Re: BUG #17170: Invalid collation created with provider icu and initdb' locale C

I wrote:
> The immediate issue is that DROP COLLATION uses the same lookup
> rule as other code paths, so it fails to see inapplicable
> pg_collation entries at all. So one approach to a fix is to
> relax its lookup rule. But I wonder if we should have prevented
> the CREATE in the first place, instead.

After a bit of experimentation, I'm inclined to the idea that
we should have rejected the CREATE COLLATION in the first place.
It's not very useful to create a collation that you're not going
to be able to use, and I think it's more user-friendly to say so
up front than to leave the user guessing about why subsequent
references don't work.

There is arguably one use-case that this fix shuts off: manually
adding ICU-based collations to template0, in the expectation
that they could get cloned into other databases where they'd be
useful. But that seems like a mighty thin argument, especially
since you could initdb with an ICU-compatible encoding if you
want to do that.

regards, tom lane