Journal
Reading latency when packets bounce regions
When a request kisses two PoPs before it reaches origin, a single p95 number becomes a polite lie. Application analytics for edge workload visibility starts by naming each hop: client to edge, edge to peer, peer to origin, and origin compute.
We ask learners to draw the path on paper before opening any vendor UI. That sketch usually reveals a cache miss story masquerading as “API slowness.” Once hops are named, you can attach spans with stable attributes — region, cache status, route family — without inventing a new label for every deploy.
A practical check: compare edge queue time against origin handler time during a quiet Tuesday and a promotional Saturday. If Saturday only inflates the edge queue, your scaling conversation belongs with capacity at the PoP, not with rewriting business logic. If both inflate, you likely have a shared dependency starving under fan-out.
Resist dashboards that average regions into one pastel line. Pastel is fine for our brand; averaged regions are not fine for incidents. Keep separate panels, then write a one-sentence narrative that product can read without translating percentiles.