What does your backup do, Sam?

What does your backup do, Sam?

We have been working on our postgres backup routines lately and spent a lot of time convincing ourselves that we have a working setup. This actually took a considerable amount of time, because we had a lot of skepticism after our recent firedrills. We think that skepticism is probably the healthy approach to backup routines anyway. Most people know that backups are important, but putting a lot of work into it has given us reasons to think about what problems that backups solve.

We aready mentioned in a previous blogpost that we have a replication server set up, and hopefully the chance that both the master and the standby go down are pretty low. So in a normal outage, we hope to be able to restore services by promoting the standby to master instead of doing a full recovery from backup, which will take longer. But there are scenarios where our standby won’t save us. The recent heartbreaking gitlab database incident is a good example. Replication had stopped in this instance and the activity was actually to get it up and running again.

Another scenario would be when your standby and master are both in the same data center. In this case, maybe a network outage would make both inaccessible and give you a complete loss of service. At this point, if you had backups, you could create new database servers. Obviously, that requires your backup to also be available off-site.

But there are more subtle type of problems where a standby can not help you. One example would be when a developer introduces a bug in an application that causes it to overwrite valid data with nonsense. In this case, the standby is just going to happily write the changes that the master does and only a backup or a log file could help you recover the lost data. Another example might be a script or person executing a database query that drops a table on the wrong server or in the wrong environment.

We’ve been thinking a little about how we could handle such outages and have some ideas of our ability to handle them, but we haven’t tested ourselves yet. An exercise that we want to do, is to delete a table from a database in QA and attempt to recover that data without losing any transactions after the table is dropped.

For example:

  1. Everything’s working fine
  2. Developer makes mistake, dropping a table
  3. Developer goes to lunch
  4. Users keep creating traffic and transactions
  5. Developer comes back from lunch, notices problem
  6. ???
  7. Developer goes home after work

We think we could manage to sort out the above incident. Our current idea is that we would use our backup and our WAL archive to do a point in time recovery to time 1. We wouldn’t do that on our current master database, because would cause us to lose the transactions between time 1 and time 5. So we’d set up an entirely new database server from the backup instead. From this new instance, we can do a full dump of any relevant tables using pg_dump. Hopefully, we can then import the generated SQL to our master database.

This is a problem you can’t really solve with a standby.

In our setup, a backup job runs pg_backup in the middle of the night. Our master server has an archive_command that it uses to store WAL segments. Both the basebackups and the wal archive are stored on and off-site, so we’ll have access even if the network in our data centers are down for some reason. We have also configured wal_keep_segments, because our backup-tests revealed that a database set up from backup was not able to start streaming replication without it. We don’t fully understand this, as all the required WAL segments are present in archive.

The first step of what a developer could be doing at time 5, would be to set up a new postgres server. They can do that by fetching the latest basebackup from either on-site or off-site, and extract it on a server. The next step would be to create a recovery.conf, setting it up with the correct restore_command to extract WAL segments from the archive, and set the recovery_target_time right before time 2.

Starting this postgres instance should then produce a database server with the same state as what was in the master at time 2. Depending on the amount of WAL segments in the archive, this could take a while.

When it’s done, the developer could use pg_dump, providing the correct --table argument and database names, which should produce a .sql file containing INSERT statements for the missing data.

They can then replay the remaining WAL by setting a later recovery_target_time and restarting the fresh database instance, which should provide them with a database server where the table has been dropped. They can use the fresh database instance to test that importing their dump fixes the problem, or at least doesn’t make it worse. After they’ve tested it and verified that it works, they can do the import on the master server and go home after work. Or that’s the idea.

A recovery is probably harder if the table has been truncated or updated wrongly, which might mean someone has to set up some sort of data reconciliation. But it might still be possible.

Being confident in our backups

One of the things we’ve done in order to be confident in our ability to restore from backup, is to set up an automated job that creates a database instance from backup every day. This database will then connect as a standby to our current master and verify that it can start streaming replication. We throw away this database after it’s been verified that the backup is good. It’s a really nice feeling when the backup-test script posts on slack that it has successfully done a point-in-time-recovery of a backup:

Making this backup-test script was a pretty simple job. It also serves as living documentation of how to do a recovery procedure. And since it’s being run every day, we can be confident that the recovery procedure actually works, unlike some 3-year old recovery procedure documentation on a wiki. A nice side-effect of this is that we actually know how long it would take us to recover from backup.

We haven’t yet completed an automated test for our off-site backup. In principle, it should work exactly the same way, but the wal archive is a different source so the recovery procedure is actually slightly different. We need to do an exercise to determine if we need to write some tooling around this recovery procedure to feel confident about it.

We do want to run firedrills on recovery situations. The table-drop scenario recovery is an example of a recovery job that is more complicated than the test that we run every day, so it would be good to do it manually a few times to verify that we can recover the data in such situations. And it could help us find the limits of what we are actually able to do. Maybe we can do more than we think.

