> Source: https://vitna.ai/docs/operations/verifying-a-deploy > Raw Markdown: https://vitna.ai/docs/operations/verifying-a-deploy.md > Corpus index: https://vitna.ai/llms.txt # Verifying a deploy A push that exits zero is not a deploy. Three checks say which commit is live. ## The three | Check | What it proves | |---|---| | The deploy log ends with a done marker | The hook ran to completion | | The service restart timestamp is newer than the push | The process actually moved | | The served page carries a change-unique marker from the pushed commit | The bytes in front of users are yours | All three, in that order. Any one of them alone can read as success while the deploy did not happen. ## Why a change-unique marker Not the title, not a version string that a cached build would also carry. Some string that exists only in the commit you just pushed. The failure this catches is the one that looks most like success: the hook aborted during the build, the previous build is still serving, and it is serving healthily. Health checks pass. The site is up. The code is old. ## What the automatic check covers The deploy job reads the health endpoint and fails unless three things hold at once: it answers `200`, its database is reachable, and the revision it reports is the commit that job pushed. That third one is the check that used to be missing, and it is the one that catches a failed build. The hook records the revision it restarted, so the endpoint reports what the running process was built from; a build that aborted leaves the previous revision reported, and the job goes red naming both. A box whose hook predates that reports no revision, and the job refuses rather than passing on a claim it cannot check. ## Re-pushing retries, but read the log first A failed deploy rewinds `refs/heads/main` on the box, because the hook only fires when a ref moves and a ref left advanced turns the obvious retry into a silent no-op. So pushing again does now retry. That is a repair, not an invitation. A push that appears to have done nothing has a reason, the reason is in the deploy log, and retrying before reading it turns one question into two.