Performance

Philosophy

do-auditlog is designed for minimal overhead during container operation. All event capture happens in memory — no file I/O during registration, invocation, or shutdown. Export is a single json.Marshal or line iteration.

Benchmarks

Benchmarked on AMD Ryzen AI MAX+ 395:

BenchmarkHookOverhead_Invocation    ~1,658 ns/op    6 allocs    (enabled)
BenchmarkHookOverhead_Disabled       ~113 ns/op    4 allocs    (disabled)
BenchmarkHookOverhead_Registration  ~21,982 ns/op   54 allocs    (full container)

Overhead: ~1.7us per cached invocation when enabled. Zero cost when disabled.

The disabled cost (~113ns, 4 allocs) is entirely from samber/do’s own hook dispatch mechanism — the plugin installs empty hooks that never execute recorder logic.

Concurrency Model

A single sync.RWMutex protects all mutable state. This reduces lock acquisition overhead from 2-4 per hook to exactly 1.

• sequence and invocationSeq are atomic.Int64 — no mutex needed for counters.
• Each hook acquires mu once, performs all mutations, then releases.
• OnEvent callback is called outside the lock to avoid blocking the hot path.
• BuildReport() uses mu.RLock() for reading — concurrent reads don’t block each other.

Disabled Mode

When Enabled: false, the plugin returns empty hooks. samber/do never calls recorder methods. The overhead is entirely samber/do’s own hook dispatch (~113ns, 4 allocs).

plugin, _ := auditlog.New(auditlog.Config{
    Enabled: false, // zero-cost hooks
})

Export Cost

Export is only paid when you need the data:

Operation	Cost
plugin.Report()	One mutex read + struct assembly
plugin.ExportToFile()	One json.MarshalIndent + file write
plugin.ExportEventsToNDJSON()	Line iteration + per-event json.Marshal
plugin.ExportToHTML()	Template render (templ)

Reproducing

go test -bench=. -benchmem ./...