Gff Gtf Pushdown Regression

GFF/GTF pushdown regression testing

Why this exists

polars_bio pushes Polars filter()/select() predicates down into the DataFusion GFF/GTF scan as an optimization. The correctness contract (issue #396 / PR #407) is:

The client-side full filter/select is the source of truth. Pushdown may only remove rows the client filter would also remove. A translation bug can cost performance, never correctness — and must never crash collect().

The classic failure (#396) was a predicate silently dropped during translation, so a filtered scan returned too many rows with no error. Unit tests alone miss this because the bug can be wrong identically on a naive reference. So this harness checks three implementations against each other on a real 11M-row GENCODE annotation:

• A — scan_gtf/gff with pushdown ON
• B — scan_gtf/gff with pushdown OFF (client-side filter/select only)
• C — oxbow (a separate Rust/Arrow GFF/GTF reader): read everything, filter in Polars — an independent ground truth

A case passes only when A == C and B == C as row sets (row order is also checked). Using oxbow as a third party is the point: it catches a bug that is wrong in the same way on both polars_bio paths.

When to run it

Run the harness after any change to:

• polars_bio/predicate_translator.py (the AST emitter / plan_predicate_pushdown)
• polars_bio/pushdown.py (apply_predicate_pushdown / apply_projection_pushdown)
• the scan call sites: io.py _overlap_source, the GFF/GTF select() wrapper (AnnotationLazyFrameWrapper), utils.py, pileup_op.py
• scan_gtf / scan_gff / register_gtf / register_gff or their options

Treat a green run (ALL PASS) as the bar for claiming a pushdown change is correct. The fast Python unit suite (tests/test_predicate_translator_units.py, tests/test_pushdown_helpers.py) is the inner loop; this harness is the real-data gate.

How to run

cd <repo root>
source .venv/bin/activate          # or however the project venv is activated
unset CONDA_PREFIX                  # maturin/uv dislike both VIRTUAL_ENV + CONDA_PREFIX set
export TQDM_DISABLE=1               # silence polars_bio's per-row progress bar
python .claude/skills/gff-gtf-pushdown-regression/scripts/bench_pushdown.py gtf gff3

• First run downloads GENCODE v50 GTF (124 MB) and GFF3 (160 MB) into ~/.cache/polars-bio-gencode (override with POLARS_BIO_GENCODE_DIR); later runs reuse them. Needs network on the first run and oxbow installed (pip install oxbow).
• Args select formats: gtf, gff3, or both. Default is gtf. Use gtf gff3 for a full check — they have slightly different feature representations and each has caught format-specific bugs.
• If you changed Rust, rebuild first: maturin develop --release. Pure-Python pushdown changes need no rebuild.
• Exit code is non-zero if any case fails, so it works as a CI/pre-merge gate, not just an interactive report.

Reading the output

[PASS] gtf chrom == chr7    rows=  570,522 A=C:True B=C:True order:True | ON 6.5s OFF 26.0s (4.0x)
--- gtf: 18/18 cases PASS ---
==== ALL PASS ====

• A=C / B=C — pushdown-on and pushdown-off both match oxbow. Both must be True.
• order:True — results also match in row order, not just as a set.
• ON / OFF (Nx) — wall time with pushdown on vs off, and the speedup. A selective predicate should be visibly faster on (e.g. chrom==chr1 & type==gene ≈ 4x); a predicate matching ~all rows trends toward 1x. If a selective predicate shows ~1x, pushdown probably is not firing — a performance regression worth investigating even when correctness passes.

A FAIL means

The changed pushdown either dropped/added rows or mangled a projection. The A.h/B.h/C.h line shows which path diverged. Almost always the fix is: make the translator decline to push that predicate (raise UnsupportedPredicate so it falls back to the client-side filter) rather than push a subtly-wrong SQL. Add a focused unit test in tests/test_predicate_translator_units.py reproducing the case, then re-run this harness.

Extending the case list

cases() in the script is the coverage. Each entry is (name, predicate, projection, attr_fields). When you add a translator branch or fix a bug, add a case here. Keep the existing edge cases — they encode hard- won failure modes and each maps to a real regression:

• str.contains, string-column ordering (chrom < "chr2") — must stay client-side (the exact #396 failure); pushing them is wrong.
• is_in([..., None]) and is_in(nulls_equal=True) — SQL IN null semantics differ from Polars; must stay client-side.
• score > nan / non-finite floats — must not crash; fall back.
• deeply nested &/| chains — must degrade gracefully, never RecursionError.
• gene_name.contains(...) -> gene_id — the #396 attribute filter+select path.
• aliased / computed projections (chrom.alias("c"), (start+1).alias("s1"), aliased attribute gene_id.alias("g")) — projection pushdown must apply the alias/computation, not silently return the raw source column.