What can go wrong, and what we do about it
Production is risky, and a tool that touches it should say plainly where it can fail. Merge Scheduler reduces the risk of late-night releases. It does not remove it, and it does not replace your own safeguards.
The short list
| If this happens | What Merge Scheduler does | What you should do |
|---|---|---|
| Our service is down at merge time | When it comes back, a release up to 15 minutes late still runs. One later than that is skipped and you are emailed, so nothing merges at a time nobody planned. | Reschedule it. A server monitor checks the service every 5 minutes and alerts us. |
| A teammate merges on GitHub without the app | Every waiting release on that branch is checked again and its owner is emailed if it would now be skipped. At merge time it checks again. | Fix or reschedule the pull request. A deploy someone else started is not waited for. |
| Your deploy is not a GitHub Actions workflow | Choose health-only mode for the repository. After the merge it waits for your deploy to go live, then watches your health URL and reverts if it fails twice in a row. It cannot see the deploy itself, only whether the site is healthy. | Use a health URL that fails when the site is really broken. Health checks are on Pro (one free try-out). |
| No deploy workflow is chosen for a repository | Every workflow the merge starts is watched, so an unrelated failing one (lint, nightly) counts as a failed deploy and can revert a good release. | Tick your deploy workflow in the repository settings. The app warns you when none is ticked. |
| The change includes a database migration | A revert undoes the code, not the data. A migration that already ran is not undone. | List hand-run or irreversible paths under "Paths that deploy by hand" to get a warning. Write migrations that can run twice. |
| The deploy fails once for a flaky reason | A failed run counts as a bad release and the change is reverted. | Make the deploy workflow reliable, or turn auto-revert off for that release and keep the email. |
| The revert itself cannot merge or its deploy fails | It stops, does not retry blindly, and emails "Needs attention" with the reason. | Fix it by hand. Branch protection that requires a review on the revert pull request will stop it. |
| Someone merged on top of your change before a revert | It does not revert, because that could undo their work. It emails you. | Revert by hand with a clear head. |
| You connect with a GitHub token instead of the app | Merges and reverts appear under your own name. We store the token encrypted and email you 14 days, 3 days and on the day before it expires. In this mode we get no GitHub events, so there is no heads-up when someone else merges outside the app. | Use a fine-grained token limited to the repos you need, replace it before it expires, and revoke it on GitHub if you think it leaked. Expired tokens make merges fail. |
| You want your own rollback to run when we revert (a down-migration, a redeploy) | On Pro you can add a webhook for the revert started event. We send one signed call and retry if it fails. We do not wait for your endpoint, and we cannot see whether it worked or undo any data. | Build and test your endpoint, make it safe to call twice, and check its result yourself. |
| A release train finds no pull request, or two | It schedules nothing and emails the owner why. If the service was down too close to the time, the run is recorded as missed. | Check Settings → the repository → Release train, and merge or schedule by hand. |
| You merge a step of a queue yourself on GitHub | The queue treats it as done and carries on with the next step. Nothing is reverted because of it. The email says it was already merged by hand, and its deploy is not watched. | Watch that deploy yourself on GitHub. |
| A step in a queue fails | By default the queue stops: the failed step is reverted if auto-revert is on, and later steps are cancelled. If you chose to carry on, the next step starts after the revert. A revert that does not work, or a cancel, always stops it. | Check the summary email. Only carry on for pull requests that do not depend on each other. |
| You miss the email | If nobody opens a release that ended badly, one reminder is sent two hours later. The Releases page always shows the outcome. On Pro you can also get a phone alarm and Slack alerts. | Open the release in the app, or set up a phone alarm for failures. |
| An incident starts and you want everything to stop | Pause all merges: anything that comes due is skipped, not merged. Reverts already under way still run. | Anyone who can schedule can pause. An owner resumes. |
| GitHub has an outage | If GitHub cannot be reached it tries again. After repeated errors it stops and tells you. | Check githubstatus.com and reschedule. |
What it never does
- It does not bypass branch protection. If GitHub says a pull request is blocked, it is skipped.
- It does not merge new commits nobody reviewed: the commit is pinned when you schedule.
- It does not store your source code. It reads pull request, workflow and check data. The only credential it can keep is a GitHub access token, and only if you choose token mode: it is stored encrypted and you can remove it at any time.
- It does not change your repository except to merge the pull request you scheduled and, if a deploy fails and you chose auto-revert, to open and merge a revert pull request.
Secrets in failure emails
A failure email shows a few log lines from the failed step. GitHub already masks the secrets it knows about. We also hide common token and password shapes before anything is stored or sent. Neither is perfect, so do not print secrets in your build logs.
About the service itself
- Merge Scheduler is built and run by one person. Support replies within 2 business days.
- It runs on a single server today. We monitor it every 5 minutes, restart it if it stops, and email the owner if something is wrong. We do our best to keep it up, but we do not promise a specific uptime. If it is down at your merge time, the merge runs late or is skipped, as described above.
- The database is backed up every day: kept for 7 days on the server, and an encrypted copy is kept off the server for 30 days of daily and 12 weeks of weekly backups.
- You can leave any time: uninstall the GitHub App and cancel your plan. Nothing is left in your repositories except revert pull requests that were opened.
Responsibility
The service is provided as is, as the terms say. You decide what to release and when, and you keep your usual safeguards: reviews, tests, backups and a way to roll back by hand.