Statistics Import and Export

On Thu, Aug 31, 2023 at 12:17 PM Corey Huinker <corey.huinker@gmail.com> wrote:

While the primary purpose of the import function(s) are to reduce downtime
during an upgrade, it is not hard to see that they could also be used to
facilitate tuning and development operations, asking questions like "how might
this query plan change if this table has 1000x rows in it?", without actually
putting those rows into the table.

Thanks. I think this may be used with postgres_fdw to import
statistics directly from the foreigns server, whenever possible,
rather than fetching the rows and building it locally. If it's known
that the stats on foreign and local servers match for a foreign table,
we will be one step closer to accurately estimating the cost of a
foreign plan locally rather than through EXPLAIN.

--
Best Wishes,
Ashutosh Bapat

Thanks. I think this may be used with postgres_fdw to import
statistics directly from the foreigns server, whenever possible,
rather than fetching the rows and building it locally. If it's known
that the stats on foreign and local servers match for a foreign table,
we will be one step closer to accurately estimating the cost of a
foreign plan locally rather than through EXPLAIN.

Yeah, that use makes sense as well, and if so then postgres_fdw would
likely need to be aware of the appropriate query for several versions back
- they change, not by much, but they do change. So now we'd have each query
text in three places: a system view, postgres_fdw, and the bin/scripts
pre-upgrade program. So I probably should consider the best way to share
those in the codebase.

Yeah, that use makes sense as well, and if so then postgres_fdw would
likely need to be aware of the appropriate query for several versions back
- they change, not by much, but they do change. So now we'd have each query
text in three places: a system view, postgres_fdw, and the bin/scripts
pre-upgrade program. So I probably should consider the best way to share
those in the codebase.

Attached is v2 of this patch.

New features:
* imports index statistics. This is not strictly accurate: it re-computes
index statistics the same as ANALYZE does, which is to say it derives those
stats entirely from table column stats, which are imported, so in that
sense we're getting index stats without touching the heap.
* now support extended statistics except for MCV, which is currently
serialized as an difficult-to-decompose bytea field.
* bare-bones CLI script pg_export_stats, which extracts stats on databases
back to v12 (tested) and could work back to v10.
* bare-bones CLI script pg_import_stats, which obviously only works on
current devel dbs, but can take exports from older versions.

On 10/31/23 08:25, Corey Huinker wrote:

Attached is v2 of this patch.

New features:
* imports index statistics. This is not strictly accurate: it
re-computes index statistics the same as ANALYZE does, which is to
say it derives those stats entirely from table column stats, which
are imported, so in that sense we're getting index stats without
touching the heap.

Maybe I just don't understand, but I'm pretty sure ANALYZE does not
derive index stats from column stats. It actually builds them from the
row sample.

* now support extended statistics except for MCV, which is currently
serialized as an difficult-to-decompose bytea field.

Doesn't pg_mcv_list_items() already do all the heavy work?

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Maybe I just don't understand, but I'm pretty sure ANALYZE does not
derive index stats from column stats. It actually builds them from the
row sample.

That is correct, my error.

* now support extended statistics except for MCV, which is currently
serialized as an difficult-to-decompose bytea field.

Doesn't pg_mcv_list_items() already do all the heavy work?

Thanks! I'll look into that.

The comment below in mcv.c made me think there was no easy way to get
output.

/*
* pg_mcv_list_out - output routine for type pg_mcv_list.
*
* MCV lists are serialized into a bytea value, so we simply call byteaout()
* to serialize the value into text. But it'd be nice to serialize that into
* a meaningful representation (e.g. for inspection by people).
*
* XXX This should probably return something meaningful, similar to what
* pg_dependencies_out does. Not sure how to deal with the deduplicated
* values, though - do we want to expand that or not?
*/

On 11/2/23 06:01, Corey Huinker wrote:

Maybe I just don't understand, but I'm pretty sure ANALYZE does not
derive index stats from column stats. It actually builds them from the
row sample.

That is correct, my error.
 

* now support extended statistics except for MCV, which is currently
serialized as an difficult-to-decompose bytea field.

Doesn't pg_mcv_list_items() already do all the heavy work?

Thanks! I'll look into that.

The comment below in mcv.c made me think there was no easy way to get
output.

/*
 * pg_mcv_list_out      - output routine for type pg_mcv_list.
 *
 * MCV lists are serialized into a bytea value, so we simply call byteaout()
 * to serialize the value into text. But it'd be nice to serialize that into
 * a meaningful representation (e.g. for inspection by people).
 *
 * XXX This should probably return something meaningful, similar to what
 * pg_dependencies_out does. Not sure how to deal with the deduplicated
 * values, though - do we want to expand that or not?
 */

Yeah, that was the simplest output function possible, it didn't seem
worth it to implement something more advanced. pg_mcv_list_items() is
more convenient for most needs, but it's quite far from the on-disk
representation.

That's actually a good question - how closely should the exported data
be to the on-disk format? I'd say we should keep it abstract, not tied
to the details of the on-disk format (which might easily change between
versions).

I'm a bit confused about the JSON schema used in pg_statistic_export
view, though. It simply serializes stakinds, stavalues, stanumbers into
arrays ... which works, but why not to use the JSON nesting? I mean,
there could be a nested document for histogram, MCV, ... with just the
correct fields.

{
...
histogram : { stavalues: [...] },
mcv : { stavalues: [...], stanumbers: [...] },
...
}

and so on. Also, what does TRIVIAL stand for?

