My approach to backing up guest VM’s from ESX3.5 is two-headed:

• Backup the data like you would a normal machine by attaching via SMB, NFS, SQL, etc.  Just get the data
• Backup the Guest VM via the ESX host (pull a bare metal snapshot of the entire system)

Why backup everything twice you ask?

Well fortunately for you and unfortunately for me I have a real world experience that will help answer this question. I had a vm guest failure and had to rebuilt a vm then restore data and settings.  Suffice to say this took forever. So here’s why you can do both…

• If data is deleted you have backup sets that have just pure data - files that are used by applications and users. If a single file gets deleted, corrupted, or anything else _bad_ the restore for this file(s) is quick and easy.  Restoration of this file(s) doesn’t need to affect every other user connected to the system in question.
• If the OS becomes unusable, system files fail, vmdk (virtual disk) files on the host get corrupted then you now have the full machine backup to simply turn back on from date X/Y/Z. Asigra actually takes a native ESX snapshot at the time of backup.

Using Asigra Televaulting (or a small handful of other backup systems) allows us to perform backups that are compressed and bit level.  So, if I have a VM that has a 20GB vmdk disk, but really only has 8GB of data in the guest system, then the backups will actually be less than 8 because all that free space will be compressed down to nothing and then the 8GB of real data will further be compressed down.

My schedules are now setup to backup real data each night, but backup the OS (again using the native ESX snapshot) once a week.  For the most part, system settings and configurations aren’t happening each day. So, if the guest VM dies, I can simply restore the VM to the last weekly backup (as of at the most 6 days) then restore the real data to to that machine (as of at the most the night before).

A note for Asigra users - I’ve had a lot of problems backing up via VI.  So far attaching directly to the ESX hosts seems to be working great.  The one fallback for this is you need to setup rules to ensure that VM’s stay on the same ESX host.  Asigra only knows that a guest VM is on the host you originally configured it to be backed up from.

So I was having some pretty significant performance problems with a vm running Windows Server 2003.  I thought it might be due to the fact that I had so many snapshots.  Fellow ESX admins over at the VMWare communities confirmed this to be the case.

So my next step would be to combine all the snapshots and get rid of my delta’s by committing them all. I was going to run some guest updates first, so I again as always, made a snapshot. Something either guest or host related happened and the guest powered off.  When I went to turn it back on VI complained:

too many levels of redo logs

uh oh! the guest wouldn’t turn on!  it turns out that this snapshot was in fact the 35th snapshot taken for this machine. This 35th snapshot didn’t complete correctly and was corrupted. it also turns out there is a 32 snapshot limit for VM guests.  Good to know VI tells you about this! <sarcasm included>

After frantically searching the web and forums for a solutions someone pointed me in the direction of a post here: http://zealkabi.blogspot.com/2008/10/virtualcenter-shows-no-snapshot-but-it.html which clearly shows the process i need to use to commit my snapshots, specifically Solution B:

If solution A did not work then next step to follow is: use vmkfstools -i to consolidate snapshots.
1. You can export the disk with vmkfstools to recreate the virtual machine:
2. Execute the following command to create a directory for the new disk:\
# mkdir /vmfs/volumes/UUID/new_RHEL5
3. Execute the following command to point vmkfstools at the last snapshot file:
# vmkfstools -i RHEL5-000001.vmdk /vmfs/volumes/UUID/new_RHEL5/new_RHEL5.vmdk

Three hours later, snapshots 32 through 1 committed and a single vmdk.  While this commit process was running I realized i could have simply told VI to run vmdisk00032.vmdk instead of the final (and corrupt) vmdisk00035.vmdk.  this would have been the quick resolution to get me back up and running, and I could have don the snapshot committal at a better time.

Huge thanks to patrickds from the VMWare communities and SANJAT KABI (http://zealkabi.blogspot.com) for their knowledge!