rust/tantivy
1
0
Fork 0
mirror of https://github.com/quickwit-oss/tantivy.git synced 2025-12-27 20:42:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

Compare commits

base: rust:fix_estimate
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-main
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

242 Commits
fix_estima ... columnar-m

Author SHA1 Message Date
Paul Masurel
937485321a Rebase and fixing unit test 2023-01-18 10:15:16 +09:00
Pascal Seitz
e3be14e7d8 start migrate Field to &str 
start migrate Field to &str in preparation of columnar
return Result for get_field
 2023-01-18 10:02:19 +09:00
Adrien Guillo
c9cb3d04bf Merge pull request #1788 from quickwit-oss/guilload/remove-std-dev-from-stats-agg 
Remove standard deviation from stats aggregation
 2023-01-16 23:16:36 -05:00
Adrien Guillo
0caaf13a90 Remove standard deviation from stats aggregation 2023-01-16 22:58:23 -05:00
Adrien Guillo
a59bd965cc Merge pull request #1794 from quickwit-oss/guilload/count-min-max-sum-aggs 
Add count, min, max, and sum aggregations
 2023-01-16 22:45:01 -05:00
Adrien Guillo
f2dad194ea Add count, min, max, and sum aggregations 2023-01-16 12:22:20 -05:00
Paul Masurel
25bad784ad Integrated fastfield codecs into columnar. (#1782) 
Introduced asymetric OptionalCodec / SerializableOptionalCodec
Removed cardinality from the columnar sstable.
Added DynamicColumn
Reorganized all files
Change DenseCodec serialization logic.
Renamed methods to rank/select
Moved versioning footer to the columnar level
 2023-01-16 17:24:49 +09:00
PSeitz
4bac945709 add ip field example (#1775) 2023-01-16 06:06:11 +01:00
trinity-1686a
16b704e190 make file_slice_for_range on sstable public (#1784) 2023-01-16 13:59:57 +09:00
PSeitz
6ca9a477f3 reuse stats for average (#1785) 
* reuse stats for average

* fix count type
 2023-01-13 23:32:27 +08:00
Shikhar Bhushan
2650111b76 EnableScoring::Disabled - optional Searcher (#1780) 2023-01-12 09:26:50 -05:00
PSeitz
1176555eff handle user input on get_docid_for_value_range (#1760) 
* handle user input on get_docid_for_value_range

fixes #1757

* pass range as parameter
 2023-01-12 14:20:16 +01:00
Adrien Guillo
f8d111a75e Merge pull request #1777 from quickwit-oss/guilload/ff-range-query-on-not-indexed-fields 
Allow range queries via fast fields on non-indexed fields
 2023-01-11 10:14:32 -05:00
Adrien Guillo
e17996f2fd Allow range queries via fast fields on non-indexed fields 2023-01-11 09:56:13 -05:00
Adrien Guillo
f3621c0487 Add license to tokenizer-api crate (#1778) 2023-01-11 05:26:41 +01:00
Adrien Guillo
14222a47a3 Fix typo (#1776) 2023-01-11 00:49:13 +09:00
Adam Reichold
8312c882a5 More cosmetic fixes for upcoming Clippy lints. (#1771) 2023-01-10 10:32:45 +01:00
Paul Masurel
7a8fce0ae7 Minor mini fixes 2023-01-10 14:15:30 +09:00
Michael Kleen
196e42f33e Add regex tokenizer (#1759) 
This adds a regex tokenizer which tokenizes the text by using a
regex pattern to split.

Co-authored-by: Michael Kleen <mkleen@gmailw.com>
 2023-01-10 13:38:37 +09:00
Adam Reichold
82a183bc2d Bump dependency on lru to from version 0.7.5 to version 0.9.0. (#1755) 2023-01-10 13:35:37 +09:00
dependabot[bot]
3090d49615 Update base64 requirement from 0.20.0 to 0.21.0 (#1769) 
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/v0.20.0...v0.21.0)

---
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>
 2023-01-10 13:35:05 +09:00
PSeitz
7c6cc818ae enable range query on fast field for u64 compatible types (#1762) 
* enable range query on fast field for u64 compatible types

* rename, update benches
 2023-01-10 04:08:26 +01:00
PSeitz
514d23a20c move tokenizer API to seperate crate (#1767) 
closes #1766

Finding tantivy tokenizers is a frustrating experience currently, since
they need be updated for each tantivy version. That's unnecessary since
the API is rather stable anyway.
 2023-01-09 06:37:38 +01:00
Paul Masurel
4f9efe654c Support for columnar (#1734) 
* Added support for dynamic fast field.

See README for more information.

* Apply suggestions from code review

Co-authored-by: PSeitz <PSeitz@users.noreply.github.com>
 2023-01-07 17:37:00 +09:00
Adam Reichold
1afa5bf3db Make construction of LevenshteinAutomatonBuilder for FuzzyTermQuery instances lazy. (#1756) 2023-01-06 12:44:49 +09:00
PSeitz
07a51eb7c8 refactor multivalue fastfield, refactor range query (#1749) 
Introduce MakeZero trait, remove make_zero from FastValue
Merge two multivalue fastfield implementations into one
prepare range query on fastfield for different types
 2023-01-05 12:09:50 +01:00
Adam Reichold
2080c370c2 Enable usage of FuzzyTermQuery for specific fields via QueryParser (#1750) 
* Make nightly Clippy mostly happy.

* Document how to produce TermSetQuery queries using QueryParser.

* Enable construction of queries using FuzzyTermQuery via the QueryParser

* Use FxHashMap instead of HashMap in the QueryParser as these hash tables are not exposed to DoS attacks.

* Use a struct instead of a tuple to improve readability.
 2023-01-04 18:11:27 +09:00
Daw-Chih Liou
b22f96624e doc: update comments in the faceted search example (#1737) 
* doc: update comments in the faceted search example

* chore: format
 2023-01-02 11:07:30 +01:00
pinkforest(she/her)
b78dc5e313 Bump prettytables (#1746) 2022-12-31 15:01:39 +01:00
Paul Masurel
3f915925af Fixing unit tests 2022-12-27 12:02:16 +09:00
Paul Masurel
9c5fef5af7 Fixing sstable proptest (#1743) 2022-12-26 16:29:33 +09:00
Paul Masurel
9948a84ebe Simplifies the count_ones definition. (#1742) 2022-12-26 16:08:01 +09:00
PSeitz
45156fd869 use group_by in translate_codec_idx_to_original_id (#1736) 2022-12-26 06:13:29 +01:00
Paul Masurel
bc959006fa Ooops. Removing ordered_floats. 2022-12-22 19:50:34 +09:00
Paul Masurel
7385a8f80c Supporting PartialCmp in VectorColumn. (#1735) 
* Supporting PartialCmp in VectorColumn.
* Apply suggestions from code review

Co-authored-by: PSeitz <PSeitz@users.noreply.github.com>
 2022-12-22 17:47:25 +09:00
Paul Masurel
13b89cba17 Adding inlines. 2022-12-22 14:29:41 +09:00
Hasnain Lakhani
f4804ce2f5 Adjust spelling of "returns" in docs for DisjunctionMaxQuery (#1733) 2022-12-22 14:04:07 +09:00
Paul Masurel
2a6d1eaf78 Added missing license. 2022-12-22 12:47:43 +09:00
Paul Masurel
540a9972bd Support for NotNaN in fast fields 2022-12-22 12:28:25 +09:00
Paul Masurel
bb48c3e488 Refactoring to prepare for the addition of dynamic fast field (#1730) 
* Refactoring to prepare for the addition of dynamic fast field

- Exposing insert_key / insert_value
- Renamed SSTable::{Reader/Writer}-> SSTable::{ValueReader/ValueWriter}
- Added a generic Dictionary object in the sstable crate
- Removing the TermDictionary wrapper from tantivy, relying directly on
  an alias of the generic Dictionary object.
- dropped the use of byteorder in sstable.
- Stopped scanning / reading the entire dictionary when streaming a range.

* Added a benchmark for streaming sstable ranges.

* CR comments.

Rename deserialize_u64 -> deserialize_vint_u64

* Removed needless allocation, split serialize into serialize and clear.
 2022-12-22 12:25:46 +09:00
Paul Masurel
3339a3ec05 Removed feature(quickwit) in tantivy-common. 2022-12-22 10:19:57 +09:00
Paul Masurel
f39165e1e7 Moving FileSlice to tantivy-common (#1729) 2022-12-21 16:35:11 +09:00
Paul Masurel
32cb1d22da Removed AsyncIoResult. (#1728) 2022-12-21 16:01:17 +09:00
Paul Masurel
4a6bf50e78 Clippy 2022-12-21 15:43:34 +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
293 changed files with 22697 additions and 3373 deletions
Expand all files Collapse all files

1
.gitattributes vendored
View File

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

2
.github/workflows/test.yml vendored
View File

@@ -48,7 +48,7 @@ jobs:
strategy:
matrix:
features: [
{ label: "all", flags: "mmap,brotli-compression,lz4-compression,snappy-compression,zstd-compression,failpoints" },
{ label: "all", flags: "mmap,stopwords,brotli-compression,lz4-compression,snappy-compression,zstd-compression,failpoints" },
{ label: "quickwit", flags: "mmap,quickwit,failpoints" }
]

1
.gitignore vendored
View File

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

39
CHANGELOG.md
View File

@@ -1,10 +1,37 @@
Tantivy 0.19
================================
#### 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)
- 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.
#### 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
================================
@@ -22,6 +49,10 @@ 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
================================

44
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"]
@@ -11,19 +11,21 @@ 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 = "0.21.0"
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"
@@ -34,17 +36,11 @@ 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"
@@ -52,14 +48,21 @@ murmurhash32 = "0.2.0"
time = { version = "0.3.10", features = ["serde-well-known"] }
smallvec = "1.8.0"
rayon = "1.5.2"
lru = "0.7.5"
lru = "0.9.0"
fastdivide = "0.4.0"
itertools = "0.10.3"
measure_time = "0.8.2"
serde_cbor = { version = "0.11.2", optional = true }
async-trait = "0.1.53"
arc-swap = "1.5.0"
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 }
tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
[target.'cfg(windows)'.dependencies]
winapi = "0.3.9"
@@ -69,10 +72,10 @@ 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]
@@ -89,8 +92,9 @@ 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"]
@@ -100,10 +104,10 @@ 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", "tokenizer-api"]
# Following the "fail" crate best practises, we isolate
# tests that define specific behavior in fail check points

18
README.md
View File

@@ -29,7 +29,7 @@ Your mileage WILL vary depending on the nature of queries and their load.
# Features
- Full-text search
- Configurable tokenizer (stemming available for 17 Latin languages with third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)), Japanese ([lindera](https://github.com/lindera-morphology/lindera-tantivy), [Vaporetto](https://crates.io/crates/vaporetto_tantivy), and [tantivy-tokenizer-tiny-segmenter](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)) and Korean ([lindera](https://github.com/lindera-morphology/lindera-tantivy) + [lindera-ko-dic-builder](https://github.com/lindera-morphology/lindera-ko-dic-builder))
- Configurable tokenizer (stemming available for 17 Latin languages) with third party support for Chinese ([tantivy-jieba](https://crates.io/crates/tantivy-jieba) and [cang-jie](https://crates.io/crates/cang-jie)), Japanese ([lindera](https://github.com/lindera-morphology/lindera-tantivy), [Vaporetto](https://crates.io/crates/vaporetto_tantivy), and [tantivy-tokenizer-tiny-segmenter](https://crates.io/crates/tantivy-tokenizer-tiny-segmenter)) and Korean ([lindera](https://github.com/lindera-morphology/lindera-tantivy) + [lindera-ko-dic-builder](https://github.com/lindera-morphology/lindera-ko-dic-builder))
- Fast (check out the :racehorse: :sparkles: [benchmark](https://tantivy-search.github.io/bench/) :sparkles: :racehorse:)
- Tiny startup time (<10ms), perfect for command-line tools
- BM25 scoring (the same as Lucene)
@@ -42,12 +42,12 @@ Your mileage WILL vary depending on the nature of queries and their load.
- Single valued and multivalued u64, i64, and f64 fast fields (equivalent of doc values in Lucene)
- `&[u8]` fast fields
- Text, i64, u64, f64, dates, and hierarchical facet fields
- LZ4 compressed document store
- Compressed document store (LZ4, Zstd, None, Brotli, Snap)
- Range queries
- Faceted search
- Configurable indexing (optional term frequency and position indexing)
- JSON Field
- Aggregation Collector: range buckets, average, and stats metrics
- Aggregation Collector: histogram, range buckets, average, and stats metrics
- LogMergePolicy with deletes
- Searcher Warmer API
- Cheesy logo with a horse
@@ -58,7 +58,7 @@ Distributed search is out of the scope of Tantivy, but if you are looking for th
# 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,
@@ -81,9 +81,17 @@ 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.
## Tokenizer
When implementing a tokenizer for tantivy depend on the `tantivy-tokenizer-api` crate.
## 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

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

23
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())?;
@@ -87,22 +86,20 @@ impl BitUnpacker {
}
#[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 addr_in_bits = idx * self.num_bits;
let addr = addr_in_bits >> 3;
let addr_in_bits = idx * self.num_bits as u32;
let addr = (addr_in_bits >> 3) as usize;
let bit_shift = addr_in_bits & 7;
debug_assert!(
addr + 8 <= data.len() as u64,
addr + 8 <= data.len(),
"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 bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
let val_shifted = (val_unshifted_unmasked >> bit_shift) as u64;
let val_shifted = val_unshifted_unmasked >> bit_shift;
val_shifted & self.mask
}
}
@@ -130,7 +127,7 @@ 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);
}
}

8
bitpacker/src/blocked_bitpacker.rs
View File

@@ -84,7 +84,7 @@ impl BlockedBitpacker {
#[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();
}
}
@@ -126,11 +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()

122
bitpacker/src/lib.rs
View File

@@ -1,6 +1,8 @@
mod bitpacker;
mod blocked_bitpacker;
use std::cmp::Ordering;
pub use crate::bitpacker::{BitPacker, BitUnpacker};
pub use crate::blocked_bitpacker::BlockedBitpacker;
@@ -37,44 +39,104 @@ pub fn compute_num_bits(n: u64) -> u8 {
}
}
/// Computes the (min, max) of an iterator of `PartialOrd` values.
///
/// For values implementing `Ord` (in a way consistent to their `PartialOrd` impl),
/// this function behaves as expected.
///
/// For values with partial ordering, the behavior is non-trivial and may
/// depends on the order of the values.
/// For floats however, it simply returns the same results as if NaN were
/// skipped.
pub fn minmax<I, T>(mut vals: I) -> Option<(T, T)>
where
I: Iterator<Item = T>,
T: Copy + Ord,
T: Copy + PartialOrd,
{
if let Some(first_el) = vals.next() {
return Some(vals.fold((first_el, first_el), |(min_val, max_val), el| {
(min_val.min(el), max_val.max(el))
}));
let first_el = vals.find(|val| {
// We use this to make sure we skip all NaN values when
// working with a float type.
val.partial_cmp(val) == Some(Ordering::Equal)
})?;
let mut min_so_far: T = first_el;
let mut max_so_far: T = first_el;
for val in vals {
if val.partial_cmp(&min_so_far) == Some(Ordering::Less) {
min_so_far = val;
}
if val.partial_cmp(&max_so_far) == Some(Ordering::Greater) {
max_so_far = val;
}
}
None
Some((min_so_far, max_so_far))
}
#[test]
fn test_compute_num_bits() {
assert_eq!(compute_num_bits(1), 1u8);
assert_eq!(compute_num_bits(0), 0u8);
assert_eq!(compute_num_bits(2), 2u8);
assert_eq!(compute_num_bits(3), 2u8);
assert_eq!(compute_num_bits(4), 3u8);
assert_eq!(compute_num_bits(255), 8u8);
assert_eq!(compute_num_bits(256), 9u8);
assert_eq!(compute_num_bits(5_000_000_000), 33u8);
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_minmax_empty() {
let vals: Vec<u32> = vec![];
assert_eq!(minmax(vals.into_iter()), None);
}
#[test]
fn test_compute_num_bits() {
assert_eq!(compute_num_bits(1), 1u8);
assert_eq!(compute_num_bits(0), 0u8);
assert_eq!(compute_num_bits(2), 2u8);
assert_eq!(compute_num_bits(3), 2u8);
assert_eq!(compute_num_bits(4), 3u8);
assert_eq!(compute_num_bits(255), 8u8);
assert_eq!(compute_num_bits(256), 9u8);
assert_eq!(compute_num_bits(5_000_000_000), 33u8);
}
#[test]
fn test_minmax_one() {
assert_eq!(minmax(vec![1].into_iter()), Some((1, 1)));
}
#[test]
fn test_minmax_empty() {
let vals: Vec<u32> = vec![];
assert_eq!(minmax(vals.into_iter()), None);
}
#[test]
fn test_minmax_two() {
assert_eq!(minmax(vec![1, 2].into_iter()), Some((1, 2)));
assert_eq!(minmax(vec![2, 1].into_iter()), Some((1, 2)));
#[test]
fn test_minmax_one() {
assert_eq!(minmax(vec![1].into_iter()), Some((1, 1)));
}
#[test]
fn test_minmax_two() {
assert_eq!(minmax(vec![1, 2].into_iter()), Some((1, 2)));
assert_eq!(minmax(vec![2, 1].into_iter()), Some((1, 2)));
}
#[test]
fn test_minmax_nan() {
assert_eq!(
minmax(vec![f64::NAN, 1f64, 2f64].into_iter()),
Some((1f64, 2f64))
);
assert_eq!(
minmax(vec![2f64, f64::NAN, 1f64].into_iter()),
Some((1f64, 2f64))
);
assert_eq!(
minmax(vec![2f64, 1f64, f64::NAN].into_iter()),
Some((1f64, 2f64))
);
}
#[test]
fn test_minmax_inf() {
assert_eq!(
minmax(vec![f64::INFINITY, 1f64, 2f64].into_iter()),
Some((1f64, f64::INFINITY))
);
assert_eq!(
minmax(vec![-f64::INFINITY, 1f64, 2f64].into_iter()),
Some((-f64::INFINITY, 2f64))
);
assert_eq!(
minmax(vec![2f64, f64::INFINITY, 1f64].into_iter()),
Some((1f64, f64::INFINITY))
);
assert_eq!(
minmax(vec![2f64, 1f64, -f64::INFINITY].into_iter()),
Some((-f64::INFINITY, 2f64))
);
}
}

32
columnar/Cargo.toml Normal file
View File

@@ -0,0 +1,32 @@
[package]
name = "tantivy-columnar"
version = "0.1.0"
edition = "2021"
license = "MIT"
[dependencies]
stacker = { path = "../stacker", package="tantivy-stacker"}
serde_json = "1"
thiserror = "1"
fnv = "1"
sstable = { path = "../sstable", package = "tantivy-sstable" }
common = { path = "../common", package = "tantivy-common" }
itertools = "0.10"
log = "0.4"
tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
prettytable-rs = {version="0.10.0", optional= true}
rand = {version="0.8.3", optional= true}
fastdivide = "0.4"
measure_time = { version="0.8.2", optional=true}
[dev-dependencies]
proptest = "1"
more-asserts = "0.3.0"
rand = "0.8.3"
# temporary
[workspace]
members = []
[features]
unstable = []

109
columnar/README.md Normal file
View File

@@ -0,0 +1,109 @@
# Columnar format
This crate describes columnar format used in tantivy.
## Goals
This format is special in the following way.
- it needs to be compact
- accessing a specific column does not require to load the entire columnar. It can be done in 2 to 3 random access.
- 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.
- it is designed to allow range queries.
# Coercion rules
Users can create a columnar by inserting rows to a `ColumnarWriter`,
and serializing it into a `Write` object.
Nothing prevents a user from recording values with different type to the same `column_name`.
In that case, `tantivy-columnar`'s behavior is as follows:
- JsonValues are grouped into 3 types (String, Number, bool).
Values that corresponds to different groups are mapped to different columns. For instance, String values are treated independently
from Number or boolean values. `tantivy-columnar` will simply emit several columns associated to a given column_name.
- Only one column for a given json value type is emitted. If number values with different number types are recorded (e.g. u64, i64, f64),
`tantivy-columnar` will pick the first type that can represents the set of appended value, with the following prioriy order (`i64`, `u64`, `f64`).
`i64` is picked over `u64` as it is likely to yield less change of types. Most use cases strictly requiring `u64` show the
restriction on 50% of the values (e.g. a 64-bit hash). On the other hand, a lot of use cases can show rare negative value.
# Columnar format
This columnar format may have more than one column (with different types) associated to the same `column_name` (see [Coercion rules](#coercion-rules) above).
The `(column_name, columne_type)` couple however uniquely identifies a column.
That couple is serialized as a column `column_key`. The format of that key is:
`[column_name][ZERO_BYTE][column_type_header: u8]`
```
COLUMNAR:=
[COLUMNAR_DATA]
[COLUMNAR_KEY_TO_DATA_INDEX]
[COLUMNAR_FOOTER];
# Columns are sorted by their column key.
COLUMNAR_DATA:=
[COLUMN_DATA]+;
COLUMNAR_FOOTER := [RANGE_SSTABLE_BYTES_LEN: 8 bytes little endian]
```
The columnar file starts by the actual column data, concatenated one after the other,
sorted by column key.
A sstable associates
`(column name, column_cardinality, column_type) to range of bytes.
Column name may not contain the zero byte `\0`.
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
This crate exposes a columnar format for tantivy.
This format is described in README.md
The crate introduces the following concepts.
`Columnar` is an equivalent of a dataframe.
It maps `column_key` to `Column`.
A `Column<T>` asssociates a `RowId` (u32) to any
number of values.
This is made possible by wrapping a `ColumnIndex` and a `ColumnValue` object.
The `ColumnValue<T>` represents a mapping that associates each `RowId` to
exactly one single value.
The `ColumnIndex` then maps each RowId to a set of `RowId` in the
`ColumnValue`.
For optimization, and compression purposes, the `ColumnIndex` has three
possible representation, each for different cardinalities.
- Full
All RowId have exactly one value. The ColumnIndex is the trivial mapping.
- Optional
All RowIds can have at most one value. The ColumnIndex is the trivial mapping `ColumnRowId -> Option<ColumnValueRowId>`.
- Multivalued
All RowIds can have any number of values.
The column index is mapping values to a range.
All these objects are implemented an unit tested independently
in their own module:
- columnar
- column_index
- column_values
- column

45
columnar/src/TODO.md Normal file
View File

@@ -0,0 +1,45 @@
# zero to one
* merges
* full still needs a num_values
* replug u128
* add dictionary encoded stuff
* fix multivalued
* find a way to make columnar work with strict types
* plug to tantivy
- indexing
- aggregations
- merge
# Perf and Size
* re-add ZSTD compression for dictionaries
no systematic monotonic mapping
consider removing multilinear
f32?
adhoc solution for bool?
add metrics helper for aggregate. sum(row_id)
review inline absence/presence
improv perf of select using PDEP
compare with roaring bitmap/elias fano etc etc.
SIMD range? (see blog post)
Add alignment?
Consider another codec to bridge the gap between few and 5k elements
# Cleanup and rationalization
in benchmark, unify percent vs ratio, f32 vs f64.
investigate if should have better errors? io::Error is overused at the moment.
rename rank/select in unit tests
Review the public API via cargo doc
go through TODOs
remove all doc_id occurences -> row_id
use the rank & select naming in unit tests branch.
multi-linear -> blockwise
linear codec -> simply a multiplication for the index column
# Other
fix enhance column-cli
# Santa claus
autodetect datetime ipaddr, plug customizable tokenizer.

40
columnar/src/column/dictionary_encoded.rs Normal file
View File

@@ -0,0 +1,40 @@
use std::io;
use std::ops::Deref;
use std::sync::Arc;
use sstable::{Dictionary, VoidSSTable};
use crate::column::Column;
use crate::column_index::ColumnIndex;
/// Dictionary encoded column.
#[derive(Clone)]
pub struct BytesColumn {
pub(crate) dictionary: Arc<Dictionary<VoidSSTable>>,
pub(crate) term_ord_column: Column<u64>,
}
impl BytesColumn {
/// Returns `false` if the term does not exist (e.g. `term_ord` is greater or equal to the
/// overll number of terms).
pub fn term_ord_to_str(&self, term_ord: u64, output: &mut Vec<u8>) -> io::Result<bool> {
self.dictionary.ord_to_term(term_ord, output)
}
pub fn term_ords(&self) -> &Column<u64> {
&self.term_ord_column
}
}
impl Deref for BytesColumn {
type Target = ColumnIndex<'static>;
fn deref(&self) -> &Self::Target {
&**self.term_ords()
}
}
#[cfg(test)]
mod tests {
use crate::{ColumnarReader, ColumnarWriter};
}

56
columnar/src/column/mod.rs Normal file
View File

@@ -0,0 +1,56 @@
mod dictionary_encoded;
mod serialize;
use std::ops::Deref;
use std::sync::Arc;
use common::BinarySerializable;
pub use dictionary_encoded::BytesColumn;
pub use serialize::{open_column_bytes, open_column_u64, serialize_column_u64};
use crate::column_index::ColumnIndex;
use crate::column_values::ColumnValues;
use crate::{Cardinality, RowId};
#[derive(Clone)]
pub struct Column<T> {
pub idx: ColumnIndex<'static>,
pub values: Arc<dyn ColumnValues<T>>,
}
use crate::column_index::Set;
impl<T: PartialOrd> Column<T> {
pub fn first(&self, row_id: RowId) -> Option<T> {
match &self.idx {
ColumnIndex::Full => Some(self.values.get_val(row_id)),
ColumnIndex::Optional(opt_idx) => {
let value_row_idx = opt_idx.rank_if_exists(row_id)?;
Some(self.values.get_val(value_row_idx))
}
ColumnIndex::Multivalued(_multivalued_index) => {
todo!();
}
}
}
}
impl<T> Deref for Column<T> {
type Target = ColumnIndex<'static>;
fn deref(&self) -> &Self::Target {
&self.idx
}
}
impl BinarySerializable for Cardinality {
fn serialize<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
self.to_code().serialize(writer)
}
fn deserialize<R: std::io::Read>(reader: &mut R) -> std::io::Result<Self> {
let cardinality_code = u8::deserialize(reader)?;
let cardinality = Cardinality::try_from_code(cardinality_code)?;
Ok(cardinality)
}
}

54
columnar/src/column/serialize.rs Normal file
View File

@@ -0,0 +1,54 @@
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::{CountingWriter, OwnedBytes};
use sstable::Dictionary;
use crate::column::{BytesColumn, Column};
use crate::column_index::{serialize_column_index, SerializableColumnIndex};
use crate::column_values::{
serialize_column_values, ColumnValues, MonotonicallyMappableToU64, ALL_CODEC_TYPES,
};
pub fn serialize_column_u64<T: MonotonicallyMappableToU64>(
column_index: SerializableColumnIndex<'_>,
column_values: &impl ColumnValues<T>,
output: &mut impl Write,
) -> io::Result<()> {
let mut counting_writer = CountingWriter::wrap(output);
serialize_column_index(column_index, &mut counting_writer)?;
let column_index_num_bytes = counting_writer.written_bytes() as u32;
let output = counting_writer.finish();
serialize_column_values(column_values, &ALL_CODEC_TYPES[..], output)?;
output.write_all(&column_index_num_bytes.to_le_bytes())?;
Ok(())
}
pub fn open_column_u64<T: MonotonicallyMappableToU64>(bytes: OwnedBytes) -> io::Result<Column<T>> {
let (body, column_index_num_bytes_payload) = bytes.rsplit(4);
let column_index_num_bytes = u32::from_le_bytes(
column_index_num_bytes_payload
.as_slice()
.try_into()
.unwrap(),
);
let (column_index_data, column_values_data) = body.split(column_index_num_bytes as usize);
let column_index = crate::column_index::open_column_index(column_index_data)?;
let column_values = crate::column_values::open_u64_mapped(column_values_data)?;
Ok(Column {
idx: column_index,
values: column_values,
})
}
pub fn open_column_bytes(data: OwnedBytes) -> io::Result<BytesColumn> {
let (body, dictionary_len_bytes) = data.rsplit(4);
let dictionary_len = u32::from_le_bytes(dictionary_len_bytes.as_slice().try_into().unwrap());
let (dictionary_bytes, column_bytes) = body.split(dictionary_len as usize);
let dictionary = Arc::new(Dictionary::from_bytes(dictionary_bytes)?);
let term_ord_column = crate::column::open_column_u64::<u64>(column_bytes)?;
Ok(BytesColumn {
dictionary,
term_ord_column,
})
}

