Research: production compression replay and published evidence

[awksedgreep]

Parent: #28. Depends on the harness, sweep, and candidate profiles.

Summary

Validate candidate profiles under production-shaped replay and publish reproducible evidence suitable for operator guidance and technical articles.

Replay scenarios

• Sustained, burst, trickle, and backfill ingestion.
• Concurrent narrow/wide queries and field discovery.
• Quiet periods available for settling and continuous load without idle time.
• Retention, WAL checkpoints, backups, restart, SIGTERM, and kill-9 during optimization.
• CPU throttling, constrained memory, slow storage, and supported ARM/x86 classes.

Published evidence

• Corpus/generator identity and privacy methodology.
• Exact hardware, CPU features, memory, thread count, commit, toolchain, and commands.
• Foreground latency/throughput, settled ratio, time to settle, optimizer CPU/memory/write amplification, query effects, and operational file accounting.
• Raw machine-readable measurements plus a human-readable comparison.
• Clear statements of tradeoffs and workloads where another profile wins.

Acceptance criteria

• Every advertised profile claim is reproducible from public or generated inputs.
• Reports distinguish logical raw bytes, data payload, index, WAL, and whole-file size.
• Regression thresholds protect the selected profiles in CI or a documented periodic benchmark gate.
• Documentation and blog-ready material are generated from the same evidence rather than manually copied figures.