Integration

Every connection needs limits for authentication, queues, heartbeats, fan-out, and shutdown.

Compatibility comes from additive change, explicit semantics, and a deprecation path clients can observe.

Batching fixes duplicate fetch shape only when caching scope and authorization remain request-local.

Validate issuer, audience, signature, expiry, and scopes at the resource boundary.

Treat the schema as reviewed source code and test implementations and clients against it.

Put timeouts, quotas, translation, idempotency, and observability behind an anti-corruption layer.

Separate syntactic validation, domain rules, and authorization, then return stable problem details.

Prefer compatible evolution; version only when semantics truly conflict and plan convergence.

Bind a client key to request identity and persist the decided response for the retry window.

Sign payloads, preserve event identity, acknowledge quickly, and make processing replay-safe.

Batch APIs need per-item outcomes, bounded size, deterministic identity, and explicit atomicity.

Stream with hard limits, quarantine content, verify type independently, and decouple scanning from trust.

Deadlines cross service boundaries; status codes must separate retryable transport conditions from domain outcomes.

Token buckets express sustained rate and burst capacity while keeping admission decisions cheap.

A cursor must encode a stable total ordering, not merely an offset into changing rows.