ReviewOS

stacks/ts-cloud

Unify releases, promotion, zero-downtime deployment, and rollback

#151
Closed chrisbbreuer opened this 24 days ago · 0 comments
24 days ago

Roadmap relationship

Outcome

Give every deployable type one immutable release model with provenance, promotion between environments, health-gated activation, provider-aware zero-downtime strategies, and reliable rollback to a known artifact and configuration.

Current gap

Compute sites use atomic release directories, serverless uses S3 release snapshots, container work has separate task/image behavior, and advanced canary/blue-green abstractions are not integrated into the primary workflow. History, actor/trigger, config provenance, promotion, and rollback guarantees differ by mode.

User experience

  • Deployment/release detail shows source SHA, artifact digest, config revision/diff, build provenance, actor/trigger, steps, health checks, domains, duration, and resulting resource versions.
  • Promote an already-built immutable artifact from one environment to another after reviewing environment-specific config and approvals.
  • Choose only supported strategies—rolling, atomic/symlink, blue-green, or canary—with an explanation of capacity/cost and rollback behavior.
  • Rollback selects a release, previews config/data caveats, monitors activation health, and preserves the failed release for investigation.

Scope

  • Define a common release record and artifact identity for static, compute, serverless zip/image, ECS/container, and Compose deploys.
  • Store immutable manifests, resolved non-secret configuration hashes, artifact digests, and provenance/attestations.
  • Implement promotion without rebuilding, environment approvals, and optional manual production gates.
  • Add health-gated activation, drain/grace periods, automatic rollback policy, and strategy capability detection.
  • Unify history, compare, redeploy, rollback, retention, and release pinning across CLI/API/dashboard.

Technical direction

  • Separate build from release activation and make both durable jobs.
  • Use content-addressed artifacts and verify digests before activation.
  • Provider drivers implement strategy primitives and return structured traffic/health transitions.
  • Keep database migrations/hooks explicit with forward/backward compatibility metadata; never imply data rollback.

Safety and operability

  • Production promotion and rollback show exact target, artifact, config diff, health gates, and any irreversible hooks.
  • Automatic rollback has bounded attempts and stops on ambiguous state.
  • Artifacts and provenance are access-controlled and secrets remain write-only references.
  • Retention never deletes active, pinned, or rollback-required artifacts.

Acceptance criteria

  • The same tested artifact can be promoted from staging to production without rebuild.
  • A failed health gate restores the prior active release for supported strategies and records every transition.
  • History/compare/rollback behavior is consistent across at least static, compute, serverless, and container paths.
  • Unsupported zero-downtime strategies are unavailable with an actionable explanation.

Verification

  • Release state-machine, artifact digest, promotion, retention, health gate, drain, and rollback failure tests.
  • Driver contract tests for strategy capability and transition reporting.
  • End-to-end healthy promotion and failed canary/rolling activation with rollback.

Non-goals

  • Rolling back database state automatically.
  • Claiming zero downtime for workloads without health/readiness support.
  • Rebuilding artifacts during promotion.

Sign in to comment on this issue.