Alexey Ermakov <alexey(dot)ermakov(at)postgresql-consulting(dot)com> writes:
> I think we should change that estimation and in case when child node
> expected to return only one row (and really 0 rows) we should use it's
> total_cost as base for estimation of startup_cost.
This would make sense as part of the overall rethinking of startup_cost
discussed in
https://www.postgresql.org/message-id/flat/31065.1481742760%40sss.pgh.pa.us
but I'm disinclined to do it in isolation. Redefining startup_cost for
just one path type seems likely to create more problems than it solves.
regards, tom lane