Give every request a latency budget

When a request is slow, teams hunt for the one slow component. Usually there isn't one. There are ten components each taking a reasonable 30ms, and nobody ever added them up.

Budgets change conversations

Once the checkout path had a written budget — 250ms total, itemized per hop — every design review got sharper. "This adds a network call" stopped being an abstract concern and became "this spends 15% of the budget."

• Write the budget down, per hop, in the design doc.
• Alert on budget violations per component, not just on the total.
• Treat retries as spending the budget twice.

The tail is the product

Averages flatter you. Your p50 user had a fine experience either way; your p99 user is deciding whether to come back. Budget for the tail and the average takes care of itself.