AW: BUG #17842: Adding a qual to a working query gets bogus syntax error

Hello Tom,

thanks for the quick response.

You are probably right with your assumption of a not convertible substring, but some questions remain:

1. When the question runs succesfully, all lines are subject to the join of qj (newcol_imp and oldcol_imp heve exact the same number of lines, matching logically), but why does a limitation of the result set trigger the error?

(There really may be some substrings from declarations of numeric(4,1) which are not excluded properly from conversion to smallint.

The '0'||<rest> tries to handle an empty string)

2. Why is it reporting an syntax error (supposed at parse time) and not a runtime error (supposed at execution time)

3. Why does an uncorrelated qual like and onum < 100 trigger an error too?

( I know that a limitation can change the choosen plan, but which one cannot be seen because of the error.)

The behavior of having a succesfull full query, but getting the error with some quals seems to me very unintuitive.

PS: The exactness and correctness of this specific query is certainly not overall important for me, but I wanted to point out this very unexpected behavior for some other users.

When I find some time, I'll try to provide some anonymized data.

Best regards

Hans Buschmann

________________________________
Von: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Gesendet: Dienstag, 14. März 2023 17:20
An: Hans Buschmann
Cc: pgsql-bugs(at)lists(dot)postgresql(dot)org
Betreff: Re: BUG #17842: Adding a qual to a working query gets bogus syntax error

PG Bug reporting form <noreply(at)postgresql(dot)org> writes:
> During refactoring our application I got a bogus syntax error when adding a
> simple where-clause to a working query.

My guess is that this:

> ('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen

sometimes produces a string that fails to cast to smallint, which you
accidentally don't notice because it never gets evaluated for troublesome
values of nline --- until you add the "nlen > 0" condition. There isn't
anything in your query that prevents that from being evaluated fairly
early.

This isn't a bug, or at least you have provided no data that would
motivate (or indeed allow) anyone else to poke into it more closely.

regards, tom lane