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Mon, Nov 6, 2023 at 4:16 PM Corey Huinker <corey.huinker@gmail.com> wrote:

Yeah, that use makes sense as well, and if so then postgres_fdw would likely need to be aware of the appropriate query for several versions back - they change, not by much, but they do change. So now we'd have each query text in three places: a system view, postgres_fdw, and the bin/scripts pre-upgrade program. So I probably should consider the best way to share those in the codebase.

Attached is v2 of this patch.

While applying Patch, I noticed few Indentation issues:
1) D:\Project\Postgres>git am v2-0003-Add-pg_import_rel_stats.patch
.git/rebase-apply/patch:1265: space before tab in indent.
errmsg("invalid statistics
format, stxndeprs must be array or null");
.git/rebase-apply/patch:1424: trailing whitespace.
errmsg("invalid statistics format,
stxndistinct attnums elements must be strings, but one is %s",
.git/rebase-apply/patch:1315: new blank line at EOF.
+
warning: 3 lines add whitespace errors.
Applying: Add pg_import_rel_stats().

2) D:\Project\Postgres>git am v2-0004-Add-pg_export_stats-pg_import_stats.patch
.git/rebase-apply/patch:282: trailing whitespace.
const char *export_query_v14 =
.git/rebase-apply/patch:489: trailing whitespace.
const char *export_query_v12 =
.git/rebase-apply/patch:648: trailing whitespace.
const char *export_query_v10 =
.git/rebase-apply/patch:826: trailing whitespace.

.git/rebase-apply/patch:1142: trailing whitespace.
result = PQexec(conn,
warning: squelched 4 whitespace errors
warning: 9 lines add whitespace errors.
Applying: Add pg_export_stats, pg_import_stats.

Thanks and Regards,
Shubham Khanna.

On Tue, Oct 31, 2023 at 12:55 PM Corey Huinker <corey.huinker@gmail.com> wrote:

Yeah, that use makes sense as well, and if so then postgres_fdw would likely need to be aware of the appropriate query for several versions back - they change, not by much, but they do change. So now we'd have each query text in three places: a system view, postgres_fdw, and the bin/scripts pre-upgrade program. So I probably should consider the best way to share those in the codebase.

Attached is v2 of this patch.

New features:
* imports index statistics. This is not strictly accurate: it re-computes index statistics the same as ANALYZE does, which is to say it derives those stats entirely from table column stats, which are imported, so in that sense we're getting index stats without touching the heap.
* now support extended statistics except for MCV, which is currently serialized as an difficult-to-decompose bytea field.
* bare-bones CLI script pg_export_stats, which extracts stats on databases back to v12 (tested) and could work back to v10.
* bare-bones CLI script pg_import_stats, which obviously only works on current devel dbs, but can take exports from older versions.

I did a small experiment with your patches. In a separate database
"fdw_dst" I created a table t1 and populated it with 100K rows
#create table t1 (a int, b int);
#insert into t1 select i, i + 1 from generate_series(1, 100000) i;
#analyse t1;

In database "postgres" on the same server, I created a foreign table
pointing to t1
#create server fdw_dst_server foreign data wrapper postgres_fdw
OPTIONS ( dbname 'fdw_dst', port '5432');
#create user mapping for public server fdw_dst_server ;
#create foreign table t1 (a int, b int) server fdw_dst_server;

The estimates are off
#explain select * from t1 where a = 100;
QUERY PLAN
-----------------------------------------------------------
Foreign Scan on t1 (cost=100.00..142.26 rows=13 width=8)
(1 row)

Export and import stats for table t1
$ pg_export_stats -d fdw_dst | pg_import_stats -d postgres

gives accurate estimates
#explain select * from t1 where a = 100;
QUERY PLAN
-----------------------------------------------------------
Foreign Scan on t1 (cost=100.00..1793.02 rows=1 width=8)
(1 row)

In this simple case it's working like a charm.

Then I wanted to replace all ANALYZE commands in postgres_fdw.sql with
import and export of statistics. But I can not do that since it
requires table names to match. Foreign table metadata stores the
mapping between local and remote table as well as column names. Import
can use that mapping to install the statistics appropriately. We may
want to support a command or function in postgres_fdw to import
statistics of all the tables that point to a given foreign server.
That may be some future work based on your current patches.

I have not looked at the code though.

--
Best Wishes,
Ashutosh Bapat

Yeah, that was the simplest output function possible, it didn't seem

worth it to implement something more advanced. pg_mcv_list_items() is

more convenient for most needs, but it's quite far from the on-disk
representation.

I was able to make it work.

That's actually a good question - how closely should the exported data
be to the on-disk format? I'd say we should keep it abstract, not tied
to the details of the on-disk format (which might easily change between
versions).

For the most part, I chose the exported data json types and formats in a
way that was the most accommodating to cstring input functions. So, while
so many of the statistic values are obviously only ever integers/floats,
those get stored as a numeric data type which lacks direct
numeric->int/float4/float8 functions (though we could certainly create
them, and I'm not against that), casting them to text lets us leverage
pg_strtoint16, etc.

I'm a bit confused about the JSON schema used in pg_statistic_export
view, though. It simply serializes stakinds, stavalues, stanumbers into
arrays ... which works, but why not to use the JSON nesting? I mean,
there could be a nested document for histogram, MCV, ... with just the
correct fields.

