Re: CompactCheckpointerRequestQueue versus pad bytes

I wrote:
> I had thought that we might get a performance boost here by saving fsync
> queue traffic, but I see that md.c was already not calling
> register_dirty_segment for temp rels, so there's no joy there.

Actually, wait a second.  We were smart enough to not send fsync
requests in the first place for temp rels.  But we were not smart enough
to not call ForgetRelationFsyncRequests when deleting a temp rel,
which made for an entirely useless scan through the pending-fsyncs
table.  So there could be win there, on top of not forwarding the actual
unlink operation.

			regards, tom lane