Continuous MySQL backup validation: Restoring backups | Hacker News Reader

Continuous MySQL backup validation: Restoring backups

code.facebook.com

104 pointsjivid27 comments

> We do these by taking another full dump and only storing the difference between it and the last full backup that was taken.

Do you diff against the same base, or create an incremental chain? How many diffs do you take in between recapturing a full image? At $DAYJOB we always take full backups into a fast in-house deduplicating store.

> Periodically, each peon syncs with the ORC DB to look for new jobs assigned to it

Is there no better way to handling this than polling?

> LOAD - Load the downloaded backup into the peon’s local MySQL instance. Individual tables are restored in parallel by parsing out statements pertaining to those tables from the backup file, similar to Percona's mydumper.

Presumably you can only get this parallelism by disabling FK integrity. Is it re-enabled in the following VERIFY stage?

> Do you diff against the same base, or create an incremental chain? How many diffs do you take in between recapturing a full image? At $DAYJOB we always take full backups into a fast in-house deduplicating store.

We always diff against the same base and have 5 days in between subsequent full dumps. The number of days just comes from a trade off between space occupied by the backups and time it takes to generate them.

> Is there no better way to handling this than polling?

There's definitely different ways to approach this, we find polling works well for us. We also use the same database for crash recovery, so doing the assignments through it serves both purposes.

> Presumably you can only get this parallelism by disabling FK integrity. Is it re-enabled in the following VERIFY stage?

I'm not sure what you mean by parallelism through disabling FK integrity. Splitting the backup into its tables means we can restore a subset of tables instead of the entire backup. This allows us to load individual tables concurrently, but also not have to wait to load a massive database if all we need is a few small tables.

> I'm not sure what you mean by parallelism through disabling FK integrity.

Say you have a `user` table and a `post` table with `post.user_id` being a FOREIGN KEY on `user.user_id`. Without disabling FK integrity you would not be able to restore a post without restoring the user first. When restoring in parallel this might or might not work out.

Facebook (along with almost everyone else using MySQL at massive scale) doesn't use foreign keys.

They scale poorly in MySQL, and they lose a lot of purpose in a massively sharded environment anyway. For example, say you like a status post on Facebook, or friend another user. It's very unlikely that the liked status or friended user exists on the same shard as your account, and there's no way to enforce a foreign key relationship in an inherently non-distributed database like MySQL.

So instead integrity is handled at the application layer, with additional background processes to fix the occasional integrity problem and detect integrity anomalies.

I understood it to mean that if you restore table A and table B in parallel, if there is a foreign key between them, then referential integrity checks would cause one of the loading operations to fail. How do you deal with that?

It's likely that foreign key checks aren't handled at the RDBMS level, but rather at the application level.

Why mysqldump instead of Percona's xtrabackup/innobackupex? I was under the impression that the latter had a lot of advantages and have been considering switching, wondering if there is a good reason not to.

We take logical backups, not physical, and mysqldump is the best option for that. Having logical backups means we can do logical diffs as well (see https://youtu.be/Fe2oLZ4CWD4?t=951), and we've added table checksum support to mysqldump in our branch of MySQL (https://github.com/facebook/mysql-5.6/commit/54acbbf915935a0...)

Is the FB branch version of mysqldump still single threaded? How do you cope with that?

I currently "fake it", using "START TRANSACTION WITH CONSISTENT SNAPSHOT", with multiple mysqldump processes running, where I can't get mydumper deployed.

We run a single mysqldump with --single-transaction for each database on the server, nothing special.

Thanks, how big is a single instance? All out to a single file?

Instance size varies a lot because we've got a lot of different MySQL workloads, some with very different configurations.

Remember we're backing up each database separately though, not the entire MySQL instance at once. Each database's backup is in a separate file.

Thanks

xtrabackup does differential backups as well...is there another advantage to doing logical backups? are the diffs significantly smaller?

Yes, logical backups are smaller due to lack of index overhead. And since logical backups are textual, they can also be used for other clever purposes, such as ETL pipelines.

In addition to what evanelias said, a logical dump also means we can load it into a MySQL instance running a different storage engine as well. In our case, it allows us to take a mysqldump from an InnoDB instance and load it into a MyRocks instance if we wish.

Interesting, though I was hoping for a bit more focus on the validation part - unless I'm missing something, there's nothing protecting the actual SQL dumps against bitrot.

I had to design against that particular problem recently, ended up taking a pt-table-checksum at the time-of-dump and verifying newly restored backups against that to ensure the backup's integrity.

Unfortunately that requires halting replication temporarily, so I was hoping to hear of a more ingenious solution.

There's a few different ways to verify, a few of them involving stopping replication, like you pointed out. These can also sometimes be quite expensive, so depending on the type of verification required, the verification method can be tuned.

We also implemented table checksums inside of mysqldump, allowing us to dump out the restored data and compare checksums as well, if required. https://github.com/facebook/mysql-5.6/commit/54acbbf915935a0...

Very nice, thanks for the link!

I'll definitely be seeing if I can replace my system with that; reducing the overhead to a single transaction would be a big win.

Interesting. I assumed FB would have hot replicas (kept up to date with live master-slave replication at all times) - ready to go any time a main database fails. Cascade that to another layer of slaves and there's no need to restore anything ever.

Per their post about their MySQL backup solution[1], it sounds like they sort of do this? Although it's not a slave that's ready to be promoted to master, but somehow a cold standby that is not far off from being a hot standby based on how recent the binlogs are?

The wording is a bit cryptic, but it does seem that they definitely have hot standby-esque capabilities in place, in addition to long-term storage of their incremental/full backups.

[1] - https://www.facebook.com/notes/facebook-engineering/under-th...

They do, but that doesn't help you if bad code writes corrupted data, or other unexpected disasters which take down an entire replica set at once. For large public companies "I didn't think that failure would happen!" Isn't an acceptable excuse.

This is only a useful system if you can trust that failed databases simply disappear.

In failure modes like accidental delete without WHERE clause, or a write that corrupts the validity of the business logic, it's useless, as you can watch the logs to see all your slave machines keenly and unquestioningly repeating the issue.

We had that at my previous job, fortunately for some reasons the slave was delayed by like 10min, it saved us from a very very serious problem, especially when you do backups every 6h.

Facebook has hot replicas in every region. But replicas and backups serve completely different purposes.

Replicas are for failover and read scalability. In terms of failover, when a master dies unexpectedly Facebook's automation fails over to promote a replica to be the new master in under 30 seconds and with no loss of committed data.

Backups are for when something goes horribly wrong -- i.e. due to human error -- and you need to restore the state of something (a row, table, entire db, ...) to a previous point in time. Or perhaps effectively skip one specific transaction, or set of transactions. Replicas don't help with this; as you mentioned, they're kept up-to-date with the master. So a bad statatement run on the master will also affect the replicas.

Occasionally you have some massive failure involving both concepts, like you have 4 replicas and they're all broken or corrupted in some way, then backups are helpful in that case as well.

I suppose at Facebook scale it might be infeasible, but couldn't you get the same effect by archiving log segments and a periodic binary full backup? This is precisely what I do with my PostgreSQL databases (though with some friendly automation with pg barman), I assume you could do the same with some tooling around MySQL's binlog facilities.

Yes, although if using the binlogs as-is, that's effectively incremental backup instead of differential. The disadvantages of incremental solutions are that they require more storage and take longer to restore (especially if only doing full backups every few days); the upside is less complexity.

Reply on news.ycombinator.com