40
columnar/src/column_index/mod.rs Normal file
View File

@@ -0,0 +1,40 @@
mod multivalued_index;
mod optional_index;
mod serialize;
use std::sync::Arc;
pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
use crate::column_values::ColumnValues;
use crate::{Cardinality, RowId};
#[derive(Clone)]
pub enum ColumnIndex<'a> {
Full,
Optional(OptionalIndex),
// TODO remove the Arc<dyn> apart from serialization this is not
// dynamic at all.
Multivalued(Arc<dyn ColumnValues<RowId> + 'a>),
}
impl<'a> ColumnIndex<'a> {
pub fn get_cardinality(&self) -> Cardinality {
match self {
ColumnIndex::Full => Cardinality::Full,
ColumnIndex::Optional(_) => Cardinality::Optional,
ColumnIndex::Multivalued(_) => Cardinality::Multivalued,
}
}
pub fn num_rows(&self) -> RowId {
match self {
ColumnIndex::Full => {
todo!()
}
ColumnIndex::Optional(optional_index) => optional_index.num_rows(),
ColumnIndex::Multivalued(multivalued_index) => multivalued_index.num_vals() - 1,
}
}
}

27
columnar/src/column_index/multivalued_index.rs Normal file
View File

@@ -0,0 +1,27 @@
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::OwnedBytes;
use crate::column_values::{ColumnValues, FastFieldCodecType};
use crate::RowId;
#[derive(Clone)]
pub struct MultivaluedIndex(Arc<dyn ColumnValues<RowId>>);
pub fn serialize_multivalued_index(
multivalued_index: MultivaluedIndex,
output: &mut impl Write,
) -> io::Result<()> {
crate::column_values::serialize_column_values(
&*multivalued_index.0,
&[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
output,
)?;
Ok(())
}
pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<Arc<dyn ColumnValues<RowId>>> {
todo!();
}

453
columnar/src/column_index/optional_index/mod.rs Normal file
View File

@@ -0,0 +1,453 @@
use std::io::{self, Write};
use std::ops::Range;
use std::sync::Arc;
mod set;
mod set_block;
use common::{BinarySerializable, GroupByIteratorExtended, OwnedBytes, VInt};
pub use set::{Set, SetCodec};
use set_block::{
DenseBlock, DenseBlockCodec, SparseBlock, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
};
use crate::{InvalidData, RowId};
/// The threshold for for number of elements after which we switch to dense block encoding.
///
/// We simply pick the value that minimize the size of the blocks.
const DENSE_BLOCK_THRESHOLD: u32 =
set_block::DENSE_BLOCK_NUM_BYTES / std::mem::size_of::<u16>() as u32; //< 5_120
const ELEMENTS_PER_BLOCK: u32 = u16::MAX as u32 + 1;
const BLOCK_SIZE: RowId = 1 << 16;
#[derive(Copy, Clone, Debug)]
struct BlockMeta {
non_null_rows_before_block: u32,
start_byte_offset: u32,
block_variant: BlockVariant,
}
#[derive(Clone, Copy, Debug)]
enum BlockVariant {
Dense,
Sparse { num_vals: u16 },
}
impl BlockVariant {
pub fn empty() -> Self {
Self::Sparse { num_vals: 0 }
}
pub fn num_bytes_in_block(&self) -> u32 {
match *self {
BlockVariant::Dense => set_block::DENSE_BLOCK_NUM_BYTES,
BlockVariant::Sparse { num_vals } => num_vals as u32 * 2,
}
}
}
/// This codec is inspired by roaring bitmaps.
/// In the dense blocks, however, in order to accelerate `select`
/// we interleave an offset over two bytes. (more on this lower)
///
/// 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.
#[derive(Clone)]
pub struct OptionalIndex {
num_rows: RowId,
num_non_null_rows: RowId,
block_data: OwnedBytes,
block_metas: Arc<[BlockMeta]>,
}
impl OptionalIndex {
pub fn num_rows(&self) -> RowId {
self.num_rows
}
pub fn num_non_nulls(&self) -> RowId {
self.num_non_null_rows
}
}
/// 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(Copy, Debug, Clone)]
struct RowAddr {
block_id: u16,
in_block_row_id: u16,
}
#[inline(always)]
fn row_addr_from_row_id(row_id: RowId) -> RowAddr {
RowAddr {
block_id: (row_id / BLOCK_SIZE) as u16,
in_block_row_id: (row_id % BLOCK_SIZE) as u16,
}
}
impl Set<RowId> for OptionalIndex {
// Check if value at position is not null.
#[inline]
fn contains(&self, row_id: RowId) -> bool {
let RowAddr {
block_id,
in_block_row_id,
} = row_addr_from_row_id(row_id);
let block_meta = self.block_metas[block_id as usize];
match self.block(block_meta) {
Block::Dense(dense_block) => dense_block.contains(in_block_row_id),
Block::Sparse(sparse_block) => sparse_block.contains(in_block_row_id),
}
}
#[inline]
fn rank_if_exists(&self, row_id: RowId) -> Option<RowId> {
let RowAddr {
block_id,
in_block_row_id,
} = row_addr_from_row_id(row_id);
let block_meta = self.block_metas[block_id as usize];
let block = self.block(block_meta);
let block_offset_row_id = match block {
Block::Dense(dense_block) => dense_block.rank_if_exists(in_block_row_id),
Block::Sparse(sparse_block) => sparse_block.rank_if_exists(in_block_row_id),
}? as u32;
Some(block_meta.non_null_rows_before_block + block_offset_row_id)
}
#[inline]
fn select(&self, rank: RowId) -> RowId {
let block_pos = self.find_block(rank, 0);
let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
let block_meta = self.block_metas[block_pos as usize];
let block: Block<'_> = self.block(block_meta);
let index_in_block = (rank - block_meta.non_null_rows_before_block) as u16;
let in_block_rank = match block {
Block::Dense(dense_block) => dense_block.select(index_in_block),
Block::Sparse(sparse_block) => sparse_block.select(index_in_block),
};
block_doc_idx_start + in_block_rank as u32
}
fn select_batch(&self, ranks: &[u32], output_idxs: &mut [u32]) {
let mut block_pos = 0u32;
let mut start = 0;
let group_by_it = ranks.iter().copied().group_by(move |codec_idx| {
block_pos = self.find_block(*codec_idx, block_pos);
block_pos
});
for (block_pos, block_iter) in group_by_it {
let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
let block_meta = self.block_metas[block_pos as usize];
let block: Block<'_> = self.block(block_meta);
let offset = block_meta.non_null_rows_before_block;
let indexes_in_block_iter =
block_iter.map(move |codec_idx| (codec_idx - offset) as u16);
match block {
Block::Dense(dense_block) => {
for in_offset in dense_block.select_iter(indexes_in_block_iter) {
output_idxs[start] = in_offset as u32 + block_doc_idx_start;
start += 1;
}
}
Block::Sparse(sparse_block) => {
for in_offset in sparse_block.select_iter(indexes_in_block_iter) {
output_idxs[start] = in_offset as u32 + block_doc_idx_start;
start += 1;
}
}
};
}
}
}
impl OptionalIndex {
#[inline]
fn block<'a>(&'a self, block_meta: BlockMeta) -> Block<'a> {
let BlockMeta {
start_byte_offset,
block_variant,
..
} = block_meta;
let start_byte_offset = start_byte_offset as usize;
let bytes = self.block_data.as_slice();
match block_variant {
BlockVariant::Dense => Block::Dense(DenseBlockCodec::open(
&bytes[start_byte_offset..start_byte_offset + DENSE_BLOCK_NUM_BYTES as usize],
)),
BlockVariant::Sparse { num_vals } => {
let end_byte_offset = start_byte_offset + num_vals as usize * 2;
let sparse_bytes = &bytes[start_byte_offset..end_byte_offset];
Block::Sparse(SparseBlockCodec::open(sparse_bytes))
}
}
}
#[inline]
fn find_block(&self, dense_idx: u32, start_block_pos: u32) -> u32 {
for block_pos in start_block_pos..self.block_metas.len() as u32 {
let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
if offset > dense_idx {
return block_pos - 1;
}
}
self.block_metas.len() as u32 - 1u32
}
// TODO Add a good API for the codec_idx to original_idx translation.
// The Iterator API is a probably a bad idea
}
#[derive(Copy, Clone)]
enum Block<'a> {
Dense(DenseBlock<'a>),
Sparse(SparseBlock<'a>),
}
#[derive(Debug, Copy, Clone)]
enum OptionalIndexCodec {
Dense = 0,
Sparse = 1,
}
impl OptionalIndexCodec {
fn to_code(self) -> u8 {
self as u8
}
fn try_from_code(code: u8) -> Result<Self, InvalidData> {
match code {
0 => Ok(Self::Dense),
1 => Ok(Self::Sparse),
_ => Err(InvalidData),
}
}
}
impl BinarySerializable for OptionalIndexCodec {
fn serialize<W: Write>(&self, writer: &mut W) -> io::Result<()> {
writer.write_all(&[self.to_code()])
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let optional_codec_code = u8::deserialize(reader)?;
let optional_codec = Self::try_from_code(optional_codec_code)?;
Ok(optional_codec)
}
}
fn serialize_optional_index_block(block_els: &[u16], out: &mut impl io::Write) -> io::Result<()> {
let is_sparse = is_sparse(block_els.len() as u32);
if is_sparse {
SparseBlockCodec::serialize(block_els.iter().copied(), out)?;
} else {
DenseBlockCodec::serialize(block_els.iter().copied(), out)?;
}
Ok(())
}
pub fn serialize_optional_index<'a, W: io::Write>(
serializable_optional_index: &dyn SerializableOptionalIndex<'a>,
output: &mut W,
) -> io::Result<()> {
VInt(serializable_optional_index.num_rows() as u64).serialize(output)?;
let mut rows_it = serializable_optional_index.non_null_rows();
let mut block_metadata: Vec<SerializedBlockMeta> = 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.
let Some(idx) = rows_it.next() else {
output.write_all(&0u16.to_le_bytes())?;
return Ok(());
};
let row_addr = row_addr_from_row_id(idx);
let mut current_block_id = row_addr.block_id;
current_block.push(row_addr.in_block_row_id);
for idx in rows_it {
let value_addr = row_addr_from_row_id(idx);
if current_block_id != value_addr.block_id {
serialize_optional_index_block(&current_block[..], output)?;
block_metadata.push(SerializedBlockMeta {
block_id: current_block_id,
num_non_null_rows: current_block.len() as u32,
});
current_block.clear();
current_block_id = value_addr.block_id;
}
current_block.push(value_addr.in_block_row_id);
}
// handle last block
serialize_optional_index_block(&current_block[..], output)?;
block_metadata.push(SerializedBlockMeta {
block_id: current_block_id,
num_non_null_rows: current_block.len() as u32,
});
for block in &block_metadata {
output.write_all(&block.to_bytes())?;
}
output.write_all((block_metadata.len() as u16).to_le_bytes().as_ref())?;
Ok(())
}
const SERIALIZED_BLOCK_META_NUM_BYTES: usize = 4;
#[derive(Clone, Copy, Debug)]
struct SerializedBlockMeta {
block_id: u16,
num_non_null_rows: u32, //< takes values in 1..=u16::MAX
}
// TODO unit tests
impl SerializedBlockMeta {
#[inline]
fn from_bytes(bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES]) -> SerializedBlockMeta {
let block_id = u16::from_le_bytes(bytes[0..2].try_into().unwrap());
let num_non_null_rows: u32 =
u16::from_le_bytes(bytes[2..4].try_into().unwrap()) as u32 + 1u32;
SerializedBlockMeta {
block_id,
num_non_null_rows,
}
}
#[inline]
fn to_bytes(&self) -> [u8; SERIALIZED_BLOCK_META_NUM_BYTES] {
assert!(self.num_non_null_rows > 0);
let mut bytes = [0u8; SERIALIZED_BLOCK_META_NUM_BYTES];
bytes[0..2].copy_from_slice(&self.block_id.to_le_bytes());
// 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
bytes[2..4].copy_from_slice(&((self.num_non_null_rows - 1u32) as u16).to_le_bytes());
bytes
}
}
#[inline]
fn is_sparse(num_rows_in_block: u32) -> bool {
num_rows_in_block < DENSE_BLOCK_THRESHOLD as u32
}
fn deserialize_optional_index_block_metadatas(
data: &[u8],
num_rows: u32,
) -> (Box<[BlockMeta]>, u32) {
let num_blocks = data.len() / SERIALIZED_BLOCK_META_NUM_BYTES;
let mut block_metas = Vec::with_capacity(num_blocks as usize + 1);
let mut start_byte_offset = 0;
let mut non_null_rows_before_block = 0;
for block_meta_bytes in data.chunks_exact(SERIALIZED_BLOCK_META_NUM_BYTES) {
let block_meta_bytes: [u8; SERIALIZED_BLOCK_META_NUM_BYTES] =
block_meta_bytes.try_into().unwrap();
let SerializedBlockMeta {
block_id,
num_non_null_rows,
} = SerializedBlockMeta::from_bytes(block_meta_bytes);
block_metas.resize(
block_id as usize,
BlockMeta {
non_null_rows_before_block,
start_byte_offset,
block_variant: BlockVariant::empty(),
},
);
let block_variant = if is_sparse(num_non_null_rows) {
BlockVariant::Sparse {
num_vals: num_non_null_rows as u16,
}
} else {
BlockVariant::Dense
};
block_metas.push(BlockMeta {
non_null_rows_before_block,
start_byte_offset,
block_variant,
});
start_byte_offset += block_variant.num_bytes_in_block();
non_null_rows_before_block += num_non_null_rows as u32;
}
block_metas.resize(
((num_rows + BLOCK_SIZE - 1) / BLOCK_SIZE) as usize,
BlockMeta {
non_null_rows_before_block,
start_byte_offset,
block_variant: BlockVariant::empty(),
},
);
(block_metas.into_boxed_slice(), non_null_rows_before_block)
}
pub fn open_optional_index(bytes: OwnedBytes) -> io::Result<OptionalIndex> {
let (mut bytes, num_non_empty_blocks_bytes) = bytes.rsplit(2);
let num_non_empty_block_bytes =
u16::from_le_bytes(num_non_empty_blocks_bytes.as_slice().try_into().unwrap());
let num_rows = VInt::deserialize_u64(&mut bytes)? as u32;
let block_metas_num_bytes =
num_non_empty_block_bytes as usize * SERIALIZED_BLOCK_META_NUM_BYTES;
let (block_data, block_metas) = bytes.rsplit(block_metas_num_bytes);
let (block_metas, num_non_null_rows) =
deserialize_optional_index_block_metadatas(block_metas.as_slice(), num_rows).into();
let optional_index = OptionalIndex {
num_rows,
num_non_null_rows,
block_data,
block_metas: block_metas.into(),
};
Ok(optional_index)
}
pub trait SerializableOptionalIndex<'a> {
fn num_rows(&self) -> RowId;
fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a>;
}
impl SerializableOptionalIndex<'static> for Range<u32> {
fn num_rows(&self) -> RowId {
self.end
}
fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'static> {
Box::new(self.clone())
}
}
#[cfg(test)]
mod tests;

38
columnar/src/column_index/optional_index/set.rs Normal file
View File

@@ -0,0 +1,38 @@
use std::io;
/// A codec makes it possible to serialize a set of
/// elements, and open the resulting Set representation.
pub trait SetCodec {
type Item: Copy + TryFrom<usize> + Eq + std::hash::Hash + std::fmt::Debug;
type Reader<'a>: Set<Self::Item>;
/// Serializes a set of unique sorted u16 elements.
///
/// May panic if the elements are not sorted.
fn serialize(els: impl Iterator<Item = Self::Item>, wrt: impl io::Write) -> io::Result<()>;
fn open<'a>(data: &'a [u8]) -> Self::Reader<'a>;
}
pub trait Set<T> {
/// Returns true if the elements is contained in the Set
fn contains(&self, el: T) -> bool;
/// If the set contains `el` returns its position in the sortd set of elements.
/// If the set does not contain the element, it returns `None`.
fn rank_if_exists(&self, el: T) -> Option<T>;
/// Return the rank-th value stored in this bitmap.
///
/// # Panics
///
/// May panic if rank is greater than the number of elements in the Set.
fn select(&self, rank: T) -> T;
/// Batch version of select.
/// `ranks` is assumed to be sorted.
///
/// # Panics
///
/// May panic if rank is greater than the number of elements in the Set.
fn select_batch(&self, ranks: &[T], outputs: &mut [T]);
}

8
columnar/src/column_index/optional_index/set_block/mod.rs Normal file
View File

@@ -0,0 +1,8 @@
mod set_block;
mod sparse;
pub use set_block::{DenseBlock, DenseBlockCodec, DENSE_BLOCK_NUM_BYTES};
pub use sparse::{SparseBlock, SparseBlockCodec};
#[cfg(test)]
mod tests;

271
columnar/src/column_index/optional_index/set_block/set_block.rs Normal file
View File

@@ -0,0 +1,271 @@
use std::convert::TryInto;
use std::io::{self, Write};
use common::BinarySerializable;
use crate::column_index::optional_index::{Set, SetCodec, ELEMENTS_PER_BLOCK};
#[inline(always)]
fn get_bit_at(input: u64, n: u16) -> bool {
input & (1 << n) != 0
}
#[inline]
fn set_bit_at(input: &mut u64, n: u16) {
*input |= 1 << n;
}
/// 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.
const ELEMENTS_PER_MINI_BLOCK: u16 = 64;
const MINI_BLOCK_BITVEC_NUM_BYTES: usize = 8;
const MINI_BLOCK_OFFSET_NUM_BYTES: usize = 2;
pub const MINI_BLOCK_NUM_BYTES: usize = MINI_BLOCK_BITVEC_NUM_BYTES + MINI_BLOCK_OFFSET_NUM_BYTES;
/// Number of bytes in a dense block.
pub const DENSE_BLOCK_NUM_BYTES: u32 =
(ELEMENTS_PER_BLOCK as u32 / ELEMENTS_PER_MINI_BLOCK as u32) * MINI_BLOCK_NUM_BYTES as u32;
pub struct DenseBlockCodec;
impl SetCodec for DenseBlockCodec {
type Item = u16;
type Reader<'a> = DenseBlock<'a>;
fn serialize(els: impl Iterator<Item = u16>, wrt: impl io::Write) -> io::Result<()> {
serialize_dense_codec(els, wrt)
}
#[inline]
fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
assert_eq!(data.len(), DENSE_BLOCK_NUM_BYTES as usize);
DenseBlock(data)
}
}
/// Interpreting the bitvec as a set of integer within 0..=63
/// and given an element, returns the number of elements in the
/// set lesser than the element.
///
/// # Panics
///
/// May panic or return a wrong result if el <= 64.
#[inline(always)]
fn rank_u64(bitvec: u64, el: u16) -> u16 {
debug_assert!(el < 64);
let mask = (1u64 << el) - 1;
let masked_bitvec = bitvec & mask;
masked_bitvec.count_ones() as u16
}
#[inline(always)]
fn select_u64(mut bitvec: u64, rank: u16) -> u16 {
for _ in 0..rank {
bitvec &= bitvec - 1;
}
bitvec.trailing_zeros() as u16
}
// TODO test the following solution on Intel... on Ryzen Zen <3 it is a catastrophy.
// #[target_feature(enable = "bmi2")]
// unsafe fn select_bitvec_unsafe(bitvec: u64, rank: u16) -> u16 {
// let pdep = _pdep_u64(1u64 << rank, bitvec);
// pdep.trailing_zeros() as u16
// }
#[derive(Clone, Copy, Debug)]
struct DenseMiniBlock {
bitvec: u64,
rank: u16,
}
impl DenseMiniBlock {
fn from_bytes(data: [u8; MINI_BLOCK_NUM_BYTES]) -> Self {
let bitvec = u64::from_le_bytes(data[..MINI_BLOCK_BITVEC_NUM_BYTES].try_into().unwrap());
let rank = u16::from_le_bytes(data[MINI_BLOCK_BITVEC_NUM_BYTES..].try_into().unwrap());
Self { bitvec, rank }
}
fn to_bytes(&self) -> [u8; MINI_BLOCK_NUM_BYTES] {
let mut bytes = [0u8; MINI_BLOCK_NUM_BYTES];
bytes[..MINI_BLOCK_BITVEC_NUM_BYTES].copy_from_slice(&self.bitvec.to_le_bytes());
bytes[MINI_BLOCK_BITVEC_NUM_BYTES..].copy_from_slice(&self.rank.to_le_bytes());
bytes
}
}
#[derive(Copy, Clone)]
pub struct DenseBlock<'a>(&'a [u8]);
impl<'a> Set<u16> for DenseBlock<'a> {
#[inline(always)]
fn contains(&self, el: u16) -> bool {
let mini_block_id = el / ELEMENTS_PER_MINI_BLOCK;
let bitvec = self.mini_block(mini_block_id).bitvec;
let pos_in_bitvec = el % ELEMENTS_PER_MINI_BLOCK;
get_bit_at(bitvec, pos_in_bitvec)
}
#[inline(always)]
fn rank_if_exists(&self, el: u16) -> Option<u16> {
let block_pos = el / ELEMENTS_PER_MINI_BLOCK;
let index_block = self.mini_block(block_pos);
let pos_in_block_bit_vec = el % ELEMENTS_PER_MINI_BLOCK;
let ones_in_block = rank_u64(index_block.bitvec, pos_in_block_bit_vec);
let rank = index_block.rank + ones_in_block;
if get_bit_at(index_block.bitvec, pos_in_block_bit_vec) {
Some(rank)
} else {
None
}
}
#[inline(always)]
fn select(&self, rank: u16) -> u16 {
let block_id = self.find_miniblock_containing_rank(rank, 0).unwrap();
let index_block = self.mini_block(block_id);
let in_block_rank = rank - index_block.rank;
block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
}
fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
let orig_ids = self.select_iter(ranks.iter().copied());
for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
*output = original_id;
}
}
}
impl<'a> DenseBlock<'a> {
/// Iterator verison of select.
///
/// # Panics
/// Panics if one of the rank is higher than the number of elements in the set.
pub fn select_iter<'b>(
&self,
rank_it: impl Iterator<Item = u16> + 'b,
) -> impl Iterator<Item = u16> + 'b
where
Self: 'b,
{
let mut block_id = 0u16;
let me = *self;
rank_it.map(move |rank| {
block_id = me.find_miniblock_containing_rank(rank, block_id).unwrap();
let index_block = me.mini_block(block_id);
let in_block_rank = rank - index_block.rank;
block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
})
}
}
impl<'a> DenseBlock<'a> {
#[inline]
fn mini_block(&self, mini_block_id: u16) -> DenseMiniBlock {
let data_start_pos = mini_block_id as usize * MINI_BLOCK_NUM_BYTES;
DenseMiniBlock::from_bytes(
self.0[data_start_pos..data_start_pos + MINI_BLOCK_NUM_BYTES]
.try_into()
.unwrap(),
)
}
#[inline]
fn iter_miniblocks(
&self,
from_block_id: u16,
) -> impl Iterator<Item = (u16, DenseMiniBlock)> + '_ {
self.0
.chunks_exact(MINI_BLOCK_NUM_BYTES)
.enumerate()
.skip(from_block_id as usize)
.map(|(block_id, bytes)| {
let mini_block = DenseMiniBlock::from_bytes(bytes.try_into().unwrap());
(block_id as u16, mini_block)
})
}
/// 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.
#[inline]
fn find_miniblock_containing_rank(&self, rank: u16, from_block_id: u16) -> Option<u16> {
self.iter_miniblocks(from_block_id)
.take_while(|(_, block)| block.rank <= rank)
.map(|(block_id, _)| block_id)
.last()
}
}
/// Iterator over all values, true if set, otherwise false
pub fn serialize_dense_codec(
els: impl Iterator<Item = u16>,
mut output: impl Write,
) -> io::Result<()> {
let mut non_null_rows_before: u16 = 0u16;
let mut block = 0u64;
let mut current_block_id = 0u16;
for el in els {
let block_id = el / ELEMENTS_PER_MINI_BLOCK;
let in_offset = el % ELEMENTS_PER_MINI_BLOCK;
while block_id > current_block_id {
let dense_mini_block = DenseMiniBlock {
bitvec: block,
rank: non_null_rows_before as u16,
};
output.write_all(&dense_mini_block.to_bytes())?;
non_null_rows_before += block.count_ones() as u16;
block = 0u64;
current_block_id += 1u16;
}
set_bit_at(&mut block, in_offset);
}
while current_block_id <= u16::MAX / ELEMENTS_PER_MINI_BLOCK {
block.serialize(&mut output)?;
non_null_rows_before.serialize(&mut output)?;
// This will overflow to 0 exactly if all bits are set.
// This is however not problem as we won't use this last value.
non_null_rows_before = non_null_rows_before.wrapping_add(block.count_ones() as u16);
block = 0u64;
current_block_id += 1u16;
}
Ok(())
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_select_bitvec() {
assert_eq!(select_u64(1u64, 0), 0);
assert_eq!(select_u64(2u64, 0), 1);
assert_eq!(select_u64(4u64, 0), 2);
assert_eq!(select_u64(8u64, 0), 3);
assert_eq!(select_u64(1 | 8u64, 0), 0);
assert_eq!(select_u64(1 | 8u64, 1), 3);
}
#[test]
fn test_count_ones() {
for i in 0..=63 {
assert_eq!(rank_u64(u64::MAX, i), i);
}
}
#[test]
fn test_dense() {
assert_eq!(DENSE_BLOCK_NUM_BYTES, 10_240);
}
}

112
columnar/src/column_index/optional_index/set_block/sparse.rs Normal file
View File

@@ -0,0 +1,112 @@
use crate::column_index::optional_index::{Set, SetCodec};
pub struct SparseBlockCodec;
impl SetCodec for SparseBlockCodec {
type Item = u16;
type Reader<'a> = SparseBlock<'a>;
fn serialize(
els: impl Iterator<Item = u16>,
mut wrt: impl std::io::Write,
) -> std::io::Result<()> {
for el in els {
wrt.write_all(&el.to_le_bytes())?;
}
Ok(())
}
fn open<'a>(data: &'a [u8]) -> Self::Reader<'a> {
SparseBlock(data)
}
}
#[derive(Copy, Clone)]
pub struct SparseBlock<'a>(&'a [u8]);
impl<'a> Set<u16> for SparseBlock<'a> {
#[inline(always)]
fn contains(&self, el: u16) -> bool {
self.binary_search(el).is_ok()
}
#[inline(always)]
fn rank_if_exists(&self, el: u16) -> Option<u16> {
self.binary_search(el).ok()
}
#[inline(always)]
fn select(&self, rank: u16) -> u16 {
let offset = rank as usize * 2;
u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
}
fn select_batch(&self, ranks: &[u16], outputs: &mut [u16]) {
let orig_ids = self.select_iter(ranks.iter().copied());
for (output, original_id) in outputs.iter_mut().zip(orig_ids) {
*output = original_id;
}
}
}
#[inline(always)]
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)
}
impl<'a> SparseBlock<'a> {
#[inline(always)]
fn value_at_idx(&self, data: &[u8], idx: u16) -> u16 {
let start_offset: usize = idx as usize * 2;
get_u16(data, start_offset)
}
#[inline]
fn num_vals(&self) -> u16 {
(self.0.len() / 2) as u16
}
#[inline]
#[allow(clippy::comparison_chain)]
// Looks for the element in the block. Returns the positions if found.
fn binary_search(&self, target: u16) -> Result<u16, u16> {
let data = &self.0;
let mut size = self.num_vals();
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 Ok(mid);
}
size = right - left;
}
Err(left)
}
pub fn select_iter<'b>(
&self,
iter: impl Iterator<Item = u16> + 'b,
) -> impl Iterator<Item = u16> + 'b
where
Self: 'b,
{
iter.map(|codec_id| {
let offset = codec_id as usize * 2;
u16::from_le_bytes(self.0[offset..offset + 2].try_into().unwrap())
})
}
}

