Skip to main content
Preview
Post-Deploy Verifications are in preview for select pilot organizations. Contact the Gitar account team to request access.
Gitar checks production logs after a deployment using verification plans written during PR reviews. When the pipeline reports a deployment, Gitar identifies the PRs it shipped and checks their plans at fixed checkpoints.
Post-Deploy Verifications read Amazon CloudWatch Logs only.
To enable verification, follow Set up post-deploy verifications.

Verification plans

Gitar writes the plan during a full review of a PR in a repository with verification turned on. It appears as a collapsed Post-Deploy Verification section in the Gitar dashboard comment, and later reviews of the PR update it. Each check identifies the service and log signal to verify: When a change needs no production checks, such as a docs or test-only change, the plan says so in one sentence and the PR is not checked after deployment. Gitar tags each PR with the services its changes affect. Dead code and dependency removals get no tags. A PR is only verified by deployments of a service it is tagged with.

Deployment streams

A deployment stream is one repository deployed to one environment as one service, for example acme/worker to production as worker. Each stream keeps its own deployment history. Gitar identifies the PRs a deployment shipped by comparing it with the previous deployment in the same stream. The pipeline only needs to send the deployed commit. Only PRs that Gitar reviewed and tagged with the deployed service are verified. PRs reviewed before the repository turned verification on are never matched. A PR that touches two services is verified by a deployment of each. One deployment verifies at most 100 PRs. Report deployments from the default branch only, as a deployment from a feature branch would become the baseline for the next one.

Check schedule

Each PR in a deployment is checked at five checkpoints after the deployment time: Nothing runs at deployment time, as the first checkpoint needs 15 minutes of logs. Checkpoints before a plan’s earliest checkpoint are recorded as Skipped. A later deployment in the same stream does not cancel an earlier deployment’s remaining checkpoints, since its code is still live.

Check statuses

Each check in a plan gets one of these statuses at every checkpoint: A PR’s result at a checkpoint is its worst check, in the order Fail, Inconclusive, Not exercised, Pass. Skipped checks do not count. When every check is skipped, or the plan has no checks, the result is Skipped with the reason shown.

Dashboard

Open Observability in the dashboard to see deployments and their results. Use the environment switcher to change environments. It defaults to production.

Services

The Services tab shows one card per deployment stream, with the result of its latest deployment that shipped PRs and its recent history. Filter by repository.

Deploys

The Deploys tab lists deployments newest first, with the service, commit, checkpoint progress, check counts and verdict. Filter by service, repository or verdict. Deployments of the same commit to one environment within 10 minutes are grouped as a release.

Deployment details

Select a deployment to see each PR’s result at every checkpoint, the next checkpoint and how Gitar determined the PR range. PRs deployed to other services link to those deployments. Select a result to open its checks, each with what was expected, a one-line summary, the supporting findings and log details.

Rolled-up verdicts

PRs, deployments and services show a verdict that summarizes their results. A failure at any checkpoint keeps a PR at Failing even if later checkpoints pass, so a regression that recovered is still visible. Otherwise a PR follows its latest checkpoint. A deployment takes the verdict of its worst PR, in the order shown above.