* Optimizing top K using Adrien Grand's ideas
https://jpountz.github.io/2025/08/28/compiled-vs-vectorized-search-engine-edition.html
* Suffix-sum pruning for multi-term intersection candidates
After scoring each secondary in Phase 2, check whether remaining
secondaries' block_max scores can still beat the threshold. Skip
to the next candidate early if impossible, avoiding expensive seeks
into later secondaries.
Improves three-term intersection by ~8% on the balanced benchmark
while keeping two-term performance neutral.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Claude CR comment
* Removed 16 term scorer limit.
---------
Co-authored-by: Paul Masurel <paul.masurel@datadoghq.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix O(2^n) query parser regression for deeply-nested queries
The top-level `ast()` parser used `alt((boolean_expr, single_leaf))` at
every group level. When the group contained a single leaf with no
trailing operand, `boolean_expr` would parse `occur_leaf` (recursing
into the inner group), fail at `multispace1`, backtrack, and then
`single_leaf` would re-parse `occur_leaf` from scratch. Every nesting
level doubled the work, giving O(2^n) time for queries like
`(((((title:test)))))`.
Parse `occur_leaf` once and peek ahead for a trailing operand instead
of backtracking. This keeps parsing O(n) and also avoids the duplicate
parse for simple single-leaf queries.
Fixes#2498.
Measured on the issue reproducer (release build):
depth before after
20 0.87 s <1 us
25 28.23 s <1 us
60 (years) ~5 us
Non-pathological queries are unaffected or slightly faster:
query before after
hello 650 ns 308 ns
a AND b AND c 1380 ns 1364 ns
title:rust AND (...) 3426 ns 3460 ns
All 53 existing grammar tests and 56 query_parser tests pass. Adds a
regression test at depth 60 that would not complete under the old
parser.
* Add ignored benchmark for nested query parsing at depth 20/21
Matches the depths from issue #2498 which reported 0.87 s / 1.72 s
under the regression. With the fix these parse in single-digit
microseconds. Runs via:
cargo test -p tantivy-query-grammar --release bench_deeply_nested \
-- --ignored --nocapture
* Propagate Err::Failure and Err::Incomplete from operand parser
`alt((boolean_expr, single_leaf))` only retried on `Err::Error` and
propagated `Err::Failure` and `Err::Incomplete`. The replacement was
catching all three with `Err(_)`, which would silently fall back to
a single leaf if any cut point were ever added to `operand_leaf` or
its descendants. Match specifically on `Err::Error` to preserve the
original `alt` semantics.
* Replace inline bench with binggan bench in benches/
Move the nested-query benchmark out of the query-grammar test module
and into a proper binggan benchmark at benches/query_parser_nested.rs,
registered as a harnessless bench in Cargo.toml. Keeps the correctness
regression test (depth 60) in place.
Run with: cargo bench --bench query_parser_nested
* Fix rustfmt import ordering in query_parser_nested bench
When a string cardinality aggregation is nested it end up being applied to different buckets.
Dictionary encoding relies on a different dictionaries for each segment.
As a result, during segment collection, we only collect term ordinals in a HashSet, and decode them in the
term dictionary at the end of collection.
Before this PR, this decoding phase was done once for each bucket, causing the same work to be done over and over. This PR introduce a coupon cache. The HLL sketch relies on a hash of the string values.
We populate the cache before bucket collection, and get our values from it.
This PR also rename "caching" "buffering" in aggregation (it was never caching), and does several cleanups.
* improve bench
* add more tests for new collection type
* one collector per agg request instead per bucket
In this refactoring a collector knows in which bucket of the parent
their data is in. This allows to convert the previous approach of one
collector per bucket to one collector per request.
low card bucket optimization
* reduce dynamic dispatch, faster term agg
* use radix map, fix prepare_max_bucket
use paged term map in term agg
use special no sub agg term map impl
* specialize columntype in stats
* remove stacktrace bloat, use &mut helper
increase cache to 2048
* cleanup
remove clone
move data in term req, single doc opt for stats
* add comment
* share column block accessor
* simplify fetch block in column_block_accessor
* split subaggcache into two trait impls
* move partitions to heap
* fix name, add comment
---------
Co-authored-by: Pascal Seitz <pascal.seitz@gmail.com>
* seek_exact + cost based intersection
Adds `seek_exact` and `cost` to `DocSet` for a more efficient intersection.
Unlike `seek`, `seek_exact` does not require the DocSet to advance to the next hit, if the target does not exist.
`cost` allows to address the different DocSet types and their cost
model and is used to determine the DocSet that drives the intersection.