{
...
histogram : { stavalues: [...] },
mcv : { stavalues: [...], stanumbers: [...] },
...
}

That's a very good question. I went with this format because it was fairly
straightforward to code in SQL using existing JSON/JSONB functions, and
that's what we will need if we want to export statistics on any server
currently in existence. I'm certainly not locked in with the current
format, and if it can be shown how to transform the data into a superior
format, I'd happily do so.

and so on. Also, what does TRIVIAL stand for?

It's currently serving double-duty for "there are no stats in this slot"
and the situations where the stats computation could draw no conclusions
about the data.

Attached is v3 of this patch. Key features are:

* Handles regular pg_statistic stats for any relation type.
* Handles extended statistics.
* Export views pg_statistic_export and pg_statistic_ext_export to allow
inspection of existing stats and saving those values for later use.
* Import functions pg_import_rel_stats() and pg_import_ext_stats() which
take Oids as input. This is intentional to allow stats from one object to
be imported into another object.
* User scripts pg_export_stats and pg_import stats, which offer a primitive
way to serialize all the statistics of one database and import them into
another.
* Has regression test coverage for both with a variety of data types.
* Passes my own manual test of extracting all of the stats from a v15
version of the popular "dvdrental" example database, as well as some
additional extended statistics objects, and importing them into a
development database.
* Import operations never touch the heap of any relation outside of
pg_catalog. As such, this should be significantly faster than even the most
cursory analyze operation, and therefore should be useful in upgrade
situations, allowing the database to work with "good enough" stats more
quickly, while still allowing for regular autovacuum to recalculate the
stats "for real" at some later point.

The relation statistics code was adapted from similar features in
analyze.c, but is now done in a query context. As before, the
rowcount/pagecount values are updated on pg_class in a non-transactional
fashion to avoid table bloat, while the updates to pg_statistic are
pg_statistic_ext_data are done transactionally.

The existing statistics _store() functions were leveraged wherever
practical, so much so that the extended statistics import is mostly just
adapting the existing _build() functions into _import() functions which
pull their values from JSON rather than computing the statistics.

Current concerns are:

1. I had to code a special-case exception for MCELEM stats on array data
types, so that the array_in() call uses the element type rather than the
array type. I had assumed that the existing exmaine_attribute() functions
would have properly derived the typoid for that column, but it appears to
not be the case, and I'm clearly missing how the existing code gets it
right.
2. This hasn't been tested with external custom datatypes, but if they have
a custom typanalyze function things should be ok.
3. While I think I have cataloged all of the schema-structural changes to
pg_statistic[_ext[_data]] since version 10, I may have missed a case where
the schema stayed the same, but the values are interpreted differently.
4. I don't yet have a complete vision for how these tools will be used by
pg_upgrade and pg_dump/restore, the places where these will provide the
biggest win for users.

On 13/12/2023 17:26, Corey Huinker wrote:> 4. I don't yet have a
complete vision for how these tools will be used

by pg_upgrade and pg_dump/restore, the places where these will provide
the biggest win for users.

Some issues here with docs:

func.sgml:28465: parser error : Opening and ending tag mismatch: sect1
line 26479 and sect2
</sect2>
^

Also, as I remember, we already had some attempts to invent dump/restore
statistics [1,2]. They were stopped with the problem of type
verification. What if the definition of the type has changed between the
dump and restore? As I see in the code, Importing statistics you just
check the column name and don't see into the type.

[1]: Backup and recovery of pg_statistic /messages/by-id/724322880.K8vzik8zPz@abook
/messages/by-id/724322880.K8vzik8zPz@abook
[2]: Re: Ideas about a better API for postgres_fdw remote estimates /messages/by-id/7a40707d-1758-85a2-7bb1-6e5775518e64@postgrespro.ru
/messages/by-id/7a40707d-1758-85a2-7bb1-6e5775518e64@postgrespro.ru

--
regards,
Andrei Lepikhov
Postgres Professional

On Fri, Dec 15, 2023 at 3:36 AM Andrei Lepikhov <a.lepikhov@postgrespro.ru>
wrote:

On 13/12/2023 17:26, Corey Huinker wrote:> 4. I don't yet have a
complete vision for how these tools will be used

by pg_upgrade and pg_dump/restore, the places where these will provide
the biggest win for users.

Some issues here with docs:

func.sgml:28465: parser error : Opening and ending tag mismatch: sect1
line 26479 and sect2
</sect2>
^

Apologies, will fix.

Also, as I remember, we already had some attempts to invent dump/restore
statistics [1,2]. They were stopped with the problem of type
verification. What if the definition of the type has changed between the
dump and restore? As I see in the code, Importing statistics you just
check the column name and don't see into the type.

We look up the imported statistics via column name, that is correct.

However, the values in stavalues and mcv and such are stored purely as
text, so they must be casted using the input functions for that particular
datatype. If that column definition changed, or the underlying input
function changed, the stats import of that particular table would fail. It
should be noted, however, that those same input functions were used to
bring the data into the table via restore, so it would have already failed
on that step. Either way, the structure of the table has effectively
changed, so failure to import those statistics would be a good thing.

[1] Backup and recovery of pg_statistic
/messages/by-id/724322880.K8vzik8zPz@abook

That proposal sought to serialize enough information on the old server such
that rows could be directly inserted into pg_statistic on the new server.
As was pointed out at the time, version N of a server cannot know what the
format of pg_statistic will be in version N+1.

