On 2014-12-29 12:06:07 +0200, Heikki Linnakangas wrote:
> That's a little bit better, but I have to say I'm still not impressed. There
> are so many implicit assumptions in the system. The first assumption is that
> a 32-bit node id is sufficient.
Seriously? Are we going to build facilities for replication systems with
that many nodes? It seems absolutely unrealistic that a) somebody does
this b) that we'll blindly meet the demands of such a super hypothetical
scenario.
> Then there's the assumption that the node id should be "sticky",
> i.e. it's set for the whole session. Again, I'm sure that's useful for
> many systems, but I could just as easily imagine that you'd want it to
> reset after every commit.
It's trivial to add that/reset it manually. So what?
> To be honest, I think this patch should be reverted. Instead, we should
> design a system where extensions can define their own SLRUs to store
> additional per-transaction information. That way, each extension can have as
> much space per transaction as needed, and support functions that make most
> sense with the information. Commit timestamp tracking would be one such
> extension, and for this node ID stuff, you could have another one (or
> include it in the replication extension).
If somebody wants that they should develop it. But given that we, based
on previous discussions, don't want to run user defined code in the
relevant phase during transaction commit *and* replay I don't think it'd
be all that easy to do it fast and flexible.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services