How would you build an endpoint that needs data from three slow upstream services?

Call them concurrently rather than in sequence. In an async framework, launch all three with asyncio.gather or a task group so total latency is the slowest call, not the sum. Give every call a timeout, decide which responses are optional so you can degrade instead of failing, and cache what is stable. If the client can tolerate it, return a job id instead.

Sequentially those three latencies add up, so the first move is running them together. With httpx and an async handler I gather the three coroutines, or use a task group so a failure cancels the siblings cleanly, and the endpoint then costs roughly the slowest upstream instead of the sum. Every call gets an explicit timeout, because the default in most clients is effectively forever and one hung socket will exhaust the worker pool. Then I decide what is essential. If two of the three are decoration, I collect exceptions instead of raising and render the page without them rather than throwing a five hundred, and I emit a metric so we still know it happened. Anything stable gets cached, since even sixty seconds of caching removes most of the load. If the client can tolerate it, the better shape is accepting the request, queueing the work, and returning a job id to poll, which takes the upstream latency off the request path entirely.