This patch avoids that problem by inspecting the structure of the object to
be faux-analyzed, and using that to determine what parts of the JSON to
fetch, and what datatype to cast values to in cases like mcv and
stavaluesN. The exported JSON has no oids in it whatseover, all elements
subject to casting on import have already been cast to text, and the record
returned has the server version number of the producing system, and the
import function can use that to determine how it interprets the data it
finds.

[2] Re: Ideas about a better API for postgres_fdw remote estimates

/messages/by-id/7a40707d-1758-85a2-7bb1-6e5775518e64@postgrespro.ru

This one seems to be pulling oids from the remote server, and we can't
guarantee their stability across systems, especially for objects and
operators from extensions. I tried to go the route of extracting the full
text name of an operator, but discovered that the qualified names, in
addition to being unsightly, were irrelevant because we can't insert stats
that disagree about type with the attribute/expression. So it didn't matter
what type the remote system thought it had, the local system was going to
coerce it into the expected data type or ereport() trying.

I think there is hope for having do_analyze() run a remote query fetching
the remote table's exported stats and then storing them locally, possibly
after some modification, and that would save us from having to sample a
remote table.

Hi,

I finally had time to look at the last version of the patch, so here's a
couple thoughts and questions in somewhat random order. Please take this
as a bit of a brainstorming and push back if you disagree some of my
comments.

In general, I like the goal of this patch - not having statistics is a
common issue after an upgrade, and people sometimes don't even realize
they need to run analyze. So, it's definitely worth improving.

I'm not entirely sure about the other use case - allowing people to
tweak optimizer statistics on a running cluster, to see what would be
the plan in that case. Or more precisely - I agree that would be an
interesting and useful feature, but maybe the interface should not be
the same as for the binary upgrade use case?

interfaces
----------

When I thought about the ability to dump/load statistics in the past, I
usually envisioned some sort of DDL that would do the export and import.
So for example we'd have EXPORT STATISTICS / IMPORT STATISTICS commands,
or something like that, and that'd do all the work. This would mean
stats are "first-class citizens" and it'd be fairly straightforward to
add this into pg_dump, for example. Or at least I think so ...

Alternatively we could have the usual "functional" interface, with a
functions to export/import statistics, replacing the DDL commands.

Unfortunately, none of this works for the pg_upgrade use case, because
existing cluster versions would not support this new interface, of
course. That's a significant flaw, as it'd make this useful only for
upgrades of future versions.

So I think for the pg_upgrade use case, we don't have much choice other
than using "custom" export through a view, which is what the patch does.

However, for the other use case (tweaking optimizer stats) this is not
really an issue - that always happens on the same instance, so no issue
with not having the "export" function and so on. I'd bet there are more
convenient ways to do this than using the export view. I'm sure it could
share a lot of the infrastructure, ofc.

I suggest we focus on the pg_upgrade use case for now. In particular, I
think we really need to find a good way to integrate this into
pg_upgrade. I'm not against having custom CLI commands, but it's still a
manual thing - I wonder if we could extend pg_dump to dump stats, or
make it built-in into pg_upgrade in some way (possibly disabled by
default, or something like that).

JSON format
-----------

As for the JSON format, I wonder if we need that at all? Isn't that an
unnecessary layer of indirection? Couldn't we simply dump pg_statistic
and pg_statistic_ext_data in CSV, or something like that? The amount of
new JSONB code seems to be very small, so it's OK I guess.

I'm still a bit unsure about the "right" JSON schema. I find it a bit
inconvenient that the JSON objects mimic the pg_statistic schema very
closely. In particular, it has one array for stakind values, another
array for stavalues, array for stanumbers etc. I understand generating
this JSON in SQL is fairly straightforward, and for the pg_upgrade use
case it's probably OK. But my concern is it's not very convenient for
the "manual tweaking" use case, because the "related" fields are
scattered in different parts of the JSON.

That's pretty much why I envisioned a format "grouping" the arrays for a
particular type of statistics (MCV, histogram) into the same object, as
for example in

{
"mcv" : {"values" : [...], "frequencies" : [...]}
"histogram" : {"bounds" : [...]}
}

But that's probably much harder to generate from plain SQL (at least I
think so, I haven't tried).

data missing in the export
--------------------------

I think the data needs to include more information. Maybe not for the
pg_upgrade use case, where it's mostly guaranteed not to change, but for
the "manual tweak" use case it can change. And I don't think we want two
different formats - we want one, working for everything.

Consider for example about the staopN and stacollN fields - if we clone
the stats from one table to the other, and the table uses different
collations, will that still work? Similarly, I think we should include
the type of each column, because it's absolutely not guaranteed the
import function will fail if the type changes. For example, if the type
changes from integer to text, that will work, but the ordering will
absolutely not be the same. And so on.

For the extended statistics export, I think we need to include also the
attribute names and expressions, because these can be different between
the statistics. And not only that - the statistics values reference the
attributes by positions, but if the two tables have the attributes in a
different order (when ordered by attnum), that will break stuff.

more strict checks
------------------

I think the code should be a bit more "defensive" when importing stuff,
and do at least some sanity checks. For the pg_upgrade use case this
should be mostly non-issue (except for maybe helping to detect bugs
earlier), but for the "manual tweak" use case it's much more important.

By this I mean checks like:

* making sure the frequencies in MCV lists are not obviously wrong
(outside [0,1], sum exceeding > 1.0, etc.)

