Tasks for operations on ESX hosts and VMFS datastores fail with errors about hosts losing connection or objects being in the wrong state.

When App LAyering makes a request to vSphere for file operations - like copying cached disks or flattening sparse files - we only specify the file and the datastore.  VMware will choose the host it wants to use for this operation.  If a host in the cluster - a host which has access to the datastore in question - is in maintenance mode, vCenter may still try to use it because it does have access to the datastore.  However, machines in maintenance mode or otherwise in a bad state can't usually fulfill the request, instead returning errors about losing connectivity or an invalid state.

The only workaround is to delay the work until the machine is back out of maintenance mode, or to remove the host from the cluster entirely.  There appears to be no way for VMware to explicitly not route filesystem requests to idle machines even if those machines should be invalid.