[agwa]

> If you let a disk run full weird shit happens.

Only in buggy software that ignores errors from system calls.

Obviously you can expect availability problems when you run out of space, but there's no excuse for losing data from committed transactions given that the OS will reliably report the error.

> So I do strongly hope that besides changing software, they added some disk space monitoring.

One of the action items in their incident report was improving monitoring and alerting: https://groups.google.com/a/chromium.org/g/ct-policy/c/038B7...

[evanelias]

Keep in mind the report doesn't mention what MariaDB version was being used, nor what storage engine. The description in the incident report simply doesn't align with how InnoDB behaves, at all. (InnoDB is the default storage engine in MariaDB and MySQL, and by far the most widely used storage engine in these DBs.)

For example, the author says "Whilst its primary key index still includes the ~1,300 unsequenced rows, the table itself no longer contains them." This is simply impossible in InnoDB, which uses a clustered index – the primary key B-tree literally is the table with InnoDB's design. The statement in the incident report is complete nonsense.

Also if the storage volume runs out of space, InnoDB rolls back the statement and returns an error. See https://dev.mysql.com/doc/refman/5.7/en/innodb-error-handlin... (MySQL 5.7 doc, but this also aligns with MariaDB's behavior for this situation)

So maybe they were using MyISAM, which is ancient and infamously non-crash-safe and should pretty much never be used for anything of consequence, which has been well-known in the database world for 15+ years. Who knows, they didn't say.

The report directly says "Our teams have very little expertise with MySQL/MariaDB", which also makes me question their conclusion "that it is not possible" to restore/recover.

Also this whole incident happened to Sectigo, take from that what you will.

The one thing I'd criticize MariaDB for here is that they're lacking an additional safety mechanism which MySQL has: by default MySQL will shut down the database server if a write to the binary log cannot be successfully made, ensuring that any writes on the primary can be replicated as well. (iirc this was a contribution from Facebook.)

[bawolff]

> Only in buggy software that ignores errors from system calls.

Is there any reason to conclude that is what happened?

[mobilemidget]

Please enlighten me with some software examples that are free of bugs :)

[hiciu]

It's not about being bug-free, it's about being crash-safe. You should be able to crash your software - either by pulling power cord from the server, killing process or turning disk read-only (similar to what happens when it's full), and your database should still survive.

Data corruption like that should just not happen. You have your journaled filesystems, you have your raid, zfs pools, and whatnot; all that is worth nothing if your database software can just say "I have encountered an error during write operation and now your data is inconsistent, good luck". This is exactly what journaling / write ahead log / innodb doublewrite buffer should prevent.

There is an article "Crash-only software: More than meets the eye" on lwn.net if you would like to read more about it. The postgres fsync bug is also vaguely related to the same issue, also worth reading.

[mobilemidget]

Dont get me wrong, I agree it shouldn't happen. But that is not really my point, I also say you should not be surprised. And you kind of proof my point 'postgres fsync _bug_ is also vaguely related to the same issue'. Which would make the switch to PostgreSQL because of this, which could have been avoided by monitoring, even less strong choice. Who says there isn't a fsync like bug there still?

Corruption, empty files etc have happened before in case of exhausted disk space, with PostgreSQL, mysql and so on, so place and use additional safeguards, monitoring of disk space in this case. Because even if the data wouldn't be corrupted, a full disk leads to a failed service anyway which you want to avoid too. And even in case the table didnt get corrupted, new inserts, updates will fail, which might lead to 'missing' data too, which imho is also a sort of corruption.

Cars have airbags to avoid you getting seriously hurt in case of a crash. But dont be surprised if you do get serious hurt in case of a crash. Additional sensors that alert to the driver of a car to avoid accidents are important too.