[ClusterLabs] DRBD on asymmetric-cluster
Jason Voorhees
jvoorhees1 at gmail.com
Thu Apr 7 19:05:21 CEST 2016
>> Agree. So far I haven't created any ordering constraints because it
>> isn't important to me, YET, the order for starting services. However I
>> have a question... if I don't have any ordering constraints at all, am
>> I still able to activate resources no matter the order?
>
> Sort of, but not exactly.
>
> With a colocation constraint "A with B", the cluster must assign B to a
> node before it can place A. B does not have to be started, but it does
> have to be *able* to be started, in order to be assigned to a node. So
> if something prevents B from being started (disabled in config, not
> allowed on any online node, etc.), it will not be assigned to a node,
> and A will not run.
>
> That doesn't mean that B will be started first, though. If the cluster
> needs to start both A and B, it can start them in any order. With an
> ordering constraint "B then A", B must be started first, and the start
> must complete successfully, before A can be started.
Ok, cool, it's a little more clear now.
>>> The above constraint as currently worded will have no effect. It says
>>> that clusterdataClone must be located on either nodo1 or nodo2. Since
>>> those are your only nodes, it doesn't really constrain anything.
>> Ok, the last command (location with rule) was created to allow
>> clusterdataClone start at both nodes, because without this rule the
>> resource was always in "stopped" status in both nodes. Once I added
>> this rule my clusterdataClone resource started automatically but I
>> don't understand why it choosed a node to run as Master and the other
>> one as Slave. Is it random?
>
> I don't know why the resource would be stopped without this constraint.
> Maybe you have an opt-in cluster? But in that case you can use a normal
> location constraint, you don't need a rule.
Yes, it's an asymmetric (opt-in) cluster. I don't remember if I tried
before with a normal location without rules but I'll try that again.
>
> It will choose one as master and one as slave because you have
> master-max=1. The choice, as with everything else in pacemaker, is based
> on scores, but these are transparent to the user and appears "random".
>
I see, that's where I thought a rule would be needed for constraint
location as you suggested me in a previous email. (see next line
below)
>>> If you want to prefer one node for the master role, you want to add
>>> role=master, take out the node you don't want to prefer, and set score
>>> to something less than INFINITY.
>> Well, I could add a rule to prefer nodo1 over nodo2 to run the Master
>> role (in fact, I think I already did it) but what I want it's
>> something different: I would like the Master role to follow IPService,
>> I mean, clusterdataClone become Master where IPService was previously
>> activated.
>>
>> Is this possible? Or the only way to configure constraints is that my
>> resources (IPService, Web, MTA) follow the Master role of
>> clusterdataClone?
>
> I think the latter approach makes more sense and is common. Storage is
> more complicated than an IP and thus more likely to break, so it would
> seem to be more reliable to follow where storage can successfully start.
> The exception would be if the IP is much more important to you than the
> storage and is useful without it.
>
Well, it's just that I come from a previous and long experience with
IBM PowerHA for AIX. I was used to see the IP Service being activated
on each node, then the shared storage+filesystems and applications
scripts (aka services) at the end, so I tried to emulate the same
behavior in pacemaker.
But it makes sense what you say so, I might try to configure my
master/slave resource to be activated as Master in a preferred node
and force all other resources (such as IPService, Web and MTA) to
follow it.
> You might want to look at resource sets. The syntax is a bit difficult
> to follow but it's very flexible. See pcs constraint colocation/order set.
I was reading a little bit about it but I prefer to skip this for now at least.
Thanks a lot Ken :) Now I can resume my tests with a better
understanding of how things work. I'll have some fun today!
More information about the Users
mailing list