110
columnar/src/column_index/optional_index/set_block/tests.rs Normal file
View File

@@ -0,0 +1,110 @@
use std::collections::HashMap;
use crate::column_index::optional_index::set_block::set_block::DENSE_BLOCK_NUM_BYTES;
use crate::column_index::optional_index::set_block::{DenseBlockCodec, SparseBlockCodec};
use crate::column_index::optional_index::{Set, SetCodec};
fn test_set_helper<C: SetCodec<Item = u16>>(vals: &[u16]) -> usize {
let mut buffer = Vec::new();
C::serialize(vals.iter().copied(), &mut buffer).unwrap();
let tested_set = C::open(buffer.as_slice());
let hash_set: HashMap<C::Item, C::Item> = vals
.iter()
.copied()
.enumerate()
.map(|(ord, val)| (val, C::Item::try_from(ord).ok().unwrap()))
.collect();
for val in 0u16..=u16::MAX {
assert_eq!(tested_set.contains(val), hash_set.contains_key(&val));
assert_eq!(tested_set.rank_if_exists(val), hash_set.get(&val).copied());
}
for rank in 0..vals.len() {
assert_eq!(tested_set.select(rank as u16), vals[rank]);
}
buffer.len()
}
#[test]
fn test_dense_block_set_u16_empty() {
let buffer_len = test_set_helper::<DenseBlockCodec>(&[]);
assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
}
#[test]
fn test_dense_block_set_u16_max() {
let buffer_len = test_set_helper::<DenseBlockCodec>(&[u16::MAX]);
assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
}
#[test]
fn test_sparse_block_set_u16_empty() {
let buffer_len = test_set_helper::<SparseBlockCodec>(&[]);
assert_eq!(buffer_len, 0);
}
#[test]
fn test_sparse_block_set_u16_max() {
let buffer_len = test_set_helper::<SparseBlockCodec>(&[u16::MAX]);
assert_eq!(buffer_len, 2);
}
use proptest::prelude::*;
proptest! {
#[test]
fn test_prop_test_dense(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
let vals: Vec<u16> = els.into_iter().collect();
let buffer_len = test_set_helper::<DenseBlockCodec>(&vals);
assert_eq!(buffer_len, DENSE_BLOCK_NUM_BYTES as usize);
}
#[test]
fn test_prop_test_sparse(els in proptest::collection::btree_set(0..=u16::MAX, 0..=u16::MAX as usize)) {
let vals: Vec<u16> = els.into_iter().collect();
let buffer_len = test_set_helper::<SparseBlockCodec>(&vals);
assert_eq!(buffer_len, vals.len() * 2);
}
}
#[test]
fn test_simple_translate_codec_codec_idx_to_original_idx_dense() {
let mut buffer = Vec::new();
DenseBlockCodec::serialize([1, 3, 17, 32, 30_000, 30_001].iter().copied(), &mut buffer)
.unwrap();
let tested_set = DenseBlockCodec::open(buffer.as_slice());
assert!(tested_set.contains(1));
assert_eq!(
&tested_set
.select_iter([0, 1, 2, 5].iter().copied())
.collect::<Vec<u16>>(),
&[1, 3, 17, 30_001]
);
}
#[test]
fn test_simple_translate_codec_idx_to_original_idx_sparse() {
let mut buffer = Vec::new();
SparseBlockCodec::serialize([1, 3, 17].iter().copied(), &mut buffer).unwrap();
let tested_set = SparseBlockCodec::open(buffer.as_slice());
assert!(tested_set.contains(1));
assert_eq!(
&tested_set
.select_iter([0, 1, 2].iter().copied())
.collect::<Vec<u16>>(),
&[1, 3, 17]
);
}
#[test]
fn test_simple_translate_codec_idx_to_original_idx_dense() {
let mut buffer = Vec::new();
DenseBlockCodec::serialize(0u16..150u16, &mut buffer).unwrap();
let tested_set = DenseBlockCodec::open(buffer.as_slice());
assert!(tested_set.contains(1));
let rg = 0u16..150u16;
let els: Vec<u16> = rg.clone().collect();
assert_eq!(
&tested_set.select_iter(rg.clone()).collect::<Vec<u16>>(),
&els
);
}

327
columnar/src/column_index/optional_index/tests.rs Normal file
View File

@@ -0,0 +1,327 @@
use proptest::prelude::{any, prop, *};
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use super::*;
#[test]
fn test_dense_block_threshold() {
assert_eq!(super::DENSE_BLOCK_THRESHOLD, 5_120);
}
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 test_with_random_sets_simple() {
let vals = 10..BLOCK_SIZE * 2;
let mut out: Vec<u8> = Vec::new();
serialize_optional_index(&vals.clone(), &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
let ranks: Vec<u32> = (65_472u32..65_473u32).collect();
let els: Vec<u32> = ranks.iter().copied().map(|rank| rank + 10).collect();
let mut output = vec![0u32; ranks.len()];
null_index.select_batch(&ranks[..], &mut output[..]);
assert_eq!(&output, &els);
}
#[test]
fn test_optional_index_trailing_empty_blocks() {
test_null_index(&[false]);
}
#[test]
fn test_optional_index_one_block_false() {
let mut iter = vec![false; ELEMENTS_PER_BLOCK as usize];
iter.push(true);
test_null_index(&iter[..]);
}
#[test]
fn test_optional_index_one_block_true() {
let mut iter = vec![true; ELEMENTS_PER_BLOCK as usize];
iter.push(true);
test_null_index(&iter[..]);
}
impl<'a> SerializableOptionalIndex<'a> for &'a [bool] {
fn num_rows(&self) -> RowId {
self.len() as u32
}
fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
Box::new(
self.iter()
.cloned()
.enumerate()
.filter(|(_pos, val)| *val)
.map(|(pos, _val)| pos as u32),
)
}
}
fn test_null_index(data: &[bool]) {
let mut out: Vec<u8> = Vec::new();
serialize_optional_index(&data, &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
let orig_idx_with_value: Vec<u32> = data
.iter()
.enumerate()
.filter(|(_pos, val)| **val)
.map(|(pos, _val)| pos as u32)
.collect();
let ids: Vec<u32> = (0..orig_idx_with_value.len() as u32).collect();
let mut output = vec![0u32; ids.len()];
null_index.select_batch(&ids[..], &mut output);
// assert_eq!(&output[0..100], &orig_idx_with_value[0..100]);
assert_eq!(output, 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.rank_if_exists(*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.contains(pos as u32), *value);
}
}
#[test]
fn test_optional_index_test_translation() {
let mut out = vec![];
let iter = &[true, false, true, false];
serialize_optional_index(&&iter[..], &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
let mut output = vec![0u32; 2];
null_index.select_batch(&[0, 1], &mut output);
assert_eq!(output, &[0, 2]);
}
#[test]
fn test_optional_index_translate() {
let mut out = vec![];
let iter = &[true, false, true, false];
serialize_optional_index(&&iter[..], &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
assert_eq!(null_index.rank_if_exists(0), Some(0));
assert_eq!(null_index.rank_if_exists(2), Some(1));
}
#[test]
fn test_optional_index_small() {
let mut out = vec![];
let iter = &[true, false, true, false];
serialize_optional_index(&&iter[..], &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
assert!(null_index.contains(0));
assert!(!null_index.contains(1));
assert!(null_index.contains(2));
assert!(!null_index.contains(3));
}
#[test]
fn test_optional_index_large() {
let mut docs = vec![];
docs.extend((0..ELEMENTS_PER_BLOCK).map(|_idx| false));
docs.extend((0..=1).map(|_idx| true));
let mut out = vec![];
serialize_optional_index(&&docs[..], &mut out).unwrap();
let null_index = open_optional_index(OwnedBytes::new(out)).unwrap();
assert!(!null_index.contains(0));
assert!(!null_index.contains(100));
assert!(!null_index.contains(ELEMENTS_PER_BLOCK - 1));
assert!(null_index.contains(ELEMENTS_PER_BLOCK));
assert!(null_index.contains(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) -> OptionalIndex {
let mut out = Vec::new();
let mut rng: StdRng = StdRng::from_seed([1u8; 32]);
let vals: Vec<bool> = (0..TOTAL_NUM_VALUES)
.map(|_| rng.gen_bool(fill_ratio))
.collect();
serialize_optional_index(&&vals[..], &mut out).unwrap();
let codec = open_optional_index(OwnedBytes::new(out)).unwrap();
codec
}
fn random_range_iterator(
start: u32,
end: u32,
avg_step_size: u32,
avg_deviation: 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(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
if current >= end {
None
} else {
Some(current)
}
})
}
fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
let ratio = percent as f32 / 100.0;
let step_size = (1f32 / ratio) as u32;
let deviation = step_size - 1;
random_range_iterator(0, num_values, step_size, deviation)
}
fn walk_over_data(codec: &OptionalIndex, avg_step_size: u32) -> Option<u32> {
walk_over_data_from_positions(
codec,
random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
)
}
fn walk_over_data_from_positions(
codec: &OptionalIndex,
positions: impl Iterator<Item = u32>,
) -> Option<u32> {
let mut dense_idx: Option<u32> = None;
for idx in positions {
dense_idx = dense_idx.or(codec.rank_if_exists(idx));
}
dense_idx
}
#[bench]
fn bench_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 1000));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.5f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.01f64, 0.005f32, bench);
}
#[bench]
fn bench_translate_codec_to_orig_10percent_filled_0comma005percent_hit(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.1f64, 0.005f32, bench);
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.01f64, 10f32, bench);
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_full_scan(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.01f64, 100f32, bench);
}
fn bench_translate_codec_to_orig_util(
percent_filled: f64,
percent_hit: f32,
bench: &mut Bencher,
) {
let codec = gen_bools(percent_filled);
let num_non_nulls = codec.num_non_nulls();
let idxs: Vec<u32> = if percent_hit == 100.0f32 {
(0..num_non_nulls).collect()
} else {
n_percent_step_iterator(percent_hit, num_non_nulls).collect()
};
let mut output = vec![0u32; idxs.len()];
bench.iter(|| {
codec.select_batch(&idxs[..], &mut output);
});
}
#[bench]
fn bench_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.9f64, 0.005, bench);
}
#[bench]
fn bench_translate_codec_to_orig_90percent_filled_full_scan(bench: &mut Bencher) {
bench_translate_codec_to_orig_util(0.9f64, 100.0f32, bench);
}
}

70
columnar/src/column_index/serialize.rs Normal file
View File

@@ -0,0 +1,70 @@
use std::io;
use std::io::Write;
use common::OwnedBytes;
use crate::column_index::multivalued_index::{serialize_multivalued_index, MultivaluedIndex};
use crate::column_index::optional_index::serialize_optional_index;
use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
use crate::Cardinality;
pub enum SerializableColumnIndex<'a> {
Full,
Optional(Box<dyn SerializableOptionalIndex<'a> + 'a>),
// TODO remove the Arc<dyn> apart from serialization this is not
// dynamic at all.
Multivalued(MultivaluedIndex),
}
impl<'a> SerializableColumnIndex<'a> {
pub fn get_cardinality(&self) -> Cardinality {
match self {
SerializableColumnIndex::Full => Cardinality::Full,
SerializableColumnIndex::Optional(_) => Cardinality::Optional,
SerializableColumnIndex::Multivalued(_) => Cardinality::Multivalued,
}
}
}
pub fn serialize_column_index(
column_index: SerializableColumnIndex,
output: &mut impl Write,
) -> io::Result<()> {
let cardinality = column_index.get_cardinality().to_code();
output.write_all(&[cardinality])?;
match column_index {
SerializableColumnIndex::Full => {}
SerializableColumnIndex::Optional(optional_index) => {
serialize_optional_index(&*optional_index, output)?
}
SerializableColumnIndex::Multivalued(multivalued_index) => {
serialize_multivalued_index(multivalued_index, output)?
}
}
Ok(())
}
pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex<'static>> {
if bytes.is_empty() {
return Err(io::Error::new(
io::ErrorKind::UnexpectedEof,
"Failed to deserialize column index. Empty buffer.",
));
}
let cardinality_code = bytes[0];
let cardinality = Cardinality::try_from_code(cardinality_code)?;
bytes.advance(1);
match cardinality {
Cardinality::Full => Ok(ColumnIndex::Full),
Cardinality::Optional => {
let optional_index = super::optional_index::open_optional_index(bytes)?;
Ok(ColumnIndex::Optional(optional_index))
}
Cardinality::Multivalued => {
let multivalued_index = super::multivalued_index::open_multivalued_index(bytes)?;
Ok(ColumnIndex::Multivalued(multivalued_index))
}
}
}
// TODO unit tests

115
columnar/src/column_values/bitpacked.rs Normal file
View File

@@ -0,0 +1,115 @@
use std::io::{self, Write};
use common::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use super::serialize::NormalizedHeader;
use super::{ColumnValues, FastFieldCodec, FastFieldCodecType};
/// Depending on the field type, a different
/// fast field is required.
#[derive(Clone)]
pub struct BitpackedReader {
data: OwnedBytes,
bit_unpacker: BitUnpacker,
normalized_header: NormalizedHeader,
}
impl ColumnValues for BitpackedReader {
#[inline]
fn get_val(&self, doc: u32) -> u64 {
self.bit_unpacker.get(doc, &self.data)
}
#[inline]
fn min_value(&self) -> u64 {
// The BitpackedReader assumes a normalized vector.
0
}
#[inline]
fn max_value(&self) -> u64 {
self.normalized_header.max_value
}
#[inline]
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}
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,
})
}
/// Serializes data with the BitpackedFastFieldSerializer.
///
/// 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.
///
/// Ideally, we made a shift upstream on the column so that `col.min_value() == 0`.
fn serialize(column: &dyn ColumnValues, 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)?;
}
bit_packer.close(write)?;
Ok(())
}
fn estimate(column: &dyn ColumnValues) -> Option<f32> {
let num_bits = compute_num_bits(column.max_value());
let num_bits_uncompressed = 64;
Some(num_bits as f32 / num_bits_uncompressed as f32)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::column_values::tests::create_and_validate;
fn create_and_validate_bitpacked_codec(data: &[u64], name: &str) {
create_and_validate::<BitpackedCodec>(data, name);
}
#[test]
fn test_with_codec_data_sets() {
let data_sets = crate::column_values::tests::get_codec_test_datasets();
for (mut data, name) in data_sets {
create_and_validate_bitpacked_codec(&data, name);
data.reverse();
create_and_validate::<BitpackedCodec>(&data, name);
}
}
#[test]
fn bitpacked_fast_field_rand() {
for _ in 0..500 {
let mut data = (0..1 + rand::random::<u8>() as usize)
.map(|_| rand::random::<i64>() as u64 / 2)
.collect::<Vec<_>>();
create_and_validate_bitpacked_codec(&data, "rand");
data.reverse();
create_and_validate::<BitpackedCodec>(&data, "rand");
}
}
}

188
columnar/src/column_values/blockwise_linear.rs Normal file
View File

@@ -0,0 +1,188 @@
use std::sync::Arc;
use std::{io, iter};
use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::column_values::line::Line;
use crate::column_values::serialize::NormalizedHeader;
use crate::column_values::{ColumnValues, 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: FastFieldCodecType = FastFieldCodecType::BlockwiseLinear;
type Reader = BlockwiseLinearReader;
fn open_from_bytes(
bytes: common::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 ColumnValues) -> 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 ColumnValues, 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 ColumnValues 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)
}
#[inline(always)]
fn min_value(&self) -> u64 {
// The BlockwiseLinearReader assumes a normalized vector.
0u64
}
#[inline(always)]
fn max_value(&self) -> u64 {
self.normalized_header.max_value
}
#[inline(always)]
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}

350
columnar/src/column_values/column.rs Normal file
View File

@@ -0,0 +1,350 @@
use std::marker::PhantomData;
use std::ops::{Range, RangeInclusive};
use tantivy_bitpacker::minmax;
use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
/// `ColumnValues` provides access to a dense field column.
///
/// `Column` are just a wrapper over `ColumnValues` and a `ColumnIndex`.
pub trait ColumnValues<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)))
}
}
impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd> ColumnValues<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)
}
}
/// VecColumn provides `Column` over a slice.
pub struct VecColumn<'a, T = u64> {
pub(crate) values: &'a [T],
pub(crate) min_value: T,
pub(crate) max_value: T,
}
impl<'a, T: Copy + PartialOrd + Send + Sync> ColumnValues<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 + PartialOrd + 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 ColumnValues<Output>
where
C: ColumnValues<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> ColumnValues<Output> for MonotonicMappingColumn<C, T, Input>
where
C: ColumnValues<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> ColumnValues<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::column_values::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);
}
}

19
columnar/src/column_values/column_with_cardinality.rs Normal file
View File

@@ -0,0 +1,19 @@
// 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/>.
//

43
columnar/src/column_values/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
columnar/src/column_values/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);
}
}

813
columnar/src/column_values/compact_space/mod.rs Normal file
View File

@@ -0,0 +1,813 @@
/// 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, OwnedBytes, VInt, VIntU128};
use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
use crate::column_values::compact_space::build_compact_space::get_compact_space;
use crate::column_values::ColumnValues;
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 ColumnValues<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
}
}
// TODO reenable what can be reenabled.
// #[cfg(test)]
// mod tests {
//
// use super::*;
// use crate::column::format_version::read_format_version;
// use crate::column::column_footer::read_null_index_footer;
// use crate::column::serialize::U128Header;
// use crate::column::{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: Vec<u32> = get_positions_for_value_range_helper(&decomp, 0..=1, 1..u32::MAX);
// assert!(positions.is_empty());
//
// 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()
// ),
// &[3]
// );
// assert!(get_positions_for_value_range_helper(
// &decomp,
// 99998u128..=99998u128,
// complete_range.clone()
// )
// .is_empty());
// assert_eq!(
// &get_positions_for_value_range_helper(
// &decomp,
// 333u128..=333u128,
// complete_range.clone()
// ),
// &[8]
// );
// assert_eq!(
// &get_positions_for_value_range_helper(
// &decomp,
// 332u128..=333u128,
// complete_range.clone()
// ),
// &[8]
// );
// assert_eq!(
// &get_positions_for_value_range_helper(
// &decomp,
// 332u128..=334u128,
// complete_range.clone()
// ),
// &[8]
// );
// assert_eq!(
// &get_positions_for_value_range_helper(
// &decomp,
// 333u128..=334u128,
// complete_range.clone()
// ),
// &[8]
// );
//
// assert_eq!(
// &get_positions_for_value_range_helper(
// &decomp,
// 4_000_211_221u128..=5_000_000_000u128,
// complete_range
// ),
// &[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!(
// &get_positions_for_value_range_helper(&decomp, 0..=5, complete_range.clone())
// .is_empty(),
// );
// assert_eq!(
// &get_positions_for_value_range_helper(&decomp, 0..=100, complete_range.clone()),
// &[0]
// );
// assert_eq!(
// &get_positions_for_value_range_helper(&decomp, 0..=105, complete_range),
// &[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);
// }
// }
// }
//

75
columnar/src/column_values/gcd.rs Normal file
View File

@@ -0,0 +1,75 @@
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::num::NonZeroU64;
use crate::column_values::gcd::{compute_gcd, find_gcd};
#[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);
}
}

222
columnar/src/column_values/line.rs Normal file
View File

@@ -0,0 +1,222 @@
use std::io;
use std::num::NonZeroU32;
use common::{BinarySerializable, VInt};
use crate::column_values::ColumnValues;
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 ColumnValues) -> 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::column_values::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));
}
}

230
columnar/src/column_values/linear.rs Normal file
View File

@@ -0,0 +1,230 @@
use std::io::{self, Write};
use common::{BinarySerializable, OwnedBytes};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use super::line::Line;
use super::serialize::NormalizedHeader;
use super::{ColumnValues, 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 ColumnValues 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(always)]
fn min_value(&self) -> u64 {
// The LinearReader assumes a normalized vector.
0u64
}
#[inline(always)]
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 ColumnValues, 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 ColumnValues) -> 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::column_values::tests;
fn create_and_validate(data: &[u64], name: &str) -> Option<(f32, f32)> {
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 = tests::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");
}
}
}

222
columnar/src/column_values/main.rs Normal file
View File

@@ -0,0 +1,222 @@
#[macro_use]
extern crate prettytable;
use std::collections::HashSet;
use std::env;
use std::io::BufRead;
use std::net::{IpAddr, Ipv6Addr};
use std::str::FromStr;
use common::OwnedBytes;
use fastfield_codecs::{open_u128, serialize_u128, Column, FastFieldCodecType, VecColumn};
use itertools::Itertools;
use measure_time::print_time;
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 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())
.cloned()
.unwrap();
table.add_row(Row::new(vec![Cell::new(data_set_name).style_spec("Bbb")]));
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(&format!("{codec_type:?}")).style_spec("bFg"),
Cell::new(&ratio_cell).style_spec(style),
Cell::new(&est_cell).style_spec(""),
]));
}
}
table.printstd();
}
pub fn get_codec_test_data_sets() -> Vec<(Vec<u64>, &'static str)> {
let mut data_and_names = vec![];
let data = (1000..=200_000_u64).collect::<Vec<_>>();
data_and_names.push((data, "Autoincrement"));
let mut current_cumulative = 0;
let data = (1..=200_000_u64)
.map(|num| {
let num = (num as f32 + num as f32).log10() as u64;
current_cumulative += num;
current_cumulative
})
.collect::<Vec<_>>();
// let data = (1..=200000_u64).map(|num| num + num).collect::<Vec<_>>();
data_and_names.push((data, "Monotonically increasing concave"));
let mut current_cumulative = 0;
let data = (1..=200_000_u64)
.map(|num| {
let num = (200_000.0 - num as f32).log10() as u64;
current_cumulative += num;
current_cumulative
})
.collect::<Vec<_>>();
data_and_names.push((data, "Monotonically increasing convex"));
let data = (1000..=200_000_u64)
.map(|num| num + rand::random::<u8>() as u64)
.collect::<Vec<_>>();
data_and_names.push((data, "Almost monotonically increasing"));
data_and_names
}
pub fn serialize_with_codec(
data: &[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))
}

333
columnar/src/column_values/mod.rs Normal file
View File

@@ -0,0 +1,333 @@
#![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)]
mod tests;
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::{BinarySerializable, OwnedBytes};
use compact_space::CompactSpaceDecompressor;
use monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
};
use serialize::{Header, U128Header};
mod bitpacked;
mod blockwise_linear;
mod compact_space;
mod line;
mod linear;
pub(crate) mod monotonic_mapping;
// mod monotonic_mapping_u128;
mod column;
mod column_with_cardinality;
mod gcd;
pub mod serialize;
pub use self::column::{monotonic_map_column, ColumnValues, IterColumn, VecColumn};
pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
// pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
pub use self::serialize::{serialize_and_load, serialize_column_values, NormalizedHeader};
use crate::column_values::bitpacked::BitpackedCodec;
use crate::column_values::blockwise_linear::BlockwiseLinearCodec;
use crate::column_values::linear::LinearCodec;
#[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>>> {
// todo!();
// // 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_u64_mapped<T: MonotonicallyMappableToU64>(
mut bytes: OwnedBytes,
) -> io::Result<Arc<dyn ColumnValues<T>>> {
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 ColumnValues<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(crate) trait FastFieldCodec: 'static {
/// A codex needs to provide a unique name and id, which is
/// used for debugging and de/serialization.
const CODEC_TYPE: FastFieldCodecType;
type Reader: ColumnValues<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 ColumnValues, 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(column: &dyn ColumnValues) -> Option<f32>;
}
/// The list of all available codecs for u64 convertible data.
pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
FastFieldCodecType::Linear,
];
#[cfg(all(test, feature = "unstable"))]
mod bench {
use std::sync::Arc;
use common::OwnedBytes;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::{self, Bencher};
use super::*;
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 = 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 ColumnValues>) {
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);
}
}

264
columnar/src/column_values/monotonic_mapping.rs Normal file
View File

@@ -0,0 +1,264 @@
use std::marker::PhantomData;
use fastdivide::DividerU64;
use crate::RowId;
/// 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>
{
#[inline(always)]
fn mapping(&self, val: To) -> From {
self.orig_mapping.inverse(val)
}
#[inline(always)]
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,
}
}
}
// TODO
// impl<External: MonotonicallyMappableToU128, T: MonotonicallyMappableToU128>
// StrictlyMonotonicFn<External, u128> for StrictlyMonotonicMappingToInternal<T>
// where T: MonotonicallyMappableToU128
// {
// #[inline(always)]
// fn mapping(&self, inp: External) -> u128 {
// External::to_u128(inp)
// }
// #[inline(always)]
// fn inverse(&self, out: u128) -> External {
// External::from_u128(out)
// }
// }
impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
where T: MonotonicallyMappableToU64
{
#[inline(always)]
fn mapping(&self, inp: External) -> u64 {
External::to_u64(inp)
}
#[inline(always)]
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
{
#[inline(always)]
fn mapping(&self, inp: External) -> u64 {
self.gcd_divider
.divide(External::to_u64(inp) - self.min_value)
}
#[inline(always)]
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 {
#[inline(always)]
pub(crate) fn new(min_value: u64) -> Self {
Self { min_value }
}
}
impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
for StrictlyMonotonicMappingToInternalBaseval
{
#[inline(always)]
fn mapping(&self, val: External) -> u64 {
External::to_u64(val) - self.min_value
}
#[inline(always)]
fn inverse(&self, val: u64) -> External {
External::from_u64(self.min_value + val)
}
}
impl MonotonicallyMappableToU64 for u64 {
#[inline(always)]
fn to_u64(self) -> u64 {
self
}
#[inline(always)]
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
}
}
impl MonotonicallyMappableToU64 for RowId {
#[inline(always)]
fn to_u64(self) -> u64 {
u64::from(self)
}
#[inline(always)]
fn from_u64(val: u64) -> RowId {
val as RowId
}
}
// TODO remove me.
// Tantivy should refuse NaN values and work with NotNaN internally.
impl MonotonicallyMappableToU64 for f64 {
#[inline(always)]
fn to_u64(self) -> u64 {
common::f64_to_u64(self)
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
common::u64_to_f64(val)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[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);
// TODO
// 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
columnar/src/column_values/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())
}

