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

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() determines the actual size of the
relation by asking pg_class, and the relevant values there are set by
pg_restore_relation_stats().

* Question

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).

This was one of the initial intended uses for the statistical import
functions, specifically the pg_set_(relation|attribute)_stats variants.
Those variants have gone away, but the main functional difference was that
pg_restore_relation_stats() did inplace updates (it no longer does), and
without that difference set- variants became redundant.

So your procedure should still work so long as those statistics remain in
place, but just for explain plan generation of queries in isolation - there
is no way through statistics to make a large production query that
overflows work_mem do the same on a small test database, or other effects
that are a consequence of finding real data in the tables.

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() determines the 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]https://www.postgresql.org/docs/current/row-estimation-examples.html#ROW-ESTIMATION-EXAMPLES.

David

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

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

Thanks for the clarification.

Hi,

On Tue, 8 Apr 2025 12:14:08 +0000
"Ryohei Takahashi (Fujitsu)" <r.takahashi_2@fujitsu.com> wrote:

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?

I agree that it would be helpful to add this description, since the wording
“statistics manipulation functions” might give the impression that they can
influence generated plans.

I’ve attached a patch that adds a new paragraph to the warning section of the
documentation on statistics manipulation functions:

Manually restored statistics do not guarantee that the same query plans
will be generated as in the source environment, since factors such as
relation sizes, index OIDs, and configuration parameters may affect
planner behavior.

What do you think?

Regards,
Yugo Nagata

--
Yugo Nagata <nagata@sraoss.co.jp>

Hi Nagata san,

Thank you.
Your patch is good for me.

I think this kind of documentation can prevent users confusing.

Regards,
Ryohei Takahashi

On Fri, 24 Oct 2025 00:00:42 +0000
"Ryohei Takahashi (Fujitsu)" <r.takahashi_2@fujitsu.com> wrote:

Hi Nagata san,

Thank you.
Your patch is good for me.

I think this kind of documentation can prevent users confusing.

Thank you for your review.
I have registered this patch in the commitfest app.
https://commitfest.postgresql.org/patch/6155/

Regards,
Yugo Nagata

--
Yugo Nagata <nagata@sraoss.co.jp>

I prepared a small documentation follow-up patch adding a cross-reference to the planner statistics documentation section from the statistics manipulation warning.

The patch builds cleanly with:

`make -C doc/src/sgml html`

I will send the patch to the mailing list thread as well.

On Wed, 27 May 2026 05:54:06 +0000
vellaipandiyan sm <vellaipandiyan.sm@gmail.com> wrote:

I prepared a small documentation follow-up patch adding a cross-reference to the planner statistics documentation section from the statistics manipulation warning.

The patch builds cleanly with:

`make -C doc/src/sgml html`

I will send the patch to the mailing list thread as well.

I may have missed it, but I couldn't find the patch in the thread.

Could you please point me to it?

Regards,
Yugo Nagata

--
Yugo Nagata <nagata@sraoss.co.jp>

Hi, apologies to jump-in. Bear with me to share some observations.

Please point out if there is any misunderstanding.

Summary:
Regarding the purpose, I think there are 3 concerns about this.

1)"relation sizes" is ambiguous
2)"Index OIDs" might be misleading
3)Histogram bounds feature can reinforce not guaranteeing plan

Possible alternative:
since factors such as actual relation file sizes, the creation order
of indexes, actual data range in indexes, or configuration parameters
may affect planner behavior.

Reasons:

1)"relation sizes" is ambiguous
I think "relation sizes" is ambiguous and could be confused with
pg_class.relpages. I would suggest using "actual relation file sizes"
instead. Would that be clearer?

2)"Index OIDs" might be misleading
I initially misread this as meaning that statistics are restored by
OID, but I think the intent is different. If two indexes have fuzzily
equal costs, the planner may choose between them based on the order
in which they are listed internally, which follows the order of their
OIDs. Since index OIDs are not guaranteed to be the same after
dump/restore, the chosen index could differ between environments even
with identical statistics.

If that is the correct interpretation, I would suggest rephrasing to
something like "the creation order of indexes" to avoid the
misunderstanding that statistics are identified or restored by OID.

3)Histogram bounds feature(get_actual_variable_range())
The planner uses actual data range read from indexes when a WHERE
condition value falls outside the histogram bounds. This is done via
get_actual_variable_range() in selfuncs.c. Since this reads live index
data rather than restored statistics, the same restored statistics can
produce different selectivity estimates if the actual data range in the
target environment differs.
I think this feature can strongly explain that we can not guarantee a plan
to fix statistics.

Any thoughts on this?

Kenichiro Tanaka

2026年6月5日(金) 17:21 Yugo Nagata <nagata@sraoss.co.jp>: