rust/tantivy
1
0
Fork 0
mirror of https://github.com/quickwit-oss/tantivy.git synced 2025-12-28 04:52:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

Compare commits

base: rust:ip_fastfield
Branches Tags
rust:main rust:stuhood.upstream-erased-sort-computer rust:trinity.pointard/ff-query rust:bucket_id_agg rust:paradedb/fix-agg-validation rust:flatheadmill-geo rust:paradedb-paradedb/fix-overflow-issue rust:low_card_optimization rust:paul.masurel/clippy rust:fix_size_hint rust:qw-airmail-20250916 rust:unit-test-aggregation-sum-other rust:clippy-and-cleanups rust:trinity.pointard/opti-sstable-sorted-ords-to-term rust:release_0.24 rust:dependabot/cargo/rand-0.9.1 rust:dependabot/cargo/lru-0.14.0 rust:dependabot/cargo/fs4-0.13.1 rust:tagged-user-input-ast rust:1686a/parse-agg-res rust:dependabot/cargo/rand_distr-0.5.1 rust:issue/2411-adding-panic-handler-to-rayon-merge-thread-pool rust:dependabot/cargo/nom-8 rust:upgrade_deps rust:test_parser rust:dependabot/github_actions/codecov/codecov-action-5 rust:binggan-0.12.0 rust:agg_format rust:path_collector rust:test_order_bug rust:test_columnar_multi rust:columnar_optionalmask_in_multivalues rust:prefix-phrase-query-optim rust:tantivy-with-oneshot rust:update_example rust:store-reader-enumerate rust:document-deserialize-seed rust:check_doc_mem rust:experimental-do-not-merge rust:expdonotmerge rust:cherry-pick-vec-instead-of-btreemap rust:validate_length rust:raphael_bench_avx_filter rust:raphael_optim_memset2 rust:query-field-name-avoid-alloc rust:check_bug rust:missing_test rust:list_fields rust:block-cache-control rust:trinity--multilayer-sstable rust:tracing-tantivy2 rust:tracing-tantivy rust:index_writer_err rust:default_fast_field_tokenizer rust:fmassot/add-preserve-original-ascii-folding rust:numeric_with_str_col rust:straightforward-simplify rust:fmassot/fake-test-branch rust:fmassot/dynamic-tokenizer-poc-follow-up rust:fmassot/add-box-token-filter-and-refactor rust:dynamic-tokenizer-proof-of-concept rust:fmassot/add-box-tokenizer-and-token-filter rust:tokenizer-api-0.1.1 rust:addconversion rust:removed-extra-enum-alternative-for-datetime-precision rust:issue/1981-reproduce-bm25-bug rust:tokenizer_docs rust:merge-proptests rust:columnar-todo rust:quickwit-0.5 rust:remove-byteorder rust:remove_dyn rust:pr/alexcole/1855-inject-bm25-statistics rust:evance-expose-build-query-from-user-input-ast rust:missing-specialized-method-arcdyncolumn rust:fast-u64-range-query rust:0.19.2 rust:columnar-cli2 rust:trinity--pub-term-as-slice rust:typed-column3 rust:columnar-merge3 rust:columnar-merge-exploring-fn rust:use_columnar rust:typed-column2 rust:term_agg_bucket_limit rust:range rust:typed-column rust:columnar-main rust:forced-types rust:u128-columnar rust:use_stats rust:truncation_comment rust:0.19.1 rust:trinity/yoke-documents rust:column-handle rust:columnar-cli rust:columnar rust:revert-1711-sparse_dense_index rust:commit-changes rust:sparse_dense_index_headroom rust:optional_column rust:no-column-for-multivalued rust:fmassot--bench-hdfs-with-array rust:bugfix-position-broken rust:float rust:simplify rust:debug_time rust:sparse_codecs rust:fix_estimate rust:column-reader rust:column-reader-b rust:columnreader-simple rust:col-trait-refact rust:ip_fastfield rust:column-trait-2 rust:refact-dyn-column rust:column-trait rust:refact-codec-trait rust:pr/PSeitz/1485-1 rust:fastfieldcodec rust:termmap_per_field rust:fastfield-codecs-tests rust:rayon-indexing-writers rust:issue/1251 rust:index-bench rust:fix_open_bytes_index rust:warming rust:revert-1222-issue/1195 rust:add-random-bench-for-fast-field rust:wasm-friendly-exploratory rust:revert-1159-patch-1 rust:debug-position-panic-0.15.3 rust:fix-issue-1111 rust:remove_rand rust:unused-deps rust:hfb/phrase_match_slop rust:0.15.3 rust:expose-min-max-on-multi-fast-field-reader rust:revert-1094-expose-min-max-on-multi-fast-field-reader rust:hotfix-1088 rust:owned-bytes-instead-of-rawdoc rust:rihardsk-date-ranges rust:issue/997 rust:0.13.3 rust:fail-fast-store-checkpoint rust:debugging-store rust:issue/reproduce-897 rust:issue/969-reproduce rust:troublescooter-static rust:intermediate-mergeable-troublescooter rust:barrotsteindev-filter-collector-tpredicatevalue rust:issue/938b rust:issue/938 rust:barrotsteindev-filter-collector rust:issue/922 rust:issue/896 rust:quickfix-explain rust:slog rust:limit-random-seeks rust:0.13.1 rust:issue/866b rust:bugfix-unittest-macos rust:githubactions rust:fieldnorm_readers rust:blockwand rust:nodeffeatfails rust:removedaliveiterator rust:robyoung-introduce-offset-to-top-docs rust:specialization-of-foreach-for-union-scorer rust:elshize-bmw rust:cutt rust:nrt rust:storewriter-out rust:segment_writer_in_ram2 rust:polling rust:audunhalland-query-boost rust:poll-meta rust:segment_writer_in_ram rust:bundle rust:petr-tik-n662_footer_is_incompatible_impl rust:race-condition-in-tests rust:hotfix-0.10.3-issue681 rust:remove-tru rust:kkoziara-pre-tokenized-text rust:issue/681 rust:issue/680 rust:refactoring/field rust:updating/uuid rust:fix-bench rust:hotfix-656 rust:readd-macos-ci rust:incremental-search rust:crate-division rust:streamer-with-state rust:criterion rust:python-binding-changes rust:bump-version rust:gh-pages rust:issue/554 rust:perf/experimental-union-for_each rust:issue/526b rust:softcommits rust:0.9.1-hotfix rust:revert-521-fix-non-english-stemmers rust:issue/500-hotfix rust:hotfix/0.8.2 rust:issue/483 rust:bug/merge-deted rust:issue/407 rust:issue/457 rust:segment_fruits rust:doc rust:dds/lenient rust:issue/webasm rust:issue/396 rust:broken_collector_with_result rust:dss/automaton-weight-suggestion rust:wasm rust:common-crawl rust:tantivy-imhotep
rust:0.25.0 rust:0.22.1 rust:0.24.2 rust:0.24.1 rust:tantivy-query-grammar-v0.24.0 rust:tantivy-columnar-v0.5.0 rust:tantivy-bitpacker-v0.8.0 rust:ownedbytes-v0.9.0 rust:tantivy-stacker-v0.5.0 rust:tantivy-sstable-v0.5.0 rust:tantivy-common-v0.9.0 rust:tantivy-tokenizer-api-v0.5.0 rust:0.24 rust:qw-airmail-20250224-hotfix-merge-panic rust:qw-airmail-20250219-hotfix-merge-panic rust:qw-airmail-202406 rust:qw-airmail-20240611 rust:qw-airmail-20240606 rust:quickwit-rev rust:0.22.0 rust:tantivy-stacker-v0.3.0 rust:ownedbytes-v0.7.0 rust:tantivy-tokenizer-api-v0.3.0 rust:tantivy-columnar-v0.3.0 rust:tantivy-common-v0.7.0 rust:tantivy-query-grammar-v0.22.0 rust:tantivy-sstable-v0.3.0 rust:tantivy-bitpacker-v0.6.0 rust:0.21.1 rust:0.21 rust:0.20.2 rust:0.20.1 rust:0.20 rust:quickwit-0.5-rev rust:0.19.2 rust:0.19.1 rust:0.19 rust:0.18.1 rust:0.18 rust:0.17 rust:0.16.1 rust:0.15.3 rust:0.15.2 rust:0.15.1 rust:0.15 rust:0.14 rust:0.13.3 rust:0.13.2 rust:0.13.1 rust:0.13 rust:0.12 rust:0.11.3 rust:0.11.1 rust:0.11 rust:0.10.3 rust:0.1 rust:0.10.2 rust:0.10.1 rust:0.10.0 rust:0.9.1 rust:0.9 rust:0.8.2 rust:0.8.2b rust:0.8.1 rust:0.8.0 rust:0.7.2 rust:0.7.1 rust:0.7.0 rust:0.6.1 rust:0.6.0 rust:0.5.2 rust:0.5.1 rust:0.5.0 rust:0.4.3 rust:0.4.2 rust:0.4.0 rust:0.3.1 rust:0.3.0 rust:0.2.0
...
compare: rust:columnar
Branches Tags
rust:stuhood.upstream-erased-sort-computer rust:main rust:trinity.pointard/ff-query rust:bucket_id_agg rust:paradedb/fix-agg-validation rust:flatheadmill-geo rust:paradedb-paradedb/fix-overflow-issue rust:low_card_optimization rust:paul.masurel/clippy rust:fix_size_hint rust:qw-airmail-20250916 rust:unit-test-aggregation-sum-other rust:clippy-and-cleanups rust:trinity.pointard/opti-sstable-sorted-ords-to-term rust:release_0.24 rust:dependabot/cargo/rand-0.9.1 rust:dependabot/cargo/lru-0.14.0 rust:dependabot/cargo/fs4-0.13.1 rust:tagged-user-input-ast rust:1686a/parse-agg-res rust:dependabot/cargo/rand_distr-0.5.1 rust:issue/2411-adding-panic-handler-to-rayon-merge-thread-pool rust:dependabot/cargo/nom-8 rust:upgrade_deps rust:test_parser rust:dependabot/github_actions/codecov/codecov-action-5 rust:binggan-0.12.0 rust:agg_format rust:path_collector rust:test_order_bug rust:test_columnar_multi rust:columnar_optionalmask_in_multivalues rust:prefix-phrase-query-optim rust:tantivy-with-oneshot rust:update_example rust:store-reader-enumerate rust:document-deserialize-seed rust:check_doc_mem rust:experimental-do-not-merge rust:expdonotmerge rust:cherry-pick-vec-instead-of-btreemap rust:validate_length rust:raphael_bench_avx_filter rust:raphael_optim_memset2 rust:query-field-name-avoid-alloc rust:check_bug rust:missing_test rust:list_fields rust:block-cache-control rust:trinity--multilayer-sstable rust:tracing-tantivy2 rust:tracing-tantivy rust:index_writer_err rust:default_fast_field_tokenizer rust:fmassot/add-preserve-original-ascii-folding rust:numeric_with_str_col rust:straightforward-simplify rust:fmassot/fake-test-branch rust:fmassot/dynamic-tokenizer-poc-follow-up rust:fmassot/add-box-token-filter-and-refactor rust:dynamic-tokenizer-proof-of-concept rust:fmassot/add-box-tokenizer-and-token-filter rust:tokenizer-api-0.1.1 rust:addconversion rust:removed-extra-enum-alternative-for-datetime-precision rust:issue/1981-reproduce-bm25-bug rust:tokenizer_docs rust:merge-proptests rust:columnar-todo rust:quickwit-0.5 rust:remove-byteorder rust:remove_dyn rust:pr/alexcole/1855-inject-bm25-statistics rust:evance-expose-build-query-from-user-input-ast rust:missing-specialized-method-arcdyncolumn rust:fast-u64-range-query rust:0.19.2 rust:columnar-cli2 rust:trinity--pub-term-as-slice rust:typed-column3 rust:columnar-merge3 rust:columnar-merge-exploring-fn rust:use_columnar rust:typed-column2 rust:term_agg_bucket_limit rust:range rust:typed-column rust:columnar-main rust:forced-types rust:u128-columnar rust:use_stats rust:truncation_comment rust:0.19.1 rust:trinity/yoke-documents rust:column-handle rust:columnar-cli rust:columnar rust:revert-1711-sparse_dense_index rust:commit-changes rust:sparse_dense_index_headroom rust:optional_column rust:no-column-for-multivalued rust:fmassot--bench-hdfs-with-array rust:bugfix-position-broken rust:float rust:simplify rust:debug_time rust:sparse_codecs rust:fix_estimate rust:column-reader rust:column-reader-b rust:columnreader-simple rust:col-trait-refact rust:ip_fastfield rust:column-trait-2 rust:refact-dyn-column rust:column-trait rust:refact-codec-trait rust:pr/PSeitz/1485-1 rust:fastfieldcodec rust:termmap_per_field rust:fastfield-codecs-tests rust:rayon-indexing-writers rust:issue/1251 rust:index-bench rust:fix_open_bytes_index rust:warming rust:revert-1222-issue/1195 rust:add-random-bench-for-fast-field rust:wasm-friendly-exploratory rust:revert-1159-patch-1 rust:debug-position-panic-0.15.3 rust:fix-issue-1111 rust:remove_rand rust:unused-deps rust:hfb/phrase_match_slop rust:0.15.3 rust:expose-min-max-on-multi-fast-field-reader rust:revert-1094-expose-min-max-on-multi-fast-field-reader rust:hotfix-1088 rust:owned-bytes-instead-of-rawdoc rust:rihardsk-date-ranges rust:issue/997 rust:0.13.3 rust:fail-fast-store-checkpoint rust:debugging-store rust:issue/reproduce-897 rust:issue/969-reproduce rust:troublescooter-static rust:intermediate-mergeable-troublescooter rust:barrotsteindev-filter-collector-tpredicatevalue rust:issue/938b rust:issue/938 rust:barrotsteindev-filter-collector rust:issue/922 rust:issue/896 rust:quickfix-explain rust:slog rust:limit-random-seeks rust:0.13.1 rust:issue/866b rust:bugfix-unittest-macos rust:githubactions rust:fieldnorm_readers rust:blockwand rust:nodeffeatfails rust:removedaliveiterator rust:robyoung-introduce-offset-to-top-docs rust:specialization-of-foreach-for-union-scorer rust:elshize-bmw rust:cutt rust:nrt rust:storewriter-out rust:segment_writer_in_ram2 rust:polling rust:audunhalland-query-boost rust:poll-meta rust:segment_writer_in_ram rust:bundle rust:petr-tik-n662_footer_is_incompatible_impl rust:race-condition-in-tests rust:hotfix-0.10.3-issue681 rust:remove-tru rust:kkoziara-pre-tokenized-text rust:issue/681 rust:issue/680 rust:refactoring/field rust:updating/uuid rust:fix-bench rust:hotfix-656 rust:readd-macos-ci rust:incremental-search rust:crate-division rust:streamer-with-state rust:criterion rust:python-binding-changes rust:bump-version rust:gh-pages rust:issue/554 rust:perf/experimental-union-for_each rust:issue/526b rust:softcommits rust:0.9.1-hotfix rust:revert-521-fix-non-english-stemmers rust:issue/500-hotfix rust:hotfix/0.8.2 rust:issue/483 rust:bug/merge-deted rust:issue/407 rust:issue/457 rust:segment_fruits rust:doc rust:dds/lenient rust:issue/webasm rust:issue/396 rust:broken_collector_with_result rust:dss/automaton-weight-suggestion rust:wasm rust:common-crawl rust:tantivy-imhotep
rust:0.25.0 rust:0.22.1 rust:0.24.2 rust:0.24.1 rust:tantivy-query-grammar-v0.24.0 rust:tantivy-columnar-v0.5.0 rust:tantivy-bitpacker-v0.8.0 rust:ownedbytes-v0.9.0 rust:tantivy-stacker-v0.5.0 rust:tantivy-sstable-v0.5.0 rust:tantivy-common-v0.9.0 rust:tantivy-tokenizer-api-v0.5.0 rust:0.24 rust:qw-airmail-20250224-hotfix-merge-panic rust:qw-airmail-20250219-hotfix-merge-panic rust:qw-airmail-202406 rust:qw-airmail-20240611 rust:qw-airmail-20240606 rust:quickwit-rev rust:0.22.0 rust:tantivy-stacker-v0.3.0 rust:ownedbytes-v0.7.0 rust:tantivy-tokenizer-api-v0.3.0 rust:tantivy-columnar-v0.3.0 rust:tantivy-common-v0.7.0 rust:tantivy-query-grammar-v0.22.0 rust:tantivy-sstable-v0.3.0 rust:tantivy-bitpacker-v0.6.0 rust:0.21.1 rust:0.21 rust:0.20.2 rust:0.20.1 rust:0.20 rust:quickwit-0.5-rev rust:0.19.2 rust:0.19.1 rust:0.19 rust:0.18.1 rust:0.18 rust:0.17 rust:0.16.1 rust:0.15.3 rust:0.15.2 rust:0.15.1 rust:0.15 rust:0.14 rust:0.13.3 rust:0.13.2 rust:0.13.1 rust:0.13 rust:0.12 rust:0.11.3 rust:0.11.1 rust:0.11 rust:0.10.3 rust:0.1 rust:0.10.2 rust:0.10.1 rust:0.10.0 rust:0.9.1 rust:0.9 rust:0.8.2 rust:0.8.2b rust:0.8.1 rust:0.8.0 rust:0.7.2 rust:0.7.1 rust:0.7.0 rust:0.6.1 rust:0.6.0 rust:0.5.2 rust:0.5.1 rust:0.5.0 rust:0.4.3 rust:0.4.2 rust:0.4.0 rust:0.3.1 rust:0.3.0 rust:0.2.0

363 Commits
ip_fastfie ... columnar

Author SHA1 Message Date
Paul Masurel
8828b6d310 Support for columnar 2022-12-21 12:21:30 +09:00
Paul Masurel
2b89bf9050 Support for NotNaN in fast fields 2022-12-21 12:20:48 +09:00
Paul Masurel
3580198447 Minor refactoring 2022-12-21 12:18:33 +09:00
Paul Masurel
d96a716d20 Refactoring to prepare for the addition of dynamic fast field 2022-12-21 12:16:00 +09:00
PSeitz
2ac1cc2fc0 add sparse codec (#1723) 
* add sparse codec

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

* add the -1 u16 fix for metadata num_vals

* add dense block encoding to sparse codec

* add comment, refactor u16 reading

Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-12-20 15:30:33 +01:00
PSeitz
f9171a3981 fix clippy (#1725) 
* fix clippy

* fix clippy fastfield codecs

* fix clippy bitpacker

* fix clippy common

* fix clippy stacker

* fix clippy sstable

* fmt
 2022-12-20 07:30:06 +01:00
PSeitz
a2cf6a79b4 Sparse dense index (#1716) 
* add dense codec

* benchmark fix and important optimisation

* move code to DenseIndexBlock

improve benchmark

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

* extend benchmarks

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-12-13 07:50:09 +01:00
Paul Masurel
f6e87a5319 Cargo fmt 2022-12-13 12:30:40 +09:00
Paul Masurel
f9971e15fe Fixing unit test with sstable test. 2022-12-13 12:22:44 +09:00
PSeitz
3cdc8e7472 pass index info to serialize (#1719) 2022-12-13 04:20:31 +01:00
dependabot[bot]
fbb0f8b55d Update base64 requirement from 0.13.0 to 0.20.0 (#1720) 
Updates the requirements on [base64](https://github.com/marshallpierce/rust-base64) to permit the latest version.
- [Release notes](https://github.com/marshallpierce/rust-base64/releases)
- [Changelog](https://github.com/marshallpierce/rust-base64/blob/master/RELEASE-NOTES.md)
- [Commits](https://github.com/marshallpierce/rust-base64/compare/v0.13.0...v0.20.0)

---
updated-dependencies:
- dependency-name: base64
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
 2022-12-13 11:46:23 +09:00
Paul Masurel
136a8f4124 Isolating sstable and stacker in independant crates. (#1718) 
Both crate will be used in the new (optional + dynamic) fastfield work.
 2022-12-13 11:44:17 +09:00
PSeitz
5d4535de83 Changelog fix (#1717) 2022-12-12 14:28:42 +09:00
PSeitz
2c50b02eb3 Fix max bucket limit in histogram (#1703) 
* Fix max bucket limit in histogram

The max bucket limit in histogram was broken, since some code introduced temporary filtering of buckets, which then resulted into an incorrect increment on the bucket count.
The provided solution covers more scenarios, but there are still some scenarios unhandled (See #1702).

* Apply suggestions from code review

Co-authored-by: Paul Masurel <paul@quickwit.io>

Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-12-12 04:40:15 +01:00
PSeitz
509adab79d Bump version (#1715) 
* group workspace deps

* update cargo.toml

* revert tant version

* chore: Release
 2022-12-12 04:39:43 +01:00
PSeitz
96c93a6ba3 Merge pull request #1700 from quickwit-oss/PSeitz-patch-1 
Update CHANGELOG.md
 2022-12-02 16:31:11 +01:00
boraarslan
495824361a Move split_full_path to Schema (#1692) 2022-11-29 20:56:13 +09:00
PSeitz
485a8f507e Update CHANGELOG.md 2022-11-28 15:41:31 +01:00
PSeitz
1119e59eae prepare fastfield format for null index (#1691) 
* prepare fastfield format for null index
* add format version for fastfield
* Update fastfield_codecs/src/compact_space/mod.rs
* switch to variable size footer
* serialize delta of end
 2022-11-28 17:15:24 +09:00
PSeitz
ee1f2c1f28 add aggregation support for date type (#1693) 
* add aggregation support for date type
fixes #1332

* serialize key_as_string as rfc3339 in date histogram
* update docs
* enable date for range aggregation
 2022-11-28 09:12:08 +09:00
PSeitz
600548fd26 Merge pull request #1694 from quickwit-oss/dependabot/cargo/zstd-0.12 
Update zstd requirement from 0.11 to 0.12
 2022-11-25 05:48:59 +01:00
PSeitz
9929c0c221 Merge pull request #1696 from quickwit-oss/dependabot/cargo/env_logger-0.10.0 
Update env_logger requirement from 0.9.0 to 0.10.0
 2022-11-25 03:28:10 +01:00
dependabot[bot]
f53e65648b Update env_logger requirement from 0.9.0 to 0.10.0 
Updates the requirements on [env_logger](https://github.com/rust-cli/env_logger) to permit the latest version.
- [Release notes](https://github.com/rust-cli/env_logger/releases)
- [Changelog](https://github.com/rust-cli/env_logger/blob/main/CHANGELOG.md)
- [Commits](https://github.com/rust-cli/env_logger/compare/v0.9.0...v0.10.0)

---
updated-dependencies:
- dependency-name: env_logger
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
 2022-11-24 20:07:52 +00:00
PSeitz
0281b22b77 update create_in_ram docs (#1695) 2022-11-24 17:30:09 +01:00
dependabot[bot]
a05c184830 Update zstd requirement from 0.11 to 0.12 
Updates the requirements on [zstd](https://github.com/gyscos/zstd-rs) to permit the latest version.
- [Release notes](https://github.com/gyscos/zstd-rs/releases)
- [Commits](https://github.com/gyscos/zstd-rs/commits)

---
updated-dependencies:
- dependency-name: zstd
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
 2022-11-23 20:15:32 +00:00
Paul Masurel
0b40a7fe43 Added a expand_dots JsonObjectOptions. (#1687) 
Related with quickwit#2345.
 2022-11-21 23:03:00 +09:00
trinity-1686a
e758080465 add support for TermSetQuery in query parser (#1683) 2022-11-17 16:49:49 +01:00
Paul Masurel
2a39289a1b Handle escaped dot in json path in the QueryParser. (#1682) 2022-11-16 07:18:34 +09:00
Adam Reichold
ca6231170e Make the built-in stop word lists selectable via the Language enum already used by the Stemmer filter. (#1671) 2022-11-15 17:40:25 +09:00
PSeitz
eda6e5a10a Merge pull request #1681 from quickwit-oss/ip_range_query_multi 
remove Column from MultiValuedU128FastFieldReader
 2022-11-15 09:27:46 +08:00
Pascal Seitz
8641155cbb remove column from MultiValuedU128FastFieldReader 2022-11-14 18:49:15 +08:00
PSeitz
9a090ed994 Merge pull request #1659 from quickwit-oss/ip_range_query_multi 
add support for ip range query on multivalue fastfields
 2022-11-14 15:17:41 +08:00
Pascal Seitz
b7d0dd154a fmt 2022-11-14 14:49:15 +08:00
PSeitz
ce10fab20f Apply suggestions from code review 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-11-14 14:21:53 +08:00
Pascal Seitz
e034328a8b Improve position_to_docid, refactor, add tests 2022-11-14 14:21:53 +08:00
Pascal Seitz
f811d1616b add support for ip range query on multivalue fastfields 2022-11-14 14:21:52 +08:00
PSeitz
c665b16ff0 Merge pull request #1672 from quickwit-oss/allow_range_without_indexed 
Allow range query on fastfield without INDEXED
 2022-11-14 12:45:12 +08:00
PSeitz
3b5f810051 Merge pull request #1677 from quickwit-oss/switch_to_u32 
switch total_num_val to u32
 2022-11-14 12:01:40 +08:00
trinity-1686a
5765c261aa allow warming up of the full posting list (#1673) 
* allow warming up of the full posting list

* cargo fmt
 2022-11-14 10:27:56 +09:00
Pascal Seitz
fb9f03118d switch total_num_val to u32 2022-11-11 17:35:52 +08:00
PSeitz
55a9d808d4 Merge pull request #1674 from quickwit-oss/u128_codec_header 
add header with codec type for u128
 2022-11-11 13:47:51 +08:00
Pascal Seitz
32166682b3 add header deser test 2022-11-11 13:28:12 +08:00
Pascal Seitz
e6acf8f76d add header with codec type for u128 2022-11-11 11:52:17 +08:00
Pascal Seitz
9e8a0c2cca Allow range query on fastfield without INDEXED 2022-11-10 15:56:08 +08:00
Paul Masurel
3edf0a2724 Using the manual reload policy in IndexWriter. (#1667) 2022-11-09 11:20:41 +01:00
Paul Masurel
8ca12a5683 Added stop word filter to CHANGELOG.md 2022-11-09 17:00:45 +09:00
Adam Reichold
a4b759d2fe Include stop word lists from Lucene and the Snowball project (#1666) 2022-11-09 16:57:35 +09:00
PSeitz
3e9c806890 Merge pull request #1665 from quickwit-oss/fix_num_vals 
fix num_vals on u128 value index after merge
 2022-11-07 21:46:02 +08:00
Pascal Seitz
c69a873dd3 fix num_vals on value index after merge 2022-11-07 21:05:21 +08:00
PSeitz
666afcf641 Merge pull request #1663 from PSeitz/fix_clippy 
fix clippy
 2022-11-07 18:11:20 +08:00
Pascal Seitz
38ad46e580 fix clippy 2022-11-07 16:09:55 +08:00
PSeitz
e948889f4c Merge pull request #1662 from quickwit-oss/fix_num_vals 
fix num_vals in multivalue index after merge
 2022-11-07 15:57:32 +08:00
Pascal Seitz
6e636c9cea fix num_vals in multivalue index after merge 2022-11-07 15:00:52 +08:00
PSeitz
5a610efbc1 Merge pull request #1661 from quickwit-oss/upgrade_criterion 
update criterion to 0.4
 2022-11-04 14:45:34 +08:00
Pascal Seitz
500a0d5e48 update criterion to 0.4 2022-11-04 13:26:29 +08:00
PSeitz
509a265659 add docstore version (#1652) 
* add docstore version

closes #1589

* assert for docstore version
 2022-11-04 10:19:16 +09:00
PSeitz
5b2cea1b97 Merge pull request #1656 from quickwit-oss/multival_offset_index 
move multivalue index to own file
 2022-11-02 14:03:06 +08:00
PSeitz
a5a80ffaea Update fastfield_codecs/src/column.rs 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-11-02 06:37:27 +01:00
PSeitz
0f98d91a39 Merge pull request #1646 from quickwit-oss/no_score_calls 
No score calls if score is not requested
 2022-11-01 20:09:32 +08:00
PSeitz
2af6b01c17 Update src/query/boolean_query/boolean_weight.rs 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-11-01 16:13:00 +08:00
Adam Reichold
c32ab66bbd Small improvements to StopWorldFilter (#1657) 
* Do not copy the whole set of stop words for each stream

* Make construction of StopWordFilter more flexible.
 2022-11-01 16:47:34 +09:00
PSeitz
3f3a6f9990 Merge pull request #1653 from quickwit-oss/faster_hash 
switch to fx hashmap
 2022-11-01 14:53:18 +08:00
Pascal Seitz
83325d8f3f move multivalue index to own file 
start_doc parameter in positions to docids
 2022-11-01 10:36:13 +08:00
PSeitz
4e46f4f8c4 Merge pull request #1649 from adamreichold/split-compound-words 
RFC: Add dictionary-based SplitCompoundWords token filter.
 2022-10-27 17:12:48 +08:00
Pascal Seitz
43df356010 rename to docset 2022-10-27 16:53:38 +08:00
PSeitz
6647362464 Merge pull request #1648 from adamreichold/stemmer-todo-alloc 
Avoid unconditional allocation in StemmerTokenStream.
 2022-10-27 16:50:41 +08:00
Pascal Seitz
279b1b28d3 switch to fx hashmap 2022-10-27 16:19:59 +08:00
PSeitz
7a80851e36 Merge pull request #1645 from quickwit-oss/ip_field_range_query 
add ip range query benchmark, add seek behaviour
 2022-10-27 16:13:52 +08:00
Adam Reichold
cd952429d2 Add dictionary-based SplitCompoundWords token filter. 2022-10-27 08:30:33 +02:00
PSeitz
d777c964da Merge pull request #1650 from adamreichold/fnv-rustc-hash 
Replace FNV by rustc-hash
 2022-10-27 12:11:26 +08:00
Adam Reichold
bbb058d976 Replace FNV by rustc-hash 
Both construction have similar goals but rustc-hash ist better suited for
contemporary CPU as it works one word at a time instead of byte per byte.
 2022-10-27 00:35:09 +02:00
Adam Reichold
5f7d027a52 Avoid unconditional allocation in StemmerTokenStream. 
This fixes the TODO in two ways: If the stemmer already yields an owned string,
it is used directly as the new text of the token. Otherwise, a temporary buffer
is used to copy the stemmed text (just as before) and then swapping it into the
token to reuse its existing buffer.
 2022-10-26 18:11:15 +02:00
Pascal Seitz
dfab201191 for_each_docset to iterate without score 2022-10-26 17:25:05 +08:00
PSeitz
0c2bd36fe3 Panic on duplicate field names (#1647) 
fixes #1601
 2022-10-26 16:17:33 +09:00
Pascal Seitz
af839753e0 No score calls if score is not requested 2022-10-26 12:18:35 +08:00
Pascal Seitz
fec2b63571 improve bench by adding more blanks in compact space 2022-10-25 22:09:01 +08:00
Pascal Seitz
6213ea476a pass positions parameter 2022-10-25 17:44:51 +08:00
Pascal Seitz
5e159c26bf add ip range query benchmark, add seek behaviour 2022-10-25 15:57:19 +08:00
PSeitz
a5e59ab598 Merge pull request #1644 from quickwit-oss/get_val_u32 
switch get_val() to u32
 2022-10-24 19:30:03 +08:00
Pascal Seitz
e772d3170d switch get_val() to u32 
Fixes #1638
 2022-10-24 19:05:57 +08:00
PSeitz
8c2ba7bd55 Merge pull request #1637 from quickwit-oss/ip_field_range_query 
add range query via ip fast field
 2022-10-24 18:10:47 +08:00
Pascal Seitz
02328b0151 fix proptest 2022-10-24 17:46:06 +08:00
Pascal Seitz
7cc775256c add comments, rename 2022-10-24 17:08:37 +08:00
Pascal Seitz
07b40f8b8b add proptest 2022-10-24 16:52:55 +08:00
PSeitz
9b6b6be5b9 Apply suggestions from code review 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-10-24 16:00:38 +08:00
Pascal Seitz
6bb73a527f add range query via ip fast field 2022-10-24 16:00:38 +08:00
PSeitz
03885d0f3c Merge pull request #1643 from quickwit-oss/range_query_parser 
allow more characters in range query
 2022-10-24 15:09:47 +08:00
Pascal Seitz
f2e5135870 allow more characters in range query 
closes #1642
 2022-10-21 18:05:15 +08:00
Paul Masurel
c24157f28b Bumping version format. (#1640) 
The docstore format has changed in a non-compatible manner.
 2022-10-21 15:35:35 +09:00
PSeitz
873382cdcb Merge pull request #1639 from quickwit-oss/num_vals_u32 
switch num_vals() to u32
 2022-10-21 12:36:50 +08:00
Pascal Seitz
791350091c switch num_vals() to u32 
fixes #1630
 2022-10-20 19:44:28 +08:00
Paul Masurel
483b1d13d4 Added unit test for long tokens (#1635) 
* Bugfix on long tokens and multivalue text fields.

Fixes a minor bug for the strong edge case
in which a tokenizer would emit tokens where
the last token does not cover the last position.

More importantly, this adds unit tests.

Closes #1634

* Update src/indexer/segment_writer.rs

Co-authored-by: PSeitz <PSeitz@users.noreply.github.com>

Co-authored-by: PSeitz <PSeitz@users.noreply.github.com>
 2022-10-20 15:05:37 +09:00
PSeitz
8de7fa9d95 Merge pull request #1631 from quickwit-oss/high_positions 
add test for phrase search on multi text field
 2022-10-20 10:26:00 +08:00
Paul Masurel
94313b62f8 Hotfix issue/1629 - position broken (#1633) 
* Bugfix position broken.

For Field with several FieldValues, with a
value that contained no token at all, the token position
was reinitialized to 0.

As a result, PhraseQueries can show some false positives.
In addition, after the computation of the position delta, we can
underflow u32, and end up with gigantic delta.

We haven't been able to actually explain the bug in 1629, but it
is assumed that in some corner case these delta can cause a panic.

Closes #1629
 2022-10-20 11:03:55 +09:00
Pascal Seitz
f2b2628feb add test for phrase search on multi text field 2022-10-19 16:29:56 +08:00
PSeitz
449f595832 Merge pull request #1628 from quickwit-oss/skip_index_deser 
faster skipindex deserialization, larger blocksize on sort
 2022-10-19 11:05:20 +08:00
PSeitz
c9235df059 Merge pull request #1627 from quickwit-oss/ip_field_range_query 
add range query handling for ip via term dictionary
 2022-10-19 10:53:00 +08:00
Pascal Seitz
a4485f7611 faster skipindex deserialization, larger blocksize on sort 2022-10-18 19:32:23 +08:00
Pascal Seitz
1082ff60f9 add range query handling for ip via term dictionary 
since IPs are mapped monotonically we can use the term dictionary for range queries
 2022-10-18 13:08:27 +08:00
PSeitz
491854155c Merge pull request #1625 from quickwit-oss/index_ip_field 
index ip field
 2022-10-18 11:18:17 +08:00
Christoph Herzog
96c3d54ac7 fix: Fix power of two computation on 32bit architectures (#1624) 
The current `compute_previous_power_of_two()` implementation used for
TermHashmap takes and returns `usize` , but actually only works
correclty on 64 bit architectures (aka usize == u64)

On other architectures the leading_zeros computation is run on the wrong
type (must be u64), and leads to overflows.

Fixed simply computing the leading_zeros based on a u64 value.
 2022-10-18 11:55:02 +09:00
Pascal Seitz
6800fdec9d add indexing for ip field 
Closes #1595
 2022-10-18 10:07:48 +08:00
PSeitz
c9cf9c952a Merge pull request #1614 from quickwit-oss/remove_superfluous_steps 
refactor Term
 2022-10-17 18:25:31 +08:00
Pascal Seitz
024e53a99c remove truncate 2022-10-17 12:14:35 +08:00
Pascal Seitz
8d75e451bd fix truncate, remove mutable access from term 2022-10-17 12:14:35 +08:00
Pascal Seitz
fcfd76ec55 refactor Term 
fixes some issues with Term
Remove duplicate calls to truncate or resize
Replace Magic Number 5 with constant
Enforce minimum size of 5 for metadata
Fix broken truncate docs
use constructor instead new + set calls
normalize constructor stack
replace assert on internal behavior fixes #1585
 2022-10-17 12:14:34 +08:00
PSeitz
6b7b1cc4fa Merge pull request #1623 from quickwit-oss/remove_unused_buffer 
remove unused buffer
 2022-10-14 20:36:00 +08:00
Pascal Seitz
129f7422f5 remove unused buffer 2022-10-14 20:01:10 +08:00
PSeitz
f39cce2c8b Merge pull request #1622 from quickwit-oss/term_aggregation 
add term aggregation clarification
 2022-10-14 18:09:18 +08:00
PSeitz
d2478fac8a Merge pull request #1621 from quickwit-oss/changelog 
update CHANGELOG
 2022-10-14 18:08:57 +08:00
Pascal Seitz
952b048341 add term aggregation clarification 2022-10-14 16:12:19 +08:00
PSeitz
80f9596ec8 Merge pull request #1611 from quickwit-oss/remove_token_stream_alloc 
remove tokenstream vec alloc
 2022-10-14 15:12:30 +08:00
Pascal Seitz
84f9e77e1d update CHANGELOG 2022-10-14 15:10:33 +08:00
PSeitz
a602c248fb Merge pull request #1590 from waywardmonkeys/fix-doc-warnings-quickwit 
Fix missing doc warnings when enabling feature "quickwit".
 2022-10-14 14:09:25 +08:00
PSeitz
4b9d1fe828 Merge pull request #1620 from quickwit-oss/fix_fieldnorms_indexing 
Fix missing fieldnorm indexing
 2022-10-14 13:41:38 +08:00
Pascal Seitz
63bc390b02 Fix missing fieldnorm indexing 
Fixes broken search (no results) with BM25 for u64, i64, f64, bool, bytes and date after deletion and merge.
There were no fieldnorms recorded for those field. After merge InvertedIndexReader::total_num_tokens returns 0 (Sum over the fieldnorms is 0). BM25 does not work when total_num_tokens is 0.
Fixes #1617
 2022-10-14 12:44:40 +08:00
Paul Masurel
07393c2fa0 Attempt to fix race condition in test. (#1619) 
Close #1550
 2022-10-14 10:56:37 +09:00
PSeitz
77a415cbe4 rename NothingRecorder to DocIdRecorder (#1615) 2022-10-13 15:43:40 +09:00
PSeitz
4b4c231bba Merge pull request #1612 from quickwit-oss/no_panic_please 
return Error instead panic in fastfields
 2022-10-11 18:33:00 +08:00
PSeitz
11d3409286 add missing docs for fastfield_codecs crate (#1613) 
closes #1603
 2022-10-11 18:54:24 +09:00
Pascal Seitz
9cb8cfbea8 return Error instead panic in fastfields 
fixes #1572
 2022-10-11 14:15:22 +08:00
PSeitz
8b69aab0fc avoid prepare_doc allocation (#1610) 
avoid prepare_doc allocation, ~10% more thoughput best case
 2022-10-11 14:15:55 +09:00
PSeitz
3650d1f36a Merge pull request #1553 from quickwit-oss/ip_field 
ip field
 2022-10-11 13:09:47 +08:00
Pascal Seitz
2efebdb1bb remove tokenstream vec alloc 2022-10-11 10:30:56 +08:00
François Massot
e443ca63aa Merge pull request #1608 from quickwit-oss/nigel/serialise-bytes-as-b64-#2042 
Serialise bytes as base64 strings instead of arrays.
 2022-10-10 11:51:23 +02:00
Pascal Seitz
5c9cbee29d handle IpV4 serialization case 2022-10-07 19:52:00 +08:00
Pascal Seitz
b2ca83a93c switch to ipv6, add monotonic_mapping tests 2022-10-07 18:47:55 +08:00
Nigel Andrews
3b189080d4 Use raw string literals in tests 2022-10-07 12:28:25 +02:00
Nigel Andrews
00a6586efe Replaced String::serialize for serializer.serialize_str 2022-10-07 11:55:05 +02:00
Pascal Seitz
b9b913510e fmt 2022-10-07 16:56:19 +08:00
PSeitz
534b1d33c3 use ipv6 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-10-07 16:56:00 +08:00
PSeitz
f465173872 Apply suggestions from code review 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-10-07 16:55:53 +08:00
Pascal Seitz
96315df20d use idx part only for positions_to_docid 2022-10-07 16:54:04 +08:00
Pascal Seitz
9a1609d364 add test 2022-10-07 16:25:01 +08:00
Pascal Seitz
39f4e58450 improve comment 2022-10-07 16:25:01 +08:00
Pascal Seitz
a8a36b62cd enable test 2022-10-07 16:25:01 +08:00
Pascal Seitz
226a49338f add StrictlyMonotonicFn 2022-10-07 16:25:01 +08:00
Pascal Seitz
2864bf7123 use serializer for u128 2022-10-07 16:25:01 +08:00
Pascal Seitz
5171ff611b serialize ip as u128, add test for positions_to_docid 2022-10-07 16:25:01 +08:00
Pascal Seitz
e50e74acf8 remove u128 type 2022-10-07 16:25:01 +08:00
Pascal Seitz
0b86658389 rename ip addr, use buffer 2022-10-07 16:25:01 +08:00
Pascal Seitz
5d6602a8d9 mark null handling TODO 2022-10-07 16:25:01 +08:00
Pascal Seitz
4d29ff4d01 finalize ip addr rename 2022-10-07 16:25:01 +08:00
Pascal Seitz
cdc8e3a8be group montonic mapping and inverse 
fix mapping inverse
remove ip indexing
add get_between_vals test
 2022-10-07 16:25:01 +08:00
Pascal Seitz
67f453b534 rename to iter_gen 2022-10-07 16:25:01 +08:00
Pascal Seitz
787a37bacf expect instead of unwrap 2022-10-07 16:25:01 +08:00
Pascal Seitz
f5039f1846 remove roaring 2022-10-07 16:25:01 +08:00
Pascal Seitz
eeb1f19093 rename to iter_gen 2022-10-07 16:25:01 +08:00
Pascal Seitz
087beaf328 remove null handling 2022-10-07 16:25:01 +08:00
Pascal Seitz
309449dba3 rename to IpAddr 2022-10-07 16:25:01 +08:00
Pascal Seitz
5a76e6c5d3 fix get_between_vals forwarding 
fix get_between_vals forwarding in monotonicmapping column by adding an additional conversion function Output->Input
 2022-10-07 16:25:01 +08:00
Pascal Seitz
c8713a01ed use iter api 2022-10-07 16:25:01 +08:00
Pascal Seitz
6113e0408c remove comment 2022-10-07 16:25:01 +08:00
Pascal Seitz
400a20b7af add ip field 
add u128 multivalue reader and writer
add ip to schema
add ip writers, handle merge
 2022-10-07 16:25:01 +08:00
PSeitz
5f565e77de Merge pull request #1604 from quickwit-oss/replace_cbor 
replace cbor with cborium
 2022-10-07 14:42:55 +08:00
Pascal Seitz
516e60900d remove unwrap 2022-10-07 14:22:37 +08:00
Pascal Seitz
36e1c79f37 replace cbor with cborium 
closes #1526
 2022-10-07 13:23:39 +08:00
Bruce Mitchener
c2f1c250f9 doc: Remove reference to Searcher pool. (#1598) 
The pool of searchers was removed in 23fe73a6 as part of #1411.
 2022-10-06 00:04:11 +09:00
Bruce Mitchener
c694bc039a Fix missing doc warnings when enabling feature "quickwit". 2022-10-05 20:17:10 +07:00
PSeitz
2063f1717f Merge pull request #1591 from quickwit-oss/ff_refact 
disable linear codec for multivalue values
 2022-10-05 19:39:36 +08:00
Pascal Seitz
d742275048 renames 2022-10-05 19:16:49 +08:00
PSeitz
b9f06bc287 Update src/fastfield/multivalued/mod.rs 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-10-05 19:09:19 +08:00
Pascal Seitz
8b42c4c126 disable linear codec for multivalue value index 
don't materialize index column on merge
use simpler chain() variant
 2022-10-05 19:09:17 +08:00
PSeitz
7905965800 Merge pull request #1594 from quickwit-oss/flat_map_with_buffer 
Removing alloc on all .next() in MultiValueColumn
 2022-10-05 18:34:15 +08:00
Pascal Seitz
f60a551890 add flat_map_with_buffer to Iterator trait 2022-10-05 17:44:26 +08:00
Paul Masurel
7baa6e3ec5 Removing alloc on all .next() in MultiValueColumn 2022-10-05 17:12:06 +09:00
PSeitz
2100ec5d26 Merge pull request #1593 from waywardmonkeys/doc-improvements 
Documentation improvements.
 2022-10-05 15:50:08 +08:00
Bruce Mitchener
b3bf9a5716 Documentation improvements. 2022-10-05 14:18:10 +07:00
Paul Masurel
0dc8c458e0 Flaky unit test. (#1592) 2022-10-05 16:15:48 +09:00
Nigel Andrews
e5043d78d2 added a couple of tests + make fmt 2022-10-04 12:52:44 +02:00
Nigel Andrews
6d0bb82bd2 Fix issue 1576: serialize bytes as base64 strings 2022-10-04 12:18:13 +02:00
trinity-1686a
5945dbf0bd change format for store to make it faster with small documents (#1569) 
* use new format for docstore blocks

* move index to end of block

it makes writing the block faster due to one less memcopy
 2022-10-04 09:58:55 +02:00
PSeitz
4cf911d56a Merge pull request #1587 from quickwit-oss/no_get_val_in_serialize 
remove get_val in serialization
 2022-10-04 12:56:48 +08:00
Pascal Seitz
0f5cff762f move enumerate and remove computation 2022-10-04 12:30:19 +08:00
Pascal Seitz
6d9a123cf2 remove get_val in serialization 
remove get_val in serialization and mark as unimplemented!()
replace get_val with iter in linear codec
remove MultivalueStartIndexRandomSeeker
replace MultivalueStartIndexIter with closure
Sample 100 values in linear codec
 2022-10-04 12:01:25 +08:00
PSeitz
0f4a47816a Merge pull request #1582 from quickwit-oss/faster_sorted_field_values 
use groupby instead of vec allocation
 2022-10-04 09:36:24 +08:00
Pascal Seitz
b062ab2196 use groupby instead of vec allocation 2022-10-04 09:26:26 +08:00
Bruce Mitchener
a9d2f3db23 Tantivy requires Rust 1.62 or later. (#1583) 
Tantivy needs the `total_cmp` feature to compile, which was stabilized
in Rust 1.62.
 2022-10-03 18:31:07 +09:00
Bruce Mitchener
44e03791f9 Fix warnings when doc'ing private items. (#1579) 
This also fixes a couple of typos, but plenty remain!
 2022-10-03 14:24:00 +09:00
Bruce Mitchener
2d23763e9f Use u64::from boolean more. (#1580) 
This case is inverted from the previous cases fixed.

This is from nightly clippy.
 2022-10-03 14:17:50 +09:00
Bruce Mitchener
a24ae8d924 clippy: Fix needless-borrow warnings. (#1581) 
These show on nightly clippy.
 2022-10-03 14:15:09 +09:00
PSeitz
927dff5262 Merge pull request #1578 from quickwit-oss/dead_code 
remove dead indexing code
 2022-10-03 11:25:10 +08:00
Pascal Seitz
a695edcc95 remove dead indexing code 2022-10-03 09:44:02 +08:00
Paul Masurel
b4b4f3fa73 Removing default features for zstd (#1574) 2022-09-30 13:02:46 +09:00
PSeitz
b50e4b7c20 Merge pull request #1566 from quickwit-oss/fix_docstore_sorting 
fix docstore settings for temp docstore
 2022-09-30 10:10:36 +08:00
PSeitz
f8686ab1ec improve comments 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-09-30 10:06:34 +08:00
PSeitz
2fe42719d8 Merge pull request #1570 from quickwit-oss/no_sort_on_multi 
validate index settings on create
 2022-09-30 09:17:03 +08:00
PSeitz
fadd784a25 log improvements (#1564) 2022-09-30 09:39:26 +09:00
Pascal Seitz
0e94213af0 validate index settings on create 2022-09-29 18:58:09 +08:00
PSeitz
0da2a2e70d Merge pull request #1567 from quickwit-oss/dependabot/cargo/tantivy-fst-0.4.0 
Update tantivy-fst requirement from 0.3.0 to 0.4.0
 2022-09-29 10:00:16 +08:00
dependabot[bot]
0bcdf3cbbf Update tantivy-fst requirement from 0.3.0 to 0.4.0 
Updates the requirements on [tantivy-fst](https://github.com/tantivy-search/fst) to permit the latest version.
- [Release notes](https://github.com/tantivy-search/fst/releases)
- [Commits](https://github.com/tantivy-search/fst/commits)

---
updated-dependencies:
- dependency-name: tantivy-fst
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
 2022-09-28 20:50:43 +00:00
Pascal Seitz
8f647b817f fix docstore settings for temp docstore 
fixes #1565
 2022-09-28 17:53:59 +08:00
trinity-1686a
a86b0df6f4 Add query matching terms in a set (#1539) 2022-09-28 09:43:18 +02:00
Bruce Mitchener
f842da758c Move ArcBytes,WeakArcBytes to mmap_directory. (#1555) 
When building without default features (so without mmap, etc),
there are some warnings about unused things. This fixes the
ones related to `ArcBytes` and `WeakArcBytes`, which are only
used with the `mmap_directory` code.
 2022-09-27 09:57:28 +09:00
Bruce Mitchener
97ccd6d712 Avoid slicing a string in DocParsingError. (#1559) 
Fixes #1339.
 2022-09-26 20:27:15 +09:00
Bruce Mitchener
cb252a42af docs: "associated to" -> "associated with" (#1557) 
This reads better this way.
 2022-09-26 20:23:37 +09:00
Bruce Mitchener
d9609dd6b6 POLLING_INTERVAL needn't be pub. (#1556) 
This is only used within the file watcher and is const, so it
can't be configured.
 2022-09-26 20:22:55 +09:00
Bruce Mitchener
f03667d967 Remove references to /cpp directory. (#1560) 
This was removed in 2018, so these should be fine to remove now.
 2022-09-26 20:22:28 +09:00
PSeitz
10f10a322f Merge pull request #1554 from quickwit-oss/prepare_ip_field 
prepare for ip field
 2022-09-26 16:34:24 +08:00
Pascal Seitz
f757471077 prepare for ip field 2022-09-26 16:27:35 +08:00
PSeitz
21e0adefda use binary search instead of linear for get_val in merge (#1548) 
* use binary search instead of linear for get_val in merge

* use partition_point
 2022-09-26 09:42:33 +09:00
Bruce Mitchener
ea8e6d7b1d Tidy up clippy config. (#1547) 
* Checking cfg_attr is no longer necessary.
* Don't need multiple `clippy::` prefixes on a name.
 2022-09-26 09:37:55 +09:00
PSeitz
dac7da780e Merge pull request #1545 from waywardmonkeys/remove-some-refs 
clippy: Remove borrows that the compiler will do.
 2022-09-23 15:33:23 +08:00
PSeitz
20c87903b2 fix multivalue ff index creation regression (#1543) 
fixes multivalue ff regression by avoiding using `get_val`. Line::train calls repeatedly get_val, but get_val implementation on Column for multivalues is very slow. The fix is to use the iterator instead. Longterm fix should be to remove get_val access in serialization.

Old Code

test fastfield::bench::bench_multi_value_ff_merge_few_segments                                                           ... bench:  46,103,960 ns/iter (+/- 2,066,083)
test fastfield::bench::bench_multi_value_ff_merge_many_segments                                                          ... bench:  83,073,036 ns/iter (+/- 4,373,615)
est fastfield::bench::bench_multi_value_ff_merge_many_segments_log_merge                                                ... bench:  64,178,576 ns/iter (+/- 1,466,700)

Current

running 3 tests
test fastfield::multivalued::bench::bench_multi_value_ff_merge_few_segments                                              ... bench:  57,379,523 ns/iter (+/- 3,220,787)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments                                             ... bench:  90,831,688 ns/iter (+/- 1,445,486)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments_log_merge                                   ... bench: 158,313,264 ns/iter (+/- 28,823,250)

With Fix

running 3 tests
test fastfield::multivalued::bench::bench_multi_value_ff_merge_few_segments                                              ... bench:  57,635,671 ns/iter (+/- 2,707,361)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments                                             ... bench:  91,468,712 ns/iter (+/- 11,393,581)
test fastfield::multivalued::bench::bench_multi_value_ff_merge_many_segments_log_merge                                   ... bench:  73,909,138 ns/iter (+/- 15,846,097)
 2022-09-23 15:36:29 +09:00
PSeitz
f9c3947803 Merge pull request #1546 from waywardmonkeys/use-ux-from-bool 
Use u8::from(bool), u64::from(bool).
 2022-09-23 09:06:24 +08:00
Bruce Mitchener
e9a384bb15 Use u8::from(bool), u64::from(bool). 2022-09-22 22:44:53 +07:00
Bruce Mitchener
d231671fe2 clippy: Remove borrows that the compiler will do. 
This started showing up with clippy in rust 1.64.
 2022-09-22 22:38:23 +07:00
trinity-1686a
fa3d786a2f Add support for deleting all documents matching query (#1535) 
* add support for deleting all documents matching query

#1494
 2022-09-22 21:26:09 +09:00
Paul Masurel
75aafeeb9b Added a function to deep clone RamDirectory. (#1544) 2022-09-22 12:04:02 +02:00
PSeitz
6f066c7f65 Merge pull request #1541 from quickwit-oss/add_bench 
add benchmarks for multivalued fastfield merge
 2022-09-22 15:28:00 +08:00
Pascal Seitz
22e56aaee3 add benchmarks for multivalued fastfield merge 2022-09-22 11:25:41 +08:00
Paul Masurel
d641979127 Minor refactor of fast fields (#1538) 2022-09-21 12:55:03 +09:00
Paul Masurel
1998111521 Minor refactoring fast fields (#1537) 2022-09-21 12:46:11 +09:00
PSeitz
acb2e2e282 Merge pull request #1532 from quickwit-oss/refactor_ff 
remove fast_field_cardinality from FastValue
 2022-09-21 04:00:35 +02:00
Pascal Seitz
1ff5da5eb4 remove fast_field_cardinality from FastValue 
unused and at the wrong placed
 2022-09-21 09:38:46 +08:00
Bruce Mitchener
c3b25710ad doc: Improve directory::Lock docs. (#1534) 
Update the docs to reflect the lack of LockParams, correct an error,
and improve cross-linking.
 2022-09-20 18:03:35 +09:00
PSeitz
8492010d43 Merge pull request #1531 from waywardmonkeys/improve-docs-more 
Improvements to doc linking, grammar, etc.
 2022-09-20 15:37:07 +08:00
Bruce Mitchener
cf02e32578 Improvements to doc linking, grammar, etc. 2022-09-19 18:10:22 +07:00
PSeitz
8cca1014c9 Merge pull request #1527 from waywardmonkeys/remove-stream_field-reference 
docs: Remove mentions of stream_field method.
 2022-09-19 17:16:46 +08:00
PSeitz
938f884e32 Merge pull request #1525 from waywardmonkeys/fix-etsy-logo-alt-text-readme 
README: Fix Etsy logo and alt text.
 2022-09-19 16:55:08 +08:00
PSeitz
ed68afb698 Merge pull request #1528 from quickwit-oss/ff_refact 
fix benches
 2022-09-19 11:37:08 +08:00
PSeitz
8a7962dc22 Merge pull request #1524 from waywardmonkeys/improve-docs-1 
Documentation improvements.
 2022-09-19 11:15:42 +08:00
Pascal Seitz
a06039dea8 fix benches 
move some benches to lib.rs to test unexported items
 2022-09-19 11:07:20 +08:00
Bruce Mitchener
68b6254b09 docs: Remove mentions of stream_field method. 
This method doesn't exist, so no need to mention it.
 2022-09-18 23:13:41 +07:00
Bruce Mitchener
6a88ac3fe3 Documentation improvements. 
Fix some linking, some grammar, some typos, etc.
 2022-09-18 18:05:37 +07:00
Bruce Mitchener
191b934650 README: Fix Etsy logo and alt text. 2022-09-18 15:02:35 +07:00
PSeitz
1a2ba7025a Merge pull request #1513 from quickwit-oss/ip_codec 
add ip codec
 2022-09-16 18:53:08 +08:00
Pascal Seitz
02599ebeb7 remove ip_to_u128 2022-09-16 18:16:16 +08:00
Pascal Seitz
a16b466460 merge ColumnExt with Column trait 2022-09-16 18:15:18 +08:00
Pascal Seitz
b8d8fdeb6e move benches, improve bench data 2022-09-16 16:42:23 +08:00
Pascal Seitz
12856d80fa change bench, update numbers 2022-09-16 16:41:01 +08:00
Pascal Seitz
e75472ec9a add serialize_u128, open_u128, refactor 2022-09-16 16:40:59 +08:00
Pascal Seitz
e2e6c94ba8 remove ColumnV2 2022-09-16 16:40:06 +08:00
Pascal Seitz
9f610b25af fix benches, add benches 2022-09-16 16:38:48 +08:00
Pascal Seitz
237b64025e take ColumnV2 as parameter 
improve algorithm
stricter assertions
improve names
 2022-09-16 16:38:48 +08:00
Pascal Seitz
592caeefa0 renames 2022-09-16 16:38:48 +08:00
Pascal Seitz
570009b5b1 move to mod.rs 2022-09-16 16:38:48 +08:00
Pascal Seitz
61b5110db7 use 0 as null in compact space 2022-09-16 16:38:48 +08:00
PSeitz
58af1235e4 Apply suggestions from code review 
Co-authored-by: Paul Masurel <paul@quickwit.io>
 2022-09-16 16:38:48 +08:00
Pascal Seitz
d3e7c41a1f refactor to range_mapping 2022-09-16 16:38:48 +08:00
Pascal Seitz
11275854ca unroll get range iteration 2022-09-16 16:38:48 +08:00
Pascal Seitz
3ca48cd826 fix test 2022-09-16 16:38:48 +08:00
Pascal Seitz
47dc511733 add inline 2022-09-16 16:38:48 +08:00
Pascal Seitz
cae6b28a8f remove num_vals param 2022-09-16 16:38:48 +08:00
Pascal Seitz
9aa9efe2a4 fix bench 2022-09-16 16:38:48 +08:00
Pascal Seitz
57570b38a2 use vint, forward errors, removed unused var 2022-09-16 16:38:48 +08:00
Pascal Seitz
584394db1e fix Cargo.toml 2022-09-16 16:38:48 +08:00
Pascal Seitz
3aeb026970 fix blank_size, add comments 2022-09-16 16:38:48 +08:00
Pascal Seitz
df32ee2df2 refactor, use BTreeSet for sorted deduped values 2022-09-16 16:38:48 +08:00
Pascal Seitz
762e662bfd extend proptest for get_range 2022-09-16 16:38:48 +08:00
Pascal Seitz
63b2420058 fix get_range 
change blank handling
optimize blank collection
fix off by one errors
extend tests
fix get_range
dedupe values to save space
add bench
 2022-09-16 16:38:47 +08:00
Pascal Seitz
ced21b8791 move tests 2022-09-16 16:38:02 +08:00
Pascal Seitz
bc85947105 add ip codec 2022-09-16 16:38:01 +08:00
Paul Masurel
64f08a1a5c Hiding useless symbols and removing code. (#1522) 2022-09-16 14:42:27 +09:00
Paul Masurel
e029fdfca7 Perf fix on the MonotonicMapping column (#1519) 
The Monotonic mapping was using the default implementation
for `get_range` and `.iter`.

As a result, some of the column used in merge (e.g. multivalued
fast fields) were exhibiting a very strong performance regression.
 2022-09-15 14:20:43 +09:00
Paul Masurel
817225edfb Allow for a same-thread doc compressor. (#1510) 
In addition, it isolates the doc compressor logic,
better reports io::Result.