343
columnar/src/column_values/serialize.rs Normal file
View File

@@ -0,0 +1,343 @@
// 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/>.
use std::io;
use std::num::NonZeroU64;
use std::sync::Arc;
use common::{BinarySerializable, OwnedBytes, VInt};
use log::warn;
use super::bitpacked::BitpackedCodec;
use super::blockwise_linear::BlockwiseLinearCodec;
use super::linear::LinearCodec;
use super::monotonic_mapping::{
StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalGCDBaseval,
};
use super::{
monotonic_map_column, ColumnValues, 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(crate) fn normalize_column<C: ColumnValues>(&self, from_column: C) -> impl ColumnValues {
normalize_column(from_column, self.min_value, self.gcd)
}
pub fn compute_header(
column: impl ColumnValues<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 = super::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,
})
}
}
fn normalize_column<C: ColumnValues>(
from_column: C,
min_value: u64,
gcd: Option<NonZeroU64>,
) -> impl ColumnValues {
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(crate) fn estimate<T: MonotonicallyMappableToU64>(
typed_column: impl ColumnValues<T>,
codec_type: FastFieldCodecType,
) -> Option<f32> {
let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
let min_value = column.min_value();
let gcd = super::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),
}
}
// TODO
/// Serializes u128 values with the compact space codec.
// pub fn serialize_u128_new<F: Fn() -> I, I: Iterator<Item = u128>>(
// value_index: ColumnIndex,
// 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 = ColumnFooter {
// 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_column_values<T: MonotonicallyMappableToU64>(
typed_column: impl ColumnValues<T>,
codecs: &[FastFieldCodecType],
output: &mut impl io::Write,
) -> 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)?;
Ok(())
}
fn detect_codec(
column: impl ColumnValues<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)
}
pub(crate) fn serialize_given_codec(
column: impl ColumnValues<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)?;
}
}
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 ColumnValues<T>> {
let mut buffer = Vec::new();
super::serialize_column_values(&VecColumn::from(&column), &ALL_CODEC_TYPES, &mut buffer)
.unwrap();
super::open_u64_mapped(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_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
// TODO put the header as a footer so that it serves as a padding.
// 5 bytes of header, 1 byte of value, 7 bytes of padding.
assert_eq!(buffer.len(), 5 + 1 + 7);
}
#[test]
fn test_fastfield_bool_bit_size_bitwidth_0() {
let mut buffer = Vec::new();
let col = VecColumn::from(&[true][..]);
serialize_column_values(&col, &ALL_CODEC_TYPES, &mut buffer).unwrap();
// 5 bytes of header, 0 bytes of value, 7 bytes of padding.
assert_eq!(buffer.len(), 5 + 7);
}
#[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_column_values(&col, &[FastFieldCodecType::Bitpacked], &mut buffer).unwrap();
// Values are stored over 3 bits.
assert_eq!(buffer.len(), 7 + (3 * 80 / 8) + 7);
}
}

309
columnar/src/column_values/tests.rs Normal file
View File

@@ -0,0 +1,309 @@
use proptest::prelude::*;
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use super::bitpacked::BitpackedCodec;
use super::blockwise_linear::BlockwiseLinearCodec;
use super::linear::LinearCodec;
use super::serialize::Header;
pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
data: &[u64],
name: &str,
) -> 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_column_values(&col, &[Codec::CODEC_TYPE], &mut out).unwrap();
let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
let reader = super::open_u64_mapped::<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))
}
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..=10_000_u64).collect::<Vec<_>>();
data_and_names.push((data, "simple monotonically increasing"));
data_and_names.push((
vec![5, 6, 7, 8, 9, 10, 99, 100],
"offset in linear interpol",
));
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<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)> =
tests::create_and_validate::<C>(&data, dataset_name);
let result = if let Some((estimate, actual)) = estimate_actual_opt {
format!("Estimate `{estimate}` Actual `{actual}`")
} else {
"Disabled".to_string()
};
println!("Codec {codec_name}, DataSet {dataset_name}, {result}");
}
}
#[test]
fn test_codec_bitpacking() {
test_codec::<BitpackedCodec>();
}
#[test]
fn test_codec_interpolation() {
test_codec::<LinearCodec>();
}
#[test]
fn test_codec_multi_interpolation() {
test_codec::<BlockwiseLinearCodec>();
}
use super::*;
#[test]
fn estimation_good_interpolation_case() {
let data = (10..=20000_u64).collect::<Vec<_>>();
let data: VecColumn = data.as_slice().into();
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.01);
let multi_linear_interpol_estimation = BlockwiseLinearCodec::estimate(&data).unwrap();
assert_le!(multi_linear_interpol_estimation, 0.2);
assert_lt!(linear_interpol_estimation, multi_linear_interpol_estimation);
let bitpacked_estimation = BitpackedCodec::estimate(&data).unwrap();
assert_lt!(linear_interpol_estimation, bitpacked_estimation);
}
#[test]
fn estimation_test_bad_interpolation_case() {
let data: &[u64] = &[200, 10, 10, 10, 10, 1000, 20];
let data: VecColumn = data.into();
let linear_interpol_estimation = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.34);
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: 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 = LinearCodec::estimate(&data).unwrap();
assert_le!(linear_interpol_estimation, 0.35);
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);
}
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::column_values::serialize_column_values(
&VecColumn::from(&vals),
&[codec_type],
&mut buffer,
)?;
let buffer = OwnedBytes::new(buffer);
let column = crate::column_values::open_u64_mapped::<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::column_values::serialize_column_values(
&VecColumn::from(&vals),
&[codec_type],
&mut buffer_without_gcd,
)?;
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::column_values::serialize_column_values(
&VecColumn::from(&vals),
&[codec_type],
&mut buffer,
)?;
let buffer = OwnedBytes::new(buffer);
let column = crate::column_values::open_u64_mapped::<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::column_values::serialize_column_values(
&VecColumn::from(&vals),
&[codec_type],
&mut buffer_without_gcd,
)?;
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::column_values::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);
}

126
columnar/src/columnar/column_type.rs Normal file
View File

@@ -0,0 +1,126 @@
use crate::utils::{place_bits, select_bits};
use crate::value::NumericalType;
use crate::InvalidData;
/// The column type represents the column type and can fit on 6-bits.
///
/// - bits[0..3]: Column category type.
/// - bits[3..6]: Numerical type if necessary.
#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy)]
pub enum ColumnType {
Bytes,
Numerical(NumericalType),
Bool,
}
impl ColumnType {
/// Encoded over 6 bits.
pub(crate) fn to_code(self) -> u8 {
let column_type_category;
let numerical_type_code: u8;
match self {
ColumnType::Bytes => {
column_type_category = ColumnTypeCategory::Str;
numerical_type_code = 0u8;
}
ColumnType::Numerical(numerical_type) => {
column_type_category = ColumnTypeCategory::Numerical;
numerical_type_code = numerical_type.to_code();
}
ColumnType::Bool => {
column_type_category = ColumnTypeCategory::Bool;
numerical_type_code = 0u8;
}
}
place_bits::<0, 3>(column_type_category.to_code()) | place_bits::<3, 6>(numerical_type_code)
}
pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
if select_bits::<6, 8>(code) != 0u8 {
return Err(InvalidData);
}
let column_type_category_code = select_bits::<0, 3>(code);
let numerical_type_code = select_bits::<3, 6>(code);
let column_type_category = ColumnTypeCategory::try_from_code(column_type_category_code)?;
match column_type_category {
ColumnTypeCategory::Bool => {
if numerical_type_code != 0u8 {
return Err(InvalidData);
}
Ok(ColumnType::Bool)
}
ColumnTypeCategory::Str => {
if numerical_type_code != 0u8 {
return Err(InvalidData);
}
Ok(ColumnType::Bytes)
}
ColumnTypeCategory::Numerical => {
let numerical_type = NumericalType::try_from_code(numerical_type_code)?;
Ok(ColumnType::Numerical(numerical_type))
}
}
}
}
/// Column types are grouped into different categories that
/// corresponds to the different types of `JsonValue` types.
///
/// The columnar writer will apply coercion rules to make sure that
/// at most one column exist per `ColumnTypeCategory`.
///
/// See also [README.md].
#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq, Debug)]
#[repr(u8)]
pub(crate) enum ColumnTypeCategory {
Bool = 0u8,
Str = 1u8,
Numerical = 2u8,
}
impl ColumnTypeCategory {
pub fn to_code(self) -> u8 {
self as u8
}
pub fn try_from_code(code: u8) -> Result<Self, InvalidData> {
match code {
0u8 => Ok(Self::Bool),
1u8 => Ok(Self::Str),
2u8 => Ok(Self::Numerical),
_ => Err(InvalidData),
}
}
}
#[cfg(test)]
mod tests {
use std::collections::HashSet;
use super::*;
use crate::Cardinality;
#[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 Ok(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(), 2 + 3);
}
#[test]
fn test_cardinality_to_code() {
let mut num_cardinality = 0;
for code in u8::MIN..=u8::MAX {
if let Ok(cardinality) = Cardinality::try_from_code(code) {
assert_eq!(cardinality.to_code(), code);
num_cardinality += 1;
}
}
assert_eq!(num_cardinality, 3);
}
}

73
columnar/src/columnar/format_version.rs Normal file
View File

@@ -0,0 +1,73 @@
use crate::InvalidData;
pub const VERSION_FOOTER_NUM_BYTES: usize = MAGIC_BYTES.len() + std::mem::size_of::<u32>();
/// We end the file by these 4 bytes just to somewhat identify that
/// this is indeed a columnar file.
const MAGIC_BYTES: [u8; 4] = [2, 113, 119, 066];
pub fn footer() -> [u8; VERSION_FOOTER_NUM_BYTES] {
let mut footer_bytes = [0u8; VERSION_FOOTER_NUM_BYTES];
footer_bytes[0..4].copy_from_slice(&Version::V1.to_bytes());
footer_bytes[4..8].copy_from_slice(&MAGIC_BYTES[..]);
footer_bytes
}
pub fn parse_footer(footer_bytes: [u8; VERSION_FOOTER_NUM_BYTES]) -> Result<Version, InvalidData> {
if footer_bytes[4..8] != MAGIC_BYTES {
return Err(InvalidData);
}
Version::try_from_bytes(footer_bytes[0..4].try_into().unwrap())
}
#[derive(Debug, Copy, Clone, Eq, PartialEq)]
#[repr(u32)]
pub enum Version {
V1 = 1u32,
}
impl Version {
fn to_bytes(&self) -> [u8; 4] {
(*self as u32).to_le_bytes()
}
fn try_from_bytes(bytes: [u8; 4]) -> Result<Version, InvalidData> {
let code = u32::from_le_bytes(bytes);
match code {
1u32 => Ok(Version::V1),
_ => Err(InvalidData),
}
}
}
#[cfg(test)]
mod tests {
use std::collections::HashSet;
use super::*;
#[test]
fn test_footer_dserialization() {
let parsed_version: Version = parse_footer(footer()).unwrap();
assert_eq!(Version::V1, parsed_version);
}
#[test]
fn test_version_serialization() {
let version_to_tests: Vec<u32> = [0, 1 << 8, 1 << 16, 1 << 24]
.iter()
.copied()
.flat_map(|offset| (0..255).map(move |el| el + offset))
.collect();
let mut valid_versions: HashSet<u32> = HashSet::default();
for &i in &version_to_tests {
let version_res = Version::try_from_bytes(i.to_le_bytes());
if let Ok(version) = version_res {
assert_eq!(version, Version::V1);
assert_eq!(version.to_bytes(), i.to_le_bytes());
valid_versions.insert(i);
}
}
assert_eq!(valid_versions.len(), 1);
}
}

28
columnar/src/columnar/mod.rs Normal file
View File

@@ -0,0 +1,28 @@
// 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;
mod format_version;
mod reader;
mod writer;
pub use column_type::ColumnType;
pub use reader::ColumnarReader;
pub use writer::ColumnarWriter;

121
columnar/src/columnar/reader/mod.rs Normal file
View File

@@ -0,0 +1,121 @@
use std::{io, mem};
use common::file_slice::FileSlice;
use common::BinarySerializable;
use sstable::{Dictionary, RangeSSTable};
use crate::columnar::{format_version, ColumnType};
use crate::dynamic_column::DynamicColumnHandle;
fn io_invalid_data(msg: String) -> io::Error {
io::Error::new(io::ErrorKind::InvalidData, msg)
}
/// The ColumnarReader makes it possible to access a set of columns
/// associated to field names.
pub struct ColumnarReader {
column_dictionary: Dictionary<RangeSSTable>,
column_data: FileSlice,
}
impl ColumnarReader {
/// Opens a new Columnar file.
pub fn open<F>(file_slice: F) -> io::Result<ColumnarReader>
where FileSlice: From<F> {
Self::open_inner(file_slice.into())
}
fn open_inner(file_slice: FileSlice) -> io::Result<ColumnarReader> {
let (file_slice_without_sstable_len, footer_slice) = file_slice
.split_from_end(mem::size_of::<u64>() + format_version::VERSION_FOOTER_NUM_BYTES);
let footer_bytes = footer_slice.read_bytes()?;
let (mut sstable_len_bytes, version_footer_bytes) =
footer_bytes.rsplit(format_version::VERSION_FOOTER_NUM_BYTES);
let version_footer_bytes: [u8; format_version::VERSION_FOOTER_NUM_BYTES] =
version_footer_bytes.as_slice().try_into().unwrap();
let _version = format_version::parse_footer(version_footer_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,
})
}
// TODO fix ugly API
pub fn list_columns(&self) -> io::Result<Vec<(String, DynamicColumnHandle)>> {
let mut stream = self.column_dictionary.stream()?;
let mut results = Vec::new();
while stream.advance() {
let key_bytes: &[u8] = stream.key();
let column_code: u8 = key_bytes.last().cloned().unwrap();
let column_type: ColumnType = ColumnType::try_from_code(column_code)
.map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
let range = stream.value().clone();
let column_name =
String::from_utf8_lossy(&key_bytes[..key_bytes.len() - 1]).to_string();
let file_slice = self
.column_data
.slice(range.start as usize..range.end as usize);
let column_handle = DynamicColumnHandle {
file_slice,
column_type,
};
results.push((column_name, column_handle));
}
Ok(results)
}
/// Get all columns for the given column name.
///
/// There can be more than one column associated to a given column name, provided they have
/// different types.
// TODO fix ugly API
pub fn read_columns(&self, column_name: &str) -> io::Result<Vec<DynamicColumnHandle>> {
// Each column is a associated to a given `column_key`,
// that starts by `column_name\0column_header`.
//
// Listing the columns associated to the given column name is therefore equivalent to
// listing `column_key` with the prefix `column_name\0`.
//
// This is in turn equivalent to searching for the range
// `[column_name,\0`..column_name\1)`.
// TODO can we get some more generic `prefix(..)` logic in the dictioanry.
let mut start_key = column_name.to_string();
start_key.push('\0');
let mut end_key = column_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();
assert!(key_bytes.starts_with(start_key.as_bytes()));
let column_code: u8 = key_bytes.last().cloned().unwrap();
let column_type = ColumnType::try_from_code(column_code)
.map_err(|_| io_invalid_data(format!("Unknown column code `{column_code}`")))?;
let range = stream.value().clone();
let file_slice = self
.column_data
.slice(range.start as usize..range.end as usize);
let dynamic_column_handle = DynamicColumnHandle {
file_slice,
column_type,
};
results.push(dynamic_column_handle);
}
Ok(results)
}
/// Return the number of columns in the columnar.
pub fn num_columns(&self) -> usize {
self.column_dictionary.num_terms()
}
}

346
columnar/src/columnar/writer/column_operation.rs Normal file
View File

@@ -0,0 +1,346 @@
use crate::dictionary::UnorderedId;
use crate::utils::{place_bits, pop_first_byte, select_bits};
use crate::value::NumericalValue;
use crate::{InvalidData, NumericalType, RowId};
/// When we build a columnar dataframe, we first just group
/// all mutations per column, and appends them in append-only buffer
/// in the stacker.
///
/// These ColumnOperation<T> are therefore serialize/deserialized
/// in memory.
///
/// We represents all of these operations as `ColumnOperation`.
#[derive(Eq, PartialEq, Debug, Clone, Copy)]
pub(super) enum ColumnOperation<T> {
NewDoc(RowId),
Value(T),
}
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
struct ColumnOperationMetadata {
op_type: ColumnOperationType,
len: u8,
}
impl ColumnOperationMetadata {
fn to_code(self) -> u8 {
place_bits::<0, 4>(self.len) | place_bits::<4, 8>(self.op_type.to_code())
}
fn try_from_code(code: u8) -> Result<Self, InvalidData> {
let len = select_bits::<0, 4>(code);
let typ_code = select_bits::<4, 8>(code);
let column_type = ColumnOperationType::try_from_code(typ_code)?;
Ok(ColumnOperationMetadata {
op_type: column_type,
len,
})
}
}
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
#[repr(u8)]
enum ColumnOperationType {
NewDoc = 0u8,
AddValue = 1u8,
}
impl ColumnOperationType {
pub fn to_code(self) -> u8 {
self as u8
}
pub fn try_from_code(code: u8) -> Result<Self, InvalidData> {
match code {
0 => Ok(Self::NewDoc),
1 => Ok(Self::AddValue),
_ => Err(InvalidData),
}
}
}
impl<V: SymbolValue> ColumnOperation<V> {
pub(super) fn serialize(self) -> impl AsRef<[u8]> {
let mut minibuf = MiniBuffer::default();
let column_op_metadata = match self {
ColumnOperation::NewDoc(new_doc) => {
let symbol_len = new_doc.serialize(&mut minibuf.bytes[1..]);
ColumnOperationMetadata {
op_type: ColumnOperationType::NewDoc,
len: symbol_len,
}
}
ColumnOperation::Value(val) => {
let symbol_len = val.serialize(&mut minibuf.bytes[1..]);
ColumnOperationMetadata {
op_type: ColumnOperationType::AddValue,
len: symbol_len,
}
}
};
minibuf.bytes[0] = column_op_metadata.to_code();
// +1 for the metadata
minibuf.len = 1 + column_op_metadata.len;
minibuf
}
/// Deserialize a colummn operation.
/// Returns None if the buffer is empty.
///
/// Panics if the payload is invalid:
/// this deserialize method is meant to target in memory.
pub(super) fn deserialize(bytes: &mut &[u8]) -> Option<Self> {
let column_op_metadata_byte = pop_first_byte(bytes)?;
let column_op_metadata = ColumnOperationMetadata::try_from_code(column_op_metadata_byte)
.expect("Invalid op metadata byte");
let symbol_bytes: &[u8];
(symbol_bytes, *bytes) = bytes.split_at(column_op_metadata.len as usize);
match column_op_metadata.op_type {
ColumnOperationType::NewDoc => {
let new_doc = u32::deserialize(symbol_bytes);
Some(ColumnOperation::NewDoc(new_doc))
}
ColumnOperationType::AddValue => {
let value = V::deserialize(symbol_bytes);
Some(ColumnOperation::Value(value))
}
}
}
}
impl<T> From<T> for ColumnOperation<T> {
fn from(value: T) -> Self {
ColumnOperation::Value(value)
}
}
// Serialization trait very local to the writer.
// As we write fast fields, we accumulate them in "in memory".
// In order to limit memory usage, and in order
// to benefit from the stacker, we do this by serialization our data
// as "Symbols".
#[allow(clippy::from_over_into)]
pub(super) trait SymbolValue: Clone + Copy {
// Serializes the symbol into the given buffer.
// Returns the number of bytes written into the buffer.
/// # Panics
/// May not exceed 9bytes
fn serialize(self, buffer: &mut [u8]) -> u8;
// Panics if invalid
fn deserialize(bytes: &[u8]) -> Self;
}
impl SymbolValue for bool {
fn serialize(self, buffer: &mut [u8]) -> u8 {
buffer[0] = u8::from(self);
1u8
}
fn deserialize(bytes: &[u8]) -> Self {
bytes[0] == 1u8
}
}
#[derive(Default)]
struct MiniBuffer {
pub bytes: [u8; 10],
pub len: u8,
}
impl AsRef<[u8]> for MiniBuffer {
fn as_ref(&self) -> &[u8] {
&self.bytes[..self.len as usize]
}
}
impl SymbolValue for NumericalValue {
fn deserialize(mut bytes: &[u8]) -> Self {
let type_code = pop_first_byte(&mut bytes).unwrap();
let symbol_type = NumericalType::try_from_code(type_code).unwrap();
let mut octet: [u8; 8] = [0u8; 8];
octet[..bytes.len()].copy_from_slice(bytes);
match symbol_type {
NumericalType::U64 => {
let val: u64 = u64::from_le_bytes(octet);
NumericalValue::U64(val)
}
NumericalType::I64 => {
let encoded: u64 = u64::from_le_bytes(octet);
let val: i64 = decode_zig_zag(encoded);
NumericalValue::I64(val)
}
NumericalType::F64 => {
debug_assert_eq!(bytes.len(), 8);
let val: f64 = f64::from_le_bytes(octet);
NumericalValue::F64(val)
}
}
}
/// F64: Serialize with a fixed size of 9 bytes
/// U64: Serialize without leading zeroes
/// I64: ZigZag encoded and serialize without leading zeroes
fn serialize(self, output: &mut [u8]) -> u8 {
match self {
NumericalValue::F64(val) => {
output[0] = NumericalType::F64 as u8;
output[1..9].copy_from_slice(&val.to_le_bytes());
9u8
}
NumericalValue::U64(val) => {
let len = compute_num_bytes_for_u64(val) as u8;
output[0] = NumericalType::U64 as u8;
output[1..9].copy_from_slice(&val.to_le_bytes());
len + 1u8
}
NumericalValue::I64(val) => {
let zig_zag_encoded = encode_zig_zag(val);
let len = compute_num_bytes_for_u64(zig_zag_encoded) as u8;
output[0] = NumericalType::I64 as u8;
output[1..9].copy_from_slice(&zig_zag_encoded.to_le_bytes());
len + 1u8
}
}
}
}
impl SymbolValue for u32 {
fn serialize(self, output: &mut [u8]) -> u8 {
let len = compute_num_bytes_for_u64(self as u64);
output[0..4].copy_from_slice(&self.to_le_bytes());
len as u8
}
fn deserialize(bytes: &[u8]) -> Self {
let mut quartet: [u8; 4] = [0u8; 4];
quartet[..bytes.len()].copy_from_slice(bytes);
u32::from_le_bytes(quartet)
}
}
impl SymbolValue for UnorderedId {
fn serialize(self, output: &mut [u8]) -> u8 {
self.0.serialize(output)
}
fn deserialize(bytes: &[u8]) -> Self {
UnorderedId(u32::deserialize(bytes))
}
}
fn compute_num_bytes_for_u64(val: u64) -> usize {
let msb = (64u32 - val.leading_zeros()) as usize;
(msb + 7) / 8
}
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::*;
#[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);
}
}
#[test]
fn test_column_op_metadata_byte_serialization() {
for len in 0..=15 {
for op_type in [ColumnOperationType::AddValue, ColumnOperationType::NewDoc] {
let column_op_metadata = ColumnOperationMetadata { op_type, len };
let column_op_metadata_code = column_op_metadata.to_code();
let serdeser_metadata =
ColumnOperationMetadata::try_from_code(column_op_metadata_code).unwrap();
assert_eq!(column_op_metadata, serdeser_metadata);
}
}
}
#[track_caller]
fn ser_deser_symbol(column_op: ColumnOperation<NumericalValue>) {
let buf = column_op.serialize();
let mut buffer = buf.as_ref().to_vec();
buffer.extend_from_slice(b"234234");
let mut bytes = &buffer[..];
let serdeser_symbol = ColumnOperation::deserialize(&mut bytes).unwrap();
assert_eq!(bytes.len() + buf.as_ref().len() as usize, buffer.len());
assert_eq!(column_op, 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)));
}
fn test_column_operation_unordered_aux(val: u32, expected_len: usize) {
let column_op = ColumnOperation::Value(UnorderedId(val));
let minibuf = column_op.serialize();
assert_eq!(minibuf.as_ref().len() as usize, expected_len);
let mut buf = minibuf.as_ref().to_vec();
buf.extend_from_slice(&[2, 2, 2, 2, 2, 2]);
let mut cursor = &buf[..];
let column_op_serdeser: ColumnOperation<UnorderedId> =
ColumnOperation::deserialize(&mut cursor).unwrap();
assert_eq!(column_op_serdeser, ColumnOperation::Value(UnorderedId(val)));
assert_eq!(cursor.len() + expected_len, buf.len());
}
#[test]
fn test_column_operation_unordered() {
test_column_operation_unordered_aux(300u32, 3);
test_column_operation_unordered_aux(1u32, 2);
test_column_operation_unordered_aux(0u32, 1);
}
}

265
columnar/src/columnar/writer/column_writers.rs Normal file
View File

@@ -0,0 +1,265 @@
use std::cmp::Ordering;
use stacker::{ExpUnrolledLinkedList, MemoryArena};
use crate::columnar::writer::column_operation::{ColumnOperation, SymbolValue};
use crate::dictionary::{DictionaryBuilder, UnorderedId};
use crate::{Cardinality, NumericalType, NumericalValue, RowId};
#[derive(Copy, Clone, Debug, Eq, PartialEq)]
#[repr(u8)]
enum DocumentStep {
Same = 0,
Next = 1,
Skipped = 2,
}
#[inline(always)]
fn delta_with_last_doc(last_doc_opt: Option<u32>, doc: u32) -> DocumentStep {
let expected_next_doc = last_doc_opt.map(|last_doc| last_doc + 1).unwrap_or(0u32);
match doc.cmp(&expected_next_doc) {
Ordering::Less => DocumentStep::Same,
Ordering::Equal => DocumentStep::Next,
Ordering::Greater => DocumentStep::Skipped,
}
}
#[derive(Copy, Clone, Default)]
pub 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,
}
impl ColumnWriter {
/// Returns an iterator over the Symbol that have been recorded
/// for the given column.
pub(super) fn operation_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 || ColumnOperation::deserialize(&mut cursor))
}
/// Records a change of the document being recorded.
///
/// This function will also update the cardinality of the column
/// if necessary.
pub(super) fn record<S: SymbolValue>(&mut self, doc: RowId, value: S, arena: &mut MemoryArena) {
// Difference between `doc` and the last doc.
match delta_with_last_doc(self.last_doc_opt, doc) {
DocumentStep::Same => {
// This is the last encounterred document.
self.cardinality = Cardinality::Multivalued;
}
DocumentStep::Next => {
self.last_doc_opt = Some(doc);
self.write_symbol::<S>(ColumnOperation::NewDoc(doc), arena);
}
DocumentStep::Skipped => {
self.cardinality = self.cardinality.max(Cardinality::Optional);
self.last_doc_opt = Some(doc);
self.write_symbol::<S>(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.
pub(crate) fn get_cardinality(&self, num_docs: RowId) -> Cardinality {
match delta_with_last_doc(self.last_doc_opt, num_docs) {
DocumentStep::Same | DocumentStep::Next => self.cardinality,
DocumentStep::Skipped => self.cardinality.max(Cardinality::Optional),
}
}
/// Appends a new symbol to the `ColumnWriter`.
fn write_symbol<V: SymbolValue>(
&mut self,
column_operation: ColumnOperation<V>,
arena: &mut MemoryArena,
) {
self.values
.writer(arena)
.extend_from_slice(column_operation.serialize().as_ref());
}
}
#[derive(Clone, Copy, Default)]
pub(crate) struct NumericalColumnWriter {
compatible_numerical_types: CompatibleNumericalTypes,
column_writer: ColumnWriter,
}
/// State used to store what types are still acceptable
/// after having seen a set of numerical values.
#[derive(Clone, Copy)]
struct CompatibleNumericalTypes {
all_values_within_i64_range: bool,
all_values_within_u64_range: bool,
// f64 is always acceptable.
}
impl Default for CompatibleNumericalTypes {
fn default() -> CompatibleNumericalTypes {
CompatibleNumericalTypes {
all_values_within_i64_range: true,
all_values_within_u64_range: true,
}
}
}
impl CompatibleNumericalTypes {
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 column_type_and_cardinality(&self, num_docs: RowId) -> (NumericalType, Cardinality) {
let numerical_type = self.compatible_numerical_types.to_numerical_type();
let cardinality = self.column_writer.get_cardinality(num_docs);
(numerical_type, cardinality)
}
pub fn record_numerical_value(
&mut self,
doc: RowId,
value: NumericalValue,
arena: &mut MemoryArena,
) {
self.compatible_numerical_types.accept_value(value);
self.column_writer.record(doc, value, arena);
}
pub(super) fn operation_iterator<'a>(
self,
arena: &MemoryArena,
buffer: &'a mut Vec<u8>,
) -> impl Iterator<Item = ColumnOperation<NumericalValue>> + 'a {
self.column_writer.operation_iterator(arena, buffer)
}
}
#[derive(Copy, Clone, Default)]
pub(crate) struct StrColumnWriter {
pub(crate) dictionary_id: u32,
pub(crate) column_writer: ColumnWriter,
}
impl StrColumnWriter {
pub(crate) fn with_dictionary_id(dictionary_id: u32) -> StrColumnWriter {
StrColumnWriter {
dictionary_id,
column_writer: Default::default(),
}
}
pub(crate) fn record_bytes(
&mut self,
doc: RowId,
bytes: &[u8],
dictionaries: &mut [DictionaryBuilder],
arena: &mut MemoryArena,
) {
let unordered_id = dictionaries[self.dictionary_id as usize].get_or_allocate_id(bytes);
self.column_writer.record(doc, unordered_id, arena);
}
pub(super) fn operation_iterator<'a>(
&self,
arena: &MemoryArena,
byte_buffer: &'a mut Vec<u8>,
) -> impl Iterator<Item = ColumnOperation<UnorderedId>> + 'a {
self.column_writer.operation_iterator(arena, byte_buffer)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_delta_with_last_doc() {
assert_eq!(delta_with_last_doc(None, 0u32), DocumentStep::Next);
assert_eq!(delta_with_last_doc(None, 1u32), DocumentStep::Skipped);
assert_eq!(delta_with_last_doc(None, 2u32), DocumentStep::Skipped);
assert_eq!(delta_with_last_doc(Some(0u32), 0u32), DocumentStep::Same);
assert_eq!(delta_with_last_doc(Some(1u32), 1u32), DocumentStep::Same);
assert_eq!(delta_with_last_doc(Some(1u32), 2u32), DocumentStep::Next);
assert_eq!(delta_with_last_doc(Some(1u32), 3u32), DocumentStep::Skipped);
assert_eq!(delta_with_last_doc(Some(1u32), 4u32), DocumentStep::Skipped);
}
#[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(&[1f64.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);
}
}

