The problem
The platform had a contractual availability commitment and no instrumentation capable of demonstrating whether it was being met. Incidents were resolved by whoever noticed first.
What we did
- Defined SLIs from real customer journeys rather than server-side liveness
- Instrumented tracing across the request path end to end
- Introduced incident command roles, and ran a game day before enforcing the policy
- Agreed an error-budget policy with teeth — feature freeze when the budget is spent
