benchmarking

When latency matters, I benchmark the smallest interesting unit and watch allocations. In Go, a surprising number of regressions come from accidental heap churn: converting []byte to string, building lots of temporary maps, or using fmt.Sprintf in loo