jedarden/pdftract
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
pdftract/tests/fixtures/perf/README.md
jedarden c621947686 feat(bf-1g1fd): implement CI memory-ceiling gate with cgroup MemoryMax enforcement 
Implements Tier-1 memory ceiling gate that enforces RSS budgets for PDF
extraction, analogous to cargo-bloat for binary size.

Changes:
- CI: Add memory-ceiling template with cgroup MemoryMax (1.5 GB)
- CI: Add cgroup MemoryMax enforcement to test-glibc (6 GB) and test-musl (4 GB)
- CI: Add cgroup MemoryMax + libfuzzer rss/malloc limits to fuzz workflow
- xtask: Implement memory-ceiling command with peak RSS sampling
- Add perf fixtures (100-page, 10k-page) for memory testing
- Add run-fuzz-with-limits.sh for local fuzz testing with memory caps
- Register perf fixtures in PROVENANCE.md

Memory budgets enforced:
- Buffered 100-page PDF: < 512 MB
- Streaming mode: < 256 MB (constant in page count)
- Adversarial fixtures: < 1 GB hard ceiling

Closes bf-1g1fd

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-23 13:22:55 -04:00

1.2 KiB
Raw Blame History

Performance Test Fixtures

This directory contains PDF files used for performance and memory ceiling testing.

Memory Budgets

Per the plan (Phase 0 Quality Targets), the following memory budgets apply:

Category Target Measurement
Peak RSS, 100-page vector PDF (buffered mode) < 512 MB RSS sampled at 10 ms intervals
Peak RSS, streaming/NDJSON mode (any page count) < 256 MB, constant in page count Must stay flat as page count grows
Peak RSS, adversarial fixtures < 1 GB hard ceiling Must not scale with payload size

Fixtures

100-page vector PDF

  • 100-page-vector.pdf - Synthetic 100-page PDF for buffered mode testing
  • Target: < 512 MB peak RSS

10k-page stress test

  • 10k-page.pdf - Synthetic 10,000-page PDF for streaming mode validation
  • Target: < 256 MB peak RSS in streaming mode
  • Must remain constant regardless of page count

Generating Fixtures

Fixtures can be generated using the generate_stress_pdf.py script in the tools directory.

python tools/generate_stress_pdf.py --pages 100 -o tests/fixtures/perf/100-page-vector.pdf
python tools/generate_stress_pdf.py --pages 10000 -o tests/fixtures/perf/10k-page.pdf