Shadow Deployment

Shadow Deployment im Überblick

Shadow Deployment lässt eine neue Version parallel zur Produktion laufen. Sie erhält echten Traffic, aber ihre Antworten werden verworfen – Nutzer merken nichts davon.

Der Ablauf:

Nutzer-Request
      │
      ├──────────────────┐
      ▼                  ▼
┌──────────────┐   ┌──────────────┐
│ Produktion   │   │   Shadow     │
│ (v1.0)       │   │   (v2.0)     │
└──────┬───────┘   └──────┬───────┘
       │                  │
       ▼                  ▼
   Response            Logging
   an Nutzer           (verworfen)

Vergleich Deployment-Strategien:

Technisch betrachtet

Traffic Mirroring mit Istio

apiVersion: networking.istio.io/v1beta1
kind: VirtualService
metadata:
  name: my-service
spec:
  hosts:
    - my-service
  http:
    - route:
        - destination:
            host: my-service
            subset: v1
          weight: 100
      mirror:
        host: my-service
        subset: v2
      mirrorPercentage:
        value: 100.0  # 100% des Traffics spiegeln

Nginx Traffic Mirroring

location /api {
    # Primärer Backend
    proxy_pass http://production;
    
    # Mirror zu Shadow
    mirror /mirror;
    mirror_request_body on;
}

location = /mirror {
    internal;
    proxy_pass http://shadow$request_uri;
    
    # Antwort verwerfen
    proxy_ignore_client_abort on;
}

ML-Modell Shadow Testing

class ShadowModelService:
    def __init__(self, production_model, shadow_model):
        self.production = production_model
        self.shadow = shadow_model
        self.logger = MetricsLogger()
    
    async def predict(self, input_data):
        # Production: Synchron, Antwort an Nutzer
        prod_result = await self.production.predict(input_data)
        
        # Shadow: Asynchron, nur Logging
        asyncio.create_task(
            self._shadow_predict(input_data, prod_result)
        )
        
        return prod_result
    
    async def _shadow_predict(self, input_data, prod_result):
        try:
            shadow_result = await self.shadow.predict(input_data)
            
            # Vergleich loggen
            self.logger.log({
                "input": input_data,
                "production": prod_result,
                "shadow": shadow_result,
                "match": prod_result == shadow_result,
                "shadow_latency": shadow_result.latency
            })
        except Exception as e:
            self.logger.log_error(e)

Metriken vergleichen

def analyze_shadow_results(logs):
    metrics = {
        "total_requests": len(logs),
        "match_rate": sum(l["match"] for l in logs) / len(logs),
        "shadow_avg_latency": mean(l["shadow_latency"] for l in logs),
        "prod_avg_latency": mean(l["prod_latency"] for l in logs),
        "shadow_errors": sum(1 for l in logs if l.get("shadow_error")),
    }
    
    # Alerts
    if metrics["match_rate"] < 0.95:
        alert("Shadow model divergiert von Production!")
    
    if metrics["shadow_avg_latency"] > metrics["prod_avg_latency"] * 1.5:
        alert("Shadow model ist 50% langsamer!")
    
    return metrics

Wann Shadow Deployment nutzen?