On 15.10.2016 12:56, Gandalf Corvotempesta wrote:
> Hi to all
> i would like to spin up a new shared storage
> should i use drbd 8 or 9?
> Additionally and more important: are there any ways to totally avoid
> splitbrains? Obviously, the network used for sync is fully redundant (at
> least 2 or 3 bonded interfaces)
> Is this enough?
> if i understood properly, split brain happens when the two nodes loose
> their interconnection, thus each node will detect itself as the online
> node and start to write.
> when writes differs between nodes, i have a split brain
> But if the sync network is redundant, no splitbrain should occur
> Anything else to be aware of?
Keep in mind that it is never possible to be 100% sure that no
split-brain can occur. You might always run into freak accidents.
When you say you use bonded interfaces for redundancy are the physical
ports actually connected to separate switches? For proper redundancy the
whole network path between the systems must be redundant including any
switches that sit in between.
If the switch redundancy is done using the stacking feature of those
switches there is also the possibility that the switches don't
accurately detect a failure of the primary interface and thus the
redundancy doesn't really work. This should be unlikely but it's
possible and one of the freak accident scenarios I mentioned above where
you still end up in a split-brain situation.
In a two-node cluster you can possibly avoid this scenario by making the
second route completely independent using a direct cross-over connection
between the two machines and using the RRP:
drbd-user mailing list