Monitoring and logs¶

Health¶

GET /healthz → {"ok":true,"model":"..."} (no auth). Used by the image HEALTHCHECK (30 s interval, 300 s start period for the first download).
docker inspect -f '{{.State.Health.Status}}' jeff → starting, healthy, unhealthy.

Uptime Kuma / Gatus: HTTP check on https://jeff.example.com/healthz, expect 200 and keyword "ok":true.

Stats¶

GET /stats returns batcher counters and the active configuration:

{"batches": 42, "requests": 97, "avg_batch": 2.3, "max_batch": 8,
 "avg_infer_ms": 180.4, "avg_queue_ms": 6.1,
 "backend": {"backend": "torch", "device": "cpu", "dtype": "float32", "load_seconds": 6.0}}

Warning

/stats is not authenticated upstream. It exposes configuration (no secrets). On a public domain you may block it at the proxy, see Security.

Logs¶

docker compose logs -f jeff            # Compose
docker logs -f jeff                    # plain Docker

Startup lines to look for:

[entrypoint] Model already present: knowledgator/gliformer-large-v1@main at /models/gliformer-large-v1
[entrypoint] jeff-docker 0.2.0 | upstream Jeff 1a2b3c4d5e6f
[entrypoint] device=cpu backend=torch threads=6 max_batch=8 auth=on
... jeff ready: backend=torch model=gliformer-large-v1

Log rotation is preconfigured in the Compose files (json-file, 3 × 10 MB).

Resources¶

docker stats jeff

Alert if memory stays close to JEFF_MEM_LIMIT (risk of OOM kill, exit code 137).