Skip to content

Benchmarks

Goal: one billion events processed end-to-end in a single sustained run (ADR 0002).

Method (baseline, refined in Phase 6)

Aspect Rule
Counting Producer counts acknowledged events; processor and sink count per partition
Correctness Reconciliation: produced = processed = stored, no duplicates, no gaps
Metrics Throughput (msg/s, MB/s), end-to-end latency p50/p99/p999, consumer lag, CPU/memory/disk
Environment Hardware, profile, versions and configuration recorded with each result
Caveat kind shares one host's disk and kernel; results are indicative, not production numbers

Results

Date Profile Events Duration Avg throughput p99 latency Report
none yet