[waleedlatif1]

Summary

• Presidio workers were being killed by the kernel OOM killer: uvicorn logs Child process died with no Maximum request limit line before it. The trigger is full-NER traffic carrying single texts between the client's 256KB chunk budget and spaCy's 1M max_length. Those went whole through one NER pass, which peaks about 3.3GB above the worker's baseline for about 4s.
• Workers also ran one NER pass per in-flight request (FastAPI's sync threadpool), and glibc's per-thread malloc arenas kept each thread's peak. The memory stacked until the task hit its limit. The spikes are too short to show in per-minute memory metrics.
• NER now runs in 100k-char windows for any text over NER_WINDOW_CHARS, instead of only texts over max_length. This reuses the windowing from fix(pii): mask texts over spaCy's length limit and drop encoded files from logs #9979: pattern and custom recognizers still run over the whole text, and every window's spans are kept.
• One NER pass at a time per worker (_NER_SLOT). A long text takes it per window, so it interleaves with other requests. Regex-only requests never take it.
• MALLOC_ARENA_MAX=2 in the image, and /health is async so health checks never queue behind waiting threads.

Type of Change

• Bug fix

Testing

Reproduced on a box with the exact pinned deps, the real entrypoint and 2 workers. Memory and CPU were capped in proportion to prod (MemoryMax=10G, CPUQuota=200%), with 16 concurrent clients sending 2000-text full-NER batches plus 30% single texts of 300–900KB, for 5 minutes:

	OOM kills (= Child process died)	ok / failed	peak memory
before	45	18 / 384	10.2GB (limit)
after	0	129 / 0	8.0GB

• Each part alone falls short. At the original load (12 clients, 15% large), windowing alone left 2 OOM kills and the lock alone 9, against 12 before. On the heavy load, windowing plus MALLOC_ARENA_MAX without the lock left 2.
• Output check on 150k–650k-char texts against the original: no span lost. A few extra spans at window edges, merged into the same mask or masking a cut fragment: the trade-off fix(pii): mask texts over spaCy's length limit and drop encoded files from logs #9979 accepted.
• Under that heavy load the slowest single large text took 335s. Before the fix none of them completed.
• The test-pii-redaction-e2e.ts long-text check (over 1M chars) still goes through the same windowed path.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
docs	Ready	Preview	Nov 15, 2026 9:36am UTC

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Turn on auto-fix | Re-trigger cubic}

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

This PR reduces NER memory use within each Presidio worker.

• Long texts now run NER in small overlapping windows.
• Each worker runs one NER pass at a time.

No actionable issues were found. waleedlatif1 acknowledges waiting threads and slower queued requests, including a reported 335-second large-text request. waleedlatif1 also accepts extra spans or cut fragments at window edges as the existing windowing trade-off.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Request[Analysis request] --> Regex{Regex only?}
  Regex -->|Yes| Patterns[Analyze without the NER lock]
  Regex -->|No| Long{Text over 100000 characters?}
  Long -->|Yes| Whole[Match patterns over the whole text]
  Whole --> Windows[Split into overlapping windows]
  Windows --> Slot[Acquire the worker NER lock]
  Long -->|No| Slot
  Slot --> NER[Run one analysis or short-text batch]
  NER --> Release[Release the lock]
  Release --> More{More windows?}
  More -->|Yes| Slot
  More -->|No| Results[Return results]
  Patterns --> Results

Loading

_{Reviews (1) · Last reviewed commit: "fix(pii): bound NER memory per worker so..." · Reviewed by Greptile}