Benchmarks

benchmark/run.lisp is a reproducible SBCL benchmark for handle-log-record under representative payloads. It reports the minimum wall-clock time and the bytes consed per call over several full-GC'd repetitions:

CL_SOURCE_REGISTRY="/path/to/cl-weave//:/path/to/cl-json-kit//:$(pwd)//:" \
  sbcl --script benchmark/run.lisp

Results

handle-log-record allocates zero bytes per call on every benchmarked wire format and payload. From one run on SBCL 2.6.0 / aarch64-darwin:

Benchmark	ns/call	bytes/call
text-handler, ASCII message + 3 fields	392	0
json-handler, ASCII message + 3 fields	832	0
text-handler, 256-char field value	686	0
json-handler, 256-char field value	1787	0
json-handler, double-float-heavy fields	1227	0

Timings are machine-specific and will move on your hardware. The 0 bytes/call column is the part the test suite actually asserts, in t/performance-test.lisp's allocation-bound specs, so a change that reintroduces per-call consing fails CI instead of regressing quietly.

Comparison with log4cl

benchmark/competitors.lisp runs the same methodology against log4cl, fetched via Quicklisp on first run, writing an equivalent message and fields to a discarding stream.

On that same run cl-log-kit was roughly 1.3x slower in that minimal configuration — 477 ns/call against log4cl's 370 ns/call, both allocation-free.

The difference includes work that the compared log4cl path does not do. On each call cl-log-kit escapes emitted tokens, writes a structured per-field record, and performs the deep field snapshot, canonical-key deduplication, reentrant stream serialization, and resource-limit checks described in Fields and Handlers. log4cl writes one already-formatted message string.