Re: patch: function xmltable

Hi,

On 2017-01-25 05:45:24 +0100, Pavel Stehule wrote:
> 2017-01-25 1:35 GMT+01:00 Andres Freund <andres(at)anarazel(dot)de>:
>
> > On 2017-01-24 21:32:56 -0300, Alvaro Herrera wrote:
> > > Andres Freund wrote:
> > > > Hi,
> > > >
> > > > On 2017-01-24 17:38:49 -0300, Alvaro Herrera wrote:
> > > > > +static Datum ExecEvalTableExpr(TableExprState *tstate, ExprContext
> > *econtext,
> > > > > + bool *isnull);
> > > > > +static Datum ExecEvalTableExprFast(TableExprState *exprstate,
> > ExprContext *econtext,
> > > > > + bool *isNull);
> > > > > +static Datum tabexprFetchRow(TableExprState *tstate, ExprContext
> > *econtext,
> > > > > + bool *isNull);
> > > > > +static void tabexprInitialize(TableExprState *tstate, ExprContext
> > *econtext,
> > > > > + Datum doc);
> > > > > +static void ShutdownTableExpr(Datum arg);
> > > >
> > > > To me this (and a lot of the other code) hints quite strongly that
> > > > expression evalution is the wrong approach to implementing this. What
> > > > you're essentially doing is building a vulcano style scan node. Even
> > if
> > > > we can this, we shouldn't double down on the bad decision to have these
> > > > magic expressions that return multiple rows. There's historical reason
> > > > for tSRFs, but we shouldn't add more weirdness like this.
> > >
> > > Thanks for giving it a look. I have long thought that this patch would
> > > be at odds with your overall executor work.
> >
> > Not fundamentally, but it makes it harder.
> >
>
> If you plan to hold support SRFin target list, then nothing is different.
> In last patch is executed under nodeProjectSet.

It is, because we suddenly need to call different functions - and I'm
revamping most of execQual to have an opcode dispatch based execution
model (which then also can be JITed).

> > > XMLTABLE is specified by the standard to return multiple rows ... but
> > > then as far as my reading goes, it is only supposed to be supported in
> > > the range table (FROM clause) not in the target list. I wonder if
> > > this would end up better if we only tried to support it in RT. I asked
> > > Pavel to implement it like that a few weeks ago, but ...
> >
> > Right - it makes sense in the FROM list - but then it should be an
> > executor node, instead of some expression thingy.
> >
>
> The XMLTABLE function is from user perspective, from implementation
> perspective a form of SRF function. I use own executor node, because fcinfo
> is complex already and not too enough to hold all information about result
> columns.

> The implementation as RT doesn't reduce code - it is just moving to
> different file.

You're introducing a wholly separate callback system (TableExprRoutine)
for the new functionality. And that stuff is excruciatingly close to
stuff that the normal executor already knows how to do.

> I'll try to explain my motivation. Please, check it and correct me if I am
> wrong. I don't keep on my implementation - just try to implement XMLTABLE
> be consistent with another behave and be used all time without any
> surprise.
>
> 1. Any function that produces a content can be used in target list. We
> support SRF in target list and in FROM part. Why XMLTABLE should be a
> exception?

targetlist SRFs were a big mistake. They cause a fair number of problems
code-wise. They permeated for a long while into bits of both planner and
executor, where they really shouldn't belong. Even after the recent
changes there's a fair amount of uglyness associated with them. We
can't remove tSRFs for backward compatibility reasons, but that's not
true for XMLTABLE

Greetings,

Andres Freund