rust/tantivy
1
0
Fork 0
mirror of https://github.com/quickwit-oss/tantivy.git synced 2025-12-29 21:42:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

Compare commits

base: rust:typed-column3
Branches Tags
rust:main rust:fix_size_hint rust:bucket_id_agg rust:stuhood.upstream-erased-sort-computer rust:trinity.pointard/ff-query rust:paradedb/fix-agg-validation rust:flatheadmill-geo rust:paradedb-paradedb/fix-overflow-issue rust:low_card_optimization rust:paul.masurel/clippy 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:bugfix-position-broken
Branches Tags
rust:fix_size_hint rust:main rust:bucket_id_agg rust:stuhood.upstream-erased-sort-computer rust:trinity.pointard/ff-query rust:paradedb/fix-agg-validation rust:flatheadmill-geo rust:paradedb-paradedb/fix-overflow-issue rust:low_card_optimization rust:paul.masurel/clippy 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

1 Commits
typed-colu ... bugfix-pos

Author SHA1 Message Date
Paul Masurel
727d024a23 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 10:19:41 +09:00
272 changed files with 9578 additions and 19736 deletions
Expand all files Collapse all files

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

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

33
CHANGELOG.md
View File

@@ -1,37 +1,32 @@
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)
#### 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)
- Major bugfix: Fix missing fieldnorms for u64, i64, f64, bool, bytes and date [#1620](https://github.com/quickwit-oss/tantivy/pull/1620) (@PSeitz)
- 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). (@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)
- Validate settings on create [#1570](https://github.com/quickwit-oss/tantivy/pull/1570 (@PSeitz)
- Fix interpolation overflow in linear interpolation fastfield codec [#1480](https://github.com/quickwit-oss/tantivy/pull/1480 (@PSeitz @fulmicoton)
- 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)
- 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)
- [#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)
Tantivy 0.18
================================
@@ -49,10 +44,6 @@ 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
================================

41
Cargo.toml
View File

@@ -1,6 +1,6 @@
[package]
name = "tantivy"
version = "0.19.0"
version = "0.18.0"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
categories = ["database-implementations", "data-structures"]
@@ -14,18 +14,17 @@ edition = "2021"
rust-version = "1.62"
[dependencies]
oneshot = "0.1.5"
base64 = "0.21.0"
oneshot = "0.1.3"
base64 = "0.13.0"
byteorder = "1.4.3"
crc32fast = "1.3.2"
once_cell = "1.10.0"
regex = { version = "1.5.5", default-features = false, features = ["std", "unicode"] }
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 = "0.12", optional = true, default-features = false }
zstd = { version = "0.11", optional = true, default-features = false }
snap = { version = "1.0.5", optional = true }
tempfile = { version = "3.3.0", optional = true }
log = "0.4.16"
@@ -36,11 +35,17 @@ 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"
rustc-hash = "1.1.0"
fnv = "1.0.7"
thiserror = "1.0.30"
htmlescape = "0.3.1"
fail = "0.5.0"
@@ -48,21 +53,14 @@ murmurhash32 = "0.2.0"
time = { version = "0.3.10", features = ["serde-well-known"] }
smallvec = "1.8.0"
rayon = "1.5.2"
lru = "0.9.0"
lru = "0.7.5"
fastdivide = "0.4.0"
itertools = "0.10.3"
measure_time = "0.8.2"
ciborium = { version = "0.2", optional = true}
async-trait = "0.1.53"
arc-swap = "1.5.0"
columnar = { version="0.1", path="./columnar", package ="tantivy-columnar" }
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" }
tokenizer-api = { version="0.1", path="./tokenizer-api", package="tantivy-tokenizer-api" }
[target.'cfg(windows)'.dependencies]
winapi = "0.3.9"
@@ -72,10 +70,10 @@ maplit = "1.0.2"
matches = "0.1.9"
pretty_assertions = "1.2.1"
proptest = "1.0.0"
criterion = "0.4"
criterion = "0.3.5"
test-log = "0.2.10"
env_logger = "0.10.0"
pprof = { version = "0.11.0", features = ["flamegraph", "criterion"] }
env_logger = "0.9.0"
pprof = { version = "0.10.0", features = ["flamegraph", "criterion"] }
futures = "0.3.21"
[dev-dependencies.fail]
@@ -92,9 +90,8 @@ debug-assertions = true
overflow-checks = true
[features]
default = ["mmap", "stopwords", "lz4-compression"]
default = ["mmap", "lz4-compression" ]
mmap = ["fs2", "tempfile", "memmap2"]
stopwords = []
brotli-compression = ["brotli"]
lz4-compression = ["lz4_flex"]
@@ -104,10 +101,10 @@ zstd-compression = ["zstd"]
failpoints = ["fail/failpoints"]
unstable = [] # useful for benches.
quickwit = ["sstable"]
quickwit = ["ciborium"]
[workspace]
members = ["query-grammar", "bitpacker", "common", "ownedbytes", "stacker", "sstable", "tokenizer-api", "columnar"]
members = ["query-grammar", "bitpacker", "common", "fastfield_codecs", "ownedbytes"]
# Following the "fail" crate best practises, we isolate
# tests that define specific behavior in fail check points

51
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)
@@ -41,13 +41,13 @@ Your mileage WILL vary depending on the nature of queries and their load.
- SIMD integer compression when the platform/CPU includes the SSE2 instruction set
- Single valued and multivalued u64, i64, and f64 fast fields (equivalent of doc values in Lucene)
- `&[u8]` fast fields
- Text, i64, u64, f64, dates, ip, bool, and hierarchical facet fields
- Compressed document store (LZ4, Zstd, None, Brotli, Snap)
- Text, i64, u64, f64, dates, and hierarchical facet fields
- LZ4 compressed document store
- Range queries
- Faceted search
- Configurable indexing (optional term frequency and position indexing)
- JSON Field
- Aggregation Collector: histogram, range buckets, average, and stats metrics
- Aggregation Collector: range buckets, average, and stats metrics
- LogMergePolicy with deletes
- Searcher Warmer API
- Cheesy logo with a horse
@@ -80,11 +80,10 @@ There are many ways to support this project.
# Contributing code
We use the GitHub Pull Request workflow: reference a GitHub ticket and/or include a comprehensive commit message when opening a PR.
Feel free to update CHANGELOG.md with your contribution.
## Tokenizer
## Minimum supported Rust version
When implementing a tokenizer for tantivy depend on the `tantivy-tokenizer-api` crate.
Tantivy currently requires at least Rust 1.62 or later to compile.
## Clone and build locally
@@ -92,9 +91,41 @@ Tantivy compiles on stable Rust.
To check out and run tests, you can simply run:
```bash
git clone https://github.com/quickwit-oss/tantivy.git
cd tantivy
cargo test
git clone https://github.com/quickwit-oss/tantivy.git
cd tantivy
cargo build
```
## Run tests
Some tests will not run with just `cargo test` because of `fail-rs`.
To run the tests exhaustively, run `./run-tests.sh`.
## Debug
You might find it useful to step through the programme with a debugger.
### A failing test
Make sure you haven't run `cargo clean` after the most recent `cargo test` or `cargo build` to guarantee that the `target/` directory exists. Use this bash script to find the name of the most recent debug build of Tantivy and run it under `rust-gdb`:
```bash
find target/debug/ -maxdepth 1 -executable -type f -name "tantivy*" -printf '%TY-%Tm-%Td %TT %p\n' | sort -r | cut -d " " -f 3 | xargs -I RECENT_DBG_TANTIVY rust-gdb RECENT_DBG_TANTIVY
```
Now that you are in `rust-gdb`, you can set breakpoints on lines and methods that match your source code and run the debug executable with flags that you normally pass to `cargo test` like this:
```bash
$gdb run --test-threads 1 --test $NAME_OF_TEST
```
### An example
By default, `rustc` compiles everything in the `examples/` directory in debug mode. This makes it easy for you to make examples to reproduce bugs:
```bash
rust-gdb target/debug/examples/$EXAMPLE_NAME
$ gdb run
```
# Companies Using Tantivy

18
TODO.txt
View File

@@ -1,18 +0,0 @@
Make schema_builder API fluent.
fix doc serialization and prevent compression problems
u64 , etc. shoudl return Resutl<Option> now that we support optional missing a column is really not an error
remove fastfield codecs
ditch the first_or_default trick. if it is still useful, improve its implementation.
rename FastFieldReaders::open to load
remove fast field reader
find a way to unify the two DateTime.
readd type check in the filter wrapper
add unit test on columnar list columns.
make sure sort works

12
benches/index-bench.rs
View File

@@ -34,7 +34,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let index = Index::create_in_ram(schema.clone());
let index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let doc = schema.parse_document(doc_json).unwrap();
index_writer.add_document(doc).unwrap();
}
@@ -46,7 +46,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let index = Index::create_in_ram(schema.clone());
let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let doc = schema.parse_document(doc_json).unwrap();
index_writer.add_document(doc).unwrap();
}
@@ -59,7 +59,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let index = Index::create_in_ram(schema_with_store.clone());
let index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let doc = schema.parse_document(doc_json).unwrap();
index_writer.add_document(doc).unwrap();
}
@@ -71,7 +71,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let index = Index::create_in_ram(schema_with_store.clone());
let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let doc = schema.parse_document(doc_json).unwrap();
index_writer.add_document(doc).unwrap();
}
@@ -85,7 +85,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let json_field = dynamic_schema.get_field("json").unwrap();
let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let json_val: serde_json::Map<String, serde_json::Value> =
serde_json::from_str(doc_json).unwrap();
let doc = tantivy::doc!(json_field=>json_val);
@@ -101,7 +101,7 @@ pub fn hdfs_index_benchmark(c: &mut Criterion) {
let json_field = dynamic_schema.get_field("json").unwrap();
let mut index_writer = index.writer_with_num_threads(1, 100_000_000).unwrap();
for _ in 0..NUM_REPEATS {
for doc_json in HDFS_LOGS.trim().split('\n') {
for doc_json in HDFS_LOGS.trim().split("\n") {
let json_val: serde_json::Map<String, serde_json::Value> =
serde_json::from_str(doc_json).unwrap();
let doc = tantivy::doc!(json_field=>json_val);

8
bitpacker/Cargo.toml
View File

@@ -1,6 +1,6 @@
[package]
name = "tantivy-bitpacker"
version = "0.3.0"
version = "0.2.0"
edition = "2021"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
@@ -8,14 +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
[dependencies]
[dev-dependencies]
rand = "0.8"
proptest = "1"

36
bitpacker/benches/bench.rs
View File

@@ -4,39 +4,9 @@ extern crate test;
#[cfg(test)]
mod tests {
use rand::seq::IteratorRandom;
use rand::thread_rng;
use tantivy_bitpacker::{BitPacker, BitUnpacker, BlockedBitpacker};
use tantivy_bitpacker::BlockedBitpacker;
use test::Bencher;
#[inline(never)]
fn create_bitpacked_data(bit_width: u8, num_els: u32) -> Vec<u8> {
let mut bitpacker = BitPacker::new();
let mut buffer = Vec::new();
for _ in 0..num_els {
// the values do not matter.
bitpacker.write(0u64, bit_width, &mut buffer).unwrap();
bitpacker.flush(&mut buffer).unwrap();
}
buffer
}
#[bench]
fn bench_bitpacking_read(b: &mut Bencher) {
let bit_width = 3;
let num_els = 1_000_000u32;
let bit_unpacker = BitUnpacker::new(bit_width);
let data = create_bitpacked_data(bit_width, num_els);
let idxs: Vec<u32> = (0..num_els).choose_multiple(&mut thread_rng(), 100_000);
b.iter(|| {
let mut out = 0u64;
for &idx in &idxs {
out = out.wrapping_add(bit_unpacker.get(idx, &data[..]));
}
out
});
}
#[bench]
fn bench_blockedbitp_read(b: &mut Bencher) {
let mut blocked_bitpacker = BlockedBitpacker::new();
@@ -44,9 +14,9 @@ mod tests {
blocked_bitpacker.add(val * val);
}
b.iter(|| {
let mut out = 0u64;
let mut out = 0;
for val in 0..=21500 {
out = out.wrapping_add(blocked_bitpacker.get(val));
out = blocked_bitpacker.get(val);
}
out
});

110
bitpacker/src/bitpacker.rs
View File

@@ -25,14 +25,15 @@ 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.wrapping_shl(self.mini_buffer_written as u32);
self.mini_buffer |= val_u64.wrapping_shl(self.mini_buffer_written as u32);
output.write_all(self.mini_buffer.to_le_bytes().as_ref())?;
self.mini_buffer = val.wrapping_shr((64 - self.mini_buffer_written) as u32);
self.mini_buffer = val_u64.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 << self.mini_buffer_written;
self.mini_buffer |= val_u64 << 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())?;
@@ -56,31 +57,27 @@ impl BitPacker {
pub fn close<TWrite: io::Write>(&mut self, output: &mut TWrite) -> io::Result<()> {
self.flush(output)?;
// Padding the write file to simplify reads.
output.write_all(&[0u8; 7])?;
Ok(())
}
}
#[derive(Clone, Debug, Default, Copy)]
#[derive(Clone, Debug, Default)]
pub struct BitUnpacker {
num_bits: u32,
num_bits: u64,
mask: u64,
}
impl BitUnpacker {
/// Creates a bit unpacker, that assumes the same bitwidth for all values.
///
/// The bitunpacker works by doing an unaligned read of 8 bytes.
/// For this reason, values of `num_bits` between
/// [57..63] are forbidden.
pub fn new(num_bits: u8) -> BitUnpacker {
assert!(num_bits <= 7 * 8 || num_bits == 64);
let mask: u64 = if num_bits == 64 {
!0u64
} else {
(1u64 << num_bits) - 1u64
};
BitUnpacker {
num_bits: u32::from(num_bits),
num_bits: u64::from(num_bits),
mask,
}
}
@@ -90,32 +87,22 @@ impl BitUnpacker {
}
#[inline]
pub fn get(&self, idx: u32, data: &[u8]) -> u64 {
let addr_in_bits = idx * self.num_bits;
let addr = (addr_in_bits >> 3) as usize;
if addr + 8 > data.len() {
if self.num_bits == 0 {
return 0;
}
let bit_shift = addr_in_bits & 7;
return self.get_slow_path(addr, bit_shift, data);
pub fn get(&self, idx: u64, 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 bit_shift = addr_in_bits & 7;
let bytes: [u8; 8] = (&data[addr..addr + 8]).try_into().unwrap();
debug_assert!(
addr + 8 <= data.len() as u64,
"The fast field field should have been padded with 7 bytes."
);
let bytes: [u8; 8] = (&data[(addr as usize)..(addr as usize) + 8])
.try_into()
.unwrap();
let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
let val_shifted = val_unshifted_unmasked >> bit_shift;
val_shifted & self.mask
}
#[inline(never)]
fn get_slow_path(&self, addr: usize, bit_shift: u32, data: &[u8]) -> u64 {
let mut bytes: [u8; 8] = [0u8; 8];
let available_bytes = data.len() - addr;
// This function is meant to only be called if we did not have 8 bytes to load.
debug_assert!(available_bytes < 8);
bytes[..available_bytes].copy_from_slice(&data[addr..]);
let val_unshifted_unmasked: u64 = u64::from_le_bytes(bytes);
let val_shifted = val_unshifted_unmasked >> bit_shift;
let val_shifted = (val_unshifted_unmasked >> bit_shift) as u64;
val_shifted & self.mask
}
}
@@ -124,7 +111,7 @@ impl BitUnpacker {
mod test {
use super::{BitPacker, BitUnpacker};
fn create_bitpacker(len: usize, num_bits: u8) -> (BitUnpacker, Vec<u64>, Vec<u8>) {
fn create_fastfield_bitpacker(len: usize, num_bits: u8) -> (BitUnpacker, Vec<u64>, Vec<u8>) {
let mut data = Vec::new();
let mut bitpacker = BitPacker::new();
let max_val: u64 = (1u64 << num_bits as u64) - 1u64;
@@ -135,15 +122,15 @@ mod test {
bitpacker.write(val, num_bits, &mut data).unwrap();
}
bitpacker.close(&mut data).unwrap();
assert_eq!(data.len(), ((num_bits as usize) * len + 7) / 8);
assert_eq!(data.len(), ((num_bits as usize) * len + 7) / 8 + 7);
let bitunpacker = BitUnpacker::new(num_bits);
(bitunpacker, vals, data)
}
fn test_bitpacker_util(len: usize, num_bits: u8) {
let (bitunpacker, vals, data) = create_bitpacker(len, num_bits);
let (bitunpacker, vals, data) = create_fastfield_bitpacker(len, num_bits);
for (i, val) in vals.iter().enumerate() {
assert_eq!(bitunpacker.get(i as u32, &data), *val);
assert_eq!(bitunpacker.get(i as u64, &data), *val);
}
}
@@ -155,49 +142,4 @@ mod test {
test_bitpacker_util(6, 14);
test_bitpacker_util(1000, 14);
}
use proptest::prelude::*;
fn num_bits_strategy() -> impl Strategy<Value = u8> {
prop_oneof!(Just(0), Just(1), 2u8..56u8, Just(56), Just(64),)
}
fn vals_strategy() -> impl Strategy<Value = (u8, Vec<u64>)> {
(num_bits_strategy(), 0usize..100usize).prop_flat_map(|(num_bits, len)| {
let max_val = if num_bits == 64 {
u64::MAX
} else {
(1u64 << num_bits as u32) - 1
};
let vals = proptest::collection::vec(0..=max_val, len);
vals.prop_map(move |vals| (num_bits, vals))
})
}
fn test_bitpacker_aux(num_bits: u8, vals: &[u64]) {
let mut buffer: Vec<u8> = Vec::new();
let mut bitpacker = BitPacker::new();
for &val in vals {
bitpacker.write(val, num_bits, &mut buffer).unwrap();
}
bitpacker.flush(&mut buffer).unwrap();
assert_eq!(buffer.len(), (vals.len() * num_bits as usize + 7) / 8);
let bitunpacker = BitUnpacker::new(num_bits);
let max_val = if num_bits == 64 {
u64::MAX
} else {
(1u64 << num_bits) - 1
};
for (i, val) in vals.iter().copied().enumerate() {
assert!(val <= max_val);
assert_eq!(bitunpacker.get(i as u32, &buffer), val);
}
}
proptest::proptest! {
#[test]
fn test_bitpacker_proptest((num_bits, vals) in vals_strategy()) {
test_bitpacker_aux(num_bits, &vals);
}
}
}

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 {
if self.buffer.len() == BLOCK_SIZE as usize {
self.flush();
}
}
@@ -126,11 +126,11 @@ impl BlockedBitpacker {
}
#[inline]
pub fn get(&self, idx: usize) -> u64 {
let metadata_pos = idx / BLOCK_SIZE;
let pos_in_block = idx % BLOCK_SIZE;
let metadata_pos = idx / BLOCK_SIZE as usize;
let pos_in_block = idx % BLOCK_SIZE as usize;
if let Some(metadata) = self.offset_and_bits.get(metadata_pos) {
let unpacked = BitUnpacker::new(metadata.num_bits()).get(
pos_in_block as u32,
pos_in_block as u64,
&self.compressed_blocks[metadata.offset() as usize..],
);
unpacked + metadata.base_value()

128
bitpacker/src/lib.rs
View File

@@ -1,8 +1,6 @@
mod bitpacker;
mod blocked_bitpacker;
use std::cmp::Ordering;
pub use crate::bitpacker::{BitPacker, BitUnpacker};
pub use crate::blocked_bitpacker::BlockedBitpacker;
@@ -39,104 +37,44 @@ 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 + PartialOrd,
T: Copy + Ord,
{
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;
}
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))
}));
}
Some((min_so_far, max_so_far))
None
}
#[cfg(test)]
mod tests {
use super::*;
#[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_empty() {
let vals: Vec<u32> = vec![];
assert_eq!(minmax(vals.into_iter()), None);
}
#[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))
);
}
#[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_empty() {
let vals: Vec<u32> = vec![];
assert_eq!(minmax(vals.into_iter()), None);
}
#[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)));
}

27
columnar/Cargo.toml
View File

@@ -1,27 +0,0 @@
[package]
name = "tantivy-columnar"
version = "0.1.0"
edition = "2021"
license = "MIT"
[dependencies]
itertools = "0.10.5"
log = "0.4.17"
fnv = "1.0.7"
fastdivide = "0.4.0"
rand = { version = "0.8.5", optional = true }
measure_time = { version = "0.8.2", optional = true }
prettytable-rs = { version = "0.10.0", optional = true }
stacker = { path = "../stacker", package="tantivy-stacker"}
sstable = { path = "../sstable", package = "tantivy-sstable" }
common = { path = "../common", package = "tantivy-common" }
tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
[dev-dependencies]
proptest = "1"
more-asserts = "0.3.0"
rand = "0.8.3"
[features]
unstable = []

6
columnar/Makefile
View File

@@ -1,6 +0,0 @@
test:
echo "Run test only... No examples."
cargo test --tests --lib
fmt:
cargo +nightly fmt --all

109
columnar/README.md
View File

@@ -1,109 +0,0 @@
# 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

311
columnar/benches/bench.rs
View File

@@ -1,311 +0,0 @@
#![feature(test)]
extern crate test;
#[cfg(test)]
mod tests {
use std::ops::RangeInclusive;
use std::sync::Arc;
use common::OwnedBytes;
use rand::prelude::*;
use tantivy_columnar::*;
use test::Bencher;
use super::*;
// Warning: this generates the same permutation at each call
fn generate_permutation() -> Vec<u64> {
let mut permutation: Vec<u64> = (0u64..100_000u64).collect();
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
fn generate_random() -> Vec<u64> {
let mut permutation: Vec<u64> = (0u64..100_000u64)
.map(|el| el + random::<u16>() as u64)
.collect();
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
// Warning: this generates the same permutation at each call
fn generate_permutation_gcd() -> Vec<u64> {
let mut permutation: Vec<u64> = (1u64..100_000u64).map(|el| el * 1000).collect();
permutation.shuffle(&mut StdRng::from_seed([1u8; 32]));
permutation
}
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
column: &[T],
) -> Arc<dyn Column<T>> {
let mut buffer = Vec::new();
serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
open(OwnedBytes::new(buffer)).unwrap()
}
#[bench]
fn bench_intfastfield_jumpy_veclookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = permutation[a as usize];
}
a
});
}
#[bench]
fn bench_intfastfield_jumpy_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = column.get_val(a as u32);
}
a
});
}
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 _ in 0..300_000 {
let val = rng.gen_range(1..=100);
data.push(val);
}
data.push(SINGLE_ITEM);
data.shuffle(&mut rng);
let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
data
}
fn get_u128_column_random() -> Arc<dyn Column<u128>> {
let permutation = generate_random();
let permutation = permutation.iter().map(|el| *el as u128).collect::<Vec<_>>();
get_u128_column_from_data(&permutation)
}
fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
let mut out = vec![];
let iter_gen = || data.iter().cloned();
serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
let out = OwnedBytes::new(out);
open_u128::<u128>(out).unwrap()
}
// 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 data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*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 data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*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 data = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
positions
});
}
// U128 RANGE END
#[bench]
fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
let column = get_u128_column_random();
b.iter(|| {
let mut a = 0u128;
for i in 0u64..column.num_vals() as u64 {
a += column.get_val(i as u32);
}
a
});
}
#[bench]
fn bench_intfastfield_jumpy_stride5_u128(b: &mut Bencher) {
let column = get_u128_column_random();
b.iter(|| {
let n = column.num_vals();
let mut a = 0u128;
for i in (0..n / 5).map(|val| val * 5) {
a += column.get_val(i);
}
a
});
}
#[bench]
fn bench_intfastfield_stride7_vec(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
b.iter(|| {
let mut a = 0u64;
for i in (0..n / 7).map(|val| val * 7) {
a += permutation[i as usize];
}
a
});
}
#[bench]
fn bench_intfastfield_stride7_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0;
for i in (0..n / 7).map(|val| val * 7) {
a += column.get_val(i as u32);
}
a
});
}
#[bench]
fn bench_intfastfield_scan_all_fflookup(b: &mut Bencher) {
let permutation = generate_permutation();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0u32..n as u32 {
a += column.get_val(i);
}
a
});
}
#[bench]
fn bench_intfastfield_scan_all_fflookup_gcd(b: &mut Bencher) {
let permutation = generate_permutation_gcd();
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0..n {
a += column.get_val(i as u32);
}
a
});
}
#[bench]
fn bench_intfastfield_scan_all_vec(b: &mut Bencher) {
let permutation = generate_permutation();
b.iter(|| {
let mut a = 0u64;
for i in 0..permutation.len() {
a += permutation[i as usize] as u64;
}
a
});
}
}

57
columnar/src/TODO.md
View File

@@ -1,57 +0,0 @@
# 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
* replug facets
* replug range queries
+ mutlivaued range queries restrat frm the beginning all of the time.
# 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
replug examples
replug fast_field_codecs bench
# Cleanup and rationalization
remove the 6 bit limitation of columntype. use 4 + 4 bits instead.
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
rename columnar to something more explicit, like column_dictionary or columnar_table
remove old column from the fast field API.
remove the Column traits alias.
rename fastfield -> column
document changes
rationalization FastFieldValue, HasColumnType
# Other
fix enhance column-cli
# Santa claus
autodetect datetime ipaddr, plug customizable tokenizer.

