On Wed, 2009-09-02 at 08:57 -0500, Kevin Grittner wrote:
> > (a) leaving a literal as "unknown" until you've finished
> > inferring types (current behavior)
> > (b) casting every unknown to text immediately, and then trying to
> > infer the types
>
> No, that's not it. I'm wondering why it isn't treated as text.
> Period. Full stop. Nothing to infer. Anywhere that we have implicit
> casts defined from text to something else could, of course, still
> operate; but it would be text. No guessing.
If you have very many implicit casts, I think you lose the
predictability and safety you're looking for, and/or end up with a lot
of errors that eliminate the convenience of implicit casting.
> It often seems to have the opposite effect. See the original post.
The original problem has more to do with the fact that interpreting an
unknown value as a char seems to just discard a lot of information. I
assume that's part of the standard, but it seems like a bad idea any
time you silently discard data (which is why we prevented varchar(n)
from silently truncating a while ago).
> Here I think you have answered my question. It is seen as a feature,
> since it allows people to avoid the extra keystrokes of coding
> type-specific literal values, and allows them the entertainment of
> seeing how the values get interpreted. :-)
>
> > But you can't have both of those desirable behaviors
>
> Whether they are desirable is the point of disagreement. At least I
> now understand the reasoning.
They are desirable for a system that infers types from context. I agree
that there's more safety by explicitly declaring the type of all
literals; but I disagree that using implicit casts to make up for a lack
of an "unknown" type will improve matters (either for convenience or
safety).
Regards,
Jeff Davis
