[ClusterLabs] Gracefully stop nodes one by one with disk-less sbd

[Yan Gao]

Hi Klaus,

On 8/12/19 1:39 PM, Klaus Wenninger wrote:
> On 8/9/19 9:06 PM, Yan Gao wrote:
>> On 8/9/19 6:40 PM, Andrei Borzenkov wrote:
>>> 09.08.2019 16:34, Yan Gao пишет:
>>>> Hi,
>>>>
>>>> With disk-less sbd,  it's fine to stop cluster service from the cluster
>>>> nodes all at the same time.
>>>>
>>>> But if to stop the nodes one by one, for example with a 3-node cluster,
>>>> after stopping the 2nd node, the only remaining node resets itself with:
>>>>
>>> That is sort of documented in SBD manual page:
>>>
>>> --><--
>>> However, while the cluster is in such a degraded state, it can
>>> neither successfully fence nor be shutdown cleanly (as taking the
>>> cluster below the quorum threshold will immediately cause all remaining
>>> nodes to self-fence).
>>> --><--
>>>
>>> SBD in shared-nothing mode is basically always in such degraded state
>>> and cannot tolerate loss of quorum.
>> Well, the context here is it loses quorum *expectedly* since the other
>> nodes gracefully shut down.
>>
>>>
>>>> Aug 09 14:30:20 opensuse150-1 sbd[1079]:       pcmk:    debug:
>>>> notify_parent: Not notifying parent: state transient (2)
>>>> Aug 09 14:30:20 opensuse150-1 sbd[1080]:    cluster:    debug:
>>>> notify_parent: Notifying parent: healthy
>>>> Aug 09 14:30:20 opensuse150-1 sbd[1078]:  warning: inquisitor_child:
>>>> Latency: No liveness for 4 s exceeds threshold of 3 s (healthy servants: 0)
>>>>
>>>> I can think of the way to manipulate quorum with last_man_standing and
>>>> potentially also auto_tie_breaker, not to mention
>>>> last_man_standing_window would also be a factor... But is there a better
>>>> solution?
>>>>
>>> Lack of cluster wide shutdown mode was mentioned more than once on this
>>> list. I guess the only workaround is to use higher level tools which
>>> basically simply try to stop cluster on all nodes at once. It is still
>>> susceptible to race condition.
>> Gracefully stopping nodes one by one on purpose is still a reasonable
>> need though ...
> If you do the teardown as e.g. pcs is doing it - first tear down
> pacemaker-instances and then corosync/sbd - it is at
> least possible to tear down the pacemaker-instances one-by one
> without risking a reboot due to quorum-loss.
> With kind of current sbd having in
> - 
> https://github.com/ClusterLabs/sbd/commit/824fe834c67fb7bae7feb87607381f9fa8fa2945
> - 
> https://github.com/ClusterLabs/sbd/commit/79b778debfee5b4ab2d099b2bfc7385f45597f70
> - 
> https://github.com/ClusterLabs/sbd/commit/a716a8ddd3df615009bcff3bd96dd9ae64cb5f68
> this should be pretty robust although we are still thinking
> (probably together with some heartbeat to pacemakerd
> that assures pacemakerd is checking liveness of sub-daemons
> properly) of having a cleaner way to detect graceful
> pacemaker-shutdown.
These are all good improvements, thanks!

But in this case the remaining node is not shutting down yet, or it's 
intentionally not being shut down :-) Loss of quorum is as expected, so 
is following no-quorum-policy, but self-reset is probably too much?

Regards,
   Yan