On 12/02/2012 02:39 AM, Hermes Flying wrote:
> So only Primary/Actice supported. Good to know.

What part of my reply indicated this? Quote:
---
Pacemaker does not inherently have a concept of Primary/Secondary. It
can be configured to operate in such a mode, or it can be active/active,
n-1, n-n, etc. It all comes down to how each given resource is configured.
---

> In your description of 3+nodes, they all try to become the primary. If a
> node does not have access to the external network (besides the corosync
> paths) does it still try to become primary?

Again, you misunderstood what I said. Pacemaker does not, itself, have a
concept of "primary" node. It entirely depends on how you configure a
service. If a service is set to run in one place at a time only, then it
will pick one surviving node to restart lost services. How it does that
depends on how you've configured the fail-over of that given service and
it very likely will differ from one service to the next.

> ------------------------------------------------------------------------
> *From:* Digimer <[email protected]>
> *To:* David Coulson <[email protected]>
> *Cc:* Hermes Flying <[email protected]>; General Linux-HA mailing
> list <[email protected]>
> *Sent:* Saturday, December 1, 2012 7:20 PM
> *Subject:* Re: [Linux-HA] Some help on understanding how HA issues are
> addressed by pacemaker
> 
> On 12/01/2012 08:53 AM, David Coulson wrote:
>>
>> On 12/1/12 8:48 AM, Hermes Flying wrote:
>>> Great help! Please allow me to trouble you with one last question.
>>>
>>> If I get this, when I use fencing and the corosync fails then linux-2
>>> will attempt to crash linux-1 and take over. At this point though
>>> linux-1 won't try to do anything right? Since it knows it is the
>>> primary, I mean.
>>
>> linux-1 will be powered off or crashed, so i think that speaks for itself.
> 
> Pacemaker does not inherently have a concept of Primary/Secondary. It
> can be configured to operate in such a mode, or it can be active/active,
> n-1, n-n, etc. It all comes down to how each given resource is
> configured. You can have multiple resources in multiple configurations
> on the same instance. For example, you may have a DRBD resource set to
> run on both nodes (active/active), and a VM that runs on only one node
> at a time (active/passive) that uses the DRBD storage.
> 
> So, given this, there is no inherent mechanism to say "always fence node
> X".
> 
> So imagine you have two nodes and they lose communication; Both will
> think the other has failed and both will try to fence it's peer. The
> classic analogy is "an old west shoot-out"; Fastest node lives, slower
> nodes gets powered off.
> 
> You can help ensure one node has a better chance of winning by setting a
> delay on it. So it the above config, if you had set a 5 second delay
> against node 1, then it would get a 5 second head start in fencing node
> 2 (node 2 will wait the set time before trying to fence node 1). In this
> way, you can help ensure one node lives in such a case.
> 
> Sometime a delay is needed in any case because in some fence methods,
> like IPMI-based devices, it's technically possible that both nodes will
> get their fence call out before dieing, leaving both nodes powered off.
> Setting a delay will protect against this.
> 
>>> Then you say:"Any resource previously running on linux-1 will be
>>> started on linux-2."
>>> Now at this point: By resource you mean only pacemaker and its related
>>> modules, right? Because I want  Tomcat to be up and running and
>>> receiving requests in Linux-2 as well, which will be forwarded by load
>>> balancer of linux-1. Is this correct?
>>
>> I mean 'resources managed by pacemaker'. So if you VIP was running on
>> linux-1, and it fails, and linux-2 fences it, the only place the VIP can
>> run is linux-2. linux-1 is totally down.
>>>
>>> Also in your setup of 2 NICs or 2 switches I assume that the idea is
>>> that the probability of split-brain due to network failure is very low
>>> right? Because I have read that it is not possible to avoid
>>> split-brain without adding a third node. But I may be misunderstanding
>>> this
>> A third node will eliminate split brain by definition, as quorum will
>> only be obtained if a minimum of two nodes are available.
>>
>> If you have a diverse network configuration and good change management,
>> you're probably not going to experience a split brain unless you have a
>> substantial environment failure that will probably impact your client
>> ability to access anything. Since you are not running shared storage,
>> you're not going to experience data loss which is typically the biggest
>> concern with split brain.
> 
> A couple of notes;
> 
> Only mode=1 bonding (active/passive) works reliably. No other mode is
> supported (and I've tested all and found all other modes would fail).
> 
> Quorum comes into play with 3+ nodes, but it does not eliminate the need
> for fencing. Imagine that a node was in the middle of a write to shared
> storage and then hung totally. The other two nodes declare it as failed,
> don't fence it, and proceed to clean up the shared storage to return it
> to a consistent state. Time passes and new data gets written to the same
> area of shared storage. Then the hung node thaws, has no idea time has
> passed, and just finishes the write thinking it's locks are still valid
> and that it still has quorum. Voila! Corrupt storage, despite quorum.
> 
> Clustering all all about protecting against failures. Hand-waving a
> failure as unlikely is not a good practice in HA clustering. If it's
> possible, plan for it. :)
> 
> -- 
> Digimer
> Papers and Projects: https://alteeve.ca/w/
> What if the cure for cancer is trapped in the mind of a person without
> access to education?
> 
> 


-- 
Digimer
Papers and Projects: https://alteeve.ca/w/
What if the cure for cancer is trapped in the mind of a person without
access to education?
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to