Summary
A consumer's instances were placed in a region with no Unikraft runtime, and they have sat unschedulable for four days.
The instances were accepted, so the consumer sees objects that exist and never start. Nothing tells them why, and nothing tells us either.
One production cluster runs the runtime and sixteen do not.
What success looks like
What is happening
Read from production on 2026-09-14 at 19:00Z.
Three instances in one project on us-west-1-alice are Pending, and a fourth in another project on us-east-1-alice is in the same state. All four have been unschedulable since about 11 September.
Their pod requires a node labelled unikraft.com/virtual-kubelet=true and tolerates the virtual-kubelet taint. The scheduler answers plainly:
PodScheduled=False Unschedulable
0/5 nodes are available: 2 node(s) didn't match Pod's node affinity/selector,
3 node(s) had untolerated taint(s).
us-west-1-alice has five nodes: three control-plane and two plain workers. No kraftlet node exists there, and none carries the label the pod requires. The same holds on us-east-1-alice.
Across the whole production fleet only us-central-1-lab has kraftlet nodes. More clusters are expected to gain them, so the set of valid regions will change rather than stay fixed.
The infrastructure side of this, making the runtime's presence readable outside Flux, is https://github.com/datum-cloud/infra/issues/5382.
Related to #278
Summary
A consumer's instances were placed in a region with no Unikraft runtime, and they have sat unschedulable for four days.
The instances were accepted, so the consumer sees objects that exist and never start. Nothing tells them why, and nothing tells us either.
One production cluster runs the runtime and sixteen do not.
What success looks like
What is happening
Read from production on 2026-09-14 at 19:00Z.
Three instances in one project on
us-west-1-alicearePending, and a fourth in another project onus-east-1-aliceis in the same state. All four have been unschedulable since about 11 September.Their pod requires a node labelled
unikraft.com/virtual-kubelet=trueand tolerates the virtual-kubelet taint. The scheduler answers plainly:us-west-1-alicehas five nodes: three control-plane and two plain workers. No kraftlet node exists there, and none carries the label the pod requires. The same holds onus-east-1-alice.Across the whole production fleet only
us-central-1-labhas kraftlet nodes. More clusters are expected to gain them, so the set of valid regions will change rather than stay fixed.The infrastructure side of this, making the runtime's presence readable outside Flux, is https://github.com/datum-cloud/infra/issues/5382.
Related to #278