Re: Window Functions: v07 APIs and buffering strateties

Martijn van Oosterhout <kleptog(at)svana(dot)org> writes:
> On Tue, Oct 28, 2008 at 12:38:09PM -0400, Tom Lane wrote:
>> Right offhand, I don't see any alternative but to make both ROWS and
>> RANGE reserved. It's pretty annoying since that might break existing
>> applications that have been using these as identifiers, but the SQL
>> committee seems to care little about that :-(

> Given that the only problematic case is if expr_list ends with a
> postfix operator, wouldn't it be sufficient to simply decree that in
> that case you need parentheses? Seems a lot less painful than adding
> two reserved words.

Hmm ... actually, it might be possible to fix it with a suitable
precedence declaration? The trick is to make sure that in
... ORDER BY foo ! ROWS ...
the operator is taken as postfix not infix, which is the exact opposite
of what we did for AS-less column aliases (and, in fact, is the opposite
of what bison will do by default, IIRC). So it might be possible to fix
by attaching some new precedence level to the ROWS token.

On the other hand, this would still mean that ROWS acts oddly
differently from any other column name, and in a way that would be hard
to explain or document. So I'm really not sure that this is a better
solution than reserving it.

Still another trick in our toolbox is to use merged tokens to fix the
lack of lookahead. If I read the spec correctly, the problematic cases
of ROWS and RANGE must be followed by UNBOUNDED or BETWEEN, so folding
these token pairs into four merged-tokens would get rid of the need to
reserve anything. Merged tokens create their own little oddities too
though, as I was just complaining to Peter. I wonder whether it would
be possible to improve on that problem by arranging some way for the
grammar to signal the lexer about whether lookahead is needed, and thus
not do the merging in contexts where it couldn't be appropriate.

regards, tom lane