Deploy with containers¶

Build an image from a reviewed commit and select the deployment profile first. These procedures establish an evaluation deployment; production sizing and failure testing remain your release gates.

Build an image¶

From the repository root:

INFERPLANE_IMAGE_TAG="$(git rev-parse --short=12 HEAD)"
docker build -t "inferplane:$INFERPLANE_IMAGE_TAG" .
docker build -f Dockerfile.inferplaned \
  -t "inferplaned:$INFERPLANE_IMAGE_TAG" .

The two Dockerfiles build static Go binaries into non-root distroless images. There is no release-image publication workflow in this repository; push reviewed images to your own registry and pin the resolved digest for deployment.

Run one local container¶

Start from the main example. It uses /var/lib/inferplane for state. Remove unused providers and routes, supply the configured secrets through your runtime secret mechanism, and verify pricing before sending traffic.

The image runs as UID/GID 65532. Prepare a new dedicated state directory writable by that UID, or provision a volume with equivalent ownership. Do not share a live audit WAL with another container.

mkdir -p .local/container-state
sudo chown 65532:65532 .local/container-state
sudo chmod 700 .local/container-state

# Export only the credentials referenced by your reviewed config beforehand.
docker run --rm --name inferplane \
  -p 127.0.0.1:8080:8080 -p 127.0.0.1:9090:9090 \
  --env ANTHROPIC_API_KEY --env INFERPLANE_ADMIN_TOKEN \
  --mount type=bind,src="$PWD/examples/config.json",dst=/etc/inferplane/config.json,readonly \
  --mount type=bind,src="$PWD/.local/container-state",dst=/var/lib/inferplane \
  "inferplane:$INFERPLANE_IMAGE_TAG"

This command assumes your edited config uses only the exported provider credentials. If a Bedrock route remains enabled, explicitly supply its intended identity instead of relying on an accidental default credential chain. Bind mounts on SELinux hosts may need an appropriate container label.

Issue keys through the private admin API/console, or use docker exec with mayu keys create --config /etc/inferplane/config.json. The container has no shell; invoke the binary directly.

Deploy the Helm chart¶

The chart expects an existing Secret; it never creates provider credentials. Your secret controller should provision inferplane-secrets in the selected namespace. Its default server.admin_auth.token_refs reads INFERPLANE_ADMIN_TOKEN from that Secret. Configure an alternative authentication block explicitly if needed.

Prepare a values file with your registry/digest-compatible image selection, reviewed config, resources, persistence, and network exposure. For a local-profile evaluation:

replicaCount: 1
image:
  repository: registry.example.invalid/platform/inferplane
  tag: reviewed-commit
persistence:
  enabled: true
  size: 10Gi
secrets:
  existingSecret: inferplane-secrets
resources:
  requests:
    cpu: 100m
    memory: 256Mi
  limits:
    memory: 1Gi
ingress:
  enabled: false

The resource values are an initial measurement point, not a sizing guarantee. Supply the config block as well; chart defaults are not a production policy.

helm lint charts/inferplane -f pilot-values.yaml
helm template inferplane charts/inferplane \
  --namespace inferplane -f pilot-values.yaml > rendered.yaml
# Review the rendered configuration, mounts, secret references and ports.
helm upgrade --install inferplane charts/inferplane \
  --namespace inferplane --create-namespace -f pilot-values.yaml

Local mode allows one replica. Persistent local mode uses a Recreate deployment, so plan downtime during replacement. The pod uses UID/GID and fsGroup 65532; confirm that the storage driver or pre-provisioned claim permits group writes. The gateway startup probe allows ten minutes for audit/analytics replay before liveness begins; tune it for your history. Main config changes roll pods through a checksum annotation, while separately mounted policy changes remain watched.

For multiple replicas, use the explicit shared profile and shared Helm example. It needs the same authority database/schema, initialized control planes, unique gateway identities, and enough schedulable nodes for anti-affinity. Persistent shared mode uses separate PVCs; it does not make Postgres highly available.

Before admitting users¶

Confirm readiness, denied-model behavior, pricing coverage, audit writes, secret rotation, and configured refusal on backend failure. Keep the admin port and metrics private; terminate remote traffic with reviewed TLS/network controls. See security, monitoring, and recovery.