petbox-dev / tafra / 23926518445

95%

push

github

fix: composite key collision in multi-column joins + _accel.c v3 (#29)

* perf: _accel.c v2 — power-of-2 hashing, CSR joins, single-pass output

- Replace modulo hash probing with bitmask (& mask) on power-of-2 tables
- Replace linked-list join hash table with CSR (flat contiguous row buffer)
- Eliminate double-pass join probing with growable IndexBuf output
- Unroll composite_key inner loop for 1-3 column fast paths
- Use memcpy for first_seen output, PyArray_EMPTY where safe

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* perf: _accel.c v3 — combine bitmask hashing with slot-caching joins

Merges best of v2 (power-of-2 bitmask probing, unrolled composite_key,
CSR flat row buffer) with Derrick's slot-caching two-pass join approach
(no re-probing, exact pre-allocation).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* perf: _accel.c v3 — combine bitmask hashing with slot-caching joins

Merges best of v2 (power-of-2 bitmask probing, unrolled composite_key,
CSR flat row buffer) with Derrick's slot-caching two-pass join approach
(no re-probing, exact pre-allocation).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: composite key collision in multi-column joins with asymmetric cardinalities

_build_composite_key computed per-column cardinalities independently on each
call. When left and right sides had different unique value counts, the
positional encoding multipliers diverged, mapping distinct key tuples to the
same composite integer.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* test: add regression tests for asymmetric-cardinality composite key joins

Also guard cards computation against empty encoded arrays (e.g. when all
rows on one side are null-filtered).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version to 2.2.2

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: preserve obje... (continued)

19 of 19 new or added lines in 3 files covered. (100.0%)

1616 of 1700 relevant lines covered (95.06%)

0.95 hits per line