Re: Event Triggers: adding information

Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> writes:
> Well, that's already a problem, because as Robert keeps saying, what
> goes through utility.c and what doesn't is pretty random right at the
> moment, and we shouldn't expose that behavior to users for fear of not
> being able to change it later.

It didn't feel that random to me. In most cases, the rule is that if the
system wants to execute a command that you could have typed as a user in
the prompt, then it hacks together a parsenode and call ProcessUtility.

The main exception to that rule is the CASCADE implementation, for
reasons that you detailed in that thread earlier.

> I think we could possibly ship event triggers defined as start and end
> of a *top level* command and have them working reliably in 9.3. If you
> want users to be looking at subcommands, there is a lot more work to do
> there than we have any chance of getting done for 9.3 (if it is to ship
> in 2013, that is).

By default, you only get top level in what I've been cooking. I think
the other contexts are needed for the logical replication, and maybe it
would be good enough if they are restricted to either only internal code
or event triggers coded in C. What do you think?

> Alternatively, if you want to get something into 9.3 that has not
> necessarily got a long-term-stable API, I'd be inclined to suggest that
> we forget about a SQL-level event trigger facility for now, and just put
> in some hook call points. It's pretty well established that we're
> willing to change the API seen by hook functions across major releases.

Or just a hook. That would want to have about the exact same amount of
information as the Event Trigger anyway, so I'm thinking we could maybe
do that the same way as the parsetree exposure?

Another idea would be yet another GUC, ala allow_system_table_mods, so
that only intrepid users can have at it. Well, as long as the logical
replication use case is covered, I'm done here, so I want to hear from
Simon and Andres on that (and maybe the Slony and Londiste guys, etc),
and from you to triage what is possible and what is crazy.

> TBH this might be the best thing anyway if your long-term goals have to
> do with replication, because it'd be a lot cheaper to get to the point
> where you could write the replication code and see if it all actually
> works. I'm fairly concerned that we might spend many man-months getting
> event triggers with definitions A,B,C into the system, only to find out
> later that what is really needed for logical replication is definitions
> D,E,F.

I'm worried about that too, and that's why I'm trying to remain general
and quite transparent about the way the system actually works.

Regards,
--
Dimitri Fontaine
http://2ndQuadrant.fr PostgreSQL : Expertise, Formation et Support