How do you handle a request that needs data from five services without killing tail latency?

Synchronous fan-out chains multiply latency and failure probability, so aggregate behind an API gateway or a BFF and issue the calls in parallel rather than sequentially, with per-call timeouts and breakers.

Better still, avoid the fan-out at request time: precompute a read model (CQRS) updated asynchronously from each service's events so the read is a single local lookup. Where you must call live, degrade gracefully — return partial responses with the non-critical services' data omitted rather than failing the whole request when one slow dependency times out.