On Wed, Oct 21, 2015 at 5:23 AM, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com> wrote: > Increasing the sorting cost factor (when use_remote_estimates = false) from > 1.1 to 1.2 makes the difference disappear. > > Since the startup costs for postgres_fdw are large portion of total cost, > extra 10% of rest of the cost is comparable to 1% fuzzy limit. IMO, we > shouldn't bother too much about it as the path costs are not much different.
My feeling is that cranking the sorting cost factor up to 20-25% would be a good idea, just so we have less unnecessary plan churn. I dunno if sorting always costs that much, but if a 10% cost overhead is really 1% because it only applies to a fraction of the cost, I don't think that's good. The whole point was to pick something large enough that we wouldn't take the sorted path unless we will benefit from the sort, and clearly that's not what happened here.
PFA patch with the default multiplication factor for sort bumped up to 1.2.