96
columnar/src/column/dictionary_encoded.rs
View File

@@ -1,96 +0,0 @@
use std::io;
use std::ops::Deref;
use std::sync::Arc;
use sstable::{Dictionary, VoidSSTable};
use crate::column::Column;
use crate::RowId;
/// Dictionary encoded column.
///
/// The column simply gives access to a regular u64-column that, in
/// which the values are term-ordinals.
///
/// These ordinals are ids uniquely identify the bytes that are stored in
/// the column. These ordinals are small, and sorted in the same order
/// as the term_ord_column.
#[derive(Clone)]
pub struct BytesColumn {
pub(crate) dictionary: Arc<Dictionary<VoidSSTable>>,
pub(crate) term_ord_column: Column<u64>,
}
impl BytesColumn {
/// Fills the given `output` buffer with the term associated to the ordinal `ord`.
///
/// Returns `false` if the term does not exist (e.g. `term_ord` is greater or equal to the
/// overll number of terms).
pub fn ord_to_bytes(&self, ord: u64, output: &mut Vec<u8>) -> io::Result<bool> {
self.dictionary.ord_to_term(ord, output)
}
/// Returns the number of rows in the column.
pub fn num_rows(&self) -> RowId {
self.term_ord_column.num_rows()
}
pub fn term_ords(&self, row_id: RowId) -> impl Iterator<Item = u64> + '_ {
self.term_ord_column.values(row_id)
}
/// Returns the column of ordinals
pub fn ords(&self) -> &Column<u64> {
&self.term_ord_column
}
pub fn num_terms(&self) -> usize {
self.dictionary.num_terms()
}
pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
self.dictionary.as_ref()
}
}
#[derive(Clone)]
pub struct StrColumn(BytesColumn);
impl From<BytesColumn> for StrColumn {
fn from(bytes_col: BytesColumn) -> Self {
StrColumn(bytes_col)
}
}
impl StrColumn {
pub fn dictionary(&self) -> &Dictionary<VoidSSTable> {
self.0.dictionary.as_ref()
}
/// Fills the buffer
pub fn ord_to_str(&self, term_ord: u64, output: &mut String) -> io::Result<bool> {
unsafe {
let buf = output.as_mut_vec();
if !self.0.dictionary.ord_to_term(term_ord, buf)? {
return Ok(false);
}
// TODO consider remove checks if it hurts performance.
if std::str::from_utf8(buf.as_slice()).is_err() {
buf.clear();
return Err(io::Error::new(
io::ErrorKind::InvalidData,
"Not valid utf-8",
));
}
}
Ok(true)
}
}
impl Deref for StrColumn {
type Target = BytesColumn;
fn deref(&self) -> &Self::Target {
&self.0
}
}

111
columnar/src/column/mod.rs
View File

@@ -1,111 +0,0 @@
mod dictionary_encoded;
mod serialize;
use std::fmt::Debug;
use std::ops::Deref;
use std::sync::Arc;
use common::BinarySerializable;
pub use dictionary_encoded::{BytesColumn, StrColumn};
pub use serialize::{
open_column_bytes, open_column_u128, open_column_u64, serialize_column_mappable_to_u128,
serialize_column_mappable_to_u64,
};
use crate::column_index::ColumnIndex;
use crate::column_values::ColumnValues;
use crate::{Cardinality, RowId};
#[derive(Clone)]
pub struct Column<T> {
pub idx: ColumnIndex,
pub values: Arc<dyn ColumnValues<T>>,
}
impl<T: PartialOrd + Copy + Debug + Send + Sync + 'static> Column<T> {
pub fn num_rows(&self) -> RowId {
match &self.idx {
ColumnIndex::Full => self.values.num_vals() as u32,
ColumnIndex::Optional(optional_index) => optional_index.num_rows(),
ColumnIndex::Multivalued(col_index) => {
// The multivalued index contains all value start row_id,
// and one extra value at the end with the overall number of rows.
col_index.num_rows()
}
}
}
pub fn min_value(&self) -> T {
self.values.min_value()
}
pub fn max_value(&self) -> T {
self.values.max_value()
}
pub fn first(&self, row_id: RowId) -> Option<T> {
self.values(row_id).next()
}
pub fn values(&self, row_id: RowId) -> impl Iterator<Item = T> + '_ {
self.value_row_ids(row_id)
.map(|value_row_id: RowId| self.values.get_val(value_row_id))
}
pub fn first_or_default_col(self, default_value: T) -> Arc<dyn ColumnValues<T>> {
Arc::new(FirstValueWithDefault {
column: self,
default_value,
})
}
}
impl<T> Deref for Column<T> {
type Target = ColumnIndex;
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)
}
}
// TODO simplify or optimize
struct FirstValueWithDefault<T: Copy> {
column: Column<T>,
default_value: T,
}
impl<T: PartialOrd + Debug + Send + Sync + Copy + 'static> ColumnValues<T>
for FirstValueWithDefault<T>
{
fn get_val(&self, idx: u32) -> T {
self.column.first(idx).unwrap_or(self.default_value)
}
fn min_value(&self) -> T {
self.column.values.min_value()
}
fn max_value(&self) -> T {
self.column.values.max_value()
}
fn num_vals(&self) -> u32 {
match &self.column.idx {
ColumnIndex::Full => self.column.values.num_vals(),
ColumnIndex::Optional(optional_idx) => optional_idx.num_rows(),
ColumnIndex::Multivalued(_) => todo!(),
}
}
}

102
columnar/src/column/serialize.rs
View File

@@ -1,102 +0,0 @@
use std::fmt::Debug;
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::OwnedBytes;
use sstable::Dictionary;
use crate::column::{BytesColumn, Column};
use crate::column_index::{serialize_column_index, SerializableColumnIndex};
use crate::column_values::serialize::serialize_column_values_u128;
use crate::column_values::{
serialize_column_values, ColumnValues, FastFieldCodecType, MonotonicallyMappableToU128,
MonotonicallyMappableToU64,
};
pub fn serialize_column_mappable_to_u128<
F: Fn() -> I,
I: Iterator<Item = T>,
T: MonotonicallyMappableToU128,
>(
column_index: SerializableColumnIndex<'_>,
column_values: F,
num_vals: u32,
output: &mut impl Write,
) -> io::Result<()> {
let column_index_num_bytes = serialize_column_index(column_index, output)?;
serialize_column_values_u128(
|| column_values().map(|val| val.to_u128()),
num_vals,
output,
)?;
output.write_all(&column_index_num_bytes.to_le_bytes())?;
Ok(())
}
pub fn serialize_column_mappable_to_u64<T: MonotonicallyMappableToU64 + Debug>(
column_index: SerializableColumnIndex<'_>,
column_values: &impl ColumnValues<T>,
output: &mut impl Write,
) -> io::Result<()> {
let column_index_num_bytes = serialize_column_index(column_index, output)?;
serialize_column_values(
column_values,
&[
FastFieldCodecType::Bitpacked,
FastFieldCodecType::BlockwiseLinear,
],
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_u128<T: MonotonicallyMappableToU128>(
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_u128_mapped(column_values_data)?;
Ok(Column {
idx: column_index,
values: column_values,
})
}
pub fn open_column_bytes<T: From<BytesColumn>>(data: OwnedBytes) -> io::Result<T> {
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)?;
let bytes_column = BytesColumn {
dictionary,
term_ord_column,
};
Ok(bytes_column.into())
}

60
columnar/src/column_index/mod.rs
View File

@@ -1,60 +0,0 @@
mod multivalued_index;
mod optional_index;
mod serialize;
use std::ops::Range;
pub use optional_index::{OptionalIndex, SerializableOptionalIndex, Set};
pub use serialize::{open_column_index, serialize_column_index, SerializableColumnIndex};
use crate::column_index::multivalued_index::MultiValueIndex;
use crate::{Cardinality, RowId};
#[derive(Clone)]
pub enum ColumnIndex {
Full,
Optional(OptionalIndex),
/// In addition, at index num_rows, an extra value is added
/// containing the overal number of values.
Multivalued(MultiValueIndex),
}
impl ColumnIndex {
pub fn get_cardinality(&self) -> Cardinality {
match self {
ColumnIndex::Full => Cardinality::Full,
ColumnIndex::Optional(_) => Cardinality::Optional,
ColumnIndex::Multivalued(_) => Cardinality::Multivalued,
}
}
pub fn value_row_ids(&self, row_id: RowId) -> Range<RowId> {
match self {
ColumnIndex::Full => row_id..row_id + 1,
ColumnIndex::Optional(optional_index) => {
if let Some(val) = optional_index.rank_if_exists(row_id) {
val..val + 1
} else {
0..0
}
}
ColumnIndex::Multivalued(multivalued_index) => multivalued_index.range(row_id),
}
}
pub fn select_batch_in_place(&self, rank_ids: &mut Vec<RowId>) {
match self {
ColumnIndex::Full => {
// No need to do anything:
// value_idx and row_idx are the same.
}
ColumnIndex::Optional(optional_index) => {
optional_index.select_batch(&mut rank_ids[..]);
}
ColumnIndex::Multivalued(multivalued_index) => {
// TODO important: avoid using 0u32, and restart from the beginning all of the time.
multivalued_index.select_batch_in_place(0u32, rank_ids)
}
}
}
}

132
columnar/src/column_index/multivalued_index.rs
View File

@@ -1,132 +0,0 @@
use std::io;
use std::io::Write;
use std::ops::Range;
use std::sync::Arc;
use common::OwnedBytes;
use crate::column_values::{ColumnValues, FastFieldCodecType};
use crate::RowId;
pub fn serialize_multivalued_index(
multivalued_index: &dyn ColumnValues<RowId>,
output: &mut impl Write,
) -> io::Result<()> {
crate::column_values::serialize_column_values(
&*multivalued_index,
&[FastFieldCodecType::Bitpacked, FastFieldCodecType::Linear],
output,
)?;
Ok(())
}
pub fn open_multivalued_index(bytes: OwnedBytes) -> io::Result<MultiValueIndex> {
let start_index_column: Arc<dyn ColumnValues<RowId>> =
crate::column_values::open_u64_mapped(bytes)?;
Ok(MultiValueIndex { start_index_column })
}
#[derive(Clone)]
/// Index to resolve value range for given doc_id.
/// Starts at 0.
pub struct MultiValueIndex {
start_index_column: Arc<dyn crate::ColumnValues<RowId>>,
}
impl From<Arc<dyn ColumnValues<RowId>>> for MultiValueIndex {
fn from(start_index_column: Arc<dyn ColumnValues<RowId>>) -> Self {
MultiValueIndex { start_index_column }
}
}
impl MultiValueIndex {
/// Returns `[start, end)`, such that the values associated with
/// the given document are `start..end`.
#[inline]
pub(crate) fn range(&self, row_id: RowId) -> Range<RowId> {
let start = self.start_index_column.get_val(row_id);
let end = self.start_index_column.get_val(row_id + 1);
start..end
}
/// Returns the number of documents in the index.
#[inline]
pub fn num_rows(&self) -> u32 {
self.start_index_column.num_vals() - 1
}
/// Converts a list of ranks (row ids of values) in a 1:n index to the corresponding list of
/// row_ids. Positions are converted inplace to docids.
///
/// Since there is no index for value pos -> docid, but docid -> value pos range, we scan the
/// index.
///
/// Correctness: positions needs to be sorted. idx_reader needs to contain monotonically
/// increasing positions.
///
/// TODO: Instead of a linear scan we can employ a exponential search into binary search to
/// match a docid to its value position.
#[allow(clippy::bool_to_int_with_if)]
pub(crate) fn select_batch_in_place(&self, row_start: RowId, ranks: &mut Vec<u32>) {
if ranks.is_empty() {
return;
}
let mut cur_doc = row_start;
let mut last_doc = None;
assert!(self.start_index_column.get_val(row_start) as u32 <= ranks[0]);
let mut write_doc_pos = 0;
for i in 0..ranks.len() {
let pos = ranks[i];
loop {
let end = self.start_index_column.get_val(cur_doc + 1) as u32;
if end > pos {
ranks[write_doc_pos] = cur_doc;
write_doc_pos += if last_doc == Some(cur_doc) { 0 } else { 1 };
last_doc = Some(cur_doc);
break;
}
cur_doc += 1;
}
}
ranks.truncate(write_doc_pos);
}
}
#[cfg(test)]
mod tests {
use std::ops::Range;
use std::sync::Arc;
use super::MultiValueIndex;
use crate::column_values::IterColumn;
use crate::{ColumnValues, RowId};
fn index_to_pos_helper(
index: &MultiValueIndex,
doc_id_range: Range<u32>,
positions: &[u32],
) -> Vec<u32> {
let mut positions = positions.to_vec();
index.select_batch_in_place(doc_id_range.start, &mut positions);
positions
}
#[test]
fn test_positions_to_docid() {
let offsets: Vec<RowId> = vec![0, 10, 12, 15, 22, 23]; // docid values are [0..10, 10..12, 12..15, etc.]
let column: Arc<dyn ColumnValues<RowId>> = Arc::new(IterColumn::from(offsets.into_iter()));
let index = MultiValueIndex::from(column);
assert_eq!(index.num_rows(), 5);
let positions = &[10u32, 11, 15, 20, 21, 22];
assert_eq!(index_to_pos_helper(&index, 0..5, positions), vec![1, 3, 4]);
assert_eq!(index_to_pos_helper(&index, 1..5, positions), vec![1, 3, 4]);
assert_eq!(index_to_pos_helper(&index, 0..5, &[9]), vec![0]);
assert_eq!(index_to_pos_helper(&index, 1..5, &[10]), vec![1]);
assert_eq!(index_to_pos_helper(&index, 1..5, &[11]), vec![1]);
assert_eq!(index_to_pos_helper(&index, 2..5, &[12]), vec![2]);
assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14]), vec![2]);
assert_eq!(index_to_pos_helper(&index, 2..5, &[12, 14, 15]), vec![2, 3]);
}
}

498
columnar/src/column_index/optional_index/mod.rs
View File

@@ -1,498 +0,0 @@
use std::io::{self, Write};
use std::ops::Range;
use std::sync::Arc;
mod set;
mod set_block;
use common::{BinarySerializable, OwnedBytes, VInt};
pub use set::{SelectCursor, 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,
}
}
enum BlockSelectCursor<'a> {
Dense(<DenseBlock<'a> as Set<u16>>::SelectCursor<'a>),
Sparse(<SparseBlock<'a> as Set<u16>>::SelectCursor<'a>),
}
impl<'a> BlockSelectCursor<'a> {
fn select(&mut self, rank: u16) -> u16 {
match self {
BlockSelectCursor::Dense(dense_select_cursor) => dense_select_cursor.select(rank),
BlockSelectCursor::Sparse(sparse_select_cursor) => sparse_select_cursor.select(rank),
}
}
}
pub struct OptionalIndexSelectCursor<'a> {
current_block_cursor: BlockSelectCursor<'a>,
current_block_id: u16,
// The current block is guaranteed to contain ranks < end_rank.
current_block_end_rank: RowId,
optional_index: &'a OptionalIndex,
block_doc_idx_start: RowId,
num_null_rows_before_block: RowId,
}
impl<'a> OptionalIndexSelectCursor<'a> {
fn search_and_load_block(&mut self, rank: RowId) {
if rank < self.current_block_end_rank {
// we are already in the right block
return;
}
self.current_block_id = self.optional_index.find_block(rank, self.current_block_id);
self.current_block_end_rank = self
.optional_index
.block_metas
.get(self.current_block_id as usize + 1)
.map(|block_meta| block_meta.non_null_rows_before_block)
.unwrap_or(u32::MAX);
self.block_doc_idx_start = (self.current_block_id as u32) * ELEMENTS_PER_BLOCK;
let block_meta = self.optional_index.block_metas[self.current_block_id as usize];
self.num_null_rows_before_block = block_meta.non_null_rows_before_block;
let block: Block<'_> = self.optional_index.block(block_meta);
self.current_block_cursor = match block {
Block::Dense(dense_block) => BlockSelectCursor::Dense(dense_block.select_cursor()),
Block::Sparse(sparse_block) => BlockSelectCursor::Sparse(sparse_block.select_cursor()),
};
}
}
impl<'a> SelectCursor<RowId> for OptionalIndexSelectCursor<'a> {
fn select(&mut self, rank: RowId) -> RowId {
self.search_and_load_block(rank);
let index_in_block = (rank - self.num_null_rows_before_block) as u16;
self.current_block_cursor.select(index_in_block) as RowId + self.block_doc_idx_start
}
}
impl Set<RowId> for OptionalIndex {
type SelectCursor<'b> = OptionalIndexSelectCursor<'b> where Self: 'b;
// 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 as u32) * 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_cursor<'b>(&'b self) -> OptionalIndexSelectCursor<'b> {
OptionalIndexSelectCursor {
current_block_cursor: BlockSelectCursor::Sparse(
SparseBlockCodec::open(b"").select_cursor(),
),
current_block_id: 0u16,
current_block_end_rank: 0u32, //< this is sufficient to force the first load
optional_index: self,
block_doc_idx_start: 0u32,
num_null_rows_before_block: 0u32,
}
}
}
impl OptionalIndex {
pub fn select_batch(&self, ranks: &mut [RowId]) {
let mut select_cursor = self.select_cursor();
for rank in ranks.iter_mut() {
*rank = select_cursor.select(*rank);
}
}
#[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: u16) -> u16 {
for block_pos in start_block_pos..self.block_metas.len() as u16 {
let offset = self.block_metas[block_pos as usize].non_null_rows_before_block;
if offset > dense_idx {
return block_pos - 1u16;
}
}
self.block_metas.len() as u16 - 1u16
}
// 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;

44
columnar/src/column_index/optional_index/set.rs
View File

@@ -1,44 +0,0 @@
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>;
}
/// Stateful object that makes it possible to compute several select in a row,
/// provided the rank passed as argument are increasing.
pub trait SelectCursor<T> {
// May panic if rank is greater than the number of elements in the Set,
// or if rank is < than value provided in the previous call.
fn select(&mut self, rank: T) -> T;
}
pub trait Set<T> {
type SelectCursor<'b>: SelectCursor<T>
where Self: 'b;
/// 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;
/// Creates a brand new select cursor.
fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b>;
}

269
columnar/src/column_index/optional_index/set_block/dense.rs
View File

@@ -1,269 +0,0 @@
use std::convert::TryInto;
use std::io::{self, Write};
use common::BinarySerializable;
use crate::column_index::optional_index::{SelectCursor, 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]);
pub struct DenseBlockSelectCursor<'a> {
block_id: u16,
dense_block: DenseBlock<'a>,
}
impl<'a> SelectCursor<u16> for DenseBlockSelectCursor<'a> {
#[inline]
fn select(&mut self, rank: u16) -> u16 {
self.block_id = self
.dense_block
.find_miniblock_containing_rank(rank, self.block_id)
.unwrap();
let index_block = self.dense_block.mini_block(self.block_id);
let in_block_rank = rank - index_block.rank;
self.block_id * ELEMENTS_PER_MINI_BLOCK + select_u64(index_block.bitvec, in_block_rank)
}
}
impl<'a> Set<u16> for DenseBlock<'a> {
type SelectCursor<'b> = DenseBlockSelectCursor<'a> where Self: 'b;
#[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)
}
#[inline(always)]
fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
DenseBlockSelectCursor {
block_id: 0,
dense_block: *self,
}
}
}
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);
}
}

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

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

106
columnar/src/column_index/optional_index/set_block/sparse.rs
View File

@@ -1,106 +0,0 @@
use crate::column_index::optional_index::{SelectCursor, 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> SelectCursor<u16> for SparseBlock<'a> {
#[inline]
fn select(&mut self, rank: u16) -> u16 {
<SparseBlock<'a> as Set<u16>>::select(self, rank)
}
}
impl<'a> Set<u16> for SparseBlock<'a> {
type SelectCursor<'b> = Self where Self: 'b;
#[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())
}
#[inline(always)]
fn select_cursor<'b>(&'b self) -> Self::SelectCursor<'b> {
*self
}
}
#[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)
}
}

106
columnar/src/column_index/optional_index/set_block/tests.rs
View File

@@ -1,106 +0,0 @@
use std::collections::HashMap;
use crate::column_index::optional_index::set_block::{
DenseBlockCodec, SparseBlockCodec, DENSE_BLOCK_NUM_BYTES,
};
use crate::column_index::optional_index::{SelectCursor, 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! {
#![proptest_config(ProptestConfig::with_cases(1))]
#[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));
let mut select_cursor = tested_set.select_cursor();
assert_eq!(select_cursor.select(0), 1);
assert_eq!(select_cursor.select(1), 3);
assert_eq!(select_cursor.select(2), 17);
}
#[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));
let mut select_cursor = tested_set.select_cursor();
assert_eq!(SelectCursor::select(&mut select_cursor, 0), 1);
assert_eq!(SelectCursor::select(&mut select_cursor, 1), 3);
assert_eq!(SelectCursor::select(&mut select_cursor, 2), 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 mut select_cursor = tested_set.select_cursor();
for i in 0..150 {
assert_eq!(i, select_cursor.select(i));
}
}

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

@@ -1,327 +0,0 @@
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 select_cursor = null_index.select_cursor();
for (rank, el) in ranks.iter().copied().zip(els.iter().copied()) {
assert_eq!(select_cursor.select(rank), el);
}
}
#[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 mut select_iter = null_index.select_cursor();
for i in 0..orig_idx_with_value.len() {
assert_eq!(select_iter.select(i as u32), orig_idx_with_value[i]);
}
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 select_cursor = null_index.select_cursor();
assert_eq!(select_cursor.select(0), 0);
assert_eq!(select_cursor.select(1), 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(|| {
output.copy_from_slice(&idxs[..]);
codec.select_batch(&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);
}
}

