[avilaton]

Summary

This adds optional RSS-based worker recycling.

Workers can now be configured with max_rss or --max-rss to recycle after their current resident set size exceeds a configured MiB limit. The RSS monitor runs periodically inside each worker and triggers the existing graceful worker termination path, so in-flight requests still drain under graceful_timeout before the master process respawns the worker.

Current RSS is read with psutil.Process().memory_info().rss, which adds psutil as a dependency. With workers == 0, there is no master process to respawn the worker, so exceeding max_rss stops the single worker, matching existing max_requests behavior.

Tests

• uv run pytest tests/test_config.py tests/test___main__.py tests/test_run.py -q — passed
• uv run pytest — passed
• uv run ruff check — passed
• uv run ruff format --check — passed
• uv run ty check — passed

[davidbrochart]

Current RSS is read from Linux /proc/self/status VmRSS.

Maybe use psutil's memory_info().rss?

[avilaton]

Good idea, switched to psutil.Process().memory_info().rss, which also makes it work beyond Linux.