Go Performance
Workflow
-
Identify profile source (file, URL, or generate from tests/server)
-
Analyze with go tool pprof
-
Report findings with actionable recommendations
Quick Reference
Profile Test Flag HTTP Endpoint
CPU -cpuprofile
/debug/pprof/profile
Heap -memprofile
/debug/pprof/heap
Goroutine
/debug/pprof/goroutine
Key Commands
Generate from benchmarks
go test -bench=. -cpuprofile cpu.prof -memprofile mem.prof
Analyze (interactive or one-liner)
go tool pprof -top -cum cpu.prof go tool pprof -http=:8080 cpu.prof # web UI with flame graphs
Memory analysis
go tool pprof -alloc_space -top mem.prof # bytes allocated go tool pprof -alloc_objects -top mem.prof # allocation count (GC pressure)
Compare profiles (find leaks/regressions)
go tool pprof -base old.prof new.prof
Filter results
go tool pprof -focus='mypackage.*' cpu.prof
Tips
-
Write a benchmark first to reproduce/confirm/measure the issue
-
Focus on quick wins first: easy fixes with high impact
-
If hotspots are in external libraries, still analyze - source is in $GOPATH
-
CPU profiles need 30+ seconds for meaningful data
-
For memory leaks: compare heap profiles at two points in time
-
Use -base flag to find regressions between profiles
Writing benchmarks
Every time we work on some performance issue, we should:
-
make sure there is a benchmark
-
if there isn't one, we create it, calling the function that we aim to improve
Then, we create a second benchmark, and a new production function with the changes we want.
This way we can quickly run and compare both benchmarks.
Once we're happy with the changes, we replace the old function and old benchmark with the new ones.
Commit messages should always have the benchmark results in them.