E.g. fast field range queries may do a full scan. Phrase queries load the positions to check if a we have a hit.
They both have a higher cost than their size_hint would suggest.
Improves `size_hint` estimation for intersection and union, by having a
estimation based on random distribution with a co-location factor.
Refactor range query benchmark.
Closes#2531
*Future Work*
Implement `seek_exact` for BufferedUnionScorer and RangeDocSet (fast field range queries)
Evaluate replacing `seek` with `seek_exact` to reduce code complexity
* Apply suggestions from code review
Co-authored-by: Paul Masurel <paul@quickwit.io>
* add API contract verfication
* impl seek_exact on union
* rename seek_exact
* add mixed AND OR test, fix buffered_union
* Add a proptest of BooleanQuery. (#2690)
* fix build
* Increase the document count.
* fix merge conflict
* fix debug assert
* Fix compilation errors after rebase
- Remove duplicate proptest_boolean_query module
- Remove duplicate cost() method implementations
- Fix TopDocs API usage (add .order_by_score())
- Remove duplicate imports
- Remove unused variable assignments
---------
Co-authored-by: Paul Masurel <paul@quickwit.io>
Co-authored-by: Pascal Seitz <pascal.seitz@datadoghq.com>
Co-authored-by: Stu Hood <stuhood@gmail.com>
* Refactoring of the score tweaker into `SortKeyComputer`s to unlock two features.
- Allow lazy evaluation of score. As soon as we identified that a doc won't
reach the topK threshold, we can stop the evaluation.
- Allow for a different segment level score, segment level score and their conversion.
This PR breaks public API, but fixing code is straightforward.
* Bumping tantivy version
---------
Co-authored-by: Paul Masurel <paul.masurel@datadoghq.com>
This introduce an optimization of top level term aggregation on field with a low cardinality.
We then use a Vec as the underlying map.
In addition, we buffer subaggregations.
---------
Co-authored-by: Pascal Seitz <pascal.seitz@datadoghq.com>
Co-authored-by: Paul Masurel <paul@quickwit.io>
* Initial impl
* Added `Filter` impl in `build_single_agg_segment_collector_with_reader` + Added tests
* Added `Filter(FilterBucketResult)` + Made tests work.
* Fixed type issues.
* Fixed a test.
* 8a7a73a: Pass `segment_reader`
* Added more tests.
* Improved parsing + tests
* refactoring
* Added more tests.
* refactoring: moved parsing code under QueryParser
* Use Tantivy syntax instead of ES
* Added a sanity check test.
* Simplified impl + tests
* Added back tests in a more maintable way
* nitz.
* nitz
* implemented very simple fast-path
* improved a comment
* implemented fast field support
* Used `BoundsRange`
* Improved fast field impl + tests
* Simplified execution.
* Fixed exports + nitz
* Improved the tests to check to the expected result.
* Improved test by checking the whole result JSON
* Removed brittle perf checks.
* Added efficiency verification tests.
* Added one more efficiency check test.
* Improved the efficiency tests.
* Removed unnecessary parsing code + added direct Query obj
* Fixed tests.
* Improved tests
* Fixed code structure
* Fixed lint issues
* nitz.
* nitz
* nitz.
* nitz.
* nitz.
* Added an example
* Fixed PR comments.
* Applied PR comments + nitz
* nitz.
* Improved the code.
* Fixed a perf issue.
* Added batch processing.
* Made the example more interesting
* Fixed bucket count
* Renamed Direct to CustomQuery
* Fixed lint issues.
* No need for scorer to be an `Option`
* nitz
* Used BitSet
* Added an optimization for AllQuery
* Fixed merge issues.
* Fixed lint issues.
* Added benchmark for FILTER
* Removed the Option wrapper.
* nitz.
* Applied PR comments.
* Fixed the AllQuery optimization
* Applied PR comments.
* feat: used `erased_serde` to allow filter query to be serialized
* further improved a comment
* Added back tests.
* removed an unused method
* removed an unused method
* Added documentation
* nitz.
* Added query builder.
* Fixed a comment.
* Applied PR comments.
* Fixed doctest issues.
* Added ser/de
* Removed bench in test
* Fixed a lint issue.
* add nested histogram-termagg benchmark
* Replace AggregationsWithAccessor with AggData
With AggregationsWithAccessor pre-computation and caching was done on the collector level.
If you have 10000 sub collectors (e.g. a term aggregation with sub aggregations) this is very inefficient.
`AggData` instead moves the data from the collector to a node which reflects the cardinality of the request tree instead of the cardinality of the segment collector.
