Re: 9.6 TAP tests and extensions

On 23 September 2016 at 00:32, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Craig Ringer <craig(at)2ndquadrant(dot)com> writes:
>> On 13 September 2016 at 22:02, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> Without taking a position on the merits of this patch per se, I'd like
>>> to say that I find the argument for back-patching into 9.6 and not
>>> further than that to be pretty dubious. $(prove_check) has been there
>>> since 9.4, and in the past we've often regretted it when we failed
>>> to back-patch TAP infrastructure fixes all the way back to 9.4.
>
>> No objection to backpatching, I just thought I'd be more intrusive to
>> do that than just 9.6.
>
>> Since 9.5 and older have more limited versions of PostgresNode which
>> lack safe_psql, etc, I'm not sure it's very practical for extensions
>> to bother running TAP tests on 9.4 and 9.5 anyway.
>
> Certainly there are restrictions, but I'd imagine that every new release
> will be adding features to the TAP test infrastructure for some time to
> come. I think it's silly to claim that 9.6 is the first branch where
> TAP testing is usable at all.

OK.

I didn't intend to claim 9.6 is the first branch where it's usable,
just that the differences between 9.4/9.5 and 9.6 mean that supporting
both in extensions will likely be more pain than it is worth. Mainly
due to the safe_psql change in 2c83f435. It might've been good to
backport that, but it was pretty wide reaching and at the time I
wasn't thinking in terms of using TAP tests in extensions so it didn't
occur to me.

Extension Perl code can always use some adapter/glue code to handle
9.4 and 9.5 if they want, or error if they don't, so it's not a fatal
barrier.

Also, despite what I said upthread, there's no need for normal
PGXS-using extensions to define their $(prove_installcheck)
replacement. It works as-is. The problems I was having stemmed from
the fact that I was working with a pretty nonstandard Makefile without
realising that the changes were going to affect prove. $(prove_check)
isn't useful from extensions of course since it expects a temp install
and PGXS doesn't know how to make one, but $(prove_installcheck) does
the job fine.

It's thus sufficient to apply the patch to install the perl modules to
9.4, 9.5 and 9.6. Nothing else is needed. I've attached backports for
9.4 and 9.5.

If you were really keen, we could actually backport the new TAP code
to 9.4 and 9.5 pretty much wholesale. 9.4 and 9.5 don't have the psql
method, PostgresNode, etc, so there's nothing to break. But that's for
separate consideration.

>> Extension authors can just use:
>> ifeq ($(MAJORVERSION),9.6)
>> endif
>> when defining their prove rules.
>
> That will break as soon as 10 comes out. And numerical >= tests aren't
> all that convenient in Make. It'd be much better if a test on whether
> $(prove_check) is defined would be sufficient.

Fair enough. I forgot how much numeric range tests sucked in Make.

In that case we should backpatch the installation of the Perl modules
right back to 9.4. There's not even a need to test if
$(prove_installcheck) is defined then, just need a semicolon so it
evaluates to empty e.g.

prove_installcheck:
$(prove_installcheck) ;

check: prove_check

So ... patches attached. The 9.6 patch applies to head too.

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