Re: pgsql_fdw, FDW for PostgreSQL server

I wrote:
> There are a couple of other points that make me think we need to revisit
> the PlanForeignScan API definition some more, too. ...
> So we need to break down what PlanForeignScan currently does into three
> separate steps. The first idea that comes to mind is to call them
> GetForeignRelSize, GetForeignPaths, GetForeignPlan; but maybe somebody
> has a better idea for names?

Attached is a draft patch for that. While I was working on this
I realized that we were very far short of allowing FDWs to set up
expressions of their choice for execution; there was nothing for that in
setrefs.c, nor some other places that need to post-process expressions.
I had originally supposed that fdw_private could just contain some
expression trees, but that wasn't going to work without post-processing.
So this patch attempts to cover that too, by breaking what had been
fdw_private into a "private" part and an "fdw_exprs" list that will be
subject to expression post-processing. (The alternative to this would
be to do post-processing on all of fdw_private, but that would
considerably restrict what can be in fdw_private, so it seemed better
to decree two separate fields.)

Working on this also helped me identify some other things that had been
subliminally bothering me about pgsql_fdw's qual pushdown code. That
patch is set up with the idea of pushing entire quals (boolean
RestrictInfo expressions) across to the remote side, but I think that
is probably the wrong granularity, or at least not the only mechanism
we should have. IMO it is more important to provide a structure similar
to index quals; that is, what you want to identify is RestrictInfo
expressions of the form
remote_variable operator local_expression
where the operator has to be one that the remote can execute with the
same semantics as we think it has, but the only real restriction on the
local_expression is that it be stable, because we'll execute it locally
and send only its result value across to the remote. (The SQL sent to
the remote looks like "remote_variable operator $1", or some such.)
Thus, to take an example that's said to be unsafe in the existing code
comments, there's no problem at all with
remote_timestamp_col = now()
as long as we execute now() locally.

There might be some value in pushing entire quals across too, for
clauses like "remote_variable_1 = remote_variable_2", but I believe
that these are not nearly as important as "variable = constant" and
"variable = join_variable" cases. Consider that when dealing with a
local table, only the latter two cases can be accelerated by indexes.

regards, tom lane