On Thu, Dec 17, 2009 at 1:37 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Andrew Dunstan <andrew@dunslane.net> writes:
>> Tom Lane wrote:
>>> The problem with COPY FROM is that it hard-wires a decision that there
>>> is one and only one possible result format, which I think we pretty
>>> much proved already is the wrong thing. I'm not thrilled with "RETURNING
>>> ARRAY" either, but we need to leave ourselves wiggle room to have more
>>> than one result format from the same source file.
>
>> Well, we could have "RETURNING type-expression" with "text[]" supported
>> for the first iteration.
>
>> In answer to Heiki's argument, what I wanted was exactly to return an
>> array of text for each row. Whatever we have needs to be able to handle
>> to possibility of ragged input (see previous discussion) so we can't tie
>> it down too tightly.
>
> I think that there are two likely possibilities for the result format:
>
> * "Raw" data after just the de-escaping and column separation steps.
> Array of text is probably the right thing here, at least for a text COPY
> (doesn't seem to cover the binary case though).
>
> * The data converted to some specified row type.
Agreed.
> "RETURNING type-expression" is probably not good since it looks more
> like the second case than the first --- and in fact it could be outright
> ambiguous, what if your data actually is one column that is a text
> array?
>
> If we're willing to assume these are the *only* possibilities then we
> could use "COPY FROM ..." for the first and "COPY RETURNING type-list
> FROM ..." for the second. I'm a bit uncomfortable with that assumption
> though; it seems likely that we'll want to shoehorn in some more
> alternatives later. (Like, what about the binary case?)
You might want to specify column names as well as well as types, in
this second case.
...Robert