An infrastructure release prompt should state what must be true after apply and how each claim will be checked. Compare the live resource, a fresh plan, application behavior, and relevant monitoring with the approved change. A zero-diff plan alone may miss a user-visible failure, while a healthy dashboard cannot prove the desired configuration was applied. Define rollback or forward-fix conditions before release and name the owner who can act. Some changes are not fully reversible: reducing retention later will not recreate events already deleted. The release packet should preserve plan identity, apply receipt, observed state, tests, and unresolved limits.
Infrastructure release: verify live state and rollback limits
Operational case
For Aster, pipeline run R-47 reports the queue setting changed from 47 to 71 hours. A direct provider read confirms 71 hours; a fresh plan has no unexpected actions. The reports team checks that new messages are accepted and an existing test message remains readable. Operations watches queue age and dead-letter count during the observation window. A rollback plan can restore the configured retention value, but cannot guarantee recovery of a message that expired before the change. The reviewer records the earlier blocked database replacement as resolved separately rather than pretending it never occurred.
Receipt: R-47 applied P2 to verified production workspace.
State: provider read says 71h; fresh plan shows no unexpected action.
Behavior: message acceptance and retained test record pass.
Monitor: queue age, failed delivery, dead-letter count.
Limit: reverting retention cannot resurrect expired messages.Performance and operating cost
For V postchecks across R relevant resources, review work is O(V+R), plus an observation period set by workload behavior. A fresh plan and provider read add calls but test different failure modes. Track the tested time window: a five-minute smoke test cannot establish long-term retention behavior by itself. A small controlled fixture can verify that the service still accepts and reads messages now, while retention outcomes need later observation or a safe accelerated test environment.
Common Mistakes
- Do not infer application health solely from a zero-diff infrastructure plan.
- Do not promise rollback can restore already expired data.
- Do not close the change without checking the actual resource and operator receipt.
Connected lessons
- Production prompt engineering
- Prompt Engineering
- Prompt release review: assemble the decision packet
- Code lab: block a critical prompt regression
- Infrastructure drift: distinguish emergency repair from unauthorized change
- Rollback image retention: keep every approved fallback pullable
- Infrastructure prompts: bind scope, owner, and environment
- Infrastructure plans: classify every action before apply
- Infrastructure prompts: quarantine state secrets and drift
- Infrastructure apply: separate review from execution
- Project: review an Aster queue retention change
- Infrastructure-change prompt decisions
Continue with: Backfill prompts: hold cutover until fallback is rehearsed.
