miroir¶

Purpose¶

miroir provisions miroir-local node-local volumes from loopfile-backed pools. app runs the controller and agents; config defines the node group, StorageClass and snapshot classes.

CSIStorageCapacity and the overcommit incident¶

storageCapacity.enabled: true makes the external-provisioner publish a CSIStorageCapacity object per node and storage class from the driver's GetCapacity RPC, so the scheduler can steer a WaitForFirstConsumer pod away from a node whose pool is already full. Before this was enabled, the scheduler had no such signal and placed pods however it liked, leaving CreateVolume to refuse afterward.

Over time, nearly every VolSync cache volume accreted onto a single node: 194 of 210 volumes, against 10 and 6 on the other two. #5098 then raised the kopia cache to 16Gi, correctly, since the shared repository index was 7.57Gi and growing about 25MiB a day. 194 of those resizes succeeded on the busiest node and consumed the remaining headroom:

The last two resizes and every new volume on that node were refused with a ResourceExhausted error citing the capacity x2 overcommit guard, even though real disk on that node was only 45% used. The --overcommit-ratio=2 and --free-space-ratio=20 provisioner guards were not changed; publishing CSIStorageCapacity gives the scheduler information it was missing rather than raising the limit (#5110).

This fix stops the imbalance from getting worse; it does not undo it. The 194 existing caches stay where they are until an operator deletes their PVCs by hand. VolSync then recreates each one automatically on its next sync, on a node with room, at the cost of one cold run.

Agents republish pool stats every 60s (agent.poolStatsInterval). Until the first publish after a rollout, the scheduler treats an unpublished (node, class) pair as unfit, so new miroir-local provisioning waits about a minute; already-bound volumes are unaffected.