> >> Hmm ... now that's an interesting thought. So the input converter would
> >> actively strip trailing blanks, output would add them back,
>
> > But how would the output know how many to put back?
>
> The output routine would need access to the column typmod. Which it
> would have, in simple "SELECT columnname" cases, but this is a serious
> weakness of the scheme in general.
I think it is not really a weakness of the scheme, but a weakness that typmod
is not available in some places where it would actually be needed.
One effect of this is, that all the varlena datatypes have a redundant
length info per row, even for such types that have the same length for
all rows of one table (e.g. numeric(8,2), and char(n)). In a lot of cases
that means 50-100% more disk space than actually needed.
I can see, that making typmod availabe in more places would probably
be major work (or not feasible at all), but I think it would generally be
of (great) value.
To the problem of concatenation:
Would it be feasible to alter the concatenation method to concatenate
the results of the output functions of the relevant expressions ?
Imho that would actually return the expected results more often than it
currently does, and it would fix the typmod issue for char(n) concatenation.
Andreas