Average latency on an endpoint rarely tells you where time goes. The expensive story sits in the fan-out: identity lookups, pricing calls, inventory checks, and a notification side trip that somehow stayed on the critical path.
In topology reviews we rebuild call graphs from traces and gateway logs, then annotate each edge with ownership and timeout policy. Surprises appear quickly — services that were meant to be asynchronous still block, or a “temporary” helper became a hub.
Once the graph is honest, budget conversations change. Leaders can decide which hops to collapse, which to cache, and which to move off the request path entirely.
Draw the graph you actually run, not the one from last year’s architecture day.