Health Check

Health Checks for ML Services

Types of Health Checks
| Check | Purpose | Kubernetes |
|-------|---------|------------|
| Liveness | Is the process alive? | livenessProbe |
| Readiness | Can it accept traffic? | readinessProbe |
| Startup | Has it fully started? | startupProbe |

Implementation

Basic Health Endpoint

from fastapi import FastAPI

app = FastAPI()

@app.get("/health")
def health():
    return {"status": "healthy"}

@app.get("/ready")
def ready():
    # Check dependencies
    if not model_loaded:
        return {"status": "not ready"}, 503
    if not can_connect_to_db():
        return {"status": "not ready"}, 503
    return {"status": "ready"}

Model Health Check

@app.get("/health/model")
async def model_health():
    try:
        # Run inference on test input
        start = time.time()
        result = model.predict(test_input)
        latency = time.time() - start

        return {
            "status": "healthy",
            "model_loaded": True,
            "inference_latency_ms": latency * 1000
        }
    except Exception as e:
        return {"status": "unhealthy", "error": str(e)}, 503

Kubernetes Configuration

apiVersion: apps/v1
kind: Deployment
spec:
  template:
    spec:
      containers:
      - name: llm-server
        livenessProbe:
          httpGet:
            path: /health
            port: 8000
          initialDelaySeconds: 30
          periodSeconds: 10
          failureThreshold: 3

        readinessProbe:
          httpGet:
            path: /ready
            port: 8000
          initialDelaySeconds: 60  # Model loading time
          periodSeconds: 5
          failureThreshold: 2

        startupProbe:
          httpGet:
            path: /ready
            port: 8000
          initialDelaySeconds: 0
          periodSeconds: 10
          failureThreshold: 30  # 5 minutes to start

ML-Specific Considerations
| Check | What to Verify |
|-------|----------------|
| Model loaded | Model weights in memory |
| GPU available | CUDA device accessible |
| Warm-up complete | First inference done |
| Dependencies | Vector DB, Redis connected |

Deep Health Check

@app.get("/health/deep")
async def deep_health():
    checks = {
        "model": await check_model_health(),
        "gpu": await check_gpu_health(),
        "vector_db": await check_vector_db(),
        "cache": await check_redis(),
    }

    all_healthy = all(c["healthy"] for c in checks.values())
    status_code = 200 if all_healthy else 503

    return checks, status_code

Best Practices
- Keep liveness probes simple and fast
- Use readiness to control traffic
- Set appropriate timeouts
- Log health check failures
- Monitor health endpoints