Let me just say:

Galera is bloody incredible. We had 2 out of 3 nodes crash, and it still managed to recover automatically with no downtime.

But let me also say

When it *does* fully crash... Data recovery is an _incredible_ pain in the arse.

Thank you, Galera. Wish more customers were willing to pay for 3 SQL nodes instead of just two while expecting minimal node downtime...

WAL logging is incredibly useful for recovery as well as state sync ;)

On ZFS ZIL on a dedicated SLOG gives you a solid recovery point in case of power failure. :)

How much 3 nodes cost ?

Would love to compare with my current setup With Azure SQL.

(Right now, 0 crashes and 0 data lose in 5 years. Not ONE restaure in 5 years)

Edit : to be fair, databae size is pretty small : 45 GB

@NoToJavaScript the software itself is completely free (If you don't need enterprise support), and only really need a reliable network link between the nodes as they constantly communicate with each other to keep the data consistent.

Then it's just a question of QPS you want out of your DB.

But for a DB that small, I can imagine you could make do very well with a 12c/24t cpu, 64 gigs of ram, and a speedy ssd or nvme based storage in raid1 or even 10 for those extra fast reads and not too bad writes

The software we use is MariaDB community edition and Gallera extension.

@Aldar thanks !

On my end right now using a "servreless" (Always funny this word) Azuer SQL tier. With auto scale between 4 and 12 vCpus. With geo replication in a paired region.

Average monthly cost 3.500$. I find way too expensive. But at realibility side, it's awsome with 7 days restore in point with 1 ms precision. Honestlly, I would rather pay less and have 1 minute retrore in time window.

@Aldar

Oh I checked. The machine you escribe would cost (see screenshot). And ofc, it is not replicated automaticlly

@NoToJavaScript I'd have to check for the PIT recovery, the DBs we run on Galera are about a TB in size. And our clients are fine having just daily backups.

And, sadly, Gallera cannot really be used for cross-region replication as it relies on fast operation acknowledgments between nodes, and if it doesn't receive it in time, it can one - kick the node out of the cluster, and two - block the whole DB as it waits for a consistency.

BUT, you can still replicate the normal "log-shipping" way via bin logs from one of the master nodes... Though the classic kind of replication is much less resilient and can break if you don't set it up correctly.

@Aldar Thanks !

Seems like a pretty good tool ! But with our small DB, I think for now it is not needed.

But I’ll keep it in my “tool box” !