On Jul 13, 2008, at 10:16, Tom Lane wrote:
> Hmm. I think what that actually means is that the cast from citext to
> bpchar should be AS ASSIGNMENT rather than IMPLICIT. What is
> happening
> is that the system can't figure out whether to use length(text) or
> length(bpchar) when presented with a citext argument. I had been
> thinking yesterday that it would automatically prefer length(text)
> because text is a "preferred type", but after tracing through it I see
> that that doesn't happen because citext is not thought to be of the
> string category. (We really need a way to let user-defined types
> specify their category...)
That'd be nice.
> The fact that you need all these piggyback functions is a red flag
> because what it implies is that citext will not work nicely for any
> situation where both text and bpchar functions have been provided
> --- and that includes user-added functions, so it's hopeless to think
> that you can get to a solution this way. Downgrading the cast seems
> like the right thing to me.
Yes, that works for me. I've downgraded it and can now remove the size
functions and all the tests still pass.
> The implicit cast to varchar is a bit worrisome because it creates the
> same issue if someone has provided both varchar and text versions of a
> function. However, that seems a bit pointless given the lack of
> semantic difference, and I suspect that a lot of user-written
> functions
> come only in varchar variants --- so on balance my recommendation is
> to
> keep that one as implicit.
Yes, I agree. Thanks for tracing this out, Tom, it never would have
ocurred to me -- and now I know more than I did before!
Best,
David
