I wrote:
> I left out the proposed regression tests because they fail in "make
> installcheck" mode, unless you've previously built and installed cube
> and seg, which seems like an unacceptable requirement to me. I don't
> think that leaving the code untested is a good final answer, of course.
> The idea I was toying with was to create a dummy extension for testing
> purposes by means of doing a direct INSERT into pg_extension --- which
> is ugly and would only work for superusers, but the latter is true of
> "CREATE EXTENSION cube" too. Anybody have a better idea?
I had a possibly better idea: instead of manufacturing an empty extension
with a direct INSERT, hack on the one extension that we know for sure
will be installed, namely postgres_fdw itself. So we could do, eg,
create function foo() ...
alter extension postgres_fdw add function foo();
and then test shippability of foo() with or without having listed
postgres_fdw as a shippable extension.
This is certainly pretty ugly in its own right, but it would avoid
the maintainability hazards of an explicit INSERT into pg_extension.
So on balance it seems a bit nicer than my first idea.
regards, tom lane