út 25. 2. 2020 v 22:14 odesílatel Tom Lane <tgl@xxxxxxxxxxxxx> napsal:
Paul Jungwirth <pj@xxxxxxxxxxxxxxxxxxxxxxxx> writes:
> Not that this is necessarily fatal, but you'd need to avoid parsing
> trouble with the other EXCEPT, e.g.
> SELECT 1 EXCEPT SELECT 1;
Yeah, it doesn't sound like much consideration has been given to
that ambiguity, but it's a big problem if you want to use a syntax
like this.
> Google Big Query was mentioned upthread. I see they require parens, e.g.
> SELECT ... EXCEPT (...). I don't think that actually fixes the ambiguity
> though.
Indeed it doesn't, because you can parenthesize an EXCEPT's sub-queries:
regression=# select 1 except (select 2);
?column?
----------
1
(1 row)
In principle, once you got to the SELECT keyword you could tell things
apart, but I'm afraid that might be too late for a Bison-based parser.
> So it seems they require at least one `*` in the SELECT target list. In
> fact the `*` must be the very last thing. Personally I think it should
> be as general as possible and work even without a `*` (let alone caring
> about its position).
I wonder if they aren't thinking of the EXCEPT as annotating the '*'
rather than the whole SELECT list. That seems potentially more flexible,
not less so. Consider
SELECT t1.* EXCEPT (foo, bar), t2.* EXCEPT (baz) ... FROM t1, t2, ...
This doesn't have any problem with ambiguity if t2 has a "foo" column,
or if t1 has a "baz" column; which indeed would be cases where this
sort of ability would be pretty useful, since otherwise you end up
with painful-to-rename duplicate output column names. And certainly
there is no particular need for this construct if you didn't write
a "*".
this proposal looks well
Pavel
regards, tom lane
