Thread: Can we use Statistics Import and Export feature to perforamance testing?

On Tue, 8 Apr 2025 at 12:21, Ryohei Takahashi (Fujitsu)
<r.takahashi_2@fujitsu.com> wrote:
> By using Statistics Import and Export feature, is it possible to achieve the above request by following procedure?
> (1) Export the statistics from production environment by using pg_dump --statistics-only.
> (2) On the staging environment, set the autovacuum related parameters to prevent autovacuum from running.
> (3) Import the statistics to staging environment by using the result of (1).

You could certainly test the performance, but this method isn't
guaranteed to give meaningful results just because the table stats
match. One important thing to remember is that the planner also looks
at the *actual size* of the relation and takes that into account when
scaling the statistics (see table_block_relation_estimate_size() in
tableam.c). If the table sizes don't match between the two servers
then there's no guarantees the planner will produce the same plan.

Also, there might be other subtleties regarding OIDs of indexes which
are not guaranteed to be the same after dump/restore. Given some
fuzzily close enough cost estimates (See add_path() and
compare_path_costs_fuzzily()), it is possible a plan would switch to
using another index if sorting the indexes by their OIDs didn't match
on each server. The chances of that might be fairly small, but not
zero.

You'd also need to ensure the configs are the same in terms of GUCs
that are used for costs.

You could probably use get_relation_info_hook to overwrite the sizes
and make sure the indexes are in the same order, etc.

David

Hi,


Thank you for your reply.
I understand that the access plans are not guaranteed to be the same.

Can we add these notes to the pg_dump page in the PostgreSQL Documentation
in order to prevent users from asking the same question?

Regards,
Ryohei Takahashi

On Sat, 12 Apr 2025 at 20:29, Corey Huinker <corey.huinker@gmail.com> wrote:
>>
>> at the *actual size* of the relation and takes that into account when
>> scaling the statistics (see table_block_relation_estimate_size() in
>> tableam.c). If the table sizes don't match between the two servers
>> then there's no guarantees the planner will produce the same plan.
>
> Sorry that I didn't see this thread until now. I would like to note that table_block_relation_estimate_size()
determinesthe actual size of the relation by asking pg_class, and the relevant values there are set by
pg_restore_relation_stats().

Sorry, this isn't correct. I suspect you're probably misreading the
code. On a fleeting glance, you might have seen the "relpages =
(BlockNumber) rel->rd_rel->relpages;" line and come to this
conclusion. A more careful study will reveal the truth. Check for
"curpages = RelationGetNumberOfBlocks(rel);" and an unconditional
"*pages = curpages;".

You might be getting confused because the code does look at the
pg_class fields, but that's only to estimate the tuple density. When
pg_class has those estimates, they're used to calculate the estimated
density by doing reltuples / relpages, but that average rows per page
is then applied to the *actual* number of pages in the relation. This
method allows the stats to be scaled as the table grows and that take
effect without waiting for vacuum or analyze to update the pg_class
fields. The planner checks the current size of the table every time it
plans a query. That's very well understood and documented. See [1].

David

[1] https://www.postgresql.org/docs/current/row-estimation-examples.html#ROW-ESTIMATION-EXAMPLES