Catalin Iacob <iacobcatalin@gmail.com> writes:
> On Mon, Sep 4, 2017 at 4:15 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Also, the main thing that we need xact.c's involvement for in the first
>> place is the fact that implicit transaction blocks, unlike regular ones,
>> auto-cancel on an error, leaving you outside a block not inside a failed
>> one. So I don't exactly see how savepoints would fit into that.
> I think this hits the nail on the head and should have a place in the
> official docs as I now realize I didn't grasp this distinction before
> I read this.
Yeah, it seems like we have now made this behavior official enough that
it's time to document it better. My thought is to create a new subsection
in the FE/BE Protocol chapter that explains how multi-statement Query
messages are handled, and then to link to that from appropriate places
elsewhere. If anyone thinks the reference section would be better put
somewhere else than Protocol, please say where.
regards, tom lane