In the case of the same-thread doc compressor,
the blocks are also not copied.
 2022-09-13 15:32:48 +09:00
Shikhar Bhushan
1eab12396d Make Column: Send + Sync (#1518) 2022-09-13 13:31:28 +09:00
dependabot[bot]
8006f63426 Update criterion requirement from 0.3.5 to 0.4.0 (#1517) 
Updates the requirements on [criterion](https://github.com/bheisler/criterion.rs) to permit the latest version.
- [Release notes](https://github.com/bheisler/criterion.rs/releases)
- [Changelog](https://github.com/bheisler/criterion.rs/blob/master/CHANGELOG.md)
- [Commits](https://github.com/bheisler/criterion.rs/compare/0.3.5...0.4.0)

---
updated-dependencies:
- dependency-name: criterion
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
 2022-09-13 10:02:12 +09:00
Adam Reichold
0a907d0319 Move pretty_assertions from dependencies to dev-dependencies to reduce dependency closure of downstream projects. (#1515) 2022-09-10 18:01:26 +09:00
PSeitz
45924711fd improve docs (#1514) 
fix link alias after https://github.com/rust-lang/rustfmt/pull/5262 has been merged and released.
fix dead links
 2022-09-08 22:33:59 +09:00
PSeitz
14cb817a52 Merge pull request #1509 from quickwit-oss/refact-fast-field 
refactor, fix api
 2022-09-07 22:04:32 -07:00
Pascal Seitz
edd9155b88 return Write, add documentation 2022-09-08 12:41:55 +08:00
PSeitz
9497794d40 fix positions docs (#1511) 2022-09-08 10:24:00 +09:00
Pascal Seitz
29d56111de refactor, fix api 
refactor
fix clippy
fix docs
remove unused code
fix bytesfield index api flaw
 2022-09-07 18:43:04 +08:00
Paul Masurel
4d634d61ff Expose memory usage in SingleSegmentIndexWriter (#1508) 2022-09-07 18:33:52 +09:00
PSeitz
1f3d8ca7e2 Merge pull request #1507 from quickwit-oss/improve_test 
add check to proptest
 2022-09-07 02:30:29 -07:00
PSeitz
54696da771 Merge pull request #1505 from quickwit-oss/refact-fast-field 
Refact fast field
 2022-09-07 02:07:42 -07:00
Pascal Seitz
21c2205de9 add check to proptest 2022-09-07 16:58:07 +08:00
PSeitz
9436049d85 Merge pull request #1506 from quickwit-oss/multifastfieldbench 
add benchmark for multivalue fast field
 2022-09-07 01:36:16 -07:00
Pascal Seitz
21c9a26182 add ff creation benchmark 2022-09-07 15:43:50 +08:00
Pascal Seitz
56c68f5869 add ff creation benchmark 2022-09-07 14:03:24 +08:00
Pascal Seitz
f5e66042d8 no alloc in loop 2022-09-07 12:42:16 +08:00
Pascal Seitz
bf3327acd3 add benchmark for multivalue fast field 2022-09-06 16:55:30 +08:00
PSeitz
2a6479b66d Merge pull request #1427 from quickwit-oss/empty_segments_crash 
handle empty segments for merge
 2022-09-05 22:59:06 -07:00
Pascal Seitz
9c2ef81198 fix clippy 2022-09-06 13:34:36 +08:00
Paul Masurel
c5d30a54bc CR 2022-09-06 00:16:41 +09:00
Paul Masurel
c632fc014e Refactoring fast fields codecs. 
This removes the GCD part as a codec, and
makes it so that fastfield codecs all share
the same normalization part (shift + gcd).
 2022-09-05 23:07:12 +09:00
Pascal Seitz
085e63ae43 return new segment meta 2022-09-05 15:19:01 +08:00
Pascal Seitz
f6f23ba684 optionally create segment on merge 
create a new segment only if it contains data

fixes #1189
 2022-09-05 15:07:03 +08:00
Paul Masurel
ea72cf34d6 Int based linear interpol (#1482) 
* Rename BlockwiseLinear to BlockwiseLinearLegacy

Reimplements the blockwise multilinear codec using integer arithmetics.
Added comments

* add estimate for blockwise

* Added one unit test

* use int based for linear interpol

* fix merge conflicts

* reuse code

* cargo fmt

* fix clippy

* fix test

* fix off by one

fix off by one to accurately interpolate autoincrement fields

* extend test, fix estimate

* remove legacy codec

Co-authored-by: Pascal Seitz <pascal.seitz@gmail.com>
 2022-09-05 15:53:00 +09:00
PSeitz
00657d9e99 Merge pull request #1504 from quickwit-oss/move-to-fastfield-codec 
Move to fastfield codec
 2022-09-03 05:18:35 -07:00
Paul Masurel
26876d41d7 Moving the serialization logic to the fastfield_codecs crate. 2022-09-03 00:29:52 +09:00
Paul Masurel
8e775b6c3d Refactoring dyn Column (#1502) 2022-09-02 17:26:30 +09:00
Maxim Kraynyuchenko
e1f9af4384 Added Etsy logo to readme (#1503) 2022-09-02 15:27:59 +09:00
Paul Masurel
4e350c5f1b Clippy 2022-09-02 13:05:00 +09:00
Paul Masurel
84e0c75598 Bench fixing 2022-09-02 11:15:44 +09:00
Paul Masurel
08c4412d73 Adding dragon API to build index without any thread. (#1496) 
Closes #1487
 2022-09-01 10:32:36 +09:00
Shikhar Bhushan
70e58adff9 OwnedBytes doc clarification (#1498) 
It only exposes it with the same lifetime as `&self`, which is what keeps things safe
 2022-09-01 10:32:17 +09:00
PSeitz
0d1cd119e9 Merge pull request #1497 from quickwit-oss/improve_proptest 
custom num strategy, faster test
 2022-08-31 06:25:25 -07:00
Pascal Seitz
d3dd620048 fix clippy 2022-08-31 13:13:56 +02:00
Pascal Seitz
e89c220b56 custom num strategy, faster test 
closes #1486
faster test with rand values
 2022-08-31 12:08:44 +02:00
Paul Masurel
a451f6d60d Minor refactoring. (#1495) 2022-08-31 12:00:58 +09:00
PSeitz
f740ddeee3 Merge pull request #1493 from quickwit-oss/remove_vec_impl 
remove Column impl on Vec
 2022-08-29 07:54:33 -07:00
Pascal Seitz
7a26cc9022 add VecColumn 2022-08-29 15:49:43 +02:00
Pascal Seitz
54972caa7c remove Column impl on Vec 
remove Column impl on Vec to avoid function shadowing
 2022-08-29 11:57:41 +02:00
PSeitz
5d436759b0 Merge pull request #1480 from quickwit-oss/overflow_issue 
fix overflow issue in interpolation
 2022-08-28 16:44:00 -07:00
PSeitz
6f563b1606 Merge pull request #1491 from quickwit-oss/col-trait-refact 
Introducing a column trait
 2022-08-28 10:05:25 -07:00
Pascal Seitz
095fb68fda fix doc test 2022-08-28 18:30:39 +02:00
Pascal Seitz
6316eaefc6 fix benches 2022-08-28 14:38:30 +02:00
Paul Masurel
5331be800b Introducing a column trait 2022-08-28 14:14:27 +02:00
Paul Masurel
c73b425bc1 Fixing unit tests 2022-08-27 23:20:57 +02:00
Paul Masurel
54cfd0d154 Removing Deserializer trait (#1489) 
Removing Deserializer trait and renaming the `Serializer` trait `FastFieldCodec`.
Small refactoring estimate.
 2022-08-28 04:54:55 +09:00
PSeitz
0dd62169c8 merge FastFieldCodecReader wit FastFieldDataAccess (#1485) 
* num_vals to FastFieldCodecReader

* split open_from_bytes to own trait

* rename get_u64 to ge_val

* merge traits
 2022-08-28 03:58:28 +09:00
Paul Masurel
3a9727aa91 Pleasing Clippy 2022-08-27 11:33:03 +02:00
UEDA Akira
17093e8ffe Collapse overlapped highlighted ranges (#1473) 2022-08-26 14:37:08 +09:00
Paul Masurel
03e4630cd8 Mark the CI as successful regardless of whether uploading to Coverall fails. 2022-08-26 07:35:29 +02:00
Paul Masurel
4ae0317d68 Cargo fmt 2022-08-26 00:50:07 +02:00
Paul Masurel
107b19855f Fixing the fastfield codec benchmark (#1484) 2022-08-26 05:54:14 +09:00
Paul Masurel
d8f66ba07e Rename fastfield codecs (#1483) 2022-08-26 01:19:30 +09:00
Paul Masurel
f908549245 Argument missing in bench 2022-08-25 15:42:59 +02:00
Paul Masurel
3673a5df9b Homogeneous codec names. (#1481) 2022-08-25 05:51:37 +09:00
Pascal Seitz
3984cafccc fix overflow issue in interpolation 
use saturating_sub and saturating_add to cover edge cases with values close to u64::MAX or 0 in combination with imprecise computation
 2022-08-24 20:08:13 +02:00
Paul Masurel
298b5dd726 GCD wrapper uses DividerU64 (#1478) 2022-08-25 02:29:13 +09:00
Paul Masurel
8bbb22e9bf Minor refactoring. Introducing a codec type enum. (#1477) 2022-08-25 02:21:41 +09:00
PSeitz
513f68209d Merge pull request #1476 from quickwit-oss/fix_interpol 
add proptest to ff codecs
 2022-08-24 08:01:36 -07:00
Pascal Seitz
91f2f7e722 add proptest to ff codecs 2022-08-24 16:42:40 +02:00
PSeitz
c476b530cf Merge pull request #1432 from quickwit-oss/gcd_encoding 
add gcd test for DateTime
 2022-08-24 06:50:34 -07:00
PSeitz
77dd202e19 Merge pull request #1475 from quickwit-oss/extend_ff_access 
move fastfield stats to trait
 2022-08-24 06:44:57 -07:00
Pascal Seitz
00ebff3c16 move fastfield stats to trait 2022-08-24 15:29:55 +02:00
Paul Masurel
9a6d37c42c Apply suggestions from code review 2022-08-24 21:20:17 +09:00
PSeitz
bb01e99e05 Fixes race condition in Searcher (#1464) 
Fixes a race condition in Searcher, by avoiding repeated calls to open_segment_readers and passing them instead as argument

Closes #1461
 2022-08-24 21:17:37 +09:00
PSeitz
535f1a5d83 Merge pull request #1471 from adamreichold/ci-no-nightly-no-cry 
Split test into check and test CI jobs
 2022-08-24 04:41:42 -07:00
Pascal Seitz
625f9174a7 check for size 2022-08-24 10:32:45 +02:00
Adam Reichold
11a4d97cf5 Use a job matrix to further split and deduplicate the test CI job. 2022-08-24 10:27:57 +02:00
Adam Reichold
1c3d39677a Split checking and testing to a bit more parallelism in the CI. 2022-08-24 10:27:57 +02:00
Pascal Seitz
6f65995cfd remove gcd from api 2022-08-24 10:24:09 +02:00
Pascal Seitz
e2e4190571 add gcd test for DateTime 2022-08-24 10:24:09 +02:00
PSeitz
82209c58aa reuse get_calculated_value (#1472) 2022-08-24 17:16:25 +09:00
Paul Masurel
21519788ea Build fix (#1470) 2022-08-24 07:16:38 +09:00
Shikhar Bhushan
4c6c6e4a9c ConstScoreQuery (#1463) 2022-08-24 06:37:34 +09:00
Adam Reichold
df0ac9e901 Extend facet deserialization to handle owned in addition to borrowed strings. (#1466) 2022-08-24 06:37:13 +09:00
Adam Reichold
71ab482720 RFC: Use a more general but still object-safe signature for Query::query_terms. (#1468) 
* Use a more general but still object-safe signature for Query::query_terms.

* Further constraint the generalized Query::query_terms signature to allow extracting references to terms.
 2022-08-24 06:34:07 +09:00
Adam Reichold
2ae383e452 Cache dependencies in CI to speed up build times. (#1469) 
* Cache dependencies in CI to speed up build times.

* Give cargo-nextest a try.
 2022-08-24 06:27:29 +09:00
PSeitz
8b3a6f6231 Merge pull request #1439 from quickwit-oss/fix_value_range 
fix get calculated value
 2022-08-23 10:15:13 -07:00
PSeitz
11edd6bd59 fix for api change (#1467) 2022-08-24 01:10:12 +09:00
Pascal Seitz
193a3c21f4 fix neg slope calculated value 2022-08-23 13:42:09 +02:00
PSeitz
998b1263f6 Merge pull request #1460 from quickwit-oss/merge_ff_access_iterator 
move iter to FastFieldDataAccess
 2022-08-23 02:58:10 -07:00
Pascal Seitz
72272bdf81 fix variable name 2022-08-23 11:38:27 +02:00
Pascal Seitz
c39c2d79da move iter to FastFieldDataAccess 2022-08-23 11:26:47 +02:00
Paul Masurel
67d94f5bd2 Getting rid of the gcd dependency and using NonZeroU64 in gcd. (#1459) 2022-08-23 07:25:26 +09:00
Paul Masurel
abbd934ac9 Embeds OwnedBytes into the FastFieldCodecReader. (#1458) 2022-08-23 00:02:31 +09:00
Paul Masurel
7f9ba0ee50 Minor readability refactoring in the SegmentDocIdMapping (#1451) 2022-08-22 22:44:36 +09:00
PSeitz
8edcd6f958 Merge pull request #1428 from izihawa/feature/dismax 
[feat] Implement `DisjunctionMaxQuery` and refactor `ScoreCombiner`
 2022-08-22 06:15:30 -07:00
Pasha Podolsky
f50700835d [fix] Fn -> FnOnce 2022-08-22 15:57:30 +03:00
PSeitz
494e92ca59 fix issue in composite (#1456) 
The file offsets were recorded incorrectly in some cases, e.g. when the recording looked like this [(Field 1, Index 0, Offset 0), (Field 1, Index 1, Offset 14), (Field 0, Index 0, Offset 14)]. The last file is offset 14 to end of file for field 0. But the data was converted to a vec and sorted, which changes the last file to Field 1.
 2022-08-22 17:52:12 +09:00
Paul Masurel
4a3169011d clippy (#1452) 2022-08-20 20:01:33 +09:00
Pascal Seitz
050fc5dde9 add comment for diff dance 2022-08-20 08:56:03 +02:00
Paul Masurel
ce45889add Minor codestyle change is prefix of (#1450) 
* Minor code stlye change in the Facet::is_prefix_of.

* bugfix
 2022-08-19 21:20:33 +09:00
dependabot[bot]
4875174d16 Update prettytable-rs requirement from 0.8.0 to 0.9.0 (#1446) 
Updates the requirements on [prettytable-rs](https://github.com/phsym/prettytable-rs) to permit the latest version.
- [Release notes](https://github.com/phsym/prettytable-rs/releases)
- [Commits](https://github.com/phsym/prettytable-rs/compare/v0.8.0...v0.9.0)

---
updated-dependencies:
- dependency-name: prettytable-rs
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
 2022-08-19 18:09:59 +09:00
Kanji Yomoda
0c634c5bc6 Add missing seek to RequiredOptionalScorer (#1442) 2022-08-19 18:08:52 +09:00
Paul Masurel
e25ab5d537 Minor code stlye change in the Facet::is_prefix_of. (#1449) 2022-08-19 18:05:11 +09:00
Adam Reichold
27400c9ad3 Check for the special case of the root facet as prefix of other facets. (#1448) 2022-08-19 17:45:14 +09:00
PSeitz
19074e1d5e Merge pull request #1445 from kianmeng/fix-typos-and-markdowns 
Fix typos and markdowns
 2022-08-18 00:03:37 -07:00
Kian-Meng Ang
014b1adc3e cargo +nightly fmt 2022-08-17 22:33:44 +08:00
Kian-Meng Ang
84295d5b35 cargo fmt 2022-08-15 21:07:01 +08:00
Kian-Meng Ang
625bcb4877 Fix typos and markdowns 
Found via these commands:

    codespell -L crate,ser,panting,beauti,hart,ue,atleast,childs,ond,pris,hel,mot
    markdownlint *.md doc/src/*.md --disable MD013 MD025 MD033 MD001 MD024 MD036 MD041 MD003
 2022-08-13 18:25:47 +08:00
Pascal Seitz
f01cb7d3aa remove cast 2022-08-12 19:50:06 +02:00
PSeitz
8e773ade77 Merge pull request #1444 from quickwit-oss/add-async-doc-freq 
Support for SnippetGenerator in async context
 2022-08-12 05:46:13 -07:00
Evance Soumaoro
fad3faefe2 added InvertedIndexReader::doc_freq_async and SnippetGenerator::new methods 2022-08-12 06:39:10 +00:00
Pascal Seitz
9811d15657 improve slope calculation by delaying f64 cast 2022-08-11 13:32:10 +02:00
Pascal Seitz
31ba5a3c16 fix get calculated value 
fix get calculated value by delaying cast
 2022-08-11 09:44:20 +02:00
Pasha Podolsky
71041b2314 [fix] Fix bench 2022-07-28 21:36:28 +03:00
Pasha Podolsky
09aae134e6 [feat] Implement DisjunctionMaxQuery and refactor ScoreCombiner 2022-07-28 20:47:20 +03:00
276 changed files with 19408 additions and 5872 deletions
Expand all files Collapse all files

1
.gitattributes vendored
View File

@@ -1 +0,0 @@
cpp/* linguist-vendored

4
.github/workflows/coverage.yml vendored
View File

@@ -12,12 +12,14 @@ jobs:
steps:
- uses: actions/checkout@v3
- name: Install Rust
run: rustup toolchain install nightly --component llvm-tools-preview
run: rustup toolchain install nightly --profile minimal --component llvm-tools-preview
- uses: Swatinem/rust-cache@v2
- uses: taiki-e/install-action@cargo-llvm-cov
- name: Generate code coverage
run: cargo +nightly llvm-cov --all-features --workspace --lcov --output-path lcov.info
- name: Upload coverage to Codecov
uses: codecov/codecov-action@v3
continue-on-error: true
with:
token: ${{ secrets.CODECOV_TOKEN }} # not required for public repos
files: lcov.info

3
.github/workflows/long_running.yml vendored
View File

@@ -19,11 +19,10 @@ jobs:
uses: actions-rs/toolchain@v1
with:
toolchain: stable
profile: minimal
override: true
components: rustfmt, clippy
- name: Run indexing_unsorted
run: cargo test indexing_unsorted -- --ignored
- name: Run indexing_sorted
run: cargo test indexing_sorted -- --ignored

52
.github/workflows/test.yml vendored
View File

@@ -10,34 +10,27 @@ env:
CARGO_TERM_COLOR: always
jobs:
test:
check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
- name: Install latest nightly to test also against unstable feature flag
- name: Install nightly
uses: actions-rs/toolchain@v1
with:
toolchain: nightly
override: true
profile: minimal
components: rustfmt
- name: Install stable
uses: actions-rs/toolchain@v1
with:
toolchain: stable
override: true
components: rustfmt, clippy
profile: minimal
components: clippy
- name: Build
run: cargo build --verbose --workspace
- name: Run tests
run: cargo +stable test --features mmap,brotli-compression,lz4-compression,snappy-compression,zstd-compression,failpoints --verbose --workspace
- name: Run tests quickwit feature
run: cargo +stable test --features mmap,quickwit,failpoints --verbose --workspace
- uses: Swatinem/rust-cache@v2
- name: Check Formatting
run: cargo +nightly fmt --all -- --check
@@ -48,3 +41,34 @@ jobs:
token: ${{ secrets.GITHUB_TOKEN }}
args: --tests
test:
runs-on: ubuntu-latest
strategy:
matrix:
features: [
{ label: "all", flags: "mmap,stopwords,brotli-compression,lz4-compression,snappy-compression,zstd-compression,failpoints" },
{ label: "quickwit", flags: "mmap,quickwit,failpoints" }
]
name: test-${{ matrix.features.label}}
steps:
- uses: actions/checkout@v3
- name: Install stable
uses: actions-rs/toolchain@v1
with:
toolchain: stable
profile: minimal
override: true
- uses: taiki-e/install-action@nextest
- uses: Swatinem/rust-cache@v2
- name: Run tests
run: cargo +stable nextest run --features ${{ matrix.features.flags }} --verbose --workspace
- name: Run doctests
run: cargo +stable test --doc --features ${{ matrix.features.flags }} --verbose --workspace

1
.gitignore vendored
View File

@@ -9,7 +9,6 @@ target/release
Cargo.lock
benchmark
.DS_Store
cpp/simdcomp/bitpackingbenchmark
*.bk
.idea
trace.dat

40
ARCHITECTURE.md
View File

@@ -10,6 +10,7 @@ Tantivy's bread and butter is to address the problem of full-text search :
Given a large set of textual documents, and a text query, return the K-most relevant documents in a very efficient way. To execute these queries rapidly, the tantivy needs to build an index beforehand. The relevance score implemented in the tantivy is not configurable. Tantivy uses the same score as the default similarity used in Lucene / Elasticsearch, called [BM25](https://en.wikipedia.org/wiki/Okapi_BM25).
But tantivy's scope does not stop there. Numerous features are required to power rich-search applications. For instance, one may want to:
- compute the count of documents matching a query in the different section of an e-commerce website,
- display an average price per meter square for a real estate search engine,
- take into account historical user data to rank documents in a specific way,
@@ -22,27 +23,28 @@ rapidly select all documents matching a given predicate (also known as a query)
collect some information about them ([See collector](#collector-define-what-to-do-with-matched-documents)).
Roughly speaking the design is following these guiding principles:
- Search should be O(1) in memory.
- Indexing should be O(1) in memory. (In practice it is just sublinear)
- Search should be as fast as possible
This comes at the cost of the dynamicity of the index: while it is possible to add, and delete documents from our corpus, the tantivy is designed to handle these updates in large batches.
## [core/](src/core): Index, segments, searchers.
## [core/](src/core): Index, segments, searchers
Core contains all of the high-level code to make it possible to create an index, add documents, delete documents and commit.
This is both the most high-level part of tantivy, the least performance-sensitive one, the seemingly most mundane code... And paradoxically the most complicated part.
### Index and Segments...
### Index and Segments
A tantivy index is a collection of smaller independent immutable segments.
A tantivy index is a collection of smaller independent immutable segments.
Each segment contains its own independent set of data structures.
A segment is identified by a segment id that is in fact a UUID.
The file of a segment has the format
```segment-id . ext ```
```segment-id . ext```
The extension signals which data structure (or [`SegmentComponent`](src/core/segment_component.rs)) is stored in the file.
@@ -52,17 +54,15 @@ On commit, one segment per indexing thread is written to disk, and the `meta.jso
For a better idea of how indexing works, you may read the [following blog post](https://fulmicoton.com/posts/behold-tantivy-part2/).
### Deletes
Deletes happen by deleting a "term". Tantivy does not offer any notion of primary id, so it is up to the user to use a field in their schema as if it was a primary id, and delete the associated term if they want to delete only one specific document.
On commit, tantivy will find all of the segments with documents matching this existing term and remove from [alive bitset file](src/fastfield/alive_bitset.rs) that represents the bitset of the alive document ids.
Like all segment files, this file is immutable. Because it is possible to have more than one alive bitset file at a given instant, the alive bitset filename has the format ``` segment_id . commit_opstamp . del```.
Like all segment files, this file is immutable. Because it is possible to have more than one alive bitset file at a given instant, the alive bitset filename has the format ```segment_id . commit_opstamp . del```.
An opstamp is simply an incremental id that identifies any operation applied to the index. For instance, performing a commit or adding a document.
### DocId
Within a segment, all documents are identified by a DocId that ranges within `[0, max_doc)`.
@@ -74,6 +74,7 @@ The DocIds are simply allocated in the order documents are added to the index.
In separate threads, tantivy's index writer search for opportunities to merge segments.
The point of segment merge is to:
- eventually get rid of tombstoned documents
- reduce the otherwise ever-growing number of segments.
@@ -94,7 +95,7 @@ called [`Directory`](src/directory/directory.rs).
Contrary to Lucene however, "files" are quite different from some kind of `io::Read` object.
Check out [`src/directory/directory.rs`](src/directory/directory.rs) trait for more details.
Tantivy ships two main directory implementation: the `MMapDirectory` and the `RAMDirectory`,
Tantivy ships two main directory implementation: the `MmapDirectory` and the `RamDirectory`,
but users can extend tantivy with their own implementation.
## [schema/](src/schema): What are documents?
@@ -104,6 +105,7 @@ Tantivy's document follows a very strict schema, decided before building any ind
The schema defines all of the fields that the indexes [`Document`](src/schema/document.rs) may and should contain, their types (`text`, `i64`, `u64`, `Date`, ...) as well as how it should be indexed / represented in tantivy.
Depending on the type of the field, you can decide to
- put it in the docstore
- store it as a fast field
- index it
@@ -117,9 +119,10 @@ As of today, tantivy's schema imposes a 1:1 relationship between a field that is
This is not something tantivy supports, and it is up to the user to duplicate field / concatenate fields before feeding them to tantivy.
## General information about these data structures.
## General information about these data structures
All data structures in tantivy, have:
- a writer
- a serializer
- a reader
@@ -132,7 +135,7 @@ This conversion is done by the serializer.
Finally, the reader is in charge of offering an API to read on this on-disk read-only representation.
In tantivy, readers are designed to require very little anonymous memory. The data is read straight from an mmapped file, and loading an index is as fast as mmapping its files.
## [store/](src/store): Here is my DocId, Gimme my document!
## [store/](src/store): Here is my DocId, Gimme my document
The docstore is a row-oriented storage that, for each document, stores a subset of the fields
that are marked as stored in the schema. The docstore is compressed using a general-purpose algorithm
@@ -146,6 +149,7 @@ Once the top 10 documents have been identified, we fetch them from the store, an
**Not useful for**
Fetching a document from the store is typically a "slow" operation. It usually consists in
- searching into a compact tree-like data structure to find the position of the right block.
- decompressing a small block
- returning the document from this block.
@@ -154,8 +158,7 @@ It is NOT meant to be called for every document matching a query.
As a rule of thumb, if you hit the docstore more than 100 times per search query, you are probably misusing tantivy.
## [fastfield/](src/fastfield): Here is my DocId, Gimme my value!
## [fastfield/](src/fastfield): Here is my DocId, Gimme my value
Fast fields are stored in a column-oriented storage that allows for random access.
The only compression applied is bitpacking. The column comes with two meta data.
@@ -163,7 +166,7 @@ The minimum value in the column and the number of bits per doc.
Fetching a value for a `DocId` is then as simple as computing
```
```rust
min_value + fetch_bits(num_bits * doc_id..num_bits * (doc_id+1))
```
@@ -190,7 +193,7 @@ For advanced search engine, it is possible to store all of the features required
Finally facets are a specific kind of fast field, and the associated source code is in [`fastfield/facet_reader.rs`](src/fastfield/facet_reader.rs).
# The inverted search index.
# The inverted search index
The inverted index is the core part of full-text search.
When presented a new document with the text field "Hello, happy tax payer!", tantivy breaks it into a list of so-called tokens. In addition to just splitting these strings into tokens, it might also do different kinds of operations like dropping the punctuation, converting the character to lowercase, apply stemming, etc. Tantivy makes it possible to configure the operations to be applied in the schema (tokenizer/ is the place where these operations are implemented).
@@ -215,19 +218,18 @@ The inverted index actually consists of two data structures chained together.
Where [TermInfo](src/postings/term_info.rs) is an object containing some meta data about a term.
## [termdict/](src/termdict): Here is a term, give me the [TermInfo](src/postings/term_info.rs)!
## [termdict/](src/termdict): Here is a term, give me the [TermInfo](src/postings/term_info.rs)
Tantivy's term dictionary is mainly in charge of supplying the function
[Term](src/schema/term.rs) ⟶ [TermInfo](src/postings/term_info.rs)
It is itself broken into two parts.
- [Term](src/schema/term.rs) ⟶ [TermOrdinal](src/termdict/mod.rs) is addressed by a finite state transducer, implemented by the fst crate.
- [TermOrdinal](src/termdict/mod.rs) ⟶ [TermInfo](src/postings/term_info.rs) is addressed by the term info store.
## [postings/](src/postings): Iterate over documents... very fast!
## [postings/](src/postings): Iterate over documents... very fast
A posting list makes it possible to store a sorted list of doc ids and for each doc store
a term frequency as well.
@@ -257,7 +259,6 @@ we advance the position reader by the number of term frequencies of the current
The [BM25](https://en.wikipedia.org/wiki/Okapi_BM25) formula also requires to know the number of tokens stored in a specific field for a given document. We store this information on one byte per document in the fieldnorm.
The fieldnorm is therefore compressed. Values up to 40 are encoded unchanged.
## [tokenizer/](src/tokenizer): How should we process text?
Text processing is key to a good search experience.
@@ -268,7 +269,6 @@ Text processing can be configured by selecting an off-the-shelf [`Tokenizer`](./
Tantivy's comes with few tokenizers, but external crates are offering advanced tokenizers, such as [Lindera](https://crates.io/crates/lindera) for Japanese.
## [query/](src/query): Define and compose queries
The [Query](src/query/query.rs) trait defines what a query is.

150
CHANGELOG.md
View File

@@ -1,12 +1,41 @@
Tantivy 0.19
================================
- Updated [Date Field Type](https://github.com/quickwit-oss/tantivy/pull/1396)
The `DateTime` type has been updated to hold timestamps with microseconds precision.
`DateOptions` and `DatePrecision` have been added to configure Date fields. The precision is used to hint on fast values compression. Otherwise, seconds precision is used everywhere else (i.e terms, indexing).
- Remove Searcher pool and make `Searcher` cloneable.
#### Bugfixes
- Fix missing fieldnorms for u64, i64, f64, bool, bytes and date [#1620](https://github.com/quickwit-oss/tantivy/pull/1620) (@PSeitz)
- Fix interpolation overflow in linear interpolation fastfield codec [#1480](https://github.com/quickwit-oss/tantivy/pull/1480) (@PSeitz @fulmicoton)
#### Features/Improvements
- Add support for `IN` in queryparser , e.g. `field: IN [val1 val2 val3]` [#1683](https://github.com/quickwit-oss/tantivy/pull/1683) (@trinity-1686a)
- Skip score calculation, when no scoring is required [#1646](https://github.com/quickwit-oss/tantivy/pull/1646) (@PSeitz)
- Limit fast fields to u32 (`get_val(u32)`) [#1644](https://github.com/quickwit-oss/tantivy/pull/1644) (@PSeitz)
- The `DateTime` type has been updated to hold timestamps with microseconds precision.
`DateOptions` and `DatePrecision` have been added to configure Date fields. The precision is used to hint on fast values compression. Otherwise, seconds precision is used everywhere else (i.e terms, indexing) [#1396](https://github.com/quickwit-oss/tantivy/pull/1396) (@evanxg852000)
- Add IP address field type [#1553](https://github.com/quickwit-oss/tantivy/pull/1553) (@PSeitz)
- Add boolean field type [#1382](https://github.com/quickwit-oss/tantivy/pull/1382) (@boraarslan)
- Remove Searcher pool and make `Searcher` cloneable. (@PSeitz)
- Validate settings on create [#1570](https://github.com/quickwit-oss/tantivy/pull/1570) (@PSeitz)
- Detect and apply gcd on fastfield codecs [#1418](https://github.com/quickwit-oss/tantivy/pull/1418) (@PSeitz)
- Doc store
- use separate thread to compress block store [#1389](https://github.com/quickwit-oss/tantivy/pull/1389) [#1510](https://github.com/quickwit-oss/tantivy/pull/1510) (@PSeitz @fulmicoton)
- Expose doc store cache size [#1403](https://github.com/quickwit-oss/tantivy/pull/1403) (@PSeitz)
- Enable compression levels for doc store [#1378](https://github.com/quickwit-oss/tantivy/pull/1378) (@PSeitz)
- Make block size configurable [#1374](https://github.com/quickwit-oss/tantivy/pull/1374) (@kryesh)
- Make `tantivy::TantivyError` cloneable [#1402](https://github.com/quickwit-oss/tantivy/pull/1402) (@PSeitz)
- Add support for phrase slop in query language [#1393](https://github.com/quickwit-oss/tantivy/pull/1393) (@saroh)
- Aggregation
- Add aggregation support for date type [#1693](https://github.com/quickwit-oss/tantivy/pull/1693)(@PSeitz)
- Add support for keyed parameter in range and histgram aggregations [#1424](https://github.com/quickwit-oss/tantivy/pull/1424) (@k-yomo)
- Add aggregation bucket limit [#1363](https://github.com/quickwit-oss/tantivy/pull/1363) (@PSeitz)
- Faster indexing
- [#1610](https://github.com/quickwit-oss/tantivy/pull/1610) (@PSeitz)
- [#1594](https://github.com/quickwit-oss/tantivy/pull/1594) (@PSeitz)
- [#1582](https://github.com/quickwit-oss/tantivy/pull/1582) (@PSeitz)
- [#1611](https://github.com/quickwit-oss/tantivy/pull/1611) (@PSeitz)
- Added a pre-configured stop word filter for various language [#1666](https://github.com/quickwit-oss/tantivy/pull/1666) (@adamreichold)
Tantivy 0.18
================================
- For date values `chrono` has been replaced with `time` (@uklotzde) #1304 :
- The `time` crate is re-exported as `tantivy::time` instead of `tantivy::chrono`.
- The type alias `tantivy::DateTime` has been removed.
@@ -20,8 +49,13 @@ Tantivy 0.18
- Add terms aggregation (@PSeitz)
- Add support for zstd compression (@kryesh)
Tantivy 0.18.1
================================
- Hotfix: positions computation. #1629 (@fmassot, @fulmicoton, @PSeitz)
Tantivy 0.17
================================
- LogMergePolicy now triggers merges if the ratio of deleted documents reaches a threshold (@shikhar @fulmicoton) [#115](https://github.com/quickwit-oss/tantivy/issues/115)
- Adds a searcher Warmer API (@shikhar @fulmicoton)
- Change to non-strict schema. Ignore fields in data which are not defined in schema. Previously this returned an error. #1211
@@ -36,33 +70,39 @@ Tantivy 0.17
Tantivy 0.16.2
================================
- Bugfix in FuzzyTermQuery. (transposition_cost_one was not doing anything)
Tantivy 0.16.1
========================
- Major Bugfix on multivalued fastfield. #1151
- Demux operation (@PSeitz)
Tantivy 0.16.0
=========================
- Bugfix in the filesum check. (@evanxg852000) #1127
- Bugfix in positions when the index is sorted by a field. (@appaquet) #1125
Tantivy 0.15.3
=========================
- Major bugfix. Deleting documents was broken when the index was sorted by a field. (@appaquet, @fulmicoton) #1101
- Major bugfix. Deleting documents was broken when the index was sorted by a field. (@appaquet, @fulmicoton) #1101
Tantivy 0.15.2
========================
- Major bugfix. DocStore still panics when a deleted doc is at the beginning of a block. (@appaquet) #1088
Tantivy 0.15.1
=========================
- Major bugfix. DocStore panics when first block is deleted. (@appaquet) #1077
Tantivy 0.15.0
=========================
- API Changes. Using Range instead of (start, end) in the API and internals (`FileSlice`, `OwnedBytes`, `Snippets`, ...)
This change is breaking but migration is trivial.
- Added an Histogram collector. (@fulmicoton) #994
@@ -84,9 +124,9 @@ Tantivy 0.15.0
- Updated TermMerger implementation to rely on the union feature of the FST (@scampi) #469
- Add boolean marking whether position is required in the query_terms API call (@fulmicoton). #1070
Tantivy 0.14.0
=========================
- Remove dependency to atomicwrites #833 .Implemented by @fulmicoton upon suggestion and research from @asafigan).
- Migrated tantivy error from the now deprecated `failure` crate to `thiserror` #760. (@hirevo)
- API Change. Accessing the typed value off a `Schema::Value` now returns an Option instead of panicking if the type does not match.
@@ -105,16 +145,19 @@ This version breaks compatibility and requires users to reindex everything.
Tantivy 0.13.2
===================
Bugfix. Acquiring a facet reader on a segment that does not contain any
doc with this facet returns `None`. (#896)
Tantivy 0.13.1
===================
Made `Query` and `Collector` `Send + Sync`.
Updated misc dependency versions.
Tantivy 0.13.0
======================
Tantivy 0.13 introduce a change in the index format that will require
you to reindex your index (BlockWAND information are added in the skiplist).
The index size increase is minor as this information is only added for
@@ -129,6 +172,7 @@ so that we can discuss possible solutions.
A freshly created DocSet point directly to their first doc. A sentinel value called TERMINATED marks the end of a DocSet.
`.advance()` returns the new DocId. `Scorer::skip(target)` has been replaced by `Scorer::seek(target)` and returns the resulting DocId.
As a result, iterating through DocSet now looks as follows
```rust
let mut doc = docset.doc();
while doc != TERMINATED {
@@ -136,7 +180,9 @@ while doc != TERMINATED {
doc = docset.advance();
}
```
The change made it possible to greatly simplify a lot of the docset's code.
- Misc internal optimization and introduction of the `Scorer::for_each_pruning` function. (@fulmicoton)
- Added an offset option to the Top(.*)Collectors. (@robyoung)
- Added Block WAND. Performance on TOP-K on term-unions should be greatly increased. (@fulmicoton, and special thanks
@@ -144,6 +190,7 @@ to the PISA team for answering all my questions!)
Tantivy 0.12.0
======================
- Removing static dispatch in tokenizers for simplicity. (#762)
- Added backward iteration for `TermDictionary` stream. (@halvorboe)
- Fixed a performance issue when searching for the posting lists of a missing term (@audunhalland)
@@ -154,30 +201,32 @@ Tantivy 0.12.0
## How to update?
Crates relying on custom tokenizer, or registering tokenizer in the manager will require some
minor changes. Check https://github.com/quickwit-oss/tantivy/blob/main/examples/custom_tokenizer.rs
minor changes. Check <https://github.com/quickwit-oss/tantivy/blob/main/examples/custom_tokenizer.rs>
to check for some code sample.
Tantivy 0.11.3
=======================
- Fixed DateTime as a fast field (#735)
Tantivy 0.11.2
=======================
- The future returned by `IndexWriter::merge` does not borrow `self` mutably anymore (#732)
- Exposing a constructor for `WatchHandle` (#731)
Tantivy 0.11.1
=====================
- Bug fix #729
- Bug fix #729
Tantivy 0.11.0
=====================
- Added f64 field. Internally reuse u64 code the same way i64 does (@fdb-hiroshima)
- Various bugfixes in the query parser.
- Better handling of hyphens in query parser. (#609)
- Better handling of whitespaces.
- Better handling of hyphens in query parser. (#609)
- Better handling of whitespaces.
- Closes #498 - add support for Elastic-style unbounded range queries for alphanumeric types eg. "title:>hello", "weight:>=70.5", "height:<200" (@petr-tik)
- API change around `Box<BoxableTokenizer>`. See detail in #629
- Avoid rebuilding Regex automaton whenever a regex query is reused. #639 (@brainlock)
@@ -208,7 +257,6 @@ Tantivy 0.10.1
Avoid watching the mmap directory until someone effectively creates a reader that uses
this functionality.
Tantivy 0.10.0
=====================
@@ -224,6 +272,7 @@ Tantivy 0.10.0
Minor
---------
- Switched to Rust 2018 (@uvd)
- Small simplification of the code.
Calling .freq() or .doc() when .advance() has never been called
@@ -231,8 +280,7 @@ on segment postings should panic from now on.
- Tokens exceeding `u16::max_value() - 4` chars are discarded silently instead of panicking.
- Fast fields are now preloaded when the `SegmentReader` is created.
- `IndexMeta` is now public. (@hntd187)
- `IndexWriter` `add_document`, `delete_term`. `IndexWriter` is `Sync`, making it possible to use it with a `
Arc<RwLock<IndexWriter>>`. `add_document` and `delete_term` can
- `IndexWriter` `add_document`, `delete_term`. `IndexWriter` is `Sync`, making it possible to use it with a `Arc<RwLock<IndexWriter>>`. `add_document` and `delete_term` can
only require a read lock. (@fulmicoton)
- Introducing `Opstamp` as an expressive type alias for `u64`. (@petr-tik)
- Stamper now relies on `AtomicU64` on all platforms (@petr-tik)
@@ -248,16 +296,17 @@ Your program should be usable as is.
Fast fields used to be accessed directly from the `SegmentReader`.
The API changed, you are now required to acquire your fast field reader via the
`segment_reader.fast_fields()`, and use one of the typed method:
- `.u64()`, `.i64()` if your field is single-valued ;
- `.u64s()`, `.i64s()` if your field is multi-valued ;
- `.bytes()` if your field is bytes fast field.
Tantivy 0.9.0
=====================
*0.9.0 index format is not compatible with the
previous index format.*
- MAJOR BUGFIX :
Some `Mmap` objects were being leaked, and would never get released. (@fulmicoton)
- Removed most unsafe (@fulmicoton)
@@ -301,37 +350,40 @@ To update from tantivy 0.8, you will need to go through the following steps.
```
Tantivy 0.8.2
=====================
Fixing build for x86_64 platforms. (#496)
No need to update from 0.8.1 if tantivy
is building on your platform.
Tantivy 0.8.1
=====================
Hotfix of #476.
Merge was reflecting deletes before commit was passed.
Thanks @barrotsteindev for reporting the bug.
Tantivy 0.8.0
=====================
*No change in the index format*
- API Breaking change in the collector API. (@jwolfe, @fulmicoton)
- Multithreaded search (@jwolfe, @fulmicoton)
Tantivy 0.7.1
=====================
*No change in the index format*
- Bugfix: NGramTokenizer panics on non ascii chars
- Added a space usage API
Tantivy 0.7
=====================
- Skip data for doc ids and positions (@fulmicoton),
greatly improving performance
- Tantivy error now rely on the failure crate (@drusellers)
@@ -341,15 +393,15 @@ Tantivy 0.7
Tantivy 0.6.1
=========================
- Bugfix #324. GC removing was removing file that were still in useful
- Added support for parsing AllQuery and RangeQuery via QueryParser
- AllQuery: `*`
- RangeQuery:
- Inclusive `field:[startIncl to endIncl]`
- Exclusive `field:{startExcl to endExcl}`
- Mixed `field:[startIncl to endExcl}` and vice versa
- Unbounded `field:[start to *]`, `field:[* to end]`
- AllQuery: `*`
- RangeQuery:
- Inclusive `field:[startIncl to endIncl]`
- Exclusive `field:{startExcl to endExcl}`
- Mixed `field:[startIncl to endExcl}` and vice versa
- Unbounded `field:[start to *]`, `field:[* to end]`
Tantivy 0.6
==========================
@@ -362,58 +414,53 @@ to this release!
- Approximate field norms encoded over 1 byte. (@fulmicoton)
- Compiles on stable rust (@fulmicoton)
- Add &[u8] fastfield for associating arbitrary bytes to each document (@jason-wolfe) (#270)
- Completely uncompressed
- Internally: One u64 fast field for indexes, one fast field for the bytes themselves.
- Completely uncompressed
- Internally: One u64 fast field for indexes, one fast field for the bytes themselves.
- Add NGram token support (@drusellers)
- Add Stopword Filter support (@drusellers)
- Add a FuzzyTermQuery (@drusellers)
- Add a RegexQuery (@drusellers)
- Various performance improvements (@fulmicoton)_
Tantivy 0.5.2
===========================
- bugfix #274
- bugfix #280
- bugfix #289
Tantivy 0.5.1
==========================
- bugfix #254 : tantivy failed if no documents in a segment contained a specific field.
- bugfix #254 : tantivy failed if no documents in a segment contained a specific field.
Tantivy 0.5
==========================
- Faceting
- RangeQuery
- Configurable tokenization pipeline
- Bugfix in PhraseQuery
- Various query optimisation
- Allowing very large indexes
- 64 bits file address
- Smarter encoding of the `TermInfo` objects
- 64 bits file address
- Smarter encoding of the `TermInfo` objects
Tantivy 0.4.3
==========================
- Bugfix race condition when deleting files. (#198)
Tantivy 0.4.2
==========================
- Prevent usage of AVX2 instructions (#201)
Tantivy 0.4.1
==========================
- Bugfix for non-indexed fields. (#199)
Tantivy 0.4.0
==========================
@@ -428,37 +475,31 @@ Tantivy 0.4.0
- Searching for a non-indexed field returns an explicit Error
- Phrase query for non-tokenized field are not tokenized by the query parser.
- Faster/Better indexing (@fulmicoton)
- using murmurhash2
- faster merging
- more memory efficient fast field writer (@lnicola )
- better handling of collisions
- lesser memory usage
- using murmurhash2
- faster merging
- more memory efficient fast field writer (@lnicola )
- better handling of collisions
- lesser memory usage
- Added API, most notably to iterate over ranges of terms (@fulmicoton)
- Bugfix that was preventing to unmap segment files, on index drop (@fulmicoton)
- Made the doc! macro public (@fulmicoton)
- Added an alternative implementation of the streaming dictionary (@fulmicoton)
Tantivy 0.3.1
==========================
- Expose a method to trigger files garbage collection
Tantivy 0.3
==========================
Special thanks to @Kodraus @lnicola @Ameobea @manuel-woelker @celaus
for their contribution to this release.
Thanks also to everyone in tantivy gitter chat
for their advise and company :)
https://gitter.im/tantivy-search/tantivy
<https://gitter.im/tantivy-search/tantivy>
Warning:
@@ -467,19 +508,16 @@ code and index format.
You should not expect backward compatibility before
tantivy 1.0.
New Features
------------
- Delete. You can now delete documents from an index.
- Support for windows (Thanks to @lnicola)
Various Bugfixes & small improvements
----------------------------------------
- Added CI for Windows (https://ci.appveyor.com/project/fulmicoton/tantivy)
- Added CI for Windows (<https://ci.appveyor.com/project/fulmicoton/tantivy>)
Thanks to @KodrAus ! (#108)
- Various dependy version update (Thanks to @Ameobea) #76
- Fixed several race conditions in `Index.wait_merge_threads`
@@ -491,7 +529,3 @@ Thanks to @KodrAus ! (#108)
- Building binary targets for tantivy-cli (Thanks to @KodrAus)
- Misc invisible bug fixes, and code cleanup.
- Use

46
Cargo.toml
View File

@@ -1,6 +1,6 @@
[package]
name = "tantivy"
version = "0.18.0"
version = "0.19.0"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
categories = ["database-implementations", "data-structures"]
@@ -11,40 +11,37 @@ repository = "https://github.com/quickwit-oss/tantivy"
readme = "README.md"
keywords = ["search", "information", "retrieval"]
edition = "2021"
rust-version = "1.62"
[dependencies]
oneshot = "0.1.3"
base64 = "0.13.0"
oneshot = "0.1.5"
base64 = "0.20.0"
byteorder = "1.4.3"
crc32fast = "1.3.2"
once_cell = "1.10.0"
regex = { version = "1.5.5", default-features = false, features = ["std", "unicode"] }
tantivy-fst = "0.3.0"
aho-corasick = "0.7"
tantivy-fst = "0.4.0"
memmap2 = { version = "0.5.3", optional = true }
lz4_flex = { version = "0.9.2", default-features = false, features = ["checked-decode"], optional = true }
brotli = { version = "3.3.4", optional = true }
zstd = { version = "0.11", optional = true }
zstd = { version = "0.12", optional = true, default-features = false }
snap = { version = "1.0.5", optional = true }
tempfile = { version = "3.3.0", optional = true }
log = "0.4.16"
serde = { version = "1.0.136", features = ["derive"] }
serde_json = "1.0.79"
num_cpus = "1.13.1"
fs2={ version = "0.4.3", optional = true }
fs2 = { version = "0.4.3", optional = true }
levenshtein_automata = "0.2.1"
uuid = { version = "1.0.0", features = ["v4", "serde"] }
crossbeam-channel = "0.5.4"
tantivy-query-grammar = { version="0.18.0", path="./query-grammar" }
tantivy-bitpacker = { version="0.2", path="./bitpacker" }
common = { version = "0.3", path = "./common/", package = "tantivy-common" }
fastfield_codecs = { version="0.2", path="./fastfield_codecs", default-features = false }
ownedbytes = { version="0.3", path="./ownedbytes" }
stable_deref_trait = "1.2.0"
rust-stemmers = "1.2.0"
downcast-rs = "1.2.0"
bitpacking = { version = "0.8.4", default-features = false, features = ["bitpacker4x"] }
census = "0.4.0"
fnv = "1.0.7"
rustc-hash = "1.1.0"
thiserror = "1.0.30"
htmlescape = "0.3.1"
fail = "0.5.0"
@@ -56,11 +53,16 @@ lru = "0.7.5"
fastdivide = "0.4.0"
itertools = "0.10.3"
measure_time = "0.8.2"
pretty_assertions = "1.2.1"
serde_cbor = { version = "0.11.2", optional = true }
async-trait = "0.1.53"
arc-swap = "1.5.0"
gcd = "2.1.0"
sstable = { version="0.1", path="./sstable", package ="tantivy-sstable", optional = true }
stacker = { version="0.1", path="./stacker", package ="tantivy-stacker" }
tantivy-query-grammar = { version= "0.19.0", path="./query-grammar" }
tantivy-bitpacker = { version= "0.3", path="./bitpacker" }
common = { version= "0.5", path = "./common/", package = "tantivy-common" }
fastfield_codecs = { version= "0.3", path="./fastfield_codecs", default-features = false }
ownedbytes = { version= "0.5", path="./ownedbytes" }
[target.'cfg(windows)'.dependencies]
winapi = "0.3.9"
@@ -69,11 +71,12 @@ winapi = "0.3.9"
rand = "0.8.5"
maplit = "1.0.2"
matches = "0.1.9"
pretty_assertions = "1.2.1"
proptest = "1.0.0"
criterion = "0.3.5"
criterion = "0.4"
test-log = "0.2.10"
env_logger = "0.9.0"
pprof = { version = "0.10.0", features = ["flamegraph", "criterion"] }
env_logger = "0.10.0"
pprof = { version = "0.11.0", features = ["flamegraph", "criterion"] }
futures = "0.3.21"
[dev-dependencies.fail]
@@ -90,8 +93,9 @@ debug-assertions = true
overflow-checks = true
[features]
default = ["mmap", "lz4-compression" ]
default = ["mmap", "stopwords", "lz4-compression"]
mmap = ["fs2", "tempfile", "memmap2"]
stopwords = []
brotli-compression = ["brotli"]
lz4-compression = ["lz4_flex"]
@@ -101,10 +105,10 @@ zstd-compression = ["zstd"]
failpoints = ["fail/failpoints"]
unstable = [] # useful for benches.
quickwit = ["serde_cbor"]
quickwit = ["sstable"]
[workspace]
members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes"]
members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes", "stacker", "sstable", "columnar"]
# Following the "fail" crate best practises, we isolate
# tests that define specific behavior in fail check points

25
README.md
View File

@@ -5,7 +5,6 @@
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Crates.io](https://img.shields.io/crates/v/tantivy.svg)](https://crates.io/crates/tantivy)
![Tantivy](https://tantivy-search.github.io/logo/tantivy-logo.png)
**Tantivy** is a **full-text search engine library** written in Rust.
@@ -16,7 +15,7 @@ to build such a search engine.
Tantivy is, in fact, strongly inspired by Lucene's design.
If you are looking for an alternative to Elasticsearch or Apache Solr, check out [Quickwit](https://github.com/quickwit-oss/quickwit), our search engine built on top of Tantivy.
If you are looking for an alternative to Elasticsearch or Apache Solr, check out [Quickwit](https://github.com/quickwit-oss/quickwit), our search engine built on top of Tantivy.
# Benchmark
@@ -57,10 +56,9 @@ Your mileage WILL vary depending on the nature of queries and their load.
Distributed search is out of the scope of Tantivy, but if you are looking for this feature, check out [Quickwit](https://github.com/quickwit-oss/quickwit/).
# Getting started
Tantivy works on stable Rust (>= 1.27) and supports Linux, macOS, and Windows.
Tantivy works on stable Rust and supports Linux, macOS, and Windows.
- [Tantivy's simple search example](https://tantivy-search.github.io/examples/basic_search.html)
- [tantivy-cli and its tutorial](https://github.com/quickwit-oss/tantivy-cli) - `tantivy-cli` is an actual command-line interface that makes it easy for you to create a search engine,
@@ -83,9 +81,13 @@ There are many ways to support this project.
We use the GitHub Pull Request workflow: reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
## Minimum supported Rust version
Tantivy currently requires at least Rust 1.62 or later to compile.
## Clone and build locally
Tantivy compiles on stable Rust but requires `Rust >= 1.27`.
Tantivy compiles on stable Rust.
To check out and run tests, you can simply run:
```bash
@@ -125,20 +127,23 @@ By default, `rustc` compiles everything in the `examples/` directory in debug mo
rust-gdb target/debug/examples/$EXAMPLE_NAME
$ gdb run
```
# Companies Using Tantivy
# Companies Using Tantivy
<p align="left">
<img align="center" src="doc/assets/images/etsy.png" alt="Etsy" height="25" width="auto" />&nbsp;
<img align="center" src="doc/assets/images/Nuclia.png#gh-light-mode-only" alt="Nuclia" height="25" width="auto" /> &nbsp;
<img align="center" src="doc/assets/images/humanfirst.png#gh-light-mode-only" alt="Humanfirst.ai" height="30" width="auto" />
<img align="center" src="doc/assets/images/element.io.svg#gh-light-mode-only" alt="Element.io" height="25" width="auto" />
<img align="center" src="doc/assets/images/nuclia-dark-theme.png#gh-dark-mode-only" alt="Nuclia" height="35" width="auto" /> &nbsp;
<img align="center" src="doc/assets/images/humanfirst.ai-dark-theme.png#gh-dark-mode-only" alt="Humanfirst.ai" height="25" width="auto" />&nbsp; &nbsp;
<img align="center" src="doc/assets/images/element-dark-theme.png#gh-dark-mode-only" alt="Element.io" height="25" width="auto" />
</p>
</p>
# FAQ
### Can I use Tantivy in other languages?
- Python → [tantivy-py](https://github.com/quickwit-oss/tantivy-py)
- Ruby → [tantiny](https://github.com/baygeldin/tantiny)
@@ -152,13 +157,17 @@ You can also find other bindings on [GitHub](https://github.com/search?q=tantivy
- and [more](https://github.com/search?q=tantivy)!
### On average, how much faster is Tantivy compared to Lucene?
- According to our [search latency benchmark](https://tantivy-search.github.io/bench/), Tantivy is approximately 2x faster than Lucene.
### Does tantivy support incremental indexing?
- Yes.
### How can I edit documents?
- Data in tantivy is immutable. To edit a document, the document needs to be deleted and reindexed.
### When will my documents be searchable during indexing?
- Documents will be searchable after a `commit` is called on an `IndexWriter`. Existing `IndexReader`s will also need to be reloaded in order to reflect the changes. Finally, changes are only visible to newly acquired `Searcher`.

4
bitpacker/Cargo.toml
View File

@@ -1,6 +1,6 @@
[package]
name = "tantivy-bitpacker"
version = "0.2.0"
version = "0.3.0"
edition = "2021"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
@@ -8,6 +8,8 @@ categories = []
description = """Tantivy-sub crate: bitpacking"""
repository = "https://github.com/quickwit-oss/tantivy"
keywords = []
documentation = "https://docs.rs/tantivy-bitpacker/latest/tantivy_bitpacker"
homepage = "https://github.com/quickwit-oss/tantivy"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html

25
bitpacker/src/bitpacker.rs
View File

@@ -25,15 +25,14 @@ impl BitPacker {
num_bits: u8,
output: &mut TWrite,
) -> io::Result<()> {
let val_u64 = val as u64;
let num_bits = num_bits as usize;
if self.mini_buffer_written + num_bits > 64 {
self.mini_buffer |= val_u64.wrapping_shl(self.mini_buffer_written as u32);
self.mini_buffer |= val.wrapping_shl(self.mini_buffer_written as u32);
output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
self.mini_buffer = val_u64.wrapping_shr((64 - self.mini_buffer_written) as u32);
self.mini_buffer = val.wrapping_shr((64 - self.mini_buffer_written) as u32);
self.mini_buffer_written = self.mini_buffer_written + num_bits - 64;
} else {
self.mini_buffer |= val_u64 << self.mini_buffer_written;
self.mini_buffer |= val << self.mini_buffer_written;
self.mini_buffer_written += num_bits;
if self.mini_buffer_written == 64 {
output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
@@ -82,26 +81,28 @@ impl BitUnpacker {
}
}
pub fn bit_width(&self) -> u8 {
self.num_bits as u8
}
#[inline]
pub fn get(&self, idx: u64, data: &[u8]) -> u64 {
pub fn get(&self, idx: u32, data: &[u8]) -> u64 {
if self.num_bits == 0 {
return 0u64;
}
let num_bits = self.num_bits;
let mask = self.mask;
let addr_in_bits = idx * num_bits;
let addr_in_bits = idx * self.num_bits as u32;
let addr = addr_in_bits >> 3;
let bit_shift = addr_in_bits & 7;
debug_assert!(
addr + 8 <= data.len() as u64,
addr + 8 <= data.len() as u32,
"The fast field field should have been padded with 7 bytes."
);
let bytes: [u8; 8] = (&data[(addr as usize)..(addr as usize) + 8])
.try_into()
.unwrap();
let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
let val_shifted = (val_unshifted_unmasked >> bit_shift) as u64;
val_shifted & mask
let val_shifted: u64 = val_unshifted_unmasked >> bit_shift;
val_shifted & self.mask
}
}
@@ -128,7 +129,7 @@ mod test {
fn test_bitpacker_util(len: usize, num_bits: u8) {
let (bitunpacker, vals, data) = create_fastfield_bitpacker(len, num_bits);
for (i, val) in vals.iter().enumerate() {
assert_eq!(bitunpacker.get(i as u64, &data), *val);
assert_eq!(bitunpacker.get(i as u32, &data), *val);
}
}

18
bitpacker/src/blocked_bitpacker.rs
View File

@@ -58,6 +58,10 @@ fn metadata_test() {
assert_eq!(meta.num_bits(), 6);
}
fn mem_usage<T>(items: &Vec<T>) -> usize {
items.capacity() * std::mem::size_of::<T>()
}
impl BlockedBitpacker {
pub fn new() -> Self {
let mut compressed_blocks = vec![];
@@ -73,16 +77,14 @@ impl BlockedBitpacker {
pub fn mem_usage(&self) -> usize {
std::mem::size_of::<BlockedBitpacker>()
+ self.compressed_blocks.capacity()
+ self.offset_and_bits.capacity()
* std::mem::size_of_val(&self.offset_and_bits.get(0).cloned().unwrap_or_default())
+ self.buffer.capacity()
* std::mem::size_of_val(&self.buffer.get(0).cloned().unwrap_or_default())
+ mem_usage(&self.offset_and_bits)
+ mem_usage(&self.buffer)
}
#[inline]
pub fn add(&mut self, val: u64) {
self.buffer.push(val);
if self.buffer.len() == BLOCK_SIZE as usize {
if self.buffer.len() == BLOCK_SIZE {
self.flush();
}
}
@@ -124,11 +126,11 @@ impl BlockedBitpacker {
}
#[inline]
pub fn get(&self, idx: usize) -> u64 {
let metadata_pos = idx / BLOCK_SIZE as usize;
let pos_in_block = idx % BLOCK_SIZE as usize;
let metadata_pos = idx / BLOCK_SIZE;
let pos_in_block = idx % BLOCK_SIZE;
if let Some(metadata) = self.offset_and_bits.get(metadata_pos) {
let unpacked = BitUnpacker::new(metadata.num_bits()).get(
pos_in_block as u64,
pos_in_block as u32,
&self.compressed_blocks[metadata.offset() as usize..],
);
unpacked + metadata.base_value()

26
columnar/Cargo.toml Normal file
View File

@@ -0,0 +1,26 @@
[package]
name = "tantivy-columnar"
version = "0.1.0"
edition = "2021"
[dependencies]
stacker = { path = "../stacker", package="tantivy-stacker"}
serde_json = "1"
thiserror = "1"
fnv = "1"
tantivy-fst = "0.4.0"
sstable = { path = "../sstable", package = "tantivy-sstable" }
common = { path = "../common", package = "tantivy-common" }
fastfield_codecs = { path = "../fastfield_codecs"}
ordered-float = "3.4"
itertools = "0.10"
[features]
# default = ["quickwit"]
# quickwit = ["common/quickwit"]
[dev-dependencies]
proptest = "1"

33
columnar/README.md Normal file
View File

@@ -0,0 +1,33 @@
# Columnar format
This crate describes columnar format used in tantivy.
## Goals
This format is special in the following way.
- it needs to be compact
- it does not required to be loaded in memory.
- it is designed to fit well with quickwit's strange constraint:
we need to be able to load columns rapidly.
- columns of several types can be associated with the same column name.
- it needs to support columns with different types `(str, u64, i64, f64)`
and different cardinality `(required, optional, multivalued)`.
- columns, once loaded, offer cheap random access.
# Format
A quickwit/tantivy style sstable associated
`(column names, column_cardinality, column_type) to range of bytes.
The format of the key is:
`[column_name][ZERO_BYTE][column_type_header: u8]`
Column name may not contain the zero byte.
Listing all columns associated to `column_name` can therefore
be done by listing all keys prefixed by
`[column_name][ZERO_BYTE]`
The associated range of bytes refer to a range of bytes

154
columnar/src/column_type_header.rs Normal file
View File

@@ -0,0 +1,154 @@
use crate::value::NumericalType;
#[derive(Clone, Copy, Hash, Default, Debug, PartialEq, Eq, PartialOrd, Ord)]
#[repr(u8)]
pub enum Cardinality {
#[default]
Required = 0,
Optional = 1,
Multivalued = 2,
}
impl Cardinality {
pub fn to_code(self) -> u8 {
self as u8
}
pub fn try_from_code(code: u8) -> Option<Cardinality> {
match code {
0 => Some(Cardinality::Required),
1 => Some(Cardinality::Optional),
2 => Some(Cardinality::Multivalued),
_ => None,
}
}
}
#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy)]
pub enum ColumnType {
Bytes,
Numerical(NumericalType),
}
impl ColumnType {
pub fn to_code(self) -> u8 {
match self {
ColumnType::Bytes => 0u8,
ColumnType::Numerical(numerical_type) => 1u8 | (numerical_type.to_code() << 1),
}
}
pub fn try_from_code(code: u8) -> Option<ColumnType> {
if code == 0u8 {
return Some(ColumnType::Bytes);
}
if code & 1u8 == 0u8 {
return None;
}
let numerical_type = NumericalType::try_from_code(code >> 1)?;
Some(ColumnType::Numerical(numerical_type))
}
}
/// Represents the type and cardinality of a column.
/// This is encoded over one-byte and added to a column key in the
/// columnar sstable.
///
/// Cardinality is encoded as the first two highest two bits.
/// The low 6 bits encode the column type.
#[derive(Eq, Hash, PartialEq, Debug, Copy, Clone)]
pub struct ColumnTypeAndCardinality {
pub cardinality: Cardinality,
pub typ: ColumnType,
}
#[inline]
const fn compute_mask(num_bits: u8) -> u8 {
if num_bits == 8 {
u8::MAX
} else {
(1u8 << num_bits) - 1
}
}
#[inline]
fn select_bits<const START: u8, const END: u8>(code: u8) -> u8 {
assert!(START <= END);
assert!(END <= 8);
let num_bits: u8 = END - START;
let mask: u8 = compute_mask(num_bits);
(code >> START) & mask
}
#[inline]
fn place_bits<const START: u8, const END: u8>(code: u8) -> u8 {
assert!(START <= END);
assert!(END <= 8);
let num_bits: u8 = END - START;
let mask: u8 = compute_mask(num_bits);
assert!(code <= mask);
code << START
}
impl ColumnTypeAndCardinality {
pub fn to_code(self) -> u8 {
place_bits::<6, 8>(self.cardinality.to_code()) | place_bits::<0, 6>(self.typ.to_code())
}
pub fn try_from_code(code: u8) -> Option<ColumnTypeAndCardinality> {
let typ_code = select_bits::<0, 6>(code);
let cardinality_code = select_bits::<6, 8>(code);
let cardinality = Cardinality::try_from_code(cardinality_code)?;
let typ = ColumnType::try_from_code(typ_code)?;
assert_eq!(typ.to_code(), typ_code);
Some(ColumnTypeAndCardinality { cardinality, typ })
}
}
#[cfg(test)]
mod tests {
use std::collections::HashSet;
use super::ColumnTypeAndCardinality;
use crate::column_type_header::{Cardinality, ColumnType};
#[test]
fn test_column_type_header_to_code() {
let mut column_type_header_set: HashSet<ColumnTypeAndCardinality> = HashSet::new();
for code in u8::MIN..=u8::MAX {
if let Some(column_type_header) = ColumnTypeAndCardinality::try_from_code(code) {
assert_eq!(column_type_header.to_code(), code);
assert!(column_type_header_set.insert(column_type_header));
}
}
assert_eq!(
column_type_header_set.len(),
3 /* cardinality */ * (1 + 3) // column_types
);
}
#[test]
fn test_column_type_to_code() {
let mut column_type_set: HashSet<ColumnType> = HashSet::new();
for code in u8::MIN..=u8::MAX {
if let Some(column_type) = ColumnType::try_from_code(code) {
assert_eq!(column_type.to_code(), code);
assert!(column_type_set.insert(column_type));
}
}
assert_eq!(column_type_set.len(), 1 + 3);
}
#[test]
fn test_cardinality_to_code() {
let mut num_cardinality = 0;
for code in u8::MIN..=u8::MAX {
let cardinality_opt = Cardinality::try_from_code(code);
if let Some(cardinality) = cardinality_opt {
assert_eq!(cardinality.to_code(), code);
num_cardinality += 1;
}
}
assert_eq!(num_cardinality, 3);
}
}

78
columnar/src/dictionary.rs Normal file
View File

@@ -0,0 +1,78 @@
use std::io;
use fnv::FnvHashMap;
fn fst_err_into_io_err(fst_err: tantivy_fst::Error) -> io::Error {
match fst_err {
tantivy_fst::Error::Fst(fst_err) => {
io::Error::new(io::ErrorKind::Other, format!("FST Error: {:?}", fst_err))
}
tantivy_fst::Error::Io(io_err) => io_err,
}
}
/// `DictionaryBuilder` for dictionary encoding.
///
/// It stores the different terms encounterred and assigns them a temporary value
/// we call unordered id.
///
/// Upon serialization, we will sort the ids and hence build a `UnorderedId -> Term ordinal`
/// mapping.
#[derive(Default)]
pub struct DictionaryBuilder {
dict: FnvHashMap<Vec<u8>, UnorderedId>,
}
pub struct IdMapping {
unordered_to_ord: Vec<OrderedId>,
}
impl IdMapping {
pub fn to_ord(&self, unordered: UnorderedId) -> OrderedId {
self.unordered_to_ord[unordered.0 as usize]
}
}
impl DictionaryBuilder {
/// Get or allocate an unordered id.
/// (This ID is simply an auto-incremented id.)
pub fn get_or_allocate_id(&mut self, term: &[u8]) -> UnorderedId {
if let Some(term_id) = self.dict.get(term) {
return *term_id;
}
let new_id = UnorderedId(self.dict.len() as u32);
self.dict.insert(term.to_vec(), new_id);
new_id
}
/// Serialize the dictionary into an fst, and returns the
/// `UnorderedId -> TermOrdinal` map.
pub fn serialize<'a, W: io::Write + 'a>(&self, wrt: &mut W) -> io::Result<IdMapping> {
serialize_inner(&self.dict, wrt).map_err(fst_err_into_io_err)
}
}
/// Helper function just there for error conversion.
fn serialize_inner<'a, W: io::Write + 'a>(
dict: &FnvHashMap<Vec<u8>, UnorderedId>,
wrt: &mut W,
) -> tantivy_fst::Result<IdMapping> {
let mut terms: Vec<(&[u8], UnorderedId)> =
dict.iter().map(|(k, v)| (k.as_slice(), *v)).collect();
terms.sort_unstable_by_key(|(key, _)| *key);
let mut unordered_to_ord: Vec<OrderedId> = vec![OrderedId(0u32); terms.len()];
let mut fst_builder = tantivy_fst::MapBuilder::new(wrt)?;
for (ord, (key, unordered_id)) in terms.into_iter().enumerate() {
let ordered_id = OrderedId(ord as u32);
fst_builder.insert(key, ord as u64)?;
unordered_to_ord[unordered_id.0 as usize] = ordered_id;
}
fst_builder.finish()?;
Ok(IdMapping { unordered_to_ord })
}
#[derive(Clone, Copy, Debug)]
pub struct UnorderedId(pub u32);
#[derive(Clone, Copy)]
pub struct OrderedId(pub u32);

69
columnar/src/lib.rs Normal file
View File

@@ -0,0 +1,69 @@
// Copyright (C) 2022 Quickwit, Inc.
//
// Quickwit is offered under the AGPL v3.0 and as commercial software.
// For commercial licensing, contact us at hello@quickwit.io.
//
// AGPL:
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU Affero General Public License as
// published by the Free Software Foundation, either version 3 of the
// License, or (at your option) any later version.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Affero General Public License for more details.
//
// You should have received a copy of the GNU Affero General Public License
// along with this program. If not, see <http://www.gnu.org/licenses/>.
mod column_type_header;
mod dictionary;
mod reader;
mod serializer;
mod value;
mod writer;
pub use column_type_header::Cardinality;
pub use reader::ColumnarReader;
pub use serializer::ColumnarSerializer;
pub use writer::ColumnarWriter;
pub type DocId = u32;
#[cfg(test)]
mod tests {
use std::ops::Range;
use common::file_slice::FileSlice;
use crate::column_type_header::ColumnTypeAndCardinality;
use crate::reader::ColumnarReader;
use crate::serializer::ColumnarSerializer;
use crate::value::NumericalValue;
use crate::ColumnarWriter;
#[test]
fn test_dataframe_writer() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(1u32, b"srical.value", NumericalValue::U64(1u64));
dataframe_writer.record_numerical(2u32, b"srical.value", NumericalValue::U64(2u64));
dataframe_writer.record_numerical(4u32, b"srical.value", NumericalValue::I64(2i64));
let mut buffer: Vec<u8> = Vec::new();
let serializer = ColumnarSerializer::new(&mut buffer);
dataframe_writer.serialize(5, serializer).unwrap();
let columnar_fileslice = FileSlice::from(buffer);
let columnar = ColumnarReader::open(columnar_fileslice).unwrap();
assert_eq!(columnar.num_columns(), 1);
let cols: Vec<(ColumnTypeAndCardinality, Range<u64>)> =
columnar.read_columns("srical.value").unwrap();
assert_eq!(cols.len(), 1);
// Right now this 31 bytes are spent as follows
//
// - header 14 bytes
// - vals 8 //< due to padding? could have been 1byte?.
// - null footer 6 bytes
// - version footer 3 bytes // Should be file-wide
assert_eq!(cols[0].1, 0..31);
}
}

