When a node in the infrastructure cluster is cordoned or drained under a running cell, a human has to act: cordon the inner Node, drain the inner workloads, delete the cell, let it be recreated elsewhere.
docs/design/gpucellpool-overview.md §6(a) originally promised this ("Phase 4 automates the correct order: cordon inner Node → drain inner workloads → delete cell → recreate elsewhere"). It was never built — there is no eviction code in the repo. Rather than quietly drop the claim, it is recorded as a limitation.
Note this is the one place the operator would need pods/eviction in the workload cluster. It deliberately does not hold that right today, because it waits for a GPU to be released rather than taking it away — so implementing this is also an RBAC decision, not just a code one.
When a node in the infrastructure cluster is cordoned or drained under a running cell, a human has to act: cordon the inner Node, drain the inner workloads, delete the cell, let it be recreated elsewhere.
docs/design/gpucellpool-overview.md§6(a) originally promised this ("Phase 4 automates the correct order: cordon inner Node → drain inner workloads → delete cell → recreate elsewhere"). It was never built — there is no eviction code in the repo. Rather than quietly drop the claim, it is recorded as a limitation.Note this is the one place the operator would need
pods/evictionin the workload cluster. It deliberately does not hold that right today, because it waits for a GPU to be released rather than taking it away — so implementing this is also an RBAC decision, not just a code one.