73
columnar/src/column_index/serialize.rs
View File

@@ -1,73 +0,0 @@
use std::io;
use std::io::Write;
use common::{CountingWriter, OwnedBytes};
use crate::column_index::multivalued_index::serialize_multivalued_index;
use crate::column_index::optional_index::serialize_optional_index;
use crate::column_index::{ColumnIndex, SerializableOptionalIndex};
use crate::column_values::ColumnValues;
use crate::{Cardinality, RowId};
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(Box<dyn ColumnValues<RowId> + 'a>),
}
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<u32> {
let mut output = CountingWriter::wrap(output);
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, &mut output)?
}
SerializableColumnIndex::Multivalued(multivalued_index) => {
serialize_multivalued_index(&*multivalued_index, &mut output)?
}
}
let column_index_num_bytes = output.written_bytes() as u32;
Ok(column_index_num_bytes)
}
pub fn open_column_index(mut bytes: OwnedBytes) -> io::Result<ColumnIndex> {
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 multivalue_index = super::multivalued_index::open_multivalued_index(bytes)?;
Ok(ColumnIndex::Multivalued(multivalue_index))
}
}
}
// TODO unit tests

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

@@ -1,75 +0,0 @@
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);
}
}

326
columnar/src/column_values/mod.rs
View File

@@ -1,326 +0,0 @@
#![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::fmt::Debug;
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::{BinarySerializable, OwnedBytes};
use compact_space::CompactSpaceDecompressor;
pub use monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
use monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
};
pub use monotonic_mapping_u128::MonotonicallyMappableToU128;
use serialize::{Header, U128Header};
mod bitpacked;
mod blockwise_linear;
mod compact_space;
mod line;
mod linear;
pub(crate) mod monotonic_mapping;
pub(crate) mod monotonic_mapping_u128;
mod column;
mod gcd;
pub mod serialize;
pub use self::column::{monotonic_map_column, ColumnValues, IterColumn, VecColumn};
#[cfg(test)]
pub use self::serialize::tests::serialize_and_load;
pub use self::serialize::{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_mapped<T: MonotonicallyMappableToU128 + Debug>(
mut bytes: OwnedBytes,
) -> io::Result<Arc<dyn ColumnValues<T>>> {
let header = U128Header::deserialize(&mut bytes)?;
assert_eq!(header.codec_type, U128FastFieldCodecType::CompactSpace);
let reader = CompactSpaceDecompressor::open(bytes)?;
let inverted: StrictlyMonotonicMappingInverter<StrictlyMonotonicMappingToInternal<T>> =
StrictlyMonotonicMappingToInternal::<T>::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 + Debug>(
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 + Debug>(
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>;
}
#[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);
}
}

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

@@ -1,309 +0,0 @@
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);
}

161
columnar/src/columnar/column_type.rs
View File

@@ -1,161 +0,0 @@
use std::fmt::Debug;
use std::net::Ipv6Addr;
use crate::value::NumericalType;
use crate::InvalidData;
/// The column type represents the column type.
/// Any changes need to be propagated to `COLUMN_TYPES`.
#[derive(Hash, Eq, PartialEq, Debug, Clone, Copy, Ord, PartialOrd)]
#[repr(u8)]
pub enum ColumnType {
I64 = 0u8,
U64 = 1u8,
F64 = 2u8,
Bytes = 3u8,
Str = 4u8,
Bool = 5u8,
IpAddr = 6u8,
DateTime = 7u8,
}
// The order needs to match _exactly_ the order in the enum
const COLUMN_TYPES: [ColumnType; 8] = [
ColumnType::I64,
ColumnType::U64,
ColumnType::F64,
ColumnType::Bytes,
ColumnType::Str,
ColumnType::Bool,
ColumnType::IpAddr,
ColumnType::DateTime,
];
impl ColumnType {
pub fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn try_from_code(code: u8) -> Result<ColumnType, InvalidData> {
COLUMN_TYPES.get(code as usize).copied().ok_or(InvalidData)
}
}
impl From<NumericalType> for ColumnType {
fn from(numerical_type: NumericalType) -> Self {
match numerical_type {
NumericalType::I64 => ColumnType::I64,
NumericalType::U64 => ColumnType::U64,
NumericalType::F64 => ColumnType::F64,
}
}
}
impl ColumnType {
pub fn numerical_type(&self) -> Option<NumericalType> {
match self {
ColumnType::I64 => Some(NumericalType::I64),
ColumnType::U64 => Some(NumericalType::U64),
ColumnType::F64 => Some(NumericalType::F64),
ColumnType::Bytes
| ColumnType::Str
| ColumnType::Bool
| ColumnType::IpAddr
| ColumnType::DateTime => None,
}
}
}
// TODO remove if possible
pub trait HasAssociatedColumnType: 'static + Debug + Send + Sync + Copy + PartialOrd {
fn column_type() -> ColumnType;
fn default_value() -> Self;
}
impl HasAssociatedColumnType for u64 {
fn column_type() -> ColumnType {
ColumnType::U64
}
fn default_value() -> Self {
0u64
}
}
impl HasAssociatedColumnType for i64 {
fn column_type() -> ColumnType {
ColumnType::I64
}
fn default_value() -> Self {
0i64
}
}
impl HasAssociatedColumnType for f64 {
fn column_type() -> ColumnType {
ColumnType::F64
}
fn default_value() -> Self {
Default::default()
}
}
impl HasAssociatedColumnType for bool {
fn column_type() -> ColumnType {
ColumnType::Bool
}
fn default_value() -> Self {
Default::default()
}
}
impl HasAssociatedColumnType for crate::DateTime {
fn column_type() -> ColumnType {
ColumnType::DateTime
}
fn default_value() -> Self {
Default::default()
}
}
impl HasAssociatedColumnType for Ipv6Addr {
fn column_type() -> ColumnType {
ColumnType::IpAddr
}
fn default_value() -> Self {
Ipv6Addr::from([0u8; 16])
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::Cardinality;
#[test]
fn test_column_type_to_code() {
for (code, expected_column_type) in super::COLUMN_TYPES.iter().copied().enumerate() {
if let Ok(column_type) = ColumnType::try_from_code(code as u8) {
assert_eq!(column_type, expected_column_type);
}
}
for code in COLUMN_TYPES.len() as u8..=u8::MAX {
assert!(ColumnType::try_from_code(code as u8).is_err());
}
}
#[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
View File

@@ -1,73 +0,0 @@
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);
}
}

208
columnar/src/columnar/merge.rs
View File

@@ -1,208 +0,0 @@
use std::collections::HashMap;
use std::io;
use crate::columnar::ColumnarReader;
use crate::dynamic_column::DynamicColumn;
use crate::ColumnType;
pub enum MergeDocOrder {
/// Columnar tables are simply stacked one above the other.
/// If the i-th columnar_readers has n_rows_i rows, then
/// in the resulting columnar,
/// rows [r0..n_row_0) contains the row of columnar_readers[0], in ordder
/// rows [n_row_0..n_row_0 + n_row_1 contains the row of columnar_readers[1], in order.
/// ..
Stack,
/// Some more complex mapping, that can interleaves rows from the different readers and
/// possibly drop rows.
Complex(()),
}
pub fn merge_columnar(
_columnar_readers: &[ColumnarReader],
mapping: MergeDocOrder,
_output: &mut impl io::Write,
) -> io::Result<()> {
match mapping {
MergeDocOrder::Stack => {
// implement me :)
todo!();
}
MergeDocOrder::Complex(_) => {
// for later
todo!();
}
}
}
/// Column types are grouped into different categories.
/// After merge, all columns belonging to the same category are coerced to
/// the same column type.
///
/// In practise, today, only Numerical colummns are coerced into one type today.
///
/// See also [README.md].
#[derive(Copy, Clone, Eq, PartialEq, Hash, Debug)]
#[repr(u8)]
enum ColumnTypeCategory {
Bool,
Str,
Numerical,
DateTime,
Bytes,
IpAddr,
}
impl From<ColumnType> for ColumnTypeCategory {
fn from(column_type: ColumnType) -> Self {
match column_type {
ColumnType::I64 => ColumnTypeCategory::Numerical,
ColumnType::U64 => ColumnTypeCategory::Numerical,
ColumnType::F64 => ColumnTypeCategory::Numerical,
ColumnType::Bytes => ColumnTypeCategory::Bytes,
ColumnType::Str => ColumnTypeCategory::Str,
ColumnType::Bool => ColumnTypeCategory::Bool,
ColumnType::IpAddr => ColumnTypeCategory::IpAddr,
ColumnType::DateTime => ColumnTypeCategory::DateTime,
}
}
}
fn collect_columns(
columnar_readers: &[&ColumnarReader],
) -> io::Result<HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>>> {
// Each column name may have multiple types of column associated.
// For merging we are interested in the same column type category since they can be merged.
let mut field_name_to_group: HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>> =
HashMap::new();
for columnar_reader in columnar_readers {
let column_name_and_handle = columnar_reader.list_columns()?;
for (column_name, handle) in column_name_and_handle {
let column_type_to_handles = field_name_to_group
.entry(column_name.to_string())
.or_default();
let columns = column_type_to_handles
.entry(handle.column_type().into())
.or_default();
columns.push(handle.open()?);
}
}
normalize_columns(&mut field_name_to_group);
Ok(field_name_to_group)
}
/// Coerce numerical type columns to the same type
/// TODO rename to `coerce_columns`
fn normalize_columns(map: &mut HashMap<String, HashMap<ColumnTypeCategory, Vec<DynamicColumn>>>) {
for (_field_name, type_category_to_columns) in map.iter_mut() {
for (type_category, columns) in type_category_to_columns {
if type_category == &ColumnTypeCategory::Numerical {
let casted_columns = cast_to_common_numerical_column(&columns);
*columns = casted_columns;
}
}
}
}
/// Receives a list of columns of numerical types (u64, i64, f64)
///
/// Returns a list of `DynamicColumn` which are all of the same numerical type
fn cast_to_common_numerical_column(columns: &[DynamicColumn]) -> Vec<DynamicColumn> {
assert!(columns
.iter()
.all(|column| column.column_type().numerical_type().is_some()));
let coerce_to_i64: Vec<_> = columns
.iter()
.map(|column| column.clone().coerce_to_i64())
.collect();
if coerce_to_i64.iter().all(|column| column.is_some()) {
return coerce_to_i64
.into_iter()
.map(|column| column.unwrap())
.collect();
}
let coerce_to_u64: Vec<_> = columns
.iter()
.map(|column| column.clone().coerce_to_u64())
.collect();
if coerce_to_u64.iter().all(|column| column.is_some()) {
return coerce_to_u64
.into_iter()
.map(|column| column.unwrap())
.collect();
}
columns
.iter()
.map(|column| {
column
.clone()
.coerce_to_f64()
.expect("couldn't cast column to f64")
})
.collect()
}
#[cfg(test)]
mod tests {
use super::*;
use crate::ColumnarWriter;
#[test]
fn test_column_coercion() {
// i64 type
let columnar1 = {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(1u32, "numbers", 1i64);
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(2, &mut buffer).unwrap();
ColumnarReader::open(buffer).unwrap()
};
// u64 type
let columnar2 = {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(1u32, "numbers", u64::MAX - 100);
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(2, &mut buffer).unwrap();
ColumnarReader::open(buffer).unwrap()
};
// f64 type
let columnar3 = {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(1u32, "numbers", 30.5);
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(2, &mut buffer).unwrap();
ColumnarReader::open(buffer).unwrap()
};
let column_map = collect_columns(&[&columnar1, &columnar2, &columnar3]).unwrap();
assert_eq!(column_map.len(), 1);
let cat_to_columns = column_map.get("numbers").unwrap();
assert_eq!(cat_to_columns.len(), 1);
let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
assert!(numerical.iter().all(|column| column.is_f64()));
let column_map = collect_columns(&[&columnar1, &columnar1]).unwrap();
assert_eq!(column_map.len(), 1);
let cat_to_columns = column_map.get("numbers").unwrap();
assert_eq!(cat_to_columns.len(), 1);
let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
assert!(numerical.iter().all(|column| column.is_i64()));
let column_map = collect_columns(&[&columnar2, &columnar2]).unwrap();
assert_eq!(column_map.len(), 1);
let cat_to_columns = column_map.get("numbers").unwrap();
assert_eq!(cat_to_columns.len(), 1);
let numerical = cat_to_columns.get(&ColumnTypeCategory::Numerical).unwrap();
assert!(numerical.iter().all(|column| column.is_u64()));
}
}

10
columnar/src/columnar/mod.rs
View File

@@ -1,10 +0,0 @@
mod column_type;
mod format_version;
mod merge;
mod reader;
mod writer;
pub use column_type::{ColumnType, HasAssociatedColumnType};
pub use merge::{merge_columnar, MergeDocOrder};
pub use reader::ColumnarReader;
pub use writer::ColumnarWriter;

164
columnar/src/columnar/reader/mod.rs
View File

@@ -1,164 +0,0 @@
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 Add unit tests
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 =
// The last two bytes are respectively the 0u8 separator and the column_type.
String::from_utf8_lossy(&key_bytes[..key_bytes.len() - 2]).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.
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()
}
}
#[cfg(test)]
mod tests {
use crate::{ColumnType, ColumnarReader, ColumnarWriter};
#[test]
fn test_list_columns() {
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_column_type("col1", ColumnType::Str, false);
columnar_writer.record_column_type("col2", ColumnType::U64, false);
let mut buffer = Vec::new();
columnar_writer.serialize(1, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
let columns = columnar.list_columns().unwrap();
assert_eq!(columns.len(), 2);
assert_eq!(&columns[0].0, "col1");
assert_eq!(columns[0].1.column_type(), ColumnType::Str);
assert_eq!(&columns[1].0, "col2");
assert_eq!(columns[1].1.column_type(), ColumnType::U64);
}
#[test]
fn test_list_columns_strict_typing_prevents_coercion() {
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_column_type("count", ColumnType::U64, false);
columnar_writer.record_numerical(1, "count", 1u64);
let mut buffer = Vec::new();
columnar_writer.serialize(2, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
let columns = columnar.list_columns().unwrap();
assert_eq!(columns.len(), 1);
assert_eq!(&columns[0].0, "count");
assert_eq!(columns[0].1.column_type(), ColumnType::U64);
}
#[test]
#[should_panic(expect = "Input type forbidden")]
fn test_list_columns_strict_typing_panics_on_wrong_types() {
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_column_type("count", ColumnType::U64, false);
columnar_writer.record_numerical(1, "count", 1i64);
}
}

360
columnar/src/columnar/writer/column_operation.rs
View File

@@ -1,360 +0,0 @@
use std::net::Ipv6Addr;
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, 6>(self.len) | place_bits::<6, 8>(self.op_type.to_code())
}
fn try_from_code(code: u8) -> Result<Self, InvalidData> {
let len = select_bits::<0, 6>(code);
let typ_code = select_bits::<6, 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
}
}
impl SymbolValue for Ipv6Addr {
fn serialize(self, buffer: &mut [u8]) -> u8 {
buffer[0..16].copy_from_slice(&self.octets());
16
}
fn deserialize(bytes: &[u8]) -> Self {
let octets: [u8; 16] = bytes[0..16].try_into().unwrap();
Ipv6Addr::from(octets)
}
}
#[derive(Default)]
struct MiniBuffer {
pub bytes: [u8; 17],
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);
}
}

338
columnar/src/columnar/writer/column_writers.rs
View File

@@ -1,338 +0,0 @@
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,
}
impl NumericalColumnWriter {
pub fn force_numerical_type(&mut self, numerical_type: NumericalType) {
assert!(self
.compatible_numerical_types
.is_type_accepted(numerical_type));
self.compatible_numerical_types = CompatibleNumericalTypes::StaticType(numerical_type);
}
}
/// State used to store what types are still acceptable
/// after having seen a set of numerical values.
#[derive(Clone, Copy)]
enum CompatibleNumericalTypes {
Dynamic {
all_values_within_i64_range: bool,
all_values_within_u64_range: bool,
},
StaticType(NumericalType),
}
impl Default for CompatibleNumericalTypes {
fn default() -> CompatibleNumericalTypes {
CompatibleNumericalTypes::Dynamic {
all_values_within_i64_range: true,
all_values_within_u64_range: true,
}
}
}
impl CompatibleNumericalTypes {
fn is_type_accepted(&self, numerical_type: NumericalType) -> bool {
match self {
CompatibleNumericalTypes::Dynamic {
all_values_within_i64_range,
all_values_within_u64_range,
} => match numerical_type {
NumericalType::I64 => *all_values_within_i64_range,
NumericalType::U64 => *all_values_within_u64_range,
NumericalType::F64 => true,
},
CompatibleNumericalTypes::StaticType(static_numerical_type) => {
*static_numerical_type == numerical_type
}
}
}
fn accept_value(&mut self, numerical_value: NumericalValue) {
match self {
CompatibleNumericalTypes::Dynamic {
all_values_within_i64_range,
all_values_within_u64_range,
} => match numerical_value {
NumericalValue::I64(val_i64) => {
let value_within_u64_range = val_i64 >= 0i64;
*all_values_within_u64_range &= value_within_u64_range;
}
NumericalValue::U64(val_u64) => {
let value_within_i64_range = val_u64 < i64::MAX as u64;
*all_values_within_i64_range &= value_within_i64_range;
}
NumericalValue::F64(_) => {
*all_values_within_i64_range = false;
*all_values_within_u64_range = false;
}
},
CompatibleNumericalTypes::StaticType(typ) => {
assert_eq!(
numerical_value.numerical_type(),
*typ,
"Input type forbidden. This column has been forced to type {typ:?}, received \
{numerical_value:?}"
);
}
}
}
pub fn to_numerical_type(self) -> NumericalType {
for numerical_type in [NumericalType::I64, NumericalType::U64] {
if self.is_type_accepted(numerical_type) {
return numerical_type;
}
}
NumericalType::F64
}
}
impl NumericalColumnWriter {
pub fn numerical_type(&self) -> NumericalType {
self.compatible_numerical_types.to_numerical_type()
}
pub fn cardinality(&self, num_docs: RowId) -> Cardinality {
self.column_writer.get_cardinality(num_docs)
}
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)]
pub(crate) struct StrOrBytesColumnWriter {
pub(crate) dictionary_id: u32,
pub(crate) column_writer: ColumnWriter,
// If true, when facing a multivalued cardinality,
// values associated to a given document will be sorted.
//
// This is useful for facets.
//
// If false, the order of appearance in the document will be
// observed.
pub(crate) sort_values_within_row: bool,
}
impl StrOrBytesColumnWriter {
pub(crate) fn with_dictionary_id(dictionary_id: u32) -> StrOrBytesColumnWriter {
StrOrBytesColumnWriter {
dictionary_id,
column_writer: Default::default(),
sort_values_within_row: false,
}
}
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);
}
#[test]
#[should_panic]
fn test_compatible_numerical_types_static_incompatible_type() {
let mut compatible_numerical_types =
CompatibleNumericalTypes::StaticType(NumericalType::U64);
compatible_numerical_types.accept_value(NumericalValue::I64(1i64));
}
#[test]
fn test_compatible_numerical_types_static_different_type_forbidden() {
let mut compatible_numerical_types =
CompatibleNumericalTypes::StaticType(NumericalType::U64);
compatible_numerical_types.accept_value(NumericalValue::U64(u64::MAX));
}
#[test]
fn test_compatible_numerical_types_static() {
for typ in [NumericalType::I64, NumericalType::I64, NumericalType::F64] {
let compatible_numerical_types = CompatibleNumericalTypes::StaticType(typ);
assert_eq!(compatible_numerical_types.to_numerical_type(), typ);
}
}
}

