Now that we’ve finally finished the schema change release, I wanted to give an account of what happened in this arduous process. Before I dive into the details, I want to offer a picture that best sums up our current situation and challenges:
The shipping container is MusicBrainz and the boat is our hosting infrastructure. This picture perfectly describes the sort of challenges we’ve faced over the past few days. 🙂
Here is what happened:
Because the site was recently running slow and our search servers kept crashing, Zas and I were not available to help Bitmap prepare for the schema change release. This long process was left to Bitmap and Gentlecat to take care of on their own. We quickly realized that we were not ready for the release when the due date came and thus we delayed one week.
Sunday 22 May
Finally we were ready to proceed with the Postgres 9.5 upgrade. Once we started the process, we kept running into small problems that we didn’t get in our test setups. We do not have access to enough infrastructure to have a complete clone of our production environment, so we can only do so much to prepare for all the things that might happen when we run upgrades on our production servers.
All the while we attempted to start the upgrade, our backup database server was running much slower than anticipated. In the end we figured out that a step for optimizing the database (analyzing it) wasn’t carried out. During this time the site was really slow/unusable, but by the time the problem became apparent we had started the upgrade and could not turn back.
Once the upgrade was done, optimizing the database took much much longer than usual: 3 hours! This process wasn’t started until about 1am local time, which made for a very long night before that process finished. And even then we hit snags and had to start over a couple of times. At about 4:30am we had the site running on Postgres 9.5 in read only mode. The plan was to rest and start the schema change release in the morning.
Monday 23 May
Of course we had spent all of our time working on the Postgres upgrade and site stability, so our document that we use to plan the schema change was not in place. We spent the day preparing this and other bits for the release. To get an appreciation for what this document looks like, have a look! Note that some steps could be instant, others might take hours to carry out. Others might involve a sub-step or 20 not included in the document.
In the evening we were ready to make the change. By this point our backup DB was performing much better, so the read-only site worked acceptably. Thus, we started the release. Overall, the actual release process was reasonably smooth – we hit a few snags and had to do a lot of waiting for our slow servers. At about 1am in the morning things were finally complete. We proceeded with our sanity checks to make sure things went smoothly and all of them passed.
We proceeded to put the site into read-write mode and immediately saw portions of Postgres crashing, which is really bad. With community feedback we quickly deduced that some write operations were causing Postgres back-end processes to crash. We went back to read-only mode on the site and things stabilized and we finally went to bed at 3am.
Tuesday 24 May
In the morning we quickly found the source of database trouble with the help from the Postgres people on IRC. Thanks for the swift help Johto! We found that the steps for installing the updated third party extensions into Postgres had not completed correctly. Repeating the steps by hand fixed this problem.
Sadly yesterday morning we got an email informing us that our Live Data Feed replication stream had become corrupted. 😦 This was heartbreaking news to us, since it means a great inconvenience to all of our Live Data Feed users. We immediately split into two teams: Zas, chirlu and myself to fix the root cause of the issue and Bitmap to investigate fixing the stream.
I proceeded to setup a test environment was able to quickly reproduce the problem. Zas and chirlu were an amazing support team Googling issues as I came across them. Within fairly short time we fixed the problem and deployed the fix to our database server. The problem was caused by a bug in a piece of code that we’ve been using for 13 years! A change in Postgres caused this bug to actually become a problem and corrupt our replication feed. 😦
Once the problems were fixed we needed to initiate a new data dump and check to make sure the replication stream is working correctly. Of course we found a problem that we fixed and re-started the process to dump the data. Loads of hurry-up-and-wait situations to try our patience!
When we were satisfied that things were working correctly we re-enabled the site as read-write at about 1am and allowed people to continue editing. Exhausted we stumbled into bed waiting for data dumps to sync out to the FTP site.
Wednesday 25 May
Today Bitmap was flying home and as soon as WiFi became available on his flight he started working and helping with putting the schema change to bed. We’ve verified that everything is working as expected. At last this saga comes to and end and we can all take a break and catch up on sleep!
Thank you for your patience through all of this.