[ClusterLabs] Antw: Re: Question on sharing data with DRDB

[Digimer]

On 2019-03-21 2:57 a.m., Ulrich Windl wrote:
>>>> Digimer <lists at alteeve.ca> schrieb am 20.03.2019 um 18:47 in Nachricht
> <38f790d0-4bb6-53b8-7eb4-b285ed14779d at alteeve.ca>:
>> On 2019-03-20 1:46 p.m., Valentin Vidic wrote:
>>> On Wed, Mar 20, 2019 at 01:34:52PM -0400, Digimer wrote:
>>>> Depending on your fail-over tolerances, I might add NFS to the mix and
>>>> have the NFS server run on one node or the other, exporting your ext4 FS
>>>> that sits on DRBD in single-primary mode.
>>>>
>>>> The failover (if the NFS host died) would look like this;
>>>>
>>>> 1. Lost node is fenced.
>>>> 2. DRBD is promoted from Secondary to Primary
>>>> 3. ext4 FS is mounted.
>>>> 4. Virtual IP (used for NFS) is brought up.
>>>> 5. NFS starts
>>>>
>>>> Startup and graceful migration would be the same, minus the fence.
>>>
>>> Would it be possible for DRBD to go into SplitBrain if the lost node
>>> manages to write something to local DRBD disk before it gets fenced?
>>
>> Not when DRBD is configured correctly. You sent 'fencing
> 
> Are your really saying that dual-primary no longer needs operator intervention
> after a node crash?
> That's also split-brain to me if DRBD does not know how to recover after a
> link interruption.

Correct.

With proper fencing, automated recovery is not just possible, we've seen
it work in the field dozens of times over the last several years.

When there is a link interruption, IO blocks and a fence is called. When
the lost (or slower) node is fenced, only then does IO resume. No writes
happen during the period of no communication, so there is no split
brain. When the lost / fenced node does finally return, it knows it is
behind and pulls changes from the UpToDate peer.

>> resource-and-stonith;' and set the appropriate fence handler. This tells
>> DRBD to not proceed with a write while a node is in an unknown state
>> (which happens when the node stops responding and is cleared on
>> successful fence).


-- 
Digimer
Papers and Projects: https://alteeve.com/w/
"I am, somehow, less interested in the weight and convolutions of
Einstein’s brain than in the near certainty that people of equal talent
have lived and died in cotton fields and sweatshops." - Stephen Jay Gould