778
columnar/src/columnar/writer/mod.rs
View File

@@ -1,778 +0,0 @@
mod column_operation;
mod column_writers;
mod serializer;
mod value_index;
use std::io;
use std::net::Ipv6Addr;
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, MonotonicallyMappableToU128, MonotonicallyMappableToU64, VecColumn,
};
use crate::columnar::column_type::ColumnType;
use crate::columnar::writer::column_writers::{
ColumnWriter, NumericalColumnWriter, StrOrBytesColumnWriter,
};
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,
u64_values: Vec<u64>,
ip_addr_values: Vec<Ipv6Addr>,
}
/// 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,
datetime_field_hash_map: ArenaHashMap,
bool_field_hash_map: ArenaHashMap,
ip_addr_field_hash_map: ArenaHashMap,
bytes_field_hash_map: ArenaHashMap,
str_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),
ip_addr_field_hash_map: ArenaHashMap::new(10_000),
bytes_field_hash_map: ArenaHashMap::new(10_000),
str_field_hash_map: ArenaHashMap::new(10_000),
datetime_field_hash_map: ArenaHashMap::new(10_000),
dictionaries: Vec::new(),
arena: MemoryArena::default(),
buffers: SpareBuffers::default(),
}
}
}
#[inline]
fn mutate_or_create_column<V, TMutator>(
arena_hash_map: &mut ArenaHashMap,
column_name: &str,
updater: TMutator,
) where
V: Copy + 'static,
TMutator: FnMut(Option<V>) -> V,
{
assert!(
!column_name.as_bytes().contains(&0u8),
"key may not contain the 0 byte"
);
arena_hash_map.mutate_or_create(column_name.as_bytes(), updater);
}
impl ColumnarWriter {
pub fn mem_usage(&self) -> usize {
// TODO add dictionary builders.
self.arena.mem_usage()
+ self.numerical_field_hash_map.mem_usage()
+ self.bool_field_hash_map.mem_usage()
+ self.bytes_field_hash_map.mem_usage()
+ self.str_field_hash_map.mem_usage()
+ self.ip_addr_field_hash_map.mem_usage()
+ self.datetime_field_hash_map.mem_usage()
}
/// Records a column type. This is useful to bypass the coercion process,
/// makes sure the empty is present in the resulting columnar, or set
/// the `sort_values_within_row`.
///
/// `sort_values_within_row` is only allowed for `Bytes` or `Str` columns.
pub fn record_column_type(
&mut self,
column_name: &str,
column_type: ColumnType,
sort_values_within_row: bool,
) {
if sort_values_within_row {
assert!(
column_type == ColumnType::Bytes || column_type == ColumnType::Str,
"sort_values_within_row is only allowed for Bytes and Str columns",
);
}
match column_type {
ColumnType::Str | ColumnType::Bytes => {
let (hash_map, dictionaries) = (
if column_type == ColumnType::Str {
&mut self.str_field_hash_map
} else {
&mut self.bytes_field_hash_map
},
&mut self.dictionaries,
);
mutate_or_create_column(
hash_map,
column_name,
|column_opt: Option<StrOrBytesColumnWriter>| {
let mut column_writer = if let Some(column_writer) = column_opt {
column_writer
} else {
let dictionary_id = dictionaries.len() as u32;
dictionaries.push(DictionaryBuilder::default());
StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
};
column_writer.sort_values_within_row = sort_values_within_row;
column_writer
},
);
}
ColumnType::Bool => {
mutate_or_create_column(
&mut self.bool_field_hash_map,
column_name,
|column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
);
}
ColumnType::DateTime => {
mutate_or_create_column(
&mut self.datetime_field_hash_map,
column_name,
|column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
);
}
ColumnType::I64 | ColumnType::F64 | ColumnType::U64 => {
let numerical_type = column_type.numerical_type().unwrap();
mutate_or_create_column(
&mut self.numerical_field_hash_map,
column_name,
|column_opt: Option<NumericalColumnWriter>| {
let mut column: NumericalColumnWriter = column_opt.unwrap_or_default();
column.force_numerical_type(numerical_type);
column
},
);
}
ColumnType::IpAddr => mutate_or_create_column(
&mut self.ip_addr_field_hash_map,
column_name,
|column_opt: Option<ColumnWriter>| column_opt.unwrap_or_default(),
),
}
}
pub fn record_numerical<T: Into<NumericalValue> + Copy>(
&mut self,
doc: RowId,
column_name: &str,
numerical_value: T,
) {
let (hash_map, arena) = (&mut self.numerical_field_hash_map, &mut self.arena);
mutate_or_create_column(
hash_map,
column_name,
|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_ip_addr(&mut self, doc: RowId, column_name: &str, ip_addr: Ipv6Addr) {
assert!(
!column_name.as_bytes().contains(&0u8),
"key may not contain the 0 byte"
);
let (hash_map, arena) = (&mut self.ip_addr_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, ip_addr, arena);
column
},
);
}
pub fn record_bool(&mut self, doc: RowId, column_name: &str, val: bool) {
let (hash_map, arena) = (&mut self.bool_field_hash_map, &mut self.arena);
mutate_or_create_column(hash_map, column_name, |column_opt: Option<ColumnWriter>| {
let mut column: ColumnWriter = column_opt.unwrap_or_default();
column.record(doc, val, arena);
column
});
}
pub fn record_datetime(&mut self, doc: RowId, column_name: &str, datetime: crate::DateTime) {
let (hash_map, arena) = (&mut self.datetime_field_hash_map, &mut self.arena);
mutate_or_create_column(hash_map, column_name, |column_opt: Option<ColumnWriter>| {
let mut column: ColumnWriter = column_opt.unwrap_or_default();
column.record(doc, NumericalValue::I64(datetime.timestamp_micros), arena);
column
});
}
pub fn record_str(&mut self, doc: RowId, column_name: &str, value: &str) {
let (hash_map, arena, dictionaries) = (
&mut self.str_field_hash_map,
&mut self.arena,
&mut self.dictionaries,
);
hash_map.mutate_or_create(
column_name.as_bytes(),
|column_opt: Option<StrOrBytesColumnWriter>| {
let mut column: StrOrBytesColumnWriter = column_opt.unwrap_or_else(|| {
// Each column has its own dictionary
let dictionary_id = dictionaries.len() as u32;
dictionaries.push(DictionaryBuilder::default());
StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
});
column.record_bytes(doc, value.as_bytes(), dictionaries, arena);
column
},
);
}
pub fn record_bytes(&mut self, doc: RowId, column_name: &str, value: &[u8]) {
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<StrOrBytesColumnWriter>| {
let mut column: StrOrBytesColumnWriter = column_opt.unwrap_or_else(|| {
// Each column has its own dictionary
let dictionary_id = dictionaries.len() as u32;
dictionaries.push(DictionaryBuilder::default());
StrOrBytesColumnWriter::with_dictionary_id(dictionary_id)
});
column.record_bytes(doc, value, 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 columns: Vec<(&[u8], ColumnType, Addr)> = self
.numerical_field_hash_map
.iter()
.map(|(column_name, addr, _)| {
let numerical_column_writer: NumericalColumnWriter =
self.numerical_field_hash_map.read(addr);
let column_type = numerical_column_writer.numerical_type().into();
(column_name, column_type, addr)
})
.collect();
columns.extend(
self.bytes_field_hash_map
.iter()
.map(|(term, addr, _)| (term, ColumnType::Bytes, addr)),
);
columns.extend(
self.str_field_hash_map
.iter()
.map(|(column_name, addr, _)| (column_name, ColumnType::Str, addr)),
);
columns.extend(
self.bool_field_hash_map
.iter()
.map(|(column_name, addr, _)| (column_name, ColumnType::Bool, addr)),
);
columns.extend(
self.ip_addr_field_hash_map
.iter()
.map(|(column_name, addr, _)| (column_name, ColumnType::IpAddr, addr)),
);
columns.extend(
self.datetime_field_hash_map
.iter()
.map(|(column_name, addr, _)| (column_name, ColumnType::DateTime, addr)),
);
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, column_type, addr) in columns {
match column_type {
ColumnType::Bool | ColumnType::DateTime => {
let column_writer: ColumnWriter = if column_type == ColumnType::Bool {
self.bool_field_hash_map.read(addr)
} else {
self.datetime_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,
)?;
}
ColumnType::IpAddr => {
let column_writer: ColumnWriter = self.ip_addr_field_hash_map.read(addr);
let cardinality = column_writer.get_cardinality(num_docs);
let mut column_serializer =
serializer.serialize_column(column_name, ColumnType::IpAddr);
serialize_ip_addr_column(
cardinality,
num_docs,
column_writer.operation_iterator(arena, &mut symbol_byte_buffer),
buffers,
&mut column_serializer,
)?;
}
ColumnType::Bytes | ColumnType::Str => {
let str_or_bytes_column_writer: StrOrBytesColumnWriter =
if column_type == ColumnType::Bytes {
self.bytes_field_hash_map.read(addr)
} else {
self.str_field_hash_map.read(addr)
};
let dictionary_builder =
&dictionaries[str_or_bytes_column_writer.dictionary_id as usize];
let cardinality = str_or_bytes_column_writer
.column_writer
.get_cardinality(num_docs);
let mut column_serializer =
serializer.serialize_column(column_name, column_type);
serialize_bytes_or_str_column(
cardinality,
num_docs,
str_or_bytes_column_writer.sort_values_within_row,
dictionary_builder,
str_or_bytes_column_writer
.operation_iterator(arena, &mut symbol_byte_buffer),
buffers,
&mut column_serializer,
)?;
}
ColumnType::I64 | ColumnType::F64 | ColumnType::U64 => {
let numerical_column_writer: NumericalColumnWriter =
self.numerical_field_hash_map.read(addr);
let numerical_type = column_type.numerical_type().unwrap();
let cardinality = numerical_column_writer.cardinality(num_docs);
let mut column_serializer =
serializer.serialize_column(column_name, ColumnType::from(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_or_str_column(
cardinality: Cardinality,
num_docs: RowId,
sort_values_within_row: bool,
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),
}
});
send_to_serialize_column_mappable_to_u64(
operation_iterator,
cardinality,
num_docs,
sort_values_within_row,
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,
..
} = buffers;
match numerical_type {
NumericalType::I64 => {
send_to_serialize_column_mappable_to_u64(
coerce_numerical_symbol::<i64>(op_iterator),
cardinality,
num_docs,
false,
value_index_builders,
u64_values,
wrt,
)?;
}
NumericalType::U64 => {
send_to_serialize_column_mappable_to_u64(
coerce_numerical_symbol::<u64>(op_iterator),
cardinality,
num_docs,
false,
value_index_builders,
u64_values,
wrt,
)?;
}
NumericalType::F64 => {
send_to_serialize_column_mappable_to_u64(
coerce_numerical_symbol::<f64>(op_iterator),
cardinality,
num_docs,
false,
value_index_builders,
u64_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,
u64_values,
..
} = buffers;
send_to_serialize_column_mappable_to_u64(
column_operations_it.map(|bool_column_operation| match bool_column_operation {
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
ColumnOperation::Value(bool_val) => ColumnOperation::Value(bool_val.to_u64()),
}),
cardinality,
num_docs,
false,
value_index_builders,
u64_values,
wrt,
)?;
Ok(())
}
fn serialize_ip_addr_column(
cardinality: Cardinality,
num_docs: RowId,
column_operations_it: impl Iterator<Item = ColumnOperation<Ipv6Addr>>,
buffers: &mut SpareBuffers,
wrt: &mut impl io::Write,
) -> io::Result<()> {
let SpareBuffers {
value_index_builders,
ip_addr_values,
..
} = buffers;
send_to_serialize_column_mappable_to_u128(
column_operations_it,
cardinality,
num_docs,
value_index_builders,
ip_addr_values,
wrt,
)?;
Ok(())
}
fn send_to_serialize_column_mappable_to_u128<
T: Copy + Ord + std::fmt::Debug + Send + Sync + MonotonicallyMappableToU128 + 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();
// TODO: split index and values
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);
SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
}
};
crate::column::serialize_column_mappable_to_u128(
serializable_column_index,
|| values.iter().cloned(),
values.len() as u32,
&mut wrt,
)?;
Ok(())
}
fn sort_values_within_row_in_place(
multivalued_index: &impl ColumnValues<RowId>,
values: &mut Vec<u64>,
) {
let mut start_index: usize = 0;
for end_index in multivalued_index.iter() {
let end_index = end_index as usize;
values[start_index..end_index].sort_unstable();
start_index = end_index;
}
}
fn send_to_serialize_column_mappable_to_u64(
op_iterator: impl Iterator<Item = ColumnOperation<u64>>,
cardinality: Cardinality,
num_docs: RowId,
sort_values_within_row: bool,
value_index_builders: &mut PreallocatedIndexBuilders,
values: &mut Vec<u64>,
mut wrt: impl io::Write,
) -> io::Result<()>
where
for<'a> VecColumn<'a, u64>: ColumnValues<u64>,
{
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);
if sort_values_within_row {
sort_values_within_row_in_place(&multivalued_index, values);
}
SerializableColumnIndex::Multivalued(Box::new(multivalued_index))
}
};
crate::column::serialize_column_mappable_to_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<u64>>
where T: Coerce + MonotonicallyMappableToU64 {
operation_iterator.map(|symbol| match symbol {
ColumnOperation::NewDoc(doc) => ColumnOperation::NewDoc(doc),
ColumnOperation::Value(numerical_value) => {
ColumnOperation::Value(T::coerce(numerical_value).to_u64())
}
})
}
fn consume_operation_iterator<T: Ord, 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);
}
}
}
}
#[cfg(test)]
mod tests {
use stacker::MemoryArena;
use crate::columnar::writer::column_operation::ColumnOperation;
use crate::{Cardinality, 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
View File

@@ -1,106 +0,0 @@
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::Str, &mut buffer);
assert_eq!(buffer.len(), 12);
assert_eq!(&buffer[..10], b"root\0child");
assert_eq!(buffer[10], 0u8);
assert_eq!(buffer[11], ColumnType::Str.to_code());
}
}

195
columnar/src/columnar/writer/value_index.rs
View File

@@ -1,195 +0,0 @@
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 {
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 + 1, 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, 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, 2]
);
}
}

84
columnar/src/dictionary.rs
View File

@@ -1,84 +0,0 @@
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));
}
}

241
columnar/src/dynamic_column.rs
View File

@@ -1,241 +0,0 @@
use std::io;
use std::net::Ipv6Addr;
use std::sync::Arc;
use common::file_slice::FileSlice;
use common::{HasLen, OwnedBytes};
use crate::column::{BytesColumn, Column, StrColumn};
use crate::column_values::{monotonic_map_column, StrictlyMonotonicFn};
use crate::columnar::ColumnType;
use crate::{DateTime, NumericalType};
#[derive(Clone)]
pub enum DynamicColumn {
Bool(Column<bool>),
I64(Column<i64>),
U64(Column<u64>),
F64(Column<f64>),
IpAddr(Column<Ipv6Addr>),
DateTime(Column<DateTime>),
Bytes(BytesColumn),
Str(StrColumn),
}
impl DynamicColumn {
pub fn column_type(&self) -> ColumnType {
match self {
DynamicColumn::Bool(_) => ColumnType::Bool,
DynamicColumn::I64(_) => ColumnType::I64,
DynamicColumn::U64(_) => ColumnType::U64,
DynamicColumn::F64(_) => ColumnType::F64,
DynamicColumn::IpAddr(_) => ColumnType::IpAddr,
DynamicColumn::DateTime(_) => ColumnType::DateTime,
DynamicColumn::Bytes(_) => ColumnType::Bytes,
DynamicColumn::Str(_) => ColumnType::Str,
}
}
pub fn is_numerical(&self) -> bool {
self.column_type().numerical_type().is_some()
}
pub fn is_f64(&self) -> bool {
self.column_type().numerical_type() == Some(NumericalType::F64)
}
pub fn is_i64(&self) -> bool {
self.column_type().numerical_type() == Some(NumericalType::I64)
}
pub fn is_u64(&self) -> bool {
self.column_type().numerical_type() == Some(NumericalType::U64)
}
pub fn coerce_to_f64(self) -> Option<DynamicColumn> {
match self {
DynamicColumn::I64(column) => Some(DynamicColumn::F64(Column {
idx: column.idx,
values: Arc::new(monotonic_map_column(column.values, MapI64ToF64)),
})),
DynamicColumn::U64(column) => Some(DynamicColumn::F64(Column {
idx: column.idx,
values: Arc::new(monotonic_map_column(column.values, MapU64ToF64)),
})),
DynamicColumn::F64(_) => Some(self),
_ => None,
}
}
pub fn coerce_to_i64(self) -> Option<DynamicColumn> {
match self {
DynamicColumn::U64(column) => {
if column.max_value() > i64::MAX as u64 {
return None;
}
Some(DynamicColumn::I64(Column {
idx: column.idx,
values: Arc::new(monotonic_map_column(column.values, MapU64ToI64)),
}))
}
DynamicColumn::I64(_) => Some(self),
_ => None,
}
}
pub fn coerce_to_u64(self) -> Option<DynamicColumn> {
match self {
DynamicColumn::I64(column) => {
if column.min_value() < 0 {
return None;
}
Some(DynamicColumn::U64(Column {
idx: column.idx,
values: Arc::new(monotonic_map_column(column.values, MapI64ToU64)),
}))
}
DynamicColumn::U64(_) => Some(self),
_ => None,
}
}
}
struct MapI64ToF64;
impl StrictlyMonotonicFn<i64, f64> for MapI64ToF64 {
#[inline(always)]
fn mapping(&self, inp: i64) -> f64 {
inp as f64
}
#[inline(always)]
fn inverse(&self, out: f64) -> i64 {
out as i64
}
}
struct MapU64ToF64;
impl StrictlyMonotonicFn<u64, f64> for MapU64ToF64 {
#[inline(always)]
fn mapping(&self, inp: u64) -> f64 {
inp as f64
}
#[inline(always)]
fn inverse(&self, out: f64) -> u64 {
out as u64
}
}
struct MapU64ToI64;
impl StrictlyMonotonicFn<u64, i64> for MapU64ToI64 {
#[inline(always)]
fn mapping(&self, inp: u64) -> i64 {
inp as i64
}
#[inline(always)]
fn inverse(&self, out: i64) -> u64 {
out as u64
}
}
struct MapI64ToU64;
impl StrictlyMonotonicFn<i64, u64> for MapI64ToU64 {
#[inline(always)]
fn mapping(&self, inp: i64) -> u64 {
inp as u64
}
#[inline(always)]
fn inverse(&self, out: u64) -> i64 {
out as i64
}
}
macro_rules! static_dynamic_conversions {
($typ:ty, $enum_name:ident) => {
impl Into<Option<$typ>> for DynamicColumn {
fn into(self) -> Option<$typ> {
if let DynamicColumn::$enum_name(col) = self {
Some(col)
} else {
None
}
}
}
impl From<$typ> for DynamicColumn {
fn from(typed_column: $typ) -> Self {
DynamicColumn::$enum_name(typed_column)
}
}
};
}
static_dynamic_conversions!(Column<bool>, Bool);
static_dynamic_conversions!(Column<u64>, U64);
static_dynamic_conversions!(Column<i64>, I64);
static_dynamic_conversions!(Column<f64>, F64);
static_dynamic_conversions!(Column<crate::DateTime>, DateTime);
static_dynamic_conversions!(StrColumn, Str);
static_dynamic_conversions!(BytesColumn, Bytes);
static_dynamic_conversions!(Column<Ipv6Addr>, IpAddr);
#[derive(Clone)]
pub struct DynamicColumnHandle {
pub(crate) file_slice: FileSlice,
pub(crate) column_type: ColumnType,
}
impl DynamicColumnHandle {
// TODO rename load
pub fn open(&self) -> io::Result<DynamicColumn> {
let column_bytes: OwnedBytes = self.file_slice.read_bytes()?;
self.open_internal(column_bytes)
}
// TODO rename load_async
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)
}
/// Returns the `u64` fast field reader reader associated with `fields` of types
/// Str, u64, i64, f64, or datetime.
///
/// If not, the fastfield reader will returns the u64-value associated with the original
/// FastValue.
pub fn open_u64_lenient(&self) -> io::Result<Option<Column<u64>>> {
let column_bytes = self.file_slice.read_bytes()?;
match self.column_type {
ColumnType::Str | ColumnType::Bytes => {
let column: BytesColumn = crate::column::open_column_bytes(column_bytes)?;
Ok(Some(column.term_ord_column))
}
ColumnType::Bool => Ok(None),
ColumnType::IpAddr => Ok(None),
ColumnType::I64 | ColumnType::U64 | ColumnType::F64 | ColumnType::DateTime => {
let column = crate::column::open_column_u64::<u64>(column_bytes)?;
Ok(Some(column))
}
}
}
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::<BytesColumn>(column_bytes)?.into()
}
ColumnType::Str => crate::column::open_column_bytes::<StrColumn>(column_bytes)?.into(),
ColumnType::I64 => crate::column::open_column_u64::<i64>(column_bytes)?.into(),
ColumnType::U64 => crate::column::open_column_u64::<u64>(column_bytes)?.into(),
ColumnType::F64 => crate::column::open_column_u64::<f64>(column_bytes)?.into(),
ColumnType::Bool => crate::column::open_column_u64::<bool>(column_bytes)?.into(),
ColumnType::IpAddr => crate::column::open_column_u128::<Ipv6Addr>(column_bytes)?.into(),
ColumnType::DateTime => {
crate::column::open_column_u64::<crate::DateTime>(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
}
}

84
columnar/src/lib.rs
View File

@@ -1,84 +0,0 @@
#![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 column::{BytesColumn, Column, StrColumn};
pub use column_index::ColumnIndex;
pub use column_values::{ColumnValues, MonotonicallyMappableToU128, MonotonicallyMappableToU64};
pub use columnar::{
merge_columnar, ColumnType, ColumnarReader, ColumnarWriter, HasAssociatedColumnType,
MergeDocOrder,
};
use sstable::VoidSSTable;
pub use value::{NumericalType, NumericalValue};
pub use self::dynamic_column::{DynamicColumn, DynamicColumnHandle};
pub type RowId = u32;
pub use sstable::Dictionary;
pub type Streamer<'a> = sstable::Streamer<'a, VoidSSTable>;
#[derive(Clone, Copy, PartialOrd, PartialEq, Default, Debug)]
pub struct DateTime {
pub 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;

212
columnar/src/tests.rs
View File

@@ -1,212 +0,0 @@
use std::net::Ipv6Addr;
use crate::column_values::MonotonicallyMappableToU128;
use crate::columnar::ColumnType;
use crate::dynamic_column::{DynamicColumn, DynamicColumnHandle};
use crate::value::NumericalValue;
use crate::{Cardinality, ColumnarReader, ColumnarWriter};
#[test]
fn test_dataframe_writer_str() {
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(), 158);
}
#[test]
fn test_dataframe_writer_bytes() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_bytes(1u32, "my_string", b"hello");
dataframe_writer.record_bytes(3u32, "my_string", b"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(), 158);
}
#[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(), 22);
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_u64_multivalued() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_numerical(2u32, "divisor", 2u64);
dataframe_writer.record_numerical(3u32, "divisor", 3u64);
dataframe_writer.record_numerical(4u32, "divisor", 2u64);
dataframe_writer.record_numerical(5u32, "divisor", 5u64);
dataframe_writer.record_numerical(6u32, "divisor", 2u64);
dataframe_writer.record_numerical(6u32, "divisor", 3u64);
let mut buffer: Vec<u8> = Vec::new();
dataframe_writer.serialize(7, &mut buffer).unwrap();
let columnar = ColumnarReader::open(buffer).unwrap();
assert_eq!(columnar.num_columns(), 1);
let cols: Vec<DynamicColumnHandle> = columnar.read_columns("divisor").unwrap();
assert_eq!(cols.len(), 1);
assert_eq!(cols[0].num_bytes(), 29);
let dyn_i64_col = cols[0].open().unwrap();
let DynamicColumn::I64(divisor_col) = dyn_i64_col else { panic!(); };
assert_eq!(
divisor_col.get_cardinality(),
crate::Cardinality::Multivalued
);
assert_eq!(divisor_col.num_rows(), 7);
}
#[test]
fn test_dataframe_writer_ip_addr() {
let mut dataframe_writer = ColumnarWriter::default();
dataframe_writer.record_ip_addr(1, "ip_addr", Ipv6Addr::from_u128(1001));
dataframe_writer.record_ip_addr(3, "ip_addr", Ipv6Addr::from_u128(1050));
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("ip_addr").unwrap();
assert_eq!(cols.len(), 1);
assert_eq!(cols[0].num_bytes(), 42);
assert_eq!(cols[0].column_type(), ColumnType::IpAddr);
let dyn_bool_col = cols[0].open().unwrap();
let DynamicColumn::IpAddr(ip_col) = dyn_bool_col else { panic!(); };
let vals: Vec<Option<Ipv6Addr>> = (0..5).map(|row_id| ip_col.first(row_id)).collect();
assert_eq!(
&vals,
&[
None,
Some(Ipv6Addr::from_u128(1001)),
None,
Some(Ipv6Addr::from_u128(1050)),
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(), 33);
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_str() {
let mut buffer = Vec::new();
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_str(1, "my.column", "a");
columnar_writer.record_str(3, "my.column", "c");
columnar_writer.record_str(3, "my.column2", "different_column!");
columnar_writer.record_str(4, "my.column", "b");
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!(); };
let index: Vec<Option<u64>> = (0..5).map(|row_id| str_col.ords().first(row_id)).collect();
assert_eq!(index, &[None, Some(0), None, Some(2), Some(1)]);
assert_eq!(str_col.num_rows(), 5);
let mut term_buffer = String::new();
let term_ords = str_col.ords();
assert_eq!(term_ords.first(0), None);
assert_eq!(term_ords.first(1), Some(0));
str_col.ord_to_str(0u64, &mut term_buffer).unwrap();
assert_eq!(term_buffer, "a");
assert_eq!(term_ords.first(2), None);
assert_eq!(term_ords.first(3), Some(2));
str_col.ord_to_str(2u64, &mut term_buffer).unwrap();
assert_eq!(term_buffer, "c");
assert_eq!(term_ords.first(4), Some(1));
str_col.ord_to_str(1u64, &mut term_buffer).unwrap();
assert_eq!(term_buffer, "b");
}
#[test]
fn test_dictionary_encoded_bytes() {
let mut buffer = Vec::new();
let mut columnar_writer = ColumnarWriter::default();
columnar_writer.record_bytes(1, "my.column", b"a");
columnar_writer.record_bytes(3, "my.column", b"c");
columnar_writer.record_bytes(3, "my.column2", b"different_column!");
columnar_writer.record_bytes(4, "my.column", b"b");
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::Bytes(bytes_col) = col_handles[0].open().unwrap() else { panic!(); };
let index: Vec<Option<u64>> = (0..5)
.map(|row_id| bytes_col.ords().first(row_id))
.collect();
assert_eq!(index, &[None, Some(0), None, Some(2), Some(1)]);
assert_eq!(bytes_col.num_rows(), 5);
let mut term_buffer = Vec::new();
let term_ords = bytes_col.ords();
assert_eq!(term_ords.first(0), None);
assert_eq!(term_ords.first(1), Some(0));
bytes_col
.dictionary
.ord_to_term(0u64, &mut term_buffer)
.unwrap();
assert_eq!(term_buffer, b"a");
assert_eq!(term_ords.first(2), None);
assert_eq!(term_ords.first(3), Some(2));
bytes_col
.dictionary
.ord_to_term(2u64, &mut term_buffer)
.unwrap();
assert_eq!(term_buffer, b"c");
assert_eq!(term_ords.first(4), Some(1));
bytes_col
.dictionary
.ord_to_term(1u64, &mut term_buffer)
.unwrap();
assert_eq!(term_buffer, b"b");
}