66
columnar/src/reader/mod.rs Normal file
View File

@@ -0,0 +1,66 @@
use std::ops::Range;
use std::{io, mem};
use common::file_slice::FileSlice;
use common::BinarySerializable;
use sstable::{Dictionary, SSTableRange};
use crate::column_type_header::ColumnTypeAndCardinality;
fn io_invalid_data(msg: String) -> io::Error {
io::Error::new(io::ErrorKind::InvalidData, msg) // format!("Invalid key found.
// {key_bytes:?}")));
}
pub struct ColumnarReader {
column_dictionary: Dictionary<SSTableRange>,
column_data: FileSlice,
}
impl ColumnarReader {
pub fn num_columns(&self) -> usize {
self.column_dictionary.num_terms()
}
pub fn open(file_slice: FileSlice) -> io::Result<ColumnarReader> {
let (file_slice_without_sstable_len, sstable_len_bytes) =
file_slice.split_from_end(mem::size_of::<u64>());
let mut sstable_len_bytes = sstable_len_bytes.read_bytes()?;
let sstable_len = u64::deserialize(&mut sstable_len_bytes)?;
let (column_data, sstable) =
file_slice_without_sstable_len.split_from_end(sstable_len as usize);
let column_dictionary = Dictionary::open(sstable)?;
Ok(ColumnarReader {
column_dictionary,
column_data,
})
}
pub fn read_columns(
&self,
field_name: &str,
) -> io::Result<Vec<(ColumnTypeAndCardinality, Range<u64>)>> {
let mut start_key = field_name.to_string();
start_key.push('\0');
let mut end_key = field_name.to_string();
end_key.push(1u8 as char);
let mut stream = self
.column_dictionary
.range()
.ge(start_key.as_bytes())
.lt(end_key.as_bytes())
.into_stream()?;
let mut results = Vec::new();
while stream.advance() {
let key_bytes: &[u8] = stream.key();
if !key_bytes.starts_with(start_key.as_bytes()) {
return Err(io_invalid_data(format!("Invalid key found. {key_bytes:?}")));
}
let column_code: u8 = key_bytes.last().cloned().unwrap();
let column_type_and_cardinality = ColumnTypeAndCardinality::try_from_code(column_code)
.ok_or_else(|| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
let range = stream.value().clone();
results.push((column_type_and_cardinality, range));
}
Ok(results)
}
}

39
columnar/src/serializer.rs Normal file
View File

@@ -0,0 +1,39 @@
use std::io;
use std::io::Write;
use std::ops::Range;
use common::CountingWriter;
use sstable::value::RangeWriter;
use sstable::SSTableRange;
pub struct ColumnarSerializer<W: io::Write> {
wrt: CountingWriter<W>,
sstable_range: sstable::Writer<Vec<u8>, RangeWriter>,
}
impl<W: io::Write> ColumnarSerializer<W> {
pub fn new(wrt: W) -> ColumnarSerializer<W> {
let sstable_range: sstable::Writer<Vec<u8>, RangeWriter> =
sstable::Dictionary::<SSTableRange>::builder(Vec::with_capacity(100_000)).unwrap();
ColumnarSerializer {
wrt: CountingWriter::wrap(wrt),
sstable_range,
}
}
pub fn record_column_offsets(&mut self, key: &[u8], byte_range: Range<u64>) -> io::Result<()> {
self.sstable_range.insert(key, &byte_range)
}
pub fn wrt(&mut self) -> &mut CountingWriter<W> {
&mut self.wrt
}
pub fn finalize(mut self) -> io::Result<()> {
let sstable_bytes: Vec<u8> = self.sstable_range.finish()?;
let sstable_num_bytes: u64 = sstable_bytes.len() as u64;
self.wrt.write_all(&sstable_bytes)?;
self.wrt.write_all(&sstable_num_bytes.to_le_bytes()[..])?;
Ok(())
}
}

123
columnar/src/value.rs Normal file
View File

@@ -0,0 +1,123 @@
use ordered_float::NotNan;
#[derive(Copy, Clone, Debug, PartialEq)]
pub enum NumericalValue {
I64(i64),
U64(u64),
F64(NotNan<f64>),
}
impl From<u64> for NumericalValue {
fn from(val: u64) -> NumericalValue {
NumericalValue::U64(val)
}
}
impl From<i64> for NumericalValue {
fn from(val: i64) -> Self {
NumericalValue::I64(val)
}
}
impl From<NotNan<f64>> for NumericalValue {
fn from(val: NotNan<f64>) -> Self {
NumericalValue::F64(val)
}
}
impl NumericalValue {
pub fn numerical_type(&self) -> NumericalType {
match self {
NumericalValue::F64(_) => NumericalType::F64,
NumericalValue::I64(_) => NumericalType::I64,
NumericalValue::U64(_) => NumericalType::U64,
}
}
}
impl Eq for NumericalValue {}
#[derive(Clone, Copy, Debug, Default, Hash, Eq, PartialEq)]
#[repr(u8)]
pub enum NumericalType {
#[default]
I64 = 0,
U64 = 1,
F64 = 2,
}
impl NumericalType {
pub fn to_code(self) -> u8 {
self as u8
}
pub fn try_from_code(code: u8) -> Option<NumericalType> {
match code {
0 => Some(NumericalType::I64),
1 => Some(NumericalType::U64),
2 => Some(NumericalType::F64),
_ => None,
}
}
}
/// We voluntarily avoid using `Into` here to keep this
/// implementation quirk as private as possible.
///
/// This coercion trait actually panics if it is used
/// to convert a loose types to a stricter type.
///
/// The level is strictness is somewhat arbitrary.
/// - i64
/// - u64
/// - f64.
pub(crate) trait Coerce {
fn coerce(numerical_value: NumericalValue) -> Self;
}
impl Coerce for i64 {
fn coerce(value: NumericalValue) -> Self {
match value {
NumericalValue::I64(val) => val,
NumericalValue::U64(val) => val as i64,
NumericalValue::F64(_) => unreachable!(),
}
}
}
impl Coerce for u64 {
fn coerce(value: NumericalValue) -> Self {
match value {
NumericalValue::I64(val) => val as u64,
NumericalValue::U64(val) => val,
NumericalValue::F64(_) => unreachable!(),
}
}
}
impl Coerce for NotNan<f64> {
fn coerce(value: NumericalValue) -> Self {
match value {
NumericalValue::I64(val) => unsafe { NotNan::new_unchecked(val as f64) },
NumericalValue::U64(val) => unsafe { NotNan::new_unchecked(val as f64) },
NumericalValue::F64(val) => val,
}
}
}
#[cfg(test)]
mod tests {
use super::NumericalType;
#[test]
fn test_numerical_type_code() {
let mut num_numerical_type = 0;
for code in u8::MIN..=u8::MAX {
if let Some(numerical_type) = NumericalType::try_from_code(code) {
assert_eq!(numerical_type.to_code(), code);
num_numerical_type += 1;
}
}
assert_eq!(num_numerical_type, 3);
}
}

321
columnar/src/writer/column_operation.rs Normal file
View File

@@ -0,0 +1,321 @@
use std::fmt;
use std::num::NonZeroU8;
use ordered_float::NotNan;
use thiserror::Error;
use crate::dictionary::UnorderedId;
use crate::value::NumericalValue;
use crate::DocId;
/// When we build a columnar dataframe, we first just group
/// all mutations per column, and append them in append-only object.
///
/// We represents all of these operations as `ColumnOperation`.
#[derive(Eq, PartialEq, Debug, Clone, Copy)]
pub(crate) enum ColumnOperation<T> {
NewDoc(DocId),
Value(T),
}
impl<T> From<T> for ColumnOperation<T> {
fn from(value: T) -> Self {
ColumnOperation::Value(value)
}
}
#[allow(clippy::from_over_into)]
pub(crate) trait SymbolValue: Into<MiniBuffer> + Clone + Copy + fmt::Debug {
fn deserialize(header: NonZeroU8, bytes: &mut &[u8]) -> Result<Self, ParseError>;
}
pub(crate) struct MiniBuffer {
pub bytes: [u8; 9],
pub len: usize,
}
impl MiniBuffer {
pub fn as_slice(&self) -> &[u8] {
&self.bytes[..self.len]
}
}
fn compute_header_byte(typ: SymbolType, len: usize) -> u8 {
assert!(len <= 9);
(len << 4) as u8 | typ as u8
}
impl SymbolValue for NumericalValue {
fn deserialize(header_byte: NonZeroU8, bytes: &mut &[u8]) -> Result<Self, ParseError> {
let (typ, len) = parse_header_byte(header_byte)?;
let value_bytes: &[u8];
(value_bytes, *bytes) = bytes.split_at(len);
let symbol: NumericalValue = match typ {
SymbolType::U64 => {
let mut octet: [u8; 8] = [0u8; 8];
octet[..value_bytes.len()].copy_from_slice(value_bytes);
let val: u64 = u64::from_le_bytes(octet);
NumericalValue::U64(val)
}
SymbolType::I64 => {
let mut octet: [u8; 8] = [0u8; 8];
octet[..value_bytes.len()].copy_from_slice(value_bytes);
let encoded: u64 = u64::from_le_bytes(octet);
let val: i64 = decode_zig_zag(encoded);
NumericalValue::I64(val)
}
SymbolType::Float => {
let octet: [u8; 8] =
value_bytes.try_into().map_err(|_| ParseError::InvalidLen {
typ: SymbolType::Float,
len,
})?;
let val_possibly_nan = f64::from_le_bytes(octet);
let val_not_nan = NotNan::new(val_possibly_nan)
.map_err(|_| ParseError::NaN)?;
NumericalValue::F64(val_not_nan)
}
};
Ok(symbol)
}
}
#[allow(clippy::from_over_into)]
impl Into<MiniBuffer> for NumericalValue {
fn into(self) -> MiniBuffer {
let mut bytes = [0u8; 9];
match self {
NumericalValue::F64(val) => {
let len = 8;
let header_byte = compute_header_byte(SymbolType::Float, len);
bytes[0] = header_byte;
bytes[1..].copy_from_slice(&val.to_le_bytes());
MiniBuffer {
bytes,
len: len + 1,
}
}
NumericalValue::U64(val) => {
let len = compute_num_bytes_for_u64(val);
let header_byte = compute_header_byte(SymbolType::U64, len);
bytes[0] = header_byte;
bytes[1..].copy_from_slice(&val.to_le_bytes());
MiniBuffer {
bytes,
len: len + 1,
}
}
NumericalValue::I64(val) => {
let encoded = encode_zig_zag(val);
let len = compute_num_bytes_for_u64(encoded);
let header_byte = compute_header_byte(SymbolType::I64, len);
bytes[0] = header_byte;
bytes[1..].copy_from_slice(&encoded.to_le_bytes());
MiniBuffer {
bytes,
len: len + 1,
}
}
}
}
}
#[allow(clippy::from_over_into)]
impl Into<MiniBuffer> for UnorderedId {
fn into(self) -> MiniBuffer {
let mut bytes = [0u8; 9];
let val = self.0 as u64;
let len = compute_num_bytes_for_u64(val) + 1;
bytes[0] = len as u8;
bytes[1..].copy_from_slice(&val.to_le_bytes());
MiniBuffer { bytes, len }
}
}
impl SymbolValue for UnorderedId {
fn deserialize(header: NonZeroU8, bytes: &mut &[u8]) -> Result<UnorderedId, ParseError> {
let len = header.get() as usize;
let symbol_bytes: &[u8];
(symbol_bytes, *bytes) = bytes.split_at(len);
let mut value_bytes = [0u8; 4];
value_bytes[..len - 1].copy_from_slice(&symbol_bytes[1..]);
let value = u32::from_le_bytes(value_bytes);
Ok(UnorderedId(value))
}
}
const HEADER_MASK: u8 = (1u8 << 4) - 1u8;
fn compute_num_bytes_for_u64(val: u64) -> usize {
let msb = (64u32 - val.leading_zeros()) as usize;
(msb + 7) / 8
}
fn parse_header_byte(byte: NonZeroU8) -> Result<(SymbolType, usize), ParseError> {
let len = (byte.get() as usize) >> 4;
let typ_code = byte.get() & HEADER_MASK;
let typ = SymbolType::try_from(typ_code)?;
Ok((typ, len))
}
#[derive(Error, Debug)]
pub enum ParseError {
#[error("Type byte unknown `{0}`")]
UnknownType(u8),
#[error("Invalid len for type `{len}` for type `{typ:?}`.")]
InvalidLen { typ: SymbolType, len: usize },
#[error("Missing bytes.")]
MissingBytes,
#[error("Not a number value.")]
NaN,
}
impl<V: SymbolValue> ColumnOperation<V> {
pub fn serialize(self) -> MiniBuffer {
match self {
ColumnOperation::NewDoc(doc) => {
let mut minibuf: [u8; 9] = [0u8; 9];
minibuf[0] = 0u8;
minibuf[1..5].copy_from_slice(&doc.to_le_bytes());
MiniBuffer {
bytes: minibuf,
len: 5,
}
}
ColumnOperation::Value(val) => val.into(),
}
}
pub fn deserialize(bytes: &mut &[u8]) -> Result<Self, ParseError> {
if bytes.is_empty() {
return Err(ParseError::MissingBytes);
}
let header_byte = bytes[0];
*bytes = &bytes[1..];
if let Some(header_byte) = NonZeroU8::new(header_byte) {
let value = V::deserialize(header_byte, bytes)?;
Ok(ColumnOperation::Value(value))
} else {
let doc_bytes: &[u8];
(doc_bytes, *bytes) = bytes.split_at(4);
let doc: u32 =
u32::from_le_bytes(doc_bytes.try_into().map_err(|_| ParseError::MissingBytes)?);
Ok(ColumnOperation::NewDoc(doc))
}
}
}
#[derive(Copy, Clone, Debug, Eq, PartialEq)]
#[repr(u8)]
pub enum SymbolType {
U64 = 1u8,
I64 = 2u8,
Float = 3u8,
}
impl TryFrom<u8> for SymbolType {
type Error = ParseError;
fn try_from(byte: u8) -> Result<Self, ParseError> {
match byte {
1u8 => Ok(SymbolType::U64),
2u8 => Ok(SymbolType::I64),
3u8 => Ok(SymbolType::Float),
_ => Err(ParseError::UnknownType(byte)),
}
}
}
fn encode_zig_zag(n: i64) -> u64 {
((n << 1) ^ (n >> 63)) as u64
}
fn decode_zig_zag(n: u64) -> i64 {
((n >> 1) as i64) ^ (-((n & 1) as i64))
}
#[cfg(test)]
mod tests {
use super::{SymbolType, *};
#[track_caller]
fn test_zig_zag_aux(val: i64) {
let encoded = super::encode_zig_zag(val);
assert_eq!(decode_zig_zag(encoded), val);
if let Some(abs_val) = val.checked_abs() {
let abs_val = abs_val as u64;
assert!(encoded <= abs_val * 2);
}
}
#[test]
fn test_zig_zag() {
assert_eq!(encode_zig_zag(0i64), 0u64);
assert_eq!(encode_zig_zag(-1i64), 1u64);
assert_eq!(encode_zig_zag(1i64), 2u64);
test_zig_zag_aux(0i64);
test_zig_zag_aux(i64::MIN);
test_zig_zag_aux(i64::MAX);
}
use proptest::prelude::any;
use proptest::proptest;
proptest! {
#[test]
fn test_proptest_zig_zag(val in any::<i64>()) {
test_zig_zag_aux(val);
}
}
#[track_caller]
fn ser_deser_header_byte_aux(symbol_type: SymbolType, len: usize) {
let header_byte = compute_header_byte(symbol_type, len);
let (serdeser_numerical_type, serdeser_len) =
parse_header_byte(NonZeroU8::new(header_byte).unwrap()).unwrap();
assert_eq!(symbol_type, serdeser_numerical_type);
assert_eq!(len, serdeser_len);
}
#[test]
fn test_header_byte_serialization() {
for len in 1..9 {
ser_deser_header_byte_aux(SymbolType::Float, len);
ser_deser_header_byte_aux(SymbolType::I64, len);
ser_deser_header_byte_aux(SymbolType::U64, len);
}
}
#[track_caller]
fn ser_deser_symbol(symbol: ColumnOperation<NumericalValue>) {
let buf = symbol.serialize();
let mut bytes = &buf.bytes[..];
let serdeser_symbol = ColumnOperation::deserialize(&mut bytes).unwrap();
assert_eq!(bytes.len() + buf.len, buf.bytes.len());
assert_eq!(symbol, serdeser_symbol);
}
#[test]
fn test_compute_num_bytes_for_u64() {
assert_eq!(compute_num_bytes_for_u64(0), 0);
assert_eq!(compute_num_bytes_for_u64(1), 1);
assert_eq!(compute_num_bytes_for_u64(255), 1);
assert_eq!(compute_num_bytes_for_u64(256), 2);
assert_eq!(compute_num_bytes_for_u64((1 << 16) - 1), 2);
assert_eq!(compute_num_bytes_for_u64(1 << 16), 3);
}
#[test]
fn test_symbol_serialization() {
ser_deser_symbol(ColumnOperation::NewDoc(0));
ser_deser_symbol(ColumnOperation::NewDoc(3));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(0i64)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(1i64)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(257u64)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(-257i64)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::I64(i64::MIN)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(0u64)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MIN)));
ser_deser_symbol(ColumnOperation::Value(NumericalValue::U64(u64::MAX)));
}
}

675
columnar/src/writer/mod.rs Normal file
View File

@@ -0,0 +1,675 @@
mod column_operation;
mod value_index;
use std::io::{self, Write};
use column_operation::ColumnOperation;
use common::CountingWriter;
use fastfield_codecs::serialize::ValueIndexInfo;
use fastfield_codecs::{Column, MonotonicallyMappableToU64, VecColumn};
use ordered_float::NotNan;
use stacker::{Addr, ArenaHashMap, ExpUnrolledLinkedList, MemoryArena};
use crate::column_type_header::{ColumnType, ColumnTypeAndCardinality};
use crate::dictionary::{DictionaryBuilder, IdMapping, UnorderedId};
use crate::value::{Coerce, NumericalType, NumericalValue};
use crate::writer::column_operation::SymbolValue;
use crate::writer::value_index::{IndexBuilder, SpareIndexBuilders};
use crate::{Cardinality, ColumnarSerializer, DocId};
#[derive(Copy, Clone, Default)]
struct ColumnWriter {
// Detected cardinality of the column so far.
cardinality: Cardinality,
// Last document inserted.
// None if no doc has been added yet.
last_doc_opt: Option<u32>,
// Buffer containing the serialized values.
values: ExpUnrolledLinkedList,
}
#[derive(Clone, Copy, Default)]
pub struct NumericalColumnWriter {
compatible_numerical_types: CompatibleNumericalTypes,
column_writer: ColumnWriter,
}
#[derive(Clone, Copy)]
struct CompatibleNumericalTypes {
all_values_within_i64_range: bool,
all_values_within_u64_range: bool,
}
impl Default for CompatibleNumericalTypes {
fn default() -> CompatibleNumericalTypes {
CompatibleNumericalTypes {
all_values_within_i64_range: true,
all_values_within_u64_range: true,
}
}
}
impl CompatibleNumericalTypes {
pub fn accept_value(&mut self, numerical_value: NumericalValue) {
match numerical_value {
NumericalValue::I64(val_i64) => {
let value_within_u64_range = val_i64 >= 0i64;
self.all_values_within_u64_range &= value_within_u64_range;
}
NumericalValue::U64(val_u64) => {
let value_within_i64_range = val_u64 < i64::MAX as u64;
self.all_values_within_i64_range &= value_within_i64_range;
}
NumericalValue::F64(_) => {
self.all_values_within_i64_range = false;
self.all_values_within_u64_range = false;
}
}
}
pub fn to_numerical_type(self) -> NumericalType {
if self.all_values_within_i64_range {
NumericalType::I64
} else if self.all_values_within_u64_range {
NumericalType::U64
} else {
NumericalType::F64
}
}
}
impl NumericalColumnWriter {
pub fn record_numerical_value(
&mut self,
doc: DocId,
value: NumericalValue,
arena: &mut MemoryArena,
) {
self.compatible_numerical_types.accept_value(value);
self.column_writer.record(doc, value, arena);
}
}
impl ColumnWriter {
fn symbol_iterator<'a, V: SymbolValue>(
&self,
arena: &MemoryArena,
buffer: &'a mut Vec<u8>,
) -> impl Iterator<Item = ColumnOperation<V>> + 'a {
buffer.clear();
self.values.read_to_end(arena, buffer);
let mut cursor: &[u8] = &buffer[..];
std::iter::from_fn(move || {
if cursor.is_empty() {
return None;
}
let symbol = ColumnOperation::deserialize(&mut cursor)
.expect("Failed to deserialize symbol from in-memory. This should never happen.");
Some(symbol)
})
}
fn delta_with_last_doc(&self, doc: DocId) -> u32 {
self.last_doc_opt
.map(|last_doc| doc - last_doc)
.unwrap_or(doc + 1u32)
}
/// Records a change of the document being recorded.
///
/// This function will also update the cardinality of the column
/// if necessary.
fn record(&mut self, doc: DocId, value: NumericalValue, arena: &mut MemoryArena) {
// Difference between `doc` and the last doc.
match self.delta_with_last_doc(doc) {
0 => {
// This is the last encounterred document.
self.cardinality = Cardinality::Multivalued;
}
1 => {
self.last_doc_opt = Some(doc);
self.write_symbol::<NumericalValue>(ColumnOperation::NewDoc(doc), arena);
}
_ => {
self.cardinality = self.cardinality.max(Cardinality::Optional);
self.last_doc_opt = Some(doc);
self.write_symbol::<NumericalValue>(ColumnOperation::NewDoc(doc), arena);
}
}
self.write_symbol(ColumnOperation::Value(value), arena);
}
// Get the cardinality.
// The overall number of docs in the column is necessary to
// deal with the case where the all docs contain 1 value, except some documents
// at the end of the column.
fn get_cardinality(&self, num_docs: DocId) -> Cardinality {
if self.delta_with_last_doc(num_docs) > 1 {
self.cardinality.max(Cardinality::Optional)
} else {
self.cardinality
}
}
fn write_symbol<V: SymbolValue>(
&mut self,
symbol: ColumnOperation<V>,
arena: &mut MemoryArena,
) {
self.values
.writer(arena)
.extend_from_slice(symbol.serialize().as_slice());
}
}
#[derive(Copy, Clone, Default)]
pub struct BytesColumnWriter {
dictionary_id: u32,
column_writer: ColumnWriter,
}
impl BytesColumnWriter {
pub fn with_dictionary_id(dictionary_id: u32) -> BytesColumnWriter {
BytesColumnWriter {
dictionary_id,
column_writer: Default::default(),
}
}
pub fn record_bytes(
&mut self,
doc: DocId,
bytes: &[u8],
dictionaries: &mut [DictionaryBuilder],
arena: &mut MemoryArena,
) {
let unordered_id = dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes);
let numerical_value = NumericalValue::U64(unordered_id.0 as u64);
self.column_writer.record(doc, numerical_value, arena);
}
}
pub struct ColumnarWriter {
numerical_field_hash_map: ArenaHashMap,
bytes_field_hash_map: ArenaHashMap,
arena: MemoryArena,
// Dictionaries used to store dictionary-encoded values.
dictionaries: Vec<DictionaryBuilder>,
buffers: SpareBuffers,
}
#[derive(Default)]
struct SpareBuffers {
byte_buffer: Vec<u8>,
value_index_builders: SpareIndexBuilders,
i64_values: Vec<i64>,
u64_values: Vec<u64>,
f64_values: Vec<ordered_float::NotNan<f64>>,
}
impl Default for ColumnarWriter {
fn default() -> Self {
ColumnarWriter {
numerical_field_hash_map: ArenaHashMap::new(10_000),
bytes_field_hash_map: ArenaHashMap::new(10_000),
dictionaries: Vec::new(),
arena: MemoryArena::default(),
buffers: SpareBuffers::default(),
}
}
}
#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq, Debug)]
enum BytesOrNumerical {
Bytes,
Numerical,
}
impl ColumnarWriter {
pub fn record_numerical(&mut self, doc: DocId, key: &[u8], numerical_value: NumericalValue) {
let (hash_map, arena) = (&mut self.numerical_field_hash_map, &mut self.arena);
hash_map.mutate_or_create(key, |column_opt: Option<NumericalColumnWriter>| {
let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
column.record_numerical_value(doc, numerical_value, arena);
column
});
}
pub fn record_bytes(&mut self, doc: DocId, key: &[u8], value: &[u8]) {
let (hash_map, arena, dictionaries) = (
&mut self.bytes_field_hash_map,
&mut self.arena,
&mut self.dictionaries,
);
hash_map.mutate_or_create(key, |column_opt: Option<BytesColumnWriter>| {
let mut column: BytesColumnWriter = column_opt.unwrap_or_else(|| {
let dictionary_id = dictionaries.len() as u32;
dictionaries.push(DictionaryBuilder::default());
BytesColumnWriter::with_dictionary_id(dictionary_id)
});
column.record_bytes(doc, value, dictionaries, arena);
column
});
}
pub fn serialize<W: io::Write>(
&mut self,
num_docs: DocId,
mut serializer: ColumnarSerializer<W>,
) -> io::Result<()> {
let mut field_columns: Vec<(&[u8], BytesOrNumerical, Addr)> = self
.numerical_field_hash_map
.iter()
.map(|(term, addr, _)| (term, BytesOrNumerical::Numerical, addr))
.collect();
field_columns.extend(
self.bytes_field_hash_map
.iter()
.map(|(term, addr, _)| (term, BytesOrNumerical::Bytes, addr)),
);
let mut key_buffer = Vec::new();
field_columns.sort_unstable_by_key(|(key, col_type, _)| (*key, *col_type));
let (arena, buffers, dictionaries) = (&self.arena, &mut self.buffers, &self.dictionaries);
for (key, bytes_or_numerical, addr) in field_columns {
let wrt = serializer.wrt();
let start_offset = wrt.written_bytes();
let column_type_and_cardinality: ColumnTypeAndCardinality =
match bytes_or_numerical {
BytesOrNumerical::Bytes => {
let BytesColumnWriter { dictionary_id, column_writer } =
self.bytes_field_hash_map.read(addr);
let dictionary_builder =
&dictionaries[dictionary_id as usize];
serialize_bytes_column(
&column_writer,
num_docs,
dictionary_builder,
arena,
buffers,
wrt,
)?;
ColumnTypeAndCardinality {
cardinality: column_writer.get_cardinality(num_docs),
typ: ColumnType::Bytes,
}
}
BytesOrNumerical::Numerical => {
let NumericalColumnWriter { compatible_numerical_types, column_writer } =
self.numerical_field_hash_map.read(addr);
let cardinality = column_writer.get_cardinality(num_docs);
let numerical_type = compatible_numerical_types.to_numerical_type();
serialize_numerical_column(
cardinality,
numerical_type,
&column_writer,
num_docs,
arena,
buffers,
wrt,
)?;
ColumnTypeAndCardinality {
cardinality,
typ: ColumnType::Numerical(numerical_type),
}
}
};
let end_offset = wrt.written_bytes();
let key_with_type = prepare_key(key, column_type_and_cardinality, &mut key_buffer);
serializer.record_column_offsets(key_with_type, start_offset..end_offset)?;
}
serializer.finalize()?;
Ok(())
}
}
/// Returns a key consisting of the concatenation of the key and the column_type_and_cardinality
/// code.
fn prepare_key<'a>(
key: &[u8],
column_type_cardinality: ColumnTypeAndCardinality,
buffer: &'a mut Vec<u8>,
) -> &'a [u8] {
buffer.clear();
buffer.extend_from_slice(key);
buffer.push(0u8);
buffer.push(column_type_cardinality.to_code());
&buffer[..]
}
fn serialize_bytes_column<W: io::Write>(
column_writer: &ColumnWriter,
num_docs: DocId,
dictionary_builder: &DictionaryBuilder,
arena: &MemoryArena,
buffers: &mut SpareBuffers,
wrt: &mut CountingWriter<W>,
) -> io::Result<()> {
let start_offset = wrt.written_bytes();
let id_mapping: IdMapping = dictionary_builder.serialize(wrt)?;
let dictionary_num_bytes: u32 = (wrt.written_bytes() - start_offset) as u32;
let cardinality = column_writer.get_cardinality(num_docs);
let SpareBuffers {
byte_buffer,
value_index_builders,
u64_values,
..
} = buffers;
let symbol_iterator = column_writer
.symbol_iterator(arena, byte_buffer)
.map(|symbol: ColumnOperation<UnorderedId>| {
// We map unordered ids to ordered ids.
match symbol {
ColumnOperation::Value(unordered_id) => {
let ordered_id = id_mapping.to_ord(unordered_id);
ColumnOperation::Value(ordered_id.0 as u64)
}
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
}
});
serialize_column(
symbol_iterator,
cardinality,
num_docs,
value_index_builders,
u64_values,
wrt,
)?;
wrt.write_all(&dictionary_num_bytes.to_le_bytes()[..])?;
Ok(())
}
fn serialize_numerical_column<W: io::Write>(
cardinality: Cardinality,
numerical_type: NumericalType,
column_writer: &ColumnWriter,
num_docs: DocId,
arena: &MemoryArena,
buffers: &mut SpareBuffers,
wrt: &mut W,
) -> io::Result<()> {
let SpareBuffers {
byte_buffer,
value_index_builders,
u64_values,
i64_values,
f64_values,
} = buffers;
let symbol_iterator = column_writer.symbol_iterator(arena, byte_buffer);
match numerical_type {
NumericalType::I64 => {
serialize_column(
coerce_numerical_symbol::<i64>(symbol_iterator),
cardinality,
num_docs,
value_index_builders,
i64_values,
wrt,
)?;
}
NumericalType::U64 => {
serialize_column(
coerce_numerical_symbol::<u64>(symbol_iterator),
cardinality,
num_docs,
value_index_builders,
u64_values,
wrt,
)?;
}
NumericalType::F64 => {
serialize_column(
coerce_numerical_symbol::<NotNan<f64>>(symbol_iterator),
cardinality,
num_docs,
value_index_builders,
f64_values,
wrt,
)?;
}
};
Ok(())
}
fn serialize_column<
T: Copy + Ord + Default + Send + Sync + MonotonicallyMappableToU64,
W: io::Write,
>(
symbol_iterator: impl Iterator<Item = ColumnOperation<T>>,
cardinality: Cardinality,
num_docs: DocId,
value_index_builders: &mut SpareIndexBuilders,
values: &mut Vec<T>,
wrt: &mut W,
) -> io::Result<()>
where
for<'a> VecColumn<'a, T>: Column<T>,
{
match cardinality {
Cardinality::Required => {
consume_symbol_iterator(
symbol_iterator,
value_index_builders.borrow_required_index_builder(),
values,
);
fastfield_codecs::serialize(
VecColumn::from(&values[..]),
wrt,
&fastfield_codecs::ALL_CODEC_TYPES[..],
)?;
}
Cardinality::Optional => {
let optional_index_builder = value_index_builders.borrow_optional_index_builder();
consume_symbol_iterator(symbol_iterator, optional_index_builder, values);
let optional_index = optional_index_builder.finish(num_docs);
fastfield_codecs::serialize::serialize_new(
ValueIndexInfo::SingleValue(Box::new(optional_index)),
VecColumn::from(&values[..]),
wrt,
&fastfield_codecs::ALL_CODEC_TYPES[..],
)?;
}
Cardinality::Multivalued => {
let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
consume_symbol_iterator(symbol_iterator, multivalued_index_builder, values);
let multivalued_index = multivalued_index_builder.finish(num_docs);
fastfield_codecs::serialize::serialize_new(
ValueIndexInfo::MultiValue(Box::new(multivalued_index)),
VecColumn::from(&values[..]),
wrt,
&fastfield_codecs::ALL_CODEC_TYPES[..],
)?;
}
}
Ok(())
}
fn coerce_numerical_symbol<T>(
symbol_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
) -> impl Iterator<Item = ColumnOperation<T>>
where T: Coerce {
symbol_iterator.map(|symbol| match symbol {
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
ColumnOperation::Value(numerical_value) => {
ColumnOperation::Value(Coerce::coerce(numerical_value))
}
})
}
fn consume_symbol_iterator<T, TIndexBuilder: IndexBuilder>(
symbol_iterator: impl Iterator<Item = ColumnOperation<T>>,
index_builder: &mut TIndexBuilder,
values: &mut Vec<T>,
) {
for symbol in symbol_iterator {
match symbol {
ColumnOperation::NewDoc(doc) => {
index_builder.record_doc(doc);
}
ColumnOperation::Value(value) => {
index_builder.record_value();
values.push(value);
}
}
}
}
#[cfg(test)]
mod tests {
use ordered_float::NotNan;
use stacker::MemoryArena;
use super::prepare_key;
use crate::column_type_header::{ColumnType, ColumnTypeAndCardinality};
use crate::value::{NumericalType, NumericalValue};
use crate::writer::column_operation::ColumnOperation;
use crate::writer::CompatibleNumericalTypes;
use crate::Cardinality;
#[test]
fn test_prepare_key_bytes() {
let mut buffer: Vec<u8> = b"somegarbage".to_vec();
let column_type_and_cardinality = ColumnTypeAndCardinality {
typ: ColumnType::Bytes,
cardinality: Cardinality::Optional,
};
let prepared_key = prepare_key(b"root\0child", column_type_and_cardinality, &mut buffer);
assert_eq!(prepared_key.len(), 12);
assert_eq!(&prepared_key[..10], b"root\0child");
assert_eq!(prepared_key[10], 0u8);
assert_eq!(prepared_key[11], column_type_and_cardinality.to_code());
}
#[test]
fn test_column_writer_required_simple() {
let mut arena = MemoryArena::default();
let mut column_writer = super::ColumnWriter::default();
column_writer.record(0u32, 14i64.into(), &mut arena);
column_writer.record(1u32, 15i64.into(), &mut arena);
column_writer.record(2u32, (-16i64).into(), &mut arena);
assert_eq!(column_writer.get_cardinality(3), Cardinality::Required);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.symbol_iterator(&mut arena, &mut buffer)
.collect();
assert_eq!(symbols.len(), 6);
assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
assert!(matches!(
symbols[1],
ColumnOperation::Value(NumericalValue::I64(14i64))
));
assert!(matches!(symbols[2], ColumnOperation::NewDoc(1u32)));
assert!(matches!(
symbols[3],
ColumnOperation::Value(NumericalValue::I64(15i64))
));
assert!(matches!(symbols[4], ColumnOperation::NewDoc(2u32)));
assert!(matches!(
symbols[5],
ColumnOperation::Value(NumericalValue::I64(-16i64))
));
}
#[test]
fn test_column_writer_optional_cardinality_missing_first() {
let mut arena = MemoryArena::default();
let mut column_writer = super::ColumnWriter::default();
column_writer.record(1u32, 15i64.into(), &mut arena);
column_writer.record(2u32, (-16i64).into(), &mut arena);
assert_eq!(column_writer.get_cardinality(3), Cardinality::Optional);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.symbol_iterator(&mut arena, &mut buffer)
.collect();
assert_eq!(symbols.len(), 4);
assert!(matches!(symbols[0], ColumnOperation::NewDoc(1u32)));
assert!(matches!(
symbols[1],
ColumnOperation::Value(NumericalValue::I64(15i64))
));
assert!(matches!(symbols[2], ColumnOperation::NewDoc(2u32)));
assert!(matches!(
symbols[3],
ColumnOperation::Value(NumericalValue::I64(-16i64))
));
}
#[test]
fn test_column_writer_optional_cardinality_missing_last() {
let mut arena = MemoryArena::default();
let mut column_writer = super::ColumnWriter::default();
column_writer.record(0u32, 15i64.into(), &mut arena);
assert_eq!(column_writer.get_cardinality(2), Cardinality::Optional);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.symbol_iterator(&mut arena, &mut buffer)
.collect();
assert_eq!(symbols.len(), 2);
assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
assert!(matches!(
symbols[1],
ColumnOperation::Value(NumericalValue::I64(15i64))
));
}
#[test]
fn test_column_writer_multivalued() {
let mut arena = MemoryArena::default();
let mut column_writer = super::ColumnWriter::default();
column_writer.record(0u32, 16i64.into(), &mut arena);
column_writer.record(0u32, 17i64.into(), &mut arena);
assert_eq!(column_writer.get_cardinality(1), Cardinality::Multivalued);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.symbol_iterator(&mut arena, &mut buffer)
.collect();
assert_eq!(symbols.len(), 3);
assert!(matches!(symbols[0], ColumnOperation::NewDoc(0u32)));
assert!(matches!(
symbols[1],
ColumnOperation::Value(NumericalValue::I64(16i64))
));
assert!(matches!(
symbols[2],
ColumnOperation::Value(NumericalValue::I64(17i64))
));
}
#[track_caller]
fn test_column_writer_coercion_iter_aux(
values: impl Iterator<Item = NumericalValue>,
expected_numerical_type: NumericalType,
) {
let mut compatible_numerical_types = CompatibleNumericalTypes::default();
for value in values {
compatible_numerical_types.accept_value(value);
}
assert_eq!(
compatible_numerical_types.to_numerical_type(),
expected_numerical_type
);
}
#[track_caller]
fn test_column_writer_coercion_aux(
values: &[NumericalValue],
expected_numerical_type: NumericalType,
) {
test_column_writer_coercion_iter_aux(values.iter().copied(), expected_numerical_type);
test_column_writer_coercion_iter_aux(values.iter().rev().copied(), expected_numerical_type);
}
#[test]
fn test_column_writer_coercion() {
test_column_writer_coercion_aux(&[], NumericalType::I64);
test_column_writer_coercion_aux(&[1i64.into()], NumericalType::I64);
test_column_writer_coercion_aux(&[1u64.into()], NumericalType::I64);
// We don't detect exact integer at the moment. We could!
test_column_writer_coercion_aux(&[NotNan::new(1f64).unwrap().into()], NumericalType::F64);
test_column_writer_coercion_aux(&[u64::MAX.into()], NumericalType::U64);
test_column_writer_coercion_aux(&[(i64::MAX as u64).into()], NumericalType::U64);
test_column_writer_coercion_aux(&[(1u64 << 63).into()], NumericalType::U64);
test_column_writer_coercion_aux(&[1i64.into(), 1u64.into()], NumericalType::I64);
test_column_writer_coercion_aux(&[u64::MAX.into(), (-1i64).into()], NumericalType::F64);
}
}

218
columnar/src/writer/value_index.rs Normal file
View File

@@ -0,0 +1,218 @@
use fastfield_codecs::serialize::{MultiValueIndexInfo, SingleValueIndexInfo};
use crate::DocId;
/// The `IndexBuilder` interprets a sequence of
/// calls of the form:
/// (record_doc,record_value+)*
/// and can then serialize the results into an index.
///
/// It has different implementation depending on whether the
/// cardinality is required, optional, or multivalued.
pub(crate) trait IndexBuilder {
fn record_doc(&mut self, doc: DocId);
#[inline]
fn record_value(&mut self) {}
}
/// The RequiredIndexBuilder does nothing.
#[derive(Default)]
pub struct RequiredIndexBuilder;
impl IndexBuilder for RequiredIndexBuilder {
#[inline(always)]
fn record_doc(&mut self, _doc: DocId) {}
}
#[derive(Default)]
pub struct OptionalIndexBuilder {
docs: Vec<DocId>,
}
struct SingleValueArrayIndex<'a> {
docs: &'a [DocId],
num_docs: DocId,
}
impl<'a> SingleValueIndexInfo for SingleValueArrayIndex<'a> {
fn num_vals(&self) -> u32 {
self.num_docs as u32
}
fn num_non_nulls(&self) -> u32 {
self.docs.len() as u32
}
fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
Box::new(self.docs.iter().copied())
}
}
impl OptionalIndexBuilder {
pub fn finish(&mut self, num_docs: DocId) -> impl SingleValueIndexInfo + '_ {
debug_assert!(self
.docs
.last()
.copied()
.map(|last_doc| last_doc < num_docs)
.unwrap_or(true));
SingleValueArrayIndex {
docs: &self.docs[..],
num_docs,
}
}
fn reset(&mut self) {
self.docs.clear();
}
}
impl IndexBuilder for OptionalIndexBuilder {
#[inline(always)]
fn record_doc(&mut self, doc: DocId) {
debug_assert!(self
.docs
.last()
.copied()
.map(|prev_doc| doc > prev_doc)
.unwrap_or(true));
self.docs.push(doc);
}
}
#[derive(Default)]
pub struct MultivaluedIndexBuilder {
// TODO should we switch to `start_offset`?
end_values: Vec<DocId>,
total_num_vals_seen: u32,
}
pub struct MultivaluedValueArrayIndex<'a> {
end_offsets: &'a [DocId],
}
impl<'a> MultiValueIndexInfo for MultivaluedValueArrayIndex<'a> {
fn num_docs(&self) -> u32 {
self.end_offsets.len() as u32
}
fn num_vals(&self) -> u32 {
self.end_offsets.last().copied().unwrap_or(0u32)
}
fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_> {
if self.end_offsets.is_empty() {
return Box::new(std::iter::empty());
}
let n = self.end_offsets.len();
Box::new(std::iter::once(0u32).chain(self.end_offsets[..n - 1].iter().copied()))
}
}
impl MultivaluedIndexBuilder {
pub fn finish(&mut self, num_docs: DocId) -> impl MultiValueIndexInfo + '_ {
self.end_values
.resize(num_docs as usize, self.total_num_vals_seen);
MultivaluedValueArrayIndex {
end_offsets: &self.end_values[..],
}
}
fn reset(&mut self) {
self.end_values.clear();
self.total_num_vals_seen = 0;
}
}
impl IndexBuilder for MultivaluedIndexBuilder {
fn record_doc(&mut self, doc: DocId) {
self.end_values
.resize(doc as usize, self.total_num_vals_seen);
}
fn record_value(&mut self) {
self.total_num_vals_seen += 1;
}
}
/// The `SpareIndexBuilders` is there to avoid allocating a
/// new index builder for every single column.
#[derive(Default)]
pub struct SpareIndexBuilders {
required_index_builder: RequiredIndexBuilder,
optional_index_builder: OptionalIndexBuilder,
multivalued_index_builder: MultivaluedIndexBuilder,
}
impl SpareIndexBuilders {
pub fn borrow_required_index_builder(&mut self) -> &mut RequiredIndexBuilder {
&mut self.required_index_builder
}
pub fn borrow_optional_index_builder(&mut self) -> &mut OptionalIndexBuilder {
self.optional_index_builder.reset();
&mut self.optional_index_builder
}
pub fn borrow_multivalued_index_builder(&mut self) -> &mut MultivaluedIndexBuilder {
self.multivalued_index_builder.reset();
&mut self.multivalued_index_builder
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_optional_value_index_builder() {
let mut opt_value_index_builder = OptionalIndexBuilder::default();
opt_value_index_builder.record_doc(0u32);
opt_value_index_builder.record_value();
assert_eq!(
&opt_value_index_builder
.finish(1u32)
.iter()
.collect::<Vec<u32>>(),
&[0]
);
opt_value_index_builder.reset();
opt_value_index_builder.record_doc(1u32);
opt_value_index_builder.record_value();
assert_eq!(
&opt_value_index_builder
.finish(2u32)
.iter()
.collect::<Vec<u32>>(),
&[1]
);
}
#[test]
fn test_multivalued_value_index_builder() {
let mut multivalued_value_index_builder = MultivaluedIndexBuilder::default();
multivalued_value_index_builder.record_doc(1u32);
multivalued_value_index_builder.record_value();
multivalued_value_index_builder.record_value();
multivalued_value_index_builder.record_doc(2u32);
multivalued_value_index_builder.record_value();
assert_eq!(
multivalued_value_index_builder
.finish(4u32)
.iter()
.collect::<Vec<u32>>(),
vec![0, 0, 2, 3]
);
multivalued_value_index_builder.reset();
multivalued_value_index_builder.record_doc(2u32);
multivalued_value_index_builder.record_value();
multivalued_value_index_builder.record_value();
assert_eq!(
multivalued_value_index_builder
.finish(4u32)
.iter()
.collect::<Vec<u32>>(),
vec![0, 0, 0, 2]
);
}
}

9
common/Cargo.toml
View File

@@ -1,16 +1,21 @@
[package]
name = "tantivy-common"
version = "0.3.0"
version = "0.5.0"
authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
license = "MIT"
edition = "2021"
description = "common traits and utility functions used by multiple tantivy subcrates"
documentation = "https://docs.rs/tantivy_common/"
homepage = "https://github.com/quickwit-oss/tantivy"
repository = "https://github.com/quickwit-oss/tantivy"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]
byteorder = "1.4.3"
ownedbytes = { version="0.3", path="../ownedbytes" }
ownedbytes = { version= "0.5", path="../ownedbytes" }
async-trait = "0.1"
[dev-dependencies]
proptest = "1.0.0"

20
common/src/bitset.rs
View File

@@ -151,7 +151,7 @@ impl TinySet {
if self.is_empty() {
None
} else {
let lowest = self.0.trailing_zeros() as u32;
let lowest = self.0.trailing_zeros();
self.0 ^= TinySet::singleton(lowest).0;
Some(lowest)
}
@@ -259,11 +259,7 @@ impl BitSet {
// we do not check saturated els.
let higher = el / 64u32;
let lower = el % 64u32;
self.len += if self.tinysets[higher as usize].insert_mut(lower) {
1
} else {
0
};
self.len += u64::from(self.tinysets[higher as usize].insert_mut(lower));
}
/// Inserts an element in the `BitSet`
@@ -272,11 +268,7 @@ impl BitSet {
// we do not check saturated els.
let higher = el / 64u32;
let lower = el % 64u32;
self.len -= if self.tinysets[higher as usize].remove_mut(lower) {
1
} else {
0
};
self.len -= u64::from(self.tinysets[higher as usize].remove_mut(lower));
}
/// Returns true iff the elements is in the `BitSet`.
@@ -285,7 +277,7 @@ impl BitSet {
self.tinyset(el / 64u32).contains(el % 64)
}
/// Returns the first non-empty `TinySet` associated to a bucket lower
/// Returns the first non-empty `TinySet` associated with a bucket lower
/// or greater than bucket.
///
/// Reminder: the tiny set with the bucket `bucket`, represents the
@@ -429,7 +421,7 @@ mod tests {
bitset.serialize(&mut out).unwrap();
let bitset = ReadOnlyBitSet::open(OwnedBytes::new(out));
assert_eq!(bitset.len() as usize, i as usize);
assert_eq!(bitset.len(), i as usize);
}
}
@@ -440,7 +432,7 @@ mod tests {
bitset.serialize(&mut out).unwrap();
let bitset = ReadOnlyBitSet::open(OwnedBytes::new(out));
assert_eq!(bitset.len() as usize, 64);
assert_eq!(bitset.len(), 64);
}
#[test]

84
src/directory/file_slice.rs → common/src/file_slice.rs
View File

@@ -1,23 +1,19 @@
use std::ops::{Deref, Range};
use std::sync::{Arc, Weak};
use std::ops::{Deref, Range, RangeBounds};
use std::sync::Arc;
use std::{fmt, io};
use async_trait::async_trait;
use common::HasLen;
use stable_deref_trait::StableDeref;
use ownedbytes::{OwnedBytes, StableDeref};
use crate::directory::OwnedBytes;
pub type ArcBytes = Arc<dyn Deref<Target = [u8]> + Send + Sync + 'static>;
pub type WeakArcBytes = Weak<dyn Deref<Target = [u8]> + Send + Sync + 'static>;
use crate::HasLen;
/// Objects that represents files sections in tantivy.
///
/// By contract, whatever happens to the directory file, as long as a FileHandle
/// is alive, the data associated with it cannot be altered or destroyed.
///
/// The underlying behavior is therefore specific to the `Directory` that created it.
/// Despite its name, a `FileSlice` may or may not directly map to an actual file
/// The underlying behavior is therefore specific to the `Directory` that
/// created it. Despite its name, a [`FileSlice`] may or may not directly map to an actual file
/// on the filesystem.
#[async_trait]
@@ -27,13 +23,9 @@ pub trait FileHandle: 'static + Send + Sync + HasLen + fmt::Debug {
/// This method may panic if the range requested is invalid.
fn read_bytes(&self, range: Range<usize>) -> io::Result<OwnedBytes>;
#[cfg(feature = "quickwit")]
#[doc(hidden)]
async fn read_bytes_async(
&self,
_byte_range: Range<usize>,
) -> crate::AsyncIoResult<OwnedBytes> {
Err(crate::error::AsyncIoError::AsyncUnsupported)
async fn read_bytes_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
self.read_bytes(byte_range)
}
}
@@ -45,7 +37,7 @@ impl FileHandle for &'static [u8] {
}
#[cfg(feature = "quickwit")]
async fn read_bytes_async(&self, byte_range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
async fn read_bytes_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
Ok(self.read_bytes(byte_range)?)
}
}
@@ -73,6 +65,25 @@ impl fmt::Debug for FileSlice {
}
}
#[inline]
fn combine_ranges<R: RangeBounds<usize>>(orig_range: Range<usize>, rel_range: R) -> Range<usize> {
let start: usize = orig_range.start
+ match rel_range.start_bound().cloned() {
std::ops::Bound::Included(rel_start) => rel_start,
std::ops::Bound::Excluded(rel_start) => rel_start + 1,
std::ops::Bound::Unbounded => 0,
};
assert!(start <= orig_range.end);
let end: usize = match rel_range.end_bound().cloned() {
std::ops::Bound::Included(rel_end) => orig_range.start + rel_end + 1,
std::ops::Bound::Excluded(rel_end) => orig_range.start + rel_end,
std::ops::Bound::Unbounded => orig_range.end,
};
assert!(end >= start);
assert!(end <= orig_range.end);
start..end
}
impl FileSlice {
/// Wraps a FileHandle.
pub fn new(file_handle: Arc<dyn FileHandle>) -> Self {
@@ -96,11 +107,11 @@ impl FileSlice {
///
/// Panics if `byte_range.end` exceeds the filesize.
#[must_use]
pub fn slice(&self, byte_range: Range<usize>) -> FileSlice {
assert!(byte_range.end <= self.len());
#[inline]
pub fn slice<R: RangeBounds<usize>>(&self, byte_range: R) -> FileSlice {
FileSlice {
data: self.data.clone(),
range: self.range.start + byte_range.start..self.range.start + byte_range.end,
range: combine_ranges(self.range.clone(), byte_range),
}
}
@@ -112,7 +123,7 @@ impl FileSlice {
/// Returns a `OwnedBytes` with all of the data in the `FileSlice`.
///
/// The behavior is strongly dependant on the implementation of the underlying
/// The behavior is strongly dependent on the implementation of the underlying
/// `Directory` and the `FileSliceTrait` it creates.
/// In particular, it is up to the `Directory` implementation
/// to handle caching if needed.
@@ -120,9 +131,8 @@ impl FileSlice {
self.data.read_bytes(self.range.clone())
}
#[cfg(feature = "quickwit")]
#[doc(hidden)]
pub async fn read_bytes_async(&self) -> crate::AsyncIoResult<OwnedBytes> {
pub async fn read_bytes_async(&self) -> io::Result<OwnedBytes> {
self.data.read_bytes_async(self.range.clone()).await
}
@@ -140,12 +150,8 @@ impl FileSlice {
.read_bytes(self.range.start + range.start..self.range.start + range.end)
}
#[cfg(feature = "quickwit")]
#[doc(hidden)]
pub async fn read_bytes_slice_async(
&self,
byte_range: Range<usize>,
) -> crate::AsyncIoResult<OwnedBytes> {
pub async fn read_bytes_slice_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
assert!(
self.range.start + byte_range.end <= self.range.end,
"`to` exceeds the fileslice length"
@@ -208,7 +214,7 @@ impl FileHandle for FileSlice {
}
#[cfg(feature = "quickwit")]
async fn read_bytes_async(&self, byte_range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
async fn read_bytes_async(&self, byte_range: Range<usize>) -> io::Result<OwnedBytes> {
self.read_bytes_slice_async(byte_range).await
}
}
@@ -226,7 +232,7 @@ impl FileHandle for OwnedBytes {
}
#[cfg(feature = "quickwit")]
async fn read_bytes_async(&self, range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
async fn read_bytes_async(&self, range: Range<usize>) -> io::Result<OwnedBytes> {
let bytes = self.read_bytes(range)?;
Ok(bytes)
}
@@ -237,9 +243,9 @@ mod tests {
use std::io;
use std::sync::Arc;
use common::HasLen;
use super::{FileHandle, FileSlice};
use crate::file_slice::combine_ranges;
use crate::HasLen;
#[test]
fn test_file_slice() -> io::Result<()> {
@@ -310,4 +316,18 @@ mod tests {
b"bcd"
);
}
#[test]
fn test_combine_range() {
assert_eq!(combine_ranges(1..3, 0..1), 1..2);
assert_eq!(combine_ranges(1..3, 1..), 2..3);
assert_eq!(combine_ranges(1..4, ..2), 1..3);
assert_eq!(combine_ranges(3..10, 2..5), 5..8);
}
#[test]
#[should_panic]
fn test_combine_range_panics() {
let _ = combine_ranges(3..5, 1..4);
}
}

16
common/src/lib.rs
View File

@@ -5,13 +5,17 @@ use std::ops::Deref;
pub use byteorder::LittleEndian as Endianness;
mod bitset;
pub mod file_slice;
mod serialize;
mod vint;
mod writer;
pub use bitset::*;
pub use ownedbytes::OwnedBytes;
pub use serialize::{BinarySerializable, DeserializeFrom, FixedSize};
pub use vint::{read_u32_vint, read_u32_vint_no_advance, serialize_vint_u32, write_u32_vint, VInt};
pub use vint::{
deserialize_vint_u128, read_u32_vint, read_u32_vint_no_advance, serialize_vint_u128,
serialize_vint_u32, write_u32_vint, VInt, VIntU128,
};
pub use writer::{AntiCallToken, CountingWriter, TerminatingWrite};
/// Has length trait
@@ -52,13 +56,13 @@ const HIGHEST_BIT: u64 = 1 << 63;
/// to values over 2^63, and all values end up requiring 64 bits.
///
/// # See also
/// The [reverse mapping is `u64_to_i64`](./fn.u64_to_i64.html).
/// The reverse mapping is [`u64_to_i64()`].
#[inline]
pub fn i64_to_u64(val: i64) -> u64 {
(val as u64) ^ HIGHEST_BIT
}
/// Reverse the mapping given by [`i64_to_u64`](./fn.i64_to_u64.html).
/// Reverse the mapping given by [`i64_to_u64()`].
#[inline]
pub fn u64_to_i64(val: u64) -> i64 {
(val ^ HIGHEST_BIT) as i64
@@ -80,7 +84,7 @@ pub fn u64_to_i64(val: u64) -> i64 {
/// explains the mapping in a clear manner.
///
/// # See also
/// The [reverse mapping is `u64_to_f64`](./fn.u64_to_f64.html).
/// The reverse mapping is [`u64_to_f64()`].
#[inline]
pub fn f64_to_u64(val: f64) -> u64 {
let bits = val.to_bits();
@@ -91,7 +95,7 @@ pub fn f64_to_u64(val: f64) -> u64 {
}
}
/// Reverse the mapping given by [`i64_to_u64`](./fn.i64_to_u64.html).
/// Reverse the mapping given by [`f64_to_u64()`].
#[inline]
pub fn u64_to_f64(val: u64) -> f64 {
f64::from_bits(if val & HIGHEST_BIT != 0 {

32
common/src/serialize.rs
View File

@@ -19,7 +19,7 @@ pub trait DeserializeFrom<T: BinarySerializable> {
/// Implement deserialize from &[u8] for all types which implement BinarySerializable.
///
/// TryFrom would actually be preferrable, but not possible because of the orphan
/// TryFrom would actually be preferable, but not possible because of the orphan
/// rules (not completely sure if this could be resolved)
impl<T: BinarySerializable> DeserializeFrom<T> for &[u8] {
fn deserialize(&mut self) -> io::Result<T> {
@@ -94,6 +94,20 @@ impl FixedSize for u32 {
const SIZE_IN_BYTES: usize = 4;
}
impl BinarySerializable for u16 {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
writer.write_u16::<Endianness>(*self)
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<u16> {
reader.read_u16::<Endianness>()
}
}
impl FixedSize for u16 {
const SIZE_IN_BYTES: usize = 2;
}
impl BinarySerializable for u64 {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
writer.write_u64::<Endianness>(*self)
@@ -107,6 +121,19 @@ impl FixedSize for u64 {
const SIZE_IN_BYTES: usize = 8;
}
impl BinarySerializable for u128 {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
writer.write_u128::<Endianness>(*self)
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
reader.read_u128::<Endianness>()
}
}
impl FixedSize for u128 {
const SIZE_IN_BYTES: usize = 16;
}
impl BinarySerializable for f32 {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
writer.write_f32::<Endianness>(*self)
@@ -161,8 +188,7 @@ impl FixedSize for u8 {
impl BinarySerializable for bool {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
let val = if *self { 1 } else { 0 };
writer.write_u8(val)
writer.write_u8(u8::from(*self))
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<bool> {
let val = reader.read_u8()?;

92
common/src/vint.rs
View File

@@ -5,6 +5,75 @@ use byteorder::{ByteOrder, LittleEndian};
use super::BinarySerializable;
/// Variable int serializes a u128 number
pub fn serialize_vint_u128(mut val: u128, output: &mut Vec<u8>) {
loop {
let next_byte: u8 = (val % 128u128) as u8;
val /= 128u128;
if val == 0 {
output.push(next_byte | STOP_BIT);
return;
} else {
output.push(next_byte);
}
}
}
/// Deserializes a u128 number
///
/// Returns the number and the slice after the vint
pub fn deserialize_vint_u128(data: &[u8]) -> io::Result<(u128, &[u8])> {
let mut result = 0u128;
let mut shift = 0u64;
for i in 0..19 {
let b = data[i];
result |= u128::from(b % 128u8) << shift;
if b >= STOP_BIT {
return Ok((result, &data[i + 1..]));
}
shift += 7;
}
Err(io::Error::new(
io::ErrorKind::InvalidData,
"Failed to deserialize u128 vint",
))
}
/// Wrapper over a `u128` that serializes as a variable int.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct VIntU128(pub u128);
impl BinarySerializable for VIntU128 {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
let mut buffer = vec![];
serialize_vint_u128(self.0, &mut buffer);
writer.write_all(&buffer)
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<Self> {
let mut bytes = reader.bytes();
let mut result = 0u128;
let mut shift = 0u64;
loop {
match bytes.next() {
Some(Ok(b)) => {
result |= u128::from(b % 128u8) << shift;
if b >= STOP_BIT {
return Ok(VIntU128(result));
}
shift += 7;
}
_ => {
return Err(io::Error::new(
io::ErrorKind::InvalidData,
"Reach end of buffer while reading VInt",
));
}
}
}
}
}
/// Wrapper over a `u64` that serializes as a variable int.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct VInt(pub u64);
@@ -88,7 +157,7 @@ fn vint_len(data: &[u8]) -> usize {
/// If the buffer does not start by a valid
/// vint payload
pub fn read_u32_vint(data: &mut &[u8]) -> u32 {
let (result, vlen) = read_u32_vint_no_advance(*data);
let (result, vlen) = read_u32_vint_no_advance(data);
*data = &data[vlen..];
result
}
@@ -176,6 +245,7 @@ impl BinarySerializable for VInt {
mod tests {
use super::{serialize_vint_u32, BinarySerializable, VInt};
use crate::vint::{deserialize_vint_u128, serialize_vint_u128, VIntU128};
fn aux_test_vint(val: u64) {
let mut v = [14u8; 10];
@@ -217,6 +287,26 @@ mod tests {
assert_eq!(&buffer[..len_vint], res2, "array wrong for {}", val);
}
fn aux_test_vint_u128(val: u128) {
let mut data = vec![];
serialize_vint_u128(val, &mut data);
let (deser_val, _data) = deserialize_vint_u128(&data).unwrap();
assert_eq!(val, deser_val);
let mut out = vec![];
VIntU128(val).serialize(&mut out).unwrap();
let deser_val = VIntU128::deserialize(&mut &out[..]).unwrap();
assert_eq!(val, deser_val.0);
}
#[test]
fn test_vint_u128() {
aux_test_vint_u128(0);
aux_test_vint_u128(1);
aux_test_vint_u128(u128::MAX / 3);
aux_test_vint_u128(u128::MAX);
}
#[test]
fn test_vint_u32() {
aux_test_serialize_vint_u32(0);

4
common/src/writer.rs
View File

@@ -55,14 +55,14 @@ impl<W: TerminatingWrite> TerminatingWrite for CountingWriter<W> {
}
/// Struct used to prevent from calling
/// [`terminate_ref`](trait.TerminatingWrite.html#tymethod.terminate_ref) directly
/// [`terminate_ref`](TerminatingWrite::terminate_ref) directly
///
/// The point is that while the type is public, it cannot be built by anyone
/// outside of this module.
pub struct AntiCallToken(());
/// Trait used to indicate when no more write need to be done on a writer
pub trait TerminatingWrite: Write + Send {
pub trait TerminatingWrite: Write + Send + Sync {
/// Indicate that the writer will no longer be used. Internally call terminate_ref.
fn terminate(mut self) -> io::Result<()>
where Self: Sized {

BIN
doc/assets/images/etsy.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 85 KiB

2
doc/src/SUMMARY.md
View File

@@ -1,7 +1,5 @@
# Summary
[Avant Propos](./avant-propos.md)
- [Segments](./basis.md)

4
doc/src/avant-propos.md
View File

@@ -3,7 +3,7 @@
> Tantivy is a **search** engine **library** for Rust.
If you are familiar with Lucene, it's an excellent approximation to consider tantivy as Lucene for rust. tantivy is heavily inspired by Lucene's design and
they both have the same scope and targetted use cases.
they both have the same scope and targeted use cases.
If you are not familiar with Lucene, let's break down our little tagline.
@@ -31,4 +31,4 @@ relevancy, collapsing, highlighting, spatial search.
index from a different format.
Tantivy exposes a lot of low level API to do all of these things.

17
doc/src/basis.md
View File

@@ -11,7 +11,7 @@ directory shipped with tantivy is the `MmapDirectory`.
While this design has some downsides, this greatly simplifies the source code of
tantivy. Caching is also entirely delegated to the OS.
`tantivy` works entirely (or almost) by directly reading the datastructures as they are layed on disk. As a result, the act of opening an indexing does not involve loading different datastructures from the disk into random access memory : starting a process, opening an index, and performing your first query can typically be done in a matter of milliseconds.
`tantivy` works entirely (or almost) by directly reading the datastructures as they are laid on disk. As a result, the act of opening an indexing does not involve loading different datastructures from the disk into random access memory : starting a process, opening an index, and performing your first query can typically be done in a matter of milliseconds.
This is an interesting property for a command line search engine, or for some multi-tenant log search engine : spawning a new process for each new query can be a perfectly sensible solution in some use case.
@@ -22,7 +22,6 @@ Of course this is crucial to reduce IO, and ensure that as much of our index can
Also, whenever possible its data is accessed sequentially. Of course, this is an amazing property when tantivy needs to access the data from your spinning hard disk, but this is also
critical for performance, if your data is read from and an `SSD` or even already in your pagecache.
## Segments, and the log method
That kind of compact layout comes at one cost: it prevents our datastructures from being dynamic.
@@ -51,13 +50,9 @@ to get tantivy to fit your use case:
*Example 1* You could for instance use hadoop to build a very large search index in a timely manner, copy all of the resulting segment files in the same directory and edit the `meta.json` to get a functional index.[^2]
*Example 2* You could also disable your merge policy and enforce daily segments. Removing data after one week can then be done very efficiently by just editing the `meta.json` and deleting the files associated to segment `D-7`.
*Example 2* You could also disable your merge policy and enforce daily segments. Removing data after one week can then be done very efficiently by just editing the `meta.json` and deleting the files associated with segment `D-7`.
# Merging
## Merging
As you index more and more data, your index will accumulate more and more segments.
Having a lot of small segments is not really optimal. There is a bit of redundancy in having
@@ -66,11 +61,7 @@ all these term dictionary. Also when searching, we will need to do term lookups
That's where merging or compacting comes into place. Tantivy will continuously consider merge
opportunities and start merging segments in the background.
# Indexing throughput, number of indexing threads
## Indexing throughput, number of indexing threads
[^1]: This may eventually change.

