On Tue, Jan 16, 2024 at 10:59 PM Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
Dominique Devienne <ddevienne@xxxxxxxxx> writes:
> But sure, I take your point, you can emulate statement-level (implicit)
> rollback via an explicit SAVEPOINT, and ROLLBACK to the savepoint instead.
> But my point remains, that something like what that extension does should
> be an option of PostgreSQL itself, not an extension. --DD
> PS: I'd also be happy to hear why it's not, or won't be, on technical terms.
The reason it's not going to happen is that the community (or at least
the more senior developers) still remembers what happened the last
time we tried it.
We did implement server-side auto-rollback years ago in PG 7.3,
and it was enough of a disaster that we took it out again in 7.4.
Thanks Tom. That's insightful, and obviously something I didn't know.
The problem is that now you have a switch somewhere (whether a GUC
or something else, still a switch) that fundamentally changes the
transactional semantics seen by applications. Run an application
in the wrong mode and you have a broken app. Worse, there is an
awful lot of client-side code that now has to cope with both
behaviors. We thought that would be okay ... well, it wasn't.
It was a mess. It would be a bigger mess now if we were to try it
again, because there would be even more broken client code.
OK. That speaks against making it the default for sure.
But what if the client-code explicitly opts-in to that mode/switch?
better and more efficiently if it was in the core itself? Is it a lot of code?
Basically implicit-statement-level-rollback is the norm, AFAIK, and PostgreSQL is the exception here.
This creates frictions for ports to PostrgeSQL, and cross-RDBMBS apps in general.
Thus if it was at least possible to opt-in to it, that would be a great advance IMHO.
Client backend processes are per-user-per-DB. Would such a switch be applied to the DB?
DBs are typically tailored to specific applications, this something like this would work.
Thus all backends accessing a DB that opted-in to statement-implicit-rollback would use it (by default)?
Or could that be decided on a per-client-backend basis?
I know the discussion will probably stop here. It's unlikely to happen, I get that.
I think that's a pity, especially since there's a proof of concept, which I assume if technically valid.
Thanks, --DD
