On 1 February 2018 at 17:49, Robert Haas <robertmh...@gmail.com> wrote: > One point which I want to emphasize is that the length of the MCV list > bounds the estimated frequency of non-MCVs in two ways: no non-MCV is > ever thought to be more frequent than the least-common MCVs, and > however many non-MCVs we think we have (probably fewer than we > actually have) have to fit into whatever percentage of the table is > consumed by MCVs. This would be less important if we had reliable > n_distinct estimates, but we don't. So, even throwing things into the > MCV list that are no more common than the average item can improve > planning in some cases. >
That's a good point, and a nice explanation. I think that lends more weight to the argument that we should be including as many MCVs as possible, provided there's enough evidence to justify their inclusion. Regards, Dean