2
doc/src/examples.md
View File

@@ -1,3 +1,3 @@
# Examples
- [Basic search](/examples/basic_search.html)
- [Basic search](/examples/basic_search.html)

31
doc/src/index_sorting.md
View File

@@ -1,11 +1,11 @@
- [Index Sorting](#index-sorting)
+ [Why Sorting](#why-sorting)
* [Compression](#compression)
* [Top-N Optimization](#top-n-optimization)
* [Pruning](#pruning)
* [Other](#other)
+ [Usage](#usage)
- [Why Sorting](#why-sorting)
- [Compression](#compression)
- [Top-N Optimization](#top-n-optimization)
- [Pruning](#pruning)
- [Other](#other)
- [Usage](#usage)
# Index Sorting
@@ -15,32 +15,34 @@ Tantivy allows you to sort the index according to a property.
Presorting an index has several advantages:
###### Compression
### Compression
When data is sorted it is easier to compress the data. E.g. the numbers sequence [5, 2, 3, 1, 4] would be sorted to [1, 2, 3, 4, 5].
When data is sorted it is easier to compress the data. E.g. the numbers sequence [5, 2, 3, 1, 4] would be sorted to [1, 2, 3, 4, 5].
If we apply delta encoding this list would be unsorted [5, -3, 1, -2, 3] vs. [1, 1, 1, 1, 1].
Compression ratio is mainly affected on the fast field of the sorted property, every thing else is likely unaffected.
###### Top-N Optimization
Compression ratio is mainly affected on the fast field of the sorted property, every thing else is likely unaffected.
When data is presorted by a field and search queries request sorting by the same field, we can leverage the natural order of the documents.
### Top-N Optimization
When data is presorted by a field and search queries request sorting by the same field, we can leverage the natural order of the documents.
E.g. if the data is sorted by timestamp and want the top n newest docs containing a term, we can simply leveraging the order of the docids.
Note: Tantivy 0.16 does not do this optimization yet.
###### Pruning
### Pruning
Let's say we want all documents and want to apply the filter `>= 2010-08-11`. When the data is sorted, we could make a lookup in the fast field to find the docid range and use this as the filter.
Note: Tantivy 0.16 does not do this optimization yet.
###### Other?
### Other?
In principle there are many algorithms possible that exploit the monotonically increasing nature. (aggregations maybe?)
## Usage
The index sorting can be configured setting [`sort_by_field`](https://github.com/quickwit-oss/tantivy/blob/000d76b11a139a84b16b9b95060a1c93e8b9851c/src/core/index_meta.rs#L238) on `IndexSettings` and passing it to a `IndexBuilder`. As of Tantivy 0.16 only fast fields are allowed to be used.
```
```rust
let settings = IndexSettings {
sort_by_field: Some(IndexSortByField {
field: "intval".to_string(),
@@ -58,4 +60,3 @@ let index = index_builder.create_in_ram().unwrap();
Sorting an index is applied in the serialization step. In general there are two serialization steps: [Finishing a single segment](https://github.com/quickwit-oss/tantivy/blob/000d76b11a139a84b16b9b95060a1c93e8b9851c/src/indexer/segment_writer.rs#L338) and [merging multiple segments](https://github.com/quickwit-oss/tantivy/blob/000d76b11a139a84b16b9b95060a1c93e8b9851c/src/indexer/merger.rs#L1073).
In both cases we generate a docid mapping reflecting the sort. This mapping is used when serializing the different components (doc store, fastfields, posting list, normfield, facets).

