Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Jun 13, 2017 at 10:05 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Well, the fundamental problem is that the RTE is a lousy place to keep
>> rowcount estimates. That breaks assorted desirable properties like
>> querytrees being readonly to planning/execution (not that we don't
>> end up copying them anyway, but up to now that's been because of bad
>> implementation not because the representation was broken by design).
> How does it break those properties? I don't think enrtuples is being
> modified by planning or execution as things stand.
But it needs to be changeable, unless you like the proposition that we
can never replan a query inside a trigger on the basis of new information
about how big the transition table is. Even if you're okay with that
particular restriction, the NamedTupleStore stuff is supposed to be
flexible enough to accommodate other use-cases, and some of them will
surely not be okay with an immutable estimate for the store's size.
Thomas noted upthread that there wasn't any API for injecting a rowcount
estimate from an SPI caller, which is wrong actually: there's already an
enrtuples field in EphemeralNamedRelationMetadata. So another way to
explain why this is broken is that having a field in the RTE is redundant
with that.
regards, tom lane
