Zero-Dependency Benchmark & Stress Testing (perf-benchmark)

This skill guides the agent to design fast, self-contained benchmark scripts that prove performance with real numbers—without pulling in heavy benchmark frameworks or producing fabricated metrics.

1. The Core Simulation Architecture

Every benchmark script should follow a standalone 4-phase lifecycle (exemplified by demo.py):

1. Setup Phase: Generate synthetic data in memory or fast temporary storage (e.g. SQLite / in-memory structures) in clean batches of 10,000 to avoid memory spikes.
2. Drift/Fault Injection: Inject realistic real-world discrepancies:
   • 1 dropped record (simulating network timeout during ETL sync)
   • 1 phantom/duplicate record (simulating at-least-once delivery duplicates)
   • 1 mutated field (simulating floating-point or rounding precision mismatch)
3. Execution & Measurement: Measure the core operation using time.perf_counter():

   t0 = time.perf_counter()
   result = engine.run(...)
   elapsed = time.perf_counter() - t0
   throughput = int(total_items / elapsed) if elapsed > 0 else 0

4. Cleanup: Always clean up temporary files in a finally: block.

2. Standard Benchmark Metrics to Report

Always report concrete, human-readable numbers:

• Total Volume: e.g. 50,000 items
• Execution Time: e.g. 0.23 seconds
• Throughput: e.g. ~215,000 rows/sec
• Discrepancy Precision: 100% detection rate with zero false positives.

3. Markdown Benchmark Summary Card

Output results in a compact format ready for instant copy-pasting into README.md:

### Benchmark Performa

| Metrik | Hasil Pengujian |
|---|---|
| **Volume Data** | 50.000 transaksi |
| **Waktu Eksekusi** | 0.23 detik |
| **Throughput** | ~215.000 baris/detik |
| **Akurasi Deteksi** | 3/3 selisih terdeteksi (100%) |