Re: Add CREATE support to event triggers

Alvaro Herrera wrote:
> Actually here's a different split of these patches, which I hope makes
> more sense. My intention here is that patches 0001 to 0004 are simple
> changes that can be pushed right away; they are not directly related to
> the return-creation-command feature. Patches 0005 to 0027 implement
> that feature incrementally. You can see in patch 0005 the DDL commands
> that are still not implemented in deparse (they are the ones that have
> an elog(ERROR) rather than a "command = NULL"). Patch 0006 adds calls
> in ProcessUtilitySlow() to each command, so that the object(s) being
> touched are added to the event trigger command stash.
>
> Patches from 0007 to 0027 (excepting patch 0017) implement one or a
> small number of commands in deparse. Patch 0017 is necessary
> infrastructure in ALTER TABLE to support deparsing that one.
>
> My intention with the later patches is that they would all be pushed as
> a single commit, i.e. the deparse support would be implemented for all
> commands in a fell swoop rather than piecemeal -- except possibly patch
> 0017 (the ALTER TABLE infrastructure). I split them up only for ease of
> review. Of course, before pushing we (I) need to implement deparsing
> for all the remaining commands.

Here's a new version of this series. The main change is that I've
changed deparse_utility.c to generate JSON, and the code that was in
commands/event_trigger.c to decode that JSON, so that it uses the new
Jsonb API instead. In addition, I've moved the new code that was in
commands/event_trigger.c to utils/adt/ddl_json.c. (The only entry point
of the new file is the SQL-callable pg_event_trigger_expand_command()
function, and its purpose is to expand a JSON object emitted by the
deparse_utility.c code back into a plain text SQL command.)

I have also cleaned up the code per comments from Michael Paquier and
Andres Freund:

* the GRANT support for event triggers now correctly ignores global
objects.

* COMMENT ON .. IS NULL no longer causes a crash

* renameatt() and ExecRenameStmt are consistent in returning the
objSubId as an "int" (not int32). This is what is used as objectSubId
in ObjectAddress, which is what we're using this value for.

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services