The value of firedrills

In January 2016, we set up a postgres cluster at bring. While doing the initial configuration, we designed what we thought was a pretty decent backup architecture and we set up streaming replication to a hot standby. While configuring the cluster, we made sure to verify our ability to recover from backup, as well as our ability to fail over from the master server to the standby server. This gives us redundancy in case of outages or patching, and it gives us recoverability in case of data corruption, or so we thought. We documented the procedures and all was well. At this point our postgres installation was on 9.4, and we upgraded that to 9.5 to get some new features before we really started using it.

We ran a firedrill in November where we performed a failover in our QA environment. That failover exercise was not a success, because streaming replication had stopped two weeks prior, and we hadn’t picked that up. Cue introducing better monitoring. We set the standby up from scratch again. In January 2017, we scheduled another few fire drills and in the first one, we encountered some difficulties that we learned a lot from. We spent the better part of two work days figuring out these problems, what follows is a short summary of what happened and what we learned.

Our postgres master is set up to archive its write-ahead log (WAL) to a location that is shared with the standby and the standby receives streaming replication from the master in a typical setup.

On performing the failover, we first stopped the postgres master server, ran service postgresql promote on the standby server and updated a DNS record. So far, so good – there was 45 seconds or so of downtime, as seen from the applications and everything came back up fine. Then we started working on setting up the former master as a standby server, so we’d be back to where we started and this is where we started running into trouble.

When we wrote the failover documentation, it seemed that in order to use a former master as a standby, you had to run a pg_rewind.

Our first sign that something was wrong was the unexpected output no rewind required, which we later learned is actually a good thing. Our instructions said that we’d find a recovery.done file in the postgres data directory, which we also didn’t find. Our failover documentation had expected pg_rewind to copy this file for us. At this point, we became aware of the botched failover attempt from November, and went off trying to investigate whether it had created some weird state in the cluster. We discovered that we had 5 timelines on our master server and had no idea what that meant. We created a recovery.conf manually, and attempted to start the former master as a standby. That failed, with errors in the log that the server was unable to retrieve timeline 4. After a while, we discovered this file (00000004.history) in our WAL archive and moved it to the new master. This time, the standby did indeed read timeline 4, and also timeline 5, and started trying to read WAL.

This failed repeatedly because it was attempting to retrieve a WAL segment that the master had already archived - 000000050000009A00000059. We attempted to copy this segment from the WAL archive to the master, but that didn’t help. At this point, we started wondering why the standby wasn’t simply executing its restore_command – after all, the segment was present in the archive.

We noticed that the standby logged that it started streaming from timeline 3, while the master logged that it was writing to timeline 5. We added some logging to our restore_command and discovered that the standby was actually trying to retrieve 000000030000009A00000059 (notice that 3), and not 000000050000009A00000059. These have the same WAL segment numbers, but are on different timelines. At this point we discovered a file named 000000030000009A00000059.partial in our WAL archive and started worrying about data corruption. As a last-ditch attempt for the day, we removed the .partial suffix from the file, the standby recovered it but still couldn’t start streaming replication again.

After hours, we read a bit about postgres and timelines and in the morning we tried to set up the standby by adding recovery_target_timeline='latest' to our recovery.conf. At this point, the standby was able to retrieve 000000050000009A00000059, but no other WAL files from yesterday were still available, our backup job had made a new pg_basebackup and cleaned out the old WAL.

We decided to set up a completely fresh standby from our backup and got our second nasty surprise in 2 days, when it turned out that we were missing the very first WAL segment after the pg_basebackup because of a bug in our backup-script. At this point, we set up a fresh standby from the current running master with pg_basebackup and slept very poorly.

We now know for sure that our problem with setting up the former master as a standby was due to not knowing about timelines. What we needed to know:

  • When a standby is promoted to master, it creates a new timeline
  • It will archive the last WAL segment of the old timeline with a .partial suffix
  • The .partial segment file is harmless and you’re not expected to need it
  • A postgres instance that starts up with a recovery.conf, will by default attempt to stream to the timeline it was on when it was shut down
  • Using recovery_target_timeline lets you control this behaviour
  • Using pg_rewind is not necessary when the former master shuts down gracefully
  • In effect, what pg_rewind does is to discard transactions on the old master, which the new master does not know about

In our case, we were on timeline 3 when we started. Because of the botched exercise in November, we also had a timeline 4, but it wasn’t in use anywhere. So when we promoted the standby, we created timeline 5 which started at the same WAL segment number as timeline 3 ended on.

You can learn even more about timelines, WAL and postgres standby setup in this talk and this talk by Heikki Linakangas, the author of pg_rewind (note: these are kind of scary).

After the firedrill, we made sure to keep WAL segments around for 7 days. We obviously documented recovery_target_timeline. We also set up testing of our backups, so that their viability for backup recovery can be automatically tested every day. We ran a few firedrills after this one, and those have worked out fine and we’re now pretty sure that we have a much better understanding of how WAL works.

We’re super happy that we did this firedrill. Not only do we now know a whole lot more about how postgres works, but we also feel pretty confident that we have a working backup and can sleep much better at night. Firedrills in general are awesome, and this one in particular helped us close a lot of really bad problems in our database setup.