A rollback plan is easiest to write while the old service is still healthy. Capture the version and configuration that are actually running. A file with “backup” in its name is useful only if you know which service can read it and how to load it.

Write three lines before changing anything

  • What will change: list the file, image, or DNS record and the new value.
  • How to verify it: choose a real request with a success condition, such as a page returning the expected title or a known API record loading through the public address.
  • When to revert: pick an observed failure and a time window, such as a repeated error during startup or a critical read that still fails after readiness goes green.

Keep the old version reachable

Retain the old artifact and config until the new version has worked through a normal traffic cycle. For file-based deploys, stage new files separately and check their ownership and readability. Switching a release directory is easier to reverse than copying individual files over the running release.

A rollback may have side effects of its own. Before shipping a database migration, test how both versions behave against the new schema. Before a DNS change, account for old and new answers coexisting while caches expire. Include those limits in the release note so the person on call can act on them.