543
columnar/src/columnar/writer/mod.rs Normal file
View File

@@ -0,0 +1,543 @@
mod column_operation;
mod column_writers;
mod serializer;
mod value_index;
use std::io;
use column_operation::ColumnOperation;
use common::CountingWriter;
use serializer::ColumnarSerializer;
use stacker::{Addr, ArenaHashMap, MemoryArena};
use crate::column_index::SerializableColumnIndex;
use crate::column_values::{ColumnValues, MonotonicallyMappableToU64, VecColumn};
use crate::columnar::column_type::{ColumnType, ColumnTypeCategory};
use crate::columnar::writer::column_writers::{
ColumnWriter, NumericalColumnWriter, StrColumnWriter,
};
use crate::columnar::writer::value_index::{IndexBuilder, PreallocatedIndexBuilders};
use crate::dictionary::{DictionaryBuilder, TermIdMapping, UnorderedId};
use crate::value::{Coerce, NumericalType, NumericalValue};
use crate::{Cardinality, RowId};
/// This is a set of buffers that are used to temporarily write the values into before passing them
/// to the fast field codecs.
#[derive(Default)]
struct SpareBuffers {
value_index_builders: PreallocatedIndexBuilders,
i64_values: Vec<i64>,
u64_values: Vec<u64>,
f64_values: Vec<f64>,
bool_values: Vec<bool>,
}
/// Makes it possible to create a new columnar.
///
/// ```rust
/// use tantivy_columnar::ColumnarWriter;
///
/// let mut columnar_writer = ColumnarWriter::default();
/// columnar_writer.record_str(0u32 /* doc id */, "product_name", "Red backpack");
/// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10u64);
/// columnar_writer.record_str(1u32 /* doc id */, "product_name", "Apple");
/// columnar_writer.record_numerical(0u32 /* doc id */, "price", 10.5f64); //< uh oh we ended up mixing integer and floats.
/// let mut wrt: Vec<u8> = Vec::new();
/// columnar_writer.serialize(2u32, &mut wrt).unwrap();
/// ```
pub struct ColumnarWriter {
numerical_field_hash_map: ArenaHashMap,
bool_field_hash_map: ArenaHashMap,
bytes_field_hash_map: ArenaHashMap,
arena: MemoryArena,
// Dictionaries used to store dictionary-encoded values.
dictionaries: Vec<DictionaryBuilder>,
buffers: SpareBuffers,
}
impl Default for ColumnarWriter {
fn default() -> Self {
ColumnarWriter {
numerical_field_hash_map: ArenaHashMap::new(10_000),
bool_field_hash_map: ArenaHashMap::new(10_000),
bytes_field_hash_map: ArenaHashMap::new(10_000),
dictionaries: Vec::new(),
arena: MemoryArena::default(),
buffers: SpareBuffers::default(),
}
}
}
impl ColumnarWriter {
pub fn record_numerical<T: Into<NumericalValue> + Copy>(
&mut self,
doc: RowId,
column_name: &str,
numerical_value: T,
) {
assert!(
!column_name.as_bytes().contains(&0u8),
"key may not contain the 0 byte"
);
let (hash_map, arena) = (&mut self.numerical_field_hash_map, &mut self.arena);
hash_map.mutate_or_create(
column_name.as_bytes(),
|column_opt: Option<NumericalColumnWriter>| {
let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
column.record_numerical_value(doc, numerical_value.into(), arena);
column
},
);
}
pub fn record_bool(&mut self, doc: RowId, column_name: &str, val: bool) {
assert!(
!column_name.as_bytes().contains(&0u8),
"key may not contain the 0 byte"
);
let (hash_map, arena) = (&mut self.bool_field_hash_map, &mut self.arena);
hash_map.mutate_or_create(
column_name.as_bytes(),
|column_opt: Option<ColumnWriter>| {
let mut column: ColumnWriter = column_opt.unwrap_or_default();
column.record(doc, val, arena);
column
},
);
}
pub fn record_str(&mut self, doc: RowId, column_name: &str, value: &str) {
assert!(
!column_name.as_bytes().contains(&0u8),
"key may not contain the 0 byte"
);
let (hash_map, arena, dictionaries) = (
&mut self.bytes_field_hash_map,
&mut self.arena,
&mut self.dictionaries,
);
hash_map.mutate_or_create(
column_name.as_bytes(),
|column_opt: Option<StrColumnWriter>| {
let mut column: StrColumnWriter = column_opt.unwrap_or_else(|| {
// Each column has its own dictionary
let dictionary_id = dictionaries.len() as u32;
dictionaries.push(DictionaryBuilder::default());
StrColumnWriter::with_dictionary_id(dictionary_id)
});
column.record_bytes(doc, value.as_bytes(), dictionaries, arena);
column
},
);
}
pub fn serialize(&mut self, num_docs: RowId, wrt: &mut dyn io::Write) -> io::Result<()> {
let mut serializer = ColumnarSerializer::new(wrt);
let mut field_columns: Vec<(&[u8], ColumnTypeCategory, Addr)> = self
.numerical_field_hash_map
.iter()
.map(|(term, addr, _)| (term, ColumnTypeCategory::Numerical, addr))
.collect();
field_columns.extend(
self.bytes_field_hash_map
.iter()
.map(|(term, addr, _)| (term, ColumnTypeCategory::Str, addr)),
);
field_columns.extend(
self.bool_field_hash_map
.iter()
.map(|(term, addr, _)| (term, ColumnTypeCategory::Bool, addr)),
);
field_columns.sort_unstable_by_key(|(column_name, col_type, _)| (*column_name, *col_type));
let (arena, buffers, dictionaries) = (&self.arena, &mut self.buffers, &self.dictionaries);
let mut symbol_byte_buffer: Vec<u8> = Vec::new();
for (column_name, bytes_or_numerical, addr) in field_columns {
match bytes_or_numerical {
ColumnTypeCategory::Bool => {
let column_writer: ColumnWriter = self.bool_field_hash_map.read(addr);
let cardinality = column_writer.get_cardinality(num_docs);
let mut column_serializer =
serializer.serialize_column(column_name, ColumnType::Bool);
serialize_bool_column(
cardinality,
num_docs,
column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
buffers,
&mut column_serializer,
)?;
}
ColumnTypeCategory::Str => {
let str_column_writer: StrColumnWriter = self.bytes_field_hash_map.read(addr);
let dictionary_builder =
&dictionaries[str_column_writer.dictionary_id as usize];
let cardinality = str_column_writer.column_writer.get_cardinality(num_docs);
let mut column_serializer =
serializer.serialize_column(column_name, ColumnType::Bytes);
serialize_bytes_column(
cardinality,
num_docs,
dictionary_builder,
str_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
buffers,
&mut column_serializer,
)?;
}
ColumnTypeCategory::Numerical => {
let numerical_column_writer: NumericalColumnWriter =
self.numerical_field_hash_map.read(addr);
let (numerical_type, cardinality) =
numerical_column_writer.column_type_and_cardinality(num_docs);
let mut column_serializer = serializer
.serialize_column(column_name, ColumnType::Numerical(numerical_type));
serialize_numerical_column(
cardinality,
num_docs,
numerical_type,
numerical_column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
buffers,
&mut column_serializer,
)?;
}
};
}
serializer.finalize()?;
Ok(())
}
}
fn serialize_bytes_column(
cardinality: Cardinality,
num_docs: RowId,
dictionary_builder: &DictionaryBuilder,
operation_it: impl Iterator<Item = ColumnOperation<UnorderedId>>,
buffers: &mut SpareBuffers,
wrt: impl io::Write,
) -> io::Result<()> {
let SpareBuffers {
value_index_builders,
u64_values,
..
} = buffers;
let mut counting_writer = CountingWriter::wrap(wrt);
let term_id_mapping: TermIdMapping = dictionary_builder.serialize(&mut counting_writer)?;
let dictionary_num_bytes: u32 = counting_writer.written_bytes() as u32;
let mut wrt = counting_writer.finish();
let operation_iterator = operation_it.map(|symbol: ColumnOperation<UnorderedId>| {
// We map unordered ids to ordered ids.
match symbol {
ColumnOperation::Value(unordered_id) => {
let ordered_id = term_id_mapping.to_ord(unordered_id);
ColumnOperation::Value(ordered_id.0 as u64)
}
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
}
});
serialize_column(
operation_iterator,
cardinality,
num_docs,
value_index_builders,
u64_values,
&mut wrt,
)?;
wrt.write_all(&dictionary_num_bytes.to_le_bytes()[..])?;
Ok(())
}
fn serialize_numerical_column(
cardinality: Cardinality,
num_docs: RowId,
numerical_type: NumericalType,
op_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
buffers: &mut SpareBuffers,
wrt: &mut impl io::Write,
) -> io::Result<()> {
let SpareBuffers {
value_index_builders,
u64_values,
i64_values,
f64_values,
..
} = buffers;
match numerical_type {
NumericalType::I64 => {
serialize_column(
coerce_numerical_symbol::<i64>(op_iterator),
cardinality,
num_docs,
value_index_builders,
i64_values,
wrt,
)?;
}
NumericalType::U64 => {
serialize_column(
coerce_numerical_symbol::<u64>(op_iterator),
cardinality,
num_docs,
value_index_builders,
u64_values,
wrt,
)?;
}
NumericalType::F64 => {
serialize_column(
coerce_numerical_symbol::<f64>(op_iterator),
cardinality,
num_docs,
value_index_builders,
f64_values,
wrt,
)?;
}
};
Ok(())
}
fn serialize_bool_column(
cardinality: Cardinality,
num_docs: RowId,
column_operations_it: impl Iterator<Item = ColumnOperation<bool>>,
buffers: &mut SpareBuffers,
wrt: &mut impl io::Write,
) -> io::Result<()> {
let SpareBuffers {
value_index_builders,
bool_values,
..
} = buffers;
serialize_column(
column_operations_it,
cardinality,
num_docs,
value_index_builders,
bool_values,
wrt,
)?;
Ok(())
}
fn serialize_column<
T: Copy + Default + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU64 + PartialOrd,
>(
op_iterator: impl Iterator<Item = ColumnOperation<T>>,
cardinality: Cardinality,
num_docs: RowId,
value_index_builders: &mut PreallocatedIndexBuilders,
values: &mut Vec<T>,
mut wrt: impl io::Write,
) -> io::Result<()>
where
for<'a> VecColumn<'a, T>: ColumnValues<T>,
{
values.clear();
let serializable_column_index = match cardinality {
Cardinality::Full => {
consume_operation_iterator(
op_iterator,
value_index_builders.borrow_required_index_builder(),
values,
);
SerializableColumnIndex::Full
}
Cardinality::Optional => {
let optional_index_builder = value_index_builders.borrow_optional_index_builder();
consume_operation_iterator(op_iterator, optional_index_builder, values);
let optional_index = optional_index_builder.finish(num_docs);
SerializableColumnIndex::Optional(Box::new(optional_index))
}
Cardinality::Multivalued => {
let multivalued_index_builder = value_index_builders.borrow_multivalued_index_builder();
consume_operation_iterator(op_iterator, multivalued_index_builder, values);
let multivalued_index = multivalued_index_builder.finish(num_docs);
todo!();
// SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
}
};
crate::column::serialize_column_u64(
serializable_column_index,
&VecColumn::from(&values[..]),
&mut wrt,
)?;
Ok(())
}
fn coerce_numerical_symbol<T>(
operation_iterator: impl Iterator<Item = ColumnOperation<NumericalValue>>,
) -> impl Iterator<Item = ColumnOperation<T>>
where T: Coerce {
operation_iterator.map(|symbol| match symbol {
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
ColumnOperation::Value(numerical_value) => {
ColumnOperation::Value(Coerce::coerce(numerical_value))
}
})
}
fn consume_operation_iterator<T: std::fmt::Debug, TIndexBuilder: IndexBuilder>(
operation_iterator: impl Iterator<Item = ColumnOperation<T>>,
index_builder: &mut TIndexBuilder,
values: &mut Vec<T>,
) {
for symbol in operation_iterator {
match symbol {
ColumnOperation::NewDoc(doc) => {
index_builder.record_row(doc);
}
ColumnOperation::Value(value) => {
index_builder.record_value();
values.push(value);
}
}
}
}
// /// Serializes the column with the codec with the best estimate on the data.
// fn serialize_numerical<T: MonotonicallyMappableToU64>(
// value_index: ValueIndexInfo,
// typed_column: impl Column<T>,
// output: &mut impl io::Write,
// codecs: &[FastFieldCodecType],
// ) -> io::Result<()> {
// let counting_writer = CountingWriter::wrap(output);
// serialize_value_index(value_index, output)?;
// let value_index_len = counting_writer.written_bytes();
// let output = counting_writer.finish();
// serialize_column(value_index, output)?;
// let column = monotonic_map_column(
// typed_column,
// crate::column::monotonic_mapping::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 column_header = ColumnFooter {
// value_index_len: todo!(),
// cardinality: todo!(),
// };
// 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)?;
// Ok(())
// }
#[cfg(test)]
mod tests {
use column_operation::ColumnOperation;
use stacker::MemoryArena;
use super::*;
use crate::value::NumericalValue;
#[test]
fn test_column_writer_required_simple() {
let mut arena = MemoryArena::default();
let mut column_writer = super::ColumnWriter::default();
column_writer.record(0u32, NumericalValue::from(14i64), &mut arena);
column_writer.record(1u32, NumericalValue::from(15i64), &mut arena);
column_writer.record(2u32, NumericalValue::from(-16i64), &mut arena);
assert_eq!(column_writer.get_cardinality(3), Cardinality::Full);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.operation_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, NumericalValue::from(15i64), &mut arena);
column_writer.record(2u32, NumericalValue::from(-16i64), &mut arena);
assert_eq!(column_writer.get_cardinality(3), Cardinality::Optional);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.operation_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, NumericalValue::from(15i64), &mut arena);
assert_eq!(column_writer.get_cardinality(2), Cardinality::Optional);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.operation_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, NumericalValue::from(16i64), &mut arena);
column_writer.record(0u32, NumericalValue::from(17i64), &mut arena);
assert_eq!(column_writer.get_cardinality(1), Cardinality::Multivalued);
let mut buffer = Vec::new();
let symbols: Vec<ColumnOperation<NumericalValue>> = column_writer
.operation_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))
));
}
}

106
columnar/src/columnar/writer/serializer.rs Normal file
View File

@@ -0,0 +1,106 @@
use std::io;
use std::io::Write;
use common::CountingWriter;
use sstable::value::RangeValueWriter;
use sstable::RangeSSTable;
use crate::columnar::ColumnType;
pub struct ColumnarSerializer<W: io::Write> {
wrt: CountingWriter<W>,
sstable_range: sstable::Writer<Vec<u8>, RangeValueWriter>,
prepare_key_buffer: Vec<u8>,
}
/// Returns a key consisting of the concatenation of the key and the column_type_and_cardinality
/// code.
fn prepare_key(key: &[u8], column_type: ColumnType, buffer: &mut Vec<u8>) {
buffer.clear();
buffer.extend_from_slice(key);
buffer.push(0u8);
buffer.push(column_type.to_code());
}
impl<W: io::Write> ColumnarSerializer<W> {
pub(crate) fn new(wrt: W) -> ColumnarSerializer<W> {
let sstable_range: sstable::Writer<Vec<u8>, RangeValueWriter> =
sstable::Dictionary::<RangeSSTable>::builder(Vec::with_capacity(100_000)).unwrap();
ColumnarSerializer {
wrt: CountingWriter::wrap(wrt),
sstable_range,
prepare_key_buffer: Vec::new(),
}
}
pub fn serialize_column<'a>(
&'a mut self,
column_name: &[u8],
column_type: ColumnType,
) -> impl io::Write + 'a {
let start_offset = self.wrt.written_bytes();
prepare_key(column_name, column_type, &mut self.prepare_key_buffer);
ColumnSerializer {
columnar_serializer: self,
start_offset,
}
}
pub(crate) 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()[..])?;
self.wrt
.write_all(&super::super::format_version::footer())?;
self.wrt.flush()?;
Ok(())
}
}
struct ColumnSerializer<'a, W: io::Write> {
columnar_serializer: &'a mut ColumnarSerializer<W>,
start_offset: u64,
}
impl<'a, W: io::Write> Drop for ColumnSerializer<'a, W> {
fn drop(&mut self) {
let end_offset: u64 = self.columnar_serializer.wrt.written_bytes();
let byte_range = self.start_offset..end_offset;
self.columnar_serializer.sstable_range.insert_cannot_fail(
&self.columnar_serializer.prepare_key_buffer[..],
&byte_range,
);
self.columnar_serializer.prepare_key_buffer.clear();
}
}
impl<'a, W: io::Write> io::Write for ColumnSerializer<'a, W> {
fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
self.columnar_serializer.wrt.write(buf)
}
fn flush(&mut self) -> io::Result<()> {
self.columnar_serializer.wrt.flush()
}
fn write_all(&mut self, buf: &[u8]) -> io::Result<()> {
self.columnar_serializer.wrt.write_all(buf)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::columnar::column_type::ColumnType;
#[test]
fn test_prepare_key_bytes() {
let mut buffer: Vec<u8> = b"somegarbage".to_vec();
prepare_key(b"root\0child", ColumnType::Bytes, &mut buffer);
assert_eq!(buffer.len(), 12);
assert_eq!(&buffer[..10], b"root\0child");
assert_eq!(buffer[10], 0u8);
assert_eq!(buffer[11], ColumnType::Bytes.to_code());
}
}

205
columnar/src/columnar/writer/value_index.rs Normal file
View File

@@ -0,0 +1,205 @@
use crate::column_index::SerializableOptionalIndex;
use crate::column_values::{ColumnValues, VecColumn};
use crate::RowId;
/// The `IndexBuilder` interprets a sequence of
/// calls of the form:
/// (record_doc,record_value+)*
/// and can then serialize the results into an index to associate docids with their value[s].
///
/// It has different implementation depending on whether the
/// cardinality is required, optional, or multivalued.
pub(crate) trait IndexBuilder {
fn record_row(&mut self, doc: RowId);
#[inline]
fn record_value(&mut self) {}
}
/// The FullIndexBuilder does nothing.
#[derive(Default)]
pub struct FullIndexBuilder;
impl IndexBuilder for FullIndexBuilder {
#[inline(always)]
fn record_row(&mut self, _doc: RowId) {}
}
#[derive(Default)]
pub struct OptionalIndexBuilder {
docs: Vec<RowId>,
}
struct SingleValueArrayIndex<'a> {
// RowIds with a value, in a strictly increasing order
row_ids: &'a [RowId],
num_rows: RowId,
}
impl<'a> SerializableOptionalIndex<'a> for SingleValueArrayIndex<'a> {
fn num_rows(&self) -> RowId {
self.num_rows
}
fn non_null_rows(&self) -> Box<dyn Iterator<Item = RowId> + 'a> {
Box::new(self.row_ids.iter().copied())
}
}
impl OptionalIndexBuilder {
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<'a>(&'a mut self, num_rows: RowId) -> impl SerializableOptionalIndex + 'a {
debug_assert!(self
.docs
.last()
.copied()
.map(|last_doc| last_doc < num_rows)
.unwrap_or(true));
SingleValueArrayIndex {
row_ids: &self.docs[..],
num_rows,
}
}
fn reset(&mut self) {
self.docs.clear();
}
}
impl IndexBuilder for OptionalIndexBuilder {
#[inline(always)]
fn record_row(&mut self, doc: RowId) {
debug_assert!(self
.docs
.last()
.copied()
.map(|prev_doc| doc > prev_doc)
.unwrap_or(true));
self.docs.push(doc);
}
}
#[derive(Default)]
pub struct MultivaluedIndexBuilder {
start_offsets: Vec<RowId>,
total_num_vals_seen: u32,
}
impl MultivaluedIndexBuilder {
pub fn finish(&mut self, num_docs: RowId) -> impl ColumnValues<u32> + '_ {
self.start_offsets
.resize(num_docs as usize, self.total_num_vals_seen);
VecColumn {
values: &&self.start_offsets[..],
min_value: 0,
max_value: self.start_offsets.last().copied().unwrap_or(0),
}
}
fn reset(&mut self) {
self.start_offsets.clear();
self.start_offsets.push(0u32);
self.total_num_vals_seen = 0;
}
}
impl IndexBuilder for MultivaluedIndexBuilder {
fn record_row(&mut self, row_id: RowId) {
self.start_offsets
.resize(row_id as usize + 1, 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 PreallocatedIndexBuilders {
required_index_builder: FullIndexBuilder,
optional_index_builder: OptionalIndexBuilder,
multivalued_index_builder: MultivaluedIndexBuilder,
}
impl PreallocatedIndexBuilders {
pub fn borrow_required_index_builder(&mut self) -> &mut FullIndexBuilder {
&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_row(0u32);
opt_value_index_builder.record_value();
assert_eq!(
&opt_value_index_builder
.finish(1u32)
.non_null_rows()
.collect::<Vec<u32>>(),
&[0]
);
opt_value_index_builder.reset();
opt_value_index_builder.record_row(1u32);
opt_value_index_builder.record_value();
assert_eq!(
&opt_value_index_builder
.finish(2u32)
.non_null_rows()
.collect::<Vec<u32>>(),
&[1]
);
}
#[test]
fn test_multivalued_value_index_builder() {
let mut multivalued_value_index_builder = MultivaluedIndexBuilder::default();
multivalued_value_index_builder.record_row(1u32);
multivalued_value_index_builder.record_value();
multivalued_value_index_builder.record_value();
multivalued_value_index_builder.record_row(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_row(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]
);
}
}

84
columnar/src/dictionary.rs Normal file
View File

@@ -0,0 +1,84 @@
use std::io;
use fnv::FnvHashMap;
use sstable::SSTable;
pub(crate) struct TermIdMapping {
unordered_to_ord: Vec<OrderedId>,
}
impl TermIdMapping {
pub fn to_ord(&self, unordered: UnorderedId) -> OrderedId {
self.unordered_to_ord[unordered.0 as usize]
}
}
/// When we add values, we cannot know their ordered id yet.
/// For this reason, we temporarily assign them a `UnorderedId`
/// that will be mapped to an `OrderedId` upon serialization.
#[derive(Clone, Copy, Debug, Hash, PartialEq, Eq)]
pub struct UnorderedId(pub u32);
#[derive(Clone, Copy, Hash, PartialEq, Eq, Debug)]
pub struct OrderedId(pub u32);
/// `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(crate) struct DictionaryBuilder {
dict: FnvHashMap<Vec<u8>, UnorderedId>,
}
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<TermIdMapping> {
let mut terms: Vec<(&[u8], UnorderedId)> =
self.dict.iter().map(|(k, v)| (k.as_slice(), *v)).collect();
terms.sort_unstable_by_key(|(key, _)| *key);
// TODO Remove the allocation.
let mut unordered_to_ord: Vec<OrderedId> = vec![OrderedId(0u32); terms.len()];
let mut sstable_builder = sstable::VoidSSTable::writer(wrt);
for (ord, (key, unordered_id)) in terms.into_iter().enumerate() {
let ordered_id = OrderedId(ord as u32);
sstable_builder.insert(key, &())?;
unordered_to_ord[unordered_id.0 as usize] = ordered_id;
}
sstable_builder.finish()?;
Ok(TermIdMapping { unordered_to_ord })
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_dictionary_builder() {
let mut dictionary_builder = DictionaryBuilder::default();
let hello_uid = dictionary_builder.get_or_allocate_id(b"hello");
let happy_uid = dictionary_builder.get_or_allocate_id(b"happy");
let tax_uid = dictionary_builder.get_or_allocate_id(b"tax");
let mut buffer = Vec::new();
let id_mapping = dictionary_builder.serialize(&mut buffer).unwrap();
assert_eq!(id_mapping.to_ord(hello_uid), OrderedId(1));
assert_eq!(id_mapping.to_ord(happy_uid), OrderedId(0));
assert_eq!(id_mapping.to_ord(tax_uid), OrderedId(2));
}
}

95
columnar/src/dynamic_column.rs Normal file
View File

@@ -0,0 +1,95 @@
use std::io;
use std::net::IpAddr;
use common::file_slice::FileSlice;
use common::{HasLen, OwnedBytes};
use crate::column::{BytesColumn, Column};
use crate::columnar::ColumnType;
use crate::DateTime;
#[derive(Clone)]
pub enum DynamicColumn {
Bool(Column<bool>),
I64(Column<i64>),
U64(Column<u64>),
F64(Column<f64>),
IpAddr(Column<IpAddr>),
DateTime(Column<DateTime>),
Str(BytesColumn),
}
impl From<Column<i64>> for DynamicColumn {
fn from(column_i64: Column<i64>) -> Self {
DynamicColumn::I64(column_i64)
}
}
impl From<Column<u64>> for DynamicColumn {
fn from(column_u64: Column<u64>) -> Self {
DynamicColumn::U64(column_u64)
}
}
impl From<Column<f64>> for DynamicColumn {
fn from(column_f64: Column<f64>) -> Self {
DynamicColumn::F64(column_f64)
}
}
impl From<Column<bool>> for DynamicColumn {
fn from(bool_column: Column<bool>) -> Self {
DynamicColumn::Bool(bool_column)
}
}
impl From<BytesColumn> for DynamicColumn {
fn from(dictionary_encoded_col: BytesColumn) -> Self {
DynamicColumn::Str(dictionary_encoded_col)
}
}
#[derive(Clone)]
pub struct DynamicColumnHandle {
pub(crate) file_slice: FileSlice,
pub(crate) column_type: ColumnType,
}
impl DynamicColumnHandle {
pub fn open(&self) -> io::Result<DynamicColumn> {
let column_bytes: OwnedBytes = self.file_slice.read_bytes()?;
self.open_internal(column_bytes)
}
pub async fn open_async(&self) -> io::Result<DynamicColumn> {
let column_bytes: OwnedBytes = self.file_slice.read_bytes_async().await?;
self.open_internal(column_bytes)
}
fn open_internal(&self, column_bytes: OwnedBytes) -> io::Result<DynamicColumn> {
let dynamic_column: DynamicColumn = match self.column_type {
ColumnType::Bytes => crate::column::open_column_bytes(column_bytes)?.into(),
ColumnType::Numerical(numerical_type) => match numerical_type {
crate::NumericalType::I64 => {
crate::column::open_column_u64::<i64>(column_bytes)?.into()
}
crate::NumericalType::U64 => {
crate::column::open_column_u64::<u64>(column_bytes)?.into()
}
crate::NumericalType::F64 => {
crate::column::open_column_u64::<f64>(column_bytes)?.into()
}
},
ColumnType::Bool => crate::column::open_column_u64::<bool>(column_bytes)?.into(),
};
Ok(dynamic_column)
}
pub fn num_bytes(&self) -> usize {
self.file_slice.len()
}
pub fn column_type(&self) -> ColumnType {
self.column_type
}
}

