planning-progressive-delivery
Plan how a change reaches users so that deploying is low-risk and reversible. Decouple deployment
SKILL.mdsubagents/devops-sre-advisor/skills/planning-progressive-delivery/SKILL.md
GitHubPlanning progressive delivery
Purpose
Plan how a change reaches users so that deploying is low-risk and reversible. Decouple deployment from release using feature toggles and dark launches, roll changes out gradually with blue-green or canary patterns, and keep a fast, tested rollback path (P005). The point is to separate "the code is in production" from "users see the feature", then expose the change in controlled steps you can undo.
When to use
- Planning how to release a risky or user-visible change.
- A team wants to deploy frequently without exposing unfinished features.
- Reviewing a deployment strategy for production readiness.
Skip the full apparatus for a trivial change to an internal tool where staged rollout adds no value — say so rather than imposing ceremony.
Procedure
- Decouple deployment from release (P005). Use feature toggles and dark launching so deploying code to production no longer means exposing a feature to users (CL009). This lets the team integrate and deploy continuously while controlling who sees what, and turns a release into a configuration change rather than a redeploy.
- Choose the rollout pattern for the risk:
- Blue-green — run two identical environments and switch traffic between them to release with little or no downtime and roll back fast by switching back (CL010). Use when you want an instant, whole-environment cutover and an instant reverse.
- Canary — expose a small fraction of traffic to the new version first, watch for regressions, and roll back quickly if the canary degrades (CL056). Use when you want to limit blast radius and gather real signal before full exposure. Toggles and dark launches compose with both: they gate the feature; the pattern gates the traffic.
- Instrument the rollout and define the abort condition (P005). Before exposing traffic, wire production telemetry and automated anomaly detection so a bad release is detected and acted on before customers are widely affected (CL011). State the metric and threshold that trips a rollback up front — a rollback path is only real if it is fast, tested, and has a clear trigger.
- Stage the exposure. Sequence the steps (dark launch → small canary → wider canary → full, or blue-green cutover with a hold-and-watch window), naming who/what is exposed at each step and the go/abort check between steps.
- For
review/validaterequests, check the plan has: deployment decoupled from release, a rollout pattern matched to the risk, instrumentation with a named abort condition, and a fast tested rollback. Flag any plan that relies on a big-bang cutover with no rollback as the gap to close.
Inputs
- The change, its blast radius, and how user-visible it is.
- Current deployment mechanism and whether toggles/dark launch, blue-green, or canary are available.
- The telemetry available to judge a release, and what "healthy" looks like for this service.
Output
A staged rollout plan: how deployment is decoupled from release, the chosen pattern (blue-green or
canary) with rationale, the instrumentation and named abort condition, and the rollback path —
tied to P005. For validate, a pass-or-gap line per element.
References
references/progressive-delivery-patterns-reference.md— the pattern catalogue (toggle, dark launch, blue-green, canary, rollback).- Sibling skills:
designing-deployment-pipelines(what produces the deployable),defining-slos-and-error-budgets(the burn-rate signal that can drive an abort).
Provenance
Derived from principle P005 (claims CL009, CL010, CL056, CL011; evidence EV009, EV010, EV056,
EV011) across the-devops-handbook-c4933b3c and comp500-15893c30 (both distillation-only —
paraphrase, no verbatim quotation). Pattern names are stable practice; specific tooling for toggles,
traffic shifting, and telemetry changes quickly, so verify current vendor and project documentation.