Nothing like a power outage gone wrong to test a new virtualization cluster. Last night we lost power in most of Corvallis and our UPS & Generator functioned properly in the machine room. However we had an unfortunate sequence of issues that caused some of our machines to go down, including all four of our ganeti nodes hosting 62 virtual machines went down hard. If this had happened with our old xen cluster with iSCSI, it would have taken us over an hour to get the infrastructure back in a normal state by manually restarting each VM.

But when I checked the ganeti cluster shortly after the outage, I noticed that all four nodes rebooted without any issues and the master node was already rebooting virtual machines automatically and fixing all of the DRBD block devices. Ganeti has a nice app called ganeti-watcher which is run every five minutes via cron. It has two primary functions currently (taken from ganeti-watcher(8)):

  1. Keep running all instances as marked (i.e. if they were running, restart them)
  2. Repair DRBD links by reactivating the block devices of instances which have secondaries on nodes that have rebooted.

The watcher app took around 30 minutes to bring all 62 VMs back online. The load on most of the nodes didn't go over 4 during the recovery which is quite impressive considering how much I/O its doing while VMs are booting. Normally the nodes have loads between 0.3 and 0.5. There were only 3 VMs that didn't boot cleanly because of incorrect fstab entries or incorrect kernel path settings in ganeti which was easy to fix. I was surprised we didn't have more issues like that.

While ganeti is bringing instances back online you can tail watcher.log which is generally at /var/log/ganeti/watcher.log and will show output similar to this:

2010-05-20 04:06:25,077: pid=10202 INFO Restarting (Attempt #1)
2010-05-20 04:07:16,311: pid=10202 INFO Restarting (Attempt #1)
2010-05-20 04:07:18,346: pid=10202 INFO Restarting (Attempt #1)

And once its finished will show output like this:

2010-05-20 04:35:04,066: pid=22741 INFO Restart of succeeded
2010-05-20 04:35:04,066: pid=22741 INFO Restart of succeeded
2010-05-20 04:35:04,066: pid=22741 INFO Restart of succeeded

It was great watching this system recover everything automatically with little issues and quickly. Needless to say, outages are a bad thing and its our fault that our cluster went down like this but it was great seeing this system work nearly flawlessly. We'll soon fix the power situation for our cluster so this shouldn't happen again.

Take that ESX ;-)


comments powered by Disqus