75
columnar/src/lib.rs Normal file
View File

@@ -0,0 +1,75 @@
#![cfg_attr(all(feature = "unstable", test), feature(test))]
#[cfg(test)]
#[macro_use]
extern crate more_asserts;
#[cfg(all(test, feature = "unstable"))]
extern crate test;
use std::io;
mod column;
mod column_index;
mod column_values;
mod columnar;
mod dictionary;
mod dynamic_column;
pub(crate) mod utils;
mod value;
pub use columnar::{ColumnarReader, ColumnarWriter};
pub use value::{NumericalType, NumericalValue};
// pub use self::dynamic_column::DynamicColumnHandle;
pub type RowId = u32;
#[derive(Clone, Copy)]
pub struct DateTime {
timestamp_micros: i64,
}
#[derive(Copy, Clone, Debug)]
pub struct InvalidData;
impl From<InvalidData> for io::Error {
fn from(_: InvalidData) -> Self {
io::Error::new(io::ErrorKind::InvalidData, "Invalid data")
}
}
/// Enum describing the number of values that can exist per document
/// (or per row if you will).
///
/// The cardinality must fit on 2 bits.
#[derive(Clone, Copy, Hash, Default, Debug, PartialEq, Eq, PartialOrd, Ord)]
#[repr(u8)]
pub enum Cardinality {
/// All documents contain exactly one value.
/// `Full` is the default for auto-detecting the Cardinality, since it is the most strict.
#[default]
Full = 0,
/// All documents contain at most one value.
Optional = 1,
/// All documents may contain any number of values.
Multivalued = 2,
}
impl Cardinality {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn try_from_code(code: u8) -> Result<Cardinality, InvalidData> {
match code {
0 => Ok(Cardinality::Full),
1 => Ok(Cardinality::Optional),
2 => Ok(Cardinality::Multivalued),
_ => Err(InvalidData),
}
}
}
#[cfg(test)]
mod tests;

84
columnar/src/tests.rs Normal file
View File

@@ -0,0 +1,84 @@
use crate::columnar::ColumnType;
use crate::dynamic_column::{DynamicColumn, DynamicColumnHandle};
use crate::value::NumericalValue;
use crate::{Cardinality, ColumnarReader, ColumnarWriter};
#[test]
fn test_dataframe_writer_bytes() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_str(1u32, "my_string", "hello");
dataframe_writer.record_str(3u32, "my_string", "helloeee");
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(5, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
assert_eq!(columnar.num_columns(), 1);
let cols: Vec<DynamicColumnHandle> = columnar.read_columns("my_string").unwrap();
assert_eq!(cols.len(), 1);
assert_eq!(cols[0].num_bytes(), 165);
}
#[test]
fn test_dataframe_writer_bool() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_bool(1u32, "bool.value", false);
dataframe_writer.record_bool(3u32, "bool.value", true);
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(5, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
assert_eq!(columnar.num_columns(), 1);
let cols: Vec<DynamicColumnHandle> = columnar.read_columns("bool.value").unwrap();
assert_eq!(cols.len(), 1);
assert_eq!(cols[0].num_bytes(), 29);
assert_eq!(cols[0].column_type(), ColumnType::Bool);
let dyn_bool_col = cols[0].open().unwrap();
let DynamicColumn::Bool(bool_col) = dyn_bool_col else { panic!(); };
let vals: Vec<Option<bool>> = (0..5).map(|row_id| bool_col.first(row_id)).collect();
assert_eq!(&vals, &[None, Some(false), None, Some(true), None,]);
}
#[test]
fn test_dataframe_writer_numerical() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(1u32, "srical.value", NumericalValue::U64(12u64));
dataframe_writer.record_numerical(2u32, "srical.value", NumericalValue::U64(13u64));
dataframe_writer.record_numerical(4u32, "srical.value", NumericalValue::U64(15u64));
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(6, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
assert_eq!(columnar.num_columns(), 1);
let cols: Vec<DynamicColumnHandle> = 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
assert_eq!(cols[0].num_bytes(), 40);
let column = cols[0].open().unwrap();
let DynamicColumn::I64(column_i64) = column else { panic!(); };
assert_eq!(column_i64.idx.get_cardinality(), Cardinality::Optional);
assert_eq!(column_i64.first(0), None);
assert_eq!(column_i64.first(1), Some(12i64));
assert_eq!(column_i64.first(2), Some(13i64));
assert_eq!(column_i64.first(3), None);
assert_eq!(column_i64.first(4), Some(15i64));
assert_eq!(column_i64.first(5), None);
assert_eq!(column_i64.first(6), None); //< we can change the spec for that one.
}
#[test]
fn test_dictionary_encoded() {
let mut buffer = Vec::new();
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_str(1, "my.column", "my.key");
columnar_writer.record_str(3, "my.column", "my.key2");
columnar_writer.record_str(3, "my.column2", "different_column!");
columnar_writer.serialize(5, &mut buffer).unwrap();
let columnar_reader = ColumnarReader::open(buffer).unwrap();
assert_eq!(columnar_reader.num_columns(), 2);
let col_handles = columnar_reader.read_columns("my.column").unwrap();
assert_eq!(col_handles.len(), 1);
let DynamicColumn::Str(str_col) = col_handles[0].open().unwrap() else { panic!(); };
assert_eq!(str_col.num_rows(), 5);
// let term_ords = (0..)
}

76
columnar/src/utils.rs Normal file
View File

@@ -0,0 +1,76 @@
const fn compute_mask(num_bits: u8) -> u8 {
if num_bits == 8 {
u8::MAX
} else {
(1u8 << num_bits) - 1
}
}
#[inline(always)]
#[must_use]
pub(crate) 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(always)]
#[must_use]
pub(crate) 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
}
/// Pop-front one bytes from a slice of bytes.
#[inline(always)]
pub fn pop_first_byte(bytes: &mut &[u8]) -> Option<u8> {
if bytes.is_empty() {
return None;
}
let first_byte = bytes[0];
*bytes = &bytes[1..];
Some(first_byte)
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_select_bits() {
assert_eq!(255u8, select_bits::<0, 8>(255u8));
assert_eq!(0u8, select_bits::<0, 0>(255u8));
assert_eq!(8u8, select_bits::<0, 4>(8u8));
assert_eq!(4u8, select_bits::<1, 4>(8u8));
assert_eq!(0u8, select_bits::<1, 3>(8u8));
}
#[test]
fn test_place_bits() {
assert_eq!(255u8, place_bits::<0, 8>(255u8));
assert_eq!(4u8, place_bits::<2, 3>(1u8));
assert_eq!(0u8, place_bits::<2, 2>(0u8));
}
#[test]
#[should_panic]
fn test_place_bits_overflows() {
let _ = place_bits::<1, 4>(8u8);
}
#[test]
fn test_pop_first_byte() {
let mut cursor: &[u8] = &b"abcd"[..];
assert_eq!(pop_first_byte(&mut cursor), Some(b'a'));
assert_eq!(pop_first_byte(&mut cursor), Some(b'b'));
assert_eq!(pop_first_byte(&mut cursor), Some(b'c'));
assert_eq!(pop_first_byte(&mut cursor), Some(b'd'));
assert_eq!(pop_first_byte(&mut cursor), None);
}
}

124
columnar/src/value.rs Normal file
View File

@@ -0,0 +1,124 @@
use crate::InvalidData;
#[derive(Copy, Clone, Debug, PartialEq)]
pub enum NumericalValue {
I64(i64),
U64(u64),
F64(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<f64> for NumericalValue {
fn from(val: 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) -> Result<NumericalType, InvalidData> {
match code {
0 => Ok(NumericalType::I64),
1 => Ok(NumericalType::U64),
2 => Ok(NumericalType::F64),
_ => Err(InvalidData),
}
}
}
/// We voluntarily avoid using `Into` here to keep this
/// implementation quirk as private as possible.
///
/// # Panics
/// 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 f64 {
fn coerce(value: NumericalValue) -> Self {
match value {
NumericalValue::I64(val) => val as f64,
NumericalValue::U64(val) => 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 Ok(numerical_type) = NumericalType::try_from_code(code) {
assert_eq!(numerical_type.to_code(), code);
num_numerical_type += 1;
}
}
assert_eq!(num_numerical_type, 3);
}
}

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"

8
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)
}
@@ -277,7 +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
@@ -421,7 +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);
}
}
@@ -432,7 +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]

106
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 +23,12 @@ 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> {
Err(io::Error::new(
io::ErrorKind::Unsupported,
"Async read is not supported.",
))
}
}
@@ -44,8 +39,7 @@ impl FileHandle for &'static [u8] {
Ok(OwnedBytes::new(bytes))
}
#[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 +67,34 @@ impl fmt::Debug for FileSlice {
}
}
/// Takes a range, a `RangeBounds` object, and returns
/// a `Range` that corresponds to the relative application of the
/// `RangeBounds` object to the original `Range`.
///
/// For instance, combine_ranges(`[2..11)`, `[5..7]`) returns `[7..10]`
/// as it reads, what is the sub-range that starts at the 5 element of
/// `[2..11)` and ends at the 9th element included.
///
/// This function panics, if the result would suggest something outside
/// of the bounds of the original range.
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 +118,11 @@ 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),
}
}
@@ -120,9 +142,8 @@ 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 +161,8 @@ 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"
@@ -207,8 +224,7 @@ impl FileHandle for FileSlice {
self.read_bytes_slice(range)
}
#[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
}
}
@@ -225,21 +241,20 @@ impl FileHandle for OwnedBytes {
Ok(self.slice(range))
}
#[cfg(feature = "quickwit")]
async fn read_bytes_async(&self, range: Range<usize>) -> crate::AsyncIoResult<OwnedBytes> {
let bytes = self.read_bytes(range)?;
Ok(bytes)
async fn read_bytes_async(&self, range: Range<usize>) -> io::Result<OwnedBytes> {
self.read_bytes(range)
}
}
#[cfg(test)]
mod tests {
use std::io;
use std::ops::Bound;
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 +325,23 @@ 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);
assert_eq!(combine_ranges(2..11, 5..=7), 7..10);
assert_eq!(
combine_ranges(2..11, (Bound::Excluded(5), Bound::Unbounded)),
8..11
);
}
#[test]
#[should_panic]
fn test_combine_range_panics() {
let _ = combine_ranges(3..5, 1..4);
}
}

166
common/src/group_by.rs Normal file
View File

@@ -0,0 +1,166 @@
use std::cell::RefCell;
use std::iter::Peekable;
use std::rc::Rc;
pub trait GroupByIteratorExtended: Iterator {
/// Return an `Iterator` that groups iterator elements. Consecutive elements that map to the
/// same key are assigned to the same group.
///
/// The returned Iterator item is `(K, impl Iterator)`, where Iterator are the items of the
/// group.
///
/// ```
/// use tantivy_common::GroupByIteratorExtended;
///
/// // group data into blocks of larger than zero or not.
/// let data: Vec<i32> = vec![1, 3, -2, -2, 1, 0, 1, 2];
/// // groups: |---->|------>|--------->|
///
/// let mut data_grouped = Vec::new();
/// // Note: group is an iterator
/// for (key, group) in data.into_iter().group_by(|val| *val >= 0) {
/// data_grouped.push((key, group.collect()));
/// }
/// assert_eq!(data_grouped, vec![(true, vec![1, 3]), (false, vec![-2, -2]), (true, vec![1, 0, 1, 2])]);
/// ```
fn group_by<K, F>(self, key: F) -> GroupByIterator<Self, F, K>
where
Self: Sized,
F: FnMut(&Self::Item) -> K,
K: PartialEq + Copy,
Self::Item: Copy,
{
GroupByIterator::new(self, key)
}
}
impl<I: Iterator> GroupByIteratorExtended for I {}
pub struct GroupByIterator<I, F, K: Copy>
where
I: Iterator,
F: FnMut(&I::Item) -> K,
{
// I really would like to avoid the Rc<RefCell>, but the Iterator is shared between
// `GroupByIterator` and `GroupIter`. In practice they are used consecutive and
// `GroupByIter` is finished before calling next on `GroupByIterator`. I'm not sure there
// is a solution with lifetimes for that, because we would need to enforce it in the usage
// somehow.
//
// One potential solution would be to replace the iterator approach with something similar.
inner: Rc<RefCell<GroupByShared<I, F, K>>>,
}
struct GroupByShared<I, F, K: Copy>
where
I: Iterator,
F: FnMut(&I::Item) -> K,
{
iter: Peekable<I>,
group_by_fn: F,
}
impl<I, F, K> GroupByIterator<I, F, K>
where
I: Iterator,
F: FnMut(&I::Item) -> K,
K: Copy,
{
fn new(inner: I, group_by_fn: F) -> Self {
let inner = GroupByShared {
iter: inner.peekable(),
group_by_fn,
};
Self {
inner: Rc::new(RefCell::new(inner)),
}
}
}
impl<I, F, K> Iterator for GroupByIterator<I, F, K>
where
I: Iterator,
I::Item: Copy,
F: FnMut(&I::Item) -> K,
K: Copy,
{
type Item = (K, GroupIterator<I, F, K>);
fn next(&mut self) -> Option<Self::Item> {
let mut inner = self.inner.borrow_mut();
let value = *inner.iter.peek()?;
let key = (inner.group_by_fn)(&value);
let inner = self.inner.clone();
let group_iter = GroupIterator {
inner,
group_key: key,
};
Some((key, group_iter))
}
}
pub struct GroupIterator<I, F, K: Copy>
where
I: Iterator,
F: FnMut(&I::Item) -> K,
{
inner: Rc<RefCell<GroupByShared<I, F, K>>>,
group_key: K,
}
impl<I, F, K: PartialEq + Copy> Iterator for GroupIterator<I, F, K>
where
I: Iterator,
I::Item: Copy,
F: FnMut(&I::Item) -> K,
{
type Item = I::Item;
fn next(&mut self) -> Option<Self::Item> {
let mut inner = self.inner.borrow_mut();
// peek if next value is in group
let peek_val = *inner.iter.peek()?;
if (inner.group_by_fn)(&peek_val) == self.group_key {
inner.iter.next()
} else {
None
}
}
}
#[cfg(test)]
mod tests {
use super::*;
fn group_by_collect<I: Iterator<Item = u32>>(iter: I) -> Vec<(I::Item, Vec<I::Item>)> {
iter.group_by(|val| val / 10)
.map(|(el, iter)| (el, iter.collect::<Vec<_>>()))
.collect::<Vec<_>>()
}
#[test]
fn group_by_two_groups() {
let vals = vec![1u32, 4, 15];
let grouped_vals = group_by_collect(vals.into_iter());
assert_eq!(grouped_vals, vec![(0, vec![1, 4]), (1, vec![15])]);
}
#[test]
fn group_by_test_empty() {
let vals = vec![];
let grouped_vals = group_by_collect(vals.into_iter());
assert_eq!(grouped_vals, vec![]);
}
#[test]
fn group_by_three_groups() {
let vals = vec![1u32, 4, 15, 1];
let grouped_vals = group_by_collect(vals.into_iter());
assert_eq!(
grouped_vals,
vec![(0, vec![1, 4]), (1, vec![15]), (0, vec![1])]
);
}
}

5
common/src/lib.rs
View File

@@ -5,11 +5,14 @@ use std::ops::Deref;
pub use byteorder::LittleEndian as Endianness;
mod bitset;
pub mod file_slice;
mod group_by;
mod serialize;
mod vint;
mod writer;
pub use bitset::*;
pub use group_by::GroupByIteratorExtended;
pub use ownedbytes::{OwnedBytes, StableDeref};
pub use serialize::{BinarySerializable, DeserializeFrom, FixedSize};
pub use vint::{
deserialize_vint_u128, read_u32_vint, read_u32_vint_no_advance, serialize_vint_u128,

27
common/src/serialize.rs
View File

@@ -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)

2
common/src/vint.rs
View File

@@ -157,7 +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
}

2
doc/src/basis.md
View File

@@ -50,7 +50,7 @@ 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

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();

14
examples/custom_collector.rs
View File

@@ -14,7 +14,7 @@ use fastfield_codecs::Column;
// Importing tantivy...
use tantivy::collector::{Collector, SegmentCollector};
use tantivy::query::QueryParser;
use tantivy::schema::{Field, Schema, FAST, INDEXED, TEXT};
use tantivy::schema::{Schema, FAST, INDEXED, TEXT};
use tantivy::{doc, Index, Score, SegmentReader};
#[derive(Default)]
@@ -52,11 +52,11 @@ impl Stats {
}
struct StatsCollector {
field: Field,
field: String,
}
impl StatsCollector {
fn with_field(field: Field) -> StatsCollector {
fn with_field(field: String) -> StatsCollector {
StatsCollector { field }
}
}
@@ -73,7 +73,7 @@ impl Collector for StatsCollector {
_segment_local_id: u32,
segment_reader: &SegmentReader,
) -> tantivy::Result<StatsSegmentCollector> {
let fast_field_reader = segment_reader.fast_fields().u64(self.field)?;
let fast_field_reader = segment_reader.fast_fields().u64(&self.field)?;
Ok(StatsSegmentCollector {
fast_field_reader,
stats: Stats::default(),
@@ -105,7 +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_val(doc as u64) 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;
@@ -171,7 +171,9 @@ fn main() -> tantivy::Result<()> {
// here we want to get a hit on the 'ken' in Frankenstein
let query = query_parser.parse_query("broom")?;
if let Some(stats) = searcher.search(&query, &StatsCollector::with_field(price))? {
if let Some(stats) =
searcher.search(&query, &StatsCollector::with_field("price".to_string()))?
{
println!("count: {}", stats.count());
println!("mean: {}", stats.mean());
println!("standard deviation: {}", stats.standard_deviation());

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...

20
examples/faceted_search.rs
View File

@@ -1,15 +1,17 @@
// # Basic Example
// # Faceted Search
//
// This example covers the basic functionalities of
// This example covers the faceted search functionalities of
// tantivy.
//
// We will :
// - define our schema
// = create an index in a directory
// - index few documents in our index
// - search for the best document matchings "sea whale"
// - retrieve the best document original content.
// - define a text field "name" in our schema
// - define a facet field "classification" in our schema
// - create an index in memory
// - index few documents with respective facets in our index
// - search and count the number of documents that the classifications start the facet "/Felidae"
// - Search the facet "/Felidae/Pantherinae" and count the number of documents that the
// classifications include the facet.
//
// ---
// Importing tantivy...
use tantivy::collector::FacetCollector;
@@ -21,7 +23,7 @@ fn main() -> tantivy::Result<()> {
// Let's create a temporary directory for the sake of this example
let mut schema_builder = Schema::builder();
let name = schema_builder.add_text_field("felin_name", TEXT | STORED);
let name = schema_builder.add_text_field("name", TEXT | STORED);
// this is our faceted field: its scientific classification
let classification = schema_builder.add_facet_field("classification", FacetOptions::default());

2
examples/integer_range_search.rs
View File

@@ -27,7 +27,7 @@ fn main() -> Result<()> {
reader.reload()?;
let searcher = reader.searcher();
// The end is excluded i.e. here we are searching up to 1969
let docs_in_the_sixties = RangeQuery::new_u64(year_field, 1960..1970);
let docs_in_the_sixties = RangeQuery::new_u64("year".to_string(), 1960..1970);
// Uses a Count collector to sum the total number of docs in the range
let num_60s_books = searcher.search(&docs_in_the_sixties, &Count)?;
assert_eq!(num_60s_books, 10);

73
examples/ip_field.rs Normal file
View File

@@ -0,0 +1,73 @@
// # IP Address example
//
// This example shows how the ip field can be used
// with IpV6 and IpV4.
use tantivy::collector::{Count, TopDocs};
use tantivy::query::QueryParser;
use tantivy::schema::{Schema, FAST, INDEXED, STORED, STRING};
use tantivy::Index;
fn main() -> tantivy::Result<()> {
// # Defining the schema
let mut schema_builder = Schema::builder();
let event_type = schema_builder.add_text_field("event_type", STRING | STORED);
let ip = schema_builder.add_ip_addr_field("ip", STORED | INDEXED | FAST);
let schema = schema_builder.build();
// # Indexing documents
let index = Index::create_in_ram(schema.clone());
let mut index_writer = index.writer(50_000_000)?;
let doc = schema.parse_document(
r#"{
"ip": "192.168.0.33",
"event_type": "login"
}"#,
)?;
index_writer.add_document(doc)?;
let doc = schema.parse_document(
r#"{
"ip": "192.168.0.80",
"event_type": "checkout"
}"#,
)?;
index_writer.add_document(doc)?;
let doc = schema.parse_document(
r#"{
"ip": "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
"event_type": "checkout"
}"#,
)?;
index_writer.add_document(doc)?;
index_writer.commit()?;
let reader = index.reader()?;
let searcher = reader.searcher();
let query_parser = QueryParser::for_index(&index, vec![event_type, ip]);
{
let query = query_parser.parse_query("ip:[192.168.0.0 TO 192.168.0.100]")?;
let count_docs = searcher.search(&*query, &TopDocs::with_limit(5))?;
assert_eq!(count_docs.len(), 2);
}
{
let query = query_parser.parse_query("ip:[192.168.1.0 TO 192.168.1.100]")?;
let count_docs = searcher.search(&*query, &TopDocs::with_limit(2))?;
assert_eq!(count_docs.len(), 0);
}
{
let query = query_parser.parse_query("ip:192.168.0.80")?;
let count_docs = searcher.search(&*query, &Count)?;
assert_eq!(count_docs, 1);
}
{
// IpV6 needs to be escaped because it contains `:`
let query = query_parser.parse_query("ip:\"2001:0db8:85a3:0000:0000:8a2e:0370:7334\"")?;
let count_docs = searcher.search(&*query, &Count)?;
assert_eq!(count_docs, 1);
}
Ok(())
}

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.

12
examples/warmer.rs
View File

@@ -4,7 +4,7 @@ use std::sync::{Arc, RwLock, Weak};
use tantivy::collector::TopDocs;
use tantivy::query::QueryParser;
use tantivy::schema::{Field, Schema, FAST, TEXT};
use tantivy::schema::{Schema, FAST, TEXT};
use tantivy::{
doc, DocAddress, DocId, Index, IndexReader, Opstamp, Searcher, SearcherGeneration, SegmentId,
SegmentReader, Warmer,
@@ -25,13 +25,13 @@ pub trait PriceFetcher: Send + Sync + 'static {
}
struct DynamicPriceColumn {
field: Field,
field: String,
price_cache: RwLock<HashMap<(SegmentId, Option<Opstamp>), Arc<Vec<Price>>>>,
price_fetcher: Box<dyn PriceFetcher>,
}
impl DynamicPriceColumn {
pub fn with_product_id_field<T: PriceFetcher>(field: Field, price_fetcher: T) -> Self {
pub fn with_product_id_field<T: PriceFetcher>(field: String, price_fetcher: T) -> Self {
DynamicPriceColumn {
field,
price_cache: Default::default(),
@@ -48,10 +48,10 @@ impl Warmer for DynamicPriceColumn {
fn warm(&self, searcher: &Searcher) -> tantivy::Result<()> {
for segment in searcher.segment_readers() {
let key = (segment.segment_id(), segment.delete_opstamp());
let product_id_reader = segment.fast_fields().u64(self.field)?;
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_val(doc as u64))
.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();
@@ -123,7 +123,7 @@ fn main() -> tantivy::Result<()> {
let price_table = ExternalPriceTable::default();
let price_dynamic_column = Arc::new(DynamicPriceColumn::with_product_id_field(
product_id,
"product_id".to_string(),
price_table.clone(),
));
price_table.update_price(OLIVE_OIL, 12);

12
fastfield_codecs/Cargo.toml
View File

@@ -1,18 +1,20 @@
[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/" }
ownedbytes = { version = "0.3.0", path = "../ownedbytes" }
prettytable-rs = {version="0.9.0", optional= true}
common = { version = "0.5", path = "../common/", package = "tantivy-common" }
tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
prettytable-rs = {version="0.10.0", optional= true}
rand = {version="0.8.3", optional= true}
fastdivide = "0.4"
log = "0.4"

154
fastfield_codecs/benches/bench.rs
View File

@@ -4,11 +4,11 @@ extern crate test;
#[cfg(test)]
mod tests {
use std::iter;
use std::ops::RangeInclusive;
use std::sync::Arc;
use common::OwnedBytes;
use fastfield_codecs::*;
use ownedbytes::OwnedBytes;
use rand::prelude::*;
use test::Bencher;
@@ -65,33 +65,30 @@ mod tests {
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = column.get_val(a as u64);
a = column.get_val(a as u32);
}
a
});
}
fn get_exp_data() -> Vec<u64> {
const FIFTY_PERCENT_RANGE: RangeInclusive<u64> = 1..=50;
const SINGLE_ITEM: u64 = 90;
const SINGLE_ITEM_RANGE: RangeInclusive<u64> = 90..=90;
const ONE_PERCENT_ITEM_RANGE: RangeInclusive<u64> = 49..=49;
fn get_data_50percent_item() -> Vec<u128> {
let mut rng = StdRng::from_seed([1u8; 32]);
let mut data = vec![];
for i in 0..100 {
let num = i * i;
data.extend(iter::repeat(i as u64).take(num));
for _ in 0..300_000 {
let val = rng.gen_range(1..=100);
data.push(val);
}
data.shuffle(&mut StdRng::from_seed([1u8; 32]));
data.push(SINGLE_ITEM);
// lengt = 328350
data.shuffle(&mut rng);
let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
data
}
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 get_u128_column_random() -> Arc<dyn Column<u128>> {
let permutation = generate_random();
let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
@@ -100,34 +97,123 @@ mod tests {
fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
let mut out = vec![];
serialize_u128(VecColumn::from(&data), &mut out).unwrap();
let iter_gen = || data.iter().cloned();
serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
let out = OwnedBytes::new(out);
open_u128(out).unwrap()
open_u128::<u128>(out).unwrap()
}
// U64 RANGE START
#[bench]
fn bench_intfastfield_getrange_u64_50percent_hit(b: &mut Bencher) {
let data = get_data_50percent_item();
let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
FIFTY_PERCENT_RANGE,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_intfastfield_getrange_u64_1percent_hit(b: &mut Bencher) {
let data = get_data_50percent_item();
let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
ONE_PERCENT_ITEM_RANGE,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_intfastfield_getrange_u64_single_hit(b: &mut Bencher) {
let data = get_data_50percent_item();
let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
SINGLE_ITEM_RANGE,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_intfastfield_getrange_u64_hit_all(b: &mut Bencher) {
let data = get_data_50percent_item();
let data = data.iter().map(|el| *el as u64).collect::<Vec<_>>();
let column: Arc<dyn Column<u64>> = serialize_and_load(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(0..=u64::MAX, 0..data.len() as u32, &mut positions);
positions
});
}
// U64 RANGE END
// U128 RANGE START
#[bench]
fn bench_intfastfield_getrange_u128_50percent_hit(b: &mut Bencher) {
let (major_item, _minor_item, data) = get_data_50percent_item();
let data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| column.get_between_vals(major_item..=major_item));
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
let (_major_item, minor_item, data) = get_data_50percent_item();
let data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| column.get_between_vals(minor_item..=minor_item));
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
0..data.len() as u32,
&mut positions,
);
positions
});
}
#[bench]
fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
let (_major_item, _minor_item, data) = get_data_50percent_item();
let data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| column.get_between_vals(0..=u128::MAX));
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
positions
});
}
// U128 RANGE END
#[bench]
fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
@@ -136,7 +222,7 @@ mod tests {
b.iter(|| {
let mut a = 0u128;
for i in 0u64..column.num_vals() as u64 {
a += column.get_val(i);
a += column.get_val(i as u32);
}
a
});
@@ -150,7 +236,7 @@ mod tests {
let n = column.num_vals();
let mut a = 0u128;
for i in (0..n / 5).map(|val| val * 5) {
a += column.get_val(i as u64);
a += column.get_val(i);
}
a
});
@@ -175,9 +261,9 @@ mod tests {
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
let mut a = 0;
for i in (0..n / 7).map(|val| val * 7) {
a += column.get_val(i as u64);
a += column.get_val(i as u32);
}
a
});
@@ -190,7 +276,7 @@ mod tests {
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0u64..n as u64 {
for i in 0u32..n as u32 {
a += column.get_val(i);
}
a
@@ -204,8 +290,8 @@ mod tests {
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0..n as u64 {
a += column.get_val(i);
for i in 0..n {
a += column.get_val(i as u32);
}
a
});