76
columnar/src/utils.rs
View File

@@ -1,76 +0,0 @@
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);
}
}

129
columnar/src/value.rs
View File

@@ -1,129 +0,0 @@
use crate::InvalidData;
#[derive(Copy, Clone, PartialEq, Debug)]
pub enum NumericalValue {
I64(i64),
U64(u64),
F64(f64),
}
impl NumericalValue {
pub fn numerical_type(&self) -> NumericalType {
match self {
NumericalValue::I64(_) => NumericalType::I64,
NumericalValue::U64(_) => NumericalType::U64,
NumericalValue::F64(_) => NumericalType::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)
}
}
#[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,
}
}
}
impl Coerce for crate::DateTime {
fn coerce(value: NumericalValue) -> Self {
let timestamp_micros = i64::coerce(value);
crate::DateTime { timestamp_micros }
}
}
#[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

@@ -1,21 +1,16 @@
[package]
name = "tantivy-common"
version = "0.5.0"
version = "0.3.0"
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= "0.5", path="../ownedbytes" }
async-trait = "0.1"
ownedbytes = { version="0.3", path="../ownedbytes" }
[dev-dependencies]
proptest = "1.0.0"

6
common/src/bitset.rs
View File

@@ -151,7 +151,7 @@ impl TinySet {
if self.is_empty() {
None
} else {
let lowest = self.0.trailing_zeros();
let lowest = self.0.trailing_zeros() as u32;
self.0 ^= TinySet::singleton(lowest).0;
Some(lowest)
}
@@ -421,7 +421,7 @@ mod tests {
bitset.serialize(&mut out).unwrap();
let bitset = ReadOnlyBitSet::open(OwnedBytes::new(out));
assert_eq!(bitset.len(), i as usize);
assert_eq!(bitset.len() as usize, 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(), 64);
assert_eq!(bitset.len() as usize, 64);
}
#[test]

166
common/src/group_by.rs
View File

@@ -1,166 +0,0 @@
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

@@ -5,14 +5,11 @@ 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,

14
common/src/serialize.rs
View File

@@ -94,20 +94,6 @@ 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)

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
}

73
examples-disabled/ip_field.rs
View File

@@ -1,73 +0,0 @@
// # 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(())
}

12
examples-disabled/aggregation.rs → examples/aggregation.rs
View File

@@ -13,7 +13,7 @@ use tantivy::aggregation::agg_result::AggregationResults;
use tantivy::aggregation::metric::AverageAggregation;
use tantivy::aggregation::AggregationCollector;
use tantivy::query::TermQuery;
use tantivy::schema::{self, IndexRecordOption, Schema, TextFieldIndexing};
use tantivy::schema::{self, Cardinality, IndexRecordOption, Schema, TextFieldIndexing};
use tantivy::{doc, Index, Term};
fn main() -> tantivy::Result<()> {
@@ -25,9 +25,9 @@ fn main() -> tantivy::Result<()> {
.set_stored();
let text_field = schema_builder.add_text_field("text", text_fieldtype);
let score_fieldtype =
crate::schema::NumericOptions::default().set_fast();
crate::schema::NumericOptions::default().set_fast(Cardinality::SingleValue);
let highscore_field = schema_builder.add_f64_field("highscore", score_fieldtype.clone());
let price_field = schema_builder.add_f64_field("price", score_fieldtype);
let price_field = schema_builder.add_f64_field("price", score_fieldtype.clone());
let schema = schema_builder.build();
@@ -112,18 +112,18 @@ fn main() -> tantivy::Result<()> {
],
..Default::default()
}),
sub_aggregation: sub_agg_req_1,
sub_aggregation: sub_agg_req_1.clone(),
}),
)]
.into_iter()
.collect();
let collector = AggregationCollector::from_aggs(agg_req_1, None, index.schema());
let collector = AggregationCollector::from_aggs(agg_req_1, None);
let searcher = reader.searcher();
let agg_res: AggregationResults = searcher.search(&term_query, &collector).unwrap();
let res: Value = serde_json::to_value(agg_res)?;
let res: Value = serde_json::to_value(&agg_res)?;
println!("{}", serde_json::to_string_pretty(&res)?);
Ok(())

0
examples-disabled/basic_search.rs → examples/basic_search.rs
View File

14
examples-disabled/custom_collector.rs → 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::{Schema, FAST, INDEXED, TEXT};
use tantivy::schema::{Field, Schema, FAST, INDEXED, TEXT};
use tantivy::{doc, Index, Score, SegmentReader};
#[derive(Default)]
@@ -52,11 +52,11 @@ impl Stats {
}
struct StatsCollector {
field: String,
field: Field,
}
impl StatsCollector {
fn with_field(field: String) -> StatsCollector {
fn with_field(field: Field) -> 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 f64;
let value = self.fast_field_reader.get_val(doc as u64) as f64;
self.stats.count += 1;
self.stats.sum += value;
self.stats.squared_sum += value * value;
@@ -171,9 +171,7 @@ 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".to_string()))?
{
if let Some(stats) = searcher.search(&query, &StatsCollector::with_field(price))? {
println!("count: {}", stats.count());
println!("mean: {}", stats.mean());
println!("standard deviation: {}", stats.standard_deviation());

0
examples-disabled/custom_tokenizer.rs → examples/custom_tokenizer.rs
View File

4
examples-disabled/date_time_field.rs → examples/date_time_field.rs
View File

@@ -4,7 +4,7 @@
use tantivy::collector::TopDocs;
use tantivy::query::QueryParser;
use tantivy::schema::{DateOptions, Schema, Value, INDEXED, STORED, STRING};
use tantivy::schema::{Cardinality, DateOptions, Schema, Value, INDEXED, STORED, STRING};
use tantivy::Index;
fn main() -> tantivy::Result<()> {
@@ -12,7 +12,7 @@ fn main() -> tantivy::Result<()> {
let mut schema_builder = Schema::builder();
let opts = DateOptions::from(INDEXED)
.set_stored()
.set_fast()
.set_fast(Cardinality::SingleValue)
.set_precision(tantivy::DatePrecision::Seconds);
let occurred_at = schema_builder.add_date_field("occurred_at", opts);
let event_type = schema_builder.add_text_field("event", STRING | STORED);

0
examples-disabled/deleting_updating_documents.rs → examples/deleting_updating_documents.rs
View File

20
examples-disabled/faceted_search.rs → examples/faceted_search.rs
View File

@@ -1,17 +1,15 @@
// # Faceted Search
// # Basic Example
//
// This example covers the faceted search functionalities of
// This example covers the basic functionalities of
// tantivy.
//
// We will :
// - 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.
//
// - 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.
// ---
// Importing tantivy...
use tantivy::collector::FacetCollector;
@@ -23,7 +21,7 @@ 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("name", TEXT | STORED);
let name = schema_builder.add_text_field("felin_name", TEXT | STORED);
// this is our faceted field: its scientific classification
let classification = schema_builder.add_facet_field("classification", FacetOptions::default());

0
examples-disabled/faceted_search_with_tweaked_score.rs → examples/faceted_search_with_tweaked_score.rs
View File

2
examples-disabled/integer_range_search.rs → 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".to_string(), 1960..1970);
let docs_in_the_sixties = RangeQuery::new_u64(year_field, 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);

0
examples-disabled/iterating_docs_and_positions.rs → examples/iterating_docs_and_positions.rs
View File

0
examples-disabled/json_field.rs → examples/json_field.rs
View File

0
examples-disabled/multiple_producer.rs → examples/multiple_producer.rs
View File

0
examples-disabled/pre_tokenized_text.rs → examples/pre_tokenized_text.rs
View File

0
examples-disabled/snippet.rs → examples/snippet.rs
View File

0
examples-disabled/stop_words.rs → examples/stop_words.rs
View File

12
examples-disabled/warmer.rs → 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::{Schema, FAST, TEXT};
use tantivy::schema::{Field, 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: String,
field: Field,
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: String, price_fetcher: T) -> Self {
pub fn with_product_id_field<T: PriceFetcher>(field: Field, 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))
.map(|doc| product_id_reader.get_val(doc as u64))
.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".to_string(),
product_id,
price_table.clone(),
));
price_table.update_price(OLIVE_OIL, 12);

0
examples-disabled/working_with_json.rs → examples/working_with_json.rs
View File

13
fastfield_codecs/Cargo.toml
View File

@@ -1,21 +1,18 @@
[package]
name = "fastfield_codecs"
version = "0.3.0"
version = "0.2.0"
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 = "0.5", path = "../common/", package = "tantivy-common" }
tantivy-bitpacker = { version= "0.3", path = "../bitpacker/" }
columnar = { version= "0.1", path="../columnar", package="tantivy-columnar" }
prettytable-rs = {version="0.10.0", optional= true}
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}
rand = {version="0.8.3", optional= true}
fastdivide = "0.4"
log = "0.4"

151
fastfield_codecs/benches/bench.rs
View File

@@ -4,11 +4,11 @@ extern crate test;
#[cfg(test)]
mod tests {
use std::ops::RangeInclusive;
use std::iter;
use std::sync::Arc;
use common::OwnedBytes;
use fastfield_codecs::*;
use ownedbytes::OwnedBytes;
use rand::prelude::*;
use test::Bencher;
@@ -65,30 +65,33 @@ mod tests {
b.iter(|| {
let mut a = 0u64;
for _ in 0..n {
a = column.get_val(a as u32);
a = column.get_val(a as u64);
}
a
});
}
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]);
fn get_exp_data() -> Vec<u64> {
let mut data = vec![];
for _ in 0..300_000 {
let val = rng.gen_range(1..=100);
data.push(val);
for i in 0..100 {
let num = i * i;
data.extend(iter::repeat(i as u64).take(num));
}
data.push(SINGLE_ITEM);
data.shuffle(&mut StdRng::from_seed([1u8; 32]));
data.shuffle(&mut rng);
let data = data.iter().map(|el| *el as u128).collect::<Vec<_>>();
// lengt = 328350
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<_>>();
@@ -98,122 +101,34 @@ mod tests {
fn get_u128_column_from_data(data: &[u128]) -> Arc<dyn Column<u128>> {
let mut out = vec![];
let iter_gen = || data.iter().cloned();
serialize_u128(iter_gen, data.len() as u32, &mut out).unwrap();
serialize_u128(iter_gen, data.len() as u64, &mut out).unwrap();
let out = OwnedBytes::new(out);
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 data = get_data_50percent_item();
let (major_item, _minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*FIFTY_PERCENT_RANGE.start() as u128..=*FIFTY_PERCENT_RANGE.end() as u128,
0..data.len() as u32,
&mut positions,
);
positions
});
b.iter(|| column.get_between_vals(major_item..=major_item));
}
#[bench]
fn bench_intfastfield_getrange_u128_single_hit(b: &mut Bencher) {
let data = get_data_50percent_item();
let (_major_item, minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(
*SINGLE_ITEM_RANGE.start() as u128..=*SINGLE_ITEM_RANGE.end() as u128,
0..data.len() as u32,
&mut positions,
);
positions
});
b.iter(|| column.get_between_vals(minor_item..=minor_item));
}
#[bench]
fn bench_intfastfield_getrange_u128_hit_all(b: &mut Bencher) {
let data = get_data_50percent_item();
let (_major_item, _minor_item, data) = get_data_50percent_item();
let column = get_u128_column_from_data(&data);
b.iter(|| {
let mut positions = Vec::new();
column.get_docids_for_value_range(0..=u128::MAX, 0..data.len() as u32, &mut positions);
positions
});
b.iter(|| column.get_between_vals(0..=u128::MAX));
}
// U128 RANGE END
#[bench]
fn bench_intfastfield_scan_all_fflookup_u128(b: &mut Bencher) {
@@ -222,7 +137,7 @@ mod tests {
b.iter(|| {
let mut a = 0u128;
for i in 0u64..column.num_vals() as u64 {
a += column.get_val(i as u32);
a += column.get_val(i);
}
a
});
@@ -236,7 +151,7 @@ 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);
a += column.get_val(i as u64);
}
a
});
@@ -261,9 +176,9 @@ mod tests {
let n = permutation.len();
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0;
let mut a = 0u64;
for i in (0..n / 7).map(|val| val * 7) {
a += column.get_val(i as u32);
a += column.get_val(i as u64);
}
a
});
@@ -276,7 +191,7 @@ mod tests {
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0u32..n as u32 {
for i in 0u64..n as u64 {
a += column.get_val(i);
}
a
@@ -290,8 +205,8 @@ mod tests {
let column: Arc<dyn Column<u64>> = serialize_and_load(&permutation);
b.iter(|| {
let mut a = 0u64;
for i in 0..n {
a += column.get_val(i as u32);
for i in 0..n as u64 {
a += column.get_val(i);
}
a
});

33
columnar/src/column_values/bitpacked.rs → fastfield_codecs/src/bitpacked.rs
View File

@@ -1,10 +1,10 @@
use std::io::{self, Write};
use common::OwnedBytes;
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use super::serialize::NormalizedHeader;
use super::{ColumnValues, FastFieldCodec, FastFieldCodecType};
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType};
/// Depending on the field type, a different
/// fast field is required.
@@ -15,9 +15,9 @@ pub struct BitpackedReader {
normalized_header: NormalizedHeader,
}
impl ColumnValues for BitpackedReader {
impl Column for BitpackedReader {
#[inline]
fn get_val(&self, doc: u32) -> u64 {
fn get_val(&self, doc: u64) -> u64 {
self.bit_unpacker.get(doc, &self.data)
}
#[inline]
@@ -30,7 +30,7 @@ impl ColumnValues for BitpackedReader {
self.normalized_header.max_value
}
#[inline]
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
self.normalized_header.num_vals
}
}
@@ -64,7 +64,7 @@ impl FastFieldCodec for BitpackedCodec {
/// 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<()> {
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()> {
assert_eq!(column.min_value(), 0u64);
let num_bits = compute_num_bits(column.max_value());
let mut bit_packer = BitPacker::new();
@@ -75,7 +75,7 @@ impl FastFieldCodec for BitpackedCodec {
Ok(())
}
fn estimate(column: &dyn ColumnValues) -> Option<f32> {
fn estimate(column: &dyn Column) -> 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)
@@ -85,19 +85,19 @@ impl FastFieldCodec for BitpackedCodec {
#[cfg(test)]
mod tests {
use super::*;
use crate::column_values::tests::create_and_validate;
use crate::tests::get_codec_test_datasets;
fn create_and_validate_bitpacked_codec(data: &[u64], name: &str) {
create_and_validate::<BitpackedCodec>(data, name);
fn create_and_validate(data: &[u64], name: &str) {
crate::tests::create_and_validate::<BitpackedCodec>(data, name);
}
#[test]
fn test_with_codec_data_sets() {
let data_sets = crate::column_values::tests::get_codec_test_datasets();
let data_sets = get_codec_test_datasets();
for (mut data, name) in data_sets {
create_and_validate_bitpacked_codec(&data, name);
create_and_validate(&data, name);
data.reverse();
create_and_validate::<BitpackedCodec>(&data, name);
create_and_validate(&data, name);
}
}
@@ -107,9 +107,10 @@ mod tests {
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");
create_and_validate(&data, "rand");
data.reverse();
create_and_validate::<BitpackedCodec>(&data, "rand");
create_and_validate(&data, "rand");
}
}
}

42
columnar/src/column_values/blockwise_linear.rs → fastfield_codecs/src/blockwise_linear.rs
View File

@@ -1,12 +1,13 @@
use std::sync::Arc;
use std::{io, iter};
use common::{BinarySerializable, CountingWriter, DeserializeFrom, OwnedBytes};
use common::{BinarySerializable, CountingWriter, DeserializeFrom};
use ownedbytes::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};
use crate::line::Line;
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType, VecColumn};
const CHUNK_SIZE: usize = 512;
@@ -35,18 +36,18 @@ impl BinarySerializable for Block {
}
}
fn compute_num_blocks(num_vals: u32) -> usize {
fn compute_num_blocks(num_vals: u64) -> usize {
(num_vals as usize + CHUNK_SIZE - 1) / CHUNK_SIZE
}
pub struct BlockwiseLinearCodec;
impl FastFieldCodec for BlockwiseLinearCodec {
const CODEC_TYPE: FastFieldCodecType = FastFieldCodecType::BlockwiseLinear;
const CODEC_TYPE: crate::FastFieldCodecType = FastFieldCodecType::BlockwiseLinear;
type Reader = BlockwiseLinearReader;
fn open_from_bytes(
bytes: common::OwnedBytes,
bytes: ownedbytes::OwnedBytes,
normalized_header: NormalizedHeader,
) -> io::Result<Self::Reader> {
let footer_len: u32 = (&bytes[bytes.len() - 4..]).deserialize()?;
@@ -70,14 +71,14 @@ impl FastFieldCodec for BlockwiseLinearCodec {
}
// Estimate first_chunk and extrapolate
fn estimate(column: &dyn ColumnValues) -> Option<f32> {
if column.num_vals() < 10 * CHUNK_SIZE as u32 {
fn estimate(column: &dyn crate::Column) -> Option<f32> {
if column.num_vals() < 10 * CHUNK_SIZE as u64 {
return None;
}
let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE).collect();
let mut first_chunk: Vec<u64> = column.iter().take(CHUNK_SIZE as usize).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);
let interpolated_val = line.eval(i as u64);
*buffer_val = buffer_val.wrapping_sub(interpolated_val);
}
let estimated_bit_width = first_chunk
@@ -94,12 +95,12 @@ 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() as u64 / CHUNK_SIZE as u64);
+ metadata_per_block as u64 * (column.num_vals() / 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<()> {
fn serialize(column: &dyn Column, wrt: &mut impl io::Write) -> io::Result<()> {
// The BitpackedReader assumes a normalized vector.
assert_eq!(column.min_value(), 0);
let mut buffer = Vec::with_capacity(CHUNK_SIZE);
@@ -120,7 +121,7 @@ impl FastFieldCodec for BlockwiseLinearCodec {
assert!(!buffer.is_empty());
for (i, buffer_val) in buffer.iter_mut().enumerate() {
let interpolated_val = line.eval(i as u32);
let interpolated_val = line.eval(i as u64);
*buffer_val = buffer_val.wrapping_sub(interpolated_val);
}
let bit_width = buffer.iter().copied().map(compute_num_bits).max().unwrap();
@@ -158,11 +159,11 @@ pub struct BlockwiseLinearReader {
data: OwnedBytes,
}
impl ColumnValues for BlockwiseLinearReader {
impl Column for BlockwiseLinearReader {
#[inline(always)]
fn get_val(&self, idx: u32) -> u64 {
let block_id = (idx / CHUNK_SIZE as u32) as usize;
let idx_within_block = idx % (CHUNK_SIZE as u32);
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);
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..];
@@ -170,19 +171,16 @@ impl ColumnValues 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
}
#[inline(always)]
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
self.normalized_header.num_vals
}
}

135
columnar/src/column_values/column.rs → fastfield_codecs/src/column.rs
View File

