Performance Tuning Guide

// Planning phase — fast, no side effects
const plan = await engine.evaluate(request, { ...ctx, mode: 'PLAN_ONLY' });

// Execution phase — full pipeline with evidence
if (userConfirmed) {
  const decision = await engine.evaluate(request, ctx);
}

// ❌ Registering everything — slower
const engine = new TealEngine({
  modules: [TealClassifier(), TealDrift(), TealState(), TealTemporal(), TealMonitor(), ...],
});

// ✅ Register only what your workload needs
const engine = new TealEngine({
  modules: [TealClassifier(), TealMonitor()],  // 2 modules = faster
});

// Fast path — regex only, no ONNX loading
classifier: { ensemble_mode: 'regex_only' }

// Full path — ML + regex (adds ~3ms)
classifier: { ensemble_mode: 'ensemble_union', model_path: './model.onnx' }

// High-throughput: smaller window, faster comparison
drift: { baseline_window: '1d', min_samples: 50 }

// Low-throughput: larger window for accuracy
drift: { baseline_window: '7d', min_samples: 100 }

// No ONNX model loaded — saves 20-50 MB
classifier: { ensemble_mode: 'regex_only' }

drift: {
  max_tracked_agents: 100,  // Evict oldest baselines
  storage: 'memory',        // Or 'redis' for shared state
}

response_hooks: [{
  id: 'siem-export',
  trigger: 'on_deny',  // Only fire on denials (not every request)
  dedup_window_ms: 60000,  // Suppress duplicates within 1 minute
  rate_limit: { max_per_minute: 100 },  // Prevent alert storms
}]

# Increase worker threads for high concurrency
TEALTIGER_WORKERS: 4

# Increase body size limit if needed
TEALTIGER_MAX_BODY_BYTES: 2097152  # 2 MB

# Enable keep-alive for connection reuse
TEALTIGER_KEEP_ALIVE: true

npx tealtiger benchmark --modules classifier,drift,monitor --iterations 10000

TealTiger v1.3.0 Benchmark Results
───────────────────────────────────
Modules: classifier (regex_only), drift, monitor
Iterations: 10,000

Pre-evaluation:  p50=0.3ms  p95=0.8ms  p99=1.2ms
Module dispatch:  p50=3.1ms  p95=7.2ms  p99=11.4ms
Post-hooks:      p50=0.5ms  p95=1.8ms  p99=3.1ms
Total:           p50=4.2ms  p95=9.8ms  p99=15.7ms