From: Tom Lane <tgl@sss.pgh.pa.us>
To: Michael Paquier <michael@paquier.xyz>
cc: Heikki Linnakangas <hlinnaka@iki.fi>,
        pgsql-hackers <pgsql-hackers@postgresql.org>
Subject: Re: Weird test mixup
In-reply-to: <ZfOABU4ahF2QZQ6N@paquier.xyz>
References: <10fd6cdd-c5d9-46fe-9fa1-7e661191309e@iki.fi>
 <2387221.1710454778@sss.pgh.pa.us> <ZfOABU4ahF2QZQ6N@paquier.xyz>
Comments: In-reply-to Michael Paquier <michael@paquier.xyz>
	message dated "Fri, 15 Mar 2024 07:53:57 +0900"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <2393551.1710458033.1@sss.pgh.pa.us>
Date: Thu, 14 Mar 2024 19:13:53 -0400
Message-ID: <2393552.1710458033@sss.pgh.pa.us>
Archived-At: 
 <https://www.postgresql.org/message-id/2393552.1710458033%40sss.pgh.pa.us>
Precedence: bulk

Michael Paquier <michael@paquier.xyz> writes:
> On Thu, Mar 14, 2024 at 06:19:38PM -0400, Tom Lane wrote:
>> I wonder if it'd be wise to adjust the injection point stuff so that
>> it's active in only the specific database the injection point was
>> activated in.

> It can be made optional by extending InjectionPointAttach() to
> specify a database OID or a database name.  Note that
> 041_checkpoint_at_promote.pl wants an injection point to run in the
> checkpointer, where we don't have a database requirement.

> Or we could just disable runningcheck because of the concurrency
> requirement in this test.  The test would still be able to run, just
> less times.

No, actually we *must* mark all these tests NO_INSTALLCHECK if we
stick with the current definition of injection points.  The point
of installcheck mode is that the tests are supposed to be safe to
run in a live installation.  Side-effects occurring in other
databases are completely not OK.

I can see that some tests would want to be able to inject code
cluster-wide, but I bet that's going to be a small minority.
I suggest that we invent a notion of "global" vs "local"
injection points, where a "local" one only fires in the DB it
was defined in.  Then only tests that require a global injection
point need to be NO_INSTALLCHECK.

			regards, tom lane