42
doc/src/json.md
View File

@@ -21,16 +21,17 @@ For instance, if user is a json field, the following document:
```
emits the following tokens:
- ("name", Text, "Paul")
- ("name", Text, "Masurel")
- ("address.city", Text, "Tokyo")
- ("address.country", Text, "Japan")
- ("created_at", Date, 15420648505)
- ("name", Text, "Paul")
- ("name", Text, "Masurel")
- ("address.city", Text, "Tokyo")
- ("address.country", Text, "Japan")
- ("created_at", Date, 15420648505)
# Bytes-encoding and lexicographical sort.
## Bytes-encoding and lexicographical sort
Like any other terms, these triplets are encoded into a binary format as follows.
- `json_path`: the json path is a sequence of "segments". In the example above, `address.city`
is just a debug representation of the json path `["address", "city"]`.
Its representation is done by separating segments by a unicode char `\x01`, and ending the path by `\x00`.
@@ -41,16 +42,16 @@ This representation is designed to align the natural sort of Terms with the lexi
of their binary representation (Tantivy's dictionary (whether fst or sstable) is sorted and does prefix encoding).
In the example above, the terms will be sorted as
- ("address.city", Text, "Tokyo")
- ("address.country", Text, "Japan")
- ("name", Text, "Masurel")
- ("name", Text, "Paul")
- ("created_at", Date, 15420648505)
- ("address.city", Text, "Tokyo")
- ("address.country", Text, "Japan")
- ("name", Text, "Masurel")
- ("name", Text, "Paul")
- ("created_at", Date, 15420648505)
As seen in "pitfalls", we may end up having to search for a value for a same path in several different fields. Putting the field code after the path makes it maximizes compression opportunities but also increases the chances for the two terms to end up in the actual same term dictionary block.
# Pitfalls, limitation and corner cases.
## Pitfalls, limitation and corner cases
Json gives very little information about the type of the literals it stores.
All numeric types end up mapped as a "Number" and there are no types for dates.
@@ -70,19 +71,21 @@ For instance, we do not even know if the type is a number or string based.
So the query
```
```rust
my_path.my_segment:233
```
Will be interpreted as
`(my_path.my_segment, String, 233) or (my_path.my_segment, u64, 233)`
```rust
(my_path.my_segment, String, 233) or (my_path.my_segment, u64, 233)
```
Likewise, we need to emit two tokens if the query contains an rfc3999 date.
Indeed the date could have been actually a single token inside the text of a document at ingestion time. Generally speaking, we will always at least emit a string token in query parsing, and sometimes more.
If one more json field is defined, things get even more complicated.
## Default json field
If the schema contains a text field called "text" and a json field that is set as a default field:
@@ -96,11 +99,11 @@ This is a product decision.
The user can still target the JSON field by specifying its name explicitly:
`json_dynamic.text:hello`.
## Range queries are not supported.
## Range queries are not supported
Json field do not support range queries.
## Arrays do not work like nested object.
## Arrays do not work like nested object
If json object contains an array, a search query might return more documents
than what might be expected.
@@ -120,9 +123,8 @@ Let's take an example.
Despite the array structure, a document in tantivy is a bag of terms.
The query:
```
```rust
cart.product_type:sneakers AND cart.attributes.color:red
```
Actually match the document above.

2
examples/aggregation.rs
View File

@@ -118,7 +118,7 @@ fn main() -> tantivy::Result<()> {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults = searcher.search(&term_query, &collector).unwrap();

8
examples/custom_collector.rs
View File

@@ -7,10 +7,12 @@
// Of course, you can have a look at the tantivy's built-in collectors
// such as the `CountCollector` for more examples.
use std::sync::Arc;
use fastfield_codecs::Column;
// ---
// Importing tantivy...
use tantivy::collector::{Collector, SegmentCollector};
use tantivy::fastfield::{DynamicFastFieldReader, FastFieldReader};
use tantivy::query::QueryParser;
use tantivy::schema::{Field, Schema, FAST, INDEXED, TEXT};
use tantivy::{doc, Index, Score, SegmentReader};
@@ -95,7 +97,7 @@ impl Collector for StatsCollector {
}
struct StatsSegmentCollector {
fast_field_reader: DynamicFastFieldReader<u64>,
fast_field_reader: Arc<dyn Column<u64>>,
stats: Stats,
}
@@ -103,7 +105,7 @@ impl SegmentCollector for StatsSegmentCollector {
type Fruit = Option<Stats>;
fn collect(&mut self, doc: u32, _score: Score) {
let value = self.fast_field_reader.get(doc) as f64;
let value = self.fast_field_reader.get_val(doc) as f64;
self.stats.count += 1;
self.stats.sum += value;
self.stats.squared_sum += value * value;

5
examples/custom_tokenizer.rs
View File

@@ -36,8 +36,7 @@ fn main() -> tantivy::Result<()> {
// need to be able to be able to retrieve it
// for our application.
//
// We can make our index lighter and
// by omitting `STORED` flag.
// We can make our index lighter by omitting the `STORED` flag.
let body = schema_builder.add_text_field("body", TEXT);
let schema = schema_builder.build();
@@ -50,7 +49,7 @@ fn main() -> tantivy::Result<()> {
// for your unit tests... Or this example.
let index = Index::create_in_ram(schema.clone());
// here we are registering our custome tokenizer
// here we are registering our custom tokenizer
// this will store tokens of 3 characters each
index
.tokenizers()

2
examples/deleting_updating_documents.rs
View File

@@ -113,7 +113,7 @@ fn main() -> tantivy::Result<()> {
// on its id.
//
// Note that `tantivy` does nothing to enforce the idea that
// there is only one document associated to this id.
// there is only one document associated with this id.
//
// Also you might have noticed that we apply the delete before
// having committed. This does not matter really...

4
examples/iterating_docs_and_positions.rs
View File

@@ -44,7 +44,7 @@ fn main() -> tantivy::Result<()> {
// A segment contains different data structure.
// Inverted index stands for the combination of
// - the term dictionary
// - the inverted lists associated to each terms and their positions
// - the inverted lists associated with each terms and their positions
let inverted_index = segment_reader.inverted_index(title)?;
// A `Term` is a text token associated with a field.
@@ -105,7 +105,7 @@ fn main() -> tantivy::Result<()> {
// A segment contains different data structure.
// Inverted index stands for the combination of
// - the term dictionary
// - the inverted lists associated to each terms and their positions
// - the inverted lists associated with each terms and their positions
let inverted_index = segment_reader.inverted_index(title)?;
// This segment posting object is like a cursor over the documents matching the term.

3
examples/warmer.rs
View File

@@ -3,7 +3,6 @@ use std::collections::{HashMap, HashSet};
use std::sync::{Arc, RwLock, Weak};
use tantivy::collector::TopDocs;
use tantivy::fastfield::FastFieldReader;
use tantivy::query::QueryParser;
use tantivy::schema::{Field, Schema, FAST, TEXT};
use tantivy::{
@@ -52,7 +51,7 @@ impl Warmer for DynamicPriceColumn {
let product_id_reader = segment.fast_fields().u64(self.field)?;
let product_ids: Vec<ProductId> = segment
.doc_ids_alive()
.map(|doc| product_id_reader.get(doc))
.map(|doc| product_id_reader.get_val(doc))
.collect();
let mut prices_it = self.price_fetcher.fetch_prices(&product_ids).into_iter();
let mut price_vals: Vec<Price> = Vec::new();

21
fastfield_codecs/Cargo.toml
View File

@@ -1,24 +1,35 @@
[package]
name = "fastfield_codecs"
version = "0.2.0"
version = "0.3.0"
authors = ["Pascal Seitz <pascal@quickwit.io>"]
license = "MIT"
edition = "2021"
description = "Fast field codecs used by tantivy"
documentation = "https://docs.rs/fastfield_codecs/"
homepage = "https://github.com/quickwit-oss/tantivy"
repository = "https://github.com/quickwit-oss/tantivy"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]
common = { version = "0.3", path = "../common/", package = "tantivy-common" }
tantivy-bitpacker = { version="0.2", path = "../bitpacker/" }
prettytable-rs = {version="0.8.0", optional= true}
common = { version = "0.5", path = "../common/", package = "tantivy-common" }
tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
ownedbytes = { version = "0.5", path = "../ownedbytes" }
prettytable-rs = {version="0.9.0", optional= true}
rand = {version="0.8.3", optional= true}
fastdivide = "0.4"
log = "0.4"
itertools = { version = "0.10.3" }
measure_time = { version="0.8.2", optional=true}
ordered-float = "3.4"
[dev-dependencies]
more-asserts = "0.3.0"
proptest = "1.0.0"
rand = "0.8.3"
[features]
bin = ["prettytable-rs", "rand"]
bin = ["prettytable-rs", "rand", "measure_time"]
default = ["bin"]
unstable = []

292
fastfield_codecs/benches/bench.rs
View File

@@ -4,105 +4,243 @@ extern crate test;
#[cfg(test)]
mod tests {
use fastfield_codecs::bitpacked::{BitpackedFastFieldReader, BitpackedFastFieldSerializer};
use fastfield_codecs::linearinterpol::{
LinearInterpolFastFieldReader, LinearInterpolFastFieldSerializer,
};
use fastfield_codecs::multilinearinterpol::{
MultiLinearInterpolFastFieldReader, MultiLinearInterpolFastFieldSerializer,
};
use fastfield_codecs::*;
use std::iter;
use std::sync::Arc;
fn get_data() -> Vec<u64> {
let mut data: Vec<_> = (100..55000_u64)
.map(|num| num + rand::random::<u8>() as u64)
use fastfield_codecs::*;
use ownedbytes::OwnedBytes;
use rand::prelude::*;
use test::Bencher;
use super::*;
// Warning: this generates the same permutation at each call
fn generate_permutation() -> Vec<u64> {
let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
fn generate_random() -> Vec<u64> {
let mut permutation: Vec<u64> = (0u64..100_000u64)
.map(|el| el + random::<u16>() as u64)
.collect();
data.push(99_000);
data.insert(1000, 2000);
data.insert(2000, 100);
data.insert(3000, 4100);
data.insert(4000, 100);
data.insert(5000, 800);
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
// Warning: this generates the same permutation at each call
fn generate_permutation_gcd() -> Vec<u64> {
let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
column: &[T],
) -> Arc<dyn Column<T>> {
let mut buffer = Vec::new();
serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
open(OwnedBytes::new(buffer)).unwrap()
}
#[bench]
fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = permutation[a as usize];
}
a
});
}
#[bench]
fn bench_intfastfield_jumpy_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = column.get_val(a as u32);
}
a
});
}
fn get_exp_data() -> Vec<u64> {
let mut data = vec![];
for i in 0..100 {
let num = i * i;
data.extend(iter::repeat(i as u64).take(num));
}
data.shuffle(&mut StdRng::from_seed([1u8; 32]));
// lengt = 328350
data
}
fn value_iter() -> impl Iterator<Item = u64> {
0..20_000
fn get_data_50percent_item() -> (u128, u128, Vec<u128>) {
let mut permutation = get_exp_data();
let major_item = 20;
let minor_item = 10;
permutation.extend(iter::repeat(major_item).take(permutation.len()));
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
(major_item as u128, minor_item as u128, permutation)
}
fn bench_get<S: FastFieldCodecSerializer, R: FastFieldCodecReader>(
b: &mut Bencher,
data: &[u64],
) {
let mut bytes = vec![];
S::serialize(
&mut bytes,
&data,
stats_from_vec(data),
data.iter().cloned(),
data.iter().cloned(),
)
.unwrap();
let reader = R::open_from_bytes(&bytes).unwrap();
b.iter(|| {
for pos in value_iter() {
reader.get_u64(pos as u64, &bytes);
}
});
fn get_u128_column_random() -> Arc<dyn Column<u128>> {
let permutation = generate_random();
let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
get_u128_column_from_data(&permutation)
}
fn bench_create<S: FastFieldCodecSerializer>(b: &mut Bencher, data: &[u64]) {
let mut bytes = vec![];
fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
let mut out = vec![];
let iter_gen = || data.iter().cloned();
serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
let out = OwnedBytes::new(out);
open_u128::<u128>(out).unwrap()
}
#[bench]
fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
let (major_item, _minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
S::serialize(
&mut bytes,
&data,
stats_from_vec(data),
data.iter().cloned(),
data.iter().cloned(),
)
.unwrap();
let mut positions = Vec::new();
column.get_docids_for_value_range(
major_item..=major_item,
0..data.len() as u32,
&mut positions,
);
positions
});
}
use test::Bencher;
#[bench]
fn bench_fastfield_bitpack_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<BitpackedFastFieldSerializer>(b, &data);
fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
let (_major_item, minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
minor_item..=minor_item,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_fastfield_linearinterpol_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<LinearInterpolFastFieldSerializer>(b, &data);
fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
let (_major_item, _minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
positions
});
}
#[bench]
fn bench_fastfield_multilinearinterpol_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<MultiLinearInterpolFastFieldSerializer>(b, &data);
fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
let column = get_u128_column_random();
b.iter(|| {
let mut a = 0u128;
for i in 0u64..column.num_vals() as u64 {
a += column.get_val(i as u32);
}
a
});
}
#[bench]
fn bench_fastfield_bitpack_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<BitpackedFastFieldSerializer, BitpackedFastFieldReader>(b, &data);
fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
let column = get_u128_column_random();
b.iter(|| {
let n = column.num_vals();
let mut a = 0u128;
for i in (0..n / 5).map(|val| val * 5) {
a += column.get_val(i);
}
a
});
}
#[bench]
fn bench_fastfield_linearinterpol_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<LinearInterpolFastFieldSerializer, LinearInterpolFastFieldReader>(b, &data);
fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
b.iter(|| {
let mut a = 0u64;
for i in (0..n / 7).map(|val| val * 7) {
a += permutation[i as usize];
}
a
});
}
#[bench]
fn bench_fastfield_multilinearinterpol_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<MultiLinearInterpolFastFieldSerializer, MultiLinearInterpolFastFieldReader>(
b, &data,
);
fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0;
for i in (0..n / 7).map(|val| val * 7) {
a += column.get_val(i as u32);
}
a
});
}
pub fn stats_from_vec(data: &[u64]) -> FastFieldStats {
let min_value = data.iter().cloned().min().unwrap_or(0);
let max_value = data.iter().cloned().max().unwrap_or(0);
FastFieldStats {
min_value,
max_value,
num_vals: data.len() as u64,
}
#[bench]
fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0u32..n as u32 {
a += column.get_val(i);
}
a
});
}
#[bench]
fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
let permutation = generate_permutation_gcd();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0..n {
a += column.get_val(i as u32);
}
a
});
}
#[bench]
fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
let permutation = generate_permutation();
b.iter(|| {
let mut a = 0u64;
for i in 0..permutation.len() {
a += permutation[i as usize] as u64;
}
a
});
}
}

162
fastfield_codecs/src/bitpacked.rs
View File

@@ -1,155 +1,99 @@
use std::io::{self, Write};
use common::BinarySerializable;
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::{FastFieldCodecReader, FastFieldCodecSerializer, FastFieldDataAccess, FastFieldStats};
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType};
/// Depending on the field type, a different
/// fast field is required.
#[derive(Clone)]
pub struct BitpackedFastFieldReader {
pub struct BitpackedReader {
data: OwnedBytes,
bit_unpacker: BitUnpacker,
pub min_value_u64: u64,
pub max_value_u64: u64,
normalized_header: NormalizedHeader,
}
impl FastFieldCodecReader for BitpackedFastFieldReader {
/// Opens a fast field given a file.
fn open_from_bytes(bytes: &[u8]) -> io::Result<Self> {
let (_data, mut footer) = bytes.split_at(bytes.len() - 16);
let min_value = u64::deserialize(&mut footer)?;
let amplitude = u64::deserialize(&mut footer)?;
let max_value = min_value + amplitude;
let num_bits = compute_num_bits(amplitude);
let bit_unpacker = BitUnpacker::new(num_bits);
Ok(BitpackedFastFieldReader {
min_value_u64: min_value,
max_value_u64: max_value,
bit_unpacker,
})
}
impl Column for BitpackedReader {
#[inline]
fn get_u64(&self, doc: u64, data: &[u8]) -> u64 {
self.min_value_u64 + self.bit_unpacker.get(doc, data)
fn get_val(&self, doc: u32) -> u64 {
self.bit_unpacker.get(doc, &self.data)
}
#[inline]
fn min_value(&self) -> u64 {
self.min_value_u64
// The BitpackedReader assumes a normalized vector.
0
}
#[inline]
fn max_value(&self) -> u64 {
self.max_value_u64
self.normalized_header.max_value
}
#[inline]
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}
pub struct BitpackedFastFieldSerializerLegacy<'a, W: 'a + Write> {
bit_packer: BitPacker,
write: &'a mut W,
min_value: u64,
amplitude: u64,
num_bits: u8,
}
impl<'a, W: Write> BitpackedFastFieldSerializerLegacy<'a, W> {
/// Creates a new fast field serializer.
///
/// The serializer in fact encode the values by bitpacking
/// `(val - min_value)`.
///
/// It requires a `min_value` and a `max_value` to compute
/// compute the minimum number of bits required to encode
/// values.
pub fn open(
write: &'a mut W,
min_value: u64,
max_value: u64,
) -> io::Result<BitpackedFastFieldSerializerLegacy<'a, W>> {
assert!(min_value <= max_value);
let amplitude = max_value - min_value;
let num_bits = compute_num_bits(amplitude);
let bit_packer = BitPacker::new();
Ok(BitpackedFastFieldSerializerLegacy {
bit_packer,
write,
min_value,
amplitude,
num_bits,
pub struct BitpackedCodec;
impl FastFieldCodec for BitpackedCodec {
/// The CODEC_TYPE is an enum value used for serialization.
const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::Bitpacked;
type Reader = BitpackedReader;
/// Opens a fast field given a file.
fn open_from_bytes(
data: OwnedBytes,
normalized_header: NormalizedHeader,
) -> io::Result<Self::Reader> {
let num_bits = compute_num_bits(normalized_header.max_value);
let bit_unpacker = BitUnpacker::new(num_bits);
Ok(BitpackedReader {
data,
bit_unpacker,
normalized_header,
})
}
/// Pushes a new value to the currently open u64 fast field.
#[inline]
pub fn add_val(&mut self, val: u64) -> io::Result<()> {
let val_to_write: u64 = val - self.min_value;
self.bit_packer
.write(val_to_write, self.num_bits, &mut self.write)?;
Ok(())
}
pub fn close_field(mut self) -> io::Result<()> {
self.bit_packer.close(&mut self.write)?;
self.min_value.serialize(&mut self.write)?;
self.amplitude.serialize(&mut self.write)?;
Ok(())
}
}
pub struct BitpackedFastFieldSerializer {}
impl FastFieldCodecSerializer for BitpackedFastFieldSerializer {
const NAME: &'static str = "Bitpacked";
const ID: u8 = 1;
/// Serializes data with the BitpackedFastFieldSerializer.
///
/// The serializer in fact encode the values by bitpacking
/// `(val - min_value)`.
/// The bitpacker assumes that the column has been normalized.
/// i.e. It has already been shifted by its minimum value, so that its
/// current minimum value is 0.
///
/// It requires a `min_value` and a `max_value` to compute
/// compute the minimum number of bits required to encode
/// values.
fn serialize(
write: &mut impl Write,
_fastfield_accessor: &dyn FastFieldDataAccess,
stats: FastFieldStats,
data_iter: impl Iterator<Item = u64>,
_data_iter1: impl Iterator<Item = u64>,
) -> io::Result<()> {
let mut serializer =
BitpackedFastFieldSerializerLegacy::open(write, stats.min_value, stats.max_value)?;
for val in data_iter {
serializer.add_val(val)?;
/// Ideally, we made a shift upstream on the column so that `col.min_value() == 0`.
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()> {
assert_eq!(column.min_value(), 0u64);
let num_bits = compute_num_bits(column.max_value());
let mut bit_packer = BitPacker::new();
for val in column.iter() {
bit_packer.write(val, num_bits, write)?;
}
serializer.close_field()?;
bit_packer.close(write)?;
Ok(())
}
fn is_applicable(
_fastfield_accessor: &impl FastFieldDataAccess,
_stats: FastFieldStats,
) -> bool {
true
}
fn estimate(_fastfield_accessor: &impl FastFieldDataAccess, stats: FastFieldStats) -> f32 {
let amplitude = stats.max_value - stats.min_value;
let num_bits = compute_num_bits(amplitude);
fn estimate(column: &dyn Column) -> Option<f32> {
let num_bits = compute_num_bits(column.max_value());
let num_bits_uncompressed = 64;
num_bits as f32 / num_bits_uncompressed as f32
Some(num_bits as f32 / num_bits_uncompressed as f32)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::tests::get_codec_test_data_sets;
use crate::tests::get_codec_test_datasets;
fn create_and_validate(data: &[u64], name: &str) {
crate::tests::create_and_validate::<BitpackedFastFieldSerializer, BitpackedFastFieldReader>(
data, name,
);
crate::tests::create_and_validate::<BitpackedCodec>(data, name);
}
#[test]
fn test_with_codec_data_sets() {
let data_sets = get_codec_test_data_sets();
let data_sets = get_codec_test_datasets();
for (mut data, name) in data_sets {
create_and_validate(&data, name);
data.reverse();

186
fastfield_codecs/src/blockwise_linear.rs Normal file
View File

@@ -0,0 +1,186 @@
use std::sync::Arc;
use std::{io, iter};
use common::{BinarySerializable, CountingWriter, DeserializeFrom};
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::line::Line;
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType, VecColumn};
const CHUNK_SIZE: usize = 512;
#[derive(Debug, Default)]
struct Block {
line: Line,
bit_unpacker: BitUnpacker,
data_start_offset: usize,
}
impl BinarySerializable for Block {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
self.line.serialize(writer)?;
self.bit_unpacker.bit_width().serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let line = Line::deserialize(reader)?;
let bit_width = u8::deserialize(reader)?;
Ok(Block {
line,
bit_unpacker: BitUnpacker::new(bit_width),
data_start_offset: 0,
})
}
}
fn compute_num_blocks(num_vals: u32) -> usize {
(num_vals as usize + CHUNK_SIZE - 1) / CHUNK_SIZE
}
pub struct BlockwiseLinearCodec;
impl FastFieldCodec for BlockwiseLinearCodec {
const CODEC_TYPE: crate::FastFieldCodecType = FastFieldCodecType::BlockwiseLinear;
type Reader = BlockwiseLinearReader;
fn open_from_bytes(
bytes: ownedbytes::OwnedBytes,
normalized_header: NormalizedHeader,
) -> io::Result<Self::Reader> {
let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
let footer_offset = bytes.len() - 4 - footer_len as usize;
let (data, mut footer) = bytes.split(footer_offset);
let num_blocks = compute_num_blocks(normalized_header.num_vals);
let mut blocks: Vec<Block> = iter::repeat_with(|| Block::deserialize(&mut footer))
.take(num_blocks)
.collect::<io::Result<_>>()?;
let mut start_offset = 0;
for block in &mut blocks {
block.data_start_offset = start_offset;
start_offset += (block.bit_unpacker.bit_width() as usize) * CHUNK_SIZE / 8;
}
Ok(BlockwiseLinearReader {
blocks: Arc::new(blocks),
data,
normalized_header,
})
}
// Estimate first_chunk and extrapolate
fn estimate(column: &dyn crate::Column) -> Option<f32> {
if column.num_vals() < 10 * CHUNK_SIZE as u32 {
return None;
}
let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE).collect();
let line = Line::train(&VecColumn::from(&first_chunk));
for (i, buffer_val) in first_chunk.iter_mut().enumerate() {
let interpolated_val = line.eval(i as u32);
*buffer_val = buffer_val.wrapping_sub(interpolated_val);
}
let estimated_bit_width = first_chunk
.iter()
.map(|el| ((el + 1) as f32 * 3.0) as u64)
.map(compute_num_bits)
.max()
.unwrap();
let metadata_per_block = {
let mut out = vec![];
Block::default().serialize(&mut out).unwrap();
out.len()
};
let num_bits = estimated_bit_width as u64 * column.num_vals() as u64
// function metadata per block
+ metadata_per_block as u64 * (column.num_vals() as u64 / CHUNK_SIZE as u64);
let num_bits_uncompressed = 64 * column.num_vals();
Some(num_bits as f32 / num_bits_uncompressed as f32)
}
fn serialize(column: &dyn Column, wrt: &mut impl io::Write) -> io::Result<()> {
// The BitpackedReader assumes a normalized vector.
assert_eq!(column.min_value(), 0);
let mut buffer = Vec::with_capacity(CHUNK_SIZE);
let num_vals = column.num_vals();
let num_blocks = compute_num_blocks(num_vals);
let mut blocks = Vec::with_capacity(num_blocks);
let mut vals = column.iter();
let mut bit_packer = BitPacker::new();
for _ in 0..num_blocks {
buffer.clear();
buffer.extend((&mut vals).take(CHUNK_SIZE));
let line = Line::train(&VecColumn::from(&buffer));
assert!(!buffer.is_empty());
for (i, buffer_val) in buffer.iter_mut().enumerate() {
let interpolated_val = line.eval(i as u32);
*buffer_val = buffer_val.wrapping_sub(interpolated_val);
}
let bit_width = buffer.iter().copied().map(compute_num_bits).max().unwrap();
for &buffer_val in &buffer {
bit_packer.write(buffer_val, bit_width, wrt)?;
}
blocks.push(Block {
line,
bit_unpacker: BitUnpacker::new(bit_width),
data_start_offset: 0,
});
}
bit_packer.close(wrt)?;
assert_eq!(blocks.len(), compute_num_blocks(num_vals));
let mut counting_wrt = CountingWriter::wrap(wrt);
for block in &blocks {
block.serialize(&mut counting_wrt)?;
}
let footer_len = counting_wrt.written_bytes();
(footer_len as u32).serialize(&mut counting_wrt)?;
Ok(())
}
}
#[derive(Clone)]
pub struct BlockwiseLinearReader {
blocks: Arc<Vec<Block>>,
normalized_header: NormalizedHeader,
data: OwnedBytes,
}
impl Column for BlockwiseLinearReader {
#[inline(always)]
fn get_val(&self, idx: u32) -> u64 {
let block_id = (idx / CHUNK_SIZE as u32) as usize;
let idx_within_block = idx % (CHUNK_SIZE as u32);
let block = &self.blocks[block_id];
let interpoled_val: u64 = block.line.eval(idx_within_block);
let block_bytes = &self.data[block.data_start_offset..];
let bitpacked_diff = block.bit_unpacker.get(idx_within_block, block_bytes);
interpoled_val.wrapping_add(bitpacked_diff)
}
fn min_value(&self) -> u64 {
// The BlockwiseLinearReader assumes a normalized vector.
0u64
}
fn max_value(&self) -> u64 {
self.normalized_header.max_value
}
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}

348
fastfield_codecs/src/column.rs Normal file
View File

@@ -0,0 +1,348 @@
use std::marker::PhantomData;
use std::ops::{Range, RangeInclusive};
use tantivy_bitpacker::minmax;
use crate::monotonic_mapping::StrictlyMonotonicFn;
/// `Column` provides columnar access on a field.
pub trait Column<T: PartialOrd = u64>: Send + Sync {
/// Return the value associated with the given idx.
///
/// This accessor should return as fast as possible.
///
/// # Panics
///
/// May panic if `idx` is greater than the column length.
fn get_val(&self, idx: u32) -> T;
/// Fills an output buffer with the fast field values
/// associated with the `DocId` going from
/// `start` to `start + output.len()`.
///
/// # Panics
///
/// Must panic if `start + output.len()` is greater than
/// the segment's `maxdoc`.
#[inline]
fn get_range(&self, start: u64, output: &mut [T]) {
for (out, idx) in output.iter_mut().zip(start..) {
*out = self.get_val(idx as u32);
}
}
/// Get the positions of values which are in the provided value range.
///
/// Note that position == docid for single value fast fields
#[inline]
fn get_docids_for_value_range(
&self,
value_range: RangeInclusive<T>,
doc_id_range: Range<u32>,
positions: &mut Vec<u32>,
) {
let doc_id_range = doc_id_range.start..doc_id_range.end.min(self.num_vals());
for idx in doc_id_range.start..doc_id_range.end {
let val = self.get_val(idx);
if value_range.contains(&val) {
positions.push(idx);
}
}
}
/// Returns the minimum value for this fast field.
///
/// This min_value may not be exact.
/// For instance, the min value does not take in account of possible
/// deleted document. All values are however guaranteed to be higher than
/// `.min_value()`.
fn min_value(&self) -> T;
/// Returns the maximum value for this fast field.
///
/// This max_value may not be exact.
/// For instance, the max value does not take in account of possible
/// deleted document. All values are however guaranteed to be higher than
/// `.max_value()`.
fn max_value(&self) -> T;
/// The number of values in the column.
fn num_vals(&self) -> u32;
/// Returns a iterator over the data
fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = T> + 'a> {
Box::new((0..self.num_vals()).map(|idx| self.get_val(idx)))
}
}
/// VecColumn provides `Column` over a slice.
pub struct VecColumn<'a, T = u64> {
values: &'a [T],
min_value: T,
max_value: T,
}
impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
fn get_val(&self, idx: u32) -> T {
(*self).get_val(idx)
}
fn min_value(&self) -> T {
(*self).min_value()
}
fn max_value(&self) -> T {
(*self).max_value()
}
fn num_vals(&self) -> u32 {
(*self).num_vals()
}
fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
(*self).iter()
}
fn get_range(&self, start: u64, output: &mut [T]) {
(*self).get_range(start, output)
}
}
impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
fn get_val(&self, position: u32) -> T {
self.values[position as usize]
}
fn iter(&self) -> Box<dyn Iterator<Item = T> + '_> {
Box::new(self.values.iter().copied())
}
fn min_value(&self) -> T {
self.min_value
}
fn max_value(&self) -> T {
self.max_value
}
fn num_vals(&self) -> u32 {
self.values.len() as u32
}
fn get_range(&self, start: u64, output: &mut [T]) {
output.copy_from_slice(&self.values[start as usize..][..output.len()])
}
}
impl<'a, T: Copy + Ord + Default, V> From<&'a V> for VecColumn<'a, T>
where V: AsRef<[T]> + ?Sized
{
fn from(values: &'a V) -> Self {
let values = values.as_ref();
let (min_value, max_value) = minmax(values.iter().copied()).unwrap_or_default();
Self {
values,
min_value,
max_value,
}
}
}
struct MonotonicMappingColumn<C, T, Input> {
from_column: C,
monotonic_mapping: T,
_phantom: PhantomData<Input>,
}
/// Creates a view of a column transformed by a strictly monotonic mapping. See
/// [`StrictlyMonotonicFn`].
///
/// E.g. apply a gcd monotonic_mapping([100, 200, 300]) == [1, 2, 3]
/// monotonic_mapping.mapping() is expected to be injective, and we should always have
/// monotonic_mapping.inverse(monotonic_mapping.mapping(el)) == el
///
/// The inverse of the mapping is required for:
/// `fn get_positions_for_value_range(&self, range: RangeInclusive<T>) -> Vec<u64> `
/// The user provides the original value range and we need to monotonic map them in the same way the
/// serialization does before calling the underlying column.
///
/// Note that when opening a codec, the monotonic_mapping should be the inverse of the mapping
/// during serialization. And therefore the monotonic_mapping_inv when opening is the same as
/// monotonic_mapping during serialization.
pub fn monotonic_map_column<C, T, Input, Output>(
from_column: C,
monotonic_mapping: T,
) -> impl Column<Output>
where
C: Column<Input>,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Send + Sync + Clone,
Output: PartialOrd + Send + Sync + Clone,
{
MonotonicMappingColumn {
from_column,
monotonic_mapping,
_phantom: PhantomData,
}
}
impl<C, T, Input, Output> Column<Output> for MonotonicMappingColumn<C, T, Input>
where
C: Column<Input>,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Send + Sync + Clone,
Output: PartialOrd + Send + Sync + Clone,
{
#[inline]
fn get_val(&self, idx: u32) -> Output {
let from_val = self.from_column.get_val(idx);
self.monotonic_mapping.mapping(from_val)
}
fn min_value(&self) -> Output {
let from_min_value = self.from_column.min_value();
self.monotonic_mapping.mapping(from_min_value)
}
fn max_value(&self) -> Output {
let from_max_value = self.from_column.max_value();
self.monotonic_mapping.mapping(from_max_value)
}
fn num_vals(&self) -> u32 {
self.from_column.num_vals()
}
fn iter(&self) -> Box<dyn Iterator<Item = Output> + '_> {
Box::new(
self.from_column
.iter()
.map(|el| self.monotonic_mapping.mapping(el)),
)
}
fn get_docids_for_value_range(
&self,
range: RangeInclusive<Output>,
doc_id_range: Range<u32>,
positions: &mut Vec<u32>,
) {
self.from_column.get_docids_for_value_range(
self.monotonic_mapping.inverse(range.start().clone())
..=self.monotonic_mapping.inverse(range.end().clone()),
doc_id_range,
positions,
)
}
// We voluntarily do not implement get_range as it yields a regression,
// and we do not have any specialized implementation anyway.
}
/// Wraps an iterator into a `Column`.
pub struct IterColumn<T>(T);
impl<T> From<T> for IterColumn<T>
where T: Iterator + Clone + ExactSizeIterator
{
fn from(iter: T) -> Self {
IterColumn(iter)
}
}
impl<T> Column<T::Item> for IterColumn<T>
where
T: Iterator + Clone + ExactSizeIterator + Send + Sync,
T::Item: PartialOrd,
{
fn get_val(&self, idx: u32) -> T::Item {
self.0.clone().nth(idx as usize).unwrap()
}
fn min_value(&self) -> T::Item {
self.0.clone().next().unwrap()
}
fn max_value(&self) -> T::Item {
self.0.clone().last().unwrap()
}
fn num_vals(&self) -> u32 {
self.0.len() as u32
}
fn iter(&self) -> Box<dyn Iterator<Item = T::Item> + '_> {
Box::new(self.0.clone())
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternalBaseval,
StrictlyMonotonicMappingToInternalGCDBaseval,
};
#[test]
fn test_monotonic_mapping() {
let vals = &[3u64, 5u64][..];
let col = VecColumn::from(vals);
let mapped = monotonic_map_column(col, StrictlyMonotonicMappingToInternalBaseval::new(2));
assert_eq!(mapped.min_value(), 1u64);
assert_eq!(mapped.max_value(), 3u64);
assert_eq!(mapped.num_vals(), 2);
assert_eq!(mapped.num_vals(), 2);
assert_eq!(mapped.get_val(0), 1);
assert_eq!(mapped.get_val(1), 3);
}
#[test]
fn test_range_as_col() {
let col = IterColumn::from(10..100);
assert_eq!(col.num_vals(), 90);
assert_eq!(col.max_value(), 99);
}
#[test]
fn test_monotonic_mapping_iter() {
let vals: Vec<u64> = (10..110u64).map(|el| el * 10).collect();
let col = VecColumn::from(&vals);
let mapped = monotonic_map_column(
col,
StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100),
),
);
let val_i64s: Vec<u64> = mapped.iter().collect();
for i in 0..100 {
assert_eq!(val_i64s[i as usize], mapped.get_val(i));
}
}
#[test]
fn test_monotonic_mapping_get_range() {
let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
let col = VecColumn::from(&vals);
let mapped = monotonic_map_column(
col,
StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 0),
),
);
assert_eq!(mapped.min_value(), 0u64);
assert_eq!(mapped.max_value(), 9900u64);
assert_eq!(mapped.num_vals(), 100);
let val_u64s: Vec<u64> = mapped.iter().collect();
assert_eq!(val_u64s.len(), 100);
for i in 0..100 {
assert_eq!(val_u64s[i as usize], mapped.get_val(i));
assert_eq!(val_u64s[i as usize], vals[i as usize] * 10);
}
let mut buf = [0u64; 20];
mapped.get_range(7, &mut buf[..]);
assert_eq!(&val_u64s[7..][..20], &buf);
}
}

43
fastfield_codecs/src/compact_space/blank_range.rs Normal file
View File

@@ -0,0 +1,43 @@
use std::ops::RangeInclusive;
/// The range of a blank in value space.
///
/// A blank is an unoccupied space in the data.
/// Use try_into() to construct.
/// A range has to have at least length of 3. Invalid ranges will be rejected.
///
/// Ordered by range length.
#[derive(Debug, Eq, PartialEq, Clone)]
pub(crate) struct BlankRange {
blank_range: RangeInclusive<u128>,
}
impl TryFrom<RangeInclusive<u128>> for BlankRange {
type Error = &'static str;
fn try_from(range: RangeInclusive<u128>) -> Result<Self, Self::Error> {
let blank_size = range.end().saturating_sub(*range.start());
if blank_size < 2 {
Err("invalid range")
} else {
Ok(BlankRange { blank_range: range })
}
}
}
impl BlankRange {
pub(crate) fn blank_size(&self) -> u128 {
self.blank_range.end() - self.blank_range.start() + 1
}
pub(crate) fn blank_range(&self) -> RangeInclusive<u128> {
self.blank_range.clone()
}
}
impl Ord for BlankRange {
fn cmp(&self, other: &Self) -> std::cmp::Ordering {
self.blank_size().cmp(&other.blank_size())
}
}
impl PartialOrd for BlankRange {
fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
Some(self.blank_size().cmp(&other.blank_size()))
}
}

231
fastfield_codecs/src/compact_space/build_compact_space.rs Normal file
View File

@@ -0,0 +1,231 @@
use std::collections::{BTreeSet, BinaryHeap};
use std::iter;
use std::ops::RangeInclusive;
use itertools::Itertools;
use super::blank_range::BlankRange;
use super::{CompactSpace, RangeMapping};
/// Put the blanks for the sorted values into a binary heap
fn get_blanks(values_sorted: &BTreeSet<u128>) -> BinaryHeap<BlankRange> {
let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
for (first, second) in values_sorted.iter().tuple_windows() {
// Correctness Overflow: the values are deduped and sorted (BTreeSet property), that means
// there's always space between two values.
let blank_range = first + 1..=second - 1;
let blank_range: Result<BlankRange, _> = blank_range.try_into();
if let Ok(blank_range) = blank_range {
blanks.push(blank_range);
}
}
blanks
}
struct BlankCollector {
blanks: Vec<BlankRange>,
staged_blanks_sum: u128,
}
impl BlankCollector {
fn new() -> Self {
Self {
blanks: vec![],
staged_blanks_sum: 0,
}
}
fn stage_blank(&mut self, blank: BlankRange) {
self.staged_blanks_sum += blank.blank_size();
self.blanks.push(blank);
}
fn drain(&mut self) -> impl Iterator<Item = BlankRange> + '_ {
self.staged_blanks_sum = 0;
self.blanks.drain(..)
}
fn staged_blanks_sum(&self) -> u128 {
self.staged_blanks_sum
}
fn num_staged_blanks(&self) -> usize {
self.blanks.len()
}
}
fn num_bits(val: u128) -> u8 {
(128u32 - val.leading_zeros()) as u8
}
/// Will collect blanks and add them to compact space if more bits are saved than cost from
/// metadata.
pub fn get_compact_space(
values_deduped_sorted: &BTreeSet<u128>,
total_num_values: u32,
cost_per_blank: usize,
) -> CompactSpace {
let mut compact_space_builder = CompactSpaceBuilder::new();
if values_deduped_sorted.is_empty() {
return compact_space_builder.finish();
}
let mut blanks: BinaryHeap<BlankRange> = get_blanks(values_deduped_sorted);
// Replace after stabilization of https://github.com/rust-lang/rust/issues/62924
// We start by space that's limited to min_value..=max_value
let min_value = *values_deduped_sorted.iter().next().unwrap_or(&0);
let max_value = *values_deduped_sorted.iter().last().unwrap_or(&0);
// +1 for null, in case min and max covers the whole space, we are off by one.
let mut amplitude_compact_space = (max_value - min_value).saturating_add(1);
if min_value != 0 {
compact_space_builder.add_blanks(iter::once(0..=min_value - 1));
}
if max_value != u128::MAX {
compact_space_builder.add_blanks(iter::once(max_value + 1..=u128::MAX));
}
let mut amplitude_bits: u8 = num_bits(amplitude_compact_space);
let mut blank_collector = BlankCollector::new();
// We will stage blanks until they reduce the compact space by at least 1 bit and then flush
// them if the metadata cost is lower than the total number of saved bits.
// Binary heap to process the gaps by their size
while let Some(blank_range) = blanks.pop() {
blank_collector.stage_blank(blank_range);
let staged_spaces_sum: u128 = blank_collector.staged_blanks_sum();
let amplitude_new_compact_space = amplitude_compact_space - staged_spaces_sum;
let amplitude_new_bits = num_bits(amplitude_new_compact_space);
if amplitude_bits == amplitude_new_bits {
continue;
}
let saved_bits = (amplitude_bits - amplitude_new_bits) as usize * total_num_values as usize;
// TODO: Maybe calculate exact cost of blanks and run this more expensive computation only,
// when amplitude_new_bits changes
let cost = blank_collector.num_staged_blanks() * cost_per_blank;
if cost >= saved_bits {
// Continue here, since although we walk over the blanks by size,
// we can potentially save a lot at the last bits, which are smaller blanks
//
// E.g. if the first range reduces the compact space by 1000 from 2000 to 1000, which
// saves 11-10=1 bit and the next range reduces the compact space by 950 to
// 50, which saves 10-6=4 bit
continue;
}
amplitude_compact_space = amplitude_new_compact_space;
amplitude_bits = amplitude_new_bits;
compact_space_builder.add_blanks(blank_collector.drain().map(|blank| blank.blank_range()));
}
// special case, when we don't collected any blanks because:
// * the data is empty (early exit)
// * the algorithm did decide it's not worth the cost, which can be the case for single values
//
// We drain one collected blank unconditionally, so the empty case is reserved for empty
// data, and therefore empty compact_space means the data is empty and no data is covered
// (conversely to all data) and we can assign null to it.
if compact_space_builder.is_empty() {
compact_space_builder.add_blanks(
blank_collector
.drain()
.map(|blank| blank.blank_range())
.take(1),
);
}
let compact_space = compact_space_builder.finish();
if max_value - min_value != u128::MAX {
debug_assert_eq!(
compact_space.amplitude_compact_space(),
amplitude_compact_space
);
}
compact_space
}
#[derive(Debug, Clone, Eq, PartialEq)]
struct CompactSpaceBuilder {
blanks: Vec<RangeInclusive<u128>>,
}
impl CompactSpaceBuilder {
/// Creates a new compact space builder which will initially cover the whole space.
fn new() -> Self {
Self { blanks: Vec::new() }
}
/// Assumes that repeated add_blank calls don't overlap and are not adjacent,
/// e.g. [3..=5, 5..=10] is not allowed
///
/// Both of those assumptions are true when blanks are produced from sorted values.
fn add_blanks(&mut self, blank: impl Iterator<Item = RangeInclusive<u128>>) {
self.blanks.extend(blank);
}
fn is_empty(&self) -> bool {
self.blanks.is_empty()
}
/// Convert blanks to covered space and assign null value
fn finish(mut self) -> CompactSpace {
// sort by start. ranges are not allowed to overlap
self.blanks.sort_unstable_by_key(|blank| *blank.start());
let mut covered_space = Vec::with_capacity(self.blanks.len());
// begining of the blanks
if let Some(first_blank_start) = self.blanks.first().map(RangeInclusive::start) {
if *first_blank_start != 0 {
covered_space.push(0..=first_blank_start - 1);
}
}
// Between the blanks
let between_blanks = self.blanks.iter().tuple_windows().map(|(left, right)| {
assert!(
left.end() < right.start(),
"overlapping or adjacent ranges detected"
);
*left.end() + 1..=*right.start() - 1
});
covered_space.extend(between_blanks);
// end of the blanks
if let Some(last_blank_end) = self.blanks.last().map(RangeInclusive::end) {
if *last_blank_end != u128::MAX {
covered_space.push(last_blank_end + 1..=u128::MAX);
}
}
if covered_space.is_empty() {
covered_space.push(0..=0); // empty data case
};
let mut compact_start: u64 = 1; // 0 is reserved for `null`
let mut ranges_mapping: Vec<RangeMapping> = Vec::with_capacity(covered_space.len());
for cov in covered_space {
let range_mapping = super::RangeMapping {
value_range: cov,
compact_start,
};
let covered_range_len = range_mapping.range_length();
ranges_mapping.push(range_mapping);
compact_start += covered_range_len;
}
// println!("num ranges {}", ranges_mapping.len());
CompactSpace { ranges_mapping }
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_binary_heap_pop_order() {
let mut blanks: BinaryHeap<BlankRange> = BinaryHeap::new();
blanks.push((0..=10).try_into().unwrap());
blanks.push((100..=200).try_into().unwrap());
blanks.push((100..=110).try_into().unwrap());
assert_eq!(blanks.pop().unwrap().blank_size(), 101);
assert_eq!(blanks.pop().unwrap().blank_size(), 11);
}
}

814
fastfield_codecs/src/compact_space/mod.rs Normal file
View File

@@ -0,0 +1,814 @@
/// This codec takes a large number space (u128) and reduces it to a compact number space.
///
/// It will find spaces in the number range. For example:
///
/// 100, 101, 102, 103, 104, 50000, 50001
/// could be mapped to
/// 100..104 -> 0..4
/// 50000..50001 -> 5..6
///
/// Compact space 0..=6 requires much less bits than 100..=50001
///
/// The codec is created to compress ip addresses, but may be employed in other use cases.
use std::{
cmp::Ordering,
collections::BTreeSet,
io::{self, Write},
ops::{Range, RangeInclusive},
};
use common::{BinarySerializable, CountingWriter, VInt, VIntU128};
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
use crate::compact_space::build_compact_space::get_compact_space;
use crate::Column;
mod blank_range;
mod build_compact_space;
/// The cost per blank is quite hard actually, since blanks are delta encoded, the actual cost of
/// blanks depends on the number of blanks.
///
/// The number is taken by looking at a real dataset. It is optimized for larger datasets.
const COST_PER_BLANK_IN_BITS: usize = 36;
#[derive(Debug, Clone, Eq, PartialEq)]
pub struct CompactSpace {
ranges_mapping: Vec<RangeMapping>,
}
/// Maps the range from the original space to compact_start + range.len()
#[derive(Debug, Clone, Eq, PartialEq)]
struct RangeMapping {
value_range: RangeInclusive<u128>,
compact_start: u64,
}
impl RangeMapping {
fn range_length(&self) -> u64 {
(self.value_range.end() - self.value_range.start()) as u64 + 1
}
// The last value of the compact space in this range
fn compact_end(&self) -> u64 {
self.compact_start + self.range_length() - 1
}
}
impl BinarySerializable for CompactSpace {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
VInt(self.ranges_mapping.len() as u64).serialize(writer)?;
let mut prev_value = 0;
for value_range in self
.ranges_mapping
.iter()
.map(|range_mapping| &range_mapping.value_range)
{
let blank_delta_start = value_range.start() - prev_value;
VIntU128(blank_delta_start).serialize(writer)?;
prev_value = *value_range.start();
let blank_delta_end = value_range.end() - prev_value;
VIntU128(blank_delta_end).serialize(writer)?;
prev_value = *value_range.end();
}
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let num_ranges = VInt::deserialize(reader)?.0;
let mut ranges_mapping: Vec<RangeMapping> = vec![];
let mut value = 0u128;
let mut compact_start = 1u64; // 0 is reserved for `null`
for _ in 0..num_ranges {
let blank_delta_start = VIntU128::deserialize(reader)?.0;
value += blank_delta_start;
let blank_start = value;
let blank_delta_end = VIntU128::deserialize(reader)?.0;
value += blank_delta_end;
let blank_end = value;
let range_mapping = RangeMapping {
value_range: blank_start..=blank_end,
compact_start,
};
let range_length = range_mapping.range_length();
ranges_mapping.push(range_mapping);
compact_start += range_length;
}
Ok(Self { ranges_mapping })
}
}
impl CompactSpace {
/// Amplitude is the value range of the compact space including the sentinel value used to
/// identify null values. The compact space is 0..=amplitude .
///
/// It's only used to verify we don't exceed u64 number space, which would indicate a bug.
fn amplitude_compact_space(&self) -> u128 {
self.ranges_mapping
.last()
.map(|last_range| last_range.compact_end() as u128)
.unwrap_or(1) // compact space starts at 1, 0 == null
}
fn get_range_mapping(&self, pos: usize) -> &RangeMapping {
&self.ranges_mapping[pos]
}
/// Returns either Ok(the value in the compact space) or if it is outside the compact space the
/// Err(position where it would be inserted)
fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
self.ranges_mapping
.binary_search_by(|probe| {
let value_range = &probe.value_range;
if value < *value_range.start() {
Ordering::Greater
} else if value > *value_range.end() {
Ordering::Less
} else {
Ordering::Equal
}
})
.map(|pos| {
let range_mapping = &self.ranges_mapping[pos];
let pos_in_range = (value - range_mapping.value_range.start()) as u64;
range_mapping.compact_start + pos_in_range
})
}
/// Unpacks a value from compact space u64 to u128 space
fn compact_to_u128(&self, compact: u64) -> u128 {
let pos = self
.ranges_mapping
.binary_search_by_key(&compact, |range_mapping| range_mapping.compact_start)
// Correctness: Overflow. The first range starts at compact space 0, the error from
// binary search can never be 0
.map_or_else(|e| e - 1, |v| v);
let range_mapping = &self.ranges_mapping[pos];
let diff = compact - range_mapping.compact_start;
range_mapping.value_range.start() + diff as u128
}
}
pub struct CompactSpaceCompressor {
params: IPCodecParams,
}
#[derive(Debug, Clone)]
pub struct IPCodecParams {
compact_space: CompactSpace,
bit_unpacker: BitUnpacker,
min_value: u128,
max_value: u128,
num_vals: u32,
num_bits: u8,
}
impl CompactSpaceCompressor {
/// Taking the vals as Vec may cost a lot of memory. It is used to sort the vals.
pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
let mut values_sorted = BTreeSet::new();
values_sorted.extend(iter);
let total_num_values = num_vals;
let compact_space =
get_compact_space(&values_sorted, total_num_values, COST_PER_BLANK_IN_BITS);
let amplitude_compact_space = compact_space.amplitude_compact_space();
assert!(
amplitude_compact_space <= u64::MAX as u128,
"case unsupported."
);
let num_bits = tantivy_bitpacker::compute_num_bits(amplitude_compact_space as u64);
let min_value = *values_sorted.iter().next().unwrap_or(&0);
let max_value = *values_sorted.iter().last().unwrap_or(&0);
assert_eq!(
compact_space
.u128_to_compact(max_value)
.expect("could not convert max value to compact space"),
amplitude_compact_space as u64
);
CompactSpaceCompressor {
params: IPCodecParams {
compact_space,
bit_unpacker: BitUnpacker::new(num_bits),
min_value,
max_value,
num_vals: total_num_values,
num_bits,
},
}
}
fn write_footer(self, writer: &mut impl Write) -> io::Result<()> {
let writer = &mut CountingWriter::wrap(writer);
self.params.serialize(writer)?;
let footer_len = writer.written_bytes() as u32;
footer_len.serialize(writer)?;
Ok(())
}
pub fn compress_into(
self,
vals: impl Iterator<Item = u128>,
write: &mut impl Write,
) -> io::Result<()> {
let mut bitpacker = BitPacker::default();
for val in vals {
let compact = self
.params
.compact_space
.u128_to_compact(val)
.map_err(|_| {
io::Error::new(
io::ErrorKind::InvalidData,
"Could not convert value to compact_space. This is a bug.",
)
})?;
bitpacker.write(compact, self.params.num_bits, write)?;
}
bitpacker.close(write)?;
self.write_footer(write)?;
Ok(())
}
}
#[derive(Debug, Clone)]
pub struct CompactSpaceDecompressor {
data: OwnedBytes,
params: IPCodecParams,
}
impl BinarySerializable for IPCodecParams {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
// header flags for future optional dictionary encoding
let footer_flags = 0u64;
footer_flags.serialize(writer)?;
VIntU128(self.min_value).serialize(writer)?;
VIntU128(self.max_value).serialize(writer)?;
VIntU128(self.num_vals as u128).serialize(writer)?;
self.num_bits.serialize(writer)?;
self.compact_space.serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let _header_flags = u64::deserialize(reader)?;
let min_value = VIntU128::deserialize(reader)?.0;
let max_value = VIntU128::deserialize(reader)?.0;
let num_vals = VIntU128::deserialize(reader)?.0 as u32;
let num_bits = u8::deserialize(reader)?;
let compact_space = CompactSpace::deserialize(reader)?;
Ok(Self {
compact_space,
bit_unpacker: BitUnpacker::new(num_bits),
min_value,
max_value,
num_vals,
num_bits,
})
}
}
impl Column<u128> for CompactSpaceDecompressor {
#[inline]
fn get_val(&self, doc: u32) -> u128 {
self.get(doc)
}
fn min_value(&self) -> u128 {
self.min_value()
}
fn max_value(&self) -> u128 {
self.max_value()
}
fn num_vals(&self) -> u32 {
self.params.num_vals
}
#[inline]
fn iter(&self) -> Box<dyn Iterator<Item = u128> + '_> {
Box::new(self.iter())
}
#[inline]
fn get_docids_for_value_range(
&self,
value_range: RangeInclusive<u128>,
positions_range: Range<u32>,
positions: &mut Vec<u32>,
) {
self.get_positions_for_value_range(value_range, positions_range, positions)
}
}
impl CompactSpaceDecompressor {
pub fn open(data: OwnedBytes) -> io::Result<CompactSpaceDecompressor> {
let (data_slice, footer_len_bytes) = data.split_at(data.len() - 4);
let footer_len = u32::deserialize(&mut &footer_len_bytes[..])?;
let data_footer = &data_slice[data_slice.len() - footer_len as usize..];
let params = IPCodecParams::deserialize(&mut &data_footer[..])?;
let decompressor = CompactSpaceDecompressor { data, params };
Ok(decompressor)
}
/// Converting to compact space for the decompressor is more complex, since we may get values
/// which are outside the compact space. e.g. if we map
/// 1000 => 5
/// 2000 => 6
///
/// and we want a mapping for 1005, there is no equivalent compact space. We instead return an
/// error with the index of the next range.
fn u128_to_compact(&self, value: u128) -> Result<u64, usize> {
self.params.compact_space.u128_to_compact(value)
}
fn compact_to_u128(&self, compact: u64) -> u128 {
self.params.compact_space.compact_to_u128(compact)
}
/// Comparing on compact space: Random dataset 0,24 (50% random hit) - 1.05 GElements/s
/// Comparing on compact space: Real dataset 1.08 GElements/s
///
/// Comparing on original space: Real dataset .06 GElements/s (not completely optimized)
#[inline]
pub fn get_positions_for_value_range(
&self,
value_range: RangeInclusive<u128>,
position_range: Range<u32>,
positions: &mut Vec<u32>,
) {
if value_range.start() > value_range.end() {
return;
}
let position_range = position_range.start..position_range.end.min(self.num_vals());
let from_value = *value_range.start();
let to_value = *value_range.end();
assert!(to_value >= from_value);
let compact_from = self.u128_to_compact(from_value);
let compact_to = self.u128_to_compact(to_value);
// Quick return, if both ranges fall into the same non-mapped space, the range can't cover
// any values, so we can early exit
match (compact_to, compact_from) {
(Err(pos1), Err(pos2)) if pos1 == pos2 => return,
_ => {}
}
let compact_from = compact_from.unwrap_or_else(|pos| {
// Correctness: Out of bounds, if this value is Err(last_index + 1), we early exit,
// since the to_value also mapps into the same non-mapped space
let range_mapping = self.params.compact_space.get_range_mapping(pos);
range_mapping.compact_start
});
// If there is no compact space, we go to the closest upperbound compact space
let compact_to = compact_to.unwrap_or_else(|pos| {
// Correctness: Overflow, if this value is Err(0), we early exit,
// since the from_value also mapps into the same non-mapped space
// Get end of previous range
let pos = pos - 1;
let range_mapping = self.params.compact_space.get_range_mapping(pos);
range_mapping.compact_end()
});
let range = compact_from..=compact_to;
let scan_num_docs = position_range.end - position_range.start;
let step_size = 4;
let cutoff = position_range.start + scan_num_docs - scan_num_docs % step_size;
let mut push_if_in_range = |idx, val| {
if range.contains(&val) {
positions.push(idx);
}
};
let get_val = |idx| self.params.bit_unpacker.get(idx, &self.data);
// unrolled loop
for idx in (position_range.start..cutoff).step_by(step_size as usize) {
let idx1 = idx;
let idx2 = idx + 1;
let idx3 = idx + 2;
let idx4 = idx + 3;
let val1 = get_val(idx1);
let val2 = get_val(idx2);
let val3 = get_val(idx3);
let val4 = get_val(idx4);
push_if_in_range(idx1, val1);
push_if_in_range(idx2, val2);
push_if_in_range(idx3, val3);
push_if_in_range(idx4, val4);
}
// handle rest
for idx in cutoff..position_range.end {
push_if_in_range(idx, get_val(idx));
}
}
#[inline]
fn iter_compact(&self) -> impl Iterator<Item = u64> + '_ {
(0..self.params.num_vals).map(move |idx| self.params.bit_unpacker.get(idx, &self.data))
}
#[inline]
fn iter(&self) -> impl Iterator<Item = u128> + '_ {
// TODO: Performance. It would be better to iterate on the ranges and check existence via
// the bit_unpacker.
self.iter_compact()
.map(|compact| self.compact_to_u128(compact))
}
#[inline]
pub fn get(&self, idx: u32) -> u128 {
let compact = self.params.bit_unpacker.get(idx, &self.data);
self.compact_to_u128(compact)
}
pub fn min_value(&self) -> u128 {
self.params.min_value
}
pub fn max_value(&self) -> u128 {
self.params.max_value
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::format_version::read_format_version;
use crate::null_index_footer::read_null_index_footer;
use crate::serialize::U128Header;
use crate::{open_u128, serialize_u128};
#[test]
fn compact_space_test() {
let ips = &[
2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
]
.into_iter()
.collect();
let compact_space = get_compact_space(ips, ips.len() as u32, 11);
let amplitude = compact_space.amplitude_compact_space();
assert_eq!(amplitude, 17);
assert_eq!(1, compact_space.u128_to_compact(2).unwrap());
assert_eq!(2, compact_space.u128_to_compact(3).unwrap());
assert_eq!(compact_space.u128_to_compact(100).unwrap_err(), 1);
for (num1, num2) in (0..3).tuple_windows() {
assert_eq!(
compact_space.get_range_mapping(num1).compact_end() + 1,
compact_space.get_range_mapping(num2).compact_start
);
}
let mut output: Vec<u8> = Vec::new();
compact_space.serialize(&mut output).unwrap();
assert_eq!(
compact_space,
CompactSpace::deserialize(&mut &output[..]).unwrap()
);
for ip in ips {
let compact = compact_space.u128_to_compact(*ip).unwrap();
assert_eq!(compact_space.compact_to_u128(compact), *ip);
}
}
#[test]
fn compact_space_amplitude_test() {
let ips = &[100000u128, 1000000].into_iter().collect();
let compact_space = get_compact_space(ips, ips.len() as u32, 1);
let amplitude = compact_space.amplitude_compact_space();
assert_eq!(amplitude, 2);
}
fn test_all(mut data: OwnedBytes, expected: &[u128]) {
let _header = U128Header::deserialize(&mut data);
let decompressor = CompactSpaceDecompressor::open(data).unwrap();
for (idx, expected_val) in expected.iter().cloned().enumerate() {
let val = decompressor.get(idx as u32);
assert_eq!(val, expected_val);
let test_range = |range: RangeInclusive<u128>| {
let expected_positions = expected
.iter()
.positions(|val| range.contains(val))
.map(|pos| pos as u32)
.collect::<Vec<_>>();
let mut positions = Vec::new();
decompressor.get_positions_for_value_range(
range,
0..decompressor.num_vals(),
&mut positions,
);
assert_eq!(positions, expected_positions);
};
test_range(expected_val.saturating_sub(1)..=expected_val);
test_range(expected_val..=expected_val);
test_range(expected_val..=expected_val.saturating_add(1));
test_range(expected_val.saturating_sub(1)..=expected_val.saturating_add(1));
}
}
fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
let mut out = Vec::new();
serialize_u128(
|| u128_vals.iter().cloned(),
u128_vals.len() as u32,
&mut out,
)
.unwrap();
let data = OwnedBytes::new(out);
let (data, _format_version) = read_format_version(data).unwrap();
let (data, _null_index_footer) = read_null_index_footer(data).unwrap();
test_all(data.clone(), u128_vals);
data
}
#[test]
fn test_range_1() {
let vals = &[
1u128,
100u128,
3u128,
99999u128,
100000u128,
100001u128,
4_000_211_221u128,
4_000_211_222u128,
333u128,
];
let mut data = test_aux_vals(vals);
let _header = U128Header::deserialize(&mut data);
let decomp = CompactSpaceDecompressor::open(data).unwrap();
let complete_range = 0..vals.len() as u32;
for (pos, val) in vals.iter().enumerate() {
let val = *val;
let pos = pos as u32;
let mut positions = Vec::new();
decomp.get_positions_for_value_range(val..=val, pos..pos + 1, &mut positions);
assert_eq!(positions, vec![pos]);
}
// handle docid range out of bounds
let positions = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
assert_eq!(positions, vec![]);
let positions =
get_positions_for_value_range_helper(&decomp, 0..=1, complete_range.clone());
assert_eq!(positions, vec![0]);
let positions =
get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
assert_eq!(positions, vec![0]);
let positions =
get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
assert_eq!(positions, vec![0, 2]);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
99999u128..=99999u128,
complete_range.clone()
),
vec![3]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
99999u128..=100000u128,
complete_range.clone()
),
vec![3, 4]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
99998u128..=100000u128,
complete_range.clone()
),
vec![3, 4]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
99998u128..=99999u128,
complete_range.clone()
),
vec![3]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
99998u128..=99998u128,
complete_range.clone()
),
vec![]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
333u128..=333u128,
complete_range.clone()
),
vec![8]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
332u128..=333u128,
complete_range.clone()
),
vec![8]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
332u128..=334u128,
complete_range.clone()
),
vec![8]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
333u128..=334u128,
complete_range.clone()
),
vec![8]
);
assert_eq!(
get_positions_for_value_range_helper(
&decomp,
4_000_211_221u128..=5_000_000_000u128,
complete_range
),
vec![6, 7]
);
}
#[test]
fn test_empty() {
let vals = &[];
let data = test_aux_vals(vals);
let _decomp = CompactSpaceDecompressor::open(data).unwrap();
}
#[test]
fn test_range_2() {
let vals = &[
100u128,
99999u128,
100000u128,
100001u128,
4_000_211_221u128,
4_000_211_222u128,
333u128,
];
let mut data = test_aux_vals(vals);
let _header = U128Header::deserialize(&mut data);
let decomp = CompactSpaceDecompressor::open(data).unwrap();
let complete_range = 0..vals.len() as u32;
assert_eq!(
get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone()),
vec![]
);
assert_eq!(
get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
vec![0]
);
assert_eq!(
get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
vec![0]
);
}
fn get_positions_for_value_range_helper<C: Column<T> + ?Sized, T: PartialOrd>(
column: &C,
value_range: RangeInclusive<T>,
doc_id_range: Range<u32>,
) -> Vec<u32> {
let mut positions = Vec::new();
column.get_docids_for_value_range(value_range, doc_id_range, &mut positions);
positions
}
#[test]
fn test_range_3() {
let vals = &[
200u128,
201,
202,
203,
204,
204,
206,
207,
208,
209,
210,
1_000_000,
5_000_000_000,
];
let mut out = Vec::new();
serialize_u128(|| vals.iter().cloned(), vals.len() as u32, &mut out).unwrap();
let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
let complete_range = 0..vals.len() as u32;
assert_eq!(
get_positions_for_value_range_helper(&*decomp, 199..=200, complete_range.clone()),
vec![0]
);
assert_eq!(
get_positions_for_value_range_helper(&*decomp, 199..=201, complete_range.clone()),
vec![0, 1]
);
assert_eq!(
get_positions_for_value_range_helper(&*decomp, 200..=200, complete_range.clone()),
vec![0]
);
assert_eq!(
get_positions_for_value_range_helper(&*decomp, 1_000_000..=1_000_000, complete_range),
vec![11]
);
}
#[test]
fn test_bug1() {
let vals = &[9223372036854775806];
let _data = test_aux_vals(vals);
}
#[test]
fn test_bug2() {
let vals = &[340282366920938463463374607431768211455u128];
let _data = test_aux_vals(vals);
}
#[test]
fn test_bug3() {
let vals = &[340282366920938463463374607431768211454];
let _data = test_aux_vals(vals);
}
#[test]
fn test_bug4() {
let vals = &[340282366920938463463374607431768211455, 0];
let _data = test_aux_vals(vals);
}
#[test]
fn test_first_large_gaps() {
let vals = &[1_000_000_000u128; 100];
let _data = test_aux_vals(vals);
}
use itertools::Itertools;
use proptest::prelude::*;
fn num_strategy() -> impl Strategy<Value = u128> {
prop_oneof![
1 => prop::num::u128::ANY.prop_map(|num| u128::MAX - (num % 10) ),
1 => prop::num::u128::ANY.prop_map(|num| i64::MAX as u128 + 5 - (num % 10) ),
1 => prop::num::u128::ANY.prop_map(|num| i128::MAX as u128 + 5 - (num % 10) ),
1 => prop::num::u128::ANY.prop_map(|num| num % 10 ),
20 => prop::num::u128::ANY,
]
}
proptest! {
#![proptest_config(ProptestConfig::with_cases(10))]
#[test]
fn compress_decompress_random(vals in proptest::collection::vec(num_strategy()
, 1..1000)) {
let _data = test_aux_vals(&vals);
}
}
}

39
fastfield_codecs/src/format_version.rs Normal file
View File

@@ -0,0 +1,39 @@
use std::io;
use common::BinarySerializable;
use ownedbytes::OwnedBytes;
const MAGIC_NUMBER: u16 = 4335u16;
const FASTFIELD_FORMAT_VERSION: u8 = 1;
pub(crate) fn append_format_version(output: &mut impl io::Write) -> io::Result<()> {
FASTFIELD_FORMAT_VERSION.serialize(output)?;
MAGIC_NUMBER.serialize(output)?;
Ok(())
}
pub(crate) fn read_format_version(data: OwnedBytes) -> io::Result<(OwnedBytes, u8)> {
let (data, magic_number_bytes) = data.rsplit(2);
let magic_number = u16::deserialize(&mut magic_number_bytes.as_slice())?;
if magic_number != MAGIC_NUMBER {
return Err(io::Error::new(
io::ErrorKind::InvalidData,
format!("magic number mismatch {} != {}", magic_number, MAGIC_NUMBER),
));
}
let (data, format_version_bytes) = data.rsplit(1);
let format_version = u8::deserialize(&mut format_version_bytes.as_slice())?;
if format_version > FASTFIELD_FORMAT_VERSION {
return Err(io::Error::new(
io::ErrorKind::InvalidData,
format!(
"Unsupported fastfield format version: {}. Max supported version: {}",
format_version, FASTFIELD_FORMAT_VERSION
),
));
}
Ok((data, format_version))
}

170
fastfield_codecs/src/gcd.rs Normal file
View File

@@ -0,0 +1,170 @@
use std::num::NonZeroU64;
use fastdivide::DividerU64;
/// Compute the gcd of two non null numbers.
///
/// It is recommended, but not required, to feed values such that `large >= small`.
fn compute_gcd(mut large: NonZeroU64, mut small: NonZeroU64) -> NonZeroU64 {
loop {
let rem: u64 = large.get() % small;
if let Some(new_small) = NonZeroU64::new(rem) {
(large, small) = (small, new_small);
} else {
return small;
}
}
}
// Find GCD for iterator of numbers
pub fn find_gcd(numbers: impl Iterator<Item = u64>) -> Option<NonZeroU64> {
let mut numbers = numbers.flat_map(NonZeroU64::new);
let mut gcd: NonZeroU64 = numbers.next()?;
if gcd.get() == 1 {
return Some(gcd);
}
let mut gcd_divider = DividerU64::divide_by(gcd.get());
for val in numbers {
let remainder = val.get() - (gcd_divider.divide(val.get())) * gcd.get();
if remainder == 0 {
continue;
}
gcd = compute_gcd(val, gcd);
if gcd.get() == 1 {
return Some(gcd);
}
gcd_divider = DividerU64::divide_by(gcd.get());
}
Some(gcd)
}
#[cfg(test)]
mod tests {
use std::io;
use std::num::NonZeroU64;
use ownedbytes::OwnedBytes;
use crate::gcd::{compute_gcd, find_gcd};
use crate::{FastFieldCodecType, VecColumn};
fn test_fastfield_gcd_i64_with_codec(
codec_type: FastFieldCodecType,
num_vals: usize,
) -> io::Result<()> {
let mut vals: Vec<i64> = (-4..=(num_vals as i64) - 5).map(|val| val * 1000).collect();
let mut buffer: Vec<u8> = Vec::new();
crate::serialize(VecColumn::from(&vals), &mut buffer, &[codec_type])?;
let buffer = OwnedBytes::new(buffer);
let column = crate::open::<i64>(buffer.clone())?;
assert_eq!(column.get_val(0), -4000i64);
assert_eq!(column.get_val(1), -3000i64);
assert_eq!(column.get_val(2), -2000i64);
assert_eq!(column.max_value(), (num_vals as i64 - 5) * 1000);
assert_eq!(column.min_value(), -4000i64);
// Can't apply gcd
let mut buffer_without_gcd = Vec::new();
vals.pop();
vals.push(1001i64);
crate::serialize(
VecColumn::from(&vals),
&mut buffer_without_gcd,
&[codec_type],
)?;
let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
assert!(buffer_without_gcd.len() > buffer.len());
Ok(())
}
#[test]
fn test_fastfield_gcd_i64() -> io::Result<()> {
for &codec_type in &[
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
FastFieldCodecType::Linear,
] {
test_fastfield_gcd_i64_with_codec(codec_type, 5500)?;
}
Ok(())
}
fn test_fastfield_gcd_u64_with_codec(
codec_type: FastFieldCodecType,
num_vals: usize,
) -> io::Result<()> {
let mut vals: Vec<u64> = (1..=num_vals).map(|i| i as u64 * 1000u64).collect();
let mut buffer: Vec<u8> = Vec::new();
crate::serialize(VecColumn::from(&vals), &mut buffer, &[codec_type])?;
let buffer = OwnedBytes::new(buffer);
let column = crate::open::<u64>(buffer.clone())?;
assert_eq!(column.get_val(0), 1000u64);
assert_eq!(column.get_val(1), 2000u64);
assert_eq!(column.get_val(2), 3000u64);
assert_eq!(column.max_value(), num_vals as u64 * 1000);
assert_eq!(column.min_value(), 1000u64);
// Can't apply gcd
let mut buffer_without_gcd = Vec::new();
vals.pop();
vals.push(1001u64);
crate::serialize(
VecColumn::from(&vals),
&mut buffer_without_gcd,
&[codec_type],
)?;
let buffer_without_gcd = OwnedBytes::new(buffer_without_gcd);
assert!(buffer_without_gcd.len() > buffer.len());
Ok(())
}
#[test]
fn test_fastfield_gcd_u64() -> io::Result<()> {
for &codec_type in &[
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
FastFieldCodecType::Linear,
] {
test_fastfield_gcd_u64_with_codec(codec_type, 5500)?;
}
Ok(())
}
#[test]
pub fn test_fastfield2() {
let test_fastfield = crate::serialize_and_load(&[100u64, 200u64, 300u64]);
assert_eq!(test_fastfield.get_val(0), 100);
assert_eq!(test_fastfield.get_val(1), 200);
assert_eq!(test_fastfield.get_val(2), 300);
}
#[test]
fn test_compute_gcd() {
let test_compute_gcd_aux = |large, small, expected| {
let large = NonZeroU64::new(large).unwrap();
let small = NonZeroU64::new(small).unwrap();
let expected = NonZeroU64::new(expected).unwrap();
assert_eq!(compute_gcd(small, large), expected);
assert_eq!(compute_gcd(large, small), expected);
};
test_compute_gcd_aux(1, 4, 1);
test_compute_gcd_aux(2, 4, 2);
test_compute_gcd_aux(10, 25, 5);
test_compute_gcd_aux(25, 25, 25);
}
#[test]
fn find_gcd_test() {
assert_eq!(find_gcd([0].into_iter()), None);
assert_eq!(find_gcd([0, 10].into_iter()), NonZeroU64::new(10));
assert_eq!(find_gcd([10, 0].into_iter()), NonZeroU64::new(10));
assert_eq!(find_gcd([].into_iter()), None);
assert_eq!(find_gcd([15, 30, 5, 10].into_iter()), NonZeroU64::new(5));
assert_eq!(find_gcd([15, 16, 10].into_iter()), NonZeroU64::new(1));
assert_eq!(find_gcd([0, 5, 5, 5].into_iter()), NonZeroU64::new(5));
assert_eq!(find_gcd([0, 0].into_iter()), None);
}
}

608
fastfield_codecs/src/lib.rs
View File

@@ -1,130 +1,330 @@
#![warn(missing_docs)]
#![cfg_attr(all(feature = "unstable", test), feature(test))]
//! # `fastfield_codecs`
//!
//! - Columnar storage of data for tantivy [`Column`].
//! - Encode data in different codecs.
//! - Monotonically map values to u64/u128
#[cfg(test)]
#[macro_use]
extern crate more_asserts;
#[cfg(all(test, feature = "unstable"))]
extern crate test;
use std::io;
use std::io::Write;
use std::sync::Arc;
pub mod bitpacked;
pub mod linearinterpol;
pub mod multilinearinterpol;
use common::BinarySerializable;
use compact_space::CompactSpaceDecompressor;
use format_version::read_format_version;
use monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
};
use null_index_footer::read_null_index_footer;
use ownedbytes::OwnedBytes;
use serialize::{Header, U128Header};
pub trait FastFieldCodecReader: Sized {
/// reads the metadata and returns the CodecReader
fn open_from_bytes(bytes: &[u8]) -> std::io::Result<Self>;
mod bitpacked;
mod blockwise_linear;
mod compact_space;
mod format_version;
mod line;
mod linear;
mod monotonic_mapping;
mod monotonic_mapping_u128;
#[allow(dead_code)]
mod null_index;
mod null_index_footer;
fn get_u64(&self, doc: u64, data: &[u8]) -> u64;
mod column;
mod gcd;
pub mod serialize;
fn min_value(&self) -> u64;
fn max_value(&self) -> u64;
pub use ordered_float;
use self::bitpacked::BitpackedCodec;
use self::blockwise_linear::BlockwiseLinearCodec;
pub use self::column::{monotonic_map_column, Column, IterColumn, VecColumn};
use self::linear::LinearCodec;
pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
pub use self::serialize::{
estimate, serialize, serialize_and_load, serialize_u128, NormalizedHeader,
};
#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
#[repr(u8)]
/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
pub enum FastFieldCodecType {
/// Bitpack all values in the value range. The number of bits is defined by the amplitude
/// `column.max_value() - column.min_value()`
Bitpacked = 1,
/// Linear interpolation puts a line between the first and last value and then bitpacks the
/// values by the offset from the line. The number of bits is defined by the max deviation from
/// the line.
Linear = 2,
/// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
BlockwiseLinear = 3,
}
impl BinarySerializable for FastFieldCodecType {
fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
self.to_code().serialize(wrt)
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let code = u8::deserialize(reader)?;
let codec_type: Self = Self::from_code(code)
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
Ok(codec_type)
}
}
impl FastFieldCodecType {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::Bitpacked),
2 => Some(Self::Linear),
3 => Some(Self::BlockwiseLinear),
_ => None,
}
}
}
#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
#[repr(u8)]
/// Available codecs to use to encode the u128 (via [`MonotonicallyMappableToU128`]) converted data.
pub enum U128FastFieldCodecType {
/// This codec takes a large number space (u128) and reduces it to a compact number space, by
/// removing the holes.
CompactSpace = 1,
}
impl BinarySerializable for U128FastFieldCodecType {
fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
self.to_code().serialize(wrt)
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let code = u8::deserialize(reader)?;
let codec_type: Self = Self::from_code(code)
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
Ok(codec_type)
}
}
impl U128FastFieldCodecType {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::CompactSpace),
_ => None,
}
}
}
/// Returns the correct codec reader wrapped in the `Arc` for the data.
pub fn open_u128<Item: MonotonicallyMappableToU128>(
bytes: OwnedBytes,
) -> io::Result<Arc<dyn Column<Item>>> {
let (bytes, _format_version) = read_format_version(bytes)?;
let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
let header = U128Header::deserialize(&mut bytes)?;
assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
let reader = CompactSpaceDecompressor::open(bytes)?;
let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<Item>> =
StrictlyMonotonicMappingToInternal::<Item>::new().into();
Ok(Arc::new(monotonic_map_column(reader, inverted)))
}
/// Returns the correct codec reader wrapped in the `Arc` for the data.
pub fn open<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Arc<dyn Column<T>>> {
let (bytes, _format_version) = read_format_version(bytes)?;
let (mut bytes, _null_index_footer) = read_null_index_footer(bytes)?;
let header = Header::deserialize(&mut bytes)?;
match header.codec_type {
FastFieldCodecType::Bitpacked => open_specific_codec::<BitpackedCodec, _>(bytes, &header),
FastFieldCodecType::Linear => open_specific_codec::<LinearCodec, _>(bytes, &header),
FastFieldCodecType::BlockwiseLinear => {
open_specific_codec::<BlockwiseLinearCodec, _>(bytes, &header)
}
}
}
fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
bytes: OwnedBytes,
header: &Header,
) -> io::Result<Arc<dyn Column<Item>>> {
let normalized_header = header.normalized();
let reader = C::open_from_bytes(bytes, normalized_header)?;
let min_value = header.min_value;
if let Some(gcd) = header.gcd {
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
} else {
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalBaseval::new(min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
}
}
/// The FastFieldSerializerEstimate trait is required on all variants
/// of fast field compressions, to decide which one to choose.
pub trait FastFieldCodecSerializer {
trait FastFieldCodec: 'static {
/// A codex needs to provide a unique name and id, which is
/// used for debugging and de/serialization.
const NAME: &'static str;
const ID: u8;
const CODEC_TYPE: FastFieldCodecType;
/// Check if the Codec is able to compress the data
fn is_applicable(fastfield_accessor: &impl FastFieldDataAccess, stats: FastFieldStats) -> bool;
type Reader: Column<u64> + 'static;
/// Reads the metadata and returns the CodecReader
fn open_from_bytes(bytes: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader>;
/// Serializes the data using the serializer into write.
///
/// The column iterator should be preferred over using column `get_val` method for
/// performance reasons.
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()>;
/// Returns an estimate of the compression ratio.
/// If the codec is not applicable, returns `None`.
///
/// The baseline is uncompressed 64bit data.
///
/// It could make sense to also return a value representing
/// computational complexity.
fn estimate(fastfield_accessor: &impl FastFieldDataAccess, stats: FastFieldStats) -> f32;
/// Serializes the data using the serializer into write.
/// There are multiple iterators, in case the codec needs to read the data multiple times.