@@ -1,15 +1,12 @@
use std::fmt::Debug;
use std::marker::PhantomData;
use std::ops::{Range, RangeInclusive};
use std::ops::RangeInclusive;
use tantivy_bitpacker::minmax;
use crate::column_values::monotonic_mapping::StrictlyMonotonicFn;
use crate::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 + Debug = u64>: Send + Sync {
/// `Column` provides columnar access on a field.
pub trait Column<T: PartialOrd = u64>: Send + Sync {
/// Return the value associated with the given idx.
///
/// This accessor should return as fast as possible.
@@ -17,7 +14,7 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
/// # Panics
///
/// May panic if `idx` is greater than the column length.
fn get_val(&self, idx: u32) -> T;
fn get_val(&self, idx: u64) -> T;
/// Fills an output buffer with the fast field values
/// associated with the `DocId` going from
@@ -30,27 +27,21 @@ pub trait ColumnValues<T: PartialOrd + Debug = 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 as u32);
*out = self.get_val(idx);
}
}
/// Get the positions of values which are in the provided value range.
///
/// Note that position == docid for single value fast fields
/// Return the positions of values which are in the provided range.
#[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 {
fn get_between_vals(&self, range: RangeInclusive<T>) -> Vec<u64> {
let mut vals = Vec::new();
for idx in 0..self.num_vals() {
let val = self.get_val(idx);
if value_range.contains(&val) {
positions.push(idx);
if range.contains(&val) {
vals.push(idx);
}
}
vals
}
/// Returns the minimum value for this fast field.
@@ -70,7 +61,7 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
fn max_value(&self) -> T;
/// The number of values in the column.
fn num_vals(&self) -> u32;
fn num_vals(&self) -> u64;
/// Returns a iterator over the data
fn iter<'a>(&'a self) -> Box<dyn Iterator<Item = T> + 'a> {
@@ -78,34 +69,15 @@ pub trait ColumnValues<T: PartialOrd + Debug = u64>: Send + Sync {
}
}
impl<T: Copy + PartialOrd + Debug> ColumnValues<T> for std::sync::Arc<dyn ColumnValues<T>> {
fn get_val(&self, idx: u32) -> T {
self.as_ref().get_val(idx)
}
fn min_value(&self) -> T {
self.as_ref().min_value()
}
fn max_value(&self) -> T {
self.as_ref().max_value()
}
fn num_vals(&self) -> u32 {
self.as_ref().num_vals()
}
fn iter<'b>(&'b self) -> Box<dyn Iterator<Item = T> + 'b> {
self.as_ref().iter()
}
fn get_range(&self, start: u64, output: &mut [T]) {
self.as_ref().get_range(start, output)
}
/// VecColumn provides `Column` over a slice.
pub struct VecColumn<'a, T = u64> {
values: &'a [T],
min_value: T,
max_value: T,
}
impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd + Debug> ColumnValues<T> for &'a C {
fn get_val(&self, idx: u32) -> T {
impl<'a, C: Column<T>, T: Copy + PartialOrd> Column<T> for &'a C {
fn get_val(&self, idx: u64) -> T {
(*self).get_val(idx)
}
@@ -117,7 +89,7 @@ impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd + Debug> ColumnValues
(*self).max_value()
}
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
(*self).num_vals()
}
@@ -130,15 +102,8 @@ impl<'a, C: ColumnValues<T> + ?Sized, T: Copy + PartialOrd + Debug> ColumnValues
}
}
/// 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 + Debug> ColumnValues<T> for VecColumn<'a, T> {
fn get_val(&self, position: u32) -> T {
impl<'a, T: Copy + PartialOrd + Send + Sync> Column<T> for VecColumn<'a, T> {
fn get_val(&self, position: u64) -> T {
self.values[position as usize]
}
@@ -154,8 +119,8 @@ impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> ColumnValues<T> for VecColu
self.max_value
}
fn num_vals(&self) -> u32 {
self.values.len() as u32
fn num_vals(&self) -> u64 {
self.values.len() as u64
}
fn get_range(&self, start: u64, output: &mut [T]) {
@@ -163,7 +128,7 @@ impl<'a, T: Copy + PartialOrd + Send + Sync + Debug> ColumnValues<T> for VecColu
}
}
impl<'a, T: Copy + PartialOrd + Default, V> From<&'a V> for VecColumn<'a, T>
impl<'a, T: Copy + Ord + Default, V> From<&'a V> for VecColumn<'a, T>
where V: AsRef<[T]> + ?Sized
{
fn from(values: &'a V) -> Self {
@@ -191,7 +156,7 @@ struct MonotonicMappingColumn<C, T, Input> {
/// 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> `
/// `fn get_between_vals(&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.
///
@@ -201,12 +166,12 @@ struct MonotonicMappingColumn<C, T, Input> {
pub fn monotonic_map_column<C, T, Input, Output>(
from_column: C,
monotonic_mapping: T,
) -> impl ColumnValues<Output>
) -> impl Column<Output>
where
C: ColumnValues<Input>,
C: Column<Input>,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Debug + Send + Sync + Clone,
Output: PartialOrd + Debug + Send + Sync + Clone,
Input: PartialOrd + Send + Sync + Clone,
Output: PartialOrd + Send + Sync + Clone,
{
MonotonicMappingColumn {
from_column,
@@ -215,15 +180,15 @@ where
}
}
impl<C, T, Input, Output> ColumnValues<Output> for MonotonicMappingColumn<C, T, Input>
impl<C, T, Input, Output> Column<Output> for MonotonicMappingColumn<C, T, Input>
where
C: ColumnValues<Input>,
C: Column<Input>,
T: StrictlyMonotonicFn<Input, Output> + Send + Sync,
Input: PartialOrd + Send + Debug + Sync + Clone,
Output: PartialOrd + Send + Debug + Sync + Clone,
Input: PartialOrd + Send + Sync + Clone,
Output: PartialOrd + Send + Sync + Clone,
{
#[inline]
fn get_val(&self, idx: u32) -> Output {
fn get_val(&self, idx: u64) -> Output {
let from_val = self.from_column.get_val(idx);
self.monotonic_mapping.mapping(from_val)
}
@@ -238,7 +203,7 @@ where
self.monotonic_mapping.mapping(from_max_value)
}
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
self.from_column.num_vals()
}
@@ -250,17 +215,10 @@ where
)
}
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(
fn get_between_vals(&self, range: RangeInclusive<Output>) -> Vec<u64> {
self.from_column.get_between_vals(
self.monotonic_mapping.inverse(range.start().clone())
..=self.monotonic_mapping.inverse(range.end().clone()),
doc_id_range,
positions,
)
}
@@ -268,7 +226,6 @@ where
// 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>
@@ -279,12 +236,12 @@ where T: Iterator + Clone + ExactSizeIterator
}
}
impl<T> ColumnValues<T::Item> for IterColumn<T>
impl<T> Column<T::Item> for IterColumn<T>
where
T: Iterator + Clone + ExactSizeIterator + Send + Sync,
T::Item: PartialOrd + Debug,
T::Item: PartialOrd,
{
fn get_val(&self, idx: u32) -> T::Item {
fn get_val(&self, idx: u64) -> T::Item {
self.0.clone().nth(idx as usize).unwrap()
}
@@ -296,8 +253,8 @@ where
self.0.clone().last().unwrap()
}
fn num_vals(&self) -> u32 {
self.0.len() as u32
fn num_vals(&self) -> u64 {
self.0.len() as u64
}
fn iter(&self) -> Box<dyn Iterator<Item = T::Item> + '_> {
@@ -308,7 +265,7 @@ where
#[cfg(test)]
mod tests {
use super::*;
use crate::column_values::monotonic_mapping::{
use crate::monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternalBaseval,
StrictlyMonotonicMappingToInternalGCDBaseval,
};

0
columnar/src/column_values/compact_space/blank_range.rs → fastfield_codecs/src/compact_space/blank_range.rs
View File

4
columnar/src/column_values/compact_space/build_compact_space.rs → 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: u32,
total_num_values: u64,
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;
compact_start += covered_range_len as u64;
}
// println!("num ranges {}", ranges_mapping.len());
CompactSpace { ranges_mapping }

664
columnar/src/column_values/compact_space/mod.rs → fastfield_codecs/src/compact_space/mod.rs
View File

@@ -14,14 +14,15 @@ use std::{
cmp::Ordering,
collections::BTreeSet,
io::{self, Write},
ops::{Range, RangeInclusive},
ops::RangeInclusive,
};
use common::{BinarySerializable, CountingWriter, OwnedBytes, VInt, VIntU128};
use common::{BinarySerializable, CountingWriter, VInt, VIntU128};
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{self, BitPacker, BitUnpacker};
use crate::column_values::compact_space::build_compact_space::get_compact_space;
use crate::column_values::ColumnValues;
use crate::compact_space::build_compact_space::get_compact_space;
use crate::Column;
mod blank_range;
mod build_compact_space;
@@ -96,7 +97,7 @@ impl BinarySerializable for CompactSpace {
};
let range_length = range_mapping.range_length();
ranges_mapping.push(range_mapping);
compact_start += range_length;
compact_start += range_length as u64;
}
Ok(Self { ranges_mapping })
@@ -164,13 +165,13 @@ pub struct IPCodecParams {
bit_unpacker: BitUnpacker,
min_value: u128,
max_value: u128,
num_vals: u32,
num_vals: u64,
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 {
pub fn train_from(iter: impl Iterator<Item = u128>, num_vals: u64) -> Self {
let mut values_sorted = BTreeSet::new();
values_sorted.extend(iter);
let total_num_values = num_vals;
@@ -199,7 +200,7 @@ impl CompactSpaceCompressor {
bit_unpacker: BitUnpacker::new(num_bits),
min_value,
max_value,
num_vals: total_num_values,
num_vals: total_num_values as u64,
num_bits,
},
}
@@ -266,7 +267,7 @@ 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 u32;
let num_vals = VIntU128::deserialize(reader)?.0 as u64;
let num_bits = u8::deserialize(reader)?;
let compact_space = CompactSpace::deserialize(reader)?;
@@ -281,9 +282,9 @@ impl BinarySerializable for IPCodecParams {
}
}
impl ColumnValues<u128> for CompactSpaceDecompressor {
impl Column<u128> for CompactSpaceDecompressor {
#[inline]
fn get_val(&self, doc: u32) -> u128 {
fn get_val(&self, doc: u64) -> u128 {
self.get(doc)
}
@@ -295,7 +296,7 @@ impl ColumnValues<u128> for CompactSpaceDecompressor {
self.max_value()
}
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
self.params.num_vals
}
@@ -303,15 +304,8 @@ impl ColumnValues<u128> for CompactSpaceDecompressor {
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)
fn get_between_vals(&self, range: RangeInclusive<u128>) -> Vec<u64> {
self.get_between_vals(range)
}
}
@@ -346,19 +340,12 @@ impl CompactSpaceDecompressor {
/// 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;
pub fn get_between_vals(&self, range: RangeInclusive<u128>) -> Vec<u64> {
if range.start() > range.end() {
return Vec::new();
}
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();
let from_value = *range.start();
let to_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);
@@ -366,7 +353,7 @@ 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,
(Err(pos1), Err(pos2)) if pos1 == pos2 => return Vec::new(),
_ => {}
}
@@ -388,20 +375,19 @@ impl CompactSpaceDecompressor {
});
let range = compact_from..=compact_to;
let scan_num_docs = position_range.end - position_range.start;
let mut positions = Vec::new();
let step_size = 4;
let cutoff = position_range.start + scan_num_docs - scan_num_docs % step_size;
let cutoff = self.params.num_vals - self.params.num_vals % 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);
let get_val = |idx| self.params.bit_unpacker.get(idx as u64, &self.data);
// unrolled loop
for idx in (position_range.start..cutoff).step_by(step_size as usize) {
for idx in (0..cutoff).step_by(step_size as usize) {
let idx1 = idx;
let idx2 = idx + 1;
let idx3 = idx + 2;
@@ -417,14 +403,17 @@ impl CompactSpaceDecompressor {
}
// handle rest
for idx in cutoff..position_range.end {
for idx in cutoff..self.params.num_vals {
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, &self.data))
(0..self.params.num_vals)
.map(move |idx| self.params.bit_unpacker.get(idx as u64, &self.data) as u64)
}
#[inline]
@@ -436,7 +425,7 @@ impl CompactSpaceDecompressor {
}
#[inline]
pub fn get(&self, idx: u32) -> u128 {
pub fn get(&self, idx: u64) -> u128 {
let compact = self.params.bit_unpacker.get(idx, &self.data);
self.compact_to_u128(compact)
}
@@ -450,364 +439,233 @@ impl CompactSpaceDecompressor {
}
}
// 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);
// }
// }
// }
//
#[cfg(test)]
mod tests {
use super::*;
use crate::{open_u128, serialize_u128};
#[test]
fn compact_space_test() {
let ips = &[
2u128, 4u128, 1000, 1001, 1002, 1003, 1004, 1005, 1008, 1010, 1012, 1260,
]
.into_iter()
.collect();
let compact_space = get_compact_space(ips, ips.len() as u64, 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 u64, 1);
let amplitude = compact_space.amplitude_compact_space();
assert_eq!(amplitude, 2);
}
fn test_all(data: OwnedBytes, expected: &[u128]) {
let decompressor = CompactSpaceDecompressor::open(data).unwrap();
for (idx, expected_val) in expected.iter().cloned().enumerate() {
let val = decompressor.get(idx as u64);
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)
.collect::<Vec<_>>();
let positions = decompressor.get_between_vals(range);
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 u64,
&mut out,
)
.unwrap();
let data = OwnedBytes::new(out);
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 data = test_aux_vals(vals);
let decomp = CompactSpaceDecompressor::open(data).unwrap();
let positions = decomp.get_between_vals(0..=1);
assert_eq!(positions, vec![0]);
let positions = decomp.get_between_vals(0..=2);
assert_eq!(positions, vec![0]);
let positions = decomp.get_between_vals(0..=3);
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!(
decomp.get_between_vals(4_000_211_221u128..=5_000_000_000u128),
vec![6, 7]
);
}
#[test]
fn test_empty() {
let vals = &[];
let data = test_aux_vals(vals);
let _decomp = CompactSpaceDecompressor::open(data).unwrap();
}
#[test]
fn test_range_2() {
let vals = &[
100u128,
99999u128,
100000u128,
100001u128,
4_000_211_221u128,
4_000_211_222u128,
333u128,
];
let data = test_aux_vals(vals);
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]);
}
#[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 u64, &mut out).unwrap();
let decomp = open_u128::<u128>(OwnedBytes::new(out)).unwrap();
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]);
}
#[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);
}
}
}

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

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

503
fastfield_codecs/src/lib.rs
View File

@@ -7,4 +7,505 @@
//! - Encode data in different codecs.
//! - Monotonically map values to u64/u128
pub use columnar::ColumnValues as Column;
#[cfg(test)]
#[macro_use]
extern crate more_asserts;
#[cfg(all(test, feature = "unstable"))]
extern crate test;
use std::io;
use std::io::Write;
use std::sync::Arc;
use common::BinarySerializable;
use compact_space::CompactSpaceDecompressor;
use monotonic_mapping::{
StrictlyMonotonicMappingInverter, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalBaseval, StrictlyMonotonicMappingToInternalGCDBaseval,
};
use ownedbytes::OwnedBytes;
use serialize::Header;
mod bitpacked;
mod blockwise_linear;
mod compact_space;
mod line;
mod linear;
mod monotonic_mapping;
mod monotonic_mapping_u128;
mod column;
mod gcd;
mod serialize;
use self::bitpacked::BitpackedCodec;
use self::blockwise_linear::BlockwiseLinearCodec;
pub use self::column::{monotonic_map_column, Column, VecColumn};
use self::linear::LinearCodec;
pub use self::monotonic_mapping::{MonotonicallyMappableToU64, StrictlyMonotonicFn};
pub use self::monotonic_mapping_u128::MonotonicallyMappableToU128;
pub use self::serialize::{
estimate, serialize, serialize_and_load, serialize_u128, NormalizedHeader,
};
#[derive(PartialEq, Eq, PartialOrd, Ord, Debug, Clone, Copy)]
#[repr(u8)]
/// Available codecs to use to encode the u64 (via [`MonotonicallyMappableToU64`]) converted data.
pub enum FastFieldCodecType {
/// Bitpack all values in the value range. The number of bits is defined by the amplitude
/// `column.max_value() - column.min_value()`
Bitpacked = 1,
/// Linear interpolation puts a line between the first and last value and then bitpacks the
/// values by the offset from the line. The number of bits is defined by the max deviation from
/// the line.
Linear = 2,
/// Same as [`FastFieldCodecType::Linear`], but encodes in blocks of 512 elements.
BlockwiseLinear = 3,
}
impl BinarySerializable for FastFieldCodecType {
fn serialize<W: Write>(&self, wrt: &mut W) -> io::Result<()> {
self.to_code().serialize(wrt)
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let code = u8::deserialize(reader)?;
let codec_type: Self = Self::from_code(code)
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Unknown code `{code}.`"))?;
Ok(codec_type)
}
}
impl FastFieldCodecType {
pub(crate) fn to_code(self) -> u8 {
self as u8
}
pub(crate) fn from_code(code: u8) -> Option<Self> {
match code {
1 => Some(Self::Bitpacked),
2 => Some(Self::Linear),
3 => Some(Self::BlockwiseLinear),
_ => None,
}
}
}
/// Returns the correct codec reader wrapped in the `Arc` for the data.
pub fn open_u128<Item: MonotonicallyMappableToU128>(
bytes: OwnedBytes,
) -> io::Result<Arc<dyn Column<Item>>> {
let 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>(
mut bytes: OwnedBytes,
) -> io::Result<Arc<dyn Column<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 Column<Item>>> {
let normalized_header = header.normalized();
let reader = C::open_from_bytes(bytes, normalized_header)?;
let min_value = header.min_value;
if let Some(gcd) = header.gcd {
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalGCDBaseval::new(gcd.get(), min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
} else {
let mapping = StrictlyMonotonicMappingInverter::from(
StrictlyMonotonicMappingToInternalBaseval::new(min_value),
);
Ok(Arc::new(monotonic_map_column(reader, mapping)))
}
}
/// The FastFieldSerializerEstimate trait is required on all variants
/// of fast field compressions, to decide which one to choose.
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: Column<u64> + 'static;
/// Reads the metadata and returns the CodecReader
fn open_from_bytes(bytes: OwnedBytes, header: NormalizedHeader) -> io::Result<Self::Reader>;
/// Serializes the data using the serializer into write.
///
/// The column iterator should be preferred over using column `get_val` method for
/// performance reasons.
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()>;
/// Returns an estimate of the compression ratio.
/// If the codec is not applicable, returns `None`.
///
/// The baseline is uncompressed 64bit data.
///
/// It could make sense to also return a value representing
/// computational complexity.
fn estimate(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,
FastFieldCodecType::Linear,
];
#[cfg(test)]
mod tests {
use proptest::prelude::*;
use proptest::strategy::Strategy;
use proptest::{prop_oneof, proptest};
use crate::bitpacked::BitpackedCodec;
use crate::blockwise_linear::BlockwiseLinearCodec;
use crate::linear::LinearCodec;
use crate::serialize::Header;
pub(crate) fn create_and_validate<Codec: FastFieldCodec>(
data: &[u64],
name: &str,
) -> Option<(f32, f32)> {
let col = &VecColumn::from(data);
let header = Header::compute_header(col, &[Codec::CODEC_TYPE])?;
let normalized_col = header.normalize_column(col);
let estimation = Codec::estimate(&normalized_col)?;
let mut out = Vec::new();
let col = VecColumn::from(data);
serialize(col, &mut out, &[Codec::CODEC_TYPE]).unwrap();
let 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);
for (doc, orig_val) in data.iter().copied().enumerate() {
let val = reader.get_val(doc as u64);
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<u64> = data
.iter()
.enumerate()
.filter(|(_, el)| **el == data[test_rand_idx])
.map(|(pos, _)| pos as u64)
.collect();
let positions = reader.get_between_vals(data[test_rand_idx]..=data[test_rand_idx]);
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)> =
crate::tests::create_and_validate::<C>(&data, dataset_name);
let result = if let Some((estimate, actual)) = estimate_actual_opt {
format!("Estimate `{estimate}` Actual `{actual}`")
} else {
"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);
}
}
#[cfg(all(test, feature = "unstable"))]
mod bench {
use std::sync::Arc;
use ownedbytes::OwnedBytes;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use test::{self, Bencher};
use super::*;
use crate::Column;
fn get_data() -> Vec<u64> {
let mut rng = StdRng::seed_from_u64(2u64);
let mut data: Vec<_> = (100..55000_u64)
.map(|num| num + rng.gen::<u8>() as u64)
.collect();
data.push(99_000);
data.insert(1000, 2000);
data.insert(2000, 100);
data.insert(3000, 4100);
data.insert(4000, 100);
data.insert(5000, 800);
data
}
#[inline(never)]
fn value_iter() -> impl Iterator<Item = u64> {
0..20_000
}
fn get_reader_for_bench<Codec: FastFieldCodec>(data: &[u64]) -> Codec::Reader {
let mut bytes = Vec::new();
let min_value = *data.iter().min().unwrap();
let data = data.iter().map(|el| *el - min_value).collect::<Vec<_>>();
let col = VecColumn::from(&data);
let normalized_header = crate::NormalizedHeader {
num_vals: col.num_vals(),
max_value: col.max_value(),
};
Codec::serialize(&VecColumn::from(&data), &mut bytes).unwrap();
Codec::open_from_bytes(OwnedBytes::new(bytes), normalized_header).unwrap()
}
fn bench_get<Codec: FastFieldCodec>(b: &mut Bencher, data: &[u64]) {
let col = get_reader_for_bench::<Codec>(data);
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u64);
sum = sum.wrapping_add(val);
}
sum
});
}
#[inline(never)]
fn bench_get_dynamic_helper(b: &mut Bencher, col: Arc<dyn Column>) {
b.iter(|| {
let mut sum = 0u64;
for pos in value_iter() {
let val = col.get_val(pos as u64);
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);
}
}

26
columnar/src/column_values/line.rs → fastfield_codecs/src/line.rs
View File

@@ -1,9 +1,9 @@
use std::io;
use std::num::NonZeroU32;
use std::num::NonZeroU64;
use common::{BinarySerializable, VInt};
use crate::column_values::ColumnValues;
use crate::Column;
const MID_POINT: u64 = (1u64 << 32) - 1u64;
@@ -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: NonZeroU32) -> u64 {
fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU64) -> 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: NonZeroU32) -> u64 {
return 0u64;
}
let abs_slope = (abs_dy << 32) / num_vals.get() as u64;
let abs_slope = (abs_dy << 32) / num_vals.get();
if sign {
abs_slope
} else {
@@ -62,8 +62,8 @@ fn compute_slope(y0: u64, y1: u64, num_vals: NonZeroU32) -> u64 {
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;
pub fn eval(&self, x: u64) -> u64 {
let linear_part = (x.wrapping_mul(self.slope) >> 32) as i32 as u64;
self.intercept.wrapping_add(linear_part)
}
@@ -75,7 +75,7 @@ impl Line {
Self::train_from(
first_val,
last_val,
num_vals as u32,
num_vals,
sample_positions_and_values.iter().cloned(),
)
}
@@ -84,11 +84,11 @@ impl Line {
fn train_from(
first_val: u64,
last_val: u64,
num_vals: u32,
num_vals: u64,
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) {
let idx_last_val = if let Some(idx_last_val) = NonZeroU64::new(num_vals - 1) {
idx_last_val
} else {
return Line::default();
@@ -129,7 +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 as u32)))
.map(|(pos, y)| y.wrapping_sub(line.eval(pos)))
.min_by_key(|&val| val.wrapping_sub(heuristic_shift))
.unwrap_or(0u64); //< Never happens.
line
@@ -145,7 +145,7 @@ 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 ColumnValues) -> Self {
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(
@@ -174,7 +174,7 @@ impl BinarySerializable for Line {
#[cfg(test)]
mod tests {
use super::*;
use crate::column_values::VecColumn;
use crate::VecColumn;
/// Test training a line and ensuring that the maximum difference between
/// the data points and the line is `expected`.
@@ -199,7 +199,7 @@ mod tests {
let line = Line::train(&VecColumn::from(&ys));
ys.iter()
.enumerate()
.map(|(x, y)| y.wrapping_sub(line.eval(x as u32)))
.map(|(x, y)| y.wrapping_sub(line.eval(x as u64)))
.max()
}

35
columnar/src/column_values/linear.rs → fastfield_codecs/src/linear.rs
View File

@@ -1,11 +1,12 @@
use std::io::{self, Write};
use common::{BinarySerializable, OwnedBytes};
use common::BinarySerializable;
use ownedbytes::OwnedBytes;
use tantivy_bitpacker::{compute_num_bits, BitPacker, BitUnpacker};
use super::line::Line;
use super::serialize::NormalizedHeader;
use super::{ColumnValues, FastFieldCodec, FastFieldCodecType};
use crate::line::Line;
use crate::serialize::NormalizedHeader;
use crate::{Column, FastFieldCodec, FastFieldCodecType};
/// Depending on the field type, a different
/// fast field is required.
@@ -16,27 +17,27 @@ pub struct LinearReader {
header: NormalizedHeader,
}
impl ColumnValues for LinearReader {
impl Column for LinearReader {
#[inline]
fn get_val(&self, doc: u32) -> u64 {
fn get_val(&self, doc: u64) -> 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)]
#[inline]
fn min_value(&self) -> u64 {
// The LinearReader assumes a normalized vector.
0u64
}
#[inline(always)]
#[inline]
fn max_value(&self) -> u64 {
self.header.max_value
}
#[inline]
fn num_vals(&self) -> u32 {
fn num_vals(&self) -> u64 {
self.header.num_vals
}
}
@@ -84,7 +85,7 @@ impl FastFieldCodec for LinearCodec {
}
/// Creates a new fast field serializer.
fn serialize(column: &dyn ColumnValues, write: &mut impl Write) -> io::Result<()> {
fn serialize(column: &dyn Column, write: &mut impl Write) -> io::Result<()> {
assert_eq!(column.min_value(), 0);
let line = Line::train(column);
@@ -92,7 +93,7 @@ impl FastFieldCodec for LinearCodec {
.iter()
.enumerate()
.map(|(pos, actual_value)| {
let calculated_value = line.eval(pos as u32);
let calculated_value = line.eval(pos as u64);
actual_value.wrapping_sub(calculated_value)
})
.max()
@@ -107,7 +108,7 @@ 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 u32);
let calculated_value = line.eval(pos as u64);
let offset = actual_value.wrapping_sub(calculated_value);
bit_packer.write(offset, num_bits, write)?;
}
@@ -120,7 +121,7 @@ impl FastFieldCodec for LinearCodec {
/// 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> {
fn estimate(column: &dyn Column) -> Option<f32> {
if column.num_vals() < 3 {
return None; // disable compressor for this case
}
@@ -139,7 +140,7 @@ impl FastFieldCodec for LinearCodec {
let estimated_bit_width = sample_positions_and_values
.into_iter()
.map(|(pos, actual_value)| {
let interpolated_val = line.eval(pos as u32);
let interpolated_val = line.eval(pos as u64);
actual_value.wrapping_sub(interpolated_val)
})
.map(|diff| ((diff as f32 * 1.5) * 2.0) as u64)
@@ -159,10 +160,10 @@ mod tests {
use rand::RngCore;
use super::*;
use crate::column_values::tests;
use crate::tests::get_codec_test_datasets;
fn create_and_validate(data: &[u64], name: &str) -> Option<(f32, f32)> {
tests::create_and_validate::<LinearCodec>(data, name)
crate::tests::create_and_validate::<LinearCodec>(data, name)
}
#[test]
@@ -177,7 +178,7 @@ mod tests {
#[test]
fn test_with_codec_datasets() {
let data_sets = tests::get_codec_test_datasets();
let data_sets = get_codec_test_datasets();
for (mut data, name) in data_sets {
create_and_validate(&data, name);
data.reverse();

