gunicorn workers = multiprocessing.cpu_count()*2+1 starts 33 workers in a container limited with --cpus=2: cpu_count and os.process_cpu_count() report host CPUs, not the cgroup quota
Fix: set the worker count explicitly in containers: workers = int(os.getenv("WEB_CONCURRENCY", "2")) in gunicorn_conf.py, -w N, or GUNICORN_CMD_ARGS="--workers 3". If you want auto-sizing, read the cgroup v2 quota yourself: quota, period = open('/sys/fs/cgroup/cpu.max').read().split(), then use ceil(int(quota)/int(period)) when quota != 'max'.
What we observed: the common snippet workers = multiprocessing.cpu_count() * 2 + 1 with worker_class = "uvicorn.workers.UvicornWorker" booted a wall of Booting worker with pid: ... lines for a small FastAPI service on a 16-core host, so 33 processes. Each worker opened its own Redis/httpx clients, kept its own in-memory data copy, and ran its own OpenTelemetry exporter. Shutdown printed [ERROR] Worker (pid:N) was sent SIGTERM! once per worker.
Checked with podman: podman run --rm --cpus=2 python:3.11-slim python -c "import multiprocessing,os; print(multiprocessing.cpu_count(), len(os.sched_getaffinity(0)))" prints 16 16 while /sys/fs/cgroup/cpu.max reads 200000 100000 (2 CPUs). Python 3.13's os.process_cpu_count() also printed 16 under --cpus=2, because it follows CPU affinity, not CFS quota. So you get 33 workers on 2 CPUs of quota: throttling and extra memory, with no throughput gain.
Knock-on effects that look like app bugs: per-process rate limits and caches multiply by the worker count (slowapi without a shared storage_uri, module-level dicts).