/// The iterators should be preferred over using fastfield_accessor for performance reasons.
fn serialize(
write: &mut impl Write,
fastfield_accessor: &dyn FastFieldDataAccess,
stats: FastFieldStats,
data_iter: impl Iterator<Item = u64>,
data_iter1: impl Iterator<Item = u64>,
) -> io::Result<()>;
fn estimate(column: &dyn Column) -> Option<f32>;
}
/// FastFieldDataAccess is the trait to access fast field data during serialization and estimation.
pub trait FastFieldDataAccess {
/// Return the value associated to the given position.
///
/// Whenever possible use the Iterator passed to the fastfield creation instead, for performance
/// reasons.
///
/// # Panics
///
/// May panic if `position` is greater than the index.
fn get_val(&self, position: u64) -> u64;
}
#[derive(Debug, Clone)]
/// Statistics are used in codec detection and stored in the fast field footer.
pub struct FastFieldStats {
pub min_value: u64,
pub max_value: u64,
pub num_vals: u64,
}
impl<'a> FastFieldDataAccess for &'a [u64] {
fn get_val(&self, position: u64) -> u64 {
self[position as usize]
}
}
impl FastFieldDataAccess for Vec<u64> {
fn get_val(&self, position: u64) -> u64 {
self[position as usize]
}
}
/// The list of all available codecs for u64 convertible data.
pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
FastFieldCodecType::Linear,
];
#[cfg(test)]
mod tests {
use crate::bitpacked::{BitpackedFastFieldReader, BitpackedFastFieldSerializer};
use crate::linearinterpol::{LinearInterpolFastFieldReader, LinearInterpolFastFieldSerializer};
use crate::multilinearinterpol::{
MultiLinearInterpolFastFieldReader, MultiLinearInterpolFastFieldSerializer,
};
pub fn create_and_validate<S: FastFieldCodecSerializer, R: FastFieldCodecReader>(
use proptest::prelude::*;
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use crate::bitpacked::BitpackedCodec;
use crate::blockwise_linear::BlockwiseLinearCodec;
use crate::linear::LinearCodec;
use crate::serialize::Header;
pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
data: &[u64],
name: &str,
) -> (f32, f32) {
if !S::is_applicable(&data, crate::tests::stats_from_vec(data)) {
return (f32::MAX, 0.0);
}
let estimation = S::estimate(&data, crate::tests::stats_from_vec(data));
let mut out = vec![];
S::serialize(
&mut out,
&data,
crate::tests::stats_from_vec(data),
data.iter().cloned(),
data.iter().cloned(),
)
.unwrap();
) -> Option<(f32, f32)> {
let col = &VecColumn::from(data);
let header = Header::compute_header(col, &[Codec::CODEC_TYPE])?;
let normalized_col = header.normalize_column(col);
let estimation = Codec::estimate(&normalized_col)?;
let mut out = Vec::new();
let col = VecColumn::from(data);
serialize(col, &mut out, &[Codec::CODEC_TYPE]).unwrap();
let reader = R::open_from_bytes(&out).unwrap();
for (doc, orig_val) in data.iter().enumerate() {
let val = reader.get_u64(doc as u64, &out);
if val != *orig_val {
panic!(
"val {:?} does not match orig_val {:?}, in data set {}, data {:?}",
val, orig_val, name, data
);
}
}
let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
(estimation, actual_compression)
let reader = crate::open::<u64>(OwnedBytes::new(out)).unwrap();
assert_eq!(reader.num_vals(), data.len() as u32);
for (doc, orig_val) in data.iter().copied().enumerate() {
let val = reader.get_val(doc as u32);
assert_eq!(
val, orig_val,
"val `{val}` does not match orig_val {orig_val:?}, in data set {name}, data \
`{data:?}`",
);
}
if !data.is_empty() {
let test_rand_idx = rand::thread_rng().gen_range(0..=data.len() - 1);
let expected_positions: Vec<u32> = data
.iter()
.enumerate()
.filter(|(_, el)| **el == data[test_rand_idx])
.map(|(pos, _)| pos as u32)
.collect();
let mut positions = Vec::new();
reader.get_docids_for_value_range(
data[test_rand_idx]..=data[test_rand_idx],
0..data.len() as u32,
&mut positions,
);
assert_eq!(expected_positions, positions);
}
Some((estimation, actual_compression))
}
pub fn get_codec_test_data_sets() -> Vec<(Vec<u64>, &'static str)> {
proptest! {
#![proptest_config(ProptestConfig::with_cases(100))]
#[test]
fn test_proptest_small_bitpacked(data in proptest::collection::vec(num_strategy(), 1..10)) {
create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
}
#[test]
fn test_proptest_small_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
}
#[test]
fn test_proptest_small_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..10)) {
create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
}
}
proptest! {
#![proptest_config(ProptestConfig::with_cases(10))]
#[test]
fn test_proptest_large_bitpacked(data in proptest::collection::vec(num_strategy(), 1..6000)) {
create_and_validate::<BitpackedCodec>(&data, "proptest bitpacked");
}
#[test]
fn test_proptest_large_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
create_and_validate::<LinearCodec>(&data, "proptest linearinterpol");
}
#[test]
fn test_proptest_large_blockwise_linear(data in proptest::collection::vec(num_strategy(), 1..6000)) {
create_and_validate::<BlockwiseLinearCodec>(&data, "proptest multilinearinterpol");
}
}
fn num_strategy() -> impl Strategy<Value = u64> {
prop_oneof![
1 => prop::num::u64::ANY.prop_map(|num| u64::MAX - (num % 10) ),
1 => prop::num::u64::ANY.prop_map(|num| num % 10 ),
20 => prop::num::u64::ANY,
]
}
pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
let mut data_and_names = vec![];
let data = (10..=20_u64).collect::<Vec<_>>();
let data = (10..=10_000_u64).collect::<Vec<_>>();
data_and_names.push((data, "simple monotonically increasing"));
data_and_names.push((
@@ -134,92 +334,230 @@ mod tests {
data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
data_and_names.push((vec![10], "single value"));
data_and_names.push((
vec![1572656989877777, 1170935903116329, 720575940379279, 0],
"overflow error",
));
data_and_names
}
fn test_codec<S: FastFieldCodecSerializer, R: FastFieldCodecReader>() {
let codec_name = S::NAME;
for (data, data_set_name) in get_codec_test_data_sets() {
let (estimate, actual) =
crate::tests::create_and_validate::<S, R>(&data, data_set_name);
let result = if estimate == f32::MAX {
"Disabled".to_string()
fn test_codec<C: FastFieldCodec>() {
let codec_name = format!("{:?}", C::CODEC_TYPE);
for (data, dataset_name) in get_codec_test_datasets() {
let estimate_actual_opt: Option<(f32, f32)> =
crate::tests::create_and_validate::<C>(&data, dataset_name);
let result = if let Some((estimate, actual)) = estimate_actual_opt {
format!("Estimate `{estimate}` Actual `{actual}`")
} else {
format!("Estimate {:?} Actual {:?} ", estimate, actual)
"Disabled".to_string()
};
println!(
"Codec {}, DataSet {}, {}",
codec_name, data_set_name, result
);
println!("Codec {codec_name}, DataSet {dataset_name}, {result}");
}
}
#[test]
fn test_codec_bitpacking() {
test_codec::<BitpackedFastFieldSerializer, BitpackedFastFieldReader>();
test_codec::<BitpackedCodec>();
}
#[test]
fn test_codec_interpolation() {
test_codec::<LinearInterpolFastFieldSerializer, LinearInterpolFastFieldReader>();
test_codec::<LinearCodec>();
}
#[test]
fn test_codec_multi_interpolation() {
test_codec::<MultiLinearInterpolFastFieldSerializer, MultiLinearInterpolFastFieldReader>();
test_codec::<BlockwiseLinearCodec>();
}
use super::*;
pub fn stats_from_vec(data: &[u64]) -> FastFieldStats {
let min_value = data.iter().cloned().min().unwrap_or(0);
let max_value = data.iter().cloned().max().unwrap_or(0);
FastFieldStats {
min_value,
max_value,
num_vals: data.len() as u64,
}
}
#[test]
fn estimation_good_interpolation_case() {
let data = (10..=20000_u64).collect::<Vec<_>>();
let data: VecColumn = data.as_slice().into();
let linear_interpol_estimation =
LinearInterpolFastFieldSerializer::estimate(&data, stats_from_vec(&data));
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.01);
let multi_linear_interpol_estimation =
MultiLinearInterpolFastFieldSerializer::estimate(&data, stats_from_vec(&data));
let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
assert_le!(multi_linear_interpol_estimation, 0.2);
assert_le!(linear_interpol_estimation, multi_linear_interpol_estimation);
assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
let bitpacked_estimation =
BitpackedFastFieldSerializer::estimate(&data, stats_from_vec(&data));
assert_le!(linear_interpol_estimation, bitpacked_estimation);
let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
assert_lt!(linear_interpol_estimation, bitpacked_estimation);
}
#[test]
fn estimation_test_bad_interpolation_case() {
let data = vec![200, 10, 10, 10, 10, 1000, 20];
let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
let linear_interpol_estimation =
LinearInterpolFastFieldSerializer::estimate(&data, stats_from_vec(&data));
assert_le!(linear_interpol_estimation, 0.32);
let data: VecColumn = data.into();
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.34);
let bitpacked_estimation =
BitpackedFastFieldSerializer::estimate(&data, stats_from_vec(&data));
assert_le!(bitpacked_estimation, linear_interpol_estimation);
let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
assert_lt!(bitpacked_estimation, linear_interpol_estimation);
}
#[test]
fn estimation_prefer_bitpacked() {
let data = VecColumn::from(&[10, 10, 10, 10]);
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
assert_lt!(bitpacked_estimation, linear_interpol_estimation);
}
#[test]
fn estimation_test_bad_interpolation_case_monotonically_increasing() {
let mut data = (200..=20000_u64).collect::<Vec<_>>();
let mut data: Vec<u64> = (201..=20000_u64).collect();
data.push(1_000_000);
let data: VecColumn = data.as_slice().into();
// in this case the linear interpolation can't in fact not be worse than bitpacking,
// but the estimator adds some threshold, which leads to estimated worse behavior
let linear_interpol_estimation =
LinearInterpolFastFieldSerializer::estimate(&data, stats_from_vec(&data));
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.35);
let bitpacked_estimation =
BitpackedFastFieldSerializer::estimate(&data, stats_from_vec(&data));
let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
assert_le!(bitpacked_estimation, 0.32);
assert_le!(bitpacked_estimation, linear_interpol_estimation);
}
#[test]
fn test_fast_field_codec_type_to_code() {
let mut count_codec = 0;
for code in 0..=255 {
if let Some(codec_type) = FastFieldCodecType::from_code(code) {
assert_eq!(codec_type.to_code(), code);
count_codec += 1;
}
}
assert_eq!(count_codec, 3);
}
}
#[cfg(all(test, feature = "unstable"))]
mod bench {
use std::sync::Arc;
use ownedbytes::OwnedBytes;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::{self, Bencher};
use super::*;
use crate::Column;
fn get_data() -> Vec<u64> {
let mut rng = StdRng::seed_from_u64(2u64);
let mut data: Vec<_> = (100..55000_u64)
.map(|num| num + rng.gen::<u8>() as u64)
.collect();
data.push(99_000);
data.insert(1000, 2000);
data.insert(2000, 100);
data.insert(3000, 4100);
data.insert(4000, 100);
data.insert(5000, 800);
data
}
#[inline(never)]
fn value_iter() -> impl Iterator<Item = u64> {
0..20_000
}
fn get_reader_for_bench<Codec: FastFieldCodec>(data: &[u64]) -> Codec::Reader {
let mut bytes = Vec::new();
let min_value = *data.iter().min().unwrap();
let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
let col = VecColumn::from(&data);
let normalized_header = crate::NormalizedHeader {
num_vals: col.num_vals(),
max_value: col.max_value(),
};
Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
Codec::open_from_bytes(OwnedBytes::new(bytes), normalized_header).unwrap()
}
fn bench_get<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
let col = get_reader_for_bench::<Codec>(data);
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u32);
sum = sum.wrapping_add(val);
}
sum
});
}
#[inline(never)]
fn bench_get_dynamic_helper(b: &mut Bencher, col: Arc<dyn Column>) {
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u32);
sum = sum.wrapping_add(val);
}
sum
});
}
fn bench_get_dynamic<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
let col = Arc::new(get_reader_for_bench::<Codec>(data));
bench_get_dynamic_helper(b, col);
}
fn bench_create<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
let min_value = *data.iter().min().unwrap();
let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
let mut bytes = Vec::new();
b.iter(|| {
bytes.clear();
Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
});
}
#[bench]
fn bench_fastfield_bitpack_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<BitpackedCodec>(b, &data);
}
#[bench]
fn bench_fastfield_linearinterpol_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<LinearCodec>(b, &data);
}
#[bench]
fn bench_fastfield_multilinearinterpol_create(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_create::<BlockwiseLinearCodec>(b, &data);
}
#[bench]
fn bench_fastfield_bitpack_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<BitpackedCodec>(b, &data);
}
#[bench]
fn bench_fastfield_bitpack_get_dynamic(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get_dynamic::<BitpackedCodec>(b, &data);
}
#[bench]
fn bench_fastfield_linearinterpol_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<LinearCodec>(b, &data);
}
#[bench]
fn bench_fastfield_linearinterpol_get_dynamic(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get_dynamic::<LinearCodec>(b, &data);
}
#[bench]
fn bench_fastfield_multilinearinterpol_get(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get::<BlockwiseLinearCodec>(b, &data);
}
#[bench]
fn bench_fastfield_multilinearinterpol_get_dynamic(b: &mut Bencher) {
let data: Vec<_> = get_data();
bench_get_dynamic::<BlockwiseLinearCodec>(b, &data);
}
}

222
fastfield_codecs/src/line.rs Normal file
View File

@@ -0,0 +1,222 @@
use std::io;
use std::num::NonZeroU32;
use common::{BinarySerializable, VInt};
use crate::Column;
const MID_POINT: u64 = (1u64 << 32) - 1u64;
/// `Line` describes a line function `y: ax + b` using integer
/// arithmetics.
///
/// The slope is in fact a decimal split into a 32 bit integer value,
/// and a 32-bit decimal value.
///
/// The multiplication then becomes.
/// `y = m * x >> 32 + b`
#[derive(Debug, Clone, Copy, Default)]
pub struct Line {
slope: u64,
intercept: u64,
}
/// Compute the line slope.
///
/// This function has the nice property of being
/// invariant by translation.
/// `
/// compute_slope(y0, y1)
/// = compute_slope(y0 + X % 2^64, y1 + X % 2^64)
/// `
fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU32) -> u64 {
let dy = y1.wrapping_sub(y0);
let sign = dy <= (1 << 63);
let abs_dy = if sign {
y1.wrapping_sub(y0)
} else {
y0.wrapping_sub(y1)
};
if abs_dy >= 1 << 32 {
// This is outside of realm we handle.
// Let's just bail.
return 0u64;
}
let abs_slope = (abs_dy << 32) / num_vals.get() as u64;
if sign {
abs_slope
} else {
// The complement does indeed create the
// opposite decreasing slope...
//
// Intuitively (without the bitshifts and % u64::MAX)
// ```
// (x + shift)*(u64::MAX - abs_slope)
// - (x * (u64::MAX - abs_slope))
// = - shift * abs_slope
// ```
u64::MAX - abs_slope
}
}
impl Line {
#[inline(always)]
pub fn eval(&self, x: u32) -> u64 {
let linear_part = ((x as u64).wrapping_mul(self.slope) >> 32) as i32 as u64;
self.intercept.wrapping_add(linear_part)
}
// Same as train, but the intercept is only estimated from provided sample positions
pub fn estimate(sample_positions_and_values: &[(u64, u64)]) -> Self {
let first_val = sample_positions_and_values[0].1;
let last_val = sample_positions_and_values[sample_positions_and_values.len() - 1].1;
let num_vals = sample_positions_and_values[sample_positions_and_values.len() - 1].0 + 1;
Self::train_from(
first_val,
last_val,
num_vals as u32,
sample_positions_and_values.iter().cloned(),
)
}
// Intercept is only computed from provided positions
fn train_from(
first_val: u64,
last_val: u64,
num_vals: u32,
positions_and_values: impl Iterator<Item = (u64, u64)>,
) -> Self {
// TODO replace with let else
let idx_last_val = if let Some(idx_last_val) = NonZeroU32::new(num_vals - 1) {
idx_last_val
} else {
return Line::default();
};
let y0 = first_val;
let y1 = last_val;
// We first independently pick our slope.
let slope = compute_slope(y0, y1, idx_last_val);
// We picked our slope. Note that it does not have to be perfect.
// Now we need to compute the best intercept.
//
// Intuitively, the best intercept is such that line passes through one of the
// `(i, ys[])`.
//
// The best intercept therefore has the form
// `y[i] - line.eval(i)` (using wrapping arithmetics).
// In other words, the best intercept is one of the `y - Line::eval(ys[i])`
// and our task is just to pick the one that minimizes our error.
//
// Without sorting our values, this is a difficult problem.
// We however rely on the following trick...
//
// We only focus on the case where the interpolation is half decent.
// If the line interpolation is doing its job on a dataset suited for it,
// we can hope that the maximum error won't be larger than `u64::MAX / 2`.
//
// In other words, even without the intercept the values `y - Line::eval(ys[i])` will all be
// within an interval that takes less than half of the modulo space of `u64`.
//
// Our task is therefore to identify this interval.
// Here we simply translate all of our values by `y0 - 2^63` and pick the min.
let mut line = Line {
slope,
intercept: 0,
};
let heuristic_shift = y0.wrapping_sub(MID_POINT);
line.intercept = positions_and_values
.map(|(pos, y)| y.wrapping_sub(line.eval(pos as u32)))
.min_by_key(|&val| val.wrapping_sub(heuristic_shift))
.unwrap_or(0u64); //< Never happens.
line
}
/// Returns a line that attemps to approximate a function
/// f: i in 0..[ys.num_vals()) -> ys[i].
///
/// - The approximation is always lower than the actual value.
/// Or more rigorously, formally `f(i).wrapping_sub(ys[i])` is small
/// for any i in [0..ys.len()).
/// - It computes without panicking for any value of it.
///
/// This function is only invariable by translation if all of the
/// `ys` are packaged into half of the space. (See heuristic below)
pub fn train(ys: &dyn Column) -> Self {
let first_val = ys.iter().next().unwrap();
let last_val = ys.iter().nth(ys.num_vals() as usize - 1).unwrap();
Self::train_from(
first_val,
last_val,
ys.num_vals(),
ys.iter().enumerate().map(|(pos, val)| (pos as u64, val)),
)
}
}
impl BinarySerializable for Line {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
VInt(self.slope).serialize(writer)?;
VInt(self.intercept).serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let slope = VInt::deserialize(reader)?.0;
let intercept = VInt::deserialize(reader)?.0;
Ok(Line { slope, intercept })
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::VecColumn;
/// Test training a line and ensuring that the maximum difference between
/// the data points and the line is `expected`.
///
/// This function operates translation over the data for better coverage.
#[track_caller]
fn test_line_interpol_with_translation(ys: &[u64], expected: Option<u64>) {
let mut translations = vec![0, 100, u64::MAX / 2, u64::MAX, u64::MAX - 1];
translations.extend_from_slice(ys);
for translation in translations {
let translated_ys: Vec<u64> = ys
.iter()
.copied()
.map(|y| y.wrapping_add(translation))
.collect();
let largest_err = test_eval_max_err(&translated_ys);
assert_eq!(largest_err, expected);
}
}
fn test_eval_max_err(ys: &[u64]) -> Option<u64> {
let line = Line::train(&VecColumn::from(&ys));
ys.iter()
.enumerate()
.map(|(x, y)| y.wrapping_sub(line.eval(x as u32)))
.max()
}
#[test]
fn test_train() {
test_line_interpol_with_translation(&[11, 11, 11, 12, 12, 13], Some(1));
test_line_interpol_with_translation(&[13, 12, 12, 11, 11, 11], Some(1));
test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
test_line_interpol_with_translation(&[13, 13, 12, 11, 11, 11], Some(1));
test_line_interpol_with_translation(&[u64::MAX - 1, 0, 0, 1], Some(1));
test_line_interpol_with_translation(&[u64::MAX - 1, u64::MAX, 0, 1], Some(0));
test_line_interpol_with_translation(&[0, 1, 2, 3, 5], Some(0));
test_line_interpol_with_translation(&[1, 2, 3, 4], Some(0));
let data: Vec<u64> = (0..255).collect();
test_line_interpol_with_translation(&data, Some(0));
let data: Vec<u64> = (0..255).map(|el| el * 2).collect();
test_line_interpol_with_translation(&data, Some(0));
}
}

231
fastfield_codecs/src/linear.rs Normal file
View File

@@ -0,0 +1,231 @@
use std::io::{self, Write};
use common::BinarySerializable;
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::line::Line;
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType};
/// Depending on the field type, a different
/// fast field is required.
#[derive(Clone)]
pub struct LinearReader {
data: OwnedBytes,
linear_params: LinearParams,
header: NormalizedHeader,
}
impl Column for LinearReader {
#[inline]
fn get_val(&self, doc: u32) -> u64 {
let interpoled_val: u64 = self.linear_params.line.eval(doc);
let bitpacked_diff = self.linear_params.bit_unpacker.get(doc, &self.data);
interpoled_val.wrapping_add(bitpacked_diff)
}
#[inline]
fn min_value(&self) -> u64 {
// The LinearReader assumes a normalized vector.
0u64
}
#[inline]
fn max_value(&self) -> u64 {
self.header.max_value
}
#[inline]
fn num_vals(&self) -> u32 {
self.header.num_vals
}
}
/// Fastfield serializer, which tries to guess values by linear interpolation
/// and stores the difference bitpacked.
pub struct LinearCodec;
#[derive(Debug, Clone)]
struct LinearParams {
line: Line,
bit_unpacker: BitUnpacker,
}
impl BinarySerializable for LinearParams {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
self.line.serialize(writer)?;
self.bit_unpacker.bit_width().serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let line = Line::deserialize(reader)?;
let bit_width = u8::deserialize(reader)?;
Ok(Self {
line,
bit_unpacker: BitUnpacker::new(bit_width),
})
}
}
impl FastFieldCodec for LinearCodec {
const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::Linear;
type Reader = LinearReader;
/// Opens a fast field given a file.
fn open_from_bytes(mut data: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader> {
let linear_params = LinearParams::deserialize(&mut data)?;
Ok(LinearReader {
data,
linear_params,
header,
})
}
/// Creates a new fast field serializer.
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()> {
assert_eq!(column.min_value(), 0);
let line = Line::train(column);
let max_offset_from_line = column
.iter()
.enumerate()
.map(|(pos, actual_value)| {
let calculated_value = line.eval(pos as u32);
actual_value.wrapping_sub(calculated_value)
})
.max()
.unwrap();
let num_bits = compute_num_bits(max_offset_from_line);
let linear_params = LinearParams {
line,
bit_unpacker: BitUnpacker::new(num_bits),
};
linear_params.serialize(write)?;
let mut bit_packer = BitPacker::new();
for (pos, actual_value) in column.iter().enumerate() {
let calculated_value = line.eval(pos as u32);
let offset = actual_value.wrapping_sub(calculated_value);
bit_packer.write(offset, num_bits, write)?;
}
bit_packer.close(write)?;
Ok(())
}
/// estimation for linear interpolation is hard because, you don't know
/// where the local maxima for the deviation of the calculated value are and
/// the offset to shift all values to >=0 is also unknown.
#[allow(clippy::question_mark)]
fn estimate(column: &dyn Column) -> Option<f32> {
if column.num_vals() < 3 {
return None; // disable compressor for this case
}
let limit_num_vals = column.num_vals().min(100_000);
let num_samples = 100;
let step_size = (limit_num_vals / num_samples).max(1); // 20 samples
let mut sample_positions_and_values: Vec<_> = Vec::new();
for (pos, val) in column.iter().enumerate().step_by(step_size as usize) {
sample_positions_and_values.push((pos as u64, val));
}
let line = Line::estimate(&sample_positions_and_values);
let estimated_bit_width = sample_positions_and_values
.into_iter()
.map(|(pos, actual_value)| {
let interpolated_val = line.eval(pos as u32);
actual_value.wrapping_sub(interpolated_val)
})
.map(|diff| ((diff as f32 * 1.5) * 2.0) as u64)
.map(compute_num_bits)
.max()
.unwrap_or(0);
// Extrapolate to whole column
let num_bits = (estimated_bit_width as u64 * column.num_vals() as u64) + 64;
let num_bits_uncompressed = 64 * column.num_vals();
Some(num_bits as f32 / num_bits_uncompressed as f32)
}
}
#[cfg(test)]
mod tests {
use rand::RngCore;
use super::*;
use crate::tests::get_codec_test_datasets;
fn create_and_validate(data: &[u64], name: &str) -> Option<(f32, f32)> {
crate::tests::create_and_validate::<LinearCodec>(data, name)
}
#[test]
fn test_compression() {
let data = (10..=6_000_u64).collect::<Vec<_>>();
let (estimate, actual_compression) =
create_and_validate(&data, "simple monotonically large").unwrap();
assert_le!(actual_compression, 0.001);
assert_le!(estimate, 0.02);
}
#[test]
fn test_with_codec_datasets() {
let data_sets = get_codec_test_datasets();
for (mut data, name) in data_sets {
create_and_validate(&data, name);
data.reverse();
create_and_validate(&data, name);
}
}
#[test]
fn linear_interpol_fast_field_test_large_amplitude() {
let data = vec![
i64::MAX as u64 / 2,
i64::MAX as u64 / 3,
i64::MAX as u64 / 2,
];
create_and_validate(&data, "large amplitude");
}
#[test]
fn overflow_error_test() {
let data = vec![1572656989877777, 1170935903116329, 720575940379279, 0];
create_and_validate(&data, "overflow test");
}
#[test]
fn linear_interpol_fast_concave_data() {
let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
create_and_validate(&data, "concave data");
}
#[test]
fn linear_interpol_fast_convex_data() {
let data = vec![0, 40, 60, 70, 75, 77];
create_and_validate(&data, "convex data");
}
#[test]
fn linear_interpol_fast_field_test_simple() {
let data = (10..=20_u64).collect::<Vec<_>>();
create_and_validate(&data, "simple monotonically");
}
#[test]
fn linear_interpol_fast_field_rand() {
let mut rng = rand::thread_rng();
for _ in 0..50 {
let mut data = (0..10_000).map(|_| rng.next_u64()).collect::<Vec<_>>();
create_and_validate(&data, "random");
data.reverse();
create_and_validate(&data, "random");
}
}
}

300
fastfield_codecs/src/linearinterpol.rs
View File

@@ -1,300 +0,0 @@
use std::io::{self, Read, Write};
use std::ops::Sub;
use common::{BinarySerializable, FixedSize};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::{FastFieldCodecReader, FastFieldCodecSerializer, FastFieldDataAccess, FastFieldStats};
/// Depending on the field type, a different
/// fast field is required.
#[derive(Clone)]
pub struct LinearInterpolFastFieldReader {
bit_unpacker: BitUnpacker,
pub footer: LinearInterpolFooter,
pub slope: f32,
}
#[derive(Clone, Debug)]
pub struct LinearInterpolFooter {
pub relative_max_value: u64,
pub offset: u64,
pub first_val: u64,
pub last_val: u64,
pub num_vals: u64,
pub min_value: u64,
pub max_value: u64,
}
impl BinarySerializable for LinearInterpolFooter {
fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
self.relative_max_value.serialize(write)?;
self.offset.serialize(write)?;
self.first_val.serialize(write)?;
self.last_val.serialize(write)?;
self.num_vals.serialize(write)?;
self.min_value.serialize(write)?;
self.max_value.serialize(write)?;
Ok(())
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<LinearInterpolFooter> {
Ok(LinearInterpolFooter {
relative_max_value: u64::deserialize(reader)?,
offset: u64::deserialize(reader)?,
first_val: u64::deserialize(reader)?,
last_val: u64::deserialize(reader)?,
num_vals: u64::deserialize(reader)?,
min_value: u64::deserialize(reader)?,
max_value: u64::deserialize(reader)?,
})
}
}
impl FixedSize for LinearInterpolFooter {
const SIZE_IN_BYTES: usize = 56;
}
impl FastFieldCodecReader for LinearInterpolFastFieldReader {
/// Opens a fast field given a file.
fn open_from_bytes(bytes: &[u8]) -> io::Result<Self> {
let (_data, mut footer) = bytes.split_at(bytes.len() - LinearInterpolFooter::SIZE_IN_BYTES);
let footer = LinearInterpolFooter::deserialize(&mut footer)?;
let slope = get_slope(footer.first_val, footer.last_val, footer.num_vals);
let num_bits = compute_num_bits(footer.relative_max_value);
let bit_unpacker = BitUnpacker::new(num_bits);
Ok(LinearInterpolFastFieldReader {
bit_unpacker,
footer,
slope,
})
}
#[inline]
fn get_u64(&self, doc: u64, data: &[u8]) -> u64 {
let calculated_value = get_calculated_value(self.footer.first_val, doc, self.slope);
(calculated_value + self.bit_unpacker.get(doc, data)) - self.footer.offset
}
#[inline]
fn min_value(&self) -> u64 {
self.footer.min_value
}
#[inline]
fn max_value(&self) -> u64 {
self.footer.max_value
}
}
/// Fastfield serializer, which tries to guess values by linear interpolation
/// and stores the difference bitpacked.
pub struct LinearInterpolFastFieldSerializer {}
#[inline]
fn get_slope(first_val: u64, last_val: u64, num_vals: u64) -> f32 {
if num_vals <= 1 {
return 0.0;
}
// We calculate the slope with f64 high precision and use the result in lower precision f32
// This is done in order to handle estimations for very large values like i64::MAX
((last_val as f64 - first_val as f64) / (num_vals as u64 - 1) as f64) as f32
}
#[inline]
fn get_calculated_value(first_val: u64, pos: u64, slope: f32) -> u64 {
first_val + (pos as f32 * slope) as u64
}
impl FastFieldCodecSerializer for LinearInterpolFastFieldSerializer {
const NAME: &'static str = "LinearInterpol";
const ID: u8 = 2;
/// Creates a new fast field serializer.
fn serialize(
write: &mut impl Write,
fastfield_accessor: &dyn FastFieldDataAccess,
stats: FastFieldStats,
data_iter: impl Iterator<Item = u64>,
data_iter1: impl Iterator<Item = u64>,
) -> io::Result<()> {
assert!(stats.min_value <= stats.max_value);
let first_val = fastfield_accessor.get_val(0);
let last_val = fastfield_accessor.get_val(stats.num_vals as u64 - 1);
let slope = get_slope(first_val, last_val, stats.num_vals);
// calculate offset to ensure all values are positive
let mut offset = 0;
let mut rel_positive_max = 0;
for (pos, actual_value) in data_iter1.enumerate() {
let calculated_value = get_calculated_value(first_val, pos as u64, slope);
if calculated_value > actual_value {
// negative value we need to apply an offset
// we ignore negative values in the max value calculation, because negative values
// will be offset to 0
offset = offset.max(calculated_value - actual_value);
} else {
// positive value no offset reuqired
rel_positive_max = rel_positive_max.max(actual_value - calculated_value);
}
}
// rel_positive_max will be adjusted by offset
let relative_max_value = rel_positive_max + offset;
let num_bits = compute_num_bits(relative_max_value);
let mut bit_packer = BitPacker::new();
for (pos, val) in data_iter.enumerate() {
let calculated_value = get_calculated_value(first_val, pos as u64, slope);
let diff = (val + offset) - calculated_value;
bit_packer.write(diff, num_bits, write)?;
}
bit_packer.close(write)?;
let footer = LinearInterpolFooter {
relative_max_value,
offset,
first_val,
last_val,
num_vals: stats.num_vals,
min_value: stats.min_value,
max_value: stats.max_value,
};
footer.serialize(write)?;
Ok(())
}
fn is_applicable(
_fastfield_accessor: &impl FastFieldDataAccess,
stats: FastFieldStats,
) -> bool {
if stats.num_vals < 3 {
return false; // disable compressor for this case
}
// On serialisation the offset is added to the actual value.
// We need to make sure this won't run into overflow calculation issues.
// For this we take the maximum theroretical offset and add this to the max value.
// If this doesn't overflow the algortihm should be fine
let theorethical_maximum_offset = stats.max_value - stats.min_value;
if stats
.max_value
.checked_add(theorethical_maximum_offset)
.is_none()
{
return false;
}
true
}
/// estimation for linear interpolation is hard because, you don't know
/// where the local maxima for the deviation of the calculated value are and
/// the offset to shift all values to >=0 is also unknown.
fn estimate(fastfield_accessor: &impl FastFieldDataAccess, stats: FastFieldStats) -> f32 {
let first_val = fastfield_accessor.get_val(0);
let last_val = fastfield_accessor.get_val(stats.num_vals as u64 - 1);
let slope = get_slope(first_val, last_val, stats.num_vals);
// let's sample at 0%, 5%, 10% .. 95%, 100%
let num_vals = stats.num_vals as f32 / 100.0;
let sample_positions = (0..20)
.map(|pos| (num_vals * pos as f32 * 5.0) as usize)
.collect::<Vec<_>>();
let max_distance = sample_positions
.iter()
.map(|pos| {
let calculated_value = get_calculated_value(first_val, *pos as u64, slope);
let actual_value = fastfield_accessor.get_val(*pos as u64);
distance(calculated_value, actual_value)
})
.max()
.unwrap_or(0);
// the theory would be that we don't have the actual max_distance, but we are close within
// 50% threshold.
// It is multiplied by 2 because in a log case scenario the line would be as much above as
// below. So the offset would = max_distance
//
let relative_max_value = (max_distance as f32 * 1.5) * 2.0;
let num_bits = compute_num_bits(relative_max_value as u64) as u64 * stats.num_vals as u64
+ LinearInterpolFooter::SIZE_IN_BYTES as u64;
let num_bits_uncompressed = 64 * stats.num_vals;
num_bits as f32 / num_bits_uncompressed as f32
}
}
#[inline]
fn distance<T: Sub<Output = T> + Ord>(x: T, y: T) -> T {
if x < y {
y - x
} else {
x - y
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::tests::get_codec_test_data_sets;
fn create_and_validate(data: &[u64], name: &str) -> (f32, f32) {
crate::tests::create_and_validate::<
LinearInterpolFastFieldSerializer,
LinearInterpolFastFieldReader,
>(data, name)
}
#[test]
fn test_compression() {
let data = (10..=6_000_u64).collect::<Vec<_>>();
let (estimate, actual_compression) =
create_and_validate(&data, "simple monotonically large");
assert!(actual_compression < 0.01);
assert!(estimate < 0.01);
}
#[test]
fn test_with_codec_data_sets() {
let data_sets = get_codec_test_data_sets();
for (mut data, name) in data_sets {
create_and_validate(&data, name);
data.reverse();
create_and_validate(&data, name);
}
}
#[test]
fn linear_interpol_fast_field_test_large_amplitude() {
let data = vec![
i64::MAX as u64 / 2,
i64::MAX as u64 / 3,
i64::MAX as u64 / 2,
];
create_and_validate(&data, "large amplitude");
}
#[test]
fn linear_interpol_fast_concave_data() {
let data = vec![0, 1, 2, 5, 8, 10, 20, 50];
create_and_validate(&data, "concave data");
}
#[test]
fn linear_interpol_fast_convex_data() {
let data = vec![0, 40, 60, 70, 75, 77];
create_and_validate(&data, "convex data");
}
#[test]
fn linear_interpol_fast_field_test_simple() {
let data = (10..=20_u64).collect::<Vec<_>>();
create_and_validate(&data, "simple monotonically");
}
#[test]
fn linear_interpol_fast_field_rand() {
for _ in 0..5000 {
let mut data = (0..50).map(|_| rand::random::<u64>()).collect::<Vec<_>>();
create_and_validate(&data, "random");
data.reverse();
create_and_validate(&data, "random");
}
}
}

208
fastfield_codecs/src/main.rs
View File

@@ -1,52 +1,170 @@
#[macro_use]
extern crate prettytable;
use fastfield_codecs::linearinterpol::LinearInterpolFastFieldSerializer;
use fastfield_codecs::multilinearinterpol::MultiLinearInterpolFastFieldSerializer;
use fastfield_codecs::{FastFieldCodecSerializer, FastFieldStats};
use std::collections::HashSet;
use std::env;
use std::io::BufRead;
use std::net::{IpAddr, Ipv6Addr};
use std::str::FromStr;
use fastfield_codecs::{open_u128, serialize_u128, Column, FastFieldCodecType, VecColumn};
use itertools::Itertools;
use measure_time::print_time;
use ownedbytes::OwnedBytes;
use prettytable::{Cell, Row, Table};
fn print_set_stats(ip_addrs: &[u128]) {
println!("NumIps\t{}", ip_addrs.len());
let ip_addr_set: HashSet<u128> = ip_addrs.iter().cloned().collect();
println!("NumUniqueIps\t{}", ip_addr_set.len());
let ratio_unique = ip_addr_set.len() as f64 / ip_addrs.len() as f64;
println!("RatioUniqueOverTotal\t{ratio_unique:.4}");
// histogram
let mut ip_addrs = ip_addrs.to_vec();
ip_addrs.sort();
let mut cnts: Vec<usize> = ip_addrs
.into_iter()
.dedup_with_count()
.map(|(cnt, _)| cnt)
.collect();
cnts.sort();
let top_256_cnt: usize = cnts.iter().rev().take(256).sum();
let top_128_cnt: usize = cnts.iter().rev().take(128).sum();
let top_64_cnt: usize = cnts.iter().rev().take(64).sum();
let top_8_cnt: usize = cnts.iter().rev().take(8).sum();
let total: usize = cnts.iter().sum();
println!("{}", total);
println!("{}", top_256_cnt);
println!("{}", top_128_cnt);
println!("Percentage Top8 {:02}", top_8_cnt as f32 / total as f32);
println!("Percentage Top64 {:02}", top_64_cnt as f32 / total as f32);
println!("Percentage Top128 {:02}", top_128_cnt as f32 / total as f32);
println!("Percentage Top256 {:02}", top_256_cnt as f32 / total as f32);
let mut cnts: Vec<(usize, usize)> = cnts.into_iter().dedup_with_count().collect();
cnts.sort_by(|a, b| {
if a.1 == b.1 {
a.0.cmp(&b.0)
} else {
b.1.cmp(&a.1)
}
});
}
fn ip_dataset() -> Vec<u128> {
let mut ip_addr_v4 = 0;
let stdin = std::io::stdin();
let ip_addrs: Vec<u128> = stdin
.lock()
.lines()
.flat_map(|line| {
let line = line.unwrap();
let line = line.trim();
let ip_addr = IpAddr::from_str(line.trim()).ok()?;
if ip_addr.is_ipv4() {
ip_addr_v4 += 1;
}
let ip_addr_v6: Ipv6Addr = match ip_addr {
IpAddr::V4(v4) => v4.to_ipv6_mapped(),
IpAddr::V6(v6) => v6,
};
Some(ip_addr_v6)
})
.map(|ip_v6| u128::from_be_bytes(ip_v6.octets()))
.collect();
println!("IpAddrsAny\t{}", ip_addrs.len());
println!("IpAddrsV4\t{}", ip_addr_v4);
ip_addrs
}
fn bench_ip() {
let dataset = ip_dataset();
print_set_stats(&dataset);
// Chunks
{
let mut data = vec![];
for dataset in dataset.chunks(500_000) {
serialize_u128(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
}
let compression = data.len() as f64 / (dataset.len() * 16) as f64;
println!("Compression 50_000 chunks {:.4}", compression);
println!(
"Num Bits per elem {:.2}",
(data.len() * 8) as f32 / dataset.len() as f32
);
}
let mut data = vec![];
{
print_time!("creation");
serialize_u128(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
}
let compression = data.len() as f64 / (dataset.len() * 16) as f64;
println!("Compression {:.2}", compression);
println!(
"Num Bits per elem {:.2}",
(data.len() * 8) as f32 / dataset.len() as f32
);
let decompressor = open_u128::<u128>(OwnedBytes::new(data)).unwrap();
// Sample some ranges
let mut doc_values = Vec::new();
for value in dataset.iter().take(1110).skip(1100).cloned() {
doc_values.clear();
print_time!("get range");
decompressor.get_docids_for_value_range(
value..=value,
0..decompressor.num_vals(),
&mut doc_values,
);
println!("{:?}", doc_values.len());
}
}
fn main() {
if env::args().nth(1).unwrap() == "bench_ip" {
bench_ip();
return;
}
let mut table = Table::new();
// Add a row per time
table.add_row(row!["", "Compression Ratio", "Compression Estimation"]);
for (data, data_set_name) in get_codec_test_data_sets() {
let mut results = vec![];
let res = serialize_with_codec::<LinearInterpolFastFieldSerializer>(&data);
results.push(res);
let res = serialize_with_codec::<MultiLinearInterpolFastFieldSerializer>(&data);
results.push(res);
let res = serialize_with_codec::<fastfield_codecs::bitpacked::BitpackedFastFieldSerializer>(
&data,
);
results.push(res);
// let best_estimation_codec = results
//.iter()
//.min_by(|res1, res2| res1.partial_cmp(&res2).unwrap())
//.unwrap();
let results: Vec<(f32, f32, FastFieldCodecType)> = [
serialize_with_codec(&data, FastFieldCodecType::Bitpacked),
serialize_with_codec(&data, FastFieldCodecType::Linear),
serialize_with_codec(&data, FastFieldCodecType::BlockwiseLinear),
]
.into_iter()
.flatten()
.collect();
let best_compression_ratio_codec = results
.iter()
.min_by(|res1, res2| res1.partial_cmp(res2).unwrap())
.min_by(|&res1, &res2| res1.partial_cmp(res2).unwrap())
.cloned()
.unwrap();
table.add_row(Row::new(vec![Cell::new(data_set_name).style_spec("Bbb")]));
for (is_applicable, est, comp, name) in results {
let (est_cell, ratio_cell) = if !is_applicable {
("Codec Disabled".to_string(), "".to_string())
} else {
(est.to_string(), comp.to_string())
};
for (est, comp, codec_type) in results {
let est_cell = est.to_string();
let ratio_cell = comp.to_string();
let style = if comp == best_compression_ratio_codec.1 {
"Fb"
} else {
""
};
table.add_row(Row::new(vec![
Cell::new(name).style_spec("bFg"),
Cell::new(&format!("{codec_type:?}")).style_spec("bFg"),
Cell::new(&ratio_cell).style_spec(style),
Cell::new(&est_cell).style_spec(""),
]));
@@ -91,34 +209,14 @@ pub fn get_codec_test_data_sets() -> Vec<(Vec<u64>, &'static str)> {
data_and_names
}
pub fn serialize_with_codec<S: FastFieldCodecSerializer>(
pub fn serialize_with_codec(
data: &[u64],
) -> (bool, f32, f32, &'static str) {
let is_applicable = S::is_applicable(&data, stats_from_vec(data));
if !is_applicable {
return (false, 0.0, 0.0, S::NAME);
}
let estimation = S::estimate(&data, stats_from_vec(data));
let mut out = vec![];
S::serialize(
&mut out,
&data,
stats_from_vec(data),
data.iter().cloned(),
data.iter().cloned(),
)
.unwrap();
let actual_compression = out.len() as f32 / (data.len() * 8) as f32;
(true, estimation, actual_compression, S::NAME)
}
pub fn stats_from_vec(data: &[u64]) -> FastFieldStats {
let min_value = data.iter().cloned().min().unwrap_or(0);
let max_value = data.iter().cloned().max().unwrap_or(0);
FastFieldStats {
min_value,
max_value,
num_vals: data.len() as u64,
}
codec_type: FastFieldCodecType,
) -> Option<(f32, f32, FastFieldCodecType)> {
let col = VecColumn::from(data);
let estimation = fastfield_codecs::estimate(&col, codec_type)?;
let mut out = Vec::new();
fastfield_codecs::serialize(&col, &mut out, &[codec_type]).ok()?;
let actual_compression = out.len() as f32 / (col.num_vals() * 8) as f32;
Some((estimation, actual_compression, codec_type))
}

267
fastfield_codecs/src/monotonic_mapping.rs Normal file
View File

@@ -0,0 +1,267 @@
use std::marker::PhantomData;
use fastdivide::DividerU64;
use ordered_float::NotNan;
use crate::MonotonicallyMappableToU128;
/// Monotonic maps a value to u64 value space.
/// Monotonic mapping enables `PartialOrd` on u64 space without conversion to original space.
pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
/// Converts a value to u64.
///
/// Internally all fast field values are encoded as u64.
fn to_u64(self) -> u64;
/// Converts a value from u64
///
/// Internally all fast field values are encoded as u64.
/// **Note: To be used for converting encoded Term, Posting values.**
fn from_u64(val: u64) -> Self;
}
/// Values need to be strictly monotonic mapped to a `Internal` value (u64 or u128) that can be
/// used in fast field codecs.
///
/// The monotonic mapping is required so that `PartialOrd` can be used on `Internal` without
/// converting to `External`.
///
/// All strictly monotonic functions are invertible because they are guaranteed to have a one-to-one
/// mapping from their range to their domain. The `inverse` method is required when opening a codec,
/// so a value can be converted back to its original domain (e.g. ip address or f64) from its
/// internal representation.
pub trait StrictlyMonotonicFn<External, Internal> {
/// Strictly monotonically maps the value from External to Internal.
fn mapping(&self, inp: External) -> Internal;
/// Inverse of `mapping`. Maps the value from Internal to External.
fn inverse(&self, out: Internal) -> External;
}
/// Inverts a strictly monotonic mapping from `StrictlyMonotonicFn<A, B>` to
/// `StrictlyMonotonicFn<B, A>`.
///
/// # Warning
///
/// This type comes with a footgun. A type being strictly monotonic does not impose that the inverse
/// mapping is strictly monotonic over the entire space External. e.g. a -> a * 2. Use at your own
/// risks.
pub(crate) struct StrictlyMonotonicMappingInverter<T> {
orig_mapping: T,
}
impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
fn from(orig_mapping: T) -> Self {
Self { orig_mapping }
}
}
impl<From, To, T> StrictlyMonotonicFn<To, From> for StrictlyMonotonicMappingInverter<T>
where T: StrictlyMonotonicFn<From, To>
{
fn mapping(&self, val: To) -> From {
self.orig_mapping.inverse(val)
}
fn inverse(&self, val: From) -> To {
self.orig_mapping.mapping(val)
}
}
/// Applies the strictly monotonic mapping from `T` without any additional changes.
pub(crate) struct StrictlyMonotonicMappingToInternal<T> {
_phantom: PhantomData<T>,
}
impl<T> StrictlyMonotonicMappingToInternal<T> {
pub(crate) fn new() -> StrictlyMonotonicMappingToInternal<T> {
Self {
_phantom: PhantomData,
}
}
}
impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
where T: MonotonicallyMappableToU128
{
fn mapping(&self, inp: External) -> u128 {
External::to_u128(inp)
}
fn inverse(&self, out: u128) -> External {
External::from_u128(out)
}
}
impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
where T: MonotonicallyMappableToU64
{
fn mapping(&self, inp: External) -> u64 {
External::to_u64(inp)
}
fn inverse(&self, out: u64) -> External {
External::from_u64(out)
}
}
/// Mapping dividing by gcd and a base value.
///
/// The function is assumed to be only called on values divided by passed
/// gcd value. (It is necessary for the function to be monotonic.)
pub(crate) struct StrictlyMonotonicMappingToInternalGCDBaseval {
gcd_divider: DividerU64,
gcd: u64,
min_value: u64,
}
impl StrictlyMonotonicMappingToInternalGCDBaseval {
pub(crate) fn new(gcd: u64, min_value: u64) -> Self {
let gcd_divider = DividerU64::divide_by(gcd);
Self {
gcd_divider,
gcd,
min_value,
}
}
}
impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
for StrictlyMonotonicMappingToInternalGCDBaseval
{
fn mapping(&self, inp: External) -> u64 {
self.gcd_divider
.divide(External::to_u64(inp) - self.min_value)
}
fn inverse(&self, out: u64) -> External {
External::from_u64(self.min_value + out * self.gcd)
}
}
/// Strictly monotonic mapping with a base value.
pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
min_value: u64,
}
impl StrictlyMonotonicMappingToInternalBaseval {
pub(crate) fn new(min_value: u64) -> Self {
Self { min_value }
}
}
impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
for StrictlyMonotonicMappingToInternalBaseval
{
fn mapping(&self, val: External) -> u64 {
External::to_u64(val) - self.min_value
}
fn inverse(&self, val: u64) -> External {
External::from_u64(self.min_value + val)
}
}
impl MonotonicallyMappableToU64 for u64 {
fn to_u64(self) -> u64 {
self
}
fn from_u64(val: u64) -> Self {
val
}
}
impl MonotonicallyMappableToU64 for i64 {
#[inline(always)]
fn to_u64(self) -> u64 {
common::i64_to_u64(self)
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
common::u64_to_i64(val)
}
}
impl MonotonicallyMappableToU64 for bool {
#[inline(always)]
fn to_u64(self) -> u64 {
u64::from(self)
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
val > 0
}
}
// TODO remove me.
// Tantivy should refuse NaN values and work with NotNaN internally.
impl MonotonicallyMappableToU64 for f64 {
fn to_u64(self) -> u64 {
common::f64_to_u64(self)
}
fn from_u64(val: u64) -> Self {
common::u64_to_f64(val)
}
}
impl MonotonicallyMappableToU64 for ordered_float::NotNan<f64> {
fn to_u64(self) -> u64 {
common::f64_to_u64(self.into_inner())
}
fn from_u64(val: u64) -> Self {
NotNan::new(common::u64_to_f64(val)).expect("Invalid NotNaN f64 value.")
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_from_u64_pos_inf() {
let inf_as_u64 = common::f64_to_u64(f64::INFINITY);
let inf_back_to_f64 = NotNan::from_u64(inf_as_u64);
assert_eq!(inf_back_to_f64, NotNan::new(f64::INFINITY).unwrap());
}
#[test]
fn test_from_u64_neg_inf() {
let inf_as_u64 = common::f64_to_u64(-f64::INFINITY);
let inf_back_to_f64 = NotNan::from_u64(inf_as_u64);
assert_eq!(inf_back_to_f64, NotNan::new(-f64::INFINITY).unwrap());
}
#[test]
#[should_panic(expected = "Invalid NotNaN")]
fn test_from_u64_nan_panics() {
let nan_as_u64 = common::f64_to_u64(f64::NAN);
NotNan::from_u64(nan_as_u64);
}
#[test]
fn strictly_monotonic_test() {
// identity mapping
test_round_trip(&StrictlyMonotonicMappingToInternal::<u64>::new(), 100u64);
// round trip to i64
test_round_trip(&StrictlyMonotonicMappingToInternal::<i64>::new(), 100u64);
// identity mapping
test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
// base value to i64 round trip
let mapping = StrictlyMonotonicMappingToInternalBaseval::new(100);
test_round_trip::<_, _, u64>(&mapping, 100i64);
// base value and gcd to u64 round trip
let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(10, 100);
test_round_trip::<_, _, u64>(&mapping, 100u64);
}
fn test_round_trip<T: StrictlyMonotonicFn<K, L>, K: std::fmt::Debug + Eq + Copy, L>(
mapping: &T,
test_val: K,
) {
assert_eq!(mapping.inverse(mapping.mapping(test_val)), test_val);
}
}

40
fastfield_codecs/src/monotonic_mapping_u128.rs Normal file
View File

@@ -0,0 +1,40 @@
use std::net::Ipv6Addr;
/// Montonic maps a value to u128 value space
/// Monotonic mapping enables `PartialOrd` on u128 space without conversion to original space.
pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Send + Sync {
/// Converts a value to u128.
///
/// Internally all fast field values are encoded as u64.
fn to_u128(self) -> u128;
/// Converts a value from u128
///
/// Internally all fast field values are encoded as u64.
/// **Note: To be used for converting encoded Term, Posting values.**
fn from_u128(val: u128) -> Self;
}
impl MonotonicallyMappableToU128 for u128 {
fn to_u128(self) -> u128 {
self
}
fn from_u128(val: u128) -> Self {
val
}
}
impl MonotonicallyMappableToU128 for Ipv6Addr {
fn to_u128(self) -> u128 {
ip_to_u128(self)
}
fn from_u128(val: u128) -> Self {
Ipv6Addr::from(val.to_be_bytes())
}
}
fn ip_to_u128(ip_addr: Ipv6Addr) -> u128 {
u128::from_be_bytes(ip_addr.octets())
}

427
fastfield_codecs/src/multilinearinterpol.rs
View File

@@ -1,427 +0,0 @@
//! MultiLinearInterpol compressor uses linear interpolation to guess a values and stores the
//! offset, but in blocks of 512.
//!
//! With a CHUNK_SIZE of 512 and 29 byte metadata per block, we get a overhead for metadata of 232 /
//! 512 = 0,45 bits per element. The additional space required per element in a block is the the
//! maximum deviation of the linear interpolation estimation function.
//!
//! E.g. if the maximum deviation of an element is 12, all elements cost 4bits.
//!
//! Size per block:
//! Num Elements * Maximum Deviation from Interpolation + 29 Byte Metadata
use std::io::{self, Read, Write};
use std::ops::Sub;
use common::{BinarySerializable, CountingWriter, DeserializeFrom};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::{FastFieldCodecReader, FastFieldCodecSerializer, FastFieldDataAccess, FastFieldStats};
const CHUNK_SIZE: u64 = 512;
/// Depending on the field type, a different
/// fast field is required.
#[derive(Clone)]
pub struct MultiLinearInterpolFastFieldReader {
pub footer: MultiLinearInterpolFooter,
}
#[derive(Clone, Debug, Default)]
struct Function {
// The offset in the data is required, because we have diffrent bit_widths per block
data_start_offset: u64,
// start_pos in the block will be CHUNK_SIZE * BLOCK_NUM
start_pos: u64,
// only used during serialization, 0 after deserialization
end_pos: u64,
// only used during serialization, 0 after deserialization
value_start_pos: u64,
// only used during serialization, 0 after deserialization
value_end_pos: u64,
slope: f32,
// The offset so that all values are positive when writing them
positive_val_offset: u64,
num_bits: u8,
bit_unpacker: BitUnpacker,
}
impl Function {
fn calc_slope(&mut self) {
let num_vals = self.end_pos - self.start_pos;
self.slope = get_slope(self.value_start_pos, self.value_end_pos, num_vals);
}
// split the interpolation into two function, change self and return the second split
fn split(&mut self, split_pos: u64, split_pos_value: u64) -> Function {
let mut new_function = Function {
start_pos: split_pos,
end_pos: self.end_pos,
value_start_pos: split_pos_value,
value_end_pos: self.value_end_pos,
..Default::default()
};
new_function.calc_slope();
self.end_pos = split_pos;
self.value_end_pos = split_pos_value;
self.calc_slope();
new_function
}
}
impl BinarySerializable for Function {
fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
self.data_start_offset.serialize(write)?;
self.value_start_pos.serialize(write)?;
self.positive_val_offset.serialize(write)?;
self.slope.serialize(write)?;
self.num_bits.serialize(write)?;
Ok(())
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<Function> {
let data_start_offset = u64::deserialize(reader)?;
let value_start_pos = u64::deserialize(reader)?;
let offset = u64::deserialize(reader)?;
let slope = f32::deserialize(reader)?;
let num_bits = u8::deserialize(reader)?;
let interpolation = Function {
data_start_offset,
value_start_pos,
positive_val_offset: offset,
num_bits,
bit_unpacker: BitUnpacker::new(num_bits),
slope,
..Default::default()
};
Ok(interpolation)
}
}
#[derive(Clone, Debug)]
pub struct MultiLinearInterpolFooter {
pub num_vals: u64,
pub min_value: u64,
pub max_value: u64,
interpolations: Vec<Function>,
}
impl BinarySerializable for MultiLinearInterpolFooter {
fn serialize<W: Write>(&self, write: &mut W) -> io::Result<()> {
let mut out = vec![];
self.num_vals.serialize(&mut out)?;
self.min_value.serialize(&mut out)?;
self.max_value.serialize(&mut out)?;
self.interpolations.serialize(&mut out)?;
write.write_all(&out)?;
(out.len() as u32).serialize(write)?;
Ok(())
}
fn deserialize<R: Read>(reader: &mut R) -> io::Result<MultiLinearInterpolFooter> {
let mut footer = MultiLinearInterpolFooter {
num_vals: u64::deserialize(reader)?,
min_value: u64::deserialize(reader)?,
max_value: u64::deserialize(reader)?,
interpolations: Vec::<Function>::deserialize(reader)?,
};
for (num, interpol) in footer.interpolations.iter_mut().enumerate() {
interpol.start_pos = CHUNK_SIZE * num as u64;
}
Ok(footer)
}
}
#[inline]
fn get_interpolation_position(doc: u64) -> usize {
let index = doc / CHUNK_SIZE;
index as usize
}
#[inline]
fn get_interpolation_function(doc: u64, interpolations: &[Function]) -> &Function {
&interpolations[get_interpolation_position(doc)]
}
impl FastFieldCodecReader for MultiLinearInterpolFastFieldReader {
/// Opens a fast field given a file.
fn open_from_bytes(bytes: &[u8]) -> io::Result<Self> {
let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
let (_data, mut footer) = bytes.split_at(bytes.len() - (4 + footer_len) as usize);
let footer = MultiLinearInterpolFooter::deserialize(&mut footer)?;
Ok(MultiLinearInterpolFastFieldReader { footer })
}
#[inline]
fn get_u64(&self, doc: u64, data: &[u8]) -> u64 {
let interpolation = get_interpolation_function(doc, &self.footer.interpolations);
let doc = doc - interpolation.start_pos;
let calculated_value =
get_calculated_value(interpolation.value_start_pos, doc, interpolation.slope);
let diff = interpolation
.bit_unpacker
.get(doc, &data[interpolation.data_start_offset as usize..]);
(calculated_value + diff) - interpolation.positive_val_offset
}
#[inline]
fn min_value(&self) -> u64 {
self.footer.min_value
}
#[inline]
fn max_value(&self) -> u64 {
self.footer.max_value
}
}
#[inline]
fn get_slope(first_val: u64, last_val: u64, num_vals: u64) -> f32 {
((last_val as f64 - first_val as f64) / (num_vals as u64 - 1) as f64) as f32
}
#[inline]
fn get_calculated_value(first_val: u64, pos: u64, slope: f32) -> u64 {
(first_val as i64 + (pos as f32 * slope) as i64) as u64
}
/// Same as LinearInterpolFastFieldSerializer, but working on chunks of CHUNK_SIZE elements.
pub struct MultiLinearInterpolFastFieldSerializer {}
impl FastFieldCodecSerializer for MultiLinearInterpolFastFieldSerializer {
const NAME: &'static str = "MultiLinearInterpol";
const ID: u8 = 3;
/// Creates a new fast field serializer.
fn serialize(
write: &mut impl Write,
fastfield_accessor: &dyn FastFieldDataAccess,
stats: FastFieldStats,
data_iter: impl Iterator<Item = u64>,
_data_iter1: impl Iterator<Item = u64>,
) -> io::Result<()> {
assert!(stats.min_value <= stats.max_value);
let first_val = fastfield_accessor.get_val(0);
let last_val = fastfield_accessor.get_val(stats.num_vals as u64 - 1);
let mut first_function = Function {
end_pos: stats.num_vals,
value_start_pos: first_val,
value_end_pos: last_val,
..Default::default()
};
first_function.calc_slope();
let mut interpolations = vec![first_function];
// Since we potentially apply multiple passes over the data, the data is cached.
// Multiple iteration can be expensive (merge with index sorting can add lot of overhead per
// iteration)
let data = data_iter.collect::<Vec<_>>();
//// let's split this into chunks of CHUNK_SIZE
for data_pos in (0..data.len() as u64).step_by(CHUNK_SIZE as usize).skip(1) {
let new_fun = {
let current_interpolation = interpolations.last_mut().unwrap();
current_interpolation.split(data_pos, data[data_pos as usize])
};
interpolations.push(new_fun);
}
// calculate offset and max (-> numbits) for each function
for interpolation in &mut interpolations {
let mut offset = 0;
let mut rel_positive_max = 0;
for (pos, actual_value) in data
[interpolation.start_pos as usize..interpolation.end_pos as usize]
.iter()
.cloned()
.enumerate()
{
let calculated_value = get_calculated_value(
interpolation.value_start_pos,
pos as u64,
interpolation.slope,
);
if calculated_value > actual_value {
// negative value we need to apply an offset
// we ignore negative values in the max value calculation, because negative
// values will be offset to 0
offset = offset.max(calculated_value - actual_value);
} else {
// positive value no offset reuqired
rel_positive_max = rel_positive_max.max(actual_value - calculated_value);
}
}
interpolation.positive_val_offset = offset;
interpolation.num_bits = compute_num_bits(rel_positive_max + offset);
}
let mut bit_packer = BitPacker::new();
let write = &mut CountingWriter::wrap(write);
for interpolation in &mut interpolations {
interpolation.data_start_offset = write.written_bytes();
let num_bits = interpolation.num_bits;
for (pos, actual_value) in data
[interpolation.start_pos as usize..interpolation.end_pos as usize]
.iter()
.cloned()
.enumerate()
{
let calculated_value = get_calculated_value(
interpolation.value_start_pos,
pos as u64,
interpolation.slope,
);
let diff = (actual_value + interpolation.positive_val_offset) - calculated_value;
bit_packer.write(diff, num_bits, write)?;
}
bit_packer.flush(write)?;
}
bit_packer.close(write)?;
let footer = MultiLinearInterpolFooter {
num_vals: stats.num_vals,
min_value: stats.min_value,
max_value: stats.max_value,
interpolations,
};
footer.serialize(write)?;
Ok(())
}
fn is_applicable(
_fastfield_accessor: &impl FastFieldDataAccess,
stats: FastFieldStats,
) -> bool {
if stats.num_vals < 5_000 {
return false;
}
// On serialization the offset is added to the actual value.
// We need to make sure this won't run into overflow calculation issues.
// For this we take the maximum theroretical offset and add this to the max value.
// If this doesn't overflow the algortihm should be fine
let theorethical_maximum_offset = stats.max_value - stats.min_value;
if stats
.max_value
.checked_add(theorethical_maximum_offset)
.is_none()
{
return false;
}
true
}
/// estimation for linear interpolation is hard because, you don't know
/// where the local maxima are for the deviation of the calculated value and
/// the offset is also unknown.
fn estimate(fastfield_accessor: &impl FastFieldDataAccess, stats: FastFieldStats) -> f32 {
let first_val_in_first_block = fastfield_accessor.get_val(0);
let last_elem_in_first_chunk = CHUNK_SIZE.min(stats.num_vals);
let last_val_in_first_block =
fastfield_accessor.get_val(last_elem_in_first_chunk as u64 - 1);
let slope = get_slope(
first_val_in_first_block,
last_val_in_first_block,
stats.num_vals,
);
// let's sample at 0%, 5%, 10% .. 95%, 100%, but for the first block only
let sample_positions = (0..20)
.map(|pos| (last_elem_in_first_chunk as f32 / 100.0 * pos as f32 * 5.0) as usize)
.collect::<Vec<_>>();
let max_distance = sample_positions
.iter()
.map(|pos| {
let calculated_value =
get_calculated_value(first_val_in_first_block, *pos as u64, slope);
let actual_value = fastfield_accessor.get_val(*pos as u64);
distance(calculated_value, actual_value)
})
.max()
.unwrap();
// Estimate one block and extrapolate the cost to all blocks.
// the theory would be that we don't have the actual max_distance, but we are close within
// 50% threshold.
// It is multiplied by 2 because in a log case scenario the line would be as much above as
// below. So the offset would = max_distance
//
let relative_max_value = (max_distance as f32 * 1.5) * 2.0;
let num_bits = compute_num_bits(relative_max_value as u64) as u64 * stats.num_vals as u64
// function metadata per block
+ 29 * (stats.num_vals / CHUNK_SIZE);
let num_bits_uncompressed = 64 * stats.num_vals;
num_bits as f32 / num_bits_uncompressed as f32
}
}
fn distance<T: Sub<Output = T> + Ord>(x: T, y: T) -> T {
if x < y {
y - x
} else {
x - y
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::tests::get_codec_test_data_sets;
fn create_and_validate(data: &[u64], name: &str) -> (f32, f32) {
crate::tests::create_and_validate::<
MultiLinearInterpolFastFieldSerializer,
MultiLinearInterpolFastFieldReader,
>(data, name)
}
#[test]
fn test_compression() {
let data = (10..=6_000_u64).collect::<Vec<_>>();
let (estimate, actual_compression) =
create_and_validate(&data, "simple monotonically large");
assert!(actual_compression < 0.2);
assert!(estimate < 0.20);
assert!(estimate > 0.15);
assert!(actual_compression > 0.01);
}
#[test]
fn test_with_codec_data_sets() {
let data_sets = get_codec_test_data_sets();
for (mut data, name) in data_sets {
create_and_validate(&data, name);
data.reverse();
create_and_validate(&data, name);
}
}
#[test]
fn test_simple() {
let data = (10..=20_u64).collect::<Vec<_>>();
create_and_validate(&data, "simple monotonically");
}
#[test]
fn border_cases_1() {
let data = (0..1024).collect::<Vec<_>>();
create_and_validate(&data, "border case");
}
#[test]
fn border_case_2() {
let data = (0..1025).collect::<Vec<_>>();
create_and_validate(&data, "border case");
}
#[test]
fn rand() {
for _ in 0..10 {
let mut data = (5_000..20_000)
.map(|_| rand::random::<u32>() as u64)
.collect::<Vec<_>>();
let _ = create_and_validate(&data, "random");
data.reverse();
create_and_validate(&data, "random");
}
}
}

454
fastfield_codecs/src/null_index/dense.rs Normal file
View File

@@ -0,0 +1,454 @@
use std::convert::TryInto;
use std::io::{self, Write};
use common::BinarySerializable;
use itertools::Itertools;
use ownedbytes::OwnedBytes;
use super::{get_bit_at, set_bit_at};
/// For the `DenseCodec`, `data` which contains the encoded blocks.
/// Each block consists of [u8; 12]. The first 8 bytes is a bitvec for 64 elements.
/// The last 4 bytes are the offset, the number of set bits so far.
///
/// When translating the original index to a dense index, the correct block can be computed
/// directly `orig_idx/64`. Inside the block the position is `orig_idx%64`.
///
/// When translating a dense index to the original index, we can use the offset to find the correct
/// block. Direct computation is not possible, but we can employ a linear or binary search.
#[derive(Clone)]
pub struct DenseCodec {
// data consists of blocks of 64 bits.
//
// The format is &[(u64, u32)]
// u64 is the bitvec
// u32 is the offset of the block, the number of set bits so far.
//
// At the end one block is appended, to store the number of values in the index in offset.
data: OwnedBytes,
}
const ELEMENTS_PER_BLOCK: u32 = 64;
const BLOCK_BITVEC_SIZE: usize = 8;
const BLOCK_OFFSET_SIZE: usize = 4;
const SERIALIZED_BLOCK_SIZE: usize = BLOCK_BITVEC_SIZE + BLOCK_OFFSET_SIZE;
#[inline]
fn count_ones(bitvec: u64, pos_in_bitvec: u32) -> u32 {
if pos_in_bitvec == 63 {
bitvec.count_ones()
} else {
let mask = (1u64 << (pos_in_bitvec + 1)) - 1;
let masked_bitvec = bitvec & mask;
masked_bitvec.count_ones()
}
}
#[derive(Clone, Copy)]
struct DenseIndexBlock {
bitvec: u64,
offset: u32,
}
impl From<[u8; SERIALIZED_BLOCK_SIZE]> for DenseIndexBlock {
fn from(data: [u8; SERIALIZED_BLOCK_SIZE]) -> Self {
let bitvec = u64::from_le_bytes(data[..BLOCK_BITVEC_SIZE].try_into().unwrap());
let offset = u32::from_le_bytes(data[BLOCK_BITVEC_SIZE..].try_into().unwrap());
Self { bitvec, offset }
}
}
impl DenseCodec {
/// Open the DenseCodec from OwnedBytes
pub fn open(data: OwnedBytes) -> Self {
Self { data }
}
#[inline]
/// Check if value at position is not null.
pub fn exists(&self, idx: u32) -> bool {
let block_pos = idx / ELEMENTS_PER_BLOCK;
let bitvec = self.dense_index_block(block_pos).bitvec;
let pos_in_bitvec = idx % ELEMENTS_PER_BLOCK;
get_bit_at(bitvec, pos_in_bitvec)
}
#[inline]
fn dense_index_block(&self, block_pos: u32) -> DenseIndexBlock {
dense_index_block(&self.data, block_pos)
}
/// Return the number of non-null values in an index
pub fn num_non_nulls(&self) -> u32 {
let last_block = (self.data.len() / SERIALIZED_BLOCK_SIZE) - 1;
self.dense_index_block(last_block as u32).offset
}
#[inline]
/// Translate from the original index to the codec index.
pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
let block_pos = idx / ELEMENTS_PER_BLOCK;
let index_block = self.dense_index_block(block_pos);
let pos_in_block_bit_vec = idx % ELEMENTS_PER_BLOCK;
let ones_in_block = count_ones(index_block.bitvec, pos_in_block_bit_vec);
if get_bit_at(index_block.bitvec, pos_in_block_bit_vec) {
// -1 is ok, since idx does exist, so there's at least one
Some(index_block.offset + ones_in_block - 1)
} else {
None
}
}
/// Translate positions from the codec index to the original index.
///
/// # Panics
///
/// May panic if any `idx` is greater than the max codec index.
pub fn translate_codec_idx_to_original_idx<'a>(
&'a self,
iter: impl Iterator<Item = u32> + 'a,
) -> impl Iterator<Item = u32> + 'a {
let mut block_pos = 0u32;
iter.map(move |dense_idx| {
// update block_pos to limit search scope
block_pos = find_block(dense_idx, block_pos, &self.data);
let index_block = self.dense_index_block(block_pos);
// The next offset is higher than dense_idx and therefore:
// dense_idx <= offset + num_set_bits in block
let mut num_set_bits = 0;
for idx_in_bitvec in 0..ELEMENTS_PER_BLOCK {
if get_bit_at(index_block.bitvec, idx_in_bitvec) {
num_set_bits += 1;
}
if num_set_bits == (dense_idx - index_block.offset + 1) {
let orig_idx = block_pos * ELEMENTS_PER_BLOCK + idx_in_bitvec;
return orig_idx;
}
}
panic!("Internal Error: Offset calculation in dense idx seems to be wrong.");
})
}
}
#[inline]
fn dense_index_block(data: &[u8], block_pos: u32) -> DenseIndexBlock {
let data_start_pos = block_pos as usize * SERIALIZED_BLOCK_SIZE;
let block_data: [u8; SERIALIZED_BLOCK_SIZE] = data[data_start_pos..][..SERIALIZED_BLOCK_SIZE]
.try_into()
.unwrap();
block_data.into()
}
#[inline]
/// Finds the block position containing the dense_idx.
///
/// # Correctness
/// dense_idx needs to be smaller than the number of values in the index
///
/// The last offset number is equal to the number of values in the index.
fn find_block(dense_idx: u32, mut block_pos: u32, data: &[u8]) -> u32 {
loop {
let offset = dense_index_block(data, block_pos).offset;
if offset > dense_idx {
return block_pos - 1;
}
block_pos += 1;
}
}
/// Iterator over all values, true if set, otherwise false
pub fn serialize_dense_codec(
iter: impl Iterator<Item = bool>,
mut out: impl Write,
) -> io::Result<()> {
let mut offset: u32 = 0;
for chunk in &iter.chunks(ELEMENTS_PER_BLOCK as usize) {
let mut block: u64 = 0;
for (pos, is_bit_set) in chunk.enumerate() {
if is_bit_set {
set_bit_at(&mut block, pos as u64);
}
}
block.serialize(&mut out)?;
offset.serialize(&mut out)?;
offset += block.count_ones();
}
// Add sentinal block for the offset
let block: u64 = 0;
block.serialize(&mut out)?;
offset.serialize(&mut out)?;
Ok(())
}
#[cfg(test)]
mod tests {
use proptest::prelude::{any, prop, *};
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use super::*;
fn random_bitvec() -> BoxedStrategy<Vec<bool>> {
prop_oneof![
1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..100),
1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..64),
1 => prop::collection::vec(proptest::bool::weighted(0.0), 0..100),
1 => prop::collection::vec(proptest::bool::weighted(0.0), 0..64),
8 => vec![any::<bool>()],
2 => prop::collection::vec(any::<bool>(), 0..50),
]
.boxed()
}
proptest! {
#![proptest_config(ProptestConfig::with_cases(500))]
#[test]
fn test_with_random_bitvecs(bitvec1 in random_bitvec(), bitvec2 in random_bitvec(), bitvec3 in random_bitvec()) {
let mut bitvec = Vec::new();
bitvec.extend_from_slice(&bitvec1);
bitvec.extend_from_slice(&bitvec2);
bitvec.extend_from_slice(&bitvec3);
test_null_index(bitvec);
}
}
#[test]
fn dense_codec_test_one_block_false() {
let mut iter = vec![false; 64];
iter.push(true);
test_null_index(iter);
}
fn test_null_index(data: Vec<bool>) {
let mut out = vec![];
serialize_dense_codec(data.iter().cloned(), &mut out).unwrap();
let null_index = DenseCodec::open(OwnedBytes::new(out));
let orig_idx_with_value: Vec<u32> = data
.iter()
.enumerate()
.filter(|(_pos, val)| **val)
.map(|(pos, _val)| pos as u32)
.collect();
assert_eq!(
null_index
.translate_codec_idx_to_original_idx(0..orig_idx_with_value.len() as u32)
.collect_vec(),
orig_idx_with_value
);
for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate() {
assert_eq!(
null_index.translate_to_codec_idx(*orig_idx),
Some(dense_idx as u32)
);
}
for (pos, value) in data.iter().enumerate() {
assert_eq!(null_index.exists(pos as u32), *value);
}
}
#[test]
fn dense_codec_test_translation() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_dense_codec(iter, &mut out).unwrap();
let null_index = DenseCodec::open(OwnedBytes::new(out));
assert_eq!(
null_index
.translate_codec_idx_to_original_idx(0..2)
.collect_vec(),
vec![0, 2]
);
}
#[test]
fn dense_codec_translate() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_dense_codec(iter, &mut out).unwrap();
let null_index = DenseCodec::open(OwnedBytes::new(out));
assert_eq!(null_index.translate_to_codec_idx(0), Some(0));
assert_eq!(null_index.translate_to_codec_idx(2), Some(1));
}
#[test]
fn dense_codec_test_small() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_dense_codec(iter, &mut out).unwrap();
let null_index = DenseCodec::open(OwnedBytes::new(out));
assert!(null_index.exists(0));
assert!(!null_index.exists(1));
assert!(null_index.exists(2));
assert!(!null_index.exists(3));
}
#[test]
fn dense_codec_test_large() {
let mut docs = vec![];
docs.extend((0..1000).map(|_idx| false));
docs.extend((0..=1000).map(|_idx| true));
let iter = docs.iter().cloned();
let mut out = vec![];
serialize_dense_codec(iter, &mut out).unwrap();
let null_index = DenseCodec::open(OwnedBytes::new(out));
assert!(!null_index.exists(0));
assert!(!null_index.exists(100));
assert!(!null_index.exists(999));
assert!(null_index.exists(1000));
assert!(null_index.exists(1999));
assert!(null_index.exists(2000));
assert!(!null_index.exists(2001));
}
#[test]
fn test_count_ones() {
let mut block = 0;
set_bit_at(&mut block, 0);
set_bit_at(&mut block, 2);
assert_eq!(count_ones(block, 0), 1);
assert_eq!(count_ones(block, 1), 1);
assert_eq!(count_ones(block, 2), 2);
}
}
#[cfg(all(test, feature = "unstable"))]
mod bench {
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::Bencher;
use super::*;
const TOTAL_NUM_VALUES: u32 = 1_000_000;
fn gen_bools(fill_ratio: f64) -> DenseCodec {
let mut out = Vec::new();
let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
let bools: Vec<_> = (0..TOTAL_NUM_VALUES)
.map(|_| rng.gen_bool(fill_ratio))
.collect();
serialize_dense_codec(bools.into_iter(), &mut out).unwrap();
let codec = DenseCodec::open(OwnedBytes::new(out));
codec
}
fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
let mut current = start;
std::iter::from_fn(move || {
current += rng.gen_range(1..step_size + 1);
if current >= end {
None
} else {
Some(current)
}
})
}
fn walk_over_data(codec: &DenseCodec, max_step_size: u32) -> Option<u32> {
walk_over_data_from_positions(
codec,
random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
)
}
fn walk_over_data_from_positions(
codec: &DenseCodec,
positions: impl Iterator<Item = u32>,
) -> Option<u32> {
let mut dense_idx: Option<u32> = None;
for idx in positions {
dense_idx = dense_idx.or(codec.translate_to_codec_idx(idx));
}
dense_idx
}
#[bench]
fn bench_dense_codec_translate_orig_to_codec_90percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_dense_codec_translate_orig_to_codec_50percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.5f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_dense_codec_translate_orig_to_codec_full_scan_10percent(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_dense_codec_translate_orig_to_codec_full_scan_90percent(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_dense_codec_translate_orig_to_codec_10percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_dense_codec_translate_codec_to_orig_90percent_filled_random_stride_big_step(
bench: &mut Bencher,
) {
let codec = gen_bools(0.9f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
.last()
});
}
#[bench]
fn bench_dense_codec_translate_codec_to_orig_90percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.9f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
.last()
});
}
#[bench]
fn bench_dense_codec_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(0..num_vals)
.last()
});
}
}

14
fastfield_codecs/src/null_index/mod.rs Normal file
View File

@@ -0,0 +1,14 @@
pub use dense::{serialize_dense_codec, DenseCodec};
mod dense;
mod sparse;
#[inline]
fn get_bit_at(input: u64, n: u32) -> bool {
input & (1 << n) != 0
}
#[inline]
fn set_bit_at(input: &mut u64, n: u64) {
*input |= 1 << n;
}

752
fastfield_codecs/src/null_index/sparse.rs Normal file
View File

@@ -0,0 +1,752 @@
use std::io::{self, Write};
use common::BitSet;
use ownedbytes::OwnedBytes;
use super::{serialize_dense_codec, DenseCodec};
/// `SparseCodec` is the codec for data, when only few documents have values.
/// In contrast to `DenseCodec` opening a `SparseCodec` causes runtime data to be produced, for
/// faster access.
///
/// The lower 16 bits of doc ids are stored as u16 while the upper 16 bits are given by the block
/// id. Each block contains 1<<16 docids.
///
/// # Serialized Data Layout
/// The data starts with the block data. Each block is either dense or sparse encoded, depending on
/// the number of values in the block. A block is sparse when it contains less than
/// DENSE_BLOCK_THRESHOLD (6144) values.
/// [Sparse data block | dense data block, .. #repeat*; Desc: Either a sparse or dense encoded
/// block]
/// ### Sparse block data
/// [u16 LE, .. #repeat*; Desc: Positions with values in a block]
/// ### Dense block data
/// [Dense codec for the whole block; Desc: Similar to a bitvec(0..ELEMENTS_PER_BLOCK) + Metadata
/// for faster lookups. See dense.rs]
///
/// The data is followed by block metadata, to know which area of the raw block data belongs to
/// which block. Only metadata for blocks with elements is recorded to
/// keep the overhead low for scenarios with many very sparse columns. The block metadata consists
/// of the block index and the number of values in the block. Since we don't store empty blocks
/// num_vals is incremented by 1, e.g. 0 means 1 value.
///
/// The last u16 is storing the number of metadata blocks.
/// [u16 LE, .. #repeat*; Desc: Positions with values in a block][(u16 LE, u16 LE), .. #repeat*;
/// Desc: (Block Id u16, Num Elements u16)][u16 LE; Desc: num blocks with values u16]
///
/// # Opening
/// When opening the data layout, the data is expanded to `Vec<SparseCodecBlockVariant>`, where the
/// index is the block index. For each block `byte_start` and `offset` is computed.
pub struct SparseCodec {
data: OwnedBytes,
blocks: Vec<SparseCodecBlockVariant>,
}
/// The threshold for for number of elements after which we switch to dense block encoding
const DENSE_BLOCK_THRESHOLD: u32 = 6144;
const ELEMENTS_PER_BLOCK: u32 = u16::MAX as u32 + 1;
/// 1.5 bit per Element + 12 bytes for the sentinal block
const NUM_BYTES_DENSE_BLOCK: u32 = (ELEMENTS_PER_BLOCK + ELEMENTS_PER_BLOCK / 2 + 64 + 32) / 8;
#[derive(Clone)]
enum SparseCodecBlockVariant {
Empty { offset: u32 },
Dense(DenseBlock),
Sparse(SparseBlock),
}
impl SparseCodecBlockVariant {
/// The number of non-null values that preceeded that block.
fn offset(&self) -> u32 {
match self {
SparseCodecBlockVariant::Empty { offset } => *offset,
SparseCodecBlockVariant::Dense(dense) => dense.offset,
SparseCodecBlockVariant::Sparse(sparse) => sparse.offset,
