kserve/docs/samples/rollout-strategy/rollout-strategy-example.yaml

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: rollout-strategy-example
  namespace: default
  annotations:
    serving.kserve.io/deploymentMode: "Standard"
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: "s3://my-bucket/model"
    # Example 1: Availability Mode - Direct deployment strategy for high availability
    # Configuration: maxUnavailable = 0, maxSurge = desired value
    # Behavior: New pods are created first, then old pods are terminated (zero downtime)
    deploymentStrategy:
      type: RollingUpdate
      rollingUpdate:
        maxUnavailable: "0"      # No pods unavailable during rollout
        maxSurge: "50%"          # Can create 50% more pods during rollout
    
  transformer:
    custom:
      container:
        image: my-transformer:latest
        env:
          - name: MODEL_NAME
            value: "my-model"
    # Example 2: ResourceAware Mode - Resource-efficient deployment strategy  
    # Configuration: maxSurge = 0, maxUnavailable = desired value
    # Behavior: Old pods are terminated first, then new pods are created (resource efficient)
    deploymentStrategy:
      type: RollingUpdate
      rollingUpdate:
        maxSurge: "0"            # No extra pods during rollout
        maxUnavailable: "25%"    # Up to 25% of pods can be unavailable

---
# Example 3: Using ConfigMap defaults (no deploymentStrategy specified)
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: configmap-defaults-example
  namespace: default
  annotations:
    serving.kserve.io/deploymentMode: "Standard"
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: "s3://my-bucket/model"
    # No deploymentStrategy specified - will use ConfigMap global defaults  
    # when defaultDeploymentMode is "Standard"
    # Allows administrators to set organization-wide rollout policies

---
# Example 4: Multinode deployment (Ray workload)
# Note: KServe will automatically override ANY rollout strategy to:
# maxUnavailable: "0%", maxSurge: "100%" for multinode deployments
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: multinode-example
  namespace: default
  annotations:
    serving.kserve.io/deploymentMode: "Standard"
spec:
  predictor:
    model:
      modelFormat:
        name: huggingface
      storageUri: "s3://my-bucket/llm-model"
    containers:
    - name: kserve-container
      image: my-ray-model-server:latest
      env:
      - name: RAY_NODE_COUNT    # This triggers multinode deployment
        value: "4"              # 1 head + 3 worker nodes
      - name: REQUEST_GPU_COUNT
        value: "8"
    # Even if you specify a different rollout strategy, KServe will override it
    # for multinode deployments to ensure Ray cluster stability
    deploymentStrategy:
      type: RollingUpdate
      rollingUpdate:
        maxUnavailable: "50%"  # This will be overridden to "0%"
        maxSurge: "25%"        # This will be overridden to "100%"