6
fastfield_codecs/src/bitpacked.rs
View File

@@ -1,6 +1,6 @@
use std::io::{self, Write};
use ownedbytes::OwnedBytes;
use common::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::serialize::NormalizedHeader;
@@ -17,7 +17,7 @@ pub struct BitpackedReader {
impl Column for BitpackedReader {
#[inline]
fn get_val(&self, doc: u64) -> u64 {
fn get_val(&self, doc: u32) -> u64 {
self.bit_unpacker.get(doc, &self.data)
}
#[inline]
@@ -30,7 +30,7 @@ impl Column for BitpackedReader {
self.normalized_header.max_value
}
#[inline]
fn num_vals(&self) -> u64 {
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}

28
fastfield_codecs/src/blockwise_linear.rs
View File

@@ -1,8 +1,7 @@
use std::sync::Arc;
use std::{io, iter};
use common::{BinarySerializable, CountingWriter, DeserializeFrom};
use ownedbytes::OwnedBytes;
use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::line::Line;
@@ -36,7 +35,7 @@ impl BinarySerializable for Block {
}
}
fn compute_num_blocks(num_vals: u64) -> usize {
fn compute_num_blocks(num_vals: u32) -> usize {
(num_vals as usize + CHUNK_SIZE - 1) / CHUNK_SIZE
}
@@ -47,7 +46,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
type Reader = BlockwiseLinearReader;
fn open_from_bytes(
bytes: ownedbytes::OwnedBytes,
bytes: common::OwnedBytes,
normalized_header: NormalizedHeader,
) -> io::Result<Self::Reader> {
let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
@@ -72,13 +71,13 @@ impl FastFieldCodec for BlockwiseLinearCodec {
// Estimate first_chunk and extrapolate
fn estimate(column: &dyn crate::Column) -> Option<f32> {
if column.num_vals() < 10 * CHUNK_SIZE as u64 {
if column.num_vals() < 10 * CHUNK_SIZE as u32 {
return None;
}
let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE as usize).collect();
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 u64);
let interpolated_val = line.eval(i as u32);
*buffer_val = buffer_val.wrapping_sub(interpolated_val);
}
let estimated_bit_width = first_chunk
@@ -95,7 +94,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
};
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() / CHUNK_SIZE as u64);
+ 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)
}
@@ -121,7 +120,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
assert!(!buffer.is_empty());
for (i, buffer_val) in buffer.iter_mut().enumerate() {
let interpolated_val = line.eval(i as u64);
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();
@@ -161,9 +160,9 @@ pub struct BlockwiseLinearReader {
impl Column for BlockwiseLinearReader {
#[inline(always)]
fn get_val(&self, idx: u64) -> u64 {
let block_id = (idx / CHUNK_SIZE as u64) as usize;
let idx_within_block = idx % (CHUNK_SIZE as u64);
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..];
@@ -171,16 +170,19 @@ impl Column for BlockwiseLinearReader {
interpoled_val.wrapping_add(bitpacked_diff)
}
#[inline(always)]
fn min_value(&self) -> u64 {
// The BlockwiseLinearReader assumes a normalized vector.
0u64
}
#[inline(always)]
fn max_value(&self) -> u64 {
self.normalized_header.max_value
}
fn num_vals(&self) -> u64 {
#[inline(always)]
fn num_vals(&self) -> u32 {
self.normalized_header.num_vals
}
}

184
fastfield_codecs/src/column.rs
View File

@@ -1,17 +1,21 @@
use std::fmt::{self, Debug};
use std::marker::PhantomData;
use std::ops::RangeInclusive;
use std::ops::{Range, RangeInclusive};
use tantivy_bitpacker::minmax;
pub trait Column<T: PartialOrd = u64>: Send + Sync {
/// Return the value associated to the given idx.
use crate::monotonic_mapping::StrictlyMonotonicFn;
/// `Column` provides columnar access on a field.
pub trait Column<T: PartialOrd + Debug = 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: u64) -> T;
fn get_val(&self, idx: u32) -> T;
/// Fills an output buffer with the fast field values
/// associated with the `DocId` going from
@@ -24,21 +28,28 @@ pub trait Column<T: PartialOrd = u64>: Send + Sync {
#[inline]
fn get_range(&self, start: u64, output: &mut [T]) {
for (out, idx) in output.iter_mut().zip(start..) {
*out = self.get_val(idx);
*out = self.get_val(idx as u32);
}
}
/// Return the positions of values which are in the provided range.
/// Get the positions of values which are in the provided value range.
///
/// Note that position == docid for single value fast fields
#[inline]
fn get_between_vals(&self, range: RangeInclusive<T>) -> Vec<u64> {
let mut vals = Vec::new();
for idx in 0..self.num_vals() {
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 range.contains(&val) {
vals.push(idx);
if value_range.contains(&val) {
positions.push(idx);
}
}
vals
}
/// Returns the minimum value for this fast field.
@@ -57,7 +68,8 @@ pub trait Column<T: PartialOrd = u64>: Send + Sync {
/// `.max_value()`.
fn max_value(&self) -> T;
fn num_vals(&self) -> u64;
/// 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> {
@@ -65,14 +77,15 @@ pub trait Column<T: PartialOrd = u64>: Send + Sync {
}
}
/// 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: u64) -> T {
impl<'a, C: Column<T>, T: Copy + PartialOrd + fmt::Debug> Column<T> for &'a C {
fn get_val(&self, idx: u32) -> T {
(*self).get_val(idx)
}
@@ -84,7 +97,7 @@ impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
(*self).max_value()
}
fn num_vals(&self) -> u64 {
fn num_vals(&self) -> u32 {
(*self).num_vals()
}
@@ -97,8 +110,8 @@ impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
}
}
impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
fn get_val(&self, position: u64) -> T {
impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> Column<T> for VecColumn<'a, T> {
fn get_val(&self, position: u32) -> T {
self.values[position as usize]
}
@@ -114,8 +127,8 @@ impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
self.max_value
}
fn num_vals(&self) -> u64 {
self.values.len() as u64
fn num_vals(&self) -> u32 {
self.values.len() as u32
}
fn get_range(&self, start: u64, output: &mut [T]) {
@@ -123,7 +136,7 @@ impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
}
}
impl<'a, T: Copy + Ord + Default, V> From<&'a V> for VecColumn<'a, T>
impl<'a, T: Copy + PartialOrd + Default, V> From<&'a V> for VecColumn<'a, T>
where V: AsRef<[T]> + ?Sized
{
fn from(values: &'a V) -> Self {
@@ -143,16 +156,30 @@ struct MonotonicMappingColumn<C, T, Input> {
_phantom: PhantomData<Input>,
}
/// Creates a view of a column transformed by a monotonic mapping.
pub fn monotonic_map_column<C, T, Input: PartialOrd, Output: PartialOrd>(
/// 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: Fn(Input) -> Output + Send + Sync,
Input: Send + Sync,
Output: Send + Sync,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Send + Sync + Copy + Debug,
Output: PartialOrd + Send + Sync + Copy + Debug,
{
MonotonicMappingColumn {
from_column,
@@ -161,42 +188,63 @@ where
}
}
impl<C, T, Input: PartialOrd, Output: PartialOrd> Column<Output>
for MonotonicMappingColumn<C, T, Input>
impl<C, T, Input, Output> Column<Output> for MonotonicMappingColumn<C, T, Input>
where
C: Column<Input>,
T: Fn(Input) -> Output + Send + Sync,
Input: Send + Sync,
Output: Send + Sync,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Send + Sync + Copy + Debug,
Output: PartialOrd + Send + Sync + Copy + Debug,
{
#[inline]
fn get_val(&self, idx: u64) -> Output {
fn get_val(&self, idx: u32) -> Output {
let from_val = self.from_column.get_val(idx);
(self.monotonic_mapping)(from_val)
self.monotonic_mapping.mapping(from_val)
}
fn min_value(&self) -> Output {
let from_min_value = self.from_column.min_value();
(self.monotonic_mapping)(from_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)(from_max_value)
self.monotonic_mapping.mapping(from_max_value)
}
fn num_vals(&self) -> u64 {
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(&self.monotonic_mapping))
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>,
) {
if range.start() > &self.max_value() || range.end() < &self.min_value() {
return;
}
let range = self.monotonic_mapping.inverse_coerce(range);
if range.start() > range.end() {
return;
}
self.from_column
.get_docids_for_value_range(range, 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>
@@ -210,9 +258,9 @@ where T: Iterator + Clone + ExactSizeIterator
impl<T> Column<T::Item> for IterColumn<T>
where
T: Iterator + Clone + ExactSizeIterator + Send + Sync,
T::Item: PartialOrd,
T::Item: PartialOrd + fmt::Debug,
{
fn get_val(&self, idx: u64) -> T::Item {
fn get_val(&self, idx: u32) -> T::Item {
self.0.clone().nth(idx as usize).unwrap()
}
@@ -224,8 +272,8 @@ where
self.0.clone().last().unwrap()
}
fn num_vals(&self) -> u64 {
self.0.len() as u64
fn num_vals(&self) -> u32 {
self.0.len() as u32
}
fn iter(&self) -> Box<dyn Iterator<Item = T::Item> + '_> {
@@ -236,19 +284,22 @@ where
#[cfg(test)]
mod tests {
use super::*;
use crate::MonotonicallyMappableToU64;
use crate::monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternalBaseval,
StrictlyMonotonicMappingToInternalGCDBaseval,
};
#[test]
fn test_monotonic_mapping() {
let vals = &[1u64, 3u64][..];
let vals = &[3u64, 5u64][..];
let col = VecColumn::from(vals);
let mapped = monotonic_map_column(col, |el| el + 4);
assert_eq!(mapped.min_value(), 5u64);
assert_eq!(mapped.max_value(), 7u64);
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), 5);
assert_eq!(mapped.get_val(1), 7);
assert_eq!(mapped.get_val(0), 1);
assert_eq!(mapped.get_val(1), 3);
}
#[test]
@@ -260,10 +311,15 @@ mod tests {
#[test]
fn test_monotonic_mapping_iter() {
let vals: Vec<u64> = (-1..99).map(i64::to_u64).collect();
let vals: Vec<u64> = (10..110u64).map(|el| el * 10).collect();
let col = VecColumn::from(&vals);
let mapped = monotonic_map_column(col, |el| i64::from_u64(el) * 10i64);
let val_i64s: Vec<i64> = mapped.iter().collect();
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));
}
@@ -271,20 +327,26 @@ mod tests {
#[test]
fn test_monotonic_mapping_get_range() {
let vals: Vec<u64> = (-1..99).map(i64::to_u64).collect();
let vals: Vec<u64> = (0..100u64).map(|el| el * 10).collect();
let col = VecColumn::from(&vals);
let mapped = monotonic_map_column(col, |el| i64::from_u64(el) * 10i64);
assert_eq!(mapped.min_value(), -10i64);
assert_eq!(mapped.max_value(), 980i64);
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_i64s: Vec<i64> = mapped.iter().collect();
assert_eq!(val_i64s.len(), 100);
let val_u64s: Vec<u64> = mapped.iter().collect();
assert_eq!(val_u64s.len(), 100);
for i in 0..100 {
assert_eq!(val_i64s[i as usize], mapped.get_val(i));
assert_eq!(val_i64s[i as usize], i64::from_u64(vals[i as usize]) * 10);
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 = [0i64; 20];
let mut buf = [0u64; 20];
mapped.get_range(7, &mut buf[..]);
assert_eq!(&val_i64s[7..][..20], &buf);
assert_eq!(&val_u64s[7..][..20], &buf);
}
}

4
fastfield_codecs/src/compact_space/build_compact_space.rs
View File

@@ -57,7 +57,7 @@ fn num_bits(val: u128) -> u8 {
/// metadata.
pub fn get_compact_space(
values_deduped_sorted: &BTreeSet<u128>,
total_num_values: u64,
total_num_values: u32,
cost_per_blank: usize,
) -> CompactSpace {
let mut compact_space_builder = CompactSpaceBuilder::new();
@@ -208,7 +208,7 @@ impl CompactSpaceBuilder {
};
let covered_range_len = range_mapping.range_length();
ranges_mapping.push(range_mapping);
compact_start += covered_range_len as u64;
compact_start += covered_range_len;
}
// println!("num ranges {}", ranges_mapping.len());
CompactSpace { ranges_mapping }

279
fastfield_codecs/src/compact_space/mod.rs
View File

@@ -14,11 +14,10 @@ use std::{
cmp::Ordering,
collections::BTreeSet,
io::{self, Write},
ops::RangeInclusive,
ops::{Range, RangeInclusive},
};
use common::{BinarySerializable, CountingWriter, VInt, VIntU128};
use ownedbytes::OwnedBytes;
use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
use crate::compact_space::build_compact_space::get_compact_space;
@@ -97,7 +96,7 @@ impl BinarySerializable for CompactSpace {
};
let range_length = range_mapping.range_length();
ranges_mapping.push(range_mapping);
compact_start += range_length as u64;
compact_start += range_length;
}
Ok(Self { ranges_mapping })
@@ -165,16 +164,16 @@ pub struct IPCodecParams {
bit_unpacker: BitUnpacker,
min_value: u128,
max_value: u128,
num_vals: u64,
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(column: &impl Column<u128>) -> Self {
pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u32) -> Self {
let mut values_sorted = BTreeSet::new();
values_sorted.extend(column.iter());
let total_num_values = column.num_vals();
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);
@@ -200,7 +199,7 @@ impl CompactSpaceCompressor {
bit_unpacker: BitUnpacker::new(num_bits),
min_value,
max_value,
num_vals: total_num_values as u64,
num_vals: total_num_values,
num_bits,
},
}
@@ -267,7 +266,7 @@ impl BinarySerializable for IPCodecParams {
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 u64;
let num_vals = VIntU128::deserialize(reader)?.0 as u32;
let num_bits = u8::deserialize(reader)?;
let compact_space = CompactSpace::deserialize(reader)?;
@@ -284,7 +283,7 @@ impl BinarySerializable for IPCodecParams {
impl Column<u128> for CompactSpaceDecompressor {
#[inline]
fn get_val(&self, doc: u64) -> u128 {
fn get_val(&self, doc: u32) -> u128 {
self.get(doc)
}
@@ -296,7 +295,7 @@ impl Column<u128> for CompactSpaceDecompressor {
self.max_value()
}
fn num_vals(&self) -> u64 {
fn num_vals(&self) -> u32 {
self.params.num_vals
}
@@ -304,8 +303,15 @@ impl Column<u128> for CompactSpaceDecompressor {
fn iter(&self) -> Box<dyn Iterator<Item = u128> + '_> {
Box::new(self.iter())
}
fn get_between_vals(&self, range: RangeInclusive<u128>) -> Vec<u64> {
self.get_between_vals(range)
#[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)
}
}
@@ -340,12 +346,19 @@ impl CompactSpaceDecompressor {
/// Comparing on compact space: Real dataset 1.08 GElements/s
///
/// Comparing on original space: Real dataset .06 GElements/s (not completely optimized)
pub fn get_between_vals(&self, range: RangeInclusive<u128>) -> Vec<u64> {
if range.start() > range.end() {
return Vec::new();
#[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 from_value = *range.start();
let to_value = *range.end();
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);
@@ -353,7 +366,7 @@ impl CompactSpaceDecompressor {
// 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 Vec::new(),
(Err(pos1), Err(pos2)) if pos1 == pos2 => return,
_ => {}
}
@@ -375,19 +388,20 @@ impl CompactSpaceDecompressor {
});
let range = compact_from..=compact_to;
let mut positions = Vec::new();
let scan_num_docs = position_range.end - position_range.start;
let step_size = 4;
let cutoff = self.params.num_vals - self.params.num_vals % step_size;
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 as u64, &self.data);
let get_val = |idx| self.params.bit_unpacker.get(idx, &self.data);
// unrolled loop
for idx in (0..cutoff).step_by(step_size as usize) {
for idx in (position_range.start..cutoff).step_by(step_size as usize) {
let idx1 = idx;
let idx2 = idx + 1;
let idx3 = idx + 2;
@@ -403,17 +417,14 @@ impl CompactSpaceDecompressor {
}
// handle rest
for idx in cutoff..self.params.num_vals {
for idx in cutoff..position_range.end {
push_if_in_range(idx, get_val(idx));
}
positions
}
#[inline]
fn iter_compact(&self) -> impl Iterator<Item = u64> + '_ {
(0..self.params.num_vals)
.map(move |idx| self.params.bit_unpacker.get(idx as u64, &self.data) as u64)
(0..self.params.num_vals).map(move |idx| self.params.bit_unpacker.get(idx, &self.data))
}
#[inline]
@@ -425,7 +436,7 @@ impl CompactSpaceDecompressor {
}
#[inline]
pub fn get(&self, idx: u64) -> u128 {
pub fn get(&self, idx: u32) -> u128 {
let compact = self.params.bit_unpacker.get(idx, &self.data);
self.compact_to_u128(compact)
}
@@ -442,8 +453,13 @@ impl CompactSpaceDecompressor {
#[cfg(test)]
mod tests {
use std::fmt;
use super::*;
use crate::{open_u128, serialize_u128, VecColumn};
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() {
@@ -452,7 +468,7 @@ mod tests {
]
.into_iter()
.collect();
let compact_space = get_compact_space(ips, ips.len() as u64, 11);
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());
@@ -483,24 +499,30 @@ mod tests {
#[test]
fn compact_space_amplitude_test() {
let ips = &[100000u128, 1000000].into_iter().collect();
let compact_space = get_compact_space(ips, ips.len() as u64, 1);
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(data: OwnedBytes, expected: &[u128]) {
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 u64);
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 u64)
.map(|pos| pos as u32)
.collect::<Vec<_>>();
let positions = decompressor.get_between_vals(range);
let mut positions = Vec::new();
decompressor.get_positions_for_value_range(
range,
0..decompressor.num_vals(),
&mut positions,
);
assert_eq!(positions, expected_positions);
};
@@ -513,10 +535,18 @@ mod tests {
fn test_aux_vals(u128_vals: &[u128]) -> OwnedBytes {
let mut out = Vec::new();
serialize_u128(VecColumn::from(u128_vals), &mut out).unwrap();
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
}
@@ -533,26 +563,111 @@ mod tests {
4_000_211_222u128,
333u128,
];
let data = test_aux_vals(vals);
let mut data = test_aux_vals(vals);
let _header = U128Header::deserialize(&mut data);
let decomp = CompactSpaceDecompressor::open(data).unwrap();
let positions = decomp.get_between_vals(0..=1);
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 = decomp.get_between_vals(0..=2);
let positions =
get_positions_for_value_range_helper(&decomp, 0..=2, complete_range.clone());
assert_eq!(positions, vec![0]);
let positions = decomp.get_between_vals(0..=3);
let positions =
get_positions_for_value_range_helper(&decomp, 0..=3, complete_range.clone());
assert_eq!(positions, vec![0, 2]);
assert_eq!(decomp.get_between_vals(99999u128..=99999u128), vec![3]);
assert_eq!(decomp.get_between_vals(99999u128..=100000u128), vec![3, 4]);
assert_eq!(decomp.get_between_vals(99998u128..=100000u128), vec![3, 4]);
assert_eq!(decomp.get_between_vals(99998u128..=99999u128), vec![3]);
assert_eq!(decomp.get_between_vals(99998u128..=99998u128), vec![]);
assert_eq!(decomp.get_between_vals(333u128..=333u128), vec![8]);
assert_eq!(decomp.get_between_vals(332u128..=333u128), vec![8]);
assert_eq!(decomp.get_between_vals(332u128..=334u128), vec![8]);
assert_eq!(decomp.get_between_vals(333u128..=334u128), vec![8]);
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!(
decomp.get_between_vals(4_000_211_221u128..=5_000_000_000u128),
get_positions_for_value_range_helper(
&decomp,
4_000_211_221u128..=5_000_000_000u128,
complete_range
),
vec![6, 7]
);
}
@@ -575,14 +690,32 @@ mod tests {
4_000_211_222u128,
333u128,
];
let data = test_aux_vals(vals);
let mut data = test_aux_vals(vals);
let _header = U128Header::deserialize(&mut data);
let decomp = CompactSpaceDecompressor::open(data).unwrap();
let positions = decomp.get_between_vals(0..=5);
assert_eq!(positions, vec![]);
let positions = decomp.get_between_vals(0..=100);
assert_eq!(positions, vec![0]);
let positions = decomp.get_between_vals(0..=105);
assert_eq!(positions, vec![0]);
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 + fmt::Debug>(
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]
@@ -603,13 +736,29 @@ mod tests {
5_000_000_000,
];
let mut out = Vec::new();
serialize_u128(VecColumn::from(vals), &mut out).unwrap();
let decomp = open_u128(OwnedBytes::new(out)).unwrap();
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!(decomp.get_between_vals(199..=200), vec![0]);
assert_eq!(decomp.get_between_vals(199..=201), vec![0, 1]);
assert_eq!(decomp.get_between_vals(200..=200), vec![0]);
assert_eq!(decomp.get_between_vals(1_000_000..=1_000_000), vec![11]);
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]

38
fastfield_codecs/src/format_version.rs Normal file
View File

@@ -0,0 +1,38 @@
use std::io;
use common::{BinarySerializable, 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))
}

2
fastfield_codecs/src/gcd.rs
View File

@@ -45,7 +45,7 @@ mod tests {
use std::io;
use std::num::NonZeroU64;
use ownedbytes::OwnedBytes;
use common::OwnedBytes;
use crate::gcd::{compute_gcd, find_gcd};
use crate::{FastFieldCodecType, VecColumn};

150
fastfield_codecs/src/lib.rs
View File

@@ -1,5 +1,12 @@
#![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;
@@ -7,40 +14,58 @@ extern crate more_asserts;
#[cfg(all(test, feature = "unstable"))]
extern crate test;
use std::io;
use std::io::Write;
use std::sync::Arc;
use std::{fmt, io};
use common::BinarySerializable;
use common::{BinarySerializable, OwnedBytes};
use compact_space::CompactSpaceDecompressor;
use ownedbytes::OwnedBytes;
use serialize::Header;
use format_version::read_format_version;
use monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
};
use null_index_footer::read_null_index_footer;
use serialize::{Header, U128Header};
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;
mod column;
mod gcd;
mod serialize;
pub mod serialize;
use self::bitpacked::BitpackedCodec;
use self::blockwise_linear::BlockwiseLinearCodec;
pub use self::column::{monotonic_map_column, Column, VecColumn};
pub use self::column::{monotonic_map_column, Column, IterColumn, VecColumn};
use self::linear::LinearCodec;
pub use self::monotonic_mapping::MonotonicallyMappableToU64;
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,
}
@@ -58,11 +83,11 @@ impl BinarySerializable for FastFieldCodecType {
}
impl FastFieldCodecType {
pub fn to_code(self) -> u8 {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub fn from_code(code: u8) -> Option<Self> {
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::Bitpacked),
2 => Some(Self::Linear),
@@ -72,15 +97,61 @@ impl FastFieldCodecType {
}
}
/// Returns the correct codec reader wrapped in the `Arc` for the data.
pub fn open_u128(bytes: OwnedBytes) -> io::Result<Arc<dyn Column<u128>>> {
Ok(Arc::new(CompactSpaceDecompressor::open(bytes)?))
#[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<T: MonotonicallyMappableToU64>(
mut bytes: OwnedBytes,
pub fn open_u128<Item: MonotonicallyMappableToU128 + fmt::Debug>(
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 + fmt::Debug>(
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),
@@ -91,7 +162,7 @@ pub fn open<T: MonotonicallyMappableToU64>(
}
}
fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64 + fmt::Debug>(
bytes: OwnedBytes,
header: &Header,
) -> io::Result<Arc<dyn Column<Item>>> {
@@ -99,11 +170,15 @@ fn open_specific_codec<C: FastFieldCodec, Item: MonotonicallyMappableToU64>(
let reader = C::open_from_bytes(bytes, normalized_header)?;
let min_value = header.min_value;
if let Some(gcd) = header.gcd {
let monotonic_mapping = move |val: u64| Item::from_u64(min_value + val * gcd.get());
Ok(Arc::new(monotonic_map_column(reader, monotonic_mapping)))
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
} else {
let monotonic_mapping = move |val: u64| Item::from_u64(min_value + val);
Ok(Arc::new(monotonic_map_column(reader, monotonic_mapping)))
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalBaseval::new(min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
}
}
@@ -135,6 +210,7 @@ trait FastFieldCodec: 'static {
fn estimate(column: &dyn Column) -> Option<f32>;
}
/// The list of all available codecs for u64 convertible data.
pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
@@ -143,6 +219,7 @@ pub const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
#[cfg(test)]
mod tests {
use proptest::prelude::*;
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
@@ -168,15 +245,32 @@ mod tests {
let actual_compression = out.len() as f32 / (data.len() as f32 * 8.0);
let reader = crate::open::<u64>(OwnedBytes::new(out)).unwrap();
assert_eq!(reader.num_vals(), data.len() as u64);
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 u64);
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))
}
@@ -229,6 +323,9 @@ mod tests {
pub fn get_codec_test_datasets() -> Vec<(Vec<u64>, &'static str)> {
let mut data_and_names = vec![];
let data = vec![10];
data_and_names.push((data, "minimal test"));
let data = (10..=10_000_u64).collect::<Vec<_>>();
data_and_names.push((data, "simple monotonically increasing"));
@@ -236,6 +333,9 @@ mod tests {
vec![5, 6, 7, 8, 9, 10, 99, 100],
"offset in linear interpol",
));
data_and_names.push((vec![3, 18446744073709551613, 5], "docid range regression"));
data_and_names.push((vec![5, 50, 3, 13, 1, 1000, 35], "rand small"));
data_and_names.push((vec![10], "single value"));
@@ -312,7 +412,7 @@ mod tests {
#[test]
fn estimation_test_bad_interpolation_case_monotonically_increasing() {
let mut data: Vec<u64> = (200..=20000_u64).collect();
let mut data: Vec<u64> = (201..=20000_u64).collect();
data.push(1_000_000);
let data: VecColumn = data.as_slice().into();
@@ -343,7 +443,7 @@ mod tests {
mod bench {
use std::sync::Arc;
use ownedbytes::OwnedBytes;
use common::OwnedBytes;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::{self, Bencher};
@@ -386,7 +486,7 @@ mod bench {
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u64);
let val = col.get_val(pos as u32);
sum = sum.wrapping_add(val);
}
sum
@@ -398,7 +498,7 @@ mod bench {
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u64);
let val = col.get_val(pos as u32);
sum = sum.wrapping_add(val);
}
sum

