[Pacemaker] Unique clone instance is stopped too early on move
Vladislav Bogdanov
bubble at hoster-ok.com
Fri Apr 17 06:19:42 UTC 2015
17.04.2015 00:48, Andrew Beekhof wrote:
>
>> On 22 Jan 2015, at 12:04 am, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
>>
>> 20.01.2015 02:44, Andrew Beekhof wrote:
>>>
>>>> On 16 Jan 2015, at 3:59 pm, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
>>>>
>>>> 16.01.2015 07:44, Andrew Beekhof wrote:
>>>>>
>>>>>> On 15 Jan 2015, at 3:11 pm, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
>>>>>>
>>>>>> 13.01.2015 11:32, Andrei Borzenkov wrote:
>>>>>>> On Tue, Jan 13, 2015 at 10:20 AM, Vladislav Bogdanov
>>>>>>> <bubble at hoster-ok.com> wrote:
>>>>>>>> Hi Andrew, David, all.
>>>>>>>>
>>>>>>>> I found a little bit strange operation ordering during transition execution.
>>>>>>>>
>>>>>>>> Could you please look at the following partial configuration (crmsh syntax)?
>>>>>>>>
>>>>>>>> ===
>>>>>>>> ...
>>>>>>>> clone cl-broker broker \
>>>>>>>> meta interleave=true target-role=Started
>>>>>>>> clone cl-broker-vips broker-vips \
>>>>>>>> meta clone-node-max=2 globally-unique=true interleave=true resource-stickiness=0 target-role=Started
>>>>>>>> clone cl-ctdb ctdb \
>>>>>>>> meta interleave=true target-role=Started
>>>>>>>> colocation broker-vips-with-broker inf: cl-broker-vips cl-broker
>>>>>>>> colocation broker-with-ctdb inf: cl-broker cl-ctdb
>>>>>>>> order broker-after-ctdb inf: cl-ctdb cl-broker
>>>>>>>> order broker-vips-after-broker 0: cl-broker cl-broker-vips
>>>>>>>> ...
>>>>>>>> ===
>>>>>>>>
>>>>>>>> After I put one node to standby and then back to online, I see the following transition (relevant excerpt):
>>>>>>>>
>>>>>>>> ===
>>>>>>>> * Pseudo action: cl-broker-vips_stop_0
>>>>>>>> * Resource action: broker-vips:1 stop on c-pa-0
>>>>>>>> * Pseudo action: cl-broker-vips_stopped_0
>>>>>>>> * Pseudo action: cl-ctdb_start_0
>>>>>>>> * Resource action: ctdb start on c-pa-1
>>>>>>>> * Pseudo action: cl-ctdb_running_0
>>>>>>>> * Pseudo action: cl-broker_start_0
>>>>>>>> * Resource action: ctdb monitor=10000 on c-pa-1
>>>>>>>> * Resource action: broker start on c-pa-1
>>>>>>>> * Pseudo action: cl-broker_running_0
>>>>>>>> * Pseudo action: cl-broker-vips_start_0
>>>>>>>> * Resource action: broker monitor=10000 on c-pa-1
>>>>>>>> * Resource action: broker-vips:1 start on c-pa-1
>>>>>>>> * Pseudo action: cl-broker-vips_running_0
>>>>>>>> * Resource action: broker-vips:1 monitor=30000 on c-pa-1
>>>>>>>> ===
>>>>>>>>
>>>>>>>> What could be a reason to stop unique clone instance so early for move?
>>>>>>>>
>>>>>>>
>>>>>>> Do not take it as definitive answer, but cl-broker-vips cannot run
>>>>>>> unless both other resources are started. So if you compute closure of
>>>>>>> all required transitions it looks rather logical. Having
>>>>>>> cl-broker-vips started while broker is still stopped would violate
>>>>>>> constraint.
>>>>>>
>>>>>> Problem is that broker-vips:1 is stopped on one (source) node unnecessarily early.
>>>>>
>>>>> It looks to be moving from c-pa-0 to c-pa-1
>>>>> It might be unnecessarily early, but it is what you asked for... we have to unwind the resource stack before we can build it up.
>>>>
>>>> Yes, I understand that it is valid, but could its stop be delayed until cluster is in the state when all dependencies are satisfied to start it on another node (like migration?)?
>>>
>>> No, because "we have to unwind the resource stack before we can build it up."
>>> Doing anything else would be one of those things that is trivial for a human to identify but rather complex for a computer.
>>
>> I believe there is also an issue with migration of clone instances.
>>
>> I modified pe-input to allow migration of cl-broker-vips (and also set inf score for broker-vips-after-broker
>> and make cl-broker-vips interleaved).
