Smart search, face detection and OCR were all failing with
Machine learning request to "http://immich-machine-learning:3003" failed
while the container still reported Up. Podman only sees PID 1: gunicorn's
master was alive and holding the listening socket, but its worker had died at a
WORKER TIMEOUT and was never respawned -- hence a connect timeout rather than a
refusal. The dead worker was a zombie whose remaining thread was stuck in
uninterruptible sleep in exit_mmap, so it survived SIGKILL, podman rm -f and
rm -f -t 0, and kept the container name and network alias until the host was
rebooted.
MACHINE_LEARNING_MODEL_TTL=0 addresses the cause rather than the symptom. The
default unloads models after 300s idle, so every search following a gap
reloaded four of them (CLIP, buffalo_l detection + recognition, PP-OCRv5) on a
CPU-only 4-core box and then tore those mappings back down -- and that teardown
is what wedged. Keeping them resident costs ~1-2 GB and removes the path.
IMMICH_MACHINE_LEARNING_URL is now set explicitly instead of relying on
immich's implicit default, so the ML container can be renamed without silently
losing search, and the name is a variable so a replacement can be stood up
beside a broken one without editing tasks.
Verified after deploy: zero ML failures, "in-memory cache with unloading
disabled", ping 200 from immich-server, 26/26 containers healthy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
