Re: Support for dumping extended statistics

Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com> writes:
> On 1/7/23 03:39, Bruce Momjian wrote:
>> There is certainly interest in allowing the optimizer statistics to be
>> dumped and reloaded. This could be used by pg_restore and pg_upgrade.

> Indeed, although I think it'd be better to deal with regular statistics
> (which is what 99% of systems use). Furthermore, we should probably
> think about differences between major versions - until now we could
> change on-disk format of the statistics, because we have reset them.

Yeah, it's extremely odd to be proposing dump/reload for extended
stats when we don't yet have it for plain stats. And yes, the main
stumbling block is that you need to have a plan for stats changing
across versions, or even just environmental issues. For example,
what if the target DB doesn't use the same collation as the source?
That would affect string sorting and therefore at least partially
invalidate histograms for text columns.

I actually did some work on this, probably close to ten years ago
now, and came up with some hacks that didn't pass community review.
It'd be a good idea to dig up those old discussions if you want to
re-open the topic.

regards, tom lane