52
fastfield_codecs/src/line.rs
View File

@@ -1,5 +1,5 @@
use std::io;
use std::num::NonZeroU64;
use std::num::NonZeroU32;
use common::{BinarySerializable, VInt};
@@ -29,7 +29,7 @@ pub struct Line {
/// compute_slope(y0, y1)
/// = compute_slope(y0 + X % 2^64, y1 + X % 2^64)
/// `
fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU64) -> u64 {
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 {
@@ -43,7 +43,7 @@ fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU64) -> u64 {
return 0u64;
}
let abs_slope = (abs_dy << 32) / num_vals.get();
let abs_slope = (abs_dy << 32) / num_vals.get() as u64;
if sign {
abs_slope
} else {
@@ -62,35 +62,43 @@ fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU64) -> u64 {
impl Line {
#[inline(always)]
pub fn eval(&self, x: u64) -> u64 {
let linear_part = (x.wrapping_mul(self.slope) >> 32) as i32 as u64;
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(ys: &dyn Column, sample_positions: &[u64]) -> Self {
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(
ys,
sample_positions
.iter()
.cloned()
.map(|pos| (pos, ys.get_val(pos))),
first_val,
last_val,
num_vals as u32,
sample_positions_and_values.iter().cloned(),
)
}
// Intercept is only computed from provided positions
fn train_from(ys: &dyn Column, positions_and_values: impl Iterator<Item = (u64, u64)>) -> Self {
let num_vals = if let Some(num_vals) = NonZeroU64::new(ys.num_vals() - 1) {
num_vals
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 = ys.get_val(0);
let y1 = ys.get_val(num_vals.get());
let y0 = first_val;
let y1 = last_val;
// We first independently pick our slope.
let slope = compute_slope(y0, y1, num_vals);
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.
@@ -121,7 +129,7 @@ impl Line {
};
let heuristic_shift = y0.wrapping_sub(MID_POINT);
line.intercept = positions_and_values
.map(|(pos, y)| y.wrapping_sub(line.eval(pos)))
.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
@@ -138,8 +146,12 @@ impl Line {
/// 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(
ys,
first_val,
last_val,
ys.num_vals(),
ys.iter().enumerate().map(|(pos, val)| (pos as u64, val)),
)
}
@@ -187,7 +199,7 @@ mod tests {
let line = Line::train(&VecColumn::from(&ys));
ys.iter()
.enumerate()
.map(|(x, y)| y.wrapping_sub(line.eval(x as u64)))
.map(|(x, y)| y.wrapping_sub(line.eval(x as u32)))
.max()
}

38
fastfield_codecs/src/linear.rs
View File

@@ -1,7 +1,6 @@
use std::io::{self, Write};
use common::BinarySerializable;
use ownedbytes::OwnedBytes;
use common::{BinarySerializable, OwnedBytes};
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use crate::line::Line;
@@ -19,25 +18,25 @@ pub struct LinearReader {
impl Column for LinearReader {
#[inline]
fn get_val(&self, doc: u64) -> u64 {
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]
#[inline(always)]
fn min_value(&self) -> u64 {
// The LinearReader assumes a normalized vector.
0u64
}
#[inline]
#[inline(always)]
fn max_value(&self) -> u64 {
self.header.max_value
}
#[inline]
fn num_vals(&self) -> u64 {
fn num_vals(&self) -> u32 {
self.header.num_vals
}
}
@@ -93,7 +92,7 @@ impl FastFieldCodec for LinearCodec {
.iter()
.enumerate()
.map(|(pos, actual_value)| {
let calculated_value = line.eval(pos as u64);
let calculated_value = line.eval(pos as u32);
actual_value.wrapping_sub(calculated_value)
})
.max()
@@ -108,7 +107,7 @@ impl FastFieldCodec for LinearCodec {
let mut bit_packer = BitPacker::new();
for (pos, actual_value) in column.iter().enumerate() {
let calculated_value = line.eval(pos as u64);
let calculated_value = line.eval(pos as u32);
let offset = actual_value.wrapping_sub(calculated_value);
bit_packer.write(offset, num_bits, write)?;
}
@@ -126,19 +125,21 @@ impl FastFieldCodec for LinearCodec {
return None; // disable compressor for this case
}
// let's sample at 0%, 5%, 10% .. 95%, 100%
let num_vals = column.num_vals() as f32 / 100.0;
let sample_positions = (0..20)
.map(|pos| (num_vals * pos as f32 * 5.0) as u64)
.collect::<Vec<_>>();
let limit_num_vals = column.num_vals().min(100_000);
let line = Line::estimate(column, &sample_positions);
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 estimated_bit_width = sample_positions
let line = Line::estimate(&sample_positions_and_values);
let estimated_bit_width = sample_positions_and_values
.into_iter()
.map(|pos| {
let actual_value = column.get_val(pos);
let interpolated_val = line.eval(pos as u64);
.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)
@@ -146,6 +147,7 @@ impl FastFieldCodec for LinearCodec {
.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)

19
fastfield_codecs/src/main.rs
View File

@@ -6,10 +6,10 @@ use std::io::BufRead;
use std::net::{IpAddr, Ipv6Addr};
use std::str::FromStr;
use common::OwnedBytes;
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]) {
@@ -90,7 +90,7 @@ fn bench_ip() {
{
let mut data = vec![];
for dataset in dataset.chunks(500_000) {
serialize_u128(VecColumn::from(dataset), &mut data).unwrap();
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);
@@ -101,7 +101,10 @@ fn bench_ip() {
}
let mut data = vec![];
serialize_u128(VecColumn::from(&dataset), &mut data).unwrap();
{
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);
@@ -110,11 +113,17 @@ fn bench_ip() {
(data.len() * 8) as f32 / dataset.len() as f32
);
let decompressor = open_u128(OwnedBytes::new(data)).unwrap();
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");
let doc_values = decompressor.get_between_vals(value..=value);
decompressor.get_docids_for_value_range(
value..=value,
0..decompressor.num_vals(),
&mut doc_values,
);
println!("{:?}", doc_values.len());
}
}

266
fastfield_codecs/src/monotonic_mapping.rs
View File

@@ -1,4 +1,16 @@
pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
use std::fmt;
use std::marker::PhantomData;
use std::ops::RangeInclusive;
use fastdivide::DividerU64;
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 + fmt::Debug
{
/// Converts a value to u64.
///
/// Internally all fast field values are encoded as u64.
@@ -11,11 +23,229 @@ pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync
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: Copy, Internal: Copy> {
/// 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;
/// Maps a user provded value from External to Internal.
/// It may be necessary to coerce the value if it is outside the value space.
/// In that case it tries to find the next greater value in the value space.
///
/// Returns a bool to mark if a value was outside the value space and had to be coerced _up_.
/// With that information we can detect if two values in a range both map outside the same value
/// space.
///
/// coerce_up means the next valid upper value in the value space will be chosen if the value
/// has to be coerced.
fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<Internal> {
self.mapping(*inp.start())..=self.mapping(*inp.end())
}
/// Inverse of `mapping_coerce`.
fn inverse_coerce(&self, out: RangeInclusive<Internal>) -> RangeInclusive<External> {
self.inverse(*out.start())..=self.inverse(*out.end())
}
}
/// 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>,
From: Copy,
To: Copy,
{
#[inline(always)]
fn mapping(&self, val: To) -> From {
self.orig_mapping.inverse(val)
}
#[inline(always)]
fn inverse(&self, val: From) -> To {
self.orig_mapping.mapping(val)
}
#[inline]
fn mapping_coerce(&self, inp: RangeInclusive<To>) -> RangeInclusive<From> {
self.orig_mapping.inverse_coerce(inp)
}
#[inline]
fn inverse_coerce(&self, out: RangeInclusive<From>) -> RangeInclusive<To> {
self.orig_mapping.mapping_coerce(out)
}
}
/// 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
{
#[inline(always)]
fn mapping(&self, inp: External) -> u128 {
External::to_u128(inp)
}
#[inline(always)]
fn inverse(&self, out: u128) -> External {
External::from_u128(out)
}
}
impl<External: MonotonicallyMappableToU64, T: MonotonicallyMappableToU64>
StrictlyMonotonicFn<External, u64> for StrictlyMonotonicMappingToInternal<T>
where T: MonotonicallyMappableToU64
{
#[inline(always)]
fn mapping(&self, inp: External) -> u64 {
External::to_u64(inp)
}
#[inline(always)]
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
{
#[inline(always)]
fn mapping(&self, inp: External) -> u64 {
self.gcd_divider
.divide(External::to_u64(inp) - self.min_value)
}
#[inline(always)]
fn inverse(&self, out: u64) -> External {
External::from_u64(self.min_value + out * self.gcd)
}
#[inline]
#[allow(clippy::reversed_empty_ranges)]
fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
let end = External::to_u64(*inp.end());
if end < self.min_value || inp.end() < inp.start() {
return 1..=0;
}
let map_coerce = |mut inp, coerce_up| {
let inp_lower_bound = self.inverse(0);
if inp < inp_lower_bound {
inp = inp_lower_bound;
}
let val = External::to_u64(inp);
let need_coercion = coerce_up && (val - self.min_value) % self.gcd != 0;
let mut mapped_val = self.mapping(inp);
if need_coercion {
mapped_val += 1;
}
mapped_val
};
let start = map_coerce(*inp.start(), true);
let end = map_coerce(*inp.end(), false);
start..=end
}
}
/// Strictly monotonic mapping with a base value.
pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
min_value: u64,
}
impl StrictlyMonotonicMappingToInternalBaseval {
#[inline(always)]
pub(crate) fn new(min_value: u64) -> Self {
Self { min_value }
}
}
impl<External: MonotonicallyMappableToU64> StrictlyMonotonicFn<External, u64>
for StrictlyMonotonicMappingToInternalBaseval
{
#[inline]
#[allow(clippy::reversed_empty_ranges)]
fn mapping_coerce(&self, inp: RangeInclusive<External>) -> RangeInclusive<u64> {
if External::to_u64(*inp.end()) < self.min_value {
return 1..=0;
}
let start = self.mapping(External::to_u64(*inp.start()).max(self.min_value));
let end = self.mapping(External::to_u64(*inp.end()));
start..=end
}
#[inline(always)]
fn mapping(&self, val: External) -> u64 {
External::to_u64(val) - self.min_value
}
#[inline(always)]
fn inverse(&self, val: u64) -> External {
External::from_u64(self.min_value + val)
}
}
impl MonotonicallyMappableToU64 for u64 {
#[inline(always)]
fn to_u64(self) -> u64 {
self
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
val
}
@@ -45,12 +275,46 @@ impl MonotonicallyMappableToU64 for bool {
}
}
// TODO remove me.
// Tantivy should refuse NaN values and work with NotNaN internally.
impl MonotonicallyMappableToU64 for f64 {
#[inline(always)]
fn to_u64(self) -> u64 {
common::f64_to_u64(self)
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
common::u64_to_f64(val)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[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: Copy>(
mapping: &T,
test_val: K,
) {
assert_eq!(mapping.inverse(mapping.mapping(test_val)), test_val);
}
}

43
fastfield_codecs/src/monotonic_mapping_u128.rs Normal file
View File

@@ -0,0 +1,43 @@
use std::fmt;
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 + fmt::Debug
{
/// 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())
}

500
fastfield_codecs/src/null_index/dense.rs Normal file
View File

@@ -0,0 +1,500 @@
use std::convert::TryInto;
use std::io::{self, Write};
use common::{BinarySerializable, OwnedBytes};
use itertools::Itertools;
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;
/// Interpreting the bitvec as a list of 64 bits from the low weight to the
/// high weight.
///
/// This function returns the number of bits set to 1 within
/// `[0..pos_in_vec)`.
#[inline]
fn count_ones(bitvec: u64, pos_in_bitvec: u32) -> u32 {
let mask = (1u64 << pos_in_bitvec) - 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) {
Some(index_block.offset + ones_in_block)
} 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), 0);
assert_eq!(count_ones(block, 1), 1);
assert_eq!(count_ones(block, 2), 1);
assert_eq!(count_ones(block, 3), 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,
avg_step_size: u32,
avg_deviation: 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(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
if current >= end {
None
} else {
Some(current)
}
})
}
fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
let ratio = percent as f32 / 100.0;
let step_size = (1f32 / ratio) as u32;
let deviation = step_size - 1;
random_range_iterator(0, num_values, step_size, deviation)
}
fn walk_over_data(codec: &DenseCodec, avg_step_size: u32) -> Option<u32> {
walk_over_data_from_positions(
codec,
random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
)
}
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_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 1000));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.5f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
.last()
});
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(10.0, num_non_nulls))
.last()
});
}
#[bench]
fn bench_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_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.90f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
.last()
});
}
#[bench]
fn bench_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;
}

768
fastfield_codecs/src/null_index/sparse.rs Normal file
View File

@@ -0,0 +1,768 @@
use std::io::{self, Write};
use common::{BitSet, GroupByIteratorExtended, 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.
#[inline]
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 {
#[inline]
pub fn exists(&self, idx: u32) -> bool {
self.codec.exists(idx)
}
#[inline]
pub fn translate_to_codec_idx(&self, idx: u32) -> Option<u32> {
self.codec.translate_to_codec_idx(idx)
}
#[inline]
pub fn translate_codec_idx_to_original_idx_iter<'a>(
&'a self,
iter: impl Iterator<Item = u32> + 'a,
) -> impl Iterator<Item = u32> + 'a {
self.codec.translate_codec_idx_to_original_idx(iter)
}
#[inline]
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
#[inline]
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)
}
}
}
#[inline]
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.
/// Correctness: Provided values must be in increasing values
///
/// # 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.group_by(move |codec_idx| {
block_pos = self.find_block(*codec_idx, block_pos);
block_pos
})
.flat_map(move |(block_pos, block_iter)| {
let block_doc_idx_start = block_pos * ELEMENTS_PER_BLOCK;
let block = &self.blocks[block_pos as usize];
let offset = block.offset();
let indexes_in_block_iter = block_iter.map(move |codec_idx| codec_idx - 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,
indexes_in_block_iter.collect::<Vec<_>>()
)
}
SparseCodecBlockVariant::Dense(dense) => {
Box::new(dense.translate_codec_idx_to_original_idx_iter(indexes_in_block_iter))
as Box<dyn Iterator<Item = u32>>
}
SparseCodecBlockVariant::Sparse(block) => {
Box::new(indexes_in_block_iter.map(move |idx_in_block| {
block.value_at_idx(&self.data, idx_in_block as u16) as u32
}))
}
}
.map(move |idx| idx + block_doc_idx_start)
})
}
}
#[inline]
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,
avg_step_size: u32,
avg_deviation: 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(avg_step_size - avg_deviation..=avg_step_size + avg_deviation);
if current >= end {
None
} else {
Some(current)
}
})
}
fn n_percent_step_iterator(percent: f32, num_values: u32) -> impl Iterator<Item = u32> {
let ratio = percent as f32 / 100.0;
let step_size = (1f32 / ratio) as u32;
let deviation = step_size - 1;
random_range_iterator(0, num_values, step_size, deviation)
}
fn walk_over_data(codec: &SparseCodec, avg_step_size: u32) -> Option<u32> {
walk_over_data_from_positions(
codec,
random_range_iterator(0, TOTAL_NUM_VALUES, avg_step_size, 0),
)
}
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_translate_orig_to_codec_1percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_5percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.05f64);
bench.iter(|| walk_over_data(&codec, 1000));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_1percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_10percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_full_scan_90percent_filled(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data_from_positions(&codec, 0..TOTAL_NUM_VALUES));
}
#[bench]
fn bench_translate_orig_to_codec_10percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.1f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_50percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.5f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_orig_to_codec_90percent_filled_1percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.9f64);
bench.iter(|| walk_over_data(&codec, 100));
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_0comma005percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
.last()
});
}
#[bench]
fn bench_translate_codec_to_orig_1percent_filled_10percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.01f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(10.0, num_non_nulls))
.last()
});
}
#[bench]
fn bench_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_translate_codec_to_orig_90percent_filled_0comma005percent_hit(bench: &mut Bencher) {
let codec = gen_bools(0.90f64);
let num_non_nulls = codec.num_non_nulls();
bench.iter(|| {
codec
.translate_codec_idx_to_original_idx(n_percent_step_iterator(0.005, num_non_nulls))
.last()
});
}
#[bench]
fn bench_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()
});
}
}

145
fastfield_codecs/src/null_index_footer.rs Normal file
View File

@@ -0,0 +1,145 @@
use std::io::{self, Write};
use std::ops::Range;
use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt};
#[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()
);
}
}

225
fastfield_codecs/src/serialize.rs
View File

@@ -17,38 +17,46 @@
// 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/>.
use std::io;
use std::num::NonZeroU64;
use std::sync::Arc;
use std::{fmt, io};
use common::{BinarySerializable, VInt};
use fastdivide::DividerU64;
use common::{BinarySerializable, OwnedBytes, 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,
VecColumn, ALL_CODEC_TYPES,
U128FastFieldCodecType, VecColumn, ALL_CODEC_TYPES,
};
/// The normalized header gives some parameters after applying the following
/// normalization of the vector:
/// val -> (val - min_value) / gcd
/// `val -> (val - min_value) / gcd`
///
/// By design, after normalization, `min_value = 0` and `gcd = 1`.
#[derive(Debug, Copy, Clone)]
pub struct NormalizedHeader {
pub num_vals: u64,
/// 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: u64,
pub num_vals: u32,
pub min_value: u64,
pub max_value: u64,
pub gcd: Option<NonZeroU64>,
@@ -57,8 +65,11 @@ pub(crate) struct Header {
impl Header {
pub fn normalized(self) -> NormalizedHeader {
let max_value =
(self.max_value - self.min_value) / self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
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,
@@ -66,10 +77,7 @@ impl Header {
}
pub fn normalize_column<C: Column>(&self, from_column: C) -> impl Column {
let min_value = self.min_value;
let gcd = self.gcd.map(|gcd| gcd.get()).unwrap_or(1);
let divider = DividerU64::divide_by(gcd);
monotonic_map_column(from_column, move |val| divider.divide(val - min_value))
normalize_column(from_column, self.min_value, self.gcd)
}
pub fn compute_header(
@@ -81,9 +89,8 @@ impl Header {
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 divider = DividerU64::divide_by(gcd.map(|gcd| gcd.get()).unwrap_or(1u64));
let shifted_column = monotonic_map_column(&column, |val| divider.divide(val - min_value));
let codec_type = detect_codec(shifted_column, codecs)?;
let normalized_column = normalize_column(column, min_value, gcd);
let codec_type = detect_codec(normalized_column, codecs)?;
Some(Header {
num_vals,
min_value,
@@ -94,9 +101,42 @@ impl Header {
}
}
#[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).serialize(writer)?;
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 {
@@ -109,7 +149,7 @@ impl BinarySerializable for Header {
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let num_vals = VInt::deserialize(reader)?.0;
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;
@@ -125,16 +165,21 @@ impl BinarySerializable for Header {
}
}
pub fn estimate<T: MonotonicallyMappableToU64>(
/// 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 + fmt::Debug>(
typed_column: impl Column<T>,
codec_type: FastFieldCodecType,
) -> Option<f32> {
let column = monotonic_map_column(typed_column, T::to_u64);
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 divider = DividerU64::divide_by(gcd.map(|gcd| gcd.get()).unwrap_or(1u64));
let normalized_column = monotonic_map_column(&column, |val| divider.divide(val - min_value));
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),
@@ -142,25 +187,111 @@ pub fn estimate<T: MonotonicallyMappableToU64>(
}
}
pub fn serialize_u128(
typed_column: impl Column<u128>,
/// 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<()> {
// TODO write header, to later support more codecs
let compressor = CompactSpaceCompressor::train_from(&typed_column);
compressor
.compress_into(typed_column.iter(), output)
.unwrap();
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>),
}
// TODO Remove me
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(())
}
pub fn serialize<T: MonotonicallyMappableToU64>(
/// Serializes the column with the codec with the best estimate on the data.
pub fn serialize<T: MonotonicallyMappableToU64 + fmt::Debug>(
typed_column: impl Column<T>,
output: &mut impl io::Write,
codecs: &[FastFieldCodecType],
) -> io::Result<()> {
let column = monotonic_map_column(typed_column, T::to_u64);
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 + fmt::Debug>(
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,
@@ -174,6 +305,15 @@ pub fn serialize<T: MonotonicallyMappableToU64>(
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(())
}
@@ -225,7 +365,8 @@ fn serialize_given_codec(
Ok(())
}
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
/// Helper function to serialize a column (autodetect from all codecs) and then open it
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default + fmt::Debug>(
column: &[T],
) -> Arc<dyn Column<T>> {
let mut buffer = Vec::new();
@@ -237,6 +378,18 @@ pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
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];
@@ -250,7 +403,7 @@ mod tests {
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(), 5 + 8);
assert_eq!(buffer.len(), 3 + 5 + 8 + 4 + 2);
}
#[test]
@@ -259,7 +412,7 @@ mod tests {
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(), 5 + 7);
assert_eq!(buffer.len(), 3 + 5 + 7 + 4 + 2);
}
#[test]
@@ -269,6 +422,6 @@ mod tests {
let col = VecColumn::from(&vals[..]);
serialize(col, &mut buffer, &[FastFieldCodecType::Bitpacked]).unwrap();
// Values are stored over 3 bits.
assert_eq!(buffer.len(), 7 + (3 * 80 / 8) + 7);
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]

17
ownedbytes/src/lib.rs
View File

@@ -3,7 +3,7 @@ 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 slice.
@@ -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"]

67
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;
@@ -62,6 +63,20 @@ 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
@@ -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, "*"),
}
}

65
src/aggregation/agg_req.rs
View File

@@ -51,7 +51,10 @@ use serde::{Deserialize, Serialize};
pub use super::bucket::RangeAggregation;
use super::bucket::{HistogramAggregation, TermsAggregation};
use super::metric::{AverageAggregation, StatsAggregation};
use super::metric::{
AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
SumAggregation,
};
use super::VecWithNames;
/// The top-level aggregation request structure, which contains [`Aggregation`] and their user
@@ -237,20 +240,38 @@ impl BucketAggregationType {
/// called multi-value numeric metrics aggregation.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub enum MetricAggregation {
/// Calculates the average.
/// Computes the average of the extracted values.
#[serde(rename = "avg")]
Average(AverageAggregation),
/// Calculates stats sum, average, min, max, standard_deviation on a field.
/// Counts the number of extracted values.
#[serde(rename = "value_count")]
Count(CountAggregation),
/// Finds the maximum value.
#[serde(rename = "max")]
Max(MaxAggregation),
/// Finds the minimum value.
#[serde(rename = "min")]
Min(MinAggregation),
/// Computes a collection of statistics (`min`, `max`, `sum`, `count`, and `avg`) over the
/// extracted values.
#[serde(rename = "stats")]
Stats(StatsAggregation),
/// Computes the sum of the extracted values.
#[serde(rename = "sum")]
Sum(SumAggregation),
}
impl MetricAggregation {
fn get_fast_field_names(&self, fast_field_names: &mut HashSet<String>) {
match self {
MetricAggregation::Average(avg) => fast_field_names.insert(avg.field.to_string()),
MetricAggregation::Stats(stats) => fast_field_names.insert(stats.field.to_string()),
let fast_field_name = match self {
MetricAggregation::Average(avg) => avg.field_name(),
MetricAggregation::Count(count) => count.field_name(),
MetricAggregation::Max(max) => max.field_name(),
MetricAggregation::Min(min) => min.field_name(),
MetricAggregation::Stats(stats) => stats.field_name(),
MetricAggregation::Sum(sum) => sum.field_name(),
};
fast_field_names.insert(fast_field_name.to_string());
}
}
@@ -258,6 +279,38 @@ impl MetricAggregation {
mod tests {
use super::*;
#[test]
fn test_metric_aggregations_deser() {
let agg_req_json = r#"{
"price_avg": { "avg": { "field": "price" } },
"price_count": { "value_count": { "field": "price" } },
"price_max": { "max": { "field": "price" } },
"price_min": { "min": { "field": "price" } },
"price_stats": { "stats": { "field": "price" } },
"price_sum": { "sum": { "field": "price" } }
}"#;
let agg_req: Aggregations = serde_json::from_str(agg_req_json).unwrap();
assert!(
matches!(agg_req.get("price_avg").unwrap(), Aggregation::Metric(MetricAggregation::Average(avg)) if avg.field == "price")
);
assert!(
matches!(agg_req.get("price_count").unwrap(), Aggregation::Metric(MetricAggregation::Count(count)) if count.field == "price")
);
assert!(
matches!(agg_req.get("price_max").unwrap(), Aggregation::Metric(MetricAggregation::Max(max)) if max.field == "price")
);
assert!(
matches!(agg_req.get("price_min").unwrap(), Aggregation::Metric(MetricAggregation::Min(min)) if min.field == "price")
);
assert!(
matches!(agg_req.get("price_stats").unwrap(), Aggregation::Metric(MetricAggregation::Stats(stats)) if stats.field == "price")
);
assert!(
matches!(agg_req.get("price_sum").unwrap(), Aggregation::Metric(MetricAggregation::Sum(sum)) if sum.field == "price")
);
}
#[test]
fn serialize_to_json_test() {
let agg_req1: Aggregations = vec![(

35
src/aggregation/agg_req_with_accessor.rs
View File

@@ -8,10 +8,13 @@ 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::metric::{
AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
SumAggregation,
};
use super::segment_agg_result::BucketCount;
use super::VecWithNames;
use crate::fastfield::{type_and_cardinality, FastType, MultiValuedFastFieldReader};
use crate::fastfield::{type_and_cardinality, MultiValuedFastFieldReader};
use crate::schema::{Cardinality, Type};
use crate::{InvertedIndexReader, SegmentReader, TantivyError};
@@ -91,10 +94,7 @@ impl BucketAggregationWithAccessor {
BucketAggregationType::Terms(TermsAggregation {
field: field_name, ..
}) => {
let field = reader
.schema()
.get_field(field_name)
.ok_or_else(|| TantivyError::FieldNotFound(field_name.to_string()))?;
let field = reader.schema().get_field(field_name)?;
inverted_index = Some(reader.inverted_index(field)?);
get_ff_reader_and_validate(reader, field_name, Cardinality::MultiValues)?
}
@@ -134,7 +134,11 @@ impl MetricAggregationWithAccessor {
) -> crate::Result<MetricAggregationWithAccessor> {
match &metric {
MetricAggregation::Average(AverageAggregation { field: field_name })
| MetricAggregation::Stats(StatsAggregation { field: field_name }) => {
| MetricAggregation::Count(CountAggregation { field: field_name })
| MetricAggregation::Max(MaxAggregation { field: field_name })
| MetricAggregation::Min(MinAggregation { field: field_name })
| MetricAggregation::Stats(StatsAggregation { field: field_name })
| MetricAggregation::Sum(SumAggregation { field: field_name }) => {
let (accessor, field_type) =
get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?;
@@ -188,19 +192,10 @@ fn get_ff_reader_and_validate(
field_name: &str,
cardinality: Cardinality,
) -> crate::Result<(FastFieldAccessor, Type)> {
let field = reader
.schema()
.get_field(field_name)
.ok_or_else(|| TantivyError::FieldNotFound(field_name.to_string()))?;
let field = reader.schema().get_field(field_name)?;
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 {:?}",
@@ -217,10 +212,10 @@ fn get_ff_reader_and_validate(
let ff_fields = reader.fast_fields();
match cardinality {
Cardinality::SingleValue => ff_fields
.u64_lenient(field)
.u64_lenient(field_name)
.map(|field| (FastFieldAccessor::Single(field), field_type.value_type())),
Cardinality::MultiValues => ff_fields
.u64s_lenient(field)
.u64s_lenient(field_name)
.map(|field| (FastFieldAccessor::Multi(field), field_type.value_type())),
}
}

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