>> Relevant part is:
>> clone cl-broker broker \
>> meta interleave=true target-role=Started
>> clone cl-broker-vips broker-vips \
>> meta clone-node-max=2 globally-unique=true interleave=true allow-migrate=true resource-stickiness=0 target-role=Started
>> clone cl-ctdb ctdb \
>> meta interleave=true target-role=Started
>> colocation broker-vips-with-broker inf: cl-broker-vips cl-broker
>> colocation broker-with-ctdb inf: cl-broker cl-ctdb
>> order broker-after-ctdb inf: cl-ctdb cl-broker
>> order broker-vips-after-broker inf: cl-broker cl-broker-vips
>>
>> After that (part of) transition is:
>>
>> * Resource action: broker-vips:1 migrate_to on c-pa-0
>> * Pseudo action: cl-broker-vips_stop_0
>> * Resource action: broker-vips:1 migrate_from on c-pa-1
>> * Resource action: broker-vips:1 stop on c-pa-0
>> * Pseudo action: cl-broker-vips_stopped_0
>> * Pseudo action: all_stopped
>> * Pseudo action: cl-ctdb_start_0
>> * Resource action: ctdb start on c-pa-1
>> * Pseudo action: cl-ctdb_running_0
>> * Pseudo action: cl-broker_start_0
>> * Resource action: ctdb monitor=10000 on c-pa-1
>> * Resource action: broker start on c-pa-1
>> * Pseudo action: cl-broker_running_0
>> * Pseudo action: cl-broker-vips_start_0
>> * Resource action: broker monitor=10000 on c-pa-1
>> * Pseudo action: broker-vips:1_start_0
>> * Pseudo action: cl-broker-vips_running_0
>> * Resource action: broker-vips:1 monitor=30000 on c-pa-1
>>
>> But, I would say that at least from a human logic PoV the above breaks ordering rule broker-vips-after-broker
>> (cl-broker-vips finished migrating and thus runs on c-pa-1 before cl-broker started there).
>> Technically broker-vips:1_start_0 goes at the right position, but actually resource is "started"
>> in migrate_to/mifrate_from.
>>
>>
>> I also went further and injected a pair of non-clone IPAddr2 resources into the same pe-input, and also enabled migration
>> for them (returning interleave for cl-broker-vips to false and setting ordering score for broker-vips-after-broker back to 0,
>> so all three order constraints are adjacent):
>>
>> clone cl-broker broker \
>> meta interleave=true target-role=Started
>> clone cl-broker-vips broker-vips \
>> meta clone-node-max=2 globally-unique=true interleave=false allow-migrate=true resource-stickiness=0 target-role=Started
>> clone cl-ctdb ctdb \
>> meta interleave=true target-role=Started
>> primitive broker-vip1 IPaddr2 \
>> params ip=192.168.122.70 cidr_netmask=24 nic=eth0 \
>> op start interval=0 timeout=20 \
>> op stop interval=0 timeout=20 \
>> op monitor interval=30
>> primitive broker-vip2 IPaddr2 \
>> params ip=192.168.122.71 cidr_netmask=24 nic=eth0 \
>> op start interval=0 timeout=20 \
>> op stop interval=0 timeout=20 \
>> op monitor interval=30
>> colocation broker-with-ctdb inf: cl-broker cl-ctdb
>> colocation broker-vips-with-broker inf: cl-broker-vips cl-broker
>> colocation broker-vip1-with-broker inf: broker-vip1 cl-broker
>> colocation broker-vip2-with-broker inf: broker-vip2 cl-broker
>> colocation broker-vip2-not-with-vip1 -100: broker-vip2 broker-vip1
>> order broker-after-ctdb inf: cl-ctdb cl-broker
>> order broker-vips-after-broker 0: cl-broker cl-broker-vips
>> order broker-vip1-after-broker 0: cl-broker broker-vip1
>> order broker-vip2-after-broker 0: cl-broker broker-vip2
>>
>> For broker-vip2 I see completely different output (compare with broker-vips:1):
>>
>> * Resource action: broker-vips:1 migrate_to on c-pa-0
>
> I just noticed this, since when does IPaddr2 migrate?
I just injected allow_migrate for broker-vip1, broker-vip2 and
broker-vips into the pe_input to test what would pengine do but forgot
to note that (actually cl-broker-vips definition above has it enabled
but broker-vip{1,2} misses that, damn, my fault, it should be there
too). I need to be more accurate.
For g-u clone it doesn't solve the issue btw. But for ordinary resource
it does. That makes me think that migration paths differ for g-u clone
instances.
Actually, implementing (pseudo-)migration in IPaddr2 doesn't seem to be
very complex task.
>
> Reason I noticed is because broker-vips definitely doesn’t start until the end anymore:
>
> * Resource action: broker start on c-pa-1
> * Pseudo action: cl-broker_running_0
> * Pseudo action: cl-broker-vips_start_0
> * Resource action: broker monitor=10000 on c-pa-1
> * Resource action: broker-vips:1 start on c-pa-1
Actually it is migrated at the very beginning of the transition, and
that seems to be a big issue to me, because it breaks ordering (start
became a pseudo-action, but actual work should be done in migrate_from
which is run before broker start).
>
>
>
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
>
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://bugs.clusterlabs.org
>
More information about the Pacemaker
mailing list