Software Performance: A Shape Not a Number
summary
Claim: measure latency as a shape not a number, using unsampled histograms instead of averages or lone percentiles. The talk shows Lightstep-style distributed tracing tied to histograms, a real incident with a 17.7 seconds 99th-percentile request, a critical-path finding that slow listing (not reads) and decompression/unmarshal contention cause the delay, and experiments where dropping buckets from 200 to 40 destroys signal. It stresses freshness (~tens of seconds), multi-day horizon, and handling high cardinality to triage real microservice modes.
AI-written summary · about
chapters (6)
- 0:00Introduction and goals
- 2:11Measuring system performance
- 11:21Dimensions and cardinality
- 18:02Debugging and magic
- 21:07Refining search space
- 29:56Conclusion and scorecard
related talks
45mUnderstanding Microservice Performance
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...
27mLatency SLOs Done Right
50mDesigning for Performance
01h 01mThe Performance Investigator's Field Guide
31mLatency SLOs Done Right
37mMonitoring the Health of Your App
35mTracing is More Than Traces: The Insights in Trace Aggregates