}
}
}
/// A block consists of max u16 values
#[derive(Clone)]
struct DenseBlock {
/// The number of values set before the block
offset: u32,
/// The data for the dense encoding
codec: DenseCodec,
}
impl DenseBlock {
pub fn exists(&self, idx: u32) -> bool {
self.codec.exists(idx)
}
pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
self.codec.translate_to_codec_idx(idx)
}
pub fn translate_codec_idx_to_original_idx(&self, idx: u32) -> u32 {
self.codec
.translate_codec_idx_to_original_idx(idx..=idx)
.next()
.unwrap()
}
}
/// A block consists of max u16 values
#[derive(Debug, Copy, Clone)]
struct SparseBlock {
/// The number of values in the block
num_vals: u32,
/// The number of values set before the block
offset: u32,
/// The start position of the data for the block
byte_start: u32,
}
impl SparseBlock {
fn empty_block(offset: u32) -> Self {
Self {
num_vals: 0,
byte_start: 0,
offset,
}
}
#[inline]
fn value_at_idx(&self, data: &[u8], idx: u16) -> u16 {
let start_offset: usize = self.byte_start as usize + (idx as u32 as usize * 2);
get_u16(data, start_offset)
}
#[inline]
#[allow(clippy::comparison_chain)]
// Looks for the element in the block. Returns the positions if found.
fn binary_search(&self, data: &[u8], target: u16) -> Option<u16> {
let mut size = self.num_vals as u16;
let mut left = 0;
let mut right = size;
// TODO try different implem.
// e.g. exponential search into binary search
while left < right {
let mid = left + size / 2;
// TODO do boundary check only once, and then use an
// unsafe `value_at_idx`
let mid_val = self.value_at_idx(data, mid);
if target > mid_val {
left = mid + 1;
} else if target < mid_val {
right = mid;
} else {
return Some(mid);
}
size = right - left;
}
None
}
}
#[inline]
fn get_u16(data: &[u8], byte_position: usize) -> u16 {
let bytes: [u8; 2] = data[byte_position..byte_position + 2].try_into().unwrap();
u16::from_le_bytes(bytes)
}
const SERIALIZED_BLOCK_METADATA_SIZE: usize = 4;
fn deserialize_sparse_codec_block(data: &OwnedBytes) -> Vec<SparseCodecBlockVariant> {
// The number of vals so far
let mut offset = 0;
let mut sparse_codec_blocks = Vec::new();
let num_blocks = get_u16(data, data.len() - 2);
let block_data_index_start =
data.len() - 2 - num_blocks as usize * SERIALIZED_BLOCK_METADATA_SIZE;
let mut byte_start = 0;
for block_num in 0..num_blocks as usize {
let block_data_index = block_data_index_start + SERIALIZED_BLOCK_METADATA_SIZE * block_num;
let block_idx = get_u16(data, block_data_index);
let num_vals = get_u16(data, block_data_index + 2) as u32 + 1;
sparse_codec_blocks.resize(
block_idx as usize,
SparseCodecBlockVariant::Empty { offset },
);
if is_sparse(num_vals) {
let block = SparseBlock {
num_vals,
offset,
byte_start,
};
sparse_codec_blocks.push(SparseCodecBlockVariant::Sparse(block));
byte_start += 2 * num_vals;
} else {
let block = DenseBlock {
offset,
codec: DenseCodec::open(data.slice(byte_start as usize..data.len()).clone()),
};
sparse_codec_blocks.push(SparseCodecBlockVariant::Dense(block));
// Dense blocks have a fixed size spanning ELEMENTS_PER_BLOCK.
byte_start += NUM_BYTES_DENSE_BLOCK;
}
offset += num_vals;
}
sparse_codec_blocks.push(SparseCodecBlockVariant::Empty { offset });
sparse_codec_blocks
}
/// Splits a value address into lower and upper 16bits.
/// The lower 16 bits are the value in the block
/// The upper 16 bits are the block index
#[derive(Debug, Clone, Copy)]
struct ValueAddr {
block_idx: u16,
value_in_block: u16,
}
/// Splits a idx into block index and value in the block
fn value_addr(idx: u32) -> ValueAddr {
/// Static assert number elements per block this method expects
#[allow(clippy::assertions_on_constants)]
const _: () = assert!(ELEMENTS_PER_BLOCK == (1 << 16));
let value_in_block = idx as u16;
let block_idx = (idx >> 16) as u16;
ValueAddr {
block_idx,
value_in_block,
}
}
impl SparseCodec {
/// Open the SparseCodec from OwnedBytes
pub fn open(data: OwnedBytes) -> Self {
let blocks = deserialize_sparse_codec_block(&data);
Self { data, blocks }
}
#[inline]
/// Check if value at position is not null.
pub fn exists(&self, idx: u32) -> bool {
let value_addr = value_addr(idx);
// There may be trailing nulls without data, those are not stored as blocks. It would be
// possible to create empty blocks, but for that we would need to serialize the number of
// values or pass them when opening
if let Some(block) = self.blocks.get(value_addr.block_idx as usize) {
match block {
SparseCodecBlockVariant::Empty { offset: _ } => false,
SparseCodecBlockVariant::Dense(block) => {
block.exists(value_addr.value_in_block as u32)
}
SparseCodecBlockVariant::Sparse(block) => block
.binary_search(&self.data, value_addr.value_in_block)
.is_some(),
}
} else {
false
}
}
/// Return the number of non-null values in an index
pub fn num_non_nulls(&self) -> u32 {
self.blocks.last().map(|block| block.offset()).unwrap_or(0)
}
#[inline]
/// Translate from the original index to the codec index.
pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
let value_addr = value_addr(idx);
let block = self.blocks.get(value_addr.block_idx as usize)?;
match block {
SparseCodecBlockVariant::Empty { offset: _ } => None,
SparseCodecBlockVariant::Dense(block) => block
.translate_to_codec_idx(value_addr.value_in_block as u32)
.map(|pos_in_block| pos_in_block + block.offset),
SparseCodecBlockVariant::Sparse(block) => {
let pos_in_block = block.binary_search(&self.data, value_addr.value_in_block);
pos_in_block.map(|pos_in_block: u16| block.offset + pos_in_block as u32)
}
}
}
fn find_block(&self, dense_idx: u32, mut block_pos: u32) -> u32 {
loop {
let offset = self.blocks[block_pos as usize].offset();
if offset > dense_idx {
return block_pos - 1;
}
block_pos += 1;
}
}
/// Translate positions from the codec index to the original index.
///
/// # Panics
///
/// May panic if any `idx` is greater than the max codec index.
pub fn translate_codec_idx_to_original_idx<'a>(
&'a self,
iter: impl Iterator<Item = u32> + 'a,
) -> impl Iterator<Item = u32> + 'a {
// TODO: There's a big potential performance gain, by using iterators per block instead of
// random access for each element in a block
// group_by itertools won't help though, since it requires a temporary local variable
let mut block_pos = 0u32;
iter.map(move |codec_idx| {
// update block_pos to limit search scope
block_pos = self.find_block(codec_idx, block_pos);
let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
let block = &self.blocks[block_pos as usize];
let idx_in_block = codec_idx - block.offset();
match block {
SparseCodecBlockVariant::Empty { offset: _ } => {
panic!(
"invalid input, cannot translate to original index. associated empty \
block with dense idx. block_pos {}, idx_in_block {}",
block_pos, idx_in_block
)
}
SparseCodecBlockVariant::Dense(dense) => {
dense.translate_codec_idx_to_original_idx(idx_in_block) + block_doc_idx_start
}
SparseCodecBlockVariant::Sparse(block) => {
block.value_at_idx(&self.data, idx_in_block as u16) as u32 + block_doc_idx_start
}
}
})
}
}
fn is_sparse(num_elem_in_block: u32) -> bool {
num_elem_in_block < DENSE_BLOCK_THRESHOLD
}
#[derive(Default)]
struct BlockDataSerialized {
block_idx: u16,
num_vals: u32,
}
/// Iterator over positions of set values.
pub fn serialize_sparse_codec<W: Write>(
mut iter: impl Iterator<Item = u32>,
mut out: W,
) -> io::Result<()> {
let mut block_metadata: Vec<BlockDataSerialized> = Vec::new();
let mut current_block = Vec::new();
// This if-statement for the first element ensures that
// `block_metadata` is not empty in the loop below.
if let Some(idx) = iter.next() {
let value_addr = value_addr(idx);
block_metadata.push(BlockDataSerialized {
block_idx: value_addr.block_idx,
num_vals: 1,
});
current_block.push(value_addr.value_in_block);
}
let flush_block = |current_block: &mut Vec<u16>, out: &mut W| -> io::Result<()> {
let is_sparse = is_sparse(current_block.len() as u32);
if is_sparse {
for val_in_block in current_block.iter() {
out.write_all(val_in_block.to_le_bytes().as_ref())?;
}
} else {
let mut bitset = BitSet::with_max_value(ELEMENTS_PER_BLOCK + 1);
for val_in_block in current_block.iter() {
bitset.insert(*val_in_block as u32);
}
let iter = (0..ELEMENTS_PER_BLOCK).map(|idx| bitset.contains(idx));
serialize_dense_codec(iter, out)?;
}
current_block.clear();
Ok(())
};
for idx in iter {
let value_addr = value_addr(idx);
if block_metadata[block_metadata.len() - 1].block_idx == value_addr.block_idx {
let last_idx_metadata = block_metadata.len() - 1;
block_metadata[last_idx_metadata].num_vals += 1;
} else {
// flush prev block
flush_block(&mut current_block, &mut out)?;
block_metadata.push(BlockDataSerialized {
block_idx: value_addr.block_idx,
num_vals: 1,
});
}
current_block.push(value_addr.value_in_block);
}
// handle last block
flush_block(&mut current_block, &mut out)?;
for block in &block_metadata {
out.write_all(block.block_idx.to_le_bytes().as_ref())?;
// We don't store empty blocks, therefore we can subtract 1.
// This way we will be able to use u16 when the number of elements is 1 << 16 or u16::MAX+1
out.write_all(((block.num_vals - 1) as u16).to_le_bytes().as_ref())?;
}
out.write_all((block_metadata.len() as u16).to_le_bytes().as_ref())?;
Ok(())
}
#[cfg(test)]
mod tests {
use itertools::Itertools;
use proptest::prelude::{any, prop, *};
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use super::*;
fn random_bitvec() -> BoxedStrategy<Vec<bool>> {
prop_oneof![
1 => prop::collection::vec(proptest::bool::weighted(1.0), 0..100),
1 => prop::collection::vec(proptest::bool::weighted(0.00), 0..(ELEMENTS_PER_BLOCK as usize * 3)), // empty blocks
1 => prop::collection::vec(proptest::bool::weighted(1.00), 0..(ELEMENTS_PER_BLOCK as usize + 10)), // full block
1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..100),
1 => prop::collection::vec(proptest::bool::weighted(0.01), 0..u16::MAX as usize),
8 => vec![any::<bool>()],
]
.boxed()
}
proptest! {
#![proptest_config(ProptestConfig::with_cases(50))]
#[test]
fn test_with_random_bitvecs(bitvec1 in random_bitvec(), bitvec2 in random_bitvec(), bitvec3 in random_bitvec()) {
let mut bitvec = Vec::new();
bitvec.extend_from_slice(&bitvec1);
bitvec.extend_from_slice(&bitvec2);
bitvec.extend_from_slice(&bitvec3);
test_null_index(bitvec);
}
}
#[test]
fn sparse_codec_test_one_block_false() {
let mut iter = vec![false; ELEMENTS_PER_BLOCK as usize];
iter.push(true);
test_null_index(iter);
}
#[test]
fn sparse_codec_test_one_block_true() {
let mut iter = vec![true; ELEMENTS_PER_BLOCK as usize];
iter.push(true);
test_null_index(iter);
}
fn test_null_index(data: Vec<bool>) {
let mut out = vec![];
serialize_sparse_codec(
data.iter()
.cloned()
.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let null_index = SparseCodec::open(OwnedBytes::new(out));
let orig_idx_with_value: Vec<u32> = data
.iter()
.enumerate()
.filter(|(_pos, val)| **val)
.map(|(pos, _val)| pos as u32)
.collect();
assert_eq!(
null_index
.translate_codec_idx_to_original_idx(0..orig_idx_with_value.len() as u32)
.collect_vec(),
orig_idx_with_value
);
let step_size = (orig_idx_with_value.len() / 100).max(1);
for (dense_idx, orig_idx) in orig_idx_with_value.iter().enumerate().step_by(step_size) {
assert_eq!(
null_index.translate_to_codec_idx(*orig_idx),
Some(dense_idx as u32)
);
}
// 100 samples
let step_size = (data.len() / 100).max(1);
for (pos, value) in data.iter().enumerate().step_by(step_size) {
assert_eq!(null_index.exists(pos as u32), *value);
}
}
#[test]
fn sparse_codec_test_translation() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_sparse_codec(
iter.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let null_index = SparseCodec::open(OwnedBytes::new(out));
assert_eq!(
null_index
.translate_codec_idx_to_original_idx(0..2)
.collect_vec(),
vec![0, 2]
);
}
#[test]
fn sparse_codec_translate() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_sparse_codec(
iter.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let null_index = SparseCodec::open(OwnedBytes::new(out));
assert_eq!(null_index.translate_to_codec_idx(0), Some(0));
assert_eq!(null_index.translate_to_codec_idx(2), Some(1));
}
#[test]
fn sparse_codec_test_small() {
let mut out = vec![];
let iter = ([true, false, true, false]).iter().cloned();
serialize_sparse_codec(
iter.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let null_index = SparseCodec::open(OwnedBytes::new(out));
assert!(null_index.exists(0));
assert!(!null_index.exists(1));
assert!(null_index.exists(2));
assert!(!null_index.exists(3));
}
#[test]
fn sparse_codec_test_large() {
let mut docs = vec![];
docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
docs.extend((0..=1).map(|_idx| true));
let iter = docs.iter().cloned();
let mut out = vec![];
serialize_sparse_codec(
iter.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let null_index = SparseCodec::open(OwnedBytes::new(out));
assert!(!null_index.exists(0));
assert!(!null_index.exists(100));
assert!(!null_index.exists(ELEMENTS_PER_BLOCK - 1));
assert!(null_index.exists(ELEMENTS_PER_BLOCK));
assert!(null_index.exists(ELEMENTS_PER_BLOCK + 1));
}
}
#[cfg(all(test, feature = "unstable"))]
mod bench {
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::Bencher;
use super::*;
const TOTAL_NUM_VALUES: u32 = 1_000_000;
fn gen_bools(fill_ratio: f64) -> SparseCodec {
let mut out = Vec::new();
let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
serialize_sparse_codec(
(0..TOTAL_NUM_VALUES)
.map(|_| rng.gen_bool(fill_ratio))
.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
&mut out,
)
.unwrap();
let codec = SparseCodec::open(OwnedBytes::new(out));
codec
}
fn random_range_iterator(start: u32, end: u32, step_size: u32) -> impl Iterator<Item = u32> {
let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
let mut current = start;
std::iter::from_fn(move || {
current += rng.gen_range(1..step_size + 1);
if current >= end {
None
} else {
Some(current)
}
})
}
fn walk_over_data(codec: &SparseCodec, max_step_size: u32) -> Option<u32> {
walk_over_data_from_positions(
codec,
random_range_iterator(0, TOTAL_NUM_VALUES, max_step_size),
)
}
fn walk_over_data_from_positions(
codec: &SparseCodec,
positions: impl Iterator<Item = u32>,
) -> Option<u32> {
let mut dense_idx: Option<u32> = None;
for idx in positions {
dense_idx = dense_idx.or(codec.translate_to_codec_idx(idx));
}
dense_idx
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_1percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_5percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_full_scan_10percent(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_full_scan_90percent(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_full_scan_1percent(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_10percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_sparse_codec_translate_orig_to_codec_90percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_random_stride_big_step(
bench: &mut Bencher,
) {
let codec = gen_bools(0.01f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
.last()
});
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.01f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
.last()
});
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(0..num_vals)
.last()
});
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_random_stride_big_step(
bench: &mut Bencher,
) {
let codec = gen_bools(0.90f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 50_000))
.last()
});
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_random_stride(
bench: &mut Bencher,
) {
let codec = gen_bools(0.9f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(random_range_iterator(0, num_vals, 100))
.last()
});
}
#[bench]
fn bench_sparse_codec_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
let num_vals = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(0..num_vals)
.last()
});
}
}

146
fastfield_codecs/src/null_index_footer.rs Normal file
View File

@@ -0,0 +1,146 @@
use std::io::{self, Write};
use std::ops::Range;
use common::{BinarySerializable, CountingWriter, VInt};
use ownedbytes::OwnedBytes;
#[derive(Debug, Clone, Copy, Eq, PartialEq)]
pub(crate) enum FastFieldCardinality {
Single = 1,
Multi = 2,
}
impl BinarySerializable for FastFieldCardinality {
fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
self.to_code().serialize(wrt)
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let code = u8::deserialize(reader)?;
let codec_type: Self = Self::from_code(code)
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
Ok(codec_type)
}
}
impl FastFieldCardinality {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::Single),
2 => Some(Self::Multi),
_ => None,
}
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum NullIndexCodec {
Full = 1,
}
impl BinarySerializable for NullIndexCodec {
fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
self.to_code().serialize(wrt)
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let code = u8::deserialize(reader)?;
let codec_type: Self = Self::from_code(code)
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
Ok(codec_type)
}
}
impl NullIndexCodec {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::Full),
_ => None,
}
}
}
#[derive(Debug, Clone, Eq, PartialEq)]
pub(crate) struct NullIndexFooter {
pub(crate) cardinality: FastFieldCardinality,
pub(crate) null_index_codec: NullIndexCodec,
// Unused for NullIndexCodec::Full
pub(crate) null_index_byte_range: Range<u64>,
}
impl BinarySerializable for NullIndexFooter {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
self.cardinality.serialize(writer)?;
self.null_index_codec.serialize(writer)?;
VInt(self.null_index_byte_range.start).serialize(writer)?;
VInt(self.null_index_byte_range.end - self.null_index_byte_range.start)
.serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let cardinality = FastFieldCardinality::deserialize(reader)?;
let null_index_codec = NullIndexCodec::deserialize(reader)?;
let null_index_byte_range_start = VInt::deserialize(reader)?.0;
let null_index_byte_range_end = VInt::deserialize(reader)?.0 + null_index_byte_range_start;
Ok(Self {
cardinality,
null_index_codec,
null_index_byte_range: null_index_byte_range_start..null_index_byte_range_end,
})
}
}
pub(crate) fn append_null_index_footer(
output: &mut impl io::Write,
null_index_footer: NullIndexFooter,
) -> io::Result<()> {
let mut counting_write = CountingWriter::wrap(output);
null_index_footer.serialize(&mut counting_write)?;
let footer_payload_len = counting_write.written_bytes();
BinarySerializable::serialize(&(footer_payload_len as u16), &mut counting_write)?;
Ok(())
}
pub(crate) fn read_null_index_footer(
data: OwnedBytes,
) -> io::Result<(OwnedBytes, NullIndexFooter)> {
let (data, null_footer_length_bytes) = data.rsplit(2);
let footer_length = u16::deserialize(&mut null_footer_length_bytes.as_slice())?;
let (data, null_index_footer_bytes) = data.rsplit(footer_length as usize);
let null_index_footer = NullIndexFooter::deserialize(&mut null_index_footer_bytes.as_ref())?;
Ok((data, null_index_footer))
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn null_index_footer_deser_test() {
let null_index_footer = NullIndexFooter {
cardinality: FastFieldCardinality::Single,
null_index_codec: NullIndexCodec::Full,
null_index_byte_range: 100..120,
};
let mut out = vec![];
null_index_footer.serialize(&mut out).unwrap();
assert_eq!(
null_index_footer,
NullIndexFooter::deserialize(&mut &out[..]).unwrap()
);
}
}

408
fastfield_codecs/src/serialize.rs Normal file
View File

@@ -0,0 +1,408 @@
use std::io;
use std::num::NonZeroU64;
use std::sync::Arc;
use common::{BinarySerializable, VInt};
use log::warn;
use ownedbytes::OwnedBytes;
use crate::bitpacked::BitpackedCodec;
use crate::blockwise_linear::BlockwiseLinearCodec;
use crate::compact_space::CompactSpaceCompressor;
use crate::format_version::append_format_version;
use crate::linear::LinearCodec;
use crate::monotonic_mapping::{
StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalGCDBaseval,
};
use crate::null_index_footer::{
append_null_index_footer, FastFieldCardinality, NullIndexCodec, NullIndexFooter,
};
use crate::{
monotonic_map_column, Column, FastFieldCodec, FastFieldCodecType, MonotonicallyMappableToU64,
U128FastFieldCodecType, VecColumn, ALL_CODEC_TYPES,
};
/// The normalized header gives some parameters after applying the following
/// normalization of the vector:
/// `val -> (val - min_value) / gcd`
///
/// By design, after normalization, `min_value = 0` and `gcd = 1`.
#[derive(Debug, Copy, Clone)]
pub struct NormalizedHeader {
/// The number of values in the underlying column.
pub num_vals: u32,
/// The max value of the underlying column.
pub max_value: u64,
}
#[derive(Debug, Copy, Clone)]
pub(crate) struct Header {
pub num_vals: u32,
pub min_value: u64,
pub max_value: u64,
pub gcd: Option<NonZeroU64>,
pub codec_type: FastFieldCodecType,
}
impl Header {
pub fn normalized(self) -> NormalizedHeader {
let gcd = self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
let gcd_min_val_mapping =
StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, self.min_value);
let max_value = gcd_min_val_mapping.mapping(self.max_value);
NormalizedHeader {
num_vals: self.num_vals,
max_value,
}
}
pub fn normalize_column<C: Column>(&self, from_column: C) -> impl Column {
normalize_column(from_column, self.min_value, self.gcd)
}
pub fn compute_header(
column: impl Column<u64>,
codecs: &[FastFieldCodecType],
) -> Option<Header> {
let num_vals = column.num_vals();
let min_value = column.min_value();
let max_value = column.max_value();
let gcd = crate::gcd::find_gcd(column.iter().map(|val| val - min_value))
.filter(|gcd| gcd.get() > 1u64);
let normalized_column = normalize_column(column, min_value, gcd);
let codec_type = detect_codec(normalized_column, codecs)?;
Some(Header {
num_vals,
min_value,
max_value,
gcd,
codec_type,
})
}
}
#[derive(Debug, Copy, Clone, PartialEq, Eq)]
pub(crate) struct U128Header {
pub num_vals: u32,
pub codec_type: U128FastFieldCodecType,
}
impl BinarySerializable for U128Header {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
VInt(self.num_vals as u64).serialize(writer)?;
self.codec_type.serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let num_vals = VInt::deserialize(reader)?.0 as u32;
let codec_type = U128FastFieldCodecType::deserialize(reader)?;
Ok(U128Header {
num_vals,
codec_type,
})
}
}
pub fn normalize_column<C: Column>(
from_column: C,
min_value: u64,
gcd: Option<NonZeroU64>,
) -> impl Column {
let gcd = gcd.map(|gcd| gcd.get()).unwrap_or(1);
let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd, min_value);
monotonic_map_column(from_column, mapping)
}
impl BinarySerializable for Header {
fn serialize<W: io::Write>(&self, writer: &mut W) -> io::Result<()> {
VInt(self.num_vals as u64).serialize(writer)?;
VInt(self.min_value).serialize(writer)?;
VInt(self.max_value - self.min_value).serialize(writer)?;
if let Some(gcd) = self.gcd {
VInt(gcd.get()).serialize(writer)?;
} else {
VInt(0u64).serialize(writer)?;
}
self.codec_type.serialize(writer)?;
Ok(())
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let num_vals = VInt::deserialize(reader)?.0 as u32;
let min_value = VInt::deserialize(reader)?.0;
let amplitude = VInt::deserialize(reader)?.0;
let max_value = min_value + amplitude;
let gcd_u64 = VInt::deserialize(reader)?.0;
let codec_type = FastFieldCodecType::deserialize(reader)?;
Ok(Header {
num_vals,
min_value,
max_value,
gcd: NonZeroU64::new(gcd_u64),
codec_type,
})
}
}
/// Return estimated compression for given codec in the value range [0.0..1.0], where 1.0 means no
/// compression.
pub fn estimate<T: MonotonicallyMappableToU64>(
typed_column: impl Column<T>,
codec_type: FastFieldCodecType,
) -> Option<f32> {
let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
let min_value = column.min_value();
let gcd = crate::gcd::find_gcd(column.iter().map(|val| val - min_value))
.filter(|gcd| gcd.get() > 1u64);
let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(
gcd.map(|gcd| gcd.get()).unwrap_or(1u64),
min_value,
);
let normalized_column = monotonic_map_column(&column, mapping);
match codec_type {
FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&normalized_column),
FastFieldCodecType::Linear => LinearCodec::estimate(&normalized_column),
FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&normalized_column),
}
}
/// Serializes u128 values with the compact space codec.
pub fn serialize_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
iter_gen: F,
num_vals: u32,
output: &mut impl io::Write,
) -> io::Result<()> {
serialize_u128_new(ValueIndexInfo::default(), iter_gen, num_vals, output)
}
#[allow(dead_code)]
pub enum ValueIndexInfo<'a> {
MultiValue(Box<dyn MultiValueIndexInfo + 'a>),
SingleValue(Box<dyn SingleValueIndexInfo + 'a>),
}
impl Default for ValueIndexInfo<'static> {
fn default() -> Self {
struct Dummy {}
impl SingleValueIndexInfo for Dummy {
fn num_vals(&self) -> u32 {
todo!()
}
fn num_non_nulls(&self) -> u32 {
todo!()
}
fn iter(&self) -> Box<dyn Iterator<Item = u32>> {
todo!()
}
}
Self::SingleValue(Box::new(Dummy {}))
}
}
impl<'a> ValueIndexInfo<'a> {
fn get_cardinality(&self) -> FastFieldCardinality {
match self {
ValueIndexInfo::MultiValue(_) => FastFieldCardinality::Multi,
ValueIndexInfo::SingleValue(_) => FastFieldCardinality::Single,
}
}
}
pub trait MultiValueIndexInfo {
/// The number of docs in the column.
fn num_docs(&self) -> u32;
/// The number of values in the column.
fn num_vals(&self) -> u32;
/// Return the start index of the values for each doc
fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_>;
}
pub trait SingleValueIndexInfo {
/// The number of values including nulls in the column.
fn num_vals(&self) -> u32;
/// The number of non-null values in the column.
fn num_non_nulls(&self) -> u32;
/// Return a iterator of the positions of docs with a value
fn iter(&self) -> Box<dyn Iterator<Item = u32> + '_>;
}
/// Serializes u128 values with the compact space codec.
pub fn serialize_u128_new<F: Fn() -> I, I: Iterator<Item = u128>>(
value_index: ValueIndexInfo,
iter_gen: F,
num_vals: u32,
output: &mut impl io::Write,
) -> io::Result<()> {
let header = U128Header {
num_vals,
codec_type: U128FastFieldCodecType::CompactSpace,
};
header.serialize(output)?;
let compressor = CompactSpaceCompressor::train_from(iter_gen(), num_vals);
compressor.compress_into(iter_gen(), output).unwrap();
let null_index_footer = NullIndexFooter {
cardinality: value_index.get_cardinality(),
null_index_codec: NullIndexCodec::Full,
null_index_byte_range: 0..0,
};
append_null_index_footer(output, null_index_footer)?;
append_format_version(output)?;
Ok(())
}
/// Serializes the column with the codec with the best estimate on the data.
pub fn serialize<T: MonotonicallyMappableToU64>(
typed_column: impl Column<T>,
output: &mut impl io::Write,
codecs: &[FastFieldCodecType],
) -> io::Result<()> {
serialize_new(ValueIndexInfo::default(), typed_column, output, codecs)
}
/// Serializes the column with the codec with the best estimate on the data.
pub fn serialize_new<T: MonotonicallyMappableToU64>(
value_index: ValueIndexInfo,
typed_column: impl Column<T>,
output: &mut impl io::Write,
codecs: &[FastFieldCodecType],
) -> io::Result<()> {
let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
let header = Header::compute_header(&column, codecs).ok_or_else(|| {
io::Error::new(
io::ErrorKind::InvalidInput,
format!(
"Data cannot be serialized with this list of codec. {:?}",
codecs
),
)
})?;
header.serialize(output)?;
let normalized_column = header.normalize_column(column);
assert_eq!(normalized_column.min_value(), 0u64);
serialize_given_codec(normalized_column, header.codec_type, output)?;
let null_index_footer = NullIndexFooter {
cardinality: value_index.get_cardinality(),
null_index_codec: NullIndexCodec::Full,
null_index_byte_range: 0..0,
};
append_null_index_footer(output, null_index_footer)?;
append_format_version(output)?;
Ok(())
}
fn detect_codec(
column: impl Column<u64>,
codecs: &[FastFieldCodecType],
) -> Option<FastFieldCodecType> {
let mut estimations = Vec::new();
for &codec in codecs {
let estimation_opt = match codec {
FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&column),
FastFieldCodecType::Linear => LinearCodec::estimate(&column),
FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&column),
};
if let Some(estimation) = estimation_opt {
estimations.push((estimation, codec));
}
}
if let Some(broken_estimation) = estimations.iter().find(|estimation| estimation.0.is_nan()) {
warn!(
"broken estimation for fast field codec {:?}",
broken_estimation.1
);
}
// removing nan values for codecs with broken calculations, and max values which disables
// codecs
estimations.retain(|estimation| !estimation.0.is_nan() && estimation.0 != f32::MAX);
estimations.sort_by(|(score_left, _), (score_right, _)| score_left.total_cmp(score_right));
Some(estimations.first()?.1)
}
fn serialize_given_codec(
column: impl Column<u64>,
codec_type: FastFieldCodecType,
output: &mut impl io::Write,
) -> io::Result<()> {
match codec_type {
FastFieldCodecType::Bitpacked => {
BitpackedCodec::serialize(&column, output)?;
}
FastFieldCodecType::Linear => {
LinearCodec::serialize(&column, output)?;
}
FastFieldCodecType::BlockwiseLinear => {
BlockwiseLinearCodec::serialize(&column, output)?;
}
}
output.flush()?;
Ok(())
}
/// Helper function to serialize a column (autodetect from all codecs) and then open it
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
column: &[T],
) -> Arc<dyn Column<T>> {
let mut buffer = Vec::new();
super::serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
super::open(OwnedBytes::new(buffer)).unwrap()
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_serialize_deserialize_u128_header() {
let original = U128Header {
num_vals: 11,
codec_type: U128FastFieldCodecType::CompactSpace,
};
let mut out = Vec::new();
original.serialize(&mut out).unwrap();
let restored = U128Header::deserialize(&mut &out[..]).unwrap();
assert_eq!(restored, original);
}
#[test]
fn test_serialize_deserialize() {
let original = [1u64, 5u64, 10u64];
let restored: Vec<u64> = serialize_and_load(&original[..]).iter().collect();
assert_eq!(&restored, &original[..]);
}
#[test]
fn test_fastfield_bool_size_bitwidth_1() {
let mut buffer = Vec::new();
let col = VecColumn::from(&[false, true][..]);
serialize(col, &mut buffer, &ALL_CODEC_TYPES).unwrap();
// 5 bytes of header, 1 byte of value, 7 bytes of padding.
assert_eq!(buffer.len(), 3 + 5 + 8 + 4 + 2);
}
#[test]
fn test_fastfield_bool_bit_size_bitwidth_0() {
let mut buffer = Vec::new();
let col = VecColumn::from(&[true][..]);
serialize(col, &mut buffer, &ALL_CODEC_TYPES).unwrap();
// 5 bytes of header, 0 bytes of value, 7 bytes of padding.
assert_eq!(buffer.len(), 3 + 5 + 7 + 4 + 2);
}
#[test]
fn test_fastfield_gcd() {
let mut buffer = Vec::new();
let vals: Vec<u64> = (0..80).map(|val| (val % 7) * 1_000u64).collect();
let col = VecColumn::from(&vals[..]);
serialize(col, &mut buffer, &[FastFieldCodecType::Bitpacked]).unwrap();
// Values are stored over 3 bits.
assert_eq!(buffer.len(), 3 + 7 + (3 * 80 / 8) + 7 + 4 + 2);
}
}

6
ownedbytes/Cargo.toml
View File

@@ -1,10 +1,14 @@
[package]
authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
name = "ownedbytes"
version = "0.3.0"
version = "0.5.0"
edition = "2021"
description = "Expose data as static slice"
license = "MIT"
documentation = "https://docs.rs/ownedbytes/"
homepage = "https://github.com/quickwit-oss/tantivy"
repository = "https://github.com/quickwit-oss/tantivy"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]

21
ownedbytes/src/lib.rs
View File

@@ -3,10 +3,10 @@ use std::ops::{Deref, Range};
use std::sync::Arc;
use std::{fmt, io, mem};
use stable_deref_trait::StableDeref;
pub use stable_deref_trait::StableDeref;
/// An OwnedBytes simply wraps an object that owns a slice of data and exposes
/// this data as a static slice.
/// this data as a slice.
///
/// The backing object is required to be `StableDeref`.
#[derive(Clone)]
@@ -21,7 +21,7 @@ impl OwnedBytes {
OwnedBytes::new(&[][..])
}
/// Creates an `OwnedBytes` intance given a `StableDeref` object.
/// Creates an `OwnedBytes` instance given a `StableDeref` object.
pub fn new<T: StableDeref + Deref<Target = [u8]> + 'static + Send + Sync>(
data_holder: T,
) -> OwnedBytes {
@@ -80,6 +80,21 @@ impl OwnedBytes {
(left, right)
}
/// Splits the OwnedBytes into two OwnedBytes `(left, right)`.
///
/// Right will hold `split_len` bytes.
///
/// This operation is cheap and does not require to copy any memory.
/// On the other hand, both `left` and `right` retain a handle over
/// the entire slice of memory. In other words, the memory will only
/// be released when both left and right are dropped.
#[inline]
#[must_use]
pub fn rsplit(self, split_len: usize) -> (OwnedBytes, OwnedBytes) {
let data_len = self.data.len();
self.split(data_len - split_len)
}
/// Splits the right part of the `OwnedBytes` at the given offset.
///
/// `self` is truncated to `split_len`, left with the remaining bytes.

2
query-grammar/Cargo.toml
View File

@@ -1,6 +1,6 @@
[package]
name = "tantivy-query-grammar"
version = "0.18.0"
version = "0.19.0"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
categories = ["database-implementations", "data-structures"]

2
query-grammar/src/lib.rs
View File

@@ -1,3 +1,5 @@
#![allow(clippy::derive_partial_eq_without_eq)]
mod occur;
mod query_grammar;
mod user_input_ast;

73
query-grammar/src/query_grammar.rs
View File

@@ -5,7 +5,8 @@ use combine::parser::range::{take_while, take_while1};
use combine::parser::repeat::escaped;
use combine::parser::Parser;
use combine::{
attempt, choice, eof, many, many1, one_of, optional, parser, satisfy, skip_many1, value,
attempt, between, choice, eof, many, many1, one_of, optional, parser, satisfy, sep_by,
skip_many1, value,
};
use once_cell::sync::Lazy;
use regex::Regex;
@@ -23,7 +24,7 @@ const ESCAPED_SPECIAL_CHARS_PATTERN: &str = r#"\\(\+|\^|`|:|\{|\}|"|\[|\]|\(|\)|
/// Parses a field_name
/// A field name must have at least one character and be followed by a colon.
/// All characters are allowed including special characters `SPECIAL_CHARS`, but these
/// need to be escaped with a backslack character '\'.
/// need to be escaped with a backslash character '\'.
fn field_name<'a>() -> impl Parser<&'a str, Output = String> {
static ESCAPED_SPECIAL_CHARS_RE: Lazy<Regex> =
Lazy::new(|| Regex::new(ESCAPED_SPECIAL_CHARS_PATTERN).unwrap());
@@ -62,13 +63,27 @@ fn word<'a>() -> impl Parser<&'a str, Output = String> {
})
}
// word variant that allows more characters, e.g. for range queries that don't allow field
// specifier
fn relaxed_word<'a>() -> impl Parser<&'a str, Output = String> {
(
satisfy(|c: char| {
!c.is_whitespace() && !['`', '{', '}', '"', '[', ']', '(', ')'].contains(&c)
}),
many(satisfy(|c: char| {
!c.is_whitespace() && !['{', '}', '"', '[', ']', '(', ')'].contains(&c)
})),
)
.map(|(s1, s2): (char, String)| format!("{}{}", s1, s2))
}
/// Parses a date time according to rfc3339
/// 2015-08-02T18:54:42+02
/// 2021-04-13T19:46:26.266051969+00:00
///
/// NOTE: also accepts 999999-99-99T99:99:99.266051969+99:99
/// We delegate rejecting such invalid dates to the logical AST compuation code
/// which invokes time::OffsetDateTime::parse(..., &Rfc3339) on the value to actually parse
/// We delegate rejecting such invalid dates to the logical AST computation code
/// which invokes `time::OffsetDateTime::parse(..., &Rfc3339)` on the value to actually parse
/// it (instead of merely extracting the datetime value as string as done here).
fn date_time<'a>() -> impl Parser<&'a str, Output = String> {
let two_digits = || recognize::<String, _, _>((digit(), digit()));
@@ -181,8 +196,8 @@ fn spaces1<'a>() -> impl Parser<&'a str, Output = ()> {
fn range<'a>() -> impl Parser<&'a str, Output = UserInputLeaf> {
let range_term_val = || {
attempt(date_time())
.or(word())
.or(negative_number())
.or(relaxed_word())
.or(char('*').with(value("*".to_string())))
};
@@ -250,6 +265,17 @@ fn range<'a>() -> impl Parser<&'a str, Output = UserInputLeaf> {
})
}
/// Function that parses a set out of a Stream
/// Supports ranges like: `IN [val1 val2 val3]`
fn set<'a>() -> impl Parser<&'a str, Output = UserInputLeaf> {
let term_list = between(char('['), char(']'), sep_by(term_val(), spaces()));
let set_content = ((string("IN"), spaces()), term_list).map(|(_, elements)| elements);
(optional(attempt(field_name().skip(spaces()))), set_content)
.map(|(field, elements)| UserInputLeaf::Set { field, elements })
}
fn negate(expr: UserInputAst) -> UserInputAst {
expr.unary(Occur::MustNot)
}
@@ -264,6 +290,7 @@ fn leaf<'a>() -> impl Parser<&'a str, Output = UserInputAst> {
string("NOT").skip(spaces1()).with(leaf()).map(negate),
))
.or(attempt(range().map(UserInputAst::from)))
.or(attempt(set().map(UserInputAst::from)))
.or(literal().map(UserInputAst::from))
.parse_stream(input)
.into_result()
@@ -649,6 +676,34 @@ mod test {
.expect("Cannot parse date range")
.0;
assert_eq!(res6, expected_flexible_dates);
// IP Range Unbounded
let expected_weight = UserInputLeaf::Range {
field: Some("ip".to_string()),
lower: UserInputBound::Inclusive("::1".to_string()),
upper: UserInputBound::Unbounded,
};
let res1 = range()
.parse("ip: >=::1")
.expect("Cannot parse ip v6 format")
.0;
let res2 = range()
.parse("ip:[::1 TO *}")
.expect("Cannot parse ip v6 format")
.0;
assert_eq!(res1, expected_weight);
assert_eq!(res2, expected_weight);
// IP Range Bounded
let expected_weight = UserInputLeaf::Range {
field: Some("ip".to_string()),
lower: UserInputBound::Inclusive("::0.0.0.50".to_string()),
upper: UserInputBound::Exclusive("::0.0.0.52".to_string()),
};
let res1 = range()
.parse("ip:[::0.0.0.50 TO ::0.0.0.52}")
.expect("Cannot parse ip v6 format")
.0;
assert_eq!(res1, expected_weight);
}
#[test]
@@ -705,6 +760,14 @@ mod test {
test_parse_query_to_ast_helper("+(a b) +d", "(+(*\"a\" *\"b\") +\"d\")");
}
#[test]
fn test_parse_test_query_set() {
test_parse_query_to_ast_helper("abc: IN [a b c]", r#""abc": IN ["a" "b" "c"]"#);
test_parse_query_to_ast_helper("abc: IN [1]", r#""abc": IN ["1"]"#);
test_parse_query_to_ast_helper("abc: IN []", r#""abc": IN []"#);
test_parse_query_to_ast_helper("IN [1 2]", r#"IN ["1" "2"]"#);
}
#[test]
fn test_parse_test_query_other() {
test_parse_query_to_ast_helper("(+a +b) d", "(*(+\"a\" +\"b\") *\"d\")");

17
query-grammar/src/user_input_ast.rs
View File

@@ -12,6 +12,10 @@ pub enum UserInputLeaf {
lower: UserInputBound,
upper: UserInputBound,
},
Set {
field: Option<String>,
elements: Vec<String>,
},
}
impl Debug for UserInputLeaf {
@@ -31,6 +35,19 @@ impl Debug for UserInputLeaf {
upper.display_upper(formatter)?;
Ok(())
}
UserInputLeaf::Set { field, elements } => {
if let Some(ref field) = field {
write!(formatter, "\"{}\": ", field)?;
}
write!(formatter, "IN [")?;
for (i, element) in elements.iter().enumerate() {
if i != 0 {
write!(formatter, " ")?;
}
write!(formatter, "\"{}\"", element)?;
}
write!(formatter, "]")
}
UserInputLeaf::All => write!(formatter, "*"),
}
}

14
src/aggregation/agg_req.rs
View File

@@ -1,7 +1,7 @@
//! Contains the aggregation request tree. Used to build an
//! [AggregationCollector](super::AggregationCollector).
//! [`AggregationCollector`](super::AggregationCollector).
//!
//! [Aggregations] is the top level entry point to create a request, which is a `HashMap<String,
//! [`Aggregations`] is the top level entry point to create a request, which is a `HashMap<String,
//! Aggregation>`.
//!
//! Requests are compatible with the json format of elasticsearch.
@@ -54,8 +54,8 @@ use super::bucket::{HistogramAggregation, TermsAggregation};
use super::metric::{AverageAggregation, StatsAggregation};
use super::VecWithNames;
/// The top-level aggregation request structure, which contains [Aggregation] and their user defined
/// names. It is also used in [buckets](BucketAggregation) to define sub-aggregations.
/// The top-level aggregation request structure, which contains [`Aggregation`] and their user
/// defined names. It is also used in [buckets](BucketAggregation) to define sub-aggregations.
///
/// The key is the user defined name of the aggregation.
pub type Aggregations = HashMap<String, Aggregation>;
@@ -139,15 +139,15 @@ pub fn get_fast_field_names(aggs: &Aggregations) -> HashSet<String> {
fast_field_names
}
/// Aggregation request of [BucketAggregation] or [MetricAggregation].
/// Aggregation request of [`BucketAggregation`] or [`MetricAggregation`].
///
/// An aggregation is either a bucket or a metric.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
#[serde(untagged)]
pub enum Aggregation {
/// Bucket aggregation, see [BucketAggregation] for details.
/// Bucket aggregation, see [`BucketAggregation`] for details.
Bucket(BucketAggregation),
/// Metric aggregation, see [MetricAggregation] for details.
/// Metric aggregation, see [`MetricAggregation`] for details.
Metric(MetricAggregation),
}

31
src/aggregation/agg_req_with_accessor.rs
View File

@@ -4,14 +4,14 @@ use std::rc::Rc;
use std::sync::atomic::AtomicU32;
use std::sync::Arc;
use fastfield_codecs::Column;
use super::agg_req::{Aggregation, Aggregations, BucketAggregationType, MetricAggregation};
use super::bucket::{HistogramAggregation, RangeAggregation, TermsAggregation};
use super::metric::{AverageAggregation, StatsAggregation};
use super::segment_agg_result::BucketCount;
use super::VecWithNames;
use crate::fastfield::{
type_and_cardinality, DynamicFastFieldReader, FastType, MultiValuedFastFieldReader,
};
use crate::fastfield::{type_and_cardinality, MultiValuedFastFieldReader};
use crate::schema::{Cardinality, Type};
use crate::{InvertedIndexReader, SegmentReader, TantivyError};
@@ -37,10 +37,16 @@ impl AggregationsWithAccessor {
#[derive(Clone)]
pub(crate) enum FastFieldAccessor {
Multi(MultiValuedFastFieldReader<u64>),
Single(DynamicFastFieldReader<u64>),
Single(Arc<dyn Column<u64>>),
}
impl FastFieldAccessor {
pub fn as_single(&self) -> Option<&DynamicFastFieldReader<u64>> {
pub fn as_single(&self) -> Option<&dyn Column<u64>> {
match self {
FastFieldAccessor::Multi(_) => None,
FastFieldAccessor::Single(reader) => Some(&**reader),
}
}
pub fn into_single(self) -> Option<Arc<dyn Column<u64>>> {
match self {
FastFieldAccessor::Multi(_) => None,
FastFieldAccessor::Single(reader) => Some(reader),
@@ -118,7 +124,7 @@ impl BucketAggregationWithAccessor {
pub struct MetricAggregationWithAccessor {
pub metric: MetricAggregation,
pub field_type: Type,
pub accessor: DynamicFastFieldReader<u64>,
pub accessor: Arc<dyn Column>,
}
impl MetricAggregationWithAccessor {
@@ -134,9 +140,8 @@ impl MetricAggregationWithAccessor {
Ok(MetricAggregationWithAccessor {
accessor: accessor
.as_single()
.expect("unexpected fast field cardinality")
.clone(),
.into_single()
.expect("unexpected fast field cardinality"),
field_type,
metric: metric.clone(),
})
@@ -189,13 +194,7 @@ fn get_ff_reader_and_validate(
.ok_or_else(|| TantivyError::FieldNotFound(field_name.to_string()))?;
let field_type = reader.schema().get_field_entry(field).field_type();
if let Some((ff_type, field_cardinality)) = type_and_cardinality(field_type) {
if ff_type == FastType::Date {
return Err(TantivyError::InvalidArgument(
"Unsupported field type date in aggregation".to_string(),
));
}
if let Some((_ff_type, field_cardinality)) = type_and_cardinality(field_type) {
if cardinality != field_cardinality {
return Err(TantivyError::InvalidArgument(format!(
"Invalid field cardinality on field {} expected {:?}, but got {:?}",

36
src/aggregation/agg_result.rs
View File

@@ -4,9 +4,7 @@
//! intermediate average results, which is the sum and the number of values. The actual average is
//! calculated on the step from intermediate to final aggregation result tree.
use std::collections::HashMap;
use fnv::FnvHashMap;
use rustc_hash::FxHashMap;
use serde::{Deserialize, Serialize};
use super::agg_req::BucketAggregationInternal;
@@ -14,11 +12,12 @@ use super::bucket::GetDocCount;
use super::intermediate_agg_result::{IntermediateBucketResult, IntermediateMetricResult};
use super::metric::{SingleMetricResult, Stats};
use super::Key;
use crate::schema::Schema;
use crate::TantivyError;
#[derive(Clone, Default, Debug, PartialEq, Serialize, Deserialize)]
/// The final aggegation result.
pub struct AggregationResults(pub HashMap<String, AggregationResult>);
pub struct AggregationResults(pub FxHashMap<String, AggregationResult>);
impl AggregationResults {
pub(crate) fn get_value_from_aggregation(
@@ -57,8 +56,7 @@ impl AggregationResult {
match self {
AggregationResult::BucketResult(_bucket) => Err(TantivyError::InternalError(
"Tried to retrieve value from bucket aggregation. This is not supported and \
should not happen during collection phase, but should be catched during \
validation"
should not happen during collection phase, but should be caught during validation"
.to_string(),
)),
AggregationResult::MetricResult(metric) => metric.get_value(agg_property),
@@ -114,14 +112,14 @@ pub enum BucketResult {
///
/// If there are holes depends on the request, if min_doc_count is 0, then there are no
/// holes between the first and last bucket.
/// See [HistogramAggregation](super::bucket::HistogramAggregation)
/// See [`HistogramAggregation`](super::bucket::HistogramAggregation)
buckets: BucketEntries<BucketEntry>,
},
/// This is the term result
Terms {
/// The buckets.
///
/// See [TermsAggregation](super::bucket::TermsAggregation)
/// See [`TermsAggregation`](super::bucket::TermsAggregation)
buckets: Vec<BucketEntry>,
/// The number of documents that didn’t make it into to TOP N due to shard_size or size
sum_other_doc_count: u64,
@@ -132,9 +130,12 @@ pub enum BucketResult {
}
impl BucketResult {
pub(crate) fn empty_from_req(req: &BucketAggregationInternal) -> crate::Result<Self> {
pub(crate) fn empty_from_req(
req: &BucketAggregationInternal,
schema: &Schema,
) -> crate::Result<Self> {
let empty_bucket = IntermediateBucketResult::empty_from_req(&req.bucket_agg);
empty_bucket.into_final_bucket_result(req)
empty_bucket.into_final_bucket_result(req, schema)
}
}
@@ -146,7 +147,7 @@ pub enum BucketEntries<T> {
/// Vector format bucket entries
Vec(Vec<T>),
/// HashMap format bucket entries
HashMap(FnvHashMap<String, T>),
HashMap(FxHashMap<String, T>),
}
/// This is the default entry for a bucket, which contains a key, count, and optionally
@@ -177,6 +178,9 @@ pub enum BucketEntries<T> {
/// ```
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct BucketEntry {
#[serde(skip_serializing_if = "Option::is_none")]
/// The string representation of the bucket.
pub key_as_string: Option<String>,
/// The identifier of the bucket.
pub key: Key,
/// Number of documents in the bucket.
@@ -235,10 +239,16 @@ pub struct RangeBucketEntry {
#[serde(flatten)]
/// sub-aggregations in this bucket.
pub sub_aggregation: AggregationResults,
/// The from range of the bucket. Equals f64::MIN when None.
/// The from range of the bucket. Equals `f64::MIN` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
pub from: Option<f64>,
/// The to range of the bucket. Equals f64::MAX when None.
/// The to range of the bucket. Equals `f64::MAX` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
pub to: Option<f64>,
/// The optional string representation for the `from` range.
#[serde(skip_serializing_if = "Option::is_none")]
pub from_as_string: Option<String>,
/// The optional string representation for the `to` range.
#[serde(skip_serializing_if = "Option::is_none")]
pub to_as_string: Option<String>,
}

204
src/aggregation/bucket/histogram/histogram.rs
View File

@@ -1,6 +1,7 @@
use std::cmp::Ordering;
use std::fmt::Display;
use fastfield_codecs::Column;
use itertools::Itertools;
use serde::{Deserialize, Serialize};
@@ -9,13 +10,12 @@ use crate::aggregation::agg_req_with_accessor::{
AggregationsWithAccessor, BucketAggregationWithAccessor,
};
use crate::aggregation::agg_result::BucketEntry;
use crate::aggregation::f64_from_fastfield_u64;
use crate::aggregation::intermediate_agg_result::{
IntermediateAggregationResults, IntermediateBucketResult, IntermediateHistogramBucketEntry,
};
use crate::aggregation::segment_agg_result::SegmentAggregationResultsCollector;
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader};
use crate::schema::Type;
use crate::aggregation::{f64_from_fastfield_u64, format_date};
use crate::schema::{Schema, Type};
use crate::{DocId, TantivyError};
/// Histogram is a bucket aggregation, where buckets are created dynamically for given `interval`.
@@ -37,14 +37,14 @@ use crate::{DocId, TantivyError};
/// [hard_bounds](HistogramAggregation::hard_bounds).
///
/// # Result
/// Result type is [BucketResult](crate::aggregation::agg_result::BucketResult) with
/// [BucketEntry](crate::aggregation::agg_result::BucketEntry) on the
/// AggregationCollector.
/// Result type is [`BucketResult`](crate::aggregation::agg_result::BucketResult) with
/// [`BucketEntry`](crate::aggregation::agg_result::BucketEntry) on the
/// `AggregationCollector`.
///
/// Result type is
/// [crate::aggregation::intermediate_agg_result::IntermediateBucketResult] with
/// [crate::aggregation::intermediate_agg_result::IntermediateHistogramBucketEntry] on the
/// DistributedAggregationCollector.
/// [`IntermediateBucketResult`](crate::aggregation::intermediate_agg_result::IntermediateBucketResult) with
/// [`IntermediateHistogramBucketEntry`](crate::aggregation::intermediate_agg_result::IntermediateHistogramBucketEntry) on the
/// `DistributedAggregationCollector`.
///
/// # Limitations/Compatibility
///
@@ -61,7 +61,7 @@ use crate::{DocId, TantivyError};
/// ```
///
/// Response
/// See [BucketEntry](crate::aggregation::agg_result::BucketEntry)
/// See [`BucketEntry`](crate::aggregation::agg_result::BucketEntry)
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
pub struct HistogramAggregation {
@@ -70,7 +70,7 @@ pub struct HistogramAggregation {
/// The interval to chunk your data range. Each bucket spans a value range of [0..interval).
/// Must be a positive value.
pub interval: f64,
/// Intervals implicitely defines an absolute grid of buckets `[interval * k, interval * (k +
/// Intervals implicitly defines an absolute grid of buckets `[interval * k, interval * (k +
/// 1))`.
///
/// Offset makes it possible to shift this grid into
@@ -206,6 +206,7 @@ pub struct SegmentHistogramCollector {
field_type: Type,
interval: f64,
offset: f64,
min_doc_count: u64,
first_bucket_num: i64,
bounds: HistogramBounds,
}
@@ -215,6 +216,30 @@ impl SegmentHistogramCollector {
self,
agg_with_accessor: &BucketAggregationWithAccessor,
) -> crate::Result<IntermediateBucketResult> {
// Compute the number of buckets to validate against max num buckets
// Note: We use min_doc_count here, but it's only an lowerbound here, since were are on the
// intermediate level and after merging the number of documents of a bucket could exceed
// `min_doc_count`.
{
let cut_off_buckets_front = self
.buckets
.iter()
.take_while(|bucket| bucket.doc_count <= self.min_doc_count)
.count();
let cut_off_buckets_back = self.buckets[cut_off_buckets_front..]
.iter()
.rev()
.take_while(|bucket| bucket.doc_count <= self.min_doc_count)
.count();
let estimate_num_buckets =
self.buckets.len() - cut_off_buckets_front - cut_off_buckets_back;
agg_with_accessor
.bucket_count
.add_count(estimate_num_buckets as u32);
agg_with_accessor.bucket_count.validate_bucket_count()?;
}
let mut buckets = Vec::with_capacity(
self.buckets
.iter()
@@ -251,11 +276,6 @@ impl SegmentHistogramCollector {
);
};
agg_with_accessor
.bucket_count
.add_count(buckets.len() as u32);
agg_with_accessor.bucket_count.validate_bucket_count()?;
Ok(IntermediateBucketResult::Histogram { buckets })
}
@@ -263,7 +283,7 @@ impl SegmentHistogramCollector {
req: &HistogramAggregation,
sub_aggregation: &AggregationsWithAccessor,
field_type: Type,
accessor: &DynamicFastFieldReader<u64>,
accessor: &dyn Column<u64>,
) -> crate::Result<Self> {
req.validate()?;
let min = f64_from_fastfield_u64(accessor.min_value(), &field_type);
@@ -308,6 +328,7 @@ impl SegmentHistogramCollector {
first_bucket_num,
bounds,
sub_aggregations,
min_doc_count: req.min_doc_count(),
})
}
@@ -331,10 +352,10 @@ impl SegmentHistogramCollector {
.expect("unexpected fast field cardinatility");
let mut iter = doc.chunks_exact(4);
for docs in iter.by_ref() {
let val0 = self.f64_from_fastfield_u64(accessor.get(docs[0]));
let val1 = self.f64_from_fastfield_u64(accessor.get(docs[1]));
let val2 = self.f64_from_fastfield_u64(accessor.get(docs[2]));
let val3 = self.f64_from_fastfield_u64(accessor.get(docs[3]));
let val0 = self.f64_from_fastfield_u64(accessor.get_val(docs[0]));
let val1 = self.f64_from_fastfield_u64(accessor.get_val(docs[1]));
let val2 = self.f64_from_fastfield_u64(accessor.get_val(docs[2]));
let val3 = self.f64_from_fastfield_u64(accessor.get_val(docs[3]));
let bucket_pos0 = get_bucket_num(val0);
let bucket_pos1 = get_bucket_num(val1);
@@ -370,8 +391,8 @@ impl SegmentHistogramCollector {
&bucket_with_accessor.sub_aggregation,
)?;
}
for doc in iter.remainder() {
let val = f64_from_fastfield_u64(accessor.get(*doc), &self.field_type);
for &doc in iter.remainder() {
let val = f64_from_fastfield_u64(accessor.get_val(doc), &self.field_type);
if !bounds.contains(val) {
continue;
}
@@ -380,9 +401,9 @@ impl SegmentHistogramCollector {
debug_assert_eq!(
self.buckets[bucket_pos].key,
get_bucket_val(val, self.interval, self.offset) as f64
get_bucket_val(val, self.interval, self.offset)
);
self.increment_bucket(bucket_pos, *doc, &bucket_with_accessor.sub_aggregation)?;
self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
}
if force_flush {
if let Some(sub_aggregations) = self.sub_aggregations.as_mut() {
@@ -407,7 +428,7 @@ impl SegmentHistogramCollector {
if bounds.contains(val) {
debug_assert_eq!(
self.buckets[bucket_pos].key,
get_bucket_val(val, self.interval, self.offset) as f64
get_bucket_val(val, self.interval, self.offset)
);
self.increment_bucket(bucket_pos, doc, bucket_with_accessor)?;
@@ -425,7 +446,7 @@ impl SegmentHistogramCollector {
let bucket = &mut self.buckets[bucket_pos];
bucket.doc_count += 1;
if let Some(sub_aggregation) = self.sub_aggregations.as_mut() {
(&mut sub_aggregation[bucket_pos]).collect(doc, bucket_with_accessor)?;
sub_aggregation[bucket_pos].collect(doc, bucket_with_accessor)?;
}
Ok(())
}
@@ -451,8 +472,9 @@ fn intermediate_buckets_to_final_buckets_fill_gaps(
buckets: Vec<IntermediateHistogramBucketEntry>,
histogram_req: &HistogramAggregation,
sub_aggregation: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<Vec<BucketEntry>> {
// Generate the the full list of buckets without gaps.
// Generate the full list of buckets without gaps.
//
// The bounds are the min max from the current buckets, optionally extended by
// extended_bounds from the request
@@ -491,7 +513,9 @@ fn intermediate_buckets_to_final_buckets_fill_gaps(
sub_aggregation: empty_sub_aggregation.clone(),
},
})
.map(|intermediate_bucket| intermediate_bucket.into_final_bucket_entry(sub_aggregation))
.map(|intermediate_bucket| {
intermediate_bucket.into_final_bucket_entry(sub_aggregation, schema)
})
.collect::<crate::Result<Vec<_>>>()
}
@@ -500,25 +524,48 @@ pub(crate) fn intermediate_histogram_buckets_to_final_buckets(
buckets: Vec<IntermediateHistogramBucketEntry>,
histogram_req: &HistogramAggregation,
sub_aggregation: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<Vec<BucketEntry>> {
if histogram_req.min_doc_count() == 0 {
let mut buckets = if histogram_req.min_doc_count() == 0 {
// With min_doc_count != 0, we may need to add buckets, so that there are no
// gaps, since intermediate result does not contain empty buckets (filtered to
// reduce serialization size).
intermediate_buckets_to_final_buckets_fill_gaps(buckets, histogram_req, sub_aggregation)
intermediate_buckets_to_final_buckets_fill_gaps(
buckets,
histogram_req,
sub_aggregation,
schema,
)?
} else {
buckets
.into_iter()
.filter(|histogram_bucket| histogram_bucket.doc_count >= histogram_req.min_doc_count())
.map(|histogram_bucket| histogram_bucket.into_final_bucket_entry(sub_aggregation))
.collect::<crate::Result<Vec<_>>>()
.map(|histogram_bucket| {
histogram_bucket.into_final_bucket_entry(sub_aggregation, schema)
})
.collect::<crate::Result<Vec<_>>>()?
};
// If we have a date type on the histogram buckets, we add the `key_as_string` field as rfc339
let field = schema
.get_field(&histogram_req.field)
.ok_or_else(|| TantivyError::FieldNotFound(histogram_req.field.to_string()))?;
if schema.get_field_entry(field).field_type().is_date() {
for bucket in buckets.iter_mut() {
if let crate::aggregation::Key::F64(val) = bucket.key {
let key_as_string = format_date(val as i64)?;
bucket.key_as_string = Some(key_as_string);
}
}
}
Ok(buckets)
}
/// Applies req extended_bounds/hard_bounds on the min_max value
///
/// May return (f64::MAX, f64::MIN), if there is no range.
/// May return `(f64::MAX, f64::MIN)`, if there is no range.
fn get_req_min_max(req: &HistogramAggregation, min_max: Option<(f64, f64)>) -> (f64, f64) {
let (mut min, mut max) = min_max.unwrap_or((f64::MAX, f64::MIN));
@@ -1372,6 +1419,63 @@ mod tests {
Ok(())
}
#[test]
fn histogram_date_test_single_segment() -> crate::Result<()> {
histogram_date_test_with_opt(true)
}
#[test]
fn histogram_date_test_multi_segment() -> crate::Result<()> {
histogram_date_test_with_opt(false)
}
fn histogram_date_test_with_opt(merge_segments: bool) -> crate::Result<()> {
let index = get_test_index_2_segments(merge_segments)?;
let agg_req: Aggregations = vec![(
"histogram".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Histogram(HistogramAggregation {
field: "date".to_string(),
interval: 86400000000.0, // one day in microseconds
..Default::default()
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let agg_res = exec_request(agg_req, &index)?;
let res: Value = serde_json::from_str(&serde_json::to_string(&agg_res)?)?;
assert_eq!(res["histogram"]["buckets"][0]["key"], 1546300800000000.0);
assert_eq!(
res["histogram"]["buckets"][0]["key_as_string"],
"2019-01-01T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][0]["doc_count"], 1);
assert_eq!(res["histogram"]["buckets"][1]["key"], 1546387200000000.0);
assert_eq!(
res["histogram"]["buckets"][1]["key_as_string"],
"2019-01-02T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][1]["doc_count"], 5);
assert_eq!(res["histogram"]["buckets"][2]["key"], 1546473600000000.0);
assert_eq!(
res["histogram"]["buckets"][2]["key_as_string"],
"2019-01-03T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][3], Value::Null);
Ok(())
}
#[test]
fn histogram_invalid_request() -> crate::Result<()> {
let index = get_test_index_2_segments(true)?;
@@ -1438,4 +1542,36 @@ mod tests {
Ok(())
}
#[test]
fn histogram_test_max_buckets_segments() -> crate::Result<()> {
let values = vec![0.0, 70000.0];
let index = get_test_index_from_values(true, &values)?;
let agg_req: Aggregations = vec![(
"my_interval".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Histogram(HistogramAggregation {
field: "score_f64".to_string(),
interval: 1.0,
..Default::default()
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let res = exec_request(agg_req, &index);
assert_eq!(
res.unwrap_err().to_string(),
"An invalid argument was passed: 'Aborting aggregation because too many buckets were \
created'"
.to_string()
);
Ok(())
}
}

6
src/aggregation/bucket/mod.rs
View File

@@ -1,11 +1,11 @@
//! Module for all bucket aggregations.
//!
//! BucketAggregations create buckets of documents
//! [BucketAggregation](super::agg_req::BucketAggregation).
//! [`BucketAggregation`](super::agg_req::BucketAggregation).
//!
//! Results of final buckets are [BucketResult](super::agg_result::BucketResult).
//! Results of final buckets are [`BucketResult`](super::agg_result::BucketResult).
//! Results of intermediate buckets are
//! [IntermediateBucketResult](super::intermediate_agg_result::IntermediateBucketResult)
//! [`IntermediateBucketResult`](super::intermediate_agg_result::IntermediateBucketResult)
mod histogram;
mod range;

158
src/aggregation/bucket/range.rs
View File

@@ -1,7 +1,8 @@
use std::fmt::Debug;
use std::ops::Range;
use fnv::FnvHashMap;
use fastfield_codecs::MonotonicallyMappableToU64;
use rustc_hash::FxHashMap;
use serde::{Deserialize, Serialize};
use crate::aggregation::agg_req_with_accessor::{
@@ -11,8 +12,9 @@ use crate::aggregation::intermediate_agg_result::{
IntermediateBucketResult, IntermediateRangeBucketEntry, IntermediateRangeBucketResult,
};
use crate::aggregation::segment_agg_result::{BucketCount, SegmentAggregationResultsCollector};
use crate::aggregation::{f64_from_fastfield_u64, f64_to_fastfield_u64, Key, SerializedKey};
use crate::fastfield::FastFieldReader;
use crate::aggregation::{
f64_from_fastfield_u64, f64_to_fastfield_u64, format_date, Key, SerializedKey,
};
use crate::schema::Type;
use crate::{DocId, TantivyError};
@@ -23,14 +25,14 @@ use crate::{DocId, TantivyError};
/// against each bucket range. Note that this aggregation includes the from value and excludes the
/// to value for each range.
///
/// Result type is [BucketResult](crate::aggregation::agg_result::BucketResult) with
/// [RangeBucketEntry](crate::aggregation::agg_result::RangeBucketEntry) on the
/// AggregationCollector.
/// Result type is [`BucketResult`](crate::aggregation::agg_result::BucketResult) with
/// [`RangeBucketEntry`](crate::aggregation::agg_result::RangeBucketEntry) on the
/// `AggregationCollector`.
///
/// Result type is
/// [crate::aggregation::intermediate_agg_result::IntermediateBucketResult] with
/// [crate::aggregation::intermediate_agg_result::IntermediateRangeBucketEntry] on the
/// DistributedAggregationCollector.
/// [`IntermediateBucketResult`](crate::aggregation::intermediate_agg_result::IntermediateBucketResult) with
/// [`IntermediateRangeBucketEntry`](crate::aggregation::intermediate_agg_result::IntermediateRangeBucketEntry) on the
/// `DistributedAggregationCollector`.
///
/// # Limitations/Compatibility
/// Overlapping ranges are not yet supported.
@@ -68,11 +70,11 @@ pub struct RangeAggregationRange {
#[serde(skip_serializing_if = "Option::is_none", default)]
pub key: Option<String>,
/// The from range value, which is inclusive in the range.
/// None equals to an open ended interval.
/// `None` equals to an open ended interval.
#[serde(skip_serializing_if = "Option::is_none", default)]
pub from: Option<f64>,
/// The to range value, which is not inclusive in the range.
/// None equals to an open ended interval.
/// `None` equals to an open ended interval.
#[serde(skip_serializing_if = "Option::is_none", default)]
pub to: Option<f64>,
}
@@ -102,7 +104,7 @@ impl From<Range<f64>> for RangeAggregationRange {
pub(crate) struct InternalRangeAggregationRange {
/// Custom key for the range bucket
key: Option<String>,
/// u64 range value
/// `u64` range value
range: Range<u64>,
}
@@ -132,9 +134,9 @@ pub(crate) struct SegmentRangeBucketEntry {
pub key: Key,
pub doc_count: u64,
pub sub_aggregation: Option<SegmentAggregationResultsCollector>,
/// The from range of the bucket. Equals f64::MIN when None.
/// The from range of the bucket. Equals `f64::MIN` when `None`.
pub from: Option<f64>,
/// The to range of the bucket. Equals f64::MAX when None. Open interval, `to` is not
/// The to range of the bucket. Equals `f64::MAX` when `None`. Open interval, `to` is not
/// inclusive.
pub to: Option<f64>,
}
@@ -177,12 +179,12 @@ impl SegmentRangeCollector {
) -> crate::Result<IntermediateBucketResult> {
let field_type = self.field_type;
let buckets: FnvHashMap<SerializedKey, IntermediateRangeBucketEntry> = self
let buckets: FxHashMap<SerializedKey, IntermediateRangeBucketEntry> = self
.buckets
.into_iter()
.map(move |range_bucket| {
Ok((
range_to_string(&range_bucket.range, &field_type),
range_to_string(&range_bucket.range, &field_type)?,
range_bucket
.bucket
.into_intermediate_bucket_entry(&agg_with_accessor.sub_aggregation)?,
@@ -210,8 +212,8 @@ impl SegmentRangeCollector {
let key = range
.key
.clone()
.map(|key| Key::Str(key))
.unwrap_or(range_to_key(&range.range, &field_type));
.map(|key| Ok(Key::Str(key)))
.unwrap_or_else(|| range_to_key(&range.range, &field_type))?;
let to = if range.range.end == u64::MAX {
None
} else {
@@ -229,6 +231,7 @@ impl SegmentRangeCollector {
sub_aggregation,
)?)
};
Ok(SegmentRangeAndBucketEntry {
range: range.range.clone(),
bucket: SegmentRangeBucketEntry {
@@ -262,12 +265,12 @@ impl SegmentRangeCollector {
let accessor = bucket_with_accessor
.accessor
.as_single()
.expect("unexpected fast field cardinatility");
.expect("unexpected fast field cardinality");
for docs in iter.by_ref() {
let val1 = accessor.get(docs[0]);
let val2 = accessor.get(docs[1]);
let val3 = accessor.get(docs[2]);
let val4 = accessor.get(docs[3]);
let val1 = accessor.get_val(docs[0]);
let val2 = accessor.get_val(docs[1]);
let val3 = accessor.get_val(docs[2]);
let val4 = accessor.get_val(docs[3]);
let bucket_pos1 = self.get_bucket_pos(val1);
let bucket_pos2 = self.get_bucket_pos(val2);
let bucket_pos3 = self.get_bucket_pos(val3);
@@ -278,10 +281,10 @@ impl SegmentRangeCollector {
self.increment_bucket(bucket_pos3, docs[2], &bucket_with_accessor.sub_aggregation)?;
self.increment_bucket(bucket_pos4, docs[3], &bucket_with_accessor.sub_aggregation)?;
}
for doc in iter.remainder() {
let val = accessor.get(*doc);
for &doc in iter.remainder() {
let val = accessor.get_val(doc);
let bucket_pos = self.get_bucket_pos(val);
self.increment_bucket(bucket_pos, *doc, &bucket_with_accessor.sub_aggregation)?;
self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
}
if force_flush {
for bucket in &mut self.buckets {
@@ -324,8 +327,8 @@ impl SegmentRangeCollector {
/// Converts the user provided f64 range value to fast field value space.
///
/// Internally fast field values are always stored as u64.
/// If the fast field has u64 [1,2,5], these values are stored as is in the fast field.
/// A fast field with f64 [1.0, 2.0, 5.0] is converted to u64 space, using a
/// If the fast field has u64 `[1, 2, 5]`, these values are stored as is in the fast field.
/// A fast field with f64 `[1.0, 2.0, 5.0]` is converted to u64 space, using a
/// monotonic mapping function, so the order is preserved.
///
/// Consequently, a f64 user range 1.0..3.0 needs to be converted to fast field value space using
@@ -403,33 +406,45 @@ fn extend_validate_ranges(
Ok(converted_buckets)
}
pub(crate) fn range_to_string(range: &Range<u64>, field_type: &Type) -> String {
pub(crate) fn range_to_string(range: &Range<u64>, field_type: &Type) -> crate::Result<String> {
// is_start is there for malformed requests, e.g. ig the user passes the range u64::MIN..0.0,
// it should be rendered as "*-0" and not "*-*"
let to_str = |val: u64, is_start: bool| {
if (is_start && val == u64::MIN) || (!is_start && val == u64::MAX) {
"*".to_string()
Ok("*".to_string())
} else if *field_type == Type::Date {
let val = i64::from_u64(val);
format_date(val)
} else {
f64_from_fastfield_u64(val, field_type).to_string()
Ok(f64_from_fastfield_u64(val, field_type).to_string())
}
};
format!("{}-{}", to_str(range.start, true), to_str(range.end, false))
Ok(format!(
"{}-{}",
to_str(range.start, true)?,
to_str(range.end, false)?
))
}
pub(crate) fn range_to_key(range: &Range<u64>, field_type: &Type) -> Key {
Key::Str(range_to_string(range, field_type))
pub(crate) fn range_to_key(range: &Range<u64>, field_type: &Type) -> crate::Result<Key> {
Ok(Key::Str(range_to_string(range, field_type)?))
}
#[cfg(test)]
mod tests {
use fastfield_codecs::MonotonicallyMappableToU64;
use serde_json::Value;
use super::*;
use crate::aggregation::agg_req::{
Aggregation, Aggregations, BucketAggregation, BucketAggregationType,
};
use crate::aggregation::tests::{exec_request_with_query, get_test_index_with_num_docs};
use crate::fastfield::FastValue;
use crate::aggregation::tests::{
exec_request, exec_request_with_query, get_test_index_2_segments,
get_test_index_with_num_docs,
};
pub fn get_collector_from_ranges(
ranges: Vec<RangeAggregationRange>,
@@ -567,6 +582,77 @@ mod tests {
Ok(())
}
#[test]
fn range_date_test_single_segment() -> crate::Result<()> {
range_date_test_with_opt(true)
}
#[test]
fn range_date_test_multi_segment() -> crate::Result<()> {
range_date_test_with_opt(false)
}
fn range_date_test_with_opt(merge_segments: bool) -> crate::Result<()> {
let index = get_test_index_2_segments(merge_segments)?;
let agg_req: Aggregations = vec![(
"date_ranges".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Range(RangeAggregation {
field: "date".to_string(),
ranges: vec![
RangeAggregationRange {
key: None,
from: None,
to: Some(1546300800000000.0f64),
},
RangeAggregationRange {
key: None,
from: Some(1546300800000000.0f64),
to: Some(1546387200000000.0f64),
},
],
keyed: false,
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let agg_res = exec_request(agg_req, &index)?;
let res: Value = serde_json::from_str(&serde_json::to_string(&agg_res)?)?;
assert_eq!(
res["date_ranges"]["buckets"][0]["from_as_string"],
Value::Null
);
assert_eq!(
res["date_ranges"]["buckets"][0]["key"],
"*-2019-01-01T00:00:00Z"
);
assert_eq!(
res["date_ranges"]["buckets"][1]["from_as_string"],
"2019-01-01T00:00:00Z"
);
assert_eq!(
res["date_ranges"]["buckets"][1]["to_as_string"],
"2019-01-02T00:00:00Z"
);
assert_eq!(
res["date_ranges"]["buckets"][2]["from_as_string"],
"2019-01-02T00:00:00Z"
);
assert_eq!(
res["date_ranges"]["buckets"][2]["to_as_string"],
Value::Null
);
Ok(())
}
#[test]
fn range_custom_key_keyed_buckets_test() -> crate::Result<()> {
let index = get_test_index_with_num_docs(false, 100)?;

85
src/aggregation/bucket/term_agg.rs
View File

@@ -1,7 +1,7 @@
use std::fmt::Debug;
use fnv::FnvHashMap;
use itertools::Itertools;
use rustc_hash::FxHashMap;
use serde::{Deserialize, Serialize};
use super::{CustomOrder, Order, OrderTarget};
@@ -17,7 +17,11 @@ use crate::fastfield::MultiValuedFastFieldReader;
use crate::schema::Type;
use crate::{DocId, TantivyError};
/// Creates a bucket for every unique term
/// Creates a bucket for every unique term and counts the number of occurences.
/// Note that doc_count in the response buckets equals term count here.
///
/// If the text is untokenized and single value, that means one term per document and therefore it
/// is in fact doc count.
///
/// ### Terminology
/// Shard parameters are supposed to be equivalent to elasticsearch shard parameter.
@@ -31,7 +35,7 @@ use crate::{DocId, TantivyError};
///
/// Even with a larger `segment_size` value, doc_count values for a terms aggregation may be
/// approximate. As a result, any sub-aggregations on the terms aggregation may also be approximate.
/// `sum_other_doc_count` is the number of documents that didn’t make it into the the top size
/// `sum_other_doc_count` is the number of documents that didn’t make it into the top size
/// terms. If this is greater than 0, you can be sure that the terms agg had to throw away some
/// buckets, either because they didn’t fit into size on the root node or they didn’t fit into
/// `segment_size` on the segment node.
@@ -42,14 +46,14 @@ use crate::{DocId, TantivyError};
/// each segment. It’s the sum of the size of the largest bucket on each segment that didn’t fit
/// into segment_size.
///
/// Result type is [BucketResult](crate::aggregation::agg_result::BucketResult) with
/// [TermBucketEntry](crate::aggregation::agg_result::BucketEntry) on the
/// AggregationCollector.
/// Result type is [`BucketResult`](crate::aggregation::agg_result::BucketResult) with
/// [`TermBucketEntry`](crate::aggregation::agg_result::BucketEntry) on the
/// `AggregationCollector`.
///
/// Result type is
/// [crate::aggregation::intermediate_agg_result::IntermediateBucketResult] with
/// [crate::aggregation::intermediate_agg_result::IntermediateTermBucketEntry] on the
/// DistributedAggregationCollector.
/// [`IntermediateBucketResult`](crate::aggregation::intermediate_agg_result::IntermediateBucketResult) with
/// [`IntermediateTermBucketEntry`](crate::aggregation::intermediate_agg_result::IntermediateTermBucketEntry) on the
/// `DistributedAggregationCollector`.
///
/// # Limitations/Compatibility
///
@@ -64,6 +68,25 @@ use crate::{DocId, TantivyError};
/// }
/// }
/// ```
///
/// /// # Response JSON Format
/// ```json
/// {
/// ...
/// "aggregations": {
/// "genres": {
/// "doc_count_error_upper_bound": 0,
/// "sum_other_doc_count": 0,
/// "buckets": [
/// { "key": "drumnbass", "doc_count": 6 },
/// { "key": "raggae", "doc_count": 4 },
/// { "key": "jazz", "doc_count": 2 }
/// ]
/// }
/// }
/// }
/// ```
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
pub struct TermsAggregation {
/// The field to aggregate on.
@@ -110,8 +133,8 @@ pub struct TermsAggregation {
/// Set the order. `String` is here a target, which is either "_count", "_key", or the name of
/// a metric sub_aggregation.
///
/// Single value metrics like average can be adressed by its name.
/// Multi value metrics like stats are required to adress their field by name e.g.
/// Single value metrics like average can be addressed by its name.
/// Multi value metrics like stats are required to address their field by name e.g.
/// "stats.avg"
///
/// Examples in JSON format:
@@ -176,7 +199,7 @@ impl TermsAggregationInternal {
#[derive(Clone, Debug, PartialEq)]
/// Container to store term_ids and their buckets.
struct TermBuckets {
pub(crate) entries: FnvHashMap<u32, TermBucketEntry>,
pub(crate) entries: FxHashMap<u32, TermBucketEntry>,
blueprint: Option<SegmentAggregationResultsCollector>,
}
@@ -374,7 +397,7 @@ impl SegmentTermCollector {
.expect("internal error: inverted index not loaded for term aggregation");
let term_dict = inverted_index.terms();
let mut dict: FnvHashMap<String, IntermediateTermBucketEntry> = Default::default();
let mut dict: FxHashMap<String, IntermediateTermBucketEntry> = Default::default();
let mut buffer = vec![];
for (term_id, entry) in entries {
term_dict
@@ -1106,9 +1129,9 @@ mod tests {
assert_eq!(res["my_texts"]["buckets"][0]["key"], "terma");
assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 4);
assert_eq!(res["my_texts"]["buckets"][1]["key"], "termb");
assert_eq!(res["my_texts"]["buckets"][1]["key"], "termc");
assert_eq!(res["my_texts"]["buckets"][1]["doc_count"], 0);
assert_eq!(res["my_texts"]["buckets"][2]["key"], "termc");
assert_eq!(res["my_texts"]["buckets"][2]["key"], "termb");
assert_eq!(res["my_texts"]["buckets"][2]["doc_count"], 0);
assert_eq!(res["my_texts"]["sum_other_doc_count"], 0);
assert_eq!(res["my_texts"]["doc_count_error_upper_bound"], 0);
@@ -1206,11 +1229,43 @@ mod tests {
.collect();
let res = exec_request_with_query(agg_req, &index, None);
assert!(res.is_err());
Ok(())
}
#[test]
fn terms_aggregation_multi_token_per_doc() -> crate::Result<()> {
let terms = vec!["Hello Hello", "Hallo Hallo"];
let index = get_test_index_from_terms(true, &[terms])?;
let agg_req: Aggregations = vec![(
"my_texts".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Terms(TermsAggregation {
field: "text_id".to_string(),
min_doc_count: Some(0),
..Default::default()
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let res = exec_request_with_query(agg_req, &index, None).unwrap();
assert_eq!(res["my_texts"]["buckets"][0]["key"], "hello");
assert_eq!(res["my_texts"]["buckets"][0]["doc_count"], 2);
assert_eq!(res["my_texts"]["buckets"][1]["key"], "hallo");
assert_eq!(res["my_texts"]["buckets"][1]["doc_count"], 2);
Ok(())
}
#[test]
fn test_json_format() -> crate::Result<()> {
let agg_req: Aggregations = vec![(

15
src/aggregation/collector.rs
View File

@@ -7,6 +7,7 @@ use super::intermediate_agg_result::IntermediateAggregationResults;
use super::segment_agg_result::SegmentAggregationResultsCollector;
use crate::aggregation::agg_req_with_accessor::get_aggs_with_accessor_and_validate;
use crate::collector::{Collector, SegmentCollector};
use crate::schema::Schema;
use crate::{SegmentReader, TantivyError};
/// The default max bucket count, before the aggregation fails.
@@ -16,6 +17,7 @@ pub const MAX_BUCKET_COUNT: u32 = 65000;
///
/// The collector collects all aggregations by the underlying aggregation request.
pub struct AggregationCollector {
schema: Schema,
agg: Aggregations,
