Run behind Knative

Behind Knative, or any proxy that drains connections before it stops a container, an open Responses WebSocket is a connection in flight. Knative’s queue-proxy waits for it to drain before the server even gets its SIGTERM, up to the revision’s timeoutSeconds, and counts it as concurrency, so the service does not scale to zero while one is open. The server closes a socket left idle for --server-socket-idle-close seconds (default 300, WebSocket mode), and the Playground closes its own a minute after its tab is hidden, so a tab left open holds the pod, and its GPU, for minutes rather than for the revision’s timeout. Lower it (-e IGNIS_SERVER_SOCKET_IDLE_CLOSE=120) to release the pod sooner; a client’s next request opens a new socket either way.

The Monitor is the other thing that keeps a scale-to-zero service awake: a visible Monitor tab polls /ui/metrics every few seconds and counts as traffic, so it holds the pod whatever the socket settings say; a hidden one sends nothing (GitHub #329 item 11).