Benjamin Marzinski wrote:
If you have two machines doing raid1 over a local device and a gnbd device,
you have the problem were if machine A dies after it has written to it's local
disk but not the disk on machine B. The mirror is out of sync.
This is recoverable and MD can handle this. Exactly the same can happen
on a single machine. If disk A dies during a write operation and causes
a power spike that crashes or bus-locks the machine, on recovery disk B
is in the same state as you describe.
You will lose the last file system transaction, but the journal on the
surviving node saves the day.
Besides, the mirror is always out of sync when a node recovers. The
mirror operation is always towards the last disk to be "mdadm --add"ed
DRBD was made for exactly this setup, and will (I believe) automagically handle this correctly.
Quite right, I agree. The major problem with DRBD is that cluster
manager has no control over the active/standby state of the disks. So a
custom exporting daemon is needed that moves with DRBD master. We are
finding it gets a bit complex on failover.
Regards,
James
--
Linux-cluster@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/linux-cluster
