Skip to main content
Once the daemon runs on your own box (see Deploy to a VPS), you can have it pull and apply new patch releases on its own. This page sets that up with an opt-in Compose overlay that adds a release-channel pin, Watchtower, and rolling Postgres backups.

​
Is it really “no interruption”?

For your agents, effectively yes. The SDK’s RemoteCollectorSink buffers telemetry in memory and retries with backoff, and treats HTTP 429 as backpressure. During the few-second container swap of an update, agents buffer and re-send, so no agent call fails. Two honest caveats:
  • Cost telemetry is best-effort by design (it is not billing-of-record). A long daemon outage can drop some rows; a fast update swap does not.
  • On a single node, a container recreate is not literally zero-downtime: the ingest endpoint and dashboard are unavailable for a few seconds. For true zero-downtime you need two daemon replicas behind a reverse proxy doing rolling restarts against shared Postgres (and migrations kept backward-compatible across the rollout). That is overkill for one box.

​
The release channel: pin to a minor, never :latest

The release workflow publishes three Docker tags per version, for example 0.24.0, 0.24, and latest. The overlay pins the daemon to the minor channel :0.24:
  • :0.24 auto-gets bug-fix patches (0.24.1, 0.24.2, …) but never a breaking major. You move to :0.25 deliberately.
  • :latest would auto-pull the next major and run its migrations unattended. Do not auto-track :latest in production.
A minor channel tag exists for every published minor. If you’re running an older minor than 0.24, pin the overlay to that minor’s channel instead (for example :0.23) until you deliberately move up.

​
Turn it on

The overlay (docker-compose.autoupdate.yml) layers on top of the daemon stack. From your deploy directory:
The overlay file in the repository still pins mahimairaja/voicegateway:0.10. Until that is bumped, the file you just downloaded puts you on a channel twelve minors behind what this page describes. Edit its image: line to :0.24 before you bring the stack up, and re-check it on your next pull.
That starts three things alongside the daemon: Always pass both -f files together for later up, down, and logs, so Compose keeps the merged definition.

​
Tuning

Set these in your .env next to the Compose files: Watchtower’s cadence is WATCHTOWER_POLL_INTERVAL (default 3600, hourly) in the overlay.

​
Restore from a backup

Dumps live in the voicegw-backups volume as voicegw-<timestamp>.sql. To roll back after a bad update:
1

Stop the daemon

Stop the daemon so nothing writes during the restore.
2

List available dumps

3

Restore the chosen dump

4

Pin to the known-good version and restart

Edit the image tag in docker-compose.autoupdate.yml (for example :0.24.1), then bring the stack up:
Migrations run on daemon startup and are forward-only. The Compose healthcheck gates a bad boot, but the database is shared state: keep the backups (and ideally a periodic host-level snapshot of the box) so a migration regression is always recoverable.

​
Opting back out

Bring the stack up with only the base file to drop Watchtower and the backup sidecar and return to a manually-pinned daemon: