Integrating services together

Choose the coupling style deliberately

Synchronous HTTP calls are fine for user-facing flows with tight latency budgets. Background work—billing reconciliation, search indexing, analytics rollups—usually belongs on a queue or log stream so spikes in one service do not become cascading timeouts in another. If you are unsure, default toward asynchronous handoffs for anything that can wait minutes.

Idempotency and exactly-once illusions

Networks lie; clients retry; webhooks double-fire. Design handlers so the second identical event is harmless—keys, deduplication windows, and clear upsert semantics beat “hope the vendor is polite.” Document the contract in your internal knowledge base; see documentation teams actually use.

Identity as a spine

Before you wire a fifth OAuth client, decide who issues tokens, which claims downstream services trust, and how service-to-service auth rotates. A short internal standard prevents every team inventing its own HMAC scheme. Basics live in identity and remote access basics.

Failure domains and blast radius

Draw boxes: if this integration fails, which user journeys degrade gracefully versus halt entirely? Put circuit breakers and cached fallbacks on read-mostly paths; fail loud and fast on financial writes where partial success is worse than a hard stop.

Observability across boundaries

Correlate with a trace or request ID propagated through headers. Standardize structured log fields so you can pivot from a user complaint to the offending job without grep archaeology—more in observability with signal.

Related: Automation that earns its keep · Cost of tool sprawl