Horton

Re: Statistics Import and Export

Bruce Momjian <bruce@momjian.us>

View thread
From: Bruce Momjian <bruce@momjian.us>
To: Corey Huinker <corey.huinker@gmail.com>
Cc: Jeff Davis <pgsql@j-davis.com>, jian he <jian.universality@gmail.com>, Matthias van de Meent <boekewurm+postgres@gmail.com>, Tom Lane <tgl@sss.pgh.pa.us>, Nathan Bossart <nathandbossart@gmail.com>, Magnus Hagander <magnus@hagander.net>, Stephen Frost <sfrost@snowman.net>, Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>, Peter Smith <smithpb2250@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>, alvherre@alvh.no-ip.org
Date: 2024-11-18T19:47:24Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Change pg_dump default for statistics export.

  2. pg_dump: Adjust reltuples from 0 to -1 for dumps of older versions.

  3. vacuumdb: Don't skip empty relations in --missing-stats-only mode.

  4. pg_dump: Fix query for gathering attribute stats on older versions.

  5. Prevent redeclaration of typedef TocEntry.

  6. Remove unused function parameters in pg_backup_archiver.c.

  7. pg_dump: Retrieve attribute statistics in batches.

  8. pg_dump: Reduce memory usage of dumps with statistics.

  9. Skip second WriteToc() call for custom-format dumps without data.

  10. Add relallfrozen to pg_dump statistics.

  11. Matview statistics depend on matview data.

  12. Add pg_dump --with-{schema|data|statistics} options.

  13. Stats: use schemaname/relname instead of regclass.

  14. CREATE INDEX: do update index stats if autovacuum=off.

  15. Don't convert to and from floats in pg_dump.

  16. CREATE INDEX: don't update table stats if autovacuum=off.

  17. Organize and deduplicate statistics import tests.

  18. Address stats export review comments.

  19. Address stats import review comments.

  20. Add relallfrozen to pg_class

  21. Fix pg_strtof() to not crash on NULL endptr.

  22. Use attnum to identify index columns in pg_restore_attribute_stats().

  23. pg_dump: prepare attribute stats query.

  24. Avoid unnecessary relation stats query in pg_dump.

  25. Remove redundant pg_set_*_stats() variants.

  26. Do not use in-place updates for statistics import.

  27. Fix confusion about data type of pg_class.relpages and relallvisible.

  28. Documentation fixups for dumping statistics.

  29. Trial fix for old cross-version upgrades.

  30. Transfer statistics during pg_upgrade.

  31. Lock table in ShareUpdateExclusive when importing index stats.

  32. Use in-place updates for pg_restore_relation_stats().

  33. Improve error message for replication of generated columns.

  34. pg_dump: Add dumpSchema and dumpData derivative flags.

