Benchmark Methodology

Astra benchmarks are built around control-plane realism, not synthetic throughput alone.

Ground rules

• compare Astra and etcd under the same workload shape,
• capture machine-readable output for every public benchmark claim,
• keep correctness gates alongside latency and throughput metrics,
• inspect stage telemetry rather than only aggregate p99.

What we measure

• write and read tail latency
• LIST behavior at large cardinality
• watch fanout behavior and lag
• queue wait, quorum ack, and apply timing
• CPU iowait, memory, and disk behavior
• migration correctness and cutover stability

How to read public benchmark claims

Astra benchmark claims in this repo are anchored to:

• tracked benchmark datasets in docs/research/data/
• generated public summaries such as BENCHMARKS.md
• specific validation harnesses under refs/scripts/validation/

That means the public docs can be regenerated when the underlying benchmark dataset changes.