Kubernetes¶
One Deployment, one PVC, no Service.
kubectl create secret generic multica-runtime --from-env-file=.env
kubectl apply -f deploy/kubernetes/deployment.yaml
kubectl logs -f deploy/multica-runtime
Why Recreate¶
Two pods sharing a MULTICA_DAEMON_ID compete for tasks. The Deployment uses
strategy: Recreate, so the old pod terminates before the new one starts.
Why no Service¶
The runtime is an outbound worker. It connects to Multica and your AI router and
never accepts inbound traffic. The design image's MCP/plugin ports are reached by
the agent over 127.0.0.1 inside the same pod; expose them only if you have a
specific reason to.
Security context¶
The manifest:
- runs as uid/gid
1000,runAsNonRoot: true; - sets
fsGroup: 1000so the PVC is writable; - drops all capabilities;
- disallows privilege escalation;
- uses
RuntimeDefaultseccomp.
No privileged mode, no hostPath, no Docker socket.
Storage¶
ReadWriteOnce is enough for one replica but pins the pod to a node while the
volume is attached. For rescheduling, use ReadWriteMany (EFS, CephFS, NFS).
See Persistence.
Autoscaling¶
Not supported by the bundled autoscaler (ECS/Swarm only). Run one replica, or
drive a CronJob/KEDA from the pending-task signal. See Autoscaling.
Updating¶
kubectl set image deploy/multica-runtime \
runtime=ghcr.io/proreka/multica-runtime:v0.2.0-base
kubectl rollout status deploy/multica-runtime
Sessions survive because /workspaces is a PVC.