Postgres Cluster over ZeroFS

[overflowz]

I recently had this idea, although I'm not yet sure if it's realistic or reliable. Just asking a question if someone has any comments.

How about running a Postgres server on ZeroFS with single writer and multiple readers? They all would use same data store (ZeroFS) and ZFS with L2ARC.

This would definitely require some tweaks with postgres to access the same data directory by multiple instances, etc. but I'm not exactly sure where to start / what else might be needed.

Does this sound crazy/unreliable or worth a shot? I don't have time at the moment, but whenever I do, I might start experimenting.

Cheers.

[Barre]

I think the easiest path towards something like this is having n large-ish Postgres nodes, a single ZFS pool and a single active Postgres node. When a Postgres node goes down, one other Postgres node takes over ZeroFS, imports the ZFS pool and starts to serve queries.

I guess it depends on why you want multiple readers, if it's not for scalability but for uptime, then that approach certainly makes sense.

[overflowz]

Thanks! So it does make sense and I'm a sane person, right? 😅

[Barre]

I think it does make sense yes.

I myself prefer to avoid going distributed when I can :)

So I'd probably settle on larger instances with large L2ARC to serve reads, which should be enough for the vast majority of workloads.

[overflowz]

my thinking was to be used inside kubernetes. having cheap servers (or decent at least) and whenever you need scalability, you just add another server and should be good to go (would be easier to maintain, plus high availability). While upgrading to bigger server and having another one for HA would require more advanced things to consider.

[Barre]

I think a good start would be ZeroFS to export a VFS interface that can be used from go, so that we could write a proper CSI driver.

[overflowz]

sadly, I don't follow there, I'm not experienced in that field, but I got the point.

thank you!