On Tue, 2009-11-24 at 00:54 -0800, Daniel Farina wrote:
> On Tue, Nov 24, 2009 at 12:29 AM, Hannu Krosing <hannu@krosing.net> wrote:
> > COPY stdin TO udf();
>
> If stdin becomes (is?) a legitimate source of records, then this patch
> will Just Work.
>
STDIN is a source of bytes representing a set of records. Currently, the
first argument to COPY is a source or destination of records; and the
second argument is a source or destination of bytes representing a set
of records.
I think we want the first argument to remain a source or destination of
real records with types; that is, a table or perhaps a function. And we
want the second argument to remain a source or destination of bytes;
that is, a file or perhaps a function (albeit not the same kind as the
former function).
> > COPY udf() FROM stdin;
>
> This is unaddressed, but I think it would be a good idea to consider
> enabling this kind of thing prior to application.
This makes much more sense, but it is a very different type of function
from the original proposal (which basically accepts a buffer). I agree
that it sounds useful and would be good for the sake of symmetry.
One use case may be a degree of data cleaning. For instance, you could
use a "looser" function definition, like udf(cstring, cstring, ...),
where all COPY does is break up the records into fields, and the
function can recover from type input failures using subtransactions.
Binary mode could do a similar thing with bytea.
However, I recommend that you don't try to generalize this as a data
cleanup feature that can handle ragged input. That seems like a separate
problem that will distract from the original use case.
Regards,Jeff Davis
