On Wed, Jun 30, 2021 at 11:15 AM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:
> You're right maintaining a per-partition samples and merging those might
> solve (or at least reduce) some of the problems, e.g. eliminating most
> of the I/O that'd be needed for sampling. And yeah, it's not entirely
> clear how to merge some of the statistics types (like ndistinct). But
> for a lot of the basic stats it works quite nicely, I think.
It feels like you might in some cases get very different answers.
Let's say you have 1000 partitions. In each of those partitions, there
is a particular value that appears in column X in 50% of the rows.
This value differs for every partition. So you can imagine for example
that in partition 1, X = 1 with probability 50%; in partition 2, X = 2
with probability 50%, etc. There is also a value, let's say 0, which
appears in 0.5% of the rows in every partition. It seems possible that
0 is not an MCV in any partition, or in only some of them, but it
might be more common overall than the #1 MCV of any single partition.
--
Robert Haas
EDB: http://www.enterprisedb.com