Service levels
Availability is measured per endpoint group, not per pod. The error budget is consumed by 5xx responses and by requests that exceed the latency target, because a request that takes eight seconds has failed even when it eventually succeeds.
Tracing
Every request carries a trace id through the HTTP layer, the query listener and the queue jobs it dispatches. The trace is the first artefact in an incident review: it answers which query was slow, not just which endpoint was.
Changing the contract
Additive changes ship behind a version. A field is never repurposed, and a removal is announced for one release before it is enforced. Three consumers means a breaking change is a coordinated release, not a deploy.
On-call
The runbook starts with the endpoint's budget report, then the slow-query log, then the database's own plan for the offending statement. Restarting pods is the last step, not the first.