| name | release-observability |
| description | Observe software delivery from commit through build, artifact, promotion, deployment, health, rollback, and post-release behavior so release state and failures are traceable to exact revisions. |
Release Observability
Use when teams cannot quickly answer what version is running, whether a rollout is healthy, where a delivery failed, or which release caused a regression.
Procedure
- Define the release lifecycle and identifiers that connect it: source commit, build/run ID, artifact digest/version, environment, deployment/release ID, target revision, and rollout cohort or node set where relevant.
- Emit structured events for build start/result, artifact creation, promotion, deployment start/progress/result, health gates, rollback/recovery, and final release completion.
- Preserve correlation between pipeline events and runtime telemetry so service errors or latency can be compared against exact deployment windows and revisions.
- Track deployment duration, queue/wait time, failure stage, retry count, rollback rate, lead time, and other delivery signals that reveal bottlenecks or instability.
- Make current release state queryable without reading raw pipeline logs: what is intended, what is actually running, what is progressing, and what failed.
- Separate pipeline success from runtime success. A completed deploy must still pass the health/user-behavior gates defined by the release process.
- Alert on stuck or repeatedly failing release operations when action is required, and include enough identifiers to open the right pipeline/runtime evidence immediately.
- Retain release evidence long enough to investigate delayed regressions and audit significant changes without keeping unrestricted secrets or sensitive logs.
- Validate telemetry during a failed deployment and rollback so the most important recovery path is not the least observable one.
Decision rules
- Release observability should answer “what changed?” before an incident responder searches commit history manually.
- Artifact/revision identity is more useful than mutable labels such as “latest.”
- When Fleet distributes workloads, record the rollout/revision on the selected nodes but keep Fleet's live placement registry authoritative for current location.
- Delivery metrics are diagnostic signals, not performance targets to game by removing safety gates.
Quality gate
Release observability is ready when an operator can identify the exact source/artifact deployed, see rollout state and failure stage, correlate runtime change with release events, determine whether recovery occurred, and investigate the release without reconstructing state from scattered logs.