Thread: serializable master and non-serializable hot standby: feasible set up?
Howdy!
I've been going back and forth with the GCP CloudSQL engineering team about the feasibility of a particular setup, and I'm pinging the list here hoping for a sanity check. They assure me that it's impossible and I think they must be mistaken, but I have limited experience administrating my own postgres instances. So I'm appealing to a higher authority :)
The docs outline[1][2] that a hot standby / replica with the flag `default_transaction_isolation='serializable'` is going to cause issues -- while you can connect to such an instance, basically every query against it will fail.
But! If you can somehow manage to get the replica's flags to instead use `repeatable read` or `read committed` isolation, everything seems to work well, even if the master uses `serializable` isolation. In GCP, we are having to temporarily swap the master to a lower isolation level, then stand up the replica and pin the lower isolation level flag, and then finally revert the flag change on the master. If the replica goes down, we have to repeat this process and it's a pain (not to mention data issues since our app relies on this isolation level instead of doing explicit locking in most cases).
So I know this is an awkward question to post here, but as postgres admin professionals, *would you* expect to be able to stand up a `repeatable read` replica against a `serializable` master? My expectation is that you'd simply change the setting in a .conf file on the replica and be good to go; is there something that would make this process really difficult / impossible?
Thanks so much!
On Tue, 2024-10-15 at 16:27 -0700, Jacob Biesinger wrote: > *would you* expect to be able to stand up a `repeatable read` replica against a > `serializable` master? My expectation is that you'd simply change the setting in > a .conf file on the replica and be good to go; is there something that would make > this process really difficult / impossible? I expect that to work fine, at least I cannot think of a problem with such a setup. But I have been wrong before, so test it. Yours, Laurenz Albe
Re: serializable master and non-serializable hot standby: feasible set up?
From
Jacob Biesinger
Date:
Just out of curiosity, what’s the use case for this?
We use the read-only replica as a data source for analytics workflows. The setup allows us to have fresh data without affecting performance of the production environment.
Re: serializable master and non-serializable hot standby: feasible set up?
From
Jacob Biesinger
Date:
On Tue, Oct 15, 2024 at 9:23 PM Laurenz Albe <laurenz.albe@cybertec.at> wrote:
On Tue, 2024-10-15 at 16:27 -0700, Jacob Biesinger wrote:
> *would you* expect to be able to stand up a `repeatable read` replica against a
> `serializable` master? My expectation is that you'd simply change the setting in
> a .conf file on the replica and be good to go; is there something that would make
> this process really difficult / impossible?
I expect that to work fine, at least I cannot think of a problem with such a setup.
But I have been wrong before, so test it.
The setup (serializable master, repeatable read replica) definitely works -- we've been running that way for over a year now. I guess I'm really asking "how would you go about getting the replica into the appropriate state?" Would you expect to have to downgrade the master's isolation level as I describe? Or would you expect to be able to stand up the replica using a modified conf file initially?
Thanks as always for your help!