* cross-checking that stanumbers/stavalues make sense (e.g. that MCV has
both arrays while histogram has only stavalues, that the arrays have
the same length for MCV, etc.)

* checking there are no duplicate stakind values (e.g. two MCV lists)

This is another reason why I was thinking the current JSON format may be
a bit inconvenient, because it loads the fields separately, making the
checks harder. But I guess it could be done after loading everything, as
a separate phase.

Not sure if all the checks need to be regular elog(ERROR), perhaps some
could/should be just asserts.

minor questions
---------------

1) Should the views be called pg_statistic_export or pg_stats_export?
Perhaps pg_stats_export is better, because the format is meant to be
human-readable (rather than 100% internal).

2) It's not very clear what "non-transactional update" of pg_class
fields actually means. Does that mean we update the fields in-place,
can't be rolled back, is not subject to MVCC or what? I suspect users
won't know unless the docs say that explicitly.

3) The "statistics.c" code should really document the JSON structure. Or
maybe if we plan to use this for other purposes, it should be documented
in the SGML?

Actually, this means that the use supported cases determine if the
expected JSON structure is part of the API. For pg_upgrade we could keep
it as "internal" and maybe change it as needed, but for "manual tweak"
it'd become part of the public API.

4) Why do we need the separate "replaced" flags in import_stakinds? Can
it happen that collreplaces/opreplaces differ from kindreplaces?

5) What happens in we import statistics for a table that already has
some statistics? Will this discard the existing statistics, or will this
merge them somehow? (I think we should always discard the existing
stats, and keep only the new version.)

6) What happens if we import extended stats with mismatching definition?
For example, what if the "new" statistics object does not have "mcv"
enabled, but the imported data do include MCV? What if the statistics do
have the same number of "dimensions" but not the same number of columns
and expressions?

7) The func.sgml additions in 0007 seems a bit strange, particularly the
first sentence of the paragraph.

8) While experimenting with the patch, I noticed this:

create table t (a int, b int, c text);
create statistics s on a, b, c, (a+b), (a-b) from t;

create table t2 (a text, b text, c text);
create statistics s2 on a, c from t2;

select pg_import_ext_stats(
(select oid from pg_statistic_ext where stxname = 's2'),
(select server_version_num from pg_statistic_ext_export
where ext_stats_name = 's'),
(select stats from pg_statistic_ext_export
where ext_stats_name = 's'));

WARNING: statistics import has 5 mcv dimensions, but the expects 2.
Skipping excess dimensions.
ERROR: statistics import has 5 mcv dimensions, but the expects 2.
Skipping excess dimensions.

I guess we should not trigger WARNING+ERROR with the same message.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Tue, Dec 26, 2023 at 02:18:56AM +0100, Tomas Vondra wrote:

interfaces
----------

When I thought about the ability to dump/load statistics in the past, I
usually envisioned some sort of DDL that would do the export and import.
So for example we'd have EXPORT STATISTICS / IMPORT STATISTICS commands,
or something like that, and that'd do all the work. This would mean
stats are "first-class citizens" and it'd be fairly straightforward to
add this into pg_dump, for example. Or at least I think so ...

Alternatively we could have the usual "functional" interface, with a
functions to export/import statistics, replacing the DDL commands.

Unfortunately, none of this works for the pg_upgrade use case, because
existing cluster versions would not support this new interface, of
course. That's a significant flaw, as it'd make this useful only for
upgrades of future versions.

So I think for the pg_upgrade use case, we don't have much choice other
than using "custom" export through a view, which is what the patch does.

However, for the other use case (tweaking optimizer stats) this is not
really an issue - that always happens on the same instance, so no issue
with not having the "export" function and so on. I'd bet there are more
convenient ways to do this than using the export view. I'm sure it could
share a lot of the infrastructure, ofc.

I suggest we focus on the pg_upgrade use case for now. In particular, I
think we really need to find a good way to integrate this into
pg_upgrade. I'm not against having custom CLI commands, but it's still a
manual thing - I wonder if we could extend pg_dump to dump stats, or
make it built-in into pg_upgrade in some way (possibly disabled by
default, or something like that).

I have some thoughts on this too. I understand the desire to add
something that can be used for upgrades _to_ PG 17, but I am concerned
that this will give us a cumbersome API that will hamper future
development. I think we should develop the API we want, regardless of
how useful it is for upgrades _to_ PG 17, and then figure out what
short-term hacks we can add to get it working for upgrades _to_ PG 17;
these hacks can eventually be removed. Even if they can't be removed,
they are export-only and we can continue developing the import SQL
command cleanly, and I think import is going to need the most long-term
maintenance.

I think we need a robust API to handle two cases:

* changes in how we store statistics
* changes in how how data type values are represented in the statistics

We have had such changes in the past, and I think these two issues are
what have prevented import/export of statistics up to this point.
Developing an API that doesn't cleanly handle these will cause long-term
pain.

In summary, I think we need an SQL-level command for this. I think we
need to embed the Postgres export version number into the statistics
export file (maybe in the COPY header), and then load the file via COPY
internally (not JSON) into a temporary table that we know matches the
exported Postgres version. We then need to use SQL to make any
adjustments to it before loading it into pg_statistic. Doing that
internally in JSON just isn't efficient. If people want JSON for such
cases, I suggest we add a JSON format to COPY.

