Metrics

This lab ranks in-process queues, not encodings. Times are nanoseconds. Throughput is derived: 1e9 / time_ns. Payload size (256 B or 4 KiB) is the sample, not a library score. Never call it a fixture.

What we rank

Field	Meaning	Why
TimeHandoff / handoff_median_ns	Enqueue + dequeue wall time	Primary rank. Producer-only put/s can lie.
TimeEnq / enq_median_ns	Produce cost	Split of the handoff
TimeDeq / deq_median_ns	Consume cost	Split of the handoff
avg_ops_per_sec	1e9 / mean(TimeHandoff)	Same clock, easier to read
handoff_p95_ns / p99 / p999	Tail of the same handoff series	A lock that is fast on the median and stalls on p99 is not “fast”
msgs_per_cpu_sec	n / mean(CpuTimeNs)	Spin can win latency and burn cores. Null if CpuTimeNs is missing.

Do not invent msgs_per_cpu_sec from wall time.

What we do not rank

Field	Why it exists	What to do with it
Size / SizeGzip / SizeZstd / median_size_bytes	Serializer leftovers. Payload bytes are the data type, not a result.	Removed. New CSVs and stats do not write these. Old logs may still contain Size; analysis ignores it.
mean_fidelity	FidelityScore = fraction of items that arrived in order.	Gate, not a score. If < 1, the row is invalid for ranking. Typical value is 1.0.
mean_memory_peak_bytes	Process RSS (getrusage / PeakWorkingSet64 / process.memoryUsage().rss).	Provenance. Usually the process, not the queue. Do not run a “most compact” contest on it.
StreamMode / honesty	Serializer “native vs adapted stream I/O”. There is no stream I/O in this suite. Pattern=stream means 2P2C.	Do not show an honesty column.
OpPerSecEnq / Deq / Handoff in the CSV	Convenience copies of 1e9 / Time*.	Analysis recomputes from times. Do not treat CSV ops as a second clock.

Identity (not metrics)

Language, LibraryName, LibraryVersion, Pattern, TestDataName, DataTypeInstanceCount, TypeConfigHash, Repetitions, RepetitionIndex, RunOrder, SchedulePosition, NativeKind (implementation family label).

Pattern: bytes = 1P1C, stream = 2P2C (old logs / experiment 3), plus named 1p4c / 4p1c / 4p4c. The published matrix is 1P1C and 4P4C.

Derived stats (not CSV columns)

Percentiles, std, MAD, CV, bootstrap CI, Cliff’s δ vs the fastest peer — see ANALYSIS_METHODOLOGY. runs / outliers_removed are provenance.

Missing on purpose

Tempting metric	Why it is not a warehouse column
Empty-queue wakeup	Experiment 5. Needs a different loop than “enqueue all, dequeue all”.
Occupancy / depth	Not observed by the harness.
Fairness under steal	Experiment-level, not a single number.
Encoded size / gzip	Not a queue.
Cross-language rank	Different runtimes. Directional only.

CSV ABI

New writes keep the existing header so historical logs still parse. The parser still accepts leftover serializer-benchmark names (SerializerName, TimeSer, StringOrStream, …).

Meaning of the timed columns:

CSV column	Queue meaning
TimeEnq	Enqueue ns
TimeDeq	Dequeue ns
TimeHandoff	Handoff ns
CpuTimeNs	Optional process CPU time for that repetition
FidelityScore	1.0 = every item arrived in order
MemoryPeakBytes	Peak RSS when the runner can measure it