Today, The Storehouse experienced an outage that lasted approximately 12 hours. This was caused due to updates performed late the night before and services restarting during that process.
Last night I ran upgrades of the servers that run The Storehouse, including our three ProxmoxVE nodes. When the upgrades on these nodes were complete, the nodes had an updated kernel version and needed to restart to use the new kernel.
Restarting is usually a painless process in our environment. Since the nodes are built with redundancy, one can be rebooted or offline at any time without disrupting service. Usually, we perform the rolling upgrades in the background and Storehouse customers see no disruption of affects on services.
The nodes also co-host a GlusterFS volume that we use for customer data and a few critical virtual machines. We run two pfSense routers as VM’s and store the disks on there to allow the routers to migrate between nodes. These routers (like most) are critical: without them our environment is not accessible to the outside world.
When restarting the PVE hosts for the upgrade, I do them one at a time with a delay between each to allow the Gluster volume to heal. Last night when I went to upgrade nodes, I did not wait long enough and the volume was not fully healed.
This wasn’t an issue with customer data, since no customers were uploading data during the upgrades. The router disks, however, became inconsistent and both went offline around midnight. When both went offline, the outage began.
Our monitoring system alerted me to the issues, but I slept soundly through the alerts. Due to scheduling issues in the morning, I was unable to return to the data center until 12. At that time, the routers were restored from backups and restarted, restoring service.
1. Get a backup connection.
One of the reasons it took so long to restore service was that I was unable to travel to the data center to physically connect to the network. The system relies on at least one of the two routers running at all times. When both are offline, a trip to the data center is required to fix it. While this isn’t a big deal, it takes extra time to pack and travel, and requires that I am free and able to travel.
2. Check GlusterFS during rolling upgrades.
If I had looked at the Gluster volume status, I would have noticed the inconsistent files. This isn’t an issue, as Gluster automatically will heal the files, but both nodes need to be online to do that.
3. Get better night monitoring.
It certainly didn’t help that I slept through the alerts streaming into my phone. The monitoring system was working perfectly: our offsite poller noticed the outage and was quick to alert me.
All in all, human error caused this outage.