On Sun, Sep 6, 2015 at 12:56 AM, Noah Misch <noah@leadboat.com> wrote:
> My comments have flowed out of a principle that autonomous transactions shall
> have precisely the same semantics as using another backend via dblink.
That's what I thought, too. AT is syntax sugar for dblink approach.
Syntax and performance aside, I think the only way dblink style AT
could be improved would be to have:
1) better cancel handling, especially cancel child when parent cancels
(this is a major headache with dblink)
2) automatic rejection of acquire attempts by child on exclusive
assets held by parent
#2 is what we've been discussing, but I'm pretty confused. Most
especially I'm not clear on whether parent and children share a pid or
have unique pid. If pids are unique as in dblink, #2 could be done
via a query assuming you have a way of identifying child transactions
in pg_stat_activity. The discussion has suggested that the pid is
shared, which I found odd but took on faith.
I'm also wondering if we need syntax to handle synchronous vs
asynchronous execution of the AT (the latter of which is only possible
with a separate PID, I think).
merlin
