On 25/02/2020 12:11, Laurenz Albe wrote: > On Tue, 2020-02-25 at 13:25 +0530, Robert Haas wrote: >> On Tue, Feb 25, 2020 at 12:47 PM Vladimir Sitnikov >> <sitnikov.vladimir@gmail.com> wrote: >>> Noone suggested that "commit leaves the session in a transaction state". >>> Of course, every commit should terminate the transaction. >>> However, if a commit fails (for any reason), it should produce the relevant ERROR that explains what went wrong rather than silently doing a rollback. >> >> OK, I guess I misinterpreted the proposal. That would be much less >> problematic -- any driver or application that can't handle ERROR in >> response to an attempted COMMIT would be broken already. > > I agree with that. > > There is always some chance that someone relies on COMMIT not > throwing an error when it rolls back, but I think that throwing an > error is actually less astonishing than *not* throwing one. > > So, +1 for the proposal from me.
I started this thread for some discussion and hopefully a documentation patch. But now I have moved firmly into the +1 camp. COMMIT should error if it can't commit, and then terminate the (aborted) transaction. -- Vik Fearing
OK, here is a patch that actually doesn't leave the transaction in a failed state but emits the error and rolls back the transaction.
This is far from complete as it fails a number of tests and does not cover all of the possible paths.
But I'd like to know if this is strategy will be acceptable ?
What it does is create another server error level that will emit the error and return as opposed to not returning.
I honestly haven't given much thought to the error message. At this point I just want the nod as to how to do it.