On February 1, 2017, in an effort to resolve an issue sparked by hackers, GitLab suffered a major outage when a staff member accidentally executed an
rm -rf command on the wrong server.
As a result, around 300 GB of live production data was deleted and the GitLab service became unavailable for users of the free service offering. Fortunately, the outage did not impact (paid) Enterprise customers and they were not affected by the outage that quickly became public knowledge.
As you may recall from my series on RandomGenerator, I was one of those impacted. I started noticing issues before I realized there was a situation. At the time, I thought I had made a mistake with my experimentation of the Pipeline options, since my builds were not starting at all. I probably should have known the problems I was experiencing were out of my control when my master branch would not rebuild without any changes being made.
Timeline of Events
rm -rf command, while the cause for a majority of the issues, was actually being performed to resolve an issue that occurred when a group of hackers was trying to place a heavy load on the GitLab servers. Below is an abbreviated timeline of what happened:
01/31/2017 @ 06:00 PM (UTC): Hackers used snippets to place a heavy load on the database, causing downtime.
01/31/2017 @ 10:00 PM (UTC): Database replication could not keep up, causing replication to stop working.
01/31/2017 @ 11:00 PM (UTC): Attempting to fix the database issue, the
rm -rfcommand is executed against the wrong server, removing 295.5 GB.
02/01/2017 @ 12:00 AM (UTC): Begin recovery process.
By early afternoon (in the United States), GitLab announced via Twitter that "GitLab.com should be available to the public again." When I tried re-executing a build on RandomGenerator a few hours later, everything was working as expected. This is because my changes were made before the last valid snapshot was captured.
While the trigger of this whole process was based on hackers attempting to put a heavy load on the database, the result exposed issues with GitLab's backup and recovery processes. Prior to the outage, GitLab touted having five backup and replication techniques in place to provide redundancy in the event of an unforeseen circumstance. When put to the test on February 1, none of the five were reliable, and the restoration centered around a snapshot that someone just happened to take six hours prior to the outage. As a result, items that were made after that one snapshot was created were lost indefinitely.
The fallout from the outage, while not impacting (paying) Enterprise customers, is that GitLab wasn't able to maintain the data that was being stored on their service. I can only assume that at any given time, there are corporations and individuals who are evaluating GitLab as a potential solution for their needs. Certainly, the outage and inability to recover data will impact these potential customer's views.
Earlier in my career, I worked in the insurance industry as a network programmer. There were two times a year when we participated in what we called Disaster/Recovery exercises. During those events, we arrived at an off-site location with backup tapes and documentation from the company that housed our backup information. In our room of the off-site location were servers and computers that matched our specifications to be able to recover from a series outage.
Keep in mind, this was back in the days of having mainframes, minicomputers, and microcomputer servers. There was a lot going on during this exercise. We had one day to get the systems functional and ready for testing by groups of testers who would arrive expecting their systems and data to be fully in place. Each time, we experienced minor issues, but each time, we improved and met the expectations of our customers.
From what I understand with the GitLab outage, it doesn't seem like GitLab is employing this type of disaster and recovery exercise. If they following a periodic recovery exercise, I am certain their issues would not be as drastic as the reality that was experienced.
The Future of GitLab
I do believe the GitLab can survive this situation, but it is going to require them to exercise and document disaster and recovery events or something along those lines. They are going to have to prove to the world that they have learned from this situation.
As noted above, the Enterprise customers were not impacted by the outage, but I wonder if the same issues would have happened to paying customers if the
rm -rf command was executed on one of their servers. If I was an Enterprise customer, I would be asking what steps they are taking to make sure this type of issue doesn't happen again — or at least looking at my contract to see what options exist to move toward another provider.