How do you make a service call to a dependency that might be slow or down?

Set an explicit timeout on every call, shorter than the caller's own deadline, and never rely on the client library default. Retry only idempotent operations, with exponential backoff plus jitter and a small bounded attempt count. Add a circuit breaker so repeated failures fail fast instead of queuing, and define what degraded looks like: a cached value, a partial response, or a clear error rather than a hanging request.

The first thing is a timeout on every outbound call, chosen from the latency distribution rather than picked at random, and always shorter than the deadline I was given, so I fail before my caller gives up. I propagate that remaining deadline downstream, otherwise work continues on a request nobody is waiting for. Retries only apply to idempotent calls, and even then with exponential backoff and jitter, capped at two or three attempts. The reason for the cap is amplification: if every layer retries three times, a blip at the bottom becomes an order of magnitude more load, which is how a slow dependency turns into a full outage. So I also like a retry budget, where retries are limited to a small percentage of total requests. A circuit breaker sits on top so that once failures cross a threshold I fail immediately for a cooldown and probe occasionally, which stops me piling requests onto something already struggling. And I decide up front what degraded means; on a pricing service we served the last cached price with a staleness flag rather than failing checkout.