Kubernetes (Helm)

Deploy Boltbase to Kubernetes using the official Helm chart. The chart deploys a scheduler, coordinator, Web UI, and configurable worker pools.

Installation

helm repo add boltbase https://dagu-org.github.io/dagu
helm install boltbase boltbase/boltbase

Worker Pools

The workerPools section in values.yaml defines groups of worker Deployments. Each pool creates a separate Kubernetes Deployment named <release>-boltbase-worker-<poolName>.

Default Configuration

workerPools:
  general:
    replicas: 2
    labels: {}
    resources:
      requests:
        memory: "128Mi"
        cpu: "100m"
      limits:
        memory: "256Mi"
        cpu: "200m"
    nodeSelector: {}
    tolerations: []
    affinity: {}

The default general pool has labels: {}, which means it matches DAGs without a worker_selector (an empty selector matches any worker). See Worker Labels for how matching works.

Multi-Pool Example

workerPools:
  general:
    replicas: 3
    labels: {}
    resources:
      requests:
        memory: "256Mi"
        cpu: "250m"
      limits:
        memory: "512Mi"
        cpu: "500m"

  gpu:
    replicas: 2
    labels:
      gpu: "true"
      cuda: "12.0"
    resources:
      requests:
        memory: "1Gi"
        cpu: "500m"
        nvidia.com/gpu: "1"
      limits:
        memory: "4Gi"
        cpu: "2"
        nvidia.com/gpu: "1"
    nodeSelector:
      nvidia.com/gpu.present: "true"
    tolerations:
      - key: nvidia.com/gpu
        operator: Exists
        effect: NoSchedule

This creates two Deployments:

• <release>-boltbase-worker-general — 3 replicas, no labels, matches any DAG without worker_selector
• <release>-boltbase-worker-gpu — 2 replicas, labels gpu=true,cuda=12.0, scheduled on GPU nodes

DAGs with worker_selector: {gpu: "true"} are dispatched to the gpu pool. DAGs without worker_selector are dispatched to any pool (including general).

How Pool Labels Map to Worker Labels

Each pool's labels map is converted to a --worker.labels CLI argument via the boltbase.workerLabels template helper. For example:

labels:
  gpu: "true"
  cuda: "12.0"

becomes the CLI argument --worker.labels gpu=true,cuda=12.0, which is equivalent to running:

boltbase worker --worker.labels gpu=true,cuda=12.0

These labels are then matched against worker_selector in DAG definitions. See Distributed Execution — How Dispatch Decisions Work for the full dispatch logic.

Pool Name Validation

Pool names must match ^[a-z][a-z0-9-]*$ — lowercase letters, digits, and hyphens, starting with a letter. Invalid names cause a Helm template error.

Valid: general, gpu-workers, us-east-1 Invalid: GPU, 1-workers, my_pool

Per-Pool Kubernetes Scheduling

Each pool supports standard Kubernetes scheduling fields:

workerPools:
  gpu:
    replicas: 2
    labels:
      gpu: "true"
    nodeSelector:
      nvidia.com/gpu.present: "true"
    tolerations:
      - key: nvidia.com/gpu
        operator: Exists
        effect: NoSchedule
    affinity:
      nodeAffinity:
        requiredDuringSchedulingIgnoredDuringExecution:
          nodeSelectorTerms:
            - matchExpressions:
                - key: nvidia.com/gpu.product
                  operator: In
                  values: ["NVIDIA-A100-SXM4-40GB"]

Persistence

The chart requires ReadWriteMany (RWX) storage for shared state:

persistence:
  enabled: true
  accessMode: ReadWriteMany
  size: 10Gi
  storageClass: "nfs-client"  # Your RWX-capable storage class

All worker pods mount this PVC at /data.

Local Testing

For local testing with Kind or Docker Desktop (which lack RWX storage), skip PVC validation:

helm install boltbase boltbase/boltbase \
  --set persistence.skipValidation=true \
  --set workerPools.general.replicas=1

Other Chart Components

Scheduler

scheduler:
  replicas: 1
  resources:
    requests:
      memory: "256Mi"
      cpu: "250m"

Coordinator

coordinator:
  replicas: 1
  resources:
    requests:
      memory: "128Mi"
      cpu: "100m"

Web UI

ui:
  replicas: 1
  service:
    port: 8080
  resources:
    requests:
      memory: "256Mi"
      cpu: "250m"

Authentication

auth:
  mode: "builtin"  # "none", "builtin", or "oidc"
  builtin:
    admin:
      username: "admin"
      password: "changeme"
    token:
      secret: "changeme"
      ttl: "24h"

Image

image:
  repository: ghcr.io/dagu-org/boltbase
  tag: latest
  pull_policy: IfNotPresent