I think we can then look at pg_upgrade to see if we can simulate the
export action which can use the statistics import SQL command.

--
Bruce Momjian <bruce@momjian.us> https://momjian.us
EDB https://enterprisedb.com

Only you can decide what is important to you.

Bruce Momjian <bruce@momjian.us> writes:

I think we need a robust API to handle two cases:

* changes in how we store statistics
* changes in how how data type values are represented in the statistics

We have had such changes in the past, and I think these two issues are
what have prevented import/export of statistics up to this point.
Developing an API that doesn't cleanly handle these will cause long-term
pain.

Agreed.

In summary, I think we need an SQL-level command for this.

I think a SQL command is an actively bad idea. It'll just add development
and maintenance overhead that we don't need. When I worked on this topic
years ago at Salesforce, I had things set up with simple functions, which
pg_dump would invoke by writing more or less

SELECT pg_catalog.load_statistics(....);

This has a number of advantages, not least of which is that an extension
could plausibly add compatible functions to older versions. The trick,
as you say, is to figure out what the argument lists ought to be.
Unfortunately I recall few details of what I wrote for Salesforce,
but I think I had it broken down in a way where there was a separate
function call occurring for each pg_statistic "slot", thus roughly

load_statistics(table regclass, attname text, stakind int, stavalue ...);

I might have had a separate load_statistics_xxx function for each
stakind, which would ease the issue of deciding what the datatype
of "stavalue" is. As mentioned already, we'd also need some sort of
version identifier, and we'd expect the load_statistics() functions
to be able to transform the data if the old version used a different
representation. I agree with the idea that an explicit representation
of the source table attribute's type would be wise, too.

regards, tom lane

On 12/26/23 20:19, Tom Lane wrote:

Bruce Momjian <bruce@momjian.us> writes:

I think we need a robust API to handle two cases:

* changes in how we store statistics
* changes in how how data type values are represented in the statistics

We have had such changes in the past, and I think these two issues are
what have prevented import/export of statistics up to this point.
Developing an API that doesn't cleanly handle these will cause long-term
pain.

Agreed.

I agree the format is important - we don't want to end up with a format
that's cumbersome or inconvenient to use. But I don't think the proposed
format is somewhat bad in those respects - it mostly reflects how we
store statistics and if I was designing a format for humans, it might
look a bit differently. But that's not the goal here, IMHO.

I don't quite understand the two cases above. Why should this affect how
we store statistics? Surely, making the statistics easy to use for the
optimizer is much more important than occasional export/import.

In summary, I think we need an SQL-level command for this.

I think a SQL command is an actively bad idea. It'll just add development
and maintenance overhead that we don't need. When I worked on this topic
years ago at Salesforce, I had things set up with simple functions, which
pg_dump would invoke by writing more or less

SELECT pg_catalog.load_statistics(....);

This has a number of advantages, not least of which is that an extension
could plausibly add compatible functions to older versions. The trick,
as you say, is to figure out what the argument lists ought to be.
Unfortunately I recall few details of what I wrote for Salesforce,
but I think I had it broken down in a way where there was a separate
function call occurring for each pg_statistic "slot", thus roughly

load_statistics(table regclass, attname text, stakind int, stavalue ...);

I might have had a separate load_statistics_xxx function for each
stakind, which would ease the issue of deciding what the datatype
of "stavalue" is. As mentioned already, we'd also need some sort of
version identifier, and we'd expect the load_statistics() functions
to be able to transform the data if the old version used a different
representation. I agree with the idea that an explicit representation
of the source table attribute's type would be wise, too.

Yeah, this is pretty much what I meant by "functional" interface. But if
I said maybe the format implemented by the patch is maybe too close to
how we store the statistics, then this has exactly the same issue. And
it has other issues too, I think - it breaks down the stats into
multiple function calls, so ensuring the sanity/correctness of whole
sets of statistics gets much harder, I think.

I'm not sure about the extension idea. Yes, we could have an extension
providing such functions, but do we have any precedent of making
pg_upgrade dependent on an external extension? I'd much rather have
something built-in that just works, especially if we intend to make it
the default behavior (which I think should be our aim here).

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Wed, Dec 27, 2023 at 01:08:47PM +0100, Tomas Vondra wrote:

On 12/26/23 20:19, Tom Lane wrote:

Bruce Momjian <bruce@momjian.us> writes:

I think we need a robust API to handle two cases:

* changes in how we store statistics
* changes in how how data type values are represented in the statistics

We have had such changes in the past, and I think these two issues are
what have prevented import/export of statistics up to this point.
Developing an API that doesn't cleanly handle these will cause long-term
pain.

Agreed.

I agree the format is important - we don't want to end up with a format
that's cumbersome or inconvenient to use. But I don't think the proposed
format is somewhat bad in those respects - it mostly reflects how we
store statistics and if I was designing a format for humans, it might
look a bit differently. But that's not the goal here, IMHO.

I don't quite understand the two cases above. Why should this affect how
we store statistics? Surely, making the statistics easy to use for the
optimizer is much more important than occasional export/import.

The two items above were to focus on getting a solution that can easily
handle future statistics storage changes. I figured we would want to
manipulate the data as a table internally so I am confused why we would
export JSON instead of a COPY format. I didn't think we were changing
how we internall store or use the statistics.

In summary, I think we need an SQL-level command for this.

