Latency hunts that start with one customer journey
Why whole-estate profiling wastes the first week, and how to pick a journey that will teach you something shippable.
Whole-cluster CPU graphs feel productive. They rarely tell you which change to make on Tuesday.
Anchor the first week to one journey: place order, generate quote, confirm booking. Instrument that path end to end. When p95 moves after a change, everyone can see it without arguing about methodology.
Useful evidence
- Trace spans across the hop that developers already distrust
- Database wait events for the queries on that path
- Retry and timeout settings that amplify a slow dependency
Less useful evidence
- Average CPU across twenty services
- A flame graph of a batch job that does not run during the customer spike
A mild warning
Sometimes the journey you pick is politically convenient and technically wrong. If marketing’s homepage widget is slow but finance’s reconciliation blocks warehouse pickups, reconcile first. Customer-facing does not always mean customer-critical.
For a measurement-led engagement, see our performance tuning consultation.