[jira] [Created] (CLOUDSTACK-9350) Local storage hosts get HA tasks, cause issues

["Abhinandan Prateek (JIRA)" <ji...@apache.org>]

Abhinandan Prateek created CLOUDSTACK-9350:
----------------------------------------------

             Summary: Local storage hosts get HA tasks, cause issues	
                 Key: CLOUDSTACK-9350
                 URL: https://issues.apache.org/jira/browse/CLOUDSTACK-9350
             Project: CloudStack
          Issue Type: Bug
      Security Level: Public (Anyone can view this level - this is the default.)
    Affects Versions: 4.5.1
            Reporter: Abhinandan Prateek
            Assignee: Abhinandan Prateek


When a host hits its ping time out, for whatever reason, the investigators are triggered. The KVMInvestigator sends a CheckOnHostCommand to the target host, and then to all the remaining neighbor hosts in the cluster. The CheckOnHostCommand (and also FenceCommand, the code is nearly identical) is processed by the KVM agent and simply scans through all NFS primary storage looking for the host's heartbeat in the KVMHA directory. If no heartbeat file is found, it fails the check. In the case of clusters that are local-only, these hosts will always fail the check, whether it be the target host or a neighbor checking on the target. This triggers a host 'down' event, which triggers HA tasks. The HA tasks will attempt to stop any VMs on the host, and then if the VM's offering is HA-enabled it will try to restart the VM.
Our recent issue was that a management server took extraordinarily long to rotate its logs and was slow to process some host pings. The CheckOnHostCommand was sent to a suspect host, which failed because it had no primary NFS. The neighbor checks also failed to check the suspect host's heartbeat for the same reason. Then the host was marked as down and all VMs were stopped. Multiply this by a few dozen hosts.
The immediate fix, provided in the example, is a patch to KVMInvestigator which will only attempt investigation if the host's cluster has NFS storage, which is a requirement for the host to run the check, as described above. If there is none, the host state is determined to be disconnected rather than down. This means that the host will still end up in alert state and need manual investigation, but there will be no attempt to stop or HA the VMs.
Additionally, the patch catches scenarios where a cluster might have both NFS and local storage and a host ends up in 'down' state. In this case, when the HA tasks are being created, if a VM is using local storage then the HA task generation is skipped. This VM can't be started anywhere else.
We could also make the agent side more robust, in KVMHAChecker we may not want it to return 'false' if there were zero pools passed to check for HA heartbeat. Then again, maybe we do. We decided initially to patch just the server side, because it is easier to deploy.
In the long run, I'd hope that the current HA work would supercede the current KVMInvestigator and take the cluster's ability to pass any defined checks into account before checking.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)