SpicepodCluster

A SpicepodCluster (spice.ai/v2) deploys a distributed query cluster with dedicated scheduler and executor nodes. The operator automatically manages mTLS certificate provisioning, child SpicepodSet resources, and cluster topology.

Architecture

             ┌─────────────────────┐
             │    Load Balancer    │
             └─────────────────────┘
                        │
     ┌──────────────────┼──────────────────┐
     ▼                  ▼                  ▼
┌──────────┐     ┌──────────┐     ┌──────────┐
│Scheduler │     │Scheduler │     │Scheduler │◄──► Object Store (S3)
└──────────┘     └──────────┘     └──────────┘
     ▲                  ▲                  ▲
     │    (executor-initiated connections) │
     │                  │                  │
┌──────────┐     ┌──────────┐     ┌──────────┐
│ Executor │     │ Executor │     │ Executor │────► Object Store (shuffle)
└──────────┘     └──────────┘     └──────────┘

Schedulers coordinate query planning and partition assignment; executors perform the compute work. Executors initiate connections to all schedulers and are shared across them — partition ownership is committed to a shared object store and pushed to executors via the cluster ControlStream. See Distributed Query for the full execution model.

Example

apiVersion: spice.ai/v2
kind: SpicepodCluster
metadata:
  name: my-cluster
  namespace: default
spec:
  schedulerSpec:
    replicas: 1
    spicepod:
      version: v1
      kind: Spicepod
      name: my-cluster-scheduler
    resources:
      requests:
        cpu: 100m
        memory: 256Mi
      limits:
        cpu: 500m
        memory: 512Mi
  executorSpec:
    replicas: 3
    resources:
      requests:
        cpu: 200m
        memory: 512Mi
      limits:
        cpu: "1"
        memory: 2Gi

Executors pull their Spicepod configuration from the scheduler, so no spicepod field is needed on executorSpec.

Multi-Replica Schedulers

For high availability, deploy multiple schedulers:

spec:
  schedulerSpec:
    replicas: 2

Each scheduler independently accepts and distributes work. Shared state is maintained in an S3-compatible object store.

Automatic mTLS

The operator automatically:

1. Generates a self-signed root CA certificate.
2. Issues per-node leaf certificates with appropriate SANs.
3. Configures mTLS between all cluster nodes.
4. Stores certificates in Kubernetes Secrets.
5. Tracks certificate expiry via Prometheus metrics.

Development Mode

For development and testing, mTLS can be disabled:

spec:
  allowInsecureConnections: true

Port Separation

The internal port (50052) carries cluster coordination traffic. See Distributed Query → Internal gRPC for the full RPC surface, including:

• GetAppDefinition — Executors fetch the full Spicepod definition (datasets, catalogs, views, UDFs) from the scheduler.
• ExpandSecret — Executors request secret values from the scheduler's secret store.
• GetSchedulers / AllocateInitialPartitions — Executors fetch scheduler membership and their assigned partitions at startup.
• ControlStream — Bidirectional channel carrying executor heartbeats and UpdatePartitions / RefreshDataset / CancelTasks commands.
• GetTaskHistory / GetMetrics — Federated runtime.task_history and on-demand metrics fan-out across the cluster.

Verification

# Check cluster status
kubectl get spicepodcluster my-cluster

# List child SpicepodSets
kubectl get spicepodset

# Check pods by role
kubectl get pods -l spice.ai/cluster=my-cluster
kubectl get pods -l spice.ai/cluster-role=scheduler
kubectl get pods -l spice.ai/cluster-role=executor

Configuration Inheritance

SpicepodCluster creates child SpicepodSet resources for schedulers and executors. Both schedulerSpec and executorSpec accept the common SpicepodSet spec fields: image, http, flight, metrics, replicas, resources, env, envFromSource, network, nodeAffinity, tolerations, volumeClaimTemplates, volumeMounts, serviceAccount, annotations, labels, updateStrategy, terminationGracePeriodSeconds, and a per-node cluster override.

Notable differences from a standalone SpicepodSet:

• executorSpec does not accept spicepod or probes — executors fetch the Spicepod definition from the scheduler at startup via GetAppDefinition and run without the HTTP server that probes target.
• The service toggle is not available on cluster node specs; the operator manages the headless Services required for mTLS and scheduler/executor discovery.

Per-node cluster overrides

The cluster field on schedulerSpec / executorSpec is a small subset (NodeClusterConfig) used to override cluster-internal addresses; the operator otherwise auto-populates cluster identity, role, mTLS, and scheduler discovery:

spec:
  schedulerSpec:
    cluster:
      bindAddress: "0.0.0.0:50052"

Status

kubectl get spicepodcluster my-cluster -o yaml