14
columnar/src/column_values/main.rs → 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(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
serialize_u128(|| dataset.iter().cloned(), dataset.len() as u64, &mut data).unwrap();
}
let compression = data.len() as f64 / (dataset.len() * 16) as f64;
println!("Compression 50_000 chunks {:.4}", compression);
@@ -103,7 +103,7 @@ fn bench_ip() {
let mut data = vec![];
{
print_time!("creation");
serialize_u128(|| dataset.iter().cloned(), dataset.len() as u32, &mut data).unwrap();
serialize_u128(|| dataset.iter().cloned(), dataset.len() as u64, &mut data).unwrap();
}
let compression = data.len() as f64 / (dataset.len() * 16) as f64;
@@ -115,15 +115,9 @@ fn bench_ip() {
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,
);
let doc_values = decompressor.get_between_vals(value..=value);
println!("{:?}", doc_values.len());
}
}

52
columnar/src/column_values/monotonic_mapping.rs → fastfield_codecs/src/monotonic_mapping.rs
View File

@@ -1,14 +1,12 @@
use std::fmt::Debug;
use std::marker::PhantomData;
use fastdivide::DividerU64;
use super::MonotonicallyMappableToU128;
use crate::RowId;
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 + Debug + Copy + Send + Sync {
pub trait MonotonicallyMappableToU64: 'static + PartialOrd + Copy + Send + Sync {
/// Converts a value to u64.
///
/// Internally all fast field values are encoded as u64.
@@ -58,12 +56,10 @@ impl<T> From<T> for StrictlyMonotonicMappingInverter<T> {
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)
}
@@ -86,12 +82,10 @@ 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)
}
@@ -101,12 +95,10 @@ 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)
}
@@ -134,13 +126,11 @@ impl StrictlyMonotonicMappingToInternalGCDBaseval {
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)
}
@@ -151,7 +141,6 @@ pub(crate) struct StrictlyMonotonicMappingToInternalBaseval {
min_value: u64,
}
impl StrictlyMonotonicMappingToInternalBaseval {
#[inline(always)]
pub(crate) fn new(min_value: u64) -> Self {
Self { min_value }
}
@@ -160,24 +149,20 @@ impl StrictlyMonotonicMappingToInternalBaseval {
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
}
@@ -195,20 +180,6 @@ impl MonotonicallyMappableToU64 for i64 {
}
}
impl MonotonicallyMappableToU64 for crate::DateTime {
#[inline(always)]
fn to_u64(self) -> u64 {
common::i64_to_u64(self.timestamp_micros)
}
#[inline(always)]
fn from_u64(val: u64) -> Self {
crate::DateTime {
timestamp_micros: common::u64_to_i64(val),
}
}
}
impl MonotonicallyMappableToU64 for bool {
#[inline(always)]
fn to_u64(self) -> u64 {
@@ -221,27 +192,11 @@ impl MonotonicallyMappableToU64 for bool {
}
}
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)
}
@@ -258,9 +213,8 @@ mod tests {
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);
test_round_trip(&StrictlyMonotonicMappingToInternal::<u128>::new(), 100u128);
// base value to i64 round trip
let mapping = StrictlyMonotonicMappingToInternalBaseval::new(100);

3
columnar/src/column_values/monotonic_mapping_u128.rs → fastfield_codecs/src/monotonic_mapping_u128.rs
View File

@@ -1,9 +1,8 @@
use std::fmt::Debug;
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 + Debug + Send + Sync {
pub trait MonotonicallyMappableToU128: 'static + PartialOrd + Copy + Send + Sync {
/// Converts a value to u128.
///
/// Internally all fast field values are encoded as u64.

181
columnar/src/column_values/serialize.rs → fastfield_codecs/src/serialize.rs
View File

@@ -1,22 +1,42 @@
use std::fmt::Debug;
// 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, VInt};
use log::warn;
use ownedbytes::OwnedBytes;
use super::bitpacked::BitpackedCodec;
use super::blockwise_linear::BlockwiseLinearCodec;
use super::linear::LinearCodec;
use super::monotonic_mapping::{
use crate::bitpacked::BitpackedCodec;
use crate::blockwise_linear::BlockwiseLinearCodec;
use crate::compact_space::CompactSpaceCompressor;
use crate::linear::LinearCodec;
use crate::monotonic_mapping::{
StrictlyMonotonicFn, StrictlyMonotonicMappingToInternal,
StrictlyMonotonicMappingToInternalGCDBaseval,
};
use super::{
monotonic_map_column, ColumnValues, FastFieldCodec, FastFieldCodecType,
MonotonicallyMappableToU64, U128FastFieldCodecType,
use crate::{
monotonic_map_column, Column, FastFieldCodec, FastFieldCodecType, MonotonicallyMappableToU64,
VecColumn, ALL_CODEC_TYPES,
};
use crate::column_values::compact_space::CompactSpaceCompressor;
/// The normalized header gives some parameters after applying the following
/// normalization of the vector:
@@ -26,14 +46,14 @@ use crate::column_values::compact_space::CompactSpaceCompressor;
#[derive(Debug, Copy, Clone)]
pub struct NormalizedHeader {
/// The number of values in the underlying column.
pub num_vals: u32,
pub num_vals: u64,
/// The max value of the underlying column.
pub max_value: u64,
}
#[derive(Debug, Copy, Clone)]
pub(crate) struct Header {
pub num_vals: u32,
pub num_vals: u64,
pub min_value: u64,
pub max_value: u64,
pub gcd: Option<NonZeroU64>,
@@ -53,18 +73,18 @@ impl Header {
}
}
pub(crate) fn normalize_column<C: ColumnValues>(&self, from_column: C) -> impl ColumnValues {
pub fn normalize_column<C: Column>(&self, from_column: C) -> impl Column {
normalize_column(from_column, self.min_value, self.gcd)
}
pub fn compute_header(
column: impl ColumnValues<u64>,
column: impl Column<u64>,
codecs: &[FastFieldCodecType],
) -> Option<Header> {
let num_vals = column.num_vals();
let min_value = column.min_value();
let max_value = column.max_value();
let gcd = super::gcd::find_gcd(column.iter().map(|val| val - min_value))
let gcd = crate::gcd::find_gcd(column.iter().map(|val| val - min_value))
.filter(|gcd| gcd.get() > 1u64);
let normalized_column = normalize_column(column, min_value, gcd);
let codec_type = detect_codec(normalized_column, codecs)?;
@@ -78,34 +98,11 @@ 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,
})
}
}
fn normalize_column<C: ColumnValues>(
pub fn normalize_column<C: Column>(
from_column: C,
min_value: u64,
gcd: Option<NonZeroU64>,
) -> impl ColumnValues {
) -> 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)
@@ -113,7 +110,7 @@ fn normalize_column<C: ColumnValues>(
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.num_vals).serialize(writer)?;
VInt(self.min_value).serialize(writer)?;
VInt(self.max_value - self.min_value).serialize(writer)?;
if let Some(gcd) = self.gcd {
@@ -126,7 +123,7 @@ impl BinarySerializable for Header {
}
fn deserialize<R: io::Read>(reader: &mut R) -> io::Result<Self> {
let num_vals = VInt::deserialize(reader)?.0 as u32;
let num_vals = VInt::deserialize(reader)?.0;
let min_value = VInt::deserialize(reader)?.0;
let amplitude = VInt::deserialize(reader)?.0;
let max_value = min_value + amplitude;
@@ -142,28 +139,46 @@ impl BinarySerializable for Header {
}
}
/// Return estimated compression for given codec in the value range [0.0..1.0], where 1.0 means no
/// compression.
pub fn estimate<T: MonotonicallyMappableToU64>(
typed_column: impl Column<T>,
codec_type: FastFieldCodecType,
) -> Option<f32> {
let column = monotonic_map_column(typed_column, StrictlyMonotonicMappingToInternal::<T>::new());
let min_value = column.min_value();
let gcd = crate::gcd::find_gcd(column.iter().map(|val| val - min_value))
.filter(|gcd| gcd.get() > 1u64);
let mapping = StrictlyMonotonicMappingToInternalGCDBaseval::new(
gcd.map(|gcd| gcd.get()).unwrap_or(1u64),
min_value,
);
let normalized_column = monotonic_map_column(&column, mapping);
match codec_type {
FastFieldCodecType::Bitpacked => BitpackedCodec::estimate(&normalized_column),
FastFieldCodecType::Linear => LinearCodec::estimate(&normalized_column),
FastFieldCodecType::BlockwiseLinear => BlockwiseLinearCodec::estimate(&normalized_column),
}
}
/// Serializes u128 values with the compact space codec.
pub fn serialize_column_values_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
pub fn serialize_u128<F: Fn() -> I, I: Iterator<Item = u128>>(
iter_gen: F,
num_vals: u32,
num_vals: u64,
output: &mut impl io::Write,
) -> io::Result<()> {
let header = U128Header {
num_vals,
codec_type: U128FastFieldCodecType::CompactSpace,
};
header.serialize(output)?;
// TODO write header, to later support more codecs
let compressor = CompactSpaceCompressor::train_from(iter_gen(), num_vals);
compressor.compress_into(iter_gen(), output)?;
compressor.compress_into(iter_gen(), output).unwrap();
Ok(())
}
/// Serializes the column with the codec with the best estimate on the data.
pub fn serialize_column_values<T: MonotonicallyMappableToU64 + Debug>(
typed_column: impl ColumnValues<T>,
codecs: &[FastFieldCodecType],
pub fn serialize<T: MonotonicallyMappableToU64>(
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(|| {
@@ -183,7 +198,7 @@ pub fn serialize_column_values<T: MonotonicallyMappableToU64 + Debug>(
}
fn detect_codec(
column: impl ColumnValues<u64>,
column: impl Column<u64>,
codecs: &[FastFieldCodecType],
) -> Option<FastFieldCodecType> {
let mut estimations = Vec::new();
@@ -210,8 +225,8 @@ fn detect_codec(
Some(estimations.first()?.1)
}
pub(crate) fn serialize_given_codec(
column: impl ColumnValues<u64>,
fn serialize_given_codec(
column: impl Column<u64>,
codec_type: FastFieldCodecType,
output: &mut impl io::Write,
) -> io::Result<()> {
@@ -226,43 +241,22 @@ pub(crate) fn serialize_given_codec(
BlockwiseLinearCodec::serialize(&column, output)?;
}
}
output.flush()?;
Ok(())
}
/// Helper function to serialize a column (autodetect from all codecs) and then open it
pub fn serialize_and_load<T: MonotonicallyMappableToU64 + Ord + Default>(
column: &[T],
) -> Arc<dyn Column<T>> {
let mut buffer = Vec::new();
super::serialize(VecColumn::from(&column), &mut buffer, &ALL_CODEC_TYPES).unwrap();
super::open(OwnedBytes::new(buffer)).unwrap()
}
#[cfg(test)]
pub mod tests {
use std::sync::Arc;
use common::OwnedBytes;
mod tests {
use super::*;
use crate::column_values::{open_u64_mapped, VecColumn};
const ALL_CODEC_TYPES: [FastFieldCodecType; 3] = [
FastFieldCodecType::Bitpacked,
FastFieldCodecType::Linear,
FastFieldCodecType::BlockwiseLinear,
];
/// 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();
serialize_column_values(&VecColumn::from(&column), &ALL_CODEC_TYPES, &mut buffer).unwrap();
open_u64_mapped(OwnedBytes::new(buffer)).unwrap()
}
#[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() {
@@ -275,19 +269,18 @@ pub mod tests {
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.
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 + 1);
assert_eq!(buffer.len(), 5 + 8);
}
#[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();
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);
assert_eq!(buffer.len(), 5 + 7);
}
#[test]
@@ -295,8 +288,8 @@ pub mod tests {
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();
serialize(col, &mut buffer, &[FastFieldCodecType::Bitpacked]).unwrap();
// Values are stored over 3 bits.
assert_eq!(buffer.len(), 7 + (3 * 80 / 8));
assert_eq!(buffer.len(), 7 + (3 * 80 / 8) + 7);
}
}

6
ownedbytes/Cargo.toml
View File

@@ -1,14 +1,10 @@
[package]
authors = ["Paul Masurel <paul@quickwit.io>", "Pascal Seitz <pascal@quickwit.io>"]
name = "ownedbytes"
version = "0.5.0"
version = "0.3.0"
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};
pub use stable_deref_trait::StableDeref;
use stable_deref_trait::StableDeref;
/// An OwnedBytes simply wraps an object that owns a slice of data and exposes
/// this data as a slice.
@@ -80,21 +80,6 @@ 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 = "0.19.0"
version = "0.18.0"
authors = ["Paul Masurel <paul.masurel@gmail.com>"]
license = "MIT"
categories = ["database-implementations", "data-structures"]

67
query-grammar/src/query_grammar.rs
View File

@@ -5,8 +5,7 @@ use combine::parser::range::{take_while, take_while1};
use combine::parser::repeat::escaped;
use combine::parser::Parser;
use combine::{
attempt, between, choice, eof, many, many1, one_of, optional, parser, satisfy, sep_by,
skip_many1, value,
attempt, choice, eof, many, many1, one_of, optional, parser, satisfy, skip_many1, value,
};
use once_cell::sync::Lazy;
use regex::Regex;
@@ -63,20 +62,6 @@ 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
@@ -196,8 +181,8 @@ 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())))
};
@@ -265,17 +250,6 @@ 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)
}
@@ -290,7 +264,6 @@ 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()
@@ -676,34 +649,6 @@ 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]
@@ -760,14 +705,6 @@ 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

@@ -12,10 +12,6 @@ pub enum UserInputLeaf {
lower: UserInputBound,
upper: UserInputBound,
},
Set {
field: Option<String>,
elements: Vec<String>,
},
}
impl Debug for UserInputLeaf {
@@ -35,19 +31,6 @@ 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, "*"),
}
}

2
run-tests.sh Executable file
View File

@@ -0,0 +1,2 @@
#!/bin/bash
cargo test

65
src/aggregation/agg_req.rs
View File

@@ -51,10 +51,7 @@ use serde::{Deserialize, Serialize};
pub use super::bucket::RangeAggregation;
use super::bucket::{HistogramAggregation, TermsAggregation};
use super::metric::{
AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
SumAggregation,
};
use super::metric::{AverageAggregation, StatsAggregation};
use super::VecWithNames;
/// The top-level aggregation request structure, which contains [`Aggregation`] and their user
@@ -240,38 +237,20 @@ impl BucketAggregationType {
/// called multi-value numeric metrics aggregation.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub enum MetricAggregation {
/// Computes the average of the extracted values.
/// Calculates the average.
#[serde(rename = "avg")]
Average(AverageAggregation),
/// 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.
/// Calculates stats sum, average, min, max, standard_deviation on a field.
#[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>) {
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(),
match self {
MetricAggregation::Average(avg) => fast_field_names.insert(avg.field.to_string()),
MetricAggregation::Stats(stats) => fast_field_names.insert(stats.field.to_string()),
};
fast_field_names.insert(fast_field_name.to_string());
}
}
@@ -279,38 +258,6 @@ 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![(

117
src/aggregation/agg_req_with_accessor.rs
View File

@@ -4,17 +4,15 @@ use std::rc::Rc;
use std::sync::atomic::AtomicU32;
use std::sync::Arc;
use columnar::{Column, StrColumn};
use fastfield_codecs::Column;
use super::agg_req::{Aggregation, Aggregations, BucketAggregationType, MetricAggregation};
use super::bucket::{HistogramAggregation, RangeAggregation, TermsAggregation};
use super::metric::{
AverageAggregation, CountAggregation, MaxAggregation, MinAggregation, StatsAggregation,
SumAggregation,
};
use super::metric::{AverageAggregation, StatsAggregation};
use super::segment_agg_result::BucketCount;
use super::VecWithNames;
use crate::schema::{FieldType, Type};
use crate::fastfield::{type_and_cardinality, FastType, MultiValuedFastFieldReader};
use crate::schema::{Cardinality, Type};
use crate::{InvertedIndexReader, SegmentReader, TantivyError};
#[derive(Clone, Default)]
@@ -36,12 +34,38 @@ impl AggregationsWithAccessor {
}
}
#[derive(Clone)]
pub(crate) enum FastFieldAccessor {
Multi(MultiValuedFastFieldReader<u64>),
Single(Arc<dyn Column<u64>>),
}
impl FastFieldAccessor {
pub fn as_single(&self) -> Option<&dyn Column<u64>> {
match self {
FastFieldAccessor::Multi(_) => None,
FastFieldAccessor::Single(reader) => Some(&**reader),
}
}
pub fn into_single(self) -> Option<Arc<dyn Column<u64>>> {
match self {
FastFieldAccessor::Multi(_) => None,
FastFieldAccessor::Single(reader) => Some(reader),
}
}
pub fn as_multi(&self) -> Option<&MultiValuedFastFieldReader<u64>> {
match self {
FastFieldAccessor::Multi(reader) => Some(reader),
FastFieldAccessor::Single(_) => None,
}
}
}
#[derive(Clone)]
pub struct BucketAggregationWithAccessor {
/// In general there can be buckets without fast field access, e.g. buckets that are created
/// based on search terms. So eventually this needs to be Option or moved.
pub(crate) accessor: Column<u64>,
pub(crate) str_dict_column: Option<StrColumn>,
pub(crate) accessor: FastFieldAccessor,
pub(crate) inverted_index: Option<Arc<InvertedIndexReader>>,
pub(crate) field_type: Type,
pub(crate) bucket_agg: BucketAggregationType,
pub(crate) sub_aggregation: AggregationsWithAccessor,
@@ -56,19 +80,23 @@ impl BucketAggregationWithAccessor {
bucket_count: Rc<AtomicU32>,
max_bucket_count: u32,
) -> crate::Result<BucketAggregationWithAccessor> {
let mut str_dict_column = None;
let mut inverted_index = None;
let (accessor, field_type) = match &bucket {
BucketAggregationType::Range(RangeAggregation {
field: field_name, ..
}) => get_ff_reader_and_validate(reader, field_name)?,
}) => get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?,
BucketAggregationType::Histogram(HistogramAggregation {
field: field_name, ..
}) => get_ff_reader_and_validate(reader, field_name)?,
}) => get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?,
BucketAggregationType::Terms(TermsAggregation {
field: field_name, ..
}) => {
str_dict_column = reader.fast_fields().str(&field_name)?;
get_ff_reader_and_validate(reader, field_name)?
let field = reader
.schema()
.get_field(field_name)
.ok_or_else(|| TantivyError::FieldNotFound(field_name.to_string()))?;
inverted_index = Some(reader.inverted_index(field)?);
get_ff_reader_and_validate(reader, field_name, Cardinality::MultiValues)?
}
};
let sub_aggregation = sub_aggregation.clone();
@@ -82,7 +110,7 @@ impl BucketAggregationWithAccessor {
max_bucket_count,
)?,
bucket_agg: bucket.clone(),
str_dict_column,
inverted_index,
bucket_count: BucketCount {
bucket_count,
max_bucket_count,
@@ -96,7 +124,7 @@ impl BucketAggregationWithAccessor {
pub struct MetricAggregationWithAccessor {
pub metric: MetricAggregation,
pub field_type: Type,
pub accessor: Column<u64>,
pub accessor: Arc<dyn Column>,
}
impl MetricAggregationWithAccessor {
@@ -106,15 +134,14 @@ impl MetricAggregationWithAccessor {
) -> crate::Result<MetricAggregationWithAccessor> {
match &metric {
MetricAggregation::Average(AverageAggregation { 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)?;
| MetricAggregation::Stats(StatsAggregation { field: field_name }) => {
let (accessor, field_type) =
get_ff_reader_and_validate(reader, field_name, Cardinality::SingleValue)?;
Ok(MetricAggregationWithAccessor {
accessor,
accessor: accessor
.into_single()
.expect("unexpected fast field cardinality"),
field_type,
metric: metric.clone(),
})
@@ -159,19 +186,41 @@ pub(crate) fn get_aggs_with_accessor_and_validate(
fn get_ff_reader_and_validate(
reader: &SegmentReader,
field_name: &str,
) -> crate::Result<(columnar::Column<u64>, Type)> {
let field = reader.schema().get_field(field_name)?;
// TODO we should get type metadata from columnar
let field_type = reader
cardinality: Cardinality,
) -> crate::Result<(FastFieldAccessor, Type)> {
let field = reader
.schema()
.get_field_entry(field)
.field_type()
.value_type();
// TODO Do validation
.get_field(field_name)
.ok_or_else(|| TantivyError::FieldNotFound(field_name.to_string()))?;
let field_type = reader.schema().get_field_entry(field).field_type();
if let Some((ff_type, field_cardinality)) = type_and_cardinality(field_type) {
if ff_type == FastType::Date {
return Err(TantivyError::InvalidArgument(
"Unsupported field type date in aggregation".to_string(),
));
}
if cardinality != field_cardinality {
return Err(TantivyError::InvalidArgument(format!(
"Invalid field cardinality on field {} expected {:?}, but got {:?}",
field_name, cardinality, field_cardinality
)));
}
} else {
return Err(TantivyError::InvalidArgument(format!(
"Only fast fields of type f64, u64, i64 are supported, but got {:?} ",
field_type.value_type()
)));
};
let ff_fields = reader.fast_fields();
let ff_field = ff_fields.u64_lenient(field_name)?.ok_or_else(|| {
TantivyError::InvalidArgument(format!("No fast field found for field: {}", field_name))
})?;
Ok((ff_field, field_type))
match cardinality {
Cardinality::SingleValue => ff_fields
.u64_lenient(field)
.map(|field| (FastFieldAccessor::Single(field), field_type.value_type())),
Cardinality::MultiValues => ff_fields
.u64s_lenient(field)
.map(|field| (FastFieldAccessor::Multi(field), field_type.value_type())),
}
}