max_bucket_count: u32,
}
@@ -25,8 +27,9 @@ impl AggregationCollector {
///
/// Aggregation fails when the total bucket count is higher than max_bucket_count.
/// max_bucket_count will default to `MAX_BUCKET_COUNT` (65000) when unset
pub fn from_aggs(agg: Aggregations, max_bucket_count: Option<u32>) -> Self {
pub fn from_aggs(agg: Aggregations, max_bucket_count: Option<u32>, schema: Schema) -> Self {
Self {
schema,
agg,
max_bucket_count: max_bucket_count.unwrap_or(MAX_BUCKET_COUNT),
}
@@ -39,7 +42,7 @@ impl AggregationCollector {
///
/// # Purpose
/// AggregationCollector returns `IntermediateAggregationResults` and not the final
/// `AggregationResults`, so that results from differenct indices can be merged and then converted
/// `AggregationResults`, so that results from different indices can be merged and then converted
/// into the final `AggregationResults` via the `into_final_result()` method.
pub struct DistributedAggregationCollector {
agg: Aggregations,
@@ -113,7 +116,7 @@ impl Collector for AggregationCollector {
segment_fruits: Vec<<Self::Child as SegmentCollector>::Fruit>,
) -> crate::Result<Self::Fruit> {
let res = merge_fruits(segment_fruits)?;
res.into_final_bucket_result(self.agg.clone())
res.into_final_bucket_result(self.agg.clone(), &self.schema)
}
}
@@ -131,7 +134,7 @@ fn merge_fruits(
}
}
/// AggregationSegmentCollector does the aggregation collection on a segment.
/// `AggregationSegmentCollector` does the aggregation collection on a segment.
pub struct AggregationSegmentCollector {
aggs_with_accessor: AggregationsWithAccessor,
result: SegmentAggregationResultsCollector,
@@ -139,8 +142,8 @@ pub struct AggregationSegmentCollector {
}
impl AggregationSegmentCollector {
/// Creates an AggregationSegmentCollector from an [Aggregations] request and a segment reader.
/// Also includes validation, e.g. checking field types and existence.
/// Creates an `AggregationSegmentCollector from` an [`Aggregations`] request and a segment
/// reader. Also includes validation, e.g. checking field types and existence.
pub fn from_agg_req_and_reader(
agg: &Aggregations,
reader: &SegmentReader,

18
src/aggregation/date.rs Normal file
View File

@@ -0,0 +1,18 @@
use time::format_description::well_known::Rfc3339;
use time::OffsetDateTime;
use crate::TantivyError;
pub(crate) fn format_date(val: i64) -> crate::Result<String> {
let datetime =
OffsetDateTime::from_unix_timestamp_nanos(1_000 * (val as i128)).map_err(|err| {
TantivyError::InvalidArgument(format!(
"Could not convert {:?} to OffsetDateTime, err {:?}",
val, err
))
})?;
let key_as_string = datetime
.format(&Rfc3339)
.map_err(|_err| TantivyError::InvalidArgument("Could not serialize date".to_string()))?;
Ok(key_as_string)
}

115
src/aggregation/intermediate_agg_result.rs
View File

@@ -3,15 +3,14 @@
//! indices.
use std::cmp::Ordering;
use std::collections::HashMap;
use fnv::FnvHashMap;
use itertools::Itertools;
use rustc_hash::FxHashMap;
use serde::{Deserialize, Serialize};
use super::agg_req::{
Aggregations, AggregationsInternal, BucketAggregationInternal, BucketAggregationType,
MetricAggregation,
MetricAggregation, RangeAggregation,
};
use super::agg_result::{AggregationResult, BucketResult, RangeBucketEntry};
use super::bucket::{
@@ -20,9 +19,11 @@ use super::bucket::{
};
use super::metric::{IntermediateAverage, IntermediateStats};
use super::segment_agg_result::SegmentMetricResultCollector;
use super::{Key, SerializedKey, VecWithNames};
use super::{format_date, Key, SerializedKey, VecWithNames};
use crate::aggregation::agg_result::{AggregationResults, BucketEntries, BucketEntry};
use crate::aggregation::bucket::TermsAggregationInternal;
use crate::schema::Schema;
use crate::TantivyError;
/// Contains the intermediate aggregation result, which is optimized to be merged with other
/// intermediate results.
@@ -36,29 +37,34 @@ pub struct IntermediateAggregationResults {
impl IntermediateAggregationResults {
/// Convert intermediate result and its aggregation request to the final result.
pub fn into_final_bucket_result(self, req: Aggregations) -> crate::Result<AggregationResults> {
self.into_final_bucket_result_internal(&(req.into()))
pub fn into_final_bucket_result(
self,
req: Aggregations,
schema: &Schema,
) -> crate::Result<AggregationResults> {
self.into_final_bucket_result_internal(&(req.into()), schema)
}
/// Convert intermediate result and its aggregation request to the final result.
///
/// Internal function, AggregationsInternal is used instead Aggregations, which is optimized
/// for internal processing, by splitting metric and buckets into seperate groups.
/// for internal processing, by splitting metric and buckets into separate groups.
pub(crate) fn into_final_bucket_result_internal(
self,
req: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<AggregationResults> {
// Important assumption:
// When the tree contains buckets/metric, we expect it to have all buckets/metrics from the
// request
let mut results: HashMap<String, AggregationResult> = HashMap::new();
let mut results: FxHashMap<String, AggregationResult> = FxHashMap::default();
if let Some(buckets) = self.buckets {
convert_and_add_final_buckets_to_result(&mut results, buckets, &req.buckets)?
convert_and_add_final_buckets_to_result(&mut results, buckets, &req.buckets, schema)?
} else {
// When there are no buckets, we create empty buckets, so that the serialized json
// format is constant
add_empty_final_buckets_to_result(&mut results, &req.buckets)?
add_empty_final_buckets_to_result(&mut results, &req.buckets, schema)?
};
if let Some(metrics) = self.metrics {
@@ -108,10 +114,10 @@ impl IntermediateAggregationResults {
Self { metrics, buckets }
}
/// Merge an other intermediate aggregation result into this result.
/// Merge another intermediate aggregation result into this result.
///
/// The order of the values need to be the same on both results. This is ensured when the same
/// (key values) are present on the underlying VecWithNames struct.
/// (key values) are present on the underlying `VecWithNames` struct.
pub fn merge_fruits(&mut self, other: IntermediateAggregationResults) {
if let (Some(buckets_left), Some(buckets_right)) = (&mut self.buckets, other.buckets) {
for (bucket_left, bucket_right) in
@@ -132,7 +138,7 @@ impl IntermediateAggregationResults {
}
fn convert_and_add_final_metrics_to_result(
results: &mut HashMap<String, AggregationResult>,
results: &mut FxHashMap<String, AggregationResult>,
metrics: VecWithNames<IntermediateMetricResult>,
) {
results.extend(
@@ -143,7 +149,7 @@ fn convert_and_add_final_metrics_to_result(
}
fn add_empty_final_metrics_to_result(
results: &mut HashMap<String, AggregationResult>,
results: &mut FxHashMap<String, AggregationResult>,
req_metrics: &VecWithNames<MetricAggregation>,
) -> crate::Result<()> {
results.extend(req_metrics.iter().map(|(key, req)| {
@@ -157,27 +163,30 @@ fn add_empty_final_metrics_to_result(
}
fn add_empty_final_buckets_to_result(
results: &mut HashMap<String, AggregationResult>,
results: &mut FxHashMap<String, AggregationResult>,
req_buckets: &VecWithNames<BucketAggregationInternal>,
schema: &Schema,
) -> crate::Result<()> {
let requested_buckets = req_buckets.iter();
for (key, req) in requested_buckets {
let empty_bucket = AggregationResult::BucketResult(BucketResult::empty_from_req(req)?);
let empty_bucket =
AggregationResult::BucketResult(BucketResult::empty_from_req(req, schema)?);
results.insert(key.to_string(), empty_bucket);
}
Ok(())
}
fn convert_and_add_final_buckets_to_result(
results: &mut HashMap<String, AggregationResult>,
results: &mut FxHashMap<String, AggregationResult>,
buckets: VecWithNames<IntermediateBucketResult>,
req_buckets: &VecWithNames<BucketAggregationInternal>,
schema: &Schema,
) -> crate::Result<()> {
assert_eq!(buckets.len(), req_buckets.len());
let buckets_with_request = buckets.into_iter().zip(req_buckets.values());
for ((key, bucket), req) in buckets_with_request {
let result = AggregationResult::BucketResult(bucket.into_final_bucket_result(req)?);
let result = AggregationResult::BucketResult(bucket.into_final_bucket_result(req, schema)?);
results.insert(key, result);
}
Ok(())
@@ -267,13 +276,21 @@ impl IntermediateBucketResult {
pub(crate) fn into_final_bucket_result(
self,
req: &BucketAggregationInternal,
schema: &Schema,
) -> crate::Result<BucketResult> {
match self {
IntermediateBucketResult::Range(range_res) => {
let mut buckets: Vec<RangeBucketEntry> = range_res
.buckets
.into_iter()
.map(|(_, bucket)| bucket.into_final_bucket_entry(&req.sub_aggregation))
.into_values()
.map(|bucket| {
bucket.into_final_bucket_entry(
&req.sub_aggregation,
schema,
req.as_range()
.expect("unexpected aggregation, expected histogram aggregation"),
)
})
.collect::<crate::Result<Vec<_>>>()?;
buckets.sort_by(|left, right| {
@@ -288,7 +305,7 @@ impl IntermediateBucketResult {
.keyed;
let buckets = if is_keyed {
let mut bucket_map =
FnvHashMap::with_capacity_and_hasher(buckets.len(), Default::default());
FxHashMap::with_capacity_and_hasher(buckets.len(), Default::default());
for bucket in buckets {
bucket_map.insert(bucket.key.to_string(), bucket);
}
@@ -304,11 +321,12 @@ impl IntermediateBucketResult {
req.as_histogram()
.expect("unexpected aggregation, expected histogram aggregation"),
&req.sub_aggregation,
schema,
)?;
let buckets = if req.as_histogram().unwrap().keyed {
let mut bucket_map =
FnvHashMap::with_capacity_and_hasher(buckets.len(), Default::default());
FxHashMap::with_capacity_and_hasher(buckets.len(), Default::default());
for bucket in buckets {
bucket_map.insert(bucket.key.to_string(), bucket);
}
@@ -322,6 +340,7 @@ impl IntermediateBucketResult {
req.as_term()
.expect("unexpected aggregation, expected term aggregation"),
&req.sub_aggregation,
schema,
),
}
}
@@ -396,13 +415,13 @@ impl IntermediateBucketResult {
#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
/// Range aggregation including error counts
pub struct IntermediateRangeBucketResult {
pub(crate) buckets: FnvHashMap<SerializedKey, IntermediateRangeBucketEntry>,
pub(crate) buckets: FxHashMap<SerializedKey, IntermediateRangeBucketEntry>,
}
#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize)]
/// Term aggregation including error counts
pub struct IntermediateTermBucketResult {
pub(crate) entries: FnvHashMap<String, IntermediateTermBucketEntry>,
pub(crate) entries: FxHashMap<String, IntermediateTermBucketEntry>,
pub(crate) sum_other_doc_count: u64,
pub(crate) doc_count_error_upper_bound: u64,
}
@@ -412,6 +431,7 @@ impl IntermediateTermBucketResult {
self,
req: &TermsAggregation,
sub_aggregation_req: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<BucketResult> {
let req = TermsAggregationInternal::from_req(req);
let mut buckets: Vec<BucketEntry> = self
@@ -420,11 +440,12 @@ impl IntermediateTermBucketResult {
.filter(|bucket| bucket.1.doc_count >= req.min_doc_count)
.map(|(key, entry)| {
Ok(BucketEntry {
key_as_string: None,
key: Key::Str(key),
doc_count: entry.doc_count,
sub_aggregation: entry
.sub_aggregation
.into_final_bucket_result_internal(sub_aggregation_req)?,
.into_final_bucket_result_internal(sub_aggregation_req, schema)?,
})
})
.collect::<crate::Result<_>>()?;
@@ -499,8 +520,8 @@ trait MergeFruits {
}
fn merge_maps<V: MergeFruits + Clone>(
entries_left: &mut FnvHashMap<SerializedKey, V>,
mut entries_right: FnvHashMap<SerializedKey, V>,
entries_left: &mut FxHashMap<SerializedKey, V>,
mut entries_right: FxHashMap<SerializedKey, V>,
) {
for (name, entry_left) in entries_left.iter_mut() {
if let Some(entry_right) = entries_right.remove(name) {
@@ -529,13 +550,15 @@ impl IntermediateHistogramBucketEntry {
pub(crate) fn into_final_bucket_entry(
self,
req: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<BucketEntry> {
Ok(BucketEntry {
key_as_string: None,
key: Key::F64(self.key),
doc_count: self.doc_count,
sub_aggregation: self
.sub_aggregation
.into_final_bucket_result_internal(req)?,
.into_final_bucket_result_internal(req, schema)?,
})
}
}
@@ -560,10 +583,10 @@ pub struct IntermediateRangeBucketEntry {
pub doc_count: u64,
/// The sub_aggregation in this bucket.
pub sub_aggregation: IntermediateAggregationResults,
/// The from range of the bucket. Equals f64::MIN when None.
/// The from range of the bucket. Equals `f64::MIN` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
pub from: Option<f64>,
/// The to range of the bucket. Equals f64::MAX when None.
/// The to range of the bucket. Equals `f64::MAX` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
pub to: Option<f64>,
}
@@ -572,16 +595,38 @@ impl IntermediateRangeBucketEntry {
pub(crate) fn into_final_bucket_entry(
self,
req: &AggregationsInternal,
schema: &Schema,
range_req: &RangeAggregation,
) -> crate::Result<RangeBucketEntry> {
Ok(RangeBucketEntry {
let mut range_bucket_entry = RangeBucketEntry {
key: self.key,
doc_count: self.doc_count,
sub_aggregation: self
.sub_aggregation
.into_final_bucket_result_internal(req)?,
.into_final_bucket_result_internal(req, schema)?,
to: self.to,
from: self.from,
})
to_as_string: None,
from_as_string: None,
};
// If we have a date type on the histogram buckets, we add the `key_as_string` field as
// rfc339
let field = schema
.get_field(&range_req.field)
.ok_or_else(|| TantivyError::FieldNotFound(range_req.field.to_string()))?;
if schema.get_field_entry(field).field_type().is_date() {
if let Some(val) = range_bucket_entry.to {
let key_as_string = format_date(val as i64)?;
range_bucket_entry.to_as_string = Some(key_as_string);
}
if let Some(val) = range_bucket_entry.from {
let key_as_string = format_date(val as i64)?;
range_bucket_entry.from_as_string = Some(key_as_string);
}
}
Ok(range_bucket_entry)
}
}
@@ -626,7 +671,7 @@ mod tests {
fn get_sub_test_tree(data: &[(String, u64)]) -> IntermediateAggregationResults {
let mut map = HashMap::new();
let mut buckets = FnvHashMap::default();
let mut buckets = FxHashMap::default();
for (key, doc_count) in data {
buckets.insert(
key.to_string(),
@@ -653,7 +698,7 @@ mod tests {
data: &[(String, u64, String, u64)],
) -> IntermediateAggregationResults {
let mut map = HashMap::new();
let mut buckets: FnvHashMap<_, _> = Default::default();
let mut buckets: FxHashMap<_, _> = Default::default();
for (key, doc_count, sub_aggregation_key, sub_aggregation_count) in data {
buckets.insert(
key.to_string(),

16
src/aggregation/metric/average.rs
View File

@@ -1,9 +1,9 @@
use std::fmt::Debug;
use fastfield_codecs::Column;
use serde::{Deserialize, Serialize};
use crate::aggregation::f64_from_fastfield_u64;
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader};
use crate::schema::Type;
use crate::DocId;
@@ -57,13 +57,13 @@ impl SegmentAverageCollector {
data: Default::default(),
}
}
pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &DynamicFastFieldReader<u64>) {
pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &dyn Column<u64>) {
let mut iter = doc.chunks_exact(4);
for docs in iter.by_ref() {
let val1 = field.get(docs[0]);
let val2 = field.get(docs[1]);
let val3 = field.get(docs[2]);
let val4 = field.get(docs[3]);
let val1 = field.get_val(docs[0]);
let val2 = field.get_val(docs[1]);
let val3 = field.get_val(docs[2]);
let val4 = field.get_val(docs[3]);
let val1 = f64_from_fastfield_u64(val1, &self.field_type);
let val2 = f64_from_fastfield_u64(val2, &self.field_type);
let val3 = f64_from_fastfield_u64(val3, &self.field_type);
@@ -73,8 +73,8 @@ impl SegmentAverageCollector {
self.data.collect(val3);
self.data.collect(val4);
}
for doc in iter.remainder() {
let val = field.get(*doc);
for &doc in iter.remainder() {
let val = field.get_val(doc);
let val = f64_from_fastfield_u64(val, &self.field_type);
self.data.collect(val);
}

30
src/aggregation/metric/stats.rs
View File

@@ -1,14 +1,14 @@
use fastfield_codecs::Column;
use serde::{Deserialize, Serialize};
use crate::aggregation::f64_from_fastfield_u64;
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader};
use crate::schema::Type;
use crate::{DocId, TantivyError};
/// A multi-value metric aggregation that computes stats of numeric values that are
/// extracted from the aggregated documents.
/// Supported field types are u64, i64, and f64.
/// See [Stats] for returned statistics.
/// Supported field types are `u64`, `i64`, and `f64`.
/// See [`Stats`] for returned statistics.
///
/// # JSON Format
/// ```json
@@ -43,13 +43,13 @@ pub struct Stats {
pub count: usize,
/// The sum of the fast field values.
pub sum: f64,
/// The standard deviation of the fast field values. None for count == 0.
/// The standard deviation of the fast field values. `None` for count == 0.
pub standard_deviation: Option<f64>,
/// The min value of the fast field values.
pub min: Option<f64>,
/// The max value of the fast field values.
pub max: Option<f64>,
/// The average of the values. None for count == 0.
/// The average of the values. `None` for count == 0.
pub avg: Option<f64>,
}
@@ -70,7 +70,7 @@ impl Stats {
}
}
/// IntermediateStats contains the mergeable version for stats.
/// `IntermediateStats` contains the mergeable version for stats.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct IntermediateStats {
count: usize,
@@ -163,13 +163,13 @@ impl SegmentStatsCollector {
stats: IntermediateStats::default(),
}
}
pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &DynamicFastFieldReader<u64>) {
pub(crate) fn collect_block(&mut self, doc: &[DocId], field: &dyn Column<u64>) {
let mut iter = doc.chunks_exact(4);
for docs in iter.by_ref() {
let val1 = field.get(docs[0]);
let val2 = field.get(docs[1]);
let val3 = field.get(docs[2]);
let val4 = field.get(docs[3]);
let val1 = field.get_val(docs[0]);
let val2 = field.get_val(docs[1]);
let val3 = field.get_val(docs[2]);
let val4 = field.get_val(docs[3]);
let val1 = f64_from_fastfield_u64(val1, &self.field_type);
let val2 = f64_from_fastfield_u64(val2, &self.field_type);
let val3 = f64_from_fastfield_u64(val3, &self.field_type);
@@ -179,8 +179,8 @@ impl SegmentStatsCollector {
self.stats.collect(val3);
self.stats.collect(val4);
}
for doc in iter.remainder() {
let val = field.get(*doc);
for &doc in iter.remainder() {
let val = field.get_val(doc);
let val = f64_from_fastfield_u64(val, &self.field_type);
self.stats.collect(val);
}
@@ -222,7 +222,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let reader = index.reader()?;
let searcher = reader.searcher();
@@ -300,7 +300,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults = searcher.search(&term_query, &collector).unwrap();

130
src/aggregation/mod.rs
View File

@@ -10,20 +10,19 @@
//!
//! There are two categories: [Metrics](metric) and [Buckets](bucket).
//!
//! # Usage
//!
//! ## Prerequisite
//! Currently aggregations work only on [fast fields](`crate::fastfield`). Single value fast fields
//! of type `u64`, `f64`, `i64`, `date` and fast fields on text fields.
//!
//! ## Usage
//! To use aggregations, build an aggregation request by constructing
//! [Aggregations](agg_req::Aggregations).
//! Create an [AggregationCollector] from this request. AggregationCollector implements the
//! `Collector` trait and can be passed as collector into `searcher.search()`.
//! [`Aggregations`](agg_req::Aggregations).
//! Create an [`AggregationCollector`] from this request. `AggregationCollector` implements the
//! [`Collector`](crate::collector::Collector) trait and can be passed as collector into
//! [`Searcher::search()`](crate::Searcher::search).
//!
//! #### Limitations
//!
//! Currently aggregations work only on single value fast fields of type u64, f64, i64 and
//! fast fields on text fields.
//!
//! # JSON Format
//! ## JSON Format
//! Aggregations request and result structures de/serialize into elasticsearch compatible JSON.
//!
//! ```verbatim
@@ -34,7 +33,7 @@
//! let json_response_string: String = &serde_json::to_string(&agg_res)?;
//! ```
//!
//! # Supported Aggregations
//! ## Supported Aggregations
//! - [Bucket](bucket)
//! - [Histogram](bucket::HistogramAggregation)
//! - [Range](bucket::RangeAggregation)
@@ -44,8 +43,8 @@
//! - [Stats](metric::StatsAggregation)
//!
//! # Example
//! Compute the average metric, by building [agg_req::Aggregations], which is built from an (String,
//! [agg_req::Aggregation]) iterator.
//! Compute the average metric, by building [`agg_req::Aggregations`], which is built from an
//! `(String, agg_req::Aggregation)` iterator.
//!
//! ```
//! use tantivy::aggregation::agg_req::{Aggregations, Aggregation, MetricAggregation};
@@ -54,9 +53,10 @@
//! use tantivy::query::AllQuery;
//! use tantivy::aggregation::agg_result::AggregationResults;
//! use tantivy::IndexReader;
//! use tantivy::schema::Schema;
//!
//! # #[allow(dead_code)]
//! fn aggregate_on_index(reader: &IndexReader) {
//! fn aggregate_on_index(reader: &IndexReader, schema: Schema) {
//! let agg_req: Aggregations = vec![
//! (
//! "average".to_string(),
@@ -68,7 +68,7 @@
//! .into_iter()
//! .collect();
//!
//! let collector = AggregationCollector::from_aggs(agg_req, None);
//! let collector = AggregationCollector::from_aggs(agg_req, None, schema);
//!
//! let searcher = reader.searcher();
//! let agg_res: AggregationResults = searcher.search(&AllQuery, &collector).unwrap();
@@ -143,25 +143,25 @@
//! ```
//!
//! # Distributed Aggregation
//! When the data is distributed on different [crate::Index] instances, the
//! [DistributedAggregationCollector] provides functionality to merge data between independent
//! When the data is distributed on different [`Index`](crate::Index) instances, the
//! [`DistributedAggregationCollector`] provides functionality to merge data between independent
//! search calls by returning
//! [IntermediateAggregationResults](intermediate_agg_result::IntermediateAggregationResults).
//! IntermediateAggregationResults provides the
//! [merge_fruits](intermediate_agg_result::IntermediateAggregationResults::merge_fruits) method to
//! merge multiple results. The merged result can then be converted into
//! [agg_result::AggregationResults] via the
//! [agg_result::AggregationResults::from_intermediate_and_req] method.
//! [`IntermediateAggregationResults`](intermediate_agg_result::IntermediateAggregationResults).
//! `IntermediateAggregationResults` provides the
//! [`merge_fruits`](intermediate_agg_result::IntermediateAggregationResults::merge_fruits) method
//! to merge multiple results. The merged result can then be converted into
//! [`AggregationResults`](agg_result::AggregationResults) via the
//! [`into_final_bucket_result`](intermediate_agg_result::IntermediateAggregationResults::into_final_bucket_result) method.
pub mod agg_req;
mod agg_req_with_accessor;
pub mod agg_result;
pub mod bucket;
mod collector;
mod date;
pub mod intermediate_agg_result;
pub mod metric;
mod segment_agg_result;
use std::collections::HashMap;
use std::fmt::Display;
@@ -169,10 +169,11 @@ pub use collector::{
AggregationCollector, AggregationSegmentCollector, DistributedAggregationCollector,
MAX_BUCKET_COUNT,
};
pub(crate) use date::format_date;
use fastfield_codecs::MonotonicallyMappableToU64;
use itertools::Itertools;
use serde::{Deserialize, Serialize};
use crate::fastfield::FastValue;
use crate::schema::Type;
/// Represents an associative array `(key => values)` in a very efficient manner.
@@ -260,7 +261,7 @@ impl<T: Clone> VecWithNames<T> {
}
}
/// The serialized key is used in a HashMap.
/// The serialized key is used in a `HashMap`.
pub type SerializedKey = String;
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, PartialOrd)]
@@ -269,7 +270,7 @@ pub type SerializedKey = String;
pub enum Key {
/// String key
Str(String),
/// f64 key
/// `f64` key
F64(f64),
}
@@ -282,14 +283,14 @@ impl Display for Key {
}
}
/// Invert of to_fastfield_u64. Used to convert to f64 for metrics.
/// Inverse of `to_fastfield_u64`. Used to convert to `f64` for metrics.
///
/// # Panics
/// Only u64, f64, i64 is supported
/// Only `u64`, `f64`, `date`, and `i64` are supported.
pub(crate) fn f64_from_fastfield_u64(val: u64, field_type: &Type) -> f64 {
match field_type {
Type::U64 => val as f64,
Type::I64 => i64::from_u64(val) as f64,
Type::I64 | Type::Date => i64::from_u64(val) as f64,
Type::F64 => f64::from_u64(val),
_ => {
panic!("unexpected type {:?}. This should not happen", field_type)
@@ -297,20 +298,19 @@ pub(crate) fn f64_from_fastfield_u64(val: u64, field_type: &Type) -> f64 {
}
}
/// Converts the f64 value to fast field value space.
/// Converts the `f64` value to fast field value space, which is always u64.
///
/// If the fast field has u64, values are stored as u64 in the fast field.
/// A f64 value of e.g. 2.0 therefore needs to be converted to 1u64
/// If the fast field has `u64`, values are stored unchanged as `u64` in the fast field.
///
/// If the fast field has f64 values are converted and stored to u64 using a
/// If the fast field has `f64` values are converted and stored to `u64` using a
/// monotonic mapping.
/// A f64 value of e.g. 2.0 needs to be converted using the same monotonic
/// conversion function, so that the value matches the u64 value stored in the fast
/// A `f64` value of e.g. `2.0` needs to be converted using the same monotonic
/// conversion function, so that the value matches the `u64` value stored in the fast
/// field.
pub(crate) fn f64_to_fastfield_u64(val: f64, field_type: &Type) -> Option<u64> {
match field_type {
Type::U64 => Some(val as u64),
Type::I64 => Some((val as i64).to_u64()),
Type::I64 | Type::Date => Some((val as i64).to_u64()),
Type::F64 => Some(val.to_u64()),
_ => None,
}
@@ -319,6 +319,7 @@ pub(crate) fn f64_to_fastfield_u64(val: f64, field_type: &Type) -> Option<u64> {
#[cfg(test)]
mod tests {
use serde_json::Value;
use time::OffsetDateTime;
use super::agg_req::{Aggregation, Aggregations, BucketAggregation};
use super::bucket::RangeAggregation;
@@ -334,7 +335,7 @@ mod tests {
use crate::aggregation::DistributedAggregationCollector;
use crate::query::{AllQuery, TermQuery};
use crate::schema::{Cardinality, IndexRecordOption, Schema, TextFieldIndexing, FAST, STRING};
use crate::{Index, Term};
use crate::{DateTime, Index, Term};
fn get_avg_req(field_name: &str) -> Aggregation {
Aggregation::Metric(MetricAggregation::Average(
@@ -360,7 +361,7 @@ mod tests {
index: &Index,
query: Option<(&str, &str)>,
) -> crate::Result<Value> {
let collector = AggregationCollector::from_aggs(agg_req, None);
let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
let reader = index.reader()?;
let searcher = reader.searcher();
@@ -450,9 +451,9 @@ mod tests {
text_field_id => term.to_string(),
string_field_id => term.to_string(),
score_field => i as u64,
score_field_f64 => i as f64,
score_field_f64 => i,
score_field_i64 => i as i64,
fraction_field => i as f64/100.0,
fraction_field => i/100.0,
))?;
}
index_writer.commit()?;
@@ -554,10 +555,10 @@ mod tests {
let searcher = reader.searcher();
let intermediate_agg_result = searcher.search(&AllQuery, &collector).unwrap();
intermediate_agg_result
.into_final_bucket_result(agg_req)
.into_final_bucket_result(agg_req, &index.schema())
.unwrap()
} else {
let collector = AggregationCollector::from_aggs(agg_req, None);
let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
let searcher = reader.searcher();
searcher.search(&AllQuery, &collector).unwrap()
@@ -650,6 +651,7 @@ mod tests {
.set_fast()
.set_stored();
let text_field = schema_builder.add_text_field("text", text_fieldtype);
let date_field = schema_builder.add_date_field("date", FAST);
schema_builder.add_text_field("dummy_text", STRING);
let score_fieldtype =
crate::schema::NumericOptions::default().set_fast(Cardinality::SingleValue);
@@ -667,6 +669,7 @@ mod tests {
// writing the segment
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800).unwrap()),
score_field => 1u64,
score_field_f64 => 1f64,
score_field_i64 => 1i64,
@@ -675,6 +678,7 @@ mod tests {
))?;
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400).unwrap()),
score_field => 3u64,
score_field_f64 => 3f64,
score_field_i64 => 3i64,
@@ -683,18 +687,21 @@ mod tests {
))?;
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400).unwrap()),
score_field => 5u64,
score_field_f64 => 5f64,
score_field_i64 => 5i64,
))?;
index_writer.add_document(doc!(
text_field => "nohit",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400).unwrap()),
score_field => 6u64,
score_field_f64 => 6f64,
score_field_i64 => 6i64,
))?;
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400).unwrap()),
score_field => 7u64,
score_field_f64 => 7f64,
score_field_i64 => 7i64,
@@ -702,12 +709,14 @@ mod tests {
index_writer.commit()?;
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400).unwrap()),
score_field => 11u64,
score_field_f64 => 11f64,
score_field_i64 => 11i64,
))?;
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400 + 86400).unwrap()),
score_field => 14u64,
score_field_f64 => 14f64,
score_field_i64 => 14i64,
@@ -715,6 +724,7 @@ mod tests {
index_writer.add_document(doc!(
text_field => "cool",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400 + 86400).unwrap()),
score_field => 44u64,
score_field_f64 => 44.5f64,
score_field_i64 => 44i64,
@@ -725,6 +735,7 @@ mod tests {
// no hits segment
index_writer.add_document(doc!(
text_field => "nohit",
date_field => DateTime::from_utc(OffsetDateTime::from_unix_timestamp(1_546_300_800 + 86400 + 86400).unwrap()),
score_field => 44u64,
score_field_f64 => 44.5f64,
score_field_i64 => 44i64,
@@ -797,7 +808,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults = searcher.search(&term_query, &collector).unwrap();
@@ -997,9 +1008,10 @@ mod tests {
// Test de/serialization roundtrip on intermediate_agg_result
let res: IntermediateAggregationResults =
serde_json::from_str(&serde_json::to_string(&res).unwrap()).unwrap();
res.into_final_bucket_result(agg_req.clone()).unwrap()
res.into_final_bucket_result(agg_req.clone(), &index.schema())
.unwrap()
} else {
let collector = AggregationCollector::from_aggs(agg_req.clone(), None);
let collector = AggregationCollector::from_aggs(agg_req.clone(), None, index.schema());
let searcher = reader.searcher();
searcher.search(&term_query, &collector).unwrap()
@@ -1057,7 +1069,7 @@ mod tests {
);
// Test empty result set
let collector = AggregationCollector::from_aggs(agg_req, None);
let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
let searcher = reader.searcher();
searcher.search(&query_with_no_hits, &collector).unwrap();
@@ -1122,7 +1134,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
@@ -1235,7 +1247,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1266,7 +1278,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1297,7 +1309,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1336,7 +1348,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1365,7 +1377,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req, None);
let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1394,7 +1406,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req, None);
let collector = AggregationCollector::from_aggs(agg_req, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1431,7 +1443,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1466,7 +1478,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1505,7 +1517,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1535,7 +1547,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =
@@ -1592,7 +1604,7 @@ mod tests {
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let searcher = reader.searcher();
let agg_res: AggregationResults =

6
src/aggregation/segment_agg_result.rs
View File

@@ -185,10 +185,10 @@ impl SegmentMetricResultCollector {
pub(crate) fn collect_block(&mut self, doc: &[DocId], metric: &MetricAggregationWithAccessor) {
match self {
SegmentMetricResultCollector::Average(avg_collector) => {
avg_collector.collect_block(doc, &metric.accessor);
avg_collector.collect_block(doc, &*metric.accessor);
}
SegmentMetricResultCollector::Stats(stats_collector) => {
stats_collector.collect_block(doc, &metric.accessor);
stats_collector.collect_block(doc, &*metric.accessor);
}
}
}
@@ -305,7 +305,7 @@ impl BucketCount {
}
pub(crate) fn add_count(&self, count: u32) {
self.bucket_count
.fetch_add(count as u32, std::sync::atomic::Ordering::Relaxed);
.fetch_add(count, std::sync::atomic::Ordering::Relaxed);
}
pub(crate) fn get_count(&self) -> u32 {
self.bucket_count.load(std::sync::atomic::Ordering::Relaxed)

6
src/collector/custom_score_top_collector.rs
View File

@@ -24,7 +24,7 @@ where TScore: Clone + PartialOrd
/// A custom segment scorer makes it possible to define any kind of score
/// for a given document belonging to a specific segment.
///
/// It is the segment local version of the [`CustomScorer`](./trait.CustomScorer.html).
/// It is the segment local version of the [`CustomScorer`].
pub trait CustomSegmentScorer<TScore>: 'static {
/// Computes the score of a specific `doc`.
fn score(&mut self, doc: DocId) -> TScore;
@@ -36,9 +36,9 @@ pub trait CustomSegmentScorer<TScore>: 'static {
/// Instead, it helps constructing `Self::Child` instances that will compute
/// the score at a segment scale.
pub trait CustomScorer<TScore>: Sync {
/// Type of the associated [`CustomSegmentScorer`](./trait.CustomSegmentScorer.html).
/// Type of the associated [`CustomSegmentScorer`].
type Child: CustomSegmentScorer<TScore>;
/// Builds a child scorer for a specific segment. The child scorer is associated to
/// Builds a child scorer for a specific segment. The child scorer is associated with
/// a specific segment.
fn segment_scorer(&self, segment_reader: &SegmentReader) -> crate::Result<Self::Child>;
}

22
src/collector/facet_collector.rs
View File

@@ -67,10 +67,10 @@ fn facet_depth(facet_bytes: &[u8]) -> usize {
/// (e.g. `/category/fiction`, `/category/biography`, `/category/personal_development`).
///
/// Once collection is finished, you can harvest its results in the form
/// of a `FacetCounts` object, and extract your face t counts from it.
/// of a [`FacetCounts`] object, and extract your facet counts from it.
///
/// This implementation assumes you are working with a number of facets that
/// is much hundreds of time lower than your number of documents.
/// is many hundreds of times smaller than your number of documents.
///
///
/// ```rust
@@ -91,7 +91,7 @@ fn facet_depth(facet_bytes: &[u8]) -> usize {
/// let index = Index::create_in_ram(schema);
/// {
/// let mut index_writer = index.writer(3_000_000)?;
/// // a document can be associated to any number of facets
/// // a document can be associated with any number of facets
/// index_writer.add_document(doc!(
/// title => "The Name of the Wind",
/// facet => Facet::from("/lang/en"),
@@ -231,7 +231,7 @@ impl FacetCollector {
///
/// Adding two facets within which one is the prefix of the other is forbidden.
/// If you need the correct number of unique documents for two such facets,
/// just add them in separate `FacetCollector`.
/// just add them in a separate `FacetCollector`.
pub fn add_facet<T>(&mut self, facet_from: T)
where Facet: From<T> {
let facet = Facet::from(facet_from);
@@ -338,11 +338,7 @@ impl SegmentCollector for FacetSegmentCollector {
let mut previous_collapsed_ord: usize = usize::MAX;
for &facet_ord in &self.facet_ords_buf {
let collapsed_ord = self.collapse_mapping[facet_ord as usize];
self.counts[collapsed_ord] += if collapsed_ord == previous_collapsed_ord {
0
} else {
1
};
self.counts[collapsed_ord] += u64::from(collapsed_ord != previous_collapsed_ord);
previous_collapsed_ord = collapsed_ord;
}
}
@@ -361,7 +357,7 @@ impl SegmentCollector for FacetSegmentCollector {
let mut facet = vec![];
let facet_ord = self.collapse_facet_ords[collapsed_facet_ord];
// TODO handle errors.
if facet_dict.ord_to_term(facet_ord as u64, &mut facet).is_ok() {
if facet_dict.ord_to_term(facet_ord, &mut facet).is_ok() {
if let Ok(facet) = Facet::from_encoded(facet) {
facet_counts.insert(facet, count);
}
@@ -391,7 +387,7 @@ impl<'a> Iterator for FacetChildIterator<'a> {
impl FacetCounts {
/// Returns an iterator over all of the facet count pairs inside this result.
/// See the documentation for [FacetCollector] for a usage example.
/// See the documentation for [`FacetCollector`] for a usage example.
pub fn get<T>(&self, facet_from: T) -> FacetChildIterator<'_>
where Facet: From<T> {
let facet = Facet::from(facet_from);
@@ -410,7 +406,7 @@ impl FacetCounts {
}
/// Returns a vector of top `k` facets with their counts, sorted highest-to-lowest by counts.
/// See the documentation for [FacetCollector] for a usage example.
/// See the documentation for [`FacetCollector`] for a usage example.
pub fn top_k<T>(&self, facet: T, k: usize) -> Vec<(&Facet, u64)>
where Facet: From<T> {
let mut heap = BinaryHeap::with_capacity(k);
@@ -620,7 +616,7 @@ mod tests {
.map(|mut doc| {
doc.add_facet(
facet_field,
&format!("/facet/{}", thread_rng().sample(&uniform)),
&format!("/facet/{}", thread_rng().sample(uniform)),
);
doc
})

9
src/collector/filter_collector_wrapper.rs
View File

@@ -10,9 +10,12 @@
// ---
// Importing tantivy...
use std::marker::PhantomData;
use std::sync::Arc;
use fastfield_codecs::Column;
use crate::collector::{Collector, SegmentCollector};
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader, FastValue};
use crate::fastfield::FastValue;
use crate::schema::Field;
use crate::{Score, SegmentReader, TantivyError};
@@ -158,7 +161,7 @@ where
TPredicate: 'static,
TPredicateValue: FastValue,
{
fast_field_reader: DynamicFastFieldReader<TPredicateValue>,
fast_field_reader: Arc<dyn Column<TPredicateValue>>,
segment_collector: TSegmentCollector,
predicate: TPredicate,
t_predicate_value: PhantomData<TPredicateValue>,
@@ -174,7 +177,7 @@ where
type Fruit = TSegmentCollector::Fruit;
fn collect(&mut self, doc: u32, score: Score) {
let value = self.fast_field_reader.get(doc);
let value = self.fast_field_reader.get_val(doc);
if (self.predicate)(value) {
self.segment_collector.collect(doc, score)
}

11
src/collector/histogram_collector.rs
View File

@@ -1,7 +1,10 @@
use std::sync::Arc;
use fastdivide::DividerU64;
use fastfield_codecs::Column;
use crate::collector::{Collector, SegmentCollector};
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader, FastValue};
use crate::fastfield::FastValue;
use crate::schema::{Field, Type};
use crate::{DocId, Score};
@@ -34,7 +37,7 @@ impl HistogramCollector {
/// The scale/range of the histogram is not dynamic. It is required to
/// define it by supplying following parameter:
/// - `min_value`: the minimum value that can be recorded in the histogram.
/// - `bucket_width`: the length of the interval that is associated to each buckets.
/// - `bucket_width`: the length of the interval that is associated with each buckets.
/// - `num_buckets`: The overall number of buckets.
///
/// Together, this parameters define a partition of `[min_value, min_value + num_buckets *
@@ -84,14 +87,14 @@ impl HistogramComputer {
}
pub struct SegmentHistogramCollector {
histogram_computer: HistogramComputer,
ff_reader: DynamicFastFieldReader<u64>,
ff_reader: Arc<dyn Column<u64>>,
}
impl SegmentCollector for SegmentHistogramCollector {
type Fruit = Vec<u64>;
fn collect(&mut self, doc: DocId, _score: Score) {
let value = self.ff_reader.get(doc);
let value = self.ff_reader.get_val(doc);
self.histogram_computer.add_value(value);
}

56
src/collector/mod.rs
View File

@@ -4,13 +4,13 @@
//! In tantivy jargon, we call this information your search "fruit".
//!
//! Your fruit could for instance be :
//! - [the count of matching documents](./struct.Count.html)
//! - [the top 10 documents, by relevancy or by a fast field](./struct.TopDocs.html)
//! - [facet counts](./struct.FacetCollector.html)
//! - [the count of matching documents](crate::collector::Count)
//! - [the top 10 documents, by relevancy or by a fast field](crate::collector::TopDocs)
//! - [facet counts](FacetCollector)
//!
//! At one point in your code, you will trigger the actual search operation by calling
//! [the `search(...)` method of your `Searcher` object](../struct.Searcher.html#method.search).
//! This call will look like this.
//! At some point in your code, you will trigger the actual search operation by calling
//! [`Searcher::search()`](crate::Searcher::search).
//! This call will look like this:
//!
//! ```verbatim
//! let fruit = searcher.search(&query, &collector)?;
@@ -64,7 +64,7 @@
//!
//! The `Collector` trait is implemented for up to 4 collectors.
//! If you have more than 4 collectors, you can either group them into
//! tuples of tuples `(a,(b,(c,d)))`, or rely on [`MultiCollector`](./struct.MultiCollector.html).
//! tuples of tuples `(a,(b,(c,d)))`, or rely on [`MultiCollector`].
//!
//! # Combining several collectors dynamically
//!
@@ -74,7 +74,7 @@
//!
//! Unfortunately it requires you to know at compile time your collector types.
//! If on the other hand, the collectors depend on some query parameter,
//! you can rely on `MultiCollector`'s.
//! you can rely on [`MultiCollector`]'s.
//!
//!
//! # Implementing your own collectors.
@@ -142,7 +142,7 @@ pub trait Collector: Sync + Send {
/// e.g. `usize` for the `Count` collector.
type Fruit: Fruit;
/// Type of the `SegmentCollector` associated to this collector.
/// Type of the `SegmentCollector` associated with this collector.
type Child: SegmentCollector;
/// `set_segment` is called before beginning to enumerate
@@ -156,7 +156,7 @@ pub trait Collector: Sync + Send {
/// Returns true iff the collector requires to compute scores for documents.
fn requires_scoring(&self) -> bool;
/// Combines the fruit associated to the collection of each segments
/// Combines the fruit associated with the collection of each segments
/// into one fruit.
fn merge_fruits(
&self,
@@ -170,19 +170,35 @@ pub trait Collector: Sync + Send {
segment_ord: u32,
reader: &SegmentReader,
) -> crate::Result<<Self::Child as SegmentCollector>::Fruit> {
let mut segment_collector = self.for_segment(segment_ord as u32, reader)?;
let mut segment_collector = self.for_segment(segment_ord, reader)?;
if let Some(alive_bitset) = reader.alive_bitset() {
weight.for_each(reader, &mut |doc, score| {
if alive_bitset.is_alive(doc) {
match (reader.alive_bitset(), self.requires_scoring()) {
(Some(alive_bitset), true) => {
weight.for_each(reader, &mut |doc, score| {
if alive_bitset.is_alive(doc) {
segment_collector.collect(doc, score);
}
})?;
}
(Some(alive_bitset), false) => {
weight.for_each_no_score(reader, &mut |doc| {
if alive_bitset.is_alive(doc) {
segment_collector.collect(doc, 0.0);
}
})?;
}
(None, true) => {
weight.for_each(reader, &mut |doc, score| {
segment_collector.collect(doc, score);
}
})?;
} else {
weight.for_each(reader, &mut |doc, score| {
segment_collector.collect(doc, score);
})?;
})?;
}
(None, false) => {
weight.for_each_no_score(reader, &mut |doc| {
segment_collector.collect(doc, 0.0);
})?;
}
}
Ok(segment_collector.harvest())
}
}

10
src/collector/tests.rs
View File

@@ -1,7 +1,11 @@
use std::sync::Arc;
use fastfield_codecs::Column;
use super::*;
use crate::collector::{Count, FilterCollector, TopDocs};
use crate::core::SegmentReader;
use crate::fastfield::{BytesFastFieldReader, DynamicFastFieldReader, FastFieldReader};
use crate::fastfield::BytesFastFieldReader;
use crate::query::{AllQuery, QueryParser};
use crate::schema::{Field, Schema, FAST, TEXT};
use crate::time::format_description::well_known::Rfc3339;
@@ -156,7 +160,7 @@ pub struct FastFieldTestCollector {
pub struct FastFieldSegmentCollector {
vals: Vec<u64>,
reader: DynamicFastFieldReader<u64>,
reader: Arc<dyn Column<u64>>,
}
impl FastFieldTestCollector {
@@ -197,7 +201,7 @@ impl SegmentCollector for FastFieldSegmentCollector {
type Fruit = Vec<u64>;
fn collect(&mut self, doc: DocId, _score: Score) {
let val = self.reader.get(doc);
let val = self.reader.get_val(doc);
self.vals.push(val);
}

34
src/collector/top_score_collector.rs
View File

@@ -1,6 +1,9 @@
use std::collections::BinaryHeap;
use std::fmt;
use std::marker::PhantomData;
use std::sync::Arc;
use fastfield_codecs::Column;
use super::Collector;
use crate::collector::custom_score_top_collector::CustomScoreTopCollector;
@@ -9,7 +12,7 @@ use crate::collector::tweak_score_top_collector::TweakedScoreTopCollector;
use crate::collector::{
CustomScorer, CustomSegmentScorer, ScoreSegmentTweaker, ScoreTweaker, SegmentCollector,
};
use crate::fastfield::{DynamicFastFieldReader, FastFieldReader, FastValue};
use crate::fastfield::FastValue;
use crate::query::Weight;
use crate::schema::Field;
use crate::{DocAddress, DocId, Score, SegmentOrdinal, SegmentReader, TantivyError};
@@ -129,12 +132,12 @@ impl fmt::Debug for TopDocs {
}
struct ScorerByFastFieldReader {
ff_reader: DynamicFastFieldReader<u64>,
ff_reader: Arc<dyn Column<u64>>,
}
impl CustomSegmentScorer<u64> for ScorerByFastFieldReader {
fn score(&mut self, doc: DocId) -> u64 {
self.ff_reader.get(doc)
self.ff_reader.get_val(doc)
}
}
@@ -284,7 +287,7 @@ impl TopDocs {
/// # See also
///
/// To comfortably work with `u64`s, `i64`s, `f64`s, or `date`s, please refer to
/// [.order_by_fast_field(...)](#method.order_by_fast_field) method.
/// the [.order_by_fast_field(...)](TopDocs::order_by_fast_field) method.
pub fn order_by_u64_field(
self,
field: Field,
@@ -381,7 +384,7 @@ impl TopDocs {
///
/// This method offers a convenient way to tweak or replace
/// the documents score. As suggested by the prototype you can
/// manually define your own [`ScoreTweaker`](./trait.ScoreTweaker.html)
/// manually define your own [`ScoreTweaker`]
/// and pass it as an argument, but there is a much simpler way to
/// tweak your score: you can use a closure as in the following
/// example.
@@ -398,7 +401,7 @@ impl TopDocs {
/// In the following example will will tweak our ranking a bit by
/// boosting popular products a notch.
///
/// In more serious application, this tweaking could involved running a
/// In more serious application, this tweaking could involve running a
/// learning-to-rank model over various features
///
/// ```rust
@@ -407,7 +410,6 @@ impl TopDocs {
/// # use tantivy::query::QueryParser;
/// use tantivy::SegmentReader;
/// use tantivy::collector::TopDocs;
/// use tantivy::fastfield::FastFieldReader;
/// use tantivy::schema::Field;
///
/// fn create_schema() -> Schema {
@@ -456,7 +458,7 @@ impl TopDocs {
///
/// // We can now define our actual scoring function
/// move |doc: DocId, original_score: Score| {
/// let popularity: u64 = popularity_reader.get(doc);
/// let popularity: u64 = popularity_reader.get_val(doc);
/// // Well.. For the sake of the example we use a simple logarithm
/// // function.
/// let popularity_boost_score = ((2u64 + popularity) as Score).log2();
@@ -472,7 +474,7 @@ impl TopDocs {
/// ```
///
/// # See also
/// [custom_score(...)](#method.custom_score).
/// - [custom_score(...)](TopDocs::custom_score)
pub fn tweak_score<TScore, TScoreSegmentTweaker, TScoreTweaker>(
self,
score_tweaker: TScoreTweaker,
@@ -489,8 +491,7 @@ impl TopDocs {
///
/// This method offers a convenient way to use a different score.
///
/// As suggested by the prototype you can manually define your
/// own [`CustomScorer`](./trait.CustomScorer.html)
/// As suggested by the prototype you can manually define your own [`CustomScorer`]
/// and pass it as an argument, but there is a much simpler way to
/// tweak your score: you can use a closure as in the following
/// example.
@@ -499,7 +500,7 @@ impl TopDocs {
///
/// This method only makes it possible to compute the score from a given
/// `DocId`, fastfield values for the doc and any information you could
/// have precomputed beforehands. It does not make it possible for instance
/// have precomputed beforehand. It does not make it possible for instance
/// to compute something like TfIdf as it does not have access to the list of query
/// terms present in the document, nor the term frequencies for the different terms.
///
@@ -515,7 +516,6 @@ impl TopDocs {
/// use tantivy::SegmentReader;
/// use tantivy::collector::TopDocs;
/// use tantivy::schema::Field;
/// use tantivy::fastfield::FastFieldReader;
///
/// # fn create_schema() -> Schema {
/// # let mut schema_builder = Schema::builder();
@@ -567,8 +567,8 @@ impl TopDocs {
///
/// // We can now define our actual scoring function
/// move |doc: DocId| {
/// let popularity: u64 = popularity_reader.get(doc);
/// let boosted: u64 = boosted_reader.get(doc);
/// let popularity: u64 = popularity_reader.get_val(doc);
/// let boosted: u64 = boosted_reader.get_val(doc);
/// // Score do not have to be `f64` in tantivy.
/// // Here we return a couple to get lexicographical order
/// // for free.
@@ -587,7 +587,7 @@ impl TopDocs {
/// ```
///
/// # See also
/// [tweak_score(...)](#method.tweak_score).
/// - [tweak_score(...)](TopDocs::tweak_score)
pub fn custom_score<TScore, TCustomSegmentScorer, TCustomScorer>(
self,
custom_score: TCustomScorer,
@@ -693,7 +693,7 @@ impl Collector for TopDocs {
}
}
/// Segment Collector associated to `TopDocs`.
/// Segment Collector associated with `TopDocs`.
pub struct TopScoreSegmentCollector(TopSegmentCollector<Score>);
impl SegmentCollector for TopScoreSegmentCollector {

6
src/collector/tweak_score_top_collector.rs
View File

@@ -24,7 +24,7 @@ where TScore: Clone + PartialOrd
/// A `ScoreSegmentTweaker` makes it possible to modify the default score
/// for a given document belonging to a specific segment.
///
/// It is the segment local version of the [`ScoreTweaker`](./trait.ScoreTweaker.html).
/// It is the segment local version of the [`ScoreTweaker`].
pub trait ScoreSegmentTweaker<TScore>: 'static {
/// Tweak the given `score` for the document `doc`.
fn score(&mut self, doc: DocId, score: Score) -> TScore;
@@ -37,10 +37,10 @@ pub trait ScoreSegmentTweaker<TScore>: 'static {
/// Instead, it helps constructing `Self::Child` instances that will compute
/// the score at a segment scale.
pub trait ScoreTweaker<TScore>: Sync {
/// Type of the associated [`ScoreSegmentTweaker`](./trait.ScoreSegmentTweaker.html).
/// Type of the associated [`ScoreSegmentTweaker`].
type Child: ScoreSegmentTweaker<TScore>;
/// Builds a child tweaker for a specific segment. The child scorer is associated to
/// Builds a child tweaker for a specific segment. The child scorer is associated with
/// a specific segment.
fn segment_tweaker(&self, segment_reader: &SegmentReader) -> Result<Self::Child>;
}

179
src/core/index.rs
View File

@@ -7,6 +7,7 @@ use std::sync::Arc;
use super::segment::Segment;
use super::IndexSettings;
use crate::core::single_segment_index_writer::SingleSegmentIndexWriter;
use crate::core::{
Executor, IndexMeta, SegmentId, SegmentMeta, SegmentMetaInventory, META_FILEPATH,
};
@@ -16,9 +17,9 @@ use crate::directory::MmapDirectory;
use crate::directory::{Directory, ManagedDirectory, RamDirectory, INDEX_WRITER_LOCK};
use crate::error::{DataCorruption, TantivyError};
use crate::indexer::index_writer::{MAX_NUM_THREAD, MEMORY_ARENA_NUM_BYTES_MIN};
use crate::indexer::segment_updater::save_new_metas;
use crate::indexer::segment_updater::save_metas;
use crate::reader::{IndexReader, IndexReaderBuilder};
use crate::schema::{Field, FieldType, Schema};
use crate::schema::{Cardinality, Field, FieldType, Schema};
use crate::tokenizer::{TextAnalyzer, TokenizerManager};
use crate::IndexWriter;
@@ -47,10 +48,38 @@ fn load_metas(
.map_err(From::from)
}
/// Save the index meta file.
/// This operation is atomic :
/// Either
/// - it fails, in which case an error is returned,
/// and the `meta.json` remains untouched,
/// - it succeeds, and `meta.json` is written
/// and flushed.
///
/// This method is not part of tantivy's public API
fn save_new_metas(
schema: Schema,
index_settings: IndexSettings,
directory: &dyn Directory,
) -> crate::Result<()> {
save_metas(
&IndexMeta {
index_settings,
segments: Vec::new(),
schema,
opstamp: 0u64,
payload: None,
},
directory,
)?;
directory.sync_directory()?;
Ok(())
}
/// IndexBuilder can be used to create an index.
///
/// Use in conjunction with `SchemaBuilder`. Global index settings
/// can be configured with `IndexSettings`
/// Use in conjunction with [`SchemaBuilder`][crate::schema::SchemaBuilder].
/// Global index settings can be configured with [`IndexSettings`].
///
/// # Examples
///
@@ -68,7 +97,13 @@ fn load_metas(
/// );
///
/// let schema = schema_builder.build();
/// let settings = IndexSettings{sort_by_field: Some(IndexSortByField{field:"number".to_string(), order:Order::Asc}), ..Default::default()};
/// let settings = IndexSettings{
/// sort_by_field: Some(IndexSortByField{
/// field: "number".to_string(),
/// order: Order::Asc
/// }),
/// ..Default::default()
/// };
/// let index = Index::builder().schema(schema).settings(settings).create_in_ram();
/// ```
pub struct IndexBuilder {
@@ -111,21 +146,21 @@ impl IndexBuilder {
self
}
/// Creates a new index using the `RAMDirectory`.
/// Creates a new index using the [`RamDirectory`].
///
/// The index will be allocated in anonymous memory.
/// This should only be used for unit tests.
/// This is useful for indexing small set of documents
/// for instances like unit test or temporary in memory index.
pub fn create_in_ram(self) -> Result<Index, TantivyError> {
let ram_directory = RamDirectory::create();
Ok(self
.create(ram_directory)
.expect("Creating a RAMDirectory should never fail"))
self.create(ram_directory)
}
/// Creates a new index in a given filepath.
/// The index will use the `MMapDirectory`.
/// The index will use the [`MmapDirectory`].
///
/// If a previous index was in this directory, it returns an `IndexAlreadyExists` error.
/// If a previous index was in this directory, it returns an
/// [`TantivyError::IndexAlreadyExists`] error.
#[cfg(feature = "mmap")]
pub fn create_in_dir<P: AsRef<Path>>(self, directory_path: P) -> crate::Result<Index> {
let mmap_directory: Box<dyn Directory> = Box::new(MmapDirectory::open(directory_path)?);
@@ -135,14 +170,34 @@ impl IndexBuilder {
self.create(mmap_directory)
}
/// Dragons ahead!!!
///
/// The point of this API is to let users create a simple index with a single segment
/// and without starting any thread.
///
/// Do not use this method if you are not sure what you are doing.
///
/// It expects an originally empty directory, and will not run any GC operation.
#[doc(hidden)]
pub fn single_segment_index_writer(
self,
dir: impl Into<Box<dyn Directory>>,
mem_budget: usize,
) -> crate::Result<SingleSegmentIndexWriter> {
let index = self.create(dir)?;
let index_simple_writer = SingleSegmentIndexWriter::new(index, mem_budget)?;
Ok(index_simple_writer)
}
/// Creates a new index in a temp directory.
///
/// The index will use the `MMapDirectory` in a newly created directory.
/// The temp directory will be destroyed automatically when the `Index` object
/// The index will use the [`MmapDirectory`] in a newly created directory.
/// The temp directory will be destroyed automatically when the [`Index`] object
/// is destroyed.
///
/// The temp directory is only used for testing the `MmapDirectory`.
/// For other unit tests, prefer the `RAMDirectory`, see: `create_in_ram`.
/// The temp directory is only used for testing the [`MmapDirectory`].
/// For other unit tests, prefer the [`RamDirectory`], see:
/// [`IndexBuilder::create_in_ram()`].
#[cfg(feature = "mmap")]
pub fn create_from_tempdir(self) -> crate::Result<Index> {
let mmap_directory: Box<dyn Directory> = Box::new(MmapDirectory::create_from_tempdir()?);
@@ -172,10 +227,44 @@ impl IndexBuilder {
))
}
}
fn validate(&self) -> crate::Result<()> {
if let Some(schema) = self.schema.as_ref() {
if let Some(sort_by_field) = self.index_settings.sort_by_field.as_ref() {
let schema_field = schema.get_field(&sort_by_field.field).ok_or_else(|| {
TantivyError::InvalidArgument(format!(
"Field to sort index {} not found in schema",
sort_by_field.field
))
})?;
let entry = schema.get_field_entry(schema_field);
if !entry.is_fast() {
return Err(TantivyError::InvalidArgument(format!(
"Field {} is no fast field. Field needs to be a single value fast field \
to be used to sort an index",
sort_by_field.field
)));
}
if entry.field_type().fastfield_cardinality() != Some(Cardinality::SingleValue) {
return Err(TantivyError::InvalidArgument(format!(
"Only single value fast field Cardinality supported for sorting index {}",
sort_by_field.field
)));
}
}
Ok(())
} else {
Err(TantivyError::InvalidArgument(
"no schema passed".to_string(),
))
}
}
/// Creates a new index given an implementation of the trait `Directory`.
///
/// If a directory previously existed, it will be erased.
fn create<T: Into<Box<dyn Directory>>>(self, dir: T) -> crate::Result<Index> {
self.validate()?;
let dir = dir.into();
let directory = ManagedDirectory::wrap(dir)?;
save_new_metas(
@@ -238,7 +327,7 @@ impl Index {
self.set_multithread_executor(default_num_threads)
}
/// Creates a new index using the `RamDirectory`.
/// Creates a new index using the [`RamDirectory`].
///
/// The index will be allocated in anonymous memory.
/// This is useful for indexing small set of documents
@@ -248,9 +337,10 @@ impl Index {
}
/// Creates a new index in a given filepath.
/// The index will use the `MMapDirectory`.
/// The index will use the [`MmapDirectory`].
///
/// If a previous index was in this directory, then it returns an `IndexAlreadyExists` error.
/// If a previous index was in this directory, then it returns
/// a [`TantivyError::IndexAlreadyExists`] error.
#[cfg(feature = "mmap")]
pub fn create_in_dir<P: AsRef<Path>>(
directory_path: P,
@@ -272,12 +362,13 @@ impl Index {
/// Creates a new index in a temp directory.
///
/// The index will use the `MMapDirectory` in a newly created directory.
/// The temp directory will be destroyed automatically when the `Index` object
/// The index will use the [`MmapDirectory`] in a newly created directory.
/// The temp directory will be destroyed automatically when the [`Index`] object
/// is destroyed.
///
/// The temp directory is only used for testing the `MmapDirectory`.
/// For other unit tests, prefer the `RamDirectory`, see: `create_in_ram`.
/// The temp directory is only used for testing the [`MmapDirectory`].
/// For other unit tests, prefer the [`RamDirectory`],
/// see: [`IndexBuilder::create_in_ram()`].
#[cfg(feature = "mmap")]
pub fn create_from_tempdir(schema: Schema) -> crate::Result<Index> {
IndexBuilder::new().schema(schema).create_from_tempdir()
@@ -297,7 +388,7 @@ impl Index {
builder.create(dir)
}
/// Creates a new index given a directory and an `IndexMeta`.
/// Creates a new index given a directory and an [`IndexMeta`].
fn open_from_metas(
directory: ManagedDirectory,
metas: &IndexMeta,
@@ -324,7 +415,7 @@ impl Index {
&self.tokenizers
}
/// Helper to access the tokenizer associated to a specific field.
/// Get the tokenizer associated with a specific field.
pub fn tokenizer_for_field(&self, field: Field) -> crate::Result<TextAnalyzer> {
let field_entry = self.schema.get_field_entry(field);
let field_type = field_entry.field_type();
@@ -356,14 +447,14 @@ impl Index {
})
}
/// Create a default `IndexReader` for the given index.
/// Create a default [`IndexReader`] for the given index.
///
/// See [`Index.reader_builder()`](#method.reader_builder).
/// See [`Index.reader_builder()`].
pub fn reader(&self) -> crate::Result<IndexReader> {
self.reader_builder().try_into()
}
/// Create a `IndexReader` for the given index.
/// Create a [`IndexReader`] for the given index.
///
/// Most project should create at most one reader for a given index.
/// This method is typically called only once per `Index` instance.
@@ -580,10 +671,12 @@ impl fmt::Debug for Index {
#[cfg(test)]
mod tests {
use crate::collector::Count;
use crate::directory::{RamDirectory, WatchCallback};
use crate::schema::{Field, Schema, INDEXED, TEXT};
use crate::query::TermQuery;
use crate::schema::{Field, IndexRecordOption, Schema, INDEXED, TEXT};
use crate::tokenizer::TokenizerManager;
use crate::{Directory, Index, IndexBuilder, IndexReader, IndexSettings, ReloadPolicy};
use crate::{Directory, Index, IndexBuilder, IndexReader, IndexSettings, ReloadPolicy, Term};
#[test]
fn test_indexer_for_field() {
@@ -720,7 +813,7 @@ mod tests {
let field = schema.get_field("num_likes").unwrap();
let tempdir = TempDir::new().unwrap();
let tempdir_path = PathBuf::from(tempdir.path());
let index = Index::create_in_dir(&tempdir_path, schema).unwrap();
let index = Index::create_in_dir(tempdir_path, schema).unwrap();
let reader = index
.reader_builder()
.reload_policy(ReloadPolicy::OnCommit)
@@ -849,4 +942,28 @@ mod tests {
);
Ok(())
}
#[test]
fn test_single_segment_index_writer() -> crate::Result<()> {
let mut schema_builder = Schema::builder();
let text_field = schema_builder.add_text_field("text", TEXT);
let schema = schema_builder.build();
let directory = RamDirectory::default();
let mut single_segment_index_writer = Index::builder()
.schema(schema)
.single_segment_index_writer(directory, 10_000_000)?;
for _ in 0..10 {
let doc = doc!(text_field=>"hello");
single_segment_index_writer.add_document(doc)?;
}
let index = single_segment_index_writer.finalize()?;
let searcher = index.reader()?.searcher();
let term_query = TermQuery::new(
Term::from_field_text(text_field, "hello"),
IndexRecordOption::Basic,
);
let count = searcher.search(&term_query, &Count)?;
assert_eq!(count, 10);
Ok(())
}
}

71
src/core/index_meta.rs
View File

@@ -130,10 +130,10 @@ impl SegmentMeta {
/// Returns the relative path of a component of our segment.
///
/// It just joins the segment id with the extension
/// associated to a segment component.
/// associated with a segment component.
pub fn relative_path(&self, component: SegmentComponent) -> PathBuf {
let mut path = self.id().uuid_string();
path.push_str(&*match component {
path.push_str(&match component {
SegmentComponent::Postings => ".idx".to_string(),
SegmentComponent::Positions => ".pos".to_string(),
SegmentComponent::Terms => ".term".to_string(),
@@ -235,6 +235,14 @@ impl InnerSegmentMeta {
}
}
fn return_true() -> bool {
true
}
fn is_true(val: &bool) -> bool {
*val
}
/// Search Index Settings.
///
/// Contains settings which are applied on the whole
@@ -248,6 +256,12 @@ pub struct IndexSettings {
/// The `Compressor` used to compress the doc store.
#[serde(default)]
pub docstore_compression: Compressor,
/// If set to true, docstore compression will happen on a dedicated thread.
/// (defaults: true)
#[doc(hidden)]
#[serde(default = "return_true")]
#[serde(skip_serializing_if = "is_true")]
pub docstore_compress_dedicated_thread: bool,
#[serde(default = "default_docstore_blocksize")]
/// The size of each block that will be compressed and written to disk
pub docstore_blocksize: usize,
@@ -264,6 +278,7 @@ impl Default for IndexSettings {
sort_by_field: None,
docstore_compression: Compressor::default(),
docstore_blocksize: default_docstore_blocksize(),
docstore_compress_dedicated_thread: true,
}
}
}
@@ -311,13 +326,13 @@ pub struct IndexMeta {
/// `IndexSettings` to configure index options.
#[serde(default)]
pub index_settings: IndexSettings,
/// List of `SegmentMeta` informations associated to each finalized segment of the index.
/// List of `SegmentMeta` information associated with each finalized segment of the index.
pub segments: Vec<SegmentMeta>,
/// Index `Schema`
pub schema: Schema,
/// Opstamp associated to the last `commit` operation.
/// Opstamp associated with the last `commit` operation.
pub opstamp: Opstamp,
/// Payload associated to the last commit.
/// Payload associated with the last commit.
///
/// Upon commit, clients can optionally add a small `String` payload to their commit
/// to help identify this commit.
@@ -395,7 +410,7 @@ mod tests {
use super::IndexMeta;
use crate::core::index_meta::UntrackedIndexMeta;
use crate::schema::{Schema, TEXT};
use crate::store::ZstdCompressor;
use crate::store::{Compressor, ZstdCompressor};
use crate::{IndexSettings, IndexSortByField, Order};
#[test]
@@ -447,6 +462,7 @@ mod tests {
compression_level: Some(4),
}),
docstore_blocksize: 1_000_000,
docstore_compress_dedicated_thread: true,
},
segments: Vec::new(),
schema,
@@ -485,4 +501,47 @@ mod tests {
"unknown zstd option \"bla\" at line 1 column 103".to_string()
);
}
#[test]
#[cfg(feature = "lz4-compression")]
fn test_index_settings_default() {
let mut index_settings = IndexSettings::default();
assert_eq!(
index_settings,
IndexSettings {
sort_by_field: None,
docstore_compression: Compressor::default(),
docstore_compress_dedicated_thread: true,
docstore_blocksize: 16_384
}
);
{
let index_settings_json = serde_json::to_value(&index_settings).unwrap();
assert_eq!(
index_settings_json,
serde_json::json!({
"docstore_compression": "lz4",
"docstore_blocksize": 16384
})
);
let index_settings_deser: IndexSettings =
serde_json::from_value(index_settings_json).unwrap();
assert_eq!(index_settings_deser, index_settings);
}
{
index_settings.docstore_compress_dedicated_thread = false;
let index_settings_json = serde_json::to_value(&index_settings).unwrap();
assert_eq!(
index_settings_json,
serde_json::json!({
"docstore_compression": "lz4",
"docstore_blocksize": 16384,
"docstore_compress_dedicated_thread": false,
})
);
let index_settings_deser: IndexSettings =
serde_json::from_value(index_settings_json).unwrap();
assert_eq!(index_settings_deser, index_settings);
}
}
}

63
src/core/inverted_index_reader.rs
View File

@@ -9,18 +9,17 @@ use crate::schema::{IndexRecordOption, Term};
use crate::termdict::TermDictionary;
/// The inverted index reader is in charge of accessing
/// the inverted index associated to a specific field.
/// the inverted index associated with a specific field.
///
/// # Note
///
/// It is safe to delete the segment associated to
/// It is safe to delete the segment associated with
/// an `InvertedIndexReader`. As long as it is open,
/// the `FileSlice` it is relying on should
/// the [`FileSlice`] it is relying on should
/// stay available.
///
///
/// `InvertedIndexReader` are created by calling
/// the `SegmentReader`'s [`.inverted_index(...)`] method
/// [`SegmentReader::inverted_index()`](crate::SegmentReader::inverted_index).
pub struct InvertedIndexReader {
termdict: TermDictionary,
postings_file_slice: FileSlice,
@@ -30,7 +29,7 @@ pub struct InvertedIndexReader {
}
impl InvertedIndexReader {
#[cfg_attr(feature = "cargo-clippy", allow(clippy::needless_pass_by_value))] // for symmetry
#[allow(clippy::needless_pass_by_value)] // for symmetry
pub(crate) fn new(
termdict: TermDictionary,
postings_file_slice: FileSlice,
@@ -75,7 +74,7 @@ impl InvertedIndexReader {
///
/// This is useful for enumerating through a list of terms,
/// and consuming the associated posting lists while avoiding
/// reallocating a `BlockSegmentPostings`.
/// reallocating a [`BlockSegmentPostings`].
///
/// # Warning
///
@@ -96,7 +95,7 @@ impl InvertedIndexReader {
/// Returns a block postings given a `Term`.
/// This method is for an advanced usage only.
///
/// Most user should prefer using `read_postings` instead.
/// Most users should prefer using [`Self::read_postings()`] instead.
pub fn read_block_postings(
&self,
term: &Term,
@@ -110,7 +109,7 @@ impl InvertedIndexReader {
/// Returns a block postings given a `term_info`.
/// This method is for an advanced usage only.
///
/// Most user should prefer using `read_postings` instead.
/// Most users should prefer using [`Self::read_postings()`] instead.
pub fn read_block_postings_from_terminfo(
&self,
term_info: &TermInfo,
@@ -130,7 +129,7 @@ impl InvertedIndexReader {
/// Returns a posting object given a `term_info`.
/// This method is for an advanced usage only.
///
/// Most user should prefer using `read_postings` instead.
/// Most users should prefer using [`Self::read_postings()`] instead.
pub fn read_postings_from_terminfo(
&self,
term_info: &TermInfo,
@@ -164,12 +163,12 @@ impl InvertedIndexReader {
/// or `None` if the term has never been encountered and indexed.
///
/// If the field was not indexed with the indexing options that cover
/// the requested options, the returned `SegmentPostings` the method does not fail
/// the requested options, the returned [`SegmentPostings`] the method does not fail
/// and returns a `SegmentPostings` with as much information as possible.
///
/// For instance, requesting `IndexRecordOption::Freq` for a
/// `TextIndexingOptions` that does not index position will return a `SegmentPostings`
/// with `DocId`s and frequencies.
/// For instance, requesting [`IndexRecordOption::WithFreqs`] for a
/// [`TextOptions`](crate::schema::TextOptions) that does not index position
/// will return a [`SegmentPostings`] with `DocId`s and frequencies.
pub fn read_postings(
&self,
term: &Term,
@@ -201,23 +200,16 @@ impl InvertedIndexReader {
#[cfg(feature = "quickwit")]
impl InvertedIndexReader {
pub(crate) async fn get_term_info_async(
&self,
term: &Term,
) -> crate::AsyncIoResult<Option<TermInfo>> {
pub(crate) async fn get_term_info_async(&self, term: &Term) -> io::Result<Option<TermInfo>> {
self.termdict.get_async(term.value_bytes()).await
}
/// Returns a block postings given a `Term`.
/// This method is for an advanced usage only.
///
/// Most user should prefer using `read_postings` instead.
pub async fn warm_postings(
&self,
term: &Term,
with_positions: bool,
) -> crate::AsyncIoResult<()> {
let term_info_opt = self.get_term_info_async(term).await?;
/// Most users should prefer using [`Self::read_postings()`] instead.
pub async fn warm_postings(&self, term: &Term, with_positions: bool) -> io::Result<()> {
let term_info_opt: Option<TermInfo> = self.get_term_info_async(term).await?;
if let Some(term_info) = term_info_opt {
self.postings_file_slice
.read_bytes_slice_async(term_info.postings_range.clone())
@@ -230,4 +222,25 @@ impl InvertedIndexReader {
}
Ok(())
}
/// Read the block postings for all terms.
/// This method is for an advanced usage only.
///
/// If you know which terms to pre-load, prefer using [`Self::warm_postings`] instead.
pub async fn warm_postings_full(&self, with_positions: bool) -> io::Result<()> {
self.postings_file_slice.read_bytes_async().await?;
if with_positions {
self.positions_file_slice.read_bytes_async().await?;
}
Ok(())
}
/// Returns the number of documents containing the term asynchronously.
pub async fn doc_freq_async(&self, term: &Term) -> io::Result<u32> {
Ok(self
.get_term_info_async(term)
.await?
.map(|term_info| term_info.doc_freq)
.unwrap_or(0u32))
}
}

2
src/core/mod.rs
View File

@@ -7,6 +7,7 @@ mod segment;
mod segment_component;
mod segment_id;
mod segment_reader;
mod single_segment_index_writer;
use std::path::Path;
@@ -23,6 +24,7 @@ pub use self::segment::Segment;
pub use self::segment_component::SegmentComponent;
pub use self::segment_id::SegmentId;
pub use self::segment_reader::SegmentReader;
pub use self::single_segment_index_writer::SingleSegmentIndexWriter;
/// The meta file contains all the information about the list of segments and the schema
/// of the index.

52
src/core/searcher.rs
View File

@@ -4,18 +4,18 @@ use std::{fmt, io};
use crate::collector::Collector;
use crate::core::{Executor, SegmentReader};
use crate::query::Query;
use crate::query::{EnableScoring, Query};
use crate::schema::{Document, Schema, Term};
use crate::space_usage::SearcherSpaceUsage;
use crate::store::{CacheStats, StoreReader};
use crate::{DocAddress, Index, Opstamp, SegmentId, TrackedObject};
/// Identifies the searcher generation accessed by a [Searcher].
/// Identifies the searcher generation accessed by a [`Searcher`].
///
/// While this might seem redundant, a [SearcherGeneration] contains
/// While this might seem redundant, a [`SearcherGeneration`] contains
/// both a `generation_id` AND a list of `(SegmentId, DeleteOpstamp)`.
///
/// This is on purpose. This object is used by the `Warmer` API.
/// This is on purpose. This object is used by the [`Warmer`](crate::reader::Warmer) API.
/// Having both information makes it possible to identify which
/// artifact should be refreshed or garbage collected.
///
@@ -69,20 +69,20 @@ pub struct Searcher {
}
impl Searcher {
/// Returns the `Index` associated to the `Searcher`
/// Returns the `Index` associated with the `Searcher`
pub fn index(&self) -> &Index {
&self.inner.index
}
/// [SearcherGeneration] which identifies the version of the snapshot held by this `Searcher`.
/// [`SearcherGeneration`] which identifies the version of the snapshot held by this `Searcher`.
pub fn generation(&self) -> &SearcherGeneration {
self.inner.generation.as_ref()
}
/// Fetches a document from tantivy's store given a `DocAddress`.
/// Fetches a document from tantivy's store given a [`DocAddress`].
///
/// The searcher uses the segment ordinal to route the
/// the request to the right `Segment`.
/// request to the right `Segment`.
pub fn doc(&self, doc_address: DocAddress) -> crate::Result<Document> {
let store_reader = &self.inner.store_readers[doc_address.segment_ord as usize];
store_reader.get(doc_address.doc_id)
@@ -108,7 +108,7 @@ impl Searcher {
store_reader.get_async(doc_address.doc_id).await
}
/// Access the schema associated to the index of this searcher.
/// Access the schema associated with the index of this searcher.
pub fn schema(&self) -> &Schema {
&self.inner.schema
}
@@ -134,6 +134,19 @@ impl Searcher {
Ok(total_doc_freq)
}
/// Return the overall number of documents containing
/// the given term in an asynchronous manner.
#[cfg(feature = "quickwit")]
pub async fn doc_freq_async(&self, term: &Term) -> crate::Result<u64> {
let mut total_doc_freq = 0;
for segment_reader in &self.inner.segment_readers {
let inverted_index = segment_reader.inverted_index(term.field())?;
let doc_freq = inverted_index.doc_freq_async(term).await?;
total_doc_freq += u64::from(doc_freq);
}
Ok(total_doc_freq)
}
/// Return the list of segment readers
pub fn segment_readers(&self) -> &[SegmentReader] {
&self.inner.segment_readers
@@ -148,11 +161,11 @@ impl Searcher {
///
/// Search works as follows :
///
/// First the weight object associated to the query is created.
/// First the weight object associated with the query is created.
///
/// Then, the query loops over the segments and for each segment :
/// - setup the collector and informs it that the segment being processed has changed.
/// - creates a SegmentCollector for collecting documents associated to the segment
/// - creates a SegmentCollector for collecting documents associated with the segment
/// - creates a `Scorer` object associated for this segment
/// - iterate through the matched documents and push them to the segment collector.
///
@@ -167,7 +180,7 @@ impl Searcher {
self.search_with_executor(query, collector, executor)
}
/// Same as [`search(...)`](#method.search) but multithreaded.
/// Same as [`search(...)`](Searcher::search) but multithreaded.
///
/// The current implementation is rather naive :
/// multithreading is by splitting search into as many task
@@ -186,7 +199,12 @@ impl Searcher {
executor: &Executor,
) -> crate::Result<C::Fruit> {
let scoring_enabled = collector.requires_scoring();
let weight = query.weight(self, scoring_enabled)?;
let enabled_scoring = if scoring_enabled {
EnableScoring::Enabled(self)
} else {
EnableScoring::Disabled(self.schema())
};
let weight = query.weight(enabled_scoring)?;
let segment_readers = self.segment_readers();
let fruits = executor.map(
|(segment_ord, segment_reader)| {
@@ -234,6 +252,14 @@ impl SearcherInner {
generation: TrackedObject<SearcherGeneration>,
doc_store_cache_size: usize,
) -> io::Result<SearcherInner> {
assert_eq!(
&segment_readers
.iter()
.map(|reader| (reader.segment_id(), reader.delete_opstamp()))
.collect::<BTreeMap<_, _>>(),
generation.segments(),
"Set of segments referenced by this Searcher and its SearcherGeneration must match"
);
let store_readers: Vec<StoreReader> = segment_readers
.iter()
.map(|segment_reader| segment_reader.get_store_reader(doc_store_cache_size))

2
src/core/segment.rs
View File

@@ -70,7 +70,7 @@ impl Segment {
/// Returns the relative path of a component of our segment.
///
/// It just joins the segment id with the extension
/// associated to a segment component.
/// associated with a segment component.
pub fn relative_path(&self, component: SegmentComponent) -> PathBuf {
self.meta.relative_path(component)
}

Some files were not shown because too many files have changed in this diff Show More