I think a SQL command is an actively bad idea. It'll just add development
and maintenance overhead that we don't need. When I worked on this topic
years ago at Salesforce, I had things set up with simple functions, which
pg_dump would invoke by writing more or less

SELECT pg_catalog.load_statistics(....);

This has a number of advantages, not least of which is that an extension
could plausibly add compatible functions to older versions. The trick,
as you say, is to figure out what the argument lists ought to be.
Unfortunately I recall few details of what I wrote for Salesforce,
but I think I had it broken down in a way where there was a separate
function call occurring for each pg_statistic "slot", thus roughly

load_statistics(table regclass, attname text, stakind int, stavalue ...);

I might have had a separate load_statistics_xxx function for each
stakind, which would ease the issue of deciding what the datatype
of "stavalue" is. As mentioned already, we'd also need some sort of
version identifier, and we'd expect the load_statistics() functions
to be able to transform the data if the old version used a different
representation. I agree with the idea that an explicit representation
of the source table attribute's type would be wise, too.

Yeah, this is pretty much what I meant by "functional" interface. But if
I said maybe the format implemented by the patch is maybe too close to
how we store the statistics, then this has exactly the same issue. And
it has other issues too, I think - it breaks down the stats into
multiple function calls, so ensuring the sanity/correctness of whole
sets of statistics gets much harder, I think.

I was suggesting an SQL command because this feature is going to need a
lot of options and do a lot of different things, I am afraid, and a
single function might be too complex to manage.

I'm not sure about the extension idea. Yes, we could have an extension
providing such functions, but do we have any precedent of making
pg_upgrade dependent on an external extension? I'd much rather have
something built-in that just works, especially if we intend to make it
the default behavior (which I think should be our aim here).

Uh, an extension seems nice to allow people in back branches to install
it, but not for normal usage.

--
Bruce Momjian <bruce@momjian.us> https://momjian.us
EDB https://enterprisedb.com

Only you can decide what is important to you.

On Mon, Dec 25, 2023 at 8:18 PM Tomas Vondra <tomas.vondra@enterprisedb.com>
wrote:

Hi,

I finally had time to look at the last version of the patch, so here's a
couple thoughts and questions in somewhat random order. Please take this
as a bit of a brainstorming and push back if you disagree some of my
comments.

In general, I like the goal of this patch - not having statistics is a
common issue after an upgrade, and people sometimes don't even realize
they need to run analyze. So, it's definitely worth improving.

I'm not entirely sure about the other use case - allowing people to
tweak optimizer statistics on a running cluster, to see what would be
the plan in that case. Or more precisely - I agree that would be an
interesting and useful feature, but maybe the interface should not be
the same as for the binary upgrade use case?

interfaces
----------

When I thought about the ability to dump/load statistics in the past, I
usually envisioned some sort of DDL that would do the export and import.
So for example we'd have EXPORT STATISTICS / IMPORT STATISTICS commands,
or something like that, and that'd do all the work. This would mean
stats are "first-class citizens" and it'd be fairly straightforward to
add this into pg_dump, for example. Or at least I think so ...

Alternatively we could have the usual "functional" interface, with a
functions to export/import statistics, replacing the DDL commands.

Unfortunately, none of this works for the pg_upgrade use case, because
existing cluster versions would not support this new interface, of
course. That's a significant flaw, as it'd make this useful only for
upgrades of future versions.

This was the reason I settled on the interface that I did: while we can
create whatever interface we want for importing the statistics, we would
need to be able to extract stats from databases using only the facilities
available in those same databases, and then store that in a medium that
could be conveyed across databases, either by text files or by saving them
off in a side table prior to upgrade. JSONB met the criteria.

So I think for the pg_upgrade use case, we don't have much choice other
than using "custom" export through a view, which is what the patch does.

However, for the other use case (tweaking optimizer stats) this is not
really an issue - that always happens on the same instance, so no issue
with not having the "export" function and so on. I'd bet there are more
convenient ways to do this than using the export view. I'm sure it could
share a lot of the infrastructure, ofc.

So, there is a third use case - foreign data wrappers. When analyzing a
foreign table, at least one in the postgresql_fdw family of foreign
servers, we should be able to send a query specific to the version and
dialect of that server, get back the JSONB, and import those results. That
use case may be more tangible to you than the tweak/tuning case.

JSON format
-----------

As for the JSON format, I wonder if we need that at all? Isn't that an
unnecessary layer of indirection? Couldn't we simply dump pg_statistic
and pg_statistic_ext_data in CSV, or something like that? The amount of
new JSONB code seems to be very small, so it's OK I guess.

I see a few problems with dumping pg_statistic[_ext_data]. The first is
that the importer now has to understand all of the past formats of those
two tables. The next is that the tables are chock full of Oids that don't
necessarily carry forward. I could see us having a text-ified version of
those two tables, but we'd need that for all previous iterations of those
table formats. Instead, I put the burden on the stats export to de-oid the
data and make it *_in() function friendly.

That's pretty much why I envisioned a format "grouping" the arrays for a
particular type of statistics (MCV, histogram) into the same object, as
for example in

{
"mcv" : {"values" : [...], "frequencies" : [...]}
"histogram" : {"bounds" : [...]}
}

I agree that would be a lot more readable, and probably a lot more
debuggable. But I went into this unsure if there could be more than one
stats slot of a given kind per table. Knowing that they must be unique
helps.

But that's probably much harder to generate from plain SQL (at least I
think so, I haven't tried).

