Skip to content

why-init-container.md

Latest commit

 

History

History
67 lines (50 loc) · 2.9 KB

why-init-container.md

File metadata and controls

67 lines (50 loc) · 2.9 KB

Why Init Container for Supervisor Delivery

The upstream approach

The Rust K8s driver in NVIDIA/OpenShell uses a hostPath volume to mount the supervisor binary from the node filesystem:

volumes:
  - name: supervisor
    hostPath:
      path: /opt/openshell/bin
      type: Directory

This requires the supervisor binary to be pre-staged on every node, typically by a DaemonSet or by baking it into the node image. The pods mount it read-only.

Why we diverge

OpenShift's restricted-v2 SCC (the default for workloads) blocks hostPath volume mounts. Using hostPath requires either:

  • privileged SCC (too broad, violates least-privilege)
  • A custom SCC that allows hostPath but restricts everything else (operational overhead per cluster)
  • Dropping to anyuid or hostaccess SCCs (security regression)

Additionally, a DaemonSet to pre-stage the binary on every node requires its own elevated SCC, RBAC, and lifecycle management.

Our approach

We use an init container that copies the supervisor binary from a container image into an emptyDir volume shared with the agent container:

initContainers:
  - name: supervisor-init
    image: ghcr.io/nvidia/openshell-community/supervisor:latest
    command: ["cp", "/usr/local/bin/openshell-sandbox", "/opt/openshell/bin/"]
    volumeMounts:
      - name: supervisor-bin
        mountPath: /opt/openshell/bin

containers:
  - name: agent
    command: ["/opt/openshell/bin/openshell-sandbox"]
    volumeMounts:
      - name: supervisor-bin
        mountPath: /opt/openshell/bin
        readOnly: true

volumes:
  - name: supervisor-bin
    emptyDir: {}

Trade-offs

hostPath (upstream) Init container (ours)
SCC requirement Needs hostPath access (custom or privileged SCC) Works without hostPath
Node pre-staging Required (DaemonSet or baked into node image) Not required
Cold start cost None (binary already on node) One cp command (~15MB, <1 second)
Image pull None (binary on node filesystem) One pull per node (cached after first)
Supervisor version Tied to what's on the node Tied to init container image tag
BYOC compatibility Works with any agent image Works with any agent image
Operational overhead DaemonSet lifecycle, node affinity, SCC exceptions None beyond the init container

The init container approach trades a negligible cold start cost (copying one binary) for significantly simpler operations on OpenShift. No DaemonSet, no node access, no SCC exceptions for the delivery mechanism.

Note: the agent container itself still requires elevated capabilities (SYS_ADMIN, NET_ADMIN, SYS_PTRACE, SYSLOG) for the supervisor to create network namespaces and install Landlock/seccomp. This requires a custom SCC regardless of the supervisor delivery method. The init container approach eliminates the additional SCC exception that hostPath would require on top of the capabilities.