Freeze Changes During an Incident
When a system is failing, a well-meaning deploy from another team changes the variables mid-investigation and can turn a contained incident into a wider one. A scoped freeze keeps the diagnostic signal clean without halting the whole company.
Procedure
- Scope the freeze to the affected surface, not everything: freeze deploys to the failing service and its direct dependencies, not to an unrelated docs site.
- Announce it with the exact scope and who can lift it: "Change freeze on checkout and payments-api until further notice; IC lifts it."
- Carve an explicit exception path: security fixes and the IC's own mitigation deploys are always allowed; everything else goes through the IC.
- Enforce mechanically where possible — flip the CD pipeline's gate or the deploy-freeze window, do not rely on people reading the channel.
- Record the freeze start time and scope in the decision log; if a change does land, note who approved it and why.
- When impact is contained, lift the freeze explicitly and announce it; a freeze left on silently blocks other teams' releases and breeds workarounds.
- A prod rollback or mitigation is NOT a change for freeze purposes — approve it. The freeze protects the investigation, it does not stop the fix.
- Notify teams whose releases will be blocked, by name, so they can plan around it instead of discovering the gate.
- Set a review point (e.g. at the next cadence tick) to re-confirm or lift, so the freeze cannot live past the incident.
Pitfalls
- Freezing everything org-wide "to be safe", blocking unrelated releases and creating pressure to bypass.
- An unannounced freeze that another team unknowingly violates, wasting the investigation.
- Forgetting to lift the freeze after the incident, so it persists and is overridden without notice.
- Blocking the mitigation rollback itself because "changes are frozen" — the freeze protects the fix, it does not forbid it.
- Relying on an honour-system freeze with no pipeline gate, so a well-meaning hotfix slips through.
- Leaving the freeze scoped to a service name that is actually a dependency, so the real surface stays open.
Verification
grep -E 'freeze (on|lifted)' incident/SEV*-2026-*.md
# passes when a scoped freeze has both an open announcement and a matching lift
Related: close-an-incident-and-resume-normal owns the explicit lift so the freeze never outlives the incident.