Pin to the node its model weights are actually on

FixKlusterServices
Shipped
August 7, 2026 at 7:30 AM UTC
Author
Kamo
Commit
3fc3491

The model is a hostPath under /var/lib/embedding-model on k1m1, and nothing replicates it. When k3m1 joined, the new ReplicaSet's pod was scheduled there and never started: 37 days in ContainerCreating, 26,617 failed mounts reading "hostPath type check failed ... is not a directory", which is what an absent path looks like to a Directory type check. It went unnoticed because maxUnavailable is 0, so the 130-day-old pod kept serving while its replacement could not start -- the rollout was stuck, not the service. hostPath is node-local storage, so the pod has to say which node it means. A PVC would be the better answer if this ever needs to move, but the weights are a read-only cache that is cheap to re-seed, and pinning is what makes the current layout honest.

All changes

Like what you see shipping?

Every one of these updates lands in your workspace automatically. Start free and watch it grow week after week.

Start Free ForeverView Pricing