> On 24 Jan 2020, at 08:27, Peter Eisentraut <peter.eisentr...@2ndquadrant.com> > wrote: > > I definitely want to make it work in a way that does not require writing C > code. My idea was to create a new type, perhaps called "descriptor", that > represents essentially a tuple descriptor. (It could be exactly a TupleDesc, > as this patch does, or something similar.) For the sake of discussion, we > could use JSON as the text representation of this. Then a PL/pgSQL function > or something else high level could easily be written to assemble this. > Interesting use cases are for example in the area of using PL/Perl or > PL/Python for unpacking some serialization format using existing modules in > those languages.
I do think it’s very desirable to make it usable outside of C code. > Obviously, there is a lot of leg work to be done between here and there, but > it seems doable. The purpose of this initial patch submission was to get > some opinions on the basic idea of "determine result tuple structure by > calling helper function at parse time", and so far no one has fallen off > their chair from that, so I'm encouraged. ;-) I’m interested in this development, as it makes RECORD-returning SRFs in the SELECT list a viable proposition, and that in turn allows a ValuePerCall SRF to get meaningful benefit from pipelining. (They could always pipeline, but there is no way to extract information from the RECORD that’s returned, with the sole exception of row_to_json.) I couldn’t check out that it would work though because I couldn’t apply the v2 (or v1) patch against either 12.0 or 530609a (which I think was sometime around 25th Jan). Against 12.0, I got a few rejections (prepjointree.c and clauses.c). I figured they might be inconsequential, but no: initdb then fails at CREATE VIEW pg_policies. Different rejections against 530609a, but still initdb fails. But I’m definitely very much encouraged. denty.
