DeanKelly751 Profiling
Purpose
Systematic procedure for finding performance bottlenecks through measurement.
Workflow
Step 1: Define the Problem
- What is slow? (specific endpoint, operation, user action)
- How slow? (current measurement)
- What's the target? (acceptable latency/throughput)
- When is it slow? (always, under load, specific conditions)
Step 2: Reproduce Consistently
- Isolate the slow scenario
- Create a repeatable test case
- Establish baseline measurement
Step 3: Profile at the Right Level
| Symptom | Profile type | What to look for |
|---|---|---|
| High CPU | CPU profiler / flame graph | Hot functions, tight loops |
| Growing memory | Heap profiler / allocation tracker | Leaks, large allocations |
| Slow responses | Request tracing | Time spent in each phase |
| High latency variance | Latency distribution | P99 outliers |
| Disk/network delays | I/O profiling | Blocking calls, large transfers |
Step 4: Identify the Bottleneck
Read the profile:
- Flame graphs: Wide bars = time spent there. Look for unexpected width.
- Allocation profiles: Large/frequent allocations in unexpected places.
- Trace waterfalls: Long spans, sequential calls that could be parallel.
Step 5: Optimize and Verify
- Change one thing at a time
- Re-measure with same test case
- Compare against baseline
- Verify no regression in other dimensions
Profiling Principles
- Profile in production-like conditions (data volume, concurrency)
- Profile the actual bottleneck, not what you assume is slow
- Look at distributions (p50, p95, p99), not just averages
- Low-overhead profiling for production; detailed profiling in staging
Related Skills
performance/knowledge-optimization-principles— decide what to profileperformance/execution-load-testing— profiling under loadobservability/execution-distributed-tracing— request-level profiling