  35. Disallow modifying statistics on system columns.

  36. Add missing CommandCounterIncrement() in stats import functions.

  37. Add functions pg_restore_relation_stats(), pg_restore_attribute_stats().

  38. Documentation fixup.

  39. Add functions pg_set_attribute_stats() and pg_clear_attribute_stats().

  40. Change pg_*_relation_stats() functions to return type to void.

  41. Disable autovacuum for tables in stats import tests.

  42. Allow pg_set_relation_stats() to set relpages to -1.

  43. Fixup for pg_set_relation_stats().

  44. Create functions pg_set_relation_stats, pg_clear_relation_stats.

  45. Add memory/disk usage for Window aggregate nodes in EXPLAIN.

  46. Improve performance of dumpSequenceData().

  47. Add INJECTION_POINT_CACHED() to run injection points directly from cache

  48. Improve performance of binary_upgrade_set_pg_class_oids().

  49. Improve assertion in mdwritev()

  50. CREATE INDEX: do not update stats during binary upgrade.

  51. Redefine pg_class.reltuples to be -1 before the first VACUUM or ANALYZE.

On Fri, Nov  8, 2024 at 01:25:21PM -0500, Corey Huinker wrote:
>     WHAT IS NOT DONE - EXTENDED STATISTICS
> 
>     It is a general consensus in the community that "nobody uses extended
>     statistics", though I've had difficulty getting actual figures to back this
>     up, even from my own employer. Surveying several vendors at PgConf.EU, the
>     highest estimate was that at most 1% of their customers used extended
>     statistics, though more probably should. This reinforces my belief that a
>     feature that would eliminate a major pain point in upgrades for 99% of
>     customers shouldn't be held back by the fact that the other 1% only have a
>     reduced hassle.
> 
>     However, having relation and attribute statistics carry over on major
>     version upgrades presents a slight problem: running vacuumdb
>     --analyze-in-stages after such an upgrade is completely unnecessary for
>     those without extended statistics, and would actually result in _worse_
>     statistics for the database until the last stage is complete. Granted,
>     we've had great difficulty getting users to know that vacuumdb is a thing
>     that should be run, but word has slowly spread through our own
>     documentation and those "This one simple trick will make your postgres go
>     fast post-upgrade" blog posts. Those posts will continue to lurk in search
>     results long after this feature goes into release, and it would be a rude
>     surprise to users to find out that the extra work they put in to learn
>     about a feature that helped their upgrade in 17 was suddenly detrimental
>     (albeit temporarily) in 18. We should never punish people for only being a
>     little-bit current in their knowledge. Moreover, this surprise would
>     persist even after we add extended statistics import function
>     functionality.
> 
>     I presented this problem to several people at PgConf.EU, and the consensus
>     least-bad solution was that vacuumdb should filter out tables that are not
>     missing any statistics when using options --analyze, --analyze-only, and
>     --analyze-in-stages, with an additional flag for now called --force-analyze
>     to restore the un-filtered functionality. This gives the outcome tree:
> 
>     1. Users who do not have extended statistics and do not use (or not even
>     know about) vacuumdb will be blissfully unaware, and will get better
>     post-upgrade performance.
>     2. Users who do not have extended statistics but use vacuumdb
>     --analyze-in-stages will be pleasantly surprised that the vacuumdb run is
>     almost a no-op, and completes quickly. Those who are surprised by this and
>     re-run vacuumdb --analyze-in-stages will get another no-op.
>     3. Users who have extended statistics and use vacuumdb --analyze-in-stages
>     will get a quicker vacuumdb run, as only the tables with extended stats
>     will pass the filter. Subsequent re-runs of vacuumdb --analyze-in-stages
>     would be the no-op.
>     4. Users who have extended statistics and don't use vacuumdb will still get
>     better performance than they would have without any stats imported.
> 
>     In case anyone is curious, I'm defining "missing stats" as a table/matview
>     with any of the following:
> 
>     1. A table with an attribute that lacks a corresponding pg_statistic row.
>     2. A table with an index with an expression attribute that lacks a
>     corresponding pg_statistic row (non-expression attributes just borrow the
>     pg_statistic row from the table's attribute).
>     3. A table with at least one extended statistic that does not have a
>     corresponding pg_statistic_ext_data row.
> 
>     Note that none of these criteria are concerned with the substance of the
>     statistics (ex. pg_statistic row should have mcv stats but does not),
>     merely their row-existence. 
> 
>     Some rejected alternative solutions were:
> 
>     1. Adding a new option --analyze-missing-stats. While simple, few people
>     would learn about it, knowledge of it would be drowned out by the
>     aforementioned sea of existing blog posts.
>     2. Adding --analyze-missing-stats and making --analyze-in-stages fail with
>     an error message educating the user about --analyze-missing-stats. Users
>     might not see the error, existing tooling wouldn't be able to act on the
>     error, and there are legitimate non-upgrade uses of --analyze-in-stages.
> 
>     MAIN CONCERN GOING FORWARD
> 
>     This change to vacuumdb will require some reworking of the
>     vacuum_one_database() function so that the list of tables analyzed is
>     preserved across the stages, as subsequent stages runs won't be able to
>     detect which tables were previously missing stats.

You seem to be optimizing for people using pg_upgrade, and for people
upgrading to PG 18, without adequately considering people using vacuumdb
in non-pg_upgrade situations, and people using PG 19+.  Let me explain.

First, I see little concern here for how people who use --analyze and
--analyze-only independent of pg_upgrade will be affected by this. 
While I recommend people decrease vacuum and analyze threshold during
non-peak periods:

	https://momjian.us/main/blogs/pgblog/2017.html#January_3_2017

some people might just regenerate all statistics during non-peak periods
using these options.  You can perhaps argue that --analyze-in-stages
would only be used by pg_upgrade so maybe that can be adjusted more
easily.

Second, the API for what --analyze and --analyze-only do will be very
confusing for people running, e.g., PG 20, because the average user
reading the option name will not guess it only adds missing statistics.

I think you need to rethink your approach and just accept that a mention
of the new preserving statistic behavior of pg_upgrade, and the new
vacuumdb API required, will be sufficient.  In summary, I think you need
a new --compute-missing-statistics-only that can be combined with
--analyze, --analyze-only, and --analyze-in-stages to compute only
missing statistics, and document it in the PG 18 release notes.

Frankly, we have a similar problem with partitioned tables:

	https://www.postgresql.org/docs/current/sql-analyze.html
	
	For partitioned tables, ANALYZE gathers statistics by sampling
	rows from all partitions; in addition, it will recurse into
	each partition and update its statistics. Each leaf partition
	is analyzed only once, even with multi-level partitioning. No
	statistics are collected for only the parent table (without data
	from its partitions), because with partitioning it's guaranteed
	to be empty.

-->	The autovacuum daemon does not process partitioned tables, nor
	does it process inheritance parents if only the children are ever
	modified. It is usually necessary to periodically run a manual
	ANALYZE to keep the statistics of the table hierarchy up to date.

Now, you can say partitioned table statistics are not as important as
extended statistics, but that fact remains that we have these two odd
cases where special work must be done to generate statistics.

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

  When a patient asks the doctor, "Am I going to die?", he means 
  "Am I going to die soon?"