65
src/aggregation/agg_result.rs
View File

@@ -4,7 +4,9 @@
//! intermediate average results, which is the sum and the number of values. The actual average is
//! calculated on the step from intermediate to final aggregation result tree.
use rustc_hash::FxHashMap;
use std::collections::HashMap;
use fnv::FnvHashMap;
use serde::{Deserialize, Serialize};
use super::agg_req::BucketAggregationInternal;
@@ -12,12 +14,11 @@ use super::bucket::GetDocCount;
use super::intermediate_agg_result::{IntermediateBucketResult, IntermediateMetricResult};
use super::metric::{SingleMetricResult, Stats};
use super::Key;
use crate::schema::Schema;
use crate::TantivyError;
#[derive(Clone, Default, Debug, PartialEq, Serialize, Deserialize)]
/// The final aggegation result.
pub struct AggregationResults(pub FxHashMap<String, AggregationResult>);
pub struct AggregationResults(pub HashMap<String, AggregationResult>);
impl AggregationResults {
pub(crate) fn get_value_from_aggregation(
@@ -30,7 +31,7 @@ impl AggregationResults {
} else {
// Validation is be done during request parsing, so we can't reach this state.
Err(TantivyError::InternalError(format!(
"Can't find aggregation {:?} in sub-aggregations",
"Can't find aggregation {:?} in sub_aggregations",
name
)))
}
@@ -70,51 +71,27 @@ impl AggregationResult {
pub enum MetricResult {
/// Average metric result.
Average(SingleMetricResult),
/// Count metric result.
Count(SingleMetricResult),
/// Max metric result.
Max(SingleMetricResult),
/// Min metric result.
Min(SingleMetricResult),
/// Stats metric result.
Stats(Stats),
/// Sum metric result.
Sum(SingleMetricResult),
}
impl MetricResult {
fn get_value(&self, agg_property: &str) -> crate::Result<Option<f64>> {
match self {
MetricResult::Average(avg) => Ok(avg.value),
MetricResult::Count(count) => Ok(count.value),
MetricResult::Max(max) => Ok(max.value),
MetricResult::Min(min) => Ok(min.value),
MetricResult::Stats(stats) => stats.get_value(agg_property),
MetricResult::Sum(sum) => Ok(sum.value),
}
}
}
impl From<IntermediateMetricResult> for MetricResult {
fn from(metric: IntermediateMetricResult) -> Self {
match metric {
IntermediateMetricResult::Average(intermediate_avg) => {
MetricResult::Average(intermediate_avg.finalize().into())
}
IntermediateMetricResult::Count(intermediate_count) => {
MetricResult::Count(intermediate_count.finalize().into())
}
IntermediateMetricResult::Max(intermediate_max) => {
MetricResult::Max(intermediate_max.finalize().into())
}
IntermediateMetricResult::Min(intermediate_min) => {
MetricResult::Min(intermediate_min.finalize().into())
IntermediateMetricResult::Average(avg_data) => {
MetricResult::Average(avg_data.finalize().into())
}
IntermediateMetricResult::Stats(intermediate_stats) => {
MetricResult::Stats(intermediate_stats.finalize())
}
IntermediateMetricResult::Sum(intermediate_sum) => {
MetricResult::Sum(intermediate_sum.finalize().into())
}
}
}
}
@@ -124,13 +101,13 @@ impl From<IntermediateMetricResult> for MetricResult {
#[serde(untagged)]
pub enum BucketResult {
/// This is the range entry for a bucket, which contains a key, count, from, to, and optionally
/// sub-aggregations.
/// sub_aggregations.
Range {
/// The range buckets sorted by range.
buckets: BucketEntries<RangeBucketEntry>,
},
/// This is the histogram entry for a bucket, which contains a key, count, and optionally
/// sub-aggregations.
/// sub_aggregations.
Histogram {
/// The buckets.
///
@@ -154,12 +131,9 @@ pub enum BucketResult {
}
impl BucketResult {
pub(crate) fn empty_from_req(
req: &BucketAggregationInternal,
schema: &Schema,
) -> crate::Result<Self> {
pub(crate) fn empty_from_req(req: &BucketAggregationInternal) -> crate::Result<Self> {
let empty_bucket = IntermediateBucketResult::empty_from_req(&req.bucket_agg);
empty_bucket.into_final_bucket_result(req, schema)
empty_bucket.into_final_bucket_result(req)
}
}
@@ -171,11 +145,11 @@ pub enum BucketEntries<T> {
/// Vector format bucket entries
Vec(Vec<T>),
/// HashMap format bucket entries
HashMap(FxHashMap<String, T>),
HashMap(FnvHashMap<String, T>),
}
/// This is the default entry for a bucket, which contains a key, count, and optionally
/// sub-aggregations.
/// sub_aggregations.
///
/// # JSON Format
/// ```json
@@ -202,9 +176,6 @@ pub enum BucketEntries<T> {
/// ```
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct BucketEntry {
#[serde(skip_serializing_if = "Option::is_none")]
/// The string representation of the bucket.
pub key_as_string: Option<String>,
/// The identifier of the bucket.
pub key: Key,
/// Number of documents in the bucket.
@@ -225,7 +196,7 @@ impl GetDocCount for BucketEntry {
}
/// This is the range entry for a bucket, which contains a key, count, and optionally
/// sub-aggregations.
/// sub_aggregations.
///
/// # JSON Format
/// ```json
@@ -261,7 +232,7 @@ pub struct RangeBucketEntry {
/// Number of documents in the bucket.
pub doc_count: u64,
#[serde(flatten)]
/// Sub-aggregations in this bucket.
/// sub-aggregations in this bucket.
pub sub_aggregation: AggregationResults,
/// The from range of the bucket. Equals `f64::MIN` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
@@ -269,10 +240,4 @@ pub struct RangeBucketEntry {
/// The to range of the bucket. Equals `f64::MAX` when `None`.
#[serde(skip_serializing_if = "Option::is_none")]
pub to: Option<f64>,
/// The optional string representation for the `from` range.
#[serde(skip_serializing_if = "Option::is_none")]
pub from_as_string: Option<String>,
/// The optional string representation for the `to` range.
#[serde(skip_serializing_if = "Option::is_none")]
pub to_as_string: Option<String>,
}

123
src/aggregation/bucket/histogram/date_histogram.rs
View File

@@ -1,123 +0,0 @@
use serde::{Deserialize, Serialize};
/// DateHistogramAggregation is similar to `HistogramAggregation`, but it can only be used with date type.
///
/// Currently only **fixed time** intervals are supported. Calendar-aware time intervals are not
/// supported.
///
/// Like the histogram, values are rounded down into the closest bucket.
///
/// For this calculation all fastfield values are converted to f64.
///
/// # Limitations/Compatibility
/// Only fixed time intervals are supported.
///
/// # JSON Format
/// ```json
/// {
/// "prices": {
/// "date_histogram": {
/// "field": "price",
/// "fixed_interval": "30d"
/// }
/// }
/// }
/// ```
///
/// Response
/// See [`BucketEntry`](crate::aggregation::agg_result::BucketEntry)
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
pub struct DateHistogramAggregationReq {
/// The field to aggregate on.
pub field: String,
/// The interval to chunk your data range. Each bucket spans a value range of [0..fixed_interval).
/// Accepted values
///
/// Fixed intervals are configured with the `fixed_interval` parameter.
/// In contrast to calendar-aware intervals, fixed intervals are a fixed number of SI units and never deviate, regardless of where they fall on the calendar.
/// One second is always composed of 1000ms. This allows fixed intervals to be specified in any multiple of the supported units.
/// However, it means fixed intervals cannot express other units such as months, since the duration of a month is not a fixed quantity.
/// Attempting to specify a calendar interval like month or quarter will return an Error.
///
/// The accepted units for fixed intervals are:
/// * `ms`: milliseconds
/// * `s`: seconds. Defined as 1000 milliseconds each.
/// * `m`: minutes. Defined as 60 seconds each (60_000 milliseconds).
/// * `h`: hours. Defined as 60 minutes each (3_600_000 milliseconds).
/// * `d`: days. Defined as 24 hours (86_400_000 milliseconds).
///
/// Fractional time values are not supported, but you can address this by shifting to another time unit
/// (e.g., `1.5h` could instead be specified as `90m`).
pub fixed_interval: String,
/// Intervals implicitly defines an absolute grid of buckets `[interval * k, interval * (k + 1))`.
///
pub offset: Option<String>,
/// Whether to return the buckets as a hash map
#[serde(default)]
pub keyed: bool,
}
impl DateHistogramAggregationReq {
fn validate(&self) -> crate::Result<()> {
Ok(())
}
}
#[derive(Debug, PartialEq, Eq)]
/// Errors when parsing the fixed interval for `DateHistogramAggregationReq`.
pub enum DateHistogramParseError {
/// Unit not recognized in passed String
UnitNotRecognized(String),
/// Number not found in passed String
NumberMissing(String),
/// Unit not found in passed String
UnitMissing(String),
}
fn parse_into_milliseconds(input: &str) -> Result<u64, DateHistogramParseError> {
let split_boundary = input
.char_indices()
.take_while(|(pos, el)| el.is_numeric())
.count();
let (number, unit) = input.split_at(split_boundary);
if number.is_empty() {
return Err(DateHistogramParseError::NumberMissing(input.to_string()));
}
if unit.is_empty() {
return Err(DateHistogramParseError::UnitMissing(input.to_string()));
}
let number: u64 = number.parse().unwrap();
let multiplier_from_unit = match unit {
"ms" => 1,
"s" => 1000,
"m" => 60 * 1000,
"h" => 60 * 60 * 1000,
"d" => 24 * 60 * 60 * 1000,
_ => return Err(DateHistogramParseError::UnitNotRecognized(unit.to_string())),
};
Ok(number * multiplier_from_unit)
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn parser_test() {
assert_eq!(parse_into_milliseconds("1m").unwrap(), 60_000);
assert_eq!(parse_into_milliseconds("2m").unwrap(), 120_000);
assert_eq!(
parse_into_milliseconds("2y").unwrap_err(),
DateHistogramParseError::UnitNotRecognized("y".to_string())
);
assert_eq!(
parse_into_milliseconds("2000").unwrap_err(),
DateHistogramParseError::UnitMissing("2000".to_string())
);
assert_eq!(
parse_into_milliseconds("ms").unwrap_err(),
DateHistogramParseError::NumberMissing("ms".to_string())
);
}
}

248
src/aggregation/bucket/histogram/histogram.rs
View File

@@ -1,7 +1,7 @@
use std::cmp::Ordering;
use std::fmt::Display;
use columnar::Column;
use fastfield_codecs::Column;
use itertools::Itertools;
use serde::{Deserialize, Serialize};
@@ -10,14 +10,12 @@ use crate::aggregation::agg_req_with_accessor::{
AggregationsWithAccessor, BucketAggregationWithAccessor,
};
use crate::aggregation::agg_result::BucketEntry;
use crate::aggregation::f64_from_fastfield_u64;
use crate::aggregation::intermediate_agg_result::{
IntermediateAggregationResults, IntermediateBucketResult, IntermediateHistogramBucketEntry,
};
use crate::aggregation::segment_agg_result::{
GenericSegmentAggregationResultsCollector, SegmentAggregationCollector,
};
use crate::aggregation::{f64_from_fastfield_u64, format_date};
use crate::schema::{Schema, Type};
use crate::aggregation::segment_agg_result::SegmentAggregationResultsCollector;
use crate::schema::Type;
use crate::{DocId, TantivyError};
/// Histogram is a bucket aggregation, where buckets are created dynamically for given `interval`.
@@ -64,6 +62,7 @@ use crate::{DocId, TantivyError};
///
/// Response
/// See [`BucketEntry`](crate::aggregation::agg_result::BucketEntry)
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
pub struct HistogramAggregation {
/// The field to aggregate on.
@@ -185,7 +184,7 @@ pub(crate) struct SegmentHistogramBucketEntry {
impl SegmentHistogramBucketEntry {
pub(crate) fn into_intermediate_bucket_entry(
self,
sub_aggregation: GenericSegmentAggregationResultsCollector,
sub_aggregation: SegmentAggregationResultsCollector,
agg_with_accessor: &AggregationsWithAccessor,
) -> crate::Result<IntermediateHistogramBucketEntry> {
Ok(IntermediateHistogramBucketEntry {
@@ -199,15 +198,14 @@ impl SegmentHistogramBucketEntry {
/// The collector puts values from the fast field into the correct buckets and does a conversion to
/// the correct datatype.
#[derive(Clone, Debug)]
#[derive(Clone, Debug, PartialEq)]
pub struct SegmentHistogramCollector {
/// The buckets containing the aggregation data.
buckets: Vec<SegmentHistogramBucketEntry>,
sub_aggregations: Option<Vec<GenericSegmentAggregationResultsCollector>>,
sub_aggregations: Option<Vec<SegmentAggregationResultsCollector>>,
field_type: Type,
interval: f64,
offset: f64,
min_doc_count: u64,
first_bucket_num: i64,
bounds: HistogramBounds,
}
@@ -217,30 +215,6 @@ impl SegmentHistogramCollector {
self,
agg_with_accessor: &BucketAggregationWithAccessor,
) -> crate::Result<IntermediateBucketResult> {
// Compute the number of buckets to validate against max num buckets
// Note: We use min_doc_count here, but it's only an lowerbound here, since were are on the
// intermediate level and after merging the number of documents of a bucket could exceed
// `min_doc_count`.
{
let cut_off_buckets_front = self
.buckets
.iter()
.take_while(|bucket| bucket.doc_count <= self.min_doc_count)
.count();
let cut_off_buckets_back = self.buckets[cut_off_buckets_front..]
.iter()
.rev()
.take_while(|bucket| bucket.doc_count <= self.min_doc_count)
.count();
let estimate_num_buckets =
self.buckets.len() - cut_off_buckets_front - cut_off_buckets_back;
agg_with_accessor
.bucket_count
.add_count(estimate_num_buckets as u32);
agg_with_accessor.bucket_count.validate_bucket_count()?;
}
let mut buckets = Vec::with_capacity(
self.buckets
.iter()
@@ -277,6 +251,11 @@ impl SegmentHistogramCollector {
);
};
agg_with_accessor
.bucket_count
.add_count(buckets.len() as u32);
agg_with_accessor.bucket_count.validate_bucket_count()?;
Ok(IntermediateBucketResult::Histogram { buckets })
}
@@ -284,7 +263,7 @@ impl SegmentHistogramCollector {
req: &HistogramAggregation,
sub_aggregation: &AggregationsWithAccessor,
field_type: Type,
accessor: &Column<u64>,
accessor: &dyn Column<u64>,
) -> crate::Result<Self> {
req.validate()?;
let min = f64_from_fastfield_u64(accessor.min_value(), &field_type);
@@ -301,7 +280,7 @@ impl SegmentHistogramCollector {
None
} else {
let sub_aggregation =
GenericSegmentAggregationResultsCollector::from_req_and_validate(sub_aggregation)?;
SegmentAggregationResultsCollector::from_req_and_validate(sub_aggregation)?;
Some(buckets.iter().map(|_| sub_aggregation.clone()).collect())
};
@@ -329,14 +308,13 @@ impl SegmentHistogramCollector {
first_bucket_num,
bounds,
sub_aggregations,
min_doc_count: req.min_doc_count(),
})
}
#[inline]
pub(crate) fn collect_block(
&mut self,
docs: &[DocId],
doc: &[DocId],
bucket_with_accessor: &BucketAggregationWithAccessor,
force_flush: bool,
) -> crate::Result<()> {
@@ -347,20 +325,64 @@ impl SegmentHistogramCollector {
let get_bucket_num =
|val| (get_bucket_num_f64(val, interval, offset) as i64 - first_bucket_num) as usize;
let accessor = &bucket_with_accessor.accessor;
for doc in docs {
for val in accessor.values(*doc) {
let val = self.f64_from_fastfield_u64(val);
let accessor = bucket_with_accessor
.accessor
.as_single()
.expect("unexpected fast field cardinatility");
let mut iter = doc.chunks_exact(4);
for docs in iter.by_ref() {
let val0 = self.f64_from_fastfield_u64(accessor.get_val(docs[0] as u64));
let val1 = self.f64_from_fastfield_u64(accessor.get_val(docs[1] as u64));
let val2 = self.f64_from_fastfield_u64(accessor.get_val(docs[2] as u64));
let val3 = self.f64_from_fastfield_u64(accessor.get_val(docs[3] as u64));
let bucket_pos = get_bucket_num(val);
self.increment_bucket_if_in_bounds(
val,
&bounds,
bucket_pos,
*doc,
&bucket_with_accessor.sub_aggregation,
)?;
let bucket_pos0 = get_bucket_num(val0);
let bucket_pos1 = get_bucket_num(val1);
let bucket_pos2 = get_bucket_num(val2);
let bucket_pos3 = get_bucket_num(val3);
self.increment_bucket_if_in_bounds(
val0,
&bounds,
bucket_pos0,
docs[0],
&bucket_with_accessor.sub_aggregation,
)?;
self.increment_bucket_if_in_bounds(
val1,
&bounds,
bucket_pos1,
docs[1],
&bucket_with_accessor.sub_aggregation,
)?;
self.increment_bucket_if_in_bounds(
val2,
&bounds,
bucket_pos2,
docs[2],
&bucket_with_accessor.sub_aggregation,
)?;
self.increment_bucket_if_in_bounds(
val3,
&bounds,
bucket_pos3,
docs[3],
&bucket_with_accessor.sub_aggregation,
)?;
}
for &doc in iter.remainder() {
let val = f64_from_fastfield_u64(accessor.get_val(doc as u64), &self.field_type);
if !bounds.contains(val) {
continue;
}
let bucket_pos = (get_bucket_num_f64(val, self.interval, self.offset) as i64
- self.first_bucket_num) as usize;
debug_assert_eq!(
self.buckets[bucket_pos].key,
get_bucket_val(val, self.interval, self.offset) as f64
);
self.increment_bucket(bucket_pos, doc, &bucket_with_accessor.sub_aggregation)?;
}
if force_flush {
if let Some(sub_aggregations) = self.sub_aggregations.as_mut() {
@@ -385,7 +407,7 @@ impl SegmentHistogramCollector {
if bounds.contains(val) {
debug_assert_eq!(
self.buckets[bucket_pos].key,
get_bucket_val(val, self.interval, self.offset)
get_bucket_val(val, self.interval, self.offset) as f64
);
self.increment_bucket(bucket_pos, doc, bucket_with_accessor)?;
@@ -429,7 +451,6 @@ fn intermediate_buckets_to_final_buckets_fill_gaps(
buckets: Vec<IntermediateHistogramBucketEntry>,
histogram_req: &HistogramAggregation,
sub_aggregation: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<Vec<BucketEntry>> {
// Generate the full list of buckets without gaps.
//
@@ -470,9 +491,7 @@ fn intermediate_buckets_to_final_buckets_fill_gaps(
sub_aggregation: empty_sub_aggregation.clone(),
},
})
.map(|intermediate_bucket| {
intermediate_bucket.into_final_bucket_entry(sub_aggregation, schema)
})
.map(|intermediate_bucket| intermediate_bucket.into_final_bucket_entry(sub_aggregation))
.collect::<crate::Result<Vec<_>>>()
}
@@ -481,41 +500,20 @@ pub(crate) fn intermediate_histogram_buckets_to_final_buckets(
buckets: Vec<IntermediateHistogramBucketEntry>,
histogram_req: &HistogramAggregation,
sub_aggregation: &AggregationsInternal,
schema: &Schema,
) -> crate::Result<Vec<BucketEntry>> {
let mut buckets = if histogram_req.min_doc_count() == 0 {
if histogram_req.min_doc_count() == 0 {
// With min_doc_count != 0, we may need to add buckets, so that there are no
// gaps, since intermediate result does not contain empty buckets (filtered to
// reduce serialization size).
intermediate_buckets_to_final_buckets_fill_gaps(
buckets,
histogram_req,
sub_aggregation,
schema,
)?
intermediate_buckets_to_final_buckets_fill_gaps(buckets, histogram_req, sub_aggregation)
} else {
buckets
.into_iter()
.filter(|histogram_bucket| histogram_bucket.doc_count >= histogram_req.min_doc_count())
.map(|histogram_bucket| {
histogram_bucket.into_final_bucket_entry(sub_aggregation, schema)
})
.collect::<crate::Result<Vec<_>>>()?
};
// If we have a date type on the histogram buckets, we add the `key_as_string` field as rfc339
let field = schema.get_field(&histogram_req.field)?;
if schema.get_field_entry(field).field_type().is_date() {
for bucket in buckets.iter_mut() {
if let crate::aggregation::Key::F64(val) = bucket.key {
let key_as_string = format_date(val as i64)?;
bucket.key_as_string = Some(key_as_string);
}
}
.map(|histogram_bucket| histogram_bucket.into_final_bucket_entry(sub_aggregation))
.collect::<crate::Result<Vec<_>>>()
}
Ok(buckets)
}
/// Applies req extended_bounds/hard_bounds on the min_max value
@@ -1321,6 +1319,7 @@ mod tests {
"min": Value::Null,
"max": Value::Null,
"avg": Value::Null,
"standard_deviation": Value::Null,
}
})
);
@@ -1373,63 +1372,6 @@ mod tests {
Ok(())
}
#[test]
fn histogram_date_test_single_segment() -> crate::Result<()> {
histogram_date_test_with_opt(true)
}
#[test]
fn histogram_date_test_multi_segment() -> crate::Result<()> {
histogram_date_test_with_opt(false)
}
fn histogram_date_test_with_opt(merge_segments: bool) -> crate::Result<()> {
let index = get_test_index_2_segments(merge_segments)?;
let agg_req: Aggregations = vec![(
"histogram".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Histogram(HistogramAggregation {
field: "date".to_string(),
interval: 86400000000.0, // one day in microseconds
..Default::default()
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let agg_res = exec_request(agg_req, &index)?;
let res: Value = serde_json::from_str(&serde_json::to_string(&agg_res)?)?;
assert_eq!(res["histogram"]["buckets"][0]["key"], 1546300800000000.0);
assert_eq!(
res["histogram"]["buckets"][0]["key_as_string"],
"2019-01-01T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][0]["doc_count"], 1);
assert_eq!(res["histogram"]["buckets"][1]["key"], 1546387200000000.0);
assert_eq!(
res["histogram"]["buckets"][1]["key_as_string"],
"2019-01-02T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][1]["doc_count"], 5);
assert_eq!(res["histogram"]["buckets"][2]["key"], 1546473600000000.0);
assert_eq!(
res["histogram"]["buckets"][2]["key_as_string"],
"2019-01-03T00:00:00Z"
);
assert_eq!(res["histogram"]["buckets"][3], Value::Null);
Ok(())
}
#[test]
fn histogram_invalid_request() -> crate::Result<()> {
let index = get_test_index_2_segments(true)?;
@@ -1496,36 +1438,4 @@ mod tests {
Ok(())
}
#[test]
fn histogram_test_max_buckets_segments() -> crate::Result<()> {
let values = vec![0.0, 70000.0];
let index = get_test_index_from_values(true, &values)?;
let agg_req: Aggregations = vec![(
"my_interval".to_string(),
Aggregation::Bucket(BucketAggregation {
bucket_agg: BucketAggregationType::Histogram(HistogramAggregation {
field: "score_f64".to_string(),
interval: 1.0,
..Default::default()
}),
sub_aggregation: Default::default(),
}),
)]
.into_iter()
.collect();
let res = exec_request(agg_req, &index);
assert_eq!(
res.unwrap_err().to_string(),
"An invalid argument was passed: 'Aborting aggregation because too many buckets were \
created'"
.to_string()
);
Ok(())
}
}

2
src/aggregation/bucket/histogram/mod.rs
View File

@@ -1,4 +1,2 @@
mod date_histogram;
mod histogram;
pub use date_histogram::*;
pub use histogram::*;

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