The platform forces cloneStrategyOverride: csi-clone on every cluster, and it is still open whether that should stay the default. #2908 listed this as its second fix direction. #4796 closes #2908 without deciding it.
The setting can be turned off since #4138: an empty value omits the override, and CDI then picks a strategy per StorageProfile. What's missing is the decision on the default itself.
The case for keeping csi-clone: same-class clones are the common case, and LINSTOR's native clone is the fast path for them. That's why 42778cf switched to it.
The case for changing it: the override replaces per-StorageProfile selection on every class, including ones where CSI clone is the wrong choice. With #4796, vm-disk refuses the case known to hang, a LINSTOR target without remote access from every node, but only for source.image. A source.disk clone into another class is not covered. kubernetes-nodes has its own guard (#4849).
Options I see:
- Keep
csi-clone and rely on the guards.
- Default to empty and let each StorageProfile decide, setting
cloneStrategy: csi-clone on the profiles where it is known to work.
I haven't measured the difference between the two on a cluster.
The platform forces
cloneStrategyOverride: csi-cloneon every cluster, and it is still open whether that should stay the default. #2908 listed this as its second fix direction. #4796 closes #2908 without deciding it.The setting can be turned off since #4138: an empty value omits the override, and CDI then picks a strategy per StorageProfile. What's missing is the decision on the default itself.
The case for keeping
csi-clone: same-class clones are the common case, and LINSTOR's native clone is the fast path for them. That's why 42778cf switched to it.The case for changing it: the override replaces per-StorageProfile selection on every class, including ones where CSI clone is the wrong choice. With #4796, vm-disk refuses the case known to hang, a LINSTOR target without remote access from every node, but only for
source.image. Asource.diskclone into another class is not covered. kubernetes-nodes has its own guard (#4849).Options I see:
csi-cloneand rely on the guards.cloneStrategy: csi-cloneon the profiles where it is known to work.I haven't measured the difference between the two on a cluster.