Traza Benchmarks

These values were measured by cargo run --release --bin bench; they are not estimates. The benchmark builds and starts target/release/traza-server on a free loopback port with a fresh temporary data directory.

Results

Metric	Measured	Target	Result
Sustained batched HTTP ingest (durability=wal, compaction fanout=4, max segment bytes=268435456)	65749 spans/s	>= 50,000 spans/s	PASS
Trace-by-id p95	1.280 ms	< 50 ms	PASS
Attribute-filtered query p95	5.443 ms	< 300 ms	PASS

Additional percentiles:

Query	p50	p95	p99	samples
Trace by ID	0.577 ms	1.280 ms	2.094 ms	200
Attribute filter	3.657 ms	5.443 ms	6.468 ms	100

Methodology

• Corpus: 1000000 spans, 100,000 traces with 10 spans each, 20 services, 100 indexed benchmark.group attribute values, and occasional events.
• Ingest: HTTP POST /v1/spans, 1000 spans per request, timed from the first request through the final successful response. JSON generation is intentionally inside the timed loop, so the reported rate includes client serialization and loopback HTTP overhead.
• Trace sampling: 200 deterministic trace IDs spread through the corpus; each response is parsed and checked for 10 spans.
• Filter sampling: 100 deterministic attr.benchmark.group queries with limit=100; each response is verified against the corpus construction — exact span count, the requested group value on every span, and the exact expected span ids — before its latency is recorded. A wrong answer aborts the run without writing this file.
• Percentiles: nearest-rank selection over complete request wall-clock durations measured with std::time::Instant; no warm-up samples are discarded.
• Build: Cargo release profile. Timestamp: Unix 1788687642.
• Machine context: macos/aarch64, 10 available hardware threads.
• Load conditions: 1-minute load average unavailable at the end of the run — ambient desktop load, not an idle host (the house rule is ingest.md’s). An idle rerun will likely improve the tails; the gate tripwires, not these point estimates, are the contract.
• Final server stats: {"buffer_age_seconds":null,"buffered_records":0,"bytes_on_disk":66854329,"durability":"wal","persisted_records":1000000,"record_count":1000000,"segment_count":65,"total_records":1000000,"wal_bytes":0}.

The ingest threshold is PASS. The trace p95 threshold is PASS. The filtered-query p95 threshold is PASS. Any miss remains visible in the table rather than being substituted or estimated.

Verification Notes

• Corpus declaration: 1000000 spans (1,000,000 spans).
• Every reported result is measured by this benchmark run, never estimated.
• Unsuccessful lookups are reported as misses.
• This file exists only because the run passed acceptance gate 6’s tripwires (segment format, as amended to absolute bounds): trace-lookup p50 <= 0.75 ms and attribute-filter p50 <= 6 ms on this canonical corpus. The benchmark asserts them after measuring and before writing; a run that misses either exits non-zero and writes nothing.