# Bench Prep — PREP-BENCH-IMPACT-16-001 (ImpactIndex dataset/replay)

Status: **Ready for execution** (2025-12-11)
Owners: Bench Guild · Scheduler Team
Scope: Provide deterministic dataset + replay plan for ImpactIndex throughput benchmark (resolve 10k productKeys; measure latency/throughput/memory).

## Inputs/dataset
- Snapshot file: `docs/samples/impactindex/products-10k.ndjson` (10,000 productKeys, shuffled once with seed `2025-01-01T00:00:00Z`).
- SHA256: `caa79c83b5a9affc3b9cc4e54a516281ddceff4804ce853fee3b62d7afb7ab69` (`products-10k.ndjson.sha256` included).
- Each line: `{ "productKey": "pkg:<ecosystem>/<name>@<version>", "tenant": "bench" }`.

## Benchmark procedure
- Harness location: `src/Bench/StellaOps.Bench/ImpactIndex/impact_index_bench.py`.
- Warmup: 1k lookups (excluded from metrics) to trigger caches.
- Run: process all 10k productKeys twice (cold, warm). Record per-pass statistics.
- Metrics to capture (per pass):
  - `throughput_items_per_sec`, `p95_ms`, `p99_ms`, `max_ms` for lookups.
  - `rss_mb`, `managed_mb`, `gc_gen2_count` from .NET counters.
  - `cache_hit_rate` if cache present.
- Output format: NDJSON; one object per pass with fields `{ pass: "cold"|"warm", startedAtUtc, durationMs, throughput, p95Ms, p99Ms, maxMs, rssMb, managedMb, gcGen2, cacheHitRate }`.
- Determinism: fixed seed, single-threaded option flag `--threads 1` for reproducibility; timestamps in UTC ISO-8601.

## Acceptance criteria
- Dataset and checksum published; harness reads from local sample path (no network). ?
- Benchmark run produces deterministic NDJSON for given seed and hardware profile; differences limited to ?5%.
- Cold vs warm pass metrics logged; throughput target ? 2k items/sec on reference hardware, p95 ? 25 ms.

## Next steps
- Harness command: `python src/Bench/StellaOps.Bench/ImpactIndex/impact_index_bench.py --input docs/samples/impactindex/products-10k.ndjson --output src/Bench/StellaOps.Bench/ImpactIndex/results/impactindex.ndjson --threads 1 --seed 20250101`.
- Surface metrics to perf dashboard once harness lands; otherwise store under `out/bench/impactindex/` with hashes (`results/impactindex.ndjson.sha256` present).