On Tue, Jun 6, 2017 at 9:24 PM, Jim Finnerty <jfinnert@amazon.com> wrote:
> In some MPP systems, networking costs are modeled separately from I/O costs,
> processor costs, or memory access costs. I think this is what Ashutosh may
> have been getting at with /per-packet/ costs: in a more sophisticated fdw
> cost model there could be a network cost per /packet/ that would be
> independent of the cost of reading the next page or a random page from local
> storage.
I agree. I think the question is how much we'd gain in practice if we
modeled the cost more accurately. IMHO, the bigger problem with the
FDW stuff today is that we still lack partition-wise join,
partition-wise aggregate, and asynchronous query, which means that
only relatively simple queries involving foreign tables have a chance
of getting the plan you'd really like to have. Until that's fixed, I
don't personally think it's worth spending a lot of time trying to
tweak the costing model. Of course, if somebody wants to take a run
at it and can show that the benefit is there, cool beans.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company