VORQ DocsStop and restart safely
Drain the daemon with SIGTERM, keep the state file, and know what a hard kill costs.
Stop with SIGTERM
On SIGTERM or SIGINT the daemon drains:
- It stops claiming new jobs, and
/healthzturns503. - Jobs running at a resumable backend (a poll mapping with
handle, or theopenai-responsesandopenai-batchpresets) are suspended: the work keeps running at the backend and the next start resumes polling it. - It withdraws every ask it published. The ask book has no expiry, so asks left standing would keep matching you with work.
- It waits for every other in-flight job to finish and settle, then exits.
Give your supervisor a stop timeout at least as long as your longest synchronous job. A
supervisor that sends SIGKILL too early turns the drain into a hard kill.
Keep the state file
Suspended jobs are identified by the handles in provider.state_db (default
vorqd-state.sqlite in the working directory). The file is small and only holds jobs in
flight, but it must survive a restart to be useful. In a container, keep it on a volume:
provider:
state_db: /var/lib/vorqd/state.sqlitewith /var/lib/vorqd a volume writable by the daemon's user.
If you scale to zero, suspended jobs wait until a daemon with the same state file starts again. Keep one instance running, or accept that a long wait can run past the SLA.
What happens on the next start
On its first poll sweep, the daemon lists the jobs claimed under its provider id and:
- resumes polling every job whose handle is on record (
resumed); - re-runs every other job still inside its SLA window for a model it serves (
recovered); - fails back the rest, so the client is refunded immediately;
- forgets handles for jobs that were settled, failed or reclaimed in the meantime.
Hard kill
SIGKILL gives the daemon no chance to drain. The next start recovers what it can as above,
but downtime is not given back: a job whose window closed while the daemon was down is failed,
and that counts as a missed SLA against your reputation. Asks stay on the book until the new
process publishes its own. Use SIGKILL only on a daemon that is already stuck.