New images overnight, and back again if they break.
Cleaning up by itself has been there for a while, and so have automatic backups. Updating belongs in that list, but with more care: a cleanup removes what nobody uses and a backup changes nothing, while a new image can genuinely break. So the app pulls, recreates, waits to see it come back — and puts the old image back if it does not.
Worried an update breaks something?Read the comic: new editions overnight, and the old one kept →What you can do
Per project, not everything at once
Turn it on for a compose project or a single container, in a dialog where you pick the time first. There is no "update everything" switch: the database on your home server and an image you built yourself do not want the same treatment.
A container of its own
A container that is not part of a project has Update now and Update by itself… in its menu, in the list and on its own page. It is pulled, recreated with the same settings and put back if it does not come back, like a project. One that cannot be updated on its own has the items greyed out, with the reason: it belongs to a project (update the project instead), it is a swarm task, a project schedule already has its name, or it is pinned to a digest and there is no tag to follow. A container whose tag moved because you pulled by hand counts as behind, and Recreate on the new image works for it too.
A moment, not an interval
Four in the morning by default, like a backup schedule at its own fixed time: an update should not walk an hour later every day until it happens at noon.
A missed night, caught up once
A laptop that was off at four updates when it wakes up. Off for a week? It updates once, not seven times.
Back to the version that worked
The old image id is remembered before the pull and the image is kept. If the container does not come back — a health check that never turns healthy, an error right after the start, or nothing within five minutes — the old id goes back on the name and it is recreated.
One at a time, and it stops
After the first rollback the round ends. If tonight’s pull takes down half your containers, you want the app to stop rather than finish the list.
The database first
A round updates what others depend on first, from depends_on or the depends-on label of Watchtower. What waits on it stops just before it is recreated and starts again once it is back; if that does not come back, the previous version returns. Or, as compose itself does, only what waits with restart: true.
An edited compose file is left alone
Changed the file since those services were made? Recreating would roll out those changes too, and that is a decision you make awake.
One notification in the morning
"2 updated, 1 rolled back" — not one per container. Waking up to five notifications only teaches you to dismiss them.
Try it while you watch
Update now walks exactly the same path, rollback included, so you can see it work before you trust it overnight.
What changed, one click away
Every line in the morning report links to the release notes of the version that was pulled; after a rollback, that is exactly what you want to read.
A newer patch, if you like
A schedule follows the same tag, or also a newer patch, or patch and minor. Never a major: that is something to read about awake.
Watchtower's commands
lifecycle.pre-update runs in the old container before the update, post-update in the new one after it, and enable=false keeps a schedule away. Your compose files need no change.
A cool-down
Wait until a new version is a few days old, so a release that is pulled again after a day never reaches you.
Scan first
Scan the new version and the one the containers run, together, before anything is recreated; the update only goes in when the new one has no more critical vulnerabilities. A finding you marked as not applicable or accepted no longer holds it back.
Running is not the same as working
A health check goes through starting before it turns healthy.
That difference is what the app waits for. Without a health check, running counts only after five seconds: an app that falls over at once is still running just after it starts. A container that stopped with an error is broken; one that is briefly gone counts as waiting, not broken — compose may be recreating it. Unhealthy is broken, and waiting longer will not help.
The whole story afterwards
From which image id to which, and how long it took.
The round stays on the dashboard for a day, with a rolled-back line standing out in red, and every step also lands on the Events page. Not only that a container stopped and started, but that the app did it, and why.
To a newer tag
The compose file changes along, one line.
Recreating a container on nginx:1.27.5 is not enough: the next compose up would bring 1.27.3 back. So the app changes the image line in the compose file, with your comment intact, and shows that line before it does anything. Saving goes through the same checks as the editor; if the container does not come back, the line goes back too.
All features