Craig Ringer <craig@2ndquadrant.com> writes:
> On 23 February 2016 at 21:34, Robert Haas <robertmhaas@gmail.com> wrote:
>> I believe Sridhar is imagining that someday "set autocommit to false"
>> might be a command that the server would understand.
> ... I guess. Yeah.
We've been there, we've done that. We're not doing it again.
Cf commits 26993b291, f85f43dfb, 525a48991, as well as a whole
bunch of thrashing in between the first two (grep the git logs
for "autocommit" to find most of it). It's a bit harder to locate
relevant email threads, because searching for just "autocommit"
yields too many hits; but here's one long thread from when we were
starting to realize that it wasn't working very well:
http://www.postgresql.org/message-id/flat/3E54526A.121EBEE5@tpf.co.jp
In all, this was one of the more searing experiences contributing
to what's now received project wisdom that GUCs that change
fundamental semantics are a bad idea.
> Oracle's SQL*Plus has the concept of turning autocommit off, but I suspect
> that's client-side behaviour.
The conclusion we came to back in 2002-2003 was that client-side
autocommit was the only behavior we could sanely support. I see
no reason to think that a fresh experiment in the same direction
would produce a different result.
regards, tom lane