Typed errors
QUE returns typed errors rather than generic failures: an availability rejection names the computation it ran, an auth failure names the missing session or role, a provider failure surfaces what the external system returned. The operator view and the API surface both carry these typed results so recovery decisions can be made from the error itself, not from guesswork.
Fail-closed and fail-open
Boundaries that protect customers fail closed; boundaries that protect availability fail open. Redis failure makes rate limiting deny rather than allow. An unreachable database makes booking reads return errors rather than stale data. A missing alertmanager secret makes the webhook sink refuse everything rather than accept unauthenticated payloads. Server error reporting is fail-open for availability (an unset DSN logs locally and never throws) but the reported data itself is fail-closed for correctness — malformed releases and forbidden customer keys are dropped with a stated reason.
| Dependency failure | Boundary | Behavior |
|---|---|---|
| Redis unreachable | Rate limiting | Deny (fail-closed) |
| Database unreachable | Booking reads | Typed error, no stale data (fail-closed) |
| Alertmanager secret unset | Alert webhook | 401 on every request (fail-closed) |
| Sentry DSN unset | Error reporting | Console log, no throw (fail-open) |
No raw exceptions in public surfaces
Public and customer-facing responses never expose raw exception traces. Internal failures are captured with context for operators; what customers see is the typed, actionable form. If you encounter a raw stack trace on a public surface, report it — that is a defect.