I think it would be harder, but far from impossible.

data missing in the export
--------------------------

I think the data needs to include more information. Maybe not for the
pg_upgrade use case, where it's mostly guaranteed not to change, but for
the "manual tweak" use case it can change. And I don't think we want two
different formats - we want one, working for everything.

I"m not against this at all, and I started out doing that, but the
qualified names of operators got _ugly_, and I quickly realized that what I
was generating wouldn't matter, either the input data would make sense for
the attribute's stats or it would fail trying.

Consider for example about the staopN and stacollN fields - if we clone
the stats from one table to the other, and the table uses different
collations, will that still work? Similarly, I think we should include
the type of each column, because it's absolutely not guaranteed the
import function will fail if the type changes. For example, if the type
changes from integer to text, that will work, but the ordering will
absolutely not be the same. And so on.

I can see including the type of the column, that's a lot cleaner than the
operator names for sure, and I can see us rejecting stats or sections of
stats in certain situations. Like in your example, if the collation
changed, then reject all "<" op stats but keep the "=" ones.

For the extended statistics export, I think we need to include also the
attribute names and expressions, because these can be different between
the statistics. And not only that - the statistics values reference the
attributes by positions, but if the two tables have the attributes in a
different order (when ordered by attnum), that will break stuff.

Correct me if I'm wrong, but I thought expression parse trees change _a
lot_ from version to version?

Attribute reordering is a definite vulnerability of the current
implementation, so an attribute name export might be a way to mitigate that.

* making sure the frequencies in MCV lists are not obviously wrong
(outside [0,1], sum exceeding > 1.0, etc.)

+1

* cross-checking that stanumbers/stavalues make sense (e.g. that MCV has
both arrays while histogram has only stavalues, that the arrays have
the same length for MCV, etc.)

To this end, there's an edge-case hack in the code where I have to derive
the array elemtype. I had thought that examine_attribute() or
std_typanalyze() was going to do that for me, but it didn't. Very much want
your input there.

* checking there are no duplicate stakind values (e.g. two MCV lists)

Per previous comment, it's good to learn these restrictions.

Not sure if all the checks need to be regular elog(ERROR), perhaps some
could/should be just asserts.

For this first pass, all errors were one-size fits all, safe for the
WARNING vs ERROR.

minor questions
---------------

1) Should the views be called pg_statistic_export or pg_stats_export?
Perhaps pg_stats_export is better, because the format is meant to be
human-readable (rather than 100% internal).

I have no opinion on what the best name would be, and will go with
consensus.

2) It's not very clear what "non-transactional update" of pg_class
fields actually means. Does that mean we update the fields in-place,
can't be rolled back, is not subject to MVCC or what? I suspect users
won't know unless the docs say that explicitly.

Correct. Cannot be rolled back, not subject to MVCC.

3) The "statistics.c" code should really document the JSON structure. Or
maybe if we plan to use this for other purposes, it should be documented
in the SGML?

I agree, but I also didn't expect the format to survive first contact with
reviewers, so I held back.

4) Why do we need the separate "replaced" flags in import_stakinds? Can
it happen that collreplaces/opreplaces differ from kindreplaces?

That was initially done to maximize the amount of code that could be copied
from do_analyze(). In retrospect, I like how extended statistics just
deletes all the pg_statistic_ext_data rows and replaces them and I would
like to do the same for pg_statistic before this is all done.

5) What happens in we import statistics for a table that already has
some statistics? Will this discard the existing statistics, or will this
merge them somehow? (I think we should always discard the existing
stats, and keep only the new version.)

In the case of pg_statistic_ext_data, the stats are thrown out and replaced
by the imported ones.

In the case of pg_statistic, it's basically an upsert, and any values that
were missing in the JSON are not updated on the existing row. That's
appealing in a tweak situation where you want to only alter one or two bits
of a stat, but not really useful in other situations. Per previous comment,
I'd prefer a clean slate and forcing tweaking use cases to fill in all the
blanks.

6) What happens if we import extended stats with mismatching definition?
For example, what if the "new" statistics object does not have "mcv"
enabled, but the imported data do include MCV? What if the statistics do
have the same number of "dimensions" but not the same number of columns
and expressions?

The importer is currently driven by the types of stats to be expected for
that pg_attribute/pg_statistic_ext. It only looks for things that are
possible for that stat type, and any extra JSON values are ignored.

As mentioned already, we'd also need some sort of
version identifier, and we'd expect the load_statistics() functions
to be able to transform the data if the old version used a different
representation. I agree with the idea that an explicit representation
of the source table attribute's type would be wise, too.

There is a version identifier currently (its own column not embedded in the
JSON), but I discovered that I was able to put the burden on the export
queries to spackle-over the changes in the table structures over time.
Still, I knew that we'd need the version number in there eventually.

Yeah, this is pretty much what I meant by "functional" interface. But if
I said maybe the format implemented by the patch is maybe too close to
how we store the statistics, then this has exactly the same issue. And
it has other issues too, I think - it breaks down the stats into
multiple function calls, so ensuring the sanity/correctness of whole
sets of statistics gets much harder, I think.

Export functions was my original plan, for simplicity, maintenance, etc,
but it seemed like I'd be adding quite a few functions, so the one view
made more sense for an initial version. Also, I knew that pg_dump or some
other stats exporter would have to inline the guts of those functions into
queries for older versions, and adapting a view definition seemed more
straightforward for the reader than function definitions.