golang-performance

Profile first with pprof to find the actual bottleneck—intuition is wrong about 80% of the time. Then write an atomic benchmark, measure a baseline with -benchmem -count=6, apply one optimization at a time, and verify with benchstat.

Reuse slice backing arrays with append(s[:0], ...), preallocate maps and slices with size hints, use sync.Pool for frequently allocated short-lived objects, and avoid interface boxing by using typed parameters or generics.

Use fgprof, which captures both on-CPU and off-CPU (I/O wait) time. Standard pprof CPU profiles miss time spent blocked on network, database, or filesystem calls, which is the likely cause when CPU usage is low but latency is high.

Yes. Go inserts padding to satisfy alignment requirements, so ordering fields from largest to smallest (int64 first, bools last) reduces struct size. Use the fieldalignment tool to detect structs with wasted padding automatically.

Go maps never release bucket memory when entries are deleted—a map that held millions of keys retains that allocation forever. Replace the map with a fresh make() allocation so the old bucket array becomes eligible for garbage collection.

Avoid pooling objects larger than about 32KB, since large allocations bypass pool size classes and GC handles them efficiently. Also skip pooling for infrequently used objects where pool overhead exceeds the allocation savings.