Parent scope: repository.
These are standalone scripts, not pytest targets. They can consume substantial CPU/memory/time; the import benchmark deserializes reviewed executable fixtures.
| Script | Inputs and side effects |
|---|---|
baseline_import.py |
Warmup/formal imports, optional profiling/correctness queries, temporary databases, timing/RSS/database/frame summaries; import uses set_focus=False |
sqlite_write_throughput.py |
Synthetic rows against default/optimized writers in temporary databases |
Run from the repository root with PYTHONPATH=src python benchmarks/<script>.py.
Before the import benchmark, run pytest tests/test_fixture_provenance.py and
follow fixture acceptance. The large 131k/628k
inputs use Git LFS; hydrate selected objects and rerun verification. The checksum
gate accepts a valid pointer, but benchmark loading cannot use that pointer.
The import benchmark defaults to the large samples. For a bounded check, use
--samples 8k --runs-8k 1 --skip-profile; it still loads pickle, performs a
warmup and runs correctness queries unless separately disabled. Keep outputs in
temporary/ignored locations. Dependency setup and sample hydration are separate
from running these scripts.
Use the same interpreter, platform, fixture hashes, arguments, warmup and run counts when comparing changes. RSS normalization is platform-dependent; do not compare raw resource output as though all systems use the same units.
Run pytest tests/test_baseline_import.py for parsing/RSS helper changes.
Import/runtime changes also need their core/snapshot correctness suites, and
SQLite writer/index changes need tests/test_snapshot_db.py. Faster benchmark
results do not replace replay, schema or publication correctness evidence.