Skill: performance

What I do

I teach Go performance: measure first with benchmarks and pprof, identify bottlenecks with data, then optimise allocations, concurrency, and algorithms. Never optimise without profiling evidence.

When to use me

• Investigating slow endpoints or high memory usage
• Writing benchmarks to measure before/after performance
• Profiling CPU, memory, or goroutine contention with pprof
• Reducing allocations in hot paths
• Choosing between performance trade-offs (memory vs CPU, latency vs throughput)

Core principles

1. Measure first — Never optimise without benchmark data; intuition is usually wrong
2. Profile, don't guess — Use pprof to find the actual bottleneck, not the suspected one
3. Allocations dominate — In Go, reducing allocations often gives the biggest wins
4. Benchmark before and after — Every optimisation must show measurable improvement
5. Readability over micro-optimisation — Only sacrifice clarity for proven, significant gains

Patterns & examples

Writing benchmarks:

func BenchmarkProcess(b *testing.B) {
    data := setupTestData()
    b.ResetTimer() // exclude setup from measurement

    for i := 0; i < b.N; i++ {
        process(data)
    }
}

// Run: go test -bench=BenchmarkProcess -benchmem -count=5
// Output: BenchmarkProcess-8  50000  23456 ns/op  1024 B/op  12 allocs/op

Profiling with pprof:

# CPU profile
go test -cpuprofile=cpu.prof -bench=.
go tool pprof -http=:8080 cpu.prof

# Memory profile
go test -memprofile=mem.prof -bench=.
go tool pprof -http=:8080 mem.prof

# In running server (import _ "net/http/pprof")
go tool pprof http://localhost:6060/debug/pprof/profile?seconds=30

Allocation reduction techniques:

// ❌ Allocates new slice every call
func collect(items []Item) []string {
    var names []string
    for _, item := range items {
        names = append(names, item.Name)
    }
    return names
}

// ✅ Pre-allocate with known capacity
func collect(items []Item) []string {
    names := make([]string, 0, len(items))
    for _, item := range items {
        names = append(names, item.Name)
    }
    return names
}

// ✅ Reuse buffers with sync.Pool
var bufPool = sync.Pool{
    New: func() any { return new(bytes.Buffer) },
}

func process(data []byte) string {
    buf := bufPool.Get().(*bytes.Buffer)
    defer bufPool.Put(buf)
    buf.Reset()
    buf.Write(data)
    return buf.String()
}

String building:

// ❌ O(n²) — allocates new string each iteration
result := ""
for _, s := range items {
    result += s
}

// ✅ O(n) — single allocation
var b strings.Builder
b.Grow(estimatedSize) // optional pre-allocation
for _, s := range items {
    b.WriteString(s)
}
result := b.String()

Common bottleneck locations:

Anti-patterns to avoid

• ❌ Premature optimisation — Optimising code without profiling data; wastes time, hurts readability
• ❌ Micro-benchmarks in isolation — Benchmarking a function that's called once; focus on hot paths
• ❌ Ignoring benchmem — CPU speed matters less than allocation count in GC-heavy workloads
• ❌ sync.Pool everywhere — Only helps for frequently allocated, short-lived objects; adds complexity
• ❌ Caching without eviction — Unbounded caches leak memory; always set a size limit or TTL

KB Reference

~/vaults/baphled/3. Resources/Knowledge Base/AI Development System/Skills/Performance-Profiling/Performance.md

Related skills

• benchmarking - Detailed benchmark methodology and comparison
• profiling - Deep-dive into pprof, trace, and flame graphs
• concurrency - Goroutine scheduling and contention profiling
• golang - Idiomatic Go patterns that are inherently efficient