On 2017-04-11 17:25:52 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > Tom, do you have any opinion on the volatility stuff?
>
> What was the previous behavior for such cases? If it was reasonably
> sane, we probably have to preserve it. If it was unpredictable or
> completely wacko, maybe we don't.
Previously we'd stash the result in a new tuplestore, because it
happened inside ExecMakeTableFunctionResult()'s fallback path. The
inner tuplestore (from the proper SRF) would get evaluated via the the
isDone mechanism.
That'd imo be a fair amount of work to emulate, because we'd have to
manually go over the tuplesttore.
But given that we do *not* have similar semantics for volatiles in the
targetlist, I'm quite unconvinced that that's necessary. Consider
e.g. my previous example of SELECT * FROM CAST(srf() * volatile_func() AS whatnot)
rewritten into a saner version as SELECT srf * volatile_func() FROM srf() AS srf;
here volatile_func() would before and now get re-evaluated if there's a
rewind, and would only be invoked if the row is actually evaluated.
- Andres