Oliver/hm_ai_qc_report_tool
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Gunicorn: raise max_requests and graceful_timeout

Browse source 
The previous max_requests=200 caused workers to recycle every ~5 minutes
under normal progress polling (~40 req/min), killing any in-flight
background matching/QC thread on the worker. Bumping to 5000 means a
worker only recycles after several hours, well past any single job. Also
raise graceful_timeout to 600s so in-flight threads finish on legitimate
shutdowns instead of being SIGKILL'd after 30s.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
nickviljoen 2026-04-28 13:17:43 +02:00
parent 3dd0420145
commit b140ab3860
1 changed files with 11 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
gunicorn_config.py
View file

@ -13,10 +13,18 @@ worker_class = "gthread"
worker_connections = 1000
timeout = 300
keepalive = 5
# graceful_timeout: how long gunicorn waits for in-flight work to finish on a shutdown
# (deploy, recycle). Default is 30s, which is too short for long-running matching/QC jobs
# that run as background threads inside the worker. 600s gives those jobs a chance to
# complete on graceful shutdowns; SIGKILL only fires after that.
graceful_timeout = 600
# Recycle workers after N requests to release accumulated memory
max_requests = 200
max_requests_jitter = 30
# Recycle workers after N requests to release accumulated memory.
# Set high (was 200) because each in-progress polling job fires ~40 req/min — with the
# old setting the worker recycled mid-job and killed any running matching/QC thread.
# Monitor memory in `docker stats hm-qc-app`; if it climbs over time, lower this.
max_requests = 5000
max_requests_jitter = 500
# Logging — stdout/stderr for docker logs
accesslog = "-"