Description of problem: Disconnecting the shared storage from a member
when it is running services causes the member to reboot (expected).
However, services do not fail-over to the other member.
Version-Release number: 1.2.6-1
How reproducible: 30%
Steps to Reproduce:
1. Start cluster on two members.
2. Move all services to member A
3. Disconnect shared storage from member A.
Actual results: Member A still "runs" all services, even after it is
Expected results: Member B should take over services.
Additional info: Tested on shared SCSI RAID array; fibre channel may
have different behavior given that the timeouts are different. For
instance, SCSI cable disconnects are more adequately handled in the
drivers, whereas some FC drivers continue to retry for several minutes
before giving up.
Created attachment 96584 [details]
Patch to fix failover
This was due to a strange condition which may exist outside of the
shared-storage disconnect, but is certainly exascerbated by it.
This fix breaks rolling upgrade in real-world testing. Fix it.
Ok, rolling upgrade still works - but only if it is done in a special
way. The lowest-ordered member must be upgraded *last*.
This is because the new code checks for a cluster quorum incarnation
number, while the old code does not. Since the lowest-ordered member
is the lock keeper, the check must be disabled until all other members
have restarted their services.
This workaround should be sufficient; as I believe this fix is
necessary for long-term support.
(Official) Fix for this should appear in Update 1
There was a second issue causing services not to fail over when the
high node was unplugged because the lock daemon was waiting for a
message from the quorum daemon, who was waiting for the STONITH
operations to complete. The lock client would give up prematurely -
where it should have retried the locking operation during a failover.
Fixed in 1.2.9-1
Verified, closed as errata.
The latest CVS build fixes this:
We are still experiencing this issue with 1.2.9 on top of RHEL3 U1.
Appropriate bug for your specific issue is:
Fixing product name. Clumanager on RHEL3 was part of RHCS3, not RHEL3