# Performance Hardening

This folder contains latency profiling baselines and guardrail thresholds used in CI.

## Scenarios

The profiler covers representative load patterns:

- `book_update_burst`: rapid market-data deltas with moderate detection load.
- `execution_spike`: heavier detection/execution pressure.
- `reconnect_storm`: frequent reconnect/reset behavior.

## Profiling Commands

Generate a fresh profile:

```powershell
python scripts/profile_latency.py --iterations 600 --output ops/performance/latency_baseline.json
```

Check current performance against the baseline and thresholds:

```powershell
python scripts/check_latency_regression.py \
  --baseline ops/performance/latency_baseline.json \
  --thresholds ops/performance/latency_thresholds.json \
  --iterations 600
```

CI executes the same guardrail check.

## Baseline Snapshot (2026-06-01)

Key end-to-end latency baselines from `latency_baseline.json`:

- `book_update_burst`: p95 = 0.0132 ms, p99 = 0.0198 ms
- `execution_spike`: p95 = 0.0139 ms, p99 = 0.0177 ms
- `reconnect_storm`: p95 = 0.0114 ms, p99 = 0.0134 ms

## Optimization Note

`MetricsCalculator.compute()` uses PostgreSQL SQL aggregations and percentiles, reducing Python-side row scans.

Measured benchmark (`scripts/benchmark_metrics_compute.py`):

- Python scan baseline: 12.623 ms
- SQL aggregate implementation: 11.039 ms
- Speedup: 1.14x