Date: Fri, 17 Oct 2025 13:06:16 -0500
From: Nathan Bossart <nathandbossart@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: pgsql-hackers@postgresql.org, david@justatheory.com,
	Andrew Dunstan <andrew@dunslane.net>
Subject: Re: abi-compliance-check failure due to recent changes to
 pg_{clear,restore}_{attribute,relation}_stats()
Message-ID: <aPKFmL8VjOKNVLcw@nathan>
References: <aPJ03E2itovDBcKX@nathan>
 <1713509.1760721320@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1713509.1760721320@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/aPKFmL8VjOKNVLcw%40nathan>
Precedence: bulk

On Fri, Oct 17, 2025 at 01:15:20PM -0400, Tom Lane wrote:
> FWIW, I favor the approach of having an in-tree, per-branch file
> containing the commit hash of a commit that is the current ABI
> reference for that branch.  If the file doesn't exist (which it
> wouldn't in master, and probably not in recently-forked branches),
> skip ABI checking.  I think this is superior to the discussed
> alternative of depending on git tags, because files are easy to
> change or remove, while tags are not.  In particular, I think it'd
> likely be impossible to make the ABI reference point go backwards
> if we use tags.  Maybe that's not a case we'd ever need, but I'm
> unconvinced of that.

I'm new to the topic, but IMHO the per-branch file approach is by far the
best approach.  Not only is it much more flexible, but we could even use it
as a centralized list of ABI breaks for a given branch with justification
for each.  I can't think of any strong advantages of keeping this stuff in
git metadata.  git itself uses a file for blame.ignoreRevsFile...

-- 
nathan