ubgo/logger is in the zero-allocation tier alongside zap and zerolog, while remaining slog-native. The typed hot path allocating 0 objects is enforced by a CI gate, so it cannot silently regress.
Apple M-series, Go 1.24, output to io.Discard:
| Path | ns/op | B/op | allocs/op |
|---|---|---|---|
Typed hot path (String/Int[T]/Bool, 5 fields) |
~295 | 0 | 0 |
| Disabled level (below threshold, gated out) | ~7 | 0 | 0 |
Through the slog bridge (5 attrs) |
~698 | 320 | 1 |
stdlib slog JSON (reference) |
~704 | 0 | 0 |
The slog-bridge row is published, not hidden. When you log via slog with more than 5 attributes, slog itself heap-allocates its Record attr storage — that 320 B / 1 alloc is slog's cost, not avoidable by any backend. Many "portable via slog" setups silently pay 10–40× because they pair slog with a slow backend; ubgo's bridge is ~698 ns (vs stdlib slog's own ~704 ns), so you pay slog's inherent cost and nothing extra.
Takeaway: use the native typed API (log.Info("msg", logger.String(...))) on hot paths for 0 allocations; use the slog bridge for portability and accept slog's own attr-allocation cost.
- Scalar fields (
String,Int[T],Float[T],Bool,Dur,Time) are stored unboxed in theFieldstruct — nointerface{}boxing. OnlyAnyescapes to reflection. Recordand encode buffers are pooled (sync.Pool).- The level name is a constant string (
Level.lower()), not an allocated[]byte— this was the single allocation found via memprofile and eliminated. - Disabled levels short-circuit before any field evaluation.
alloc_test.go uses testing.AllocsPerRun to assert the typed path and the disabled-level path both allocate exactly 0 objects. It runs as a dedicated non--race CI step (the race detector instruments allocations, so the assertion is meaningless under -race and is skipped there via a build-tagged constant).
Any change that reintroduces an allocation fails CI before it ships.
cd github.com/ubgo/logger
go test -run=^$ -bench=. -benchmem ./
go test -run TestZeroAlloc -count=1 ./ # the gateBenchmarks live in bench_test.go (typed, disabled-level, through-slog-bridge, stdlib-slog reference).
SyncTransport— lowest latency per call, but the caller pays sink I/O.ChannelTransport— moves I/O off the caller; bounded channel; great default.DisruptorTransport— lock-free Vyukov MPMC ring; lowest overhead under many concurrent producers (race-tested at 8 producers × 2000 records on a 64-slot ring with zero loss underBlock).
Pick DropNewest/DropOldest if predictable latency matters more than completeness under extreme load — and watch transport.Dropped() / log.Metrics().