Thread: Cost estimation in foreign data wrappers
Hello,
There is a callback function in fdw's which should also set estimates for startup and total costs for each path. Assume a fdw adds only one path (e.g. in file_fdw). I am trying to understand what can go wrong if we do a bad job in estimating these costs.
Since we have only one scan path here, it doesn't make a difference in choosing the best scan path.
By looking at the code and doing some experiments, I think this can be significant in (1) underestimating a nested loop's cost, (2) not materializing inner table in nested loop.
* Are there any other cases that this can be significant?
* Assume we are not sure about the exact cost, but we know that it is in [lower_bound, upper_bound] range, where upper_bound can be 10x lower_bound Then, what value is better to choose? lower bound? upper bound? or average?
Thanks,
-- Hadi
Hadi Moshayedi <hadi@moshayedi.net> writes:
> There is a callback function in fdw's which should also set estimates for
> startup and total costs for each path. Assume a fdw adds only one path
> (e.g. in file_fdw). I am trying to understand what can go wrong if we do a
> bad job in estimating these costs.
> Since we have only one scan path here, it doesn't make a difference in
> choosing the best scan path.
Right. But if there's more than one table in the query, it might make a
difference in terms of what join plan gets chosen. I'd say that getting
an accurate rowcount estimate is usually far more important, though.
regards, tom lane