Normalize the exact words, as specified in the specification.
Co-authored-by: ad hoc <postma.marin@protonmail.com>
483: Enhance matching words r=Kerollmops a=ManyTheFish
# Summary
Enhance milli word-matcher making it handle match computing and cropping.
# Implementation
## Computing best matches for cropping
Before we were considering that the first match of the attribute was the best one, this was accurate when only one word was searched but was missing the target when more than one word was searched.
Now we are searching for the best matches interval to crop around, the chosen interval is the one:
1) that have the highest count of unique matches
> for example, if we have a query `split the world`, then the interval `the split the split the` has 5 matches but only 2 unique matches (1 for `split` and 1 for `the`) where the interval `split of the world` has 3 matches and 3 unique matches. So the interval `split of the world` is considered better.
2) that have the minimum distance between matches
> for example, if we have a query `split the world`, then the interval `split of the world` has a distance of 3 (2 between `split` and `the`, and 1 between `the` and `world`) where the interval `split the world` has a distance of 2. So the interval `split the world` is considered better.
3) that have the highest count of ordered matches
> for example, if we have a query `split the world`, then the interval `the world split` has 2 ordered words where the interval `split the world` has 3. So the interval `split the world` is considered better.
## Cropping around the best matches interval
Before we were cropping around the interval without checking the context.
Now we are cropping around words in the same context as matching words.
This means that we will keep words that are farther from the matching words but are in the same phrase, than words that are nearer but separated by a dot.
> For instance, for the matching word `Split` the text:
`Natalie risk her future. Split The World is a book written by Emily Henry. I never read it.`
will be cropped like:
`…. Split The World is a book written by Emily Henry. …`
and not like:
`Natalie risk her future. Split The World is a book …`
Co-authored-by: ManyTheFish <many@meilisearch.com>
We need to store all the external id (primary key) in a hashmap
associated to their internal id during.
The smartstring remove heap allocation / memory usage and should
improve the cache locality.
486: Update version (v0.25.0) r=curquiza a=curquiza
v0.25.0 will be released once #478 is merged
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
472: Remove useless variables in proximity r=Kerollmops a=ManyTheFish
Was passing by plane sweep algorithm to find some inspiration, and I discover that we have useless variables that were not detected because of the recursive function.
Co-authored-by: ManyTheFish <many@meilisearch.com>
466: Bump version to 0.23.1 r=curquiza a=Kerollmops
This PR bumps the crate versions to 0.23.1. Nothing seems to be breaking in the next release.
Co-authored-by: Kerollmops <clement@meilisearch.com>
