On Sat, Aug 9, 2014 at 12:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
So I'm thinking you're right: we should rewrite this code so that only indexes having all columns distinct can match, thereby ensuring that there is only one possible interpretation of the equality semantics for the foreign key.
I've attached a patch which disallows these, though I was not quite sure about the new error message since likely this is something that should be backpacked? I wasn't sure on the policy for new translations in a minor release.
The wording of the comment above the code which disallows the duplicate column references could likely do with some meaningful explanation about the reasons why we disallow these, I thought your words would be a bit better at doing this so I didn't try very hard on that part.