It also moves the global struct shared with all aggregations in to aggregation specific structs. So each aggregation has its own space to store cached data and aggregation specific information.
This also breaks up the dependency to the elastic search aggregation structure somewhat.
Due to lifetime issues, we move the agg request specific object out of `AggData` during the collection and move it back at the end (for now). That's some unnecessary work, which costs CPU.
This allows better caching and will also pave the way for another potential optimization, by separating the collector and its storage. Currently we allocate a new collector for each sub aggregation bucket (for nested aggregations), but ideally we would have just one collector instance.
* renames
* move request data to agg request files
---------
Co-authored-by: Pascal Seitz <pascal.seitz@datadoghq.com>
* first version of extended stats along with its tests
* using IntermediateExtendStats instead of IntermediateStats with all tests passing
* Created struct for request and response
* first test with extended_stats
* kahan summation and tests with approximate equality
* version ready for merge
* removed approx dependency
* refactor for using ExtendedStats only when needed
* interim version
* refined version with code formatted
* refactored a struct
* cosmetic refactor
* fix after merge
* fix format
* added extended_stat bench
* merge and new benchmark for extended stats
* split stat segment collectors
* wrapped intermediate extended stat with a box to limit memory usage
* Revert "wrapped intermediate extended stat with a box to limit memory usage"
This reverts commit 5b4aa9f393.
* some code reformat, commented kahan summation
* refactor after review
* refactor after code review
* fix after incorrectly restoring kahan summation
* modifications for code review + bug fix in merge_fruit
* refactor assert_nearly_equals macro
* update after code review
---------
Co-authored-by: Giovanni Cuccu <gcuccu@imolainformatica.it>
* compact doc
* add any value type
* pass references when building CompactDoc
* remove OwnedValue from API
* clippy
* clippy
* fail on large documents
* fmt
* cleanup
* cleanup
* implement Value for different types
fix serde_json date Value implementation
* fmt
* cleanup
* fmt
* cleanup
* store positions instead of pos+len
* remove nodes array
* remove mediumvec
* cleanup
* infallible serialize into vec
* remove positions indirection
* remove 24MB limitation in document
use u32 for Addr
Remove the 3 byte addressing limitation and use VInt instead
* cleanup
* extend test
* cleanup, add comments
* rename, remove pub
* fix windows build (#1)
* Fix windows build
* Add doc traits
* Add field value iter
* Add value and serialization
* Adjust order
* Fix bug
* Correct type
* Fix generic bugs
* Reformat code
* Add generic to index writer which I forgot about
* Fix missing generics on single segment writer
* Add missing type export
* Add default methods for convenience
* Cleanup
* Fix more-like-this query to use standard types
* Update API and fix tests
* Add doc traits
* Add field value iter
* Add value and serialization
* Adjust order
* Fix bug
* Correct type
* Rebase main and fix conflicts
* Reformat code
* Merge upstream
* Fix missing generics on single segment writer
* Add missing type export
* Add default methods for convenience
* Cleanup
* Fix more-like-this query to use standard types
* Update API and fix tests
* Add tokenizer improvements from previous commits
* Add tokenizer improvements from previous commits
* Reformat
* Fix unit tests
* Fix unit tests
* Use enum in changes
* Stage changes
* Add new deserializer logic
* Add serializer integration
* Add document deserializer
* Implement new (de)serialization api for existing types
* Fix bugs and type errors
* Add helper implementations
* Fix errors
* Reformat code
* Add unit tests and some code organisation for serialization
* Add unit tests to deserializer
* Add some small docs
* Add support for deserializing serde values
* Reformat
* Fix typo
* Fix typo
* Change repr of facet
* Remove unused trait methods
* Add child value type
* Resolve comments
* Fix build
* Fix more build errors
* Fix more build errors
* Fix the tests I missed
* Fix examples
* fix numerical order, serialize PreTok Str
* fix coverage
* rename Document to TantivyDocument, rename DocumentAccess to Document
add Binary prefix to binary de/serialization
* fix coverage
---------
Co-authored-by: Pascal Seitz <pascal.seitz@gmail.com>
* tokenizer-api: reduce Tokenizer overhead
Previously a new `Token` for each text encountered was created, which
contains `String::with_capacity(200)`
In the new API the token_stream gets mutable access to the tokenizer,
this allows state to be shared (in this PR Token is shared).
Ideally the allocation for the BoxTokenStream would also be removed, but
this may require some lifetime tricks.
* simplify api
* move lowercase and ascii folding buffer to global
* empty Token text as default
