[ClusterLabs] iSCSILogicalUnit - scsi_id and multiple clusters

[Strahil Nikolov]

On March 6, 2020 11:06:13 AM GMT+02:00, Oyvind Albrigtsen <oalbrigt at redhat.com> wrote:
>Hi Strahil,
>
>It seems like it tries to set one based on the resource name, and from
>a quick check it seems like it also did on RHEL 7.5.
>
>https://github.com/ClusterLabs/resource-agents/blob/master/heartbeat/iSCSILogicalUnit.in#L57
>
>
>Oyvind
>
>On 05/03/20 23:15 +0200, Strahil Nikolov wrote:
>>Hey Community,
>>
>>I finaly got some time to report  an issue with iSCSILogicalUnit and
>scsi_id  ( https://github.com/ClusterLabs/resource-agents/issues/1463
>).
>>
>>The  issue was observed a while ago on RHEL 7.5 and SLES 12 SP4.
>>
>>Do you know if any change was made to:
>>A)  Either make 'scsi_id'  a mandatory option
>>B)  When 'scsi_id' is   not provided by the admin,  a random one is
>picked (permanently)
>>
>>If not, then the github issue is relevant.
>>
>>Sadly, my exam (EX436) is on monday and I can't test it before that.
>>
>>Best Rregards,
>>Strahil Nikolov
>>_______________________________________________
>>Manage your subscription:
>>https://lists.clusterlabs.org/mailman/listinfo/users
>>
>>ClusterLabs home: https://www.clusterlabs.org/
>>
>
>_______________________________________________
>Manage your subscription:
>https://lists.clusterlabs.org/mailman/listinfo/users
>
>ClusterLabs home: https://www.clusterlabs.org/

Thanks for the reply, Oyvind

In an environment with naming convention that allows globally non-unique  resource names - this could cause the behaviour I have described in github (2 clusters,  2 separate luns,  client's multipath aggregates them in 1 lun with 4 paths).

Do you think that a better approach will be to  mark 'scsi_id' as mandatory option (of course  we can bypass with  'pcs  --force') or  we should randomly set one on resource creation if the value is not defined ?
Of course  ,  the algorithm can get a little modification  - a random seed  for example.

The second & third  options will be harder  to implement as we got both pcs  & crmsh actively used among distributions,  but will add some 'dummy proofness'. For me the first option is easiest to implement.


Best Regards,
Strahil Nikolov