<!-- Generated from models/_canonical by scripts/build-model-variants.js. Edit the canonical source, not this file. Behavioural profile for GLM: scripts/model-profiles.json -->

#Task boundary

1. Implement only what the task names; no extra abstractions or files.
2. English-only comments and identifiers.
3. Stop when the checklist passes.

---

#Purpose

Rules for undoing a release. The relevant measure of a deployment system is not how rarely it fails — it is how quickly a failure is reversed.

The target: any release can be reverted in minutes, by one person, without a meeting. Everything below serves that.

---

#Detect before you can reverse

You cannot roll back what you have not noticed. Automate the decision.

yaml
# Abort and revert if the new version breaches its budget during the bake window
- alert: DeployErrorBudgetBurn
  expr: |
    sum(rate(http_requests_total{status=~"5..", version="$NEW"}[5m]))
      / sum(rate(http_requests_total{version="$NEW"}[5m])) > 0.02
  for: 3m

1. Compare the new version against the old, not against an absolute threshold — a 2% error rate may be normal for one service and catastrophic for another.
2. Bake for a defined window before promoting further.
3. Automate the rollback trigger. Relying on someone watching a dashboard fails at 2am, which is when it matters.
4. Tag every metric and log line with the version (version, git_sha, deployment_id), or you cannot attribute a regression to a deploy at all. The same label must appear on http_requests_total, http_request_duration_seconds and every business counter, or the comparison above cannot be written. → Backend/monitoring

---

#Rolling back code is the easy part

bash
kubectl rollout undo deployment/api                    # previous ReplicaSet
kubectl set image deployment/api api=$REGISTRY/api@sha256:<known-good>

Requirements for this to be fast and safe:

1. Immutable, digest-addressed artefacts. Rebuilding from a git revert takes ten minutes you do not have and may produce different bytes.
2. Keep the previous N versions available in the registry and, where applicable, in the platform's revision history.
3. The rollback path must be the same mechanism as the deploy path. A separate emergency procedure is one nobody has practised.
4. Revert the commit too, so the next deploy does not reintroduce the fault.

revisionHistoryLimit: 0 is a configuration that removes your ability to roll back at all — check it, because some Helm charts set it to save etcd space.

---

#Database changes are what actually block rollback

Code rolls back in seconds. A schema change frequently cannot roll back at all — a dropped column's data is gone.

The rule: make every migration backward compatible with the currently deployed code, then rolling back the code never requires rolling back the database.

sql
Deploy 1  Expand    add column, nullable; new code writes both, reads old
Deploy 2  Migrate   backfill; new code reads new
Deploy 3  Contract  drop the old column — only once deploy 2 is proven

Each deploy is independently reversible because the schema at every point serves both versions. → Database/migration

A destructive migration must be separated from the deploy that stops using the data, by enough time to prove the new code works.

---

#Feature flags make rollback instant

ts
if (await flags.enabled("new-checkout", { userId })) return newCheckout();
return legacyCheckout();

A flag decouples deploy from release. The fix for a bad feature becomes a configuration change — seconds, no rollout, no rebuild — instead of a redeploy.

1. Kill-switch anything risky: a new payment path, a rewritten flow, an expensive query.
2. Roll out by percentage so a fault affects 1% of users, not everyone.
3. Keep both paths working while the flag exists, and remove the flag once the new path is proven. Stale flags become dead branches nobody dares delete, and an untested legacy path is not a rollback target.
4. Flag state changes are audited: who turned what on, when.

The list is ordered deliberately: reach for the fastest mechanism the failure allows, and design so the fast ones are available. A change that can only be reversed by the last row is a change that has no rollback.

---

#Rehearse it

A rollback procedure that has never been executed is a document, not a capability.

1. Roll back in staging on a schedule, timed, following the runbook as written.
2. Include the awkward cases: a rollback with a migration in flight, a rollback of a queue-consumer change with messages in the new format.
3. Write down the decision criteria in advance — what error rate, over what window, triggers a rollback — so the choice is not made under pressure by whoever happens to be online.
4. Prefer rolling back over fixing forward during an incident. Diagnosis takes longer than reversal, and users are affected throughout.

A written trigger looks like this, and belongs in the runbook before the incident, not in a chat thread during it:

vbnet
Roll back immediately if, during the 15-minute bake window:
  - 5xx rate on the new version exceeds 2× the old version's, for 3 minutes, or
  - p99 latency on any critical route exceeds 1.5× its pre-deploy value, or
  - any `payment.*` or `auth.*` error counter is non-zero above its baseline.
Decision owner: the deployer. No approval required to roll back.

"No approval required to roll back" is the load-bearing line. A rollback that needs someone to be found is not a minutes-scale rollback.

---

#Anti-patterns

---

#Checklist

• Every metric and log line carries the deployed version
• Deploy health compares the new version against the previous one
• A bake window precedes full promotion
• Rollback triggers automatically on an error-budget breach
• Artefacts are immutable and addressed by digest
• Previous versions remain available for rollback
• Rollback uses the same mechanism as deployment
• Rolling back is accompanied by reverting the commit
• Every migration is backward compatible with the running code
• Destructive schema changes are separated from the deploy that stops using them
• Risky changes ship behind a kill switch
• Feature rollout is percentage-based
• Flags are removed once the new path is proven
• Flag changes are audited
• The rollback procedure is rehearsed on a schedule and timed
• Rollback decision criteria are written down in advance