Sunday, January 28, 2018

On InnoDB's Persistent Optimizer Statistics

As I put it in recent Facebook post, one of MySQL features that I truly hate is InnoDB's persistent statistics. I think I should clarify this statement. It's great to have a way to save statistics in between server restarts, to have better control on the way it is estimated (even on a per table basis), set it explicitly, check it with simple SELECT. These all are great additions since MySQL 5.6.2+ that I truly appreciate (even if I may not be happy with some implementation details). They helped to make plans for queries against InnoDB more predictable and allow (with some efforts applied) MySQL query optimizer to really work as "optimizer" instead of "randomizer" or "pessimizer" (as some people called it) for InnoDB tables.

What I hate about it mostly is the way innodb_stats_auto_recalc is implemented, and the fact that it is enabled by default since MySQL 5.6.6+ or so. Here is why:
  1. Even if one enables automatic statistics recalculation, she can not be sure that statistics is correct and up to date. One still really has to run ANALYZE TABLE every time after substantial changes of data to be sure, and this comes with a cost (that Percona tried to finally overcome with the fix to lp:1704195 that appeared in their Percona Server 5.7.20-18+). Or enjoy whatever bits of statistics (taken in the process of background recalculation) may be present at the moment and the resulting execution plans...
  2. The details on automatic statistics recalculation are not properly documented (if only in some comments to some bug reports). This changes to better with time (thanks to continue pressure from MySQL community, including your truly, in a form of bug reports), but still most of MySQL users are far from understanding why something happens or NOT happens when this feature is involved.
  3. Implementation introduced background thread (that does dirty reads) to do recalculation, and separate transactions against InnoDB tables where statistics is stored. This complicates implementation, analysis in gdb etc, and introduced more bugs related to coordination of work performed by this thread and other background and user threads.
  4. Recently nobody from Oracle cares much to fix bugs related to this feature.
Let me try to illustrate and prove the points above with some MySQL bug reports (as usual). Many of these bugs are still "Verified" and not fixed as of recent release of recent GA version, MySQL 5.7. The order is somewhat random:
  • Bug #70741, "InnoDB background stats thread is not properly documented" - that's one of my requests to improve documentation. Some more details were added and the bug is closed, but make sure to read the entire comment "[26 Nov 2013 13:41] Vasil Dimov" if you want to understand better how it all works.
  • Bug #70617, "Default persistent stats can cause unexpected long query times" - this is one of bugs that led me to filing the previous documentation request. Check comments by Vasil Dimov there that he made before closing it as "Not a bug"... His comments are the best documentation of the way feature is implemented that I've seen in public. Make your own conclusions.
  • Bug #78289, "Bad execution plan with innodb_stats_persistent enabled" - note that the problem started after pt-osc was applied (to overcome the problems with another feature I hate, "online" ALTER TABLE, most likely). This utility applies RENAME to the table that is altered at the last stage, and as a result statistics just disappears and you have either to wait until it is calculated again in the background, or run ANALYZE... Surely this is "Not a bug".
  • Bug #80178 and Bug #78066 are about cases when SHOW INDEXES may still give wrong results while (with persistent statistics automatic recalculation disabled) one expects the same values we see in the tables where statistics is stored, or just correct ones. Both bugs are still "Verified", even though from the comment in the latter one may assume that the problem may be fixed in recent MySQL 5.7.x.
  • Bug #75428, "InnoDB persistent statistics not persistent enough". The counter of updated rows since last recalculation does not survive restarts, and 10% threshold is not configurable, so if server restarts often and table is big enough, we may get statistics never updated. Still "Verified".
  • Bug #72368, "Empty/zero statistics for imported tablespace until explicit ANALYZE TABLE". Still "Verified", but may be fixed in versions newer than 5.6. Importing tablespace was NOT a reason for automatic statistics recalculation to ever happen for the table...
  • Bug #84940, "MySQL Server crash possibly introduced in InnoDB statistics calculation". This regression bug in 5.6.35 and 5.7.17 was quickly fixed in the next releases, but still caused troubles for some time.
  • Bug #82184, "Table status inconsistent, requires ANALYZE TABLE executed twice". As Andrii Nikitin stated himself, "Most probably second ANALYZE is needed to give some time to purge thread remove old versions of the rows.", in case the table has huge blobs. The bug is still "Verified".
  • Bug #71814, "Persistent stats activity conflicts with mysqldump import of same info". The bug is "Closed" without any reason stated in public (what a surprise...). Note the following comment by Shane Bester (who actually verified and explained the bug):
    "Personally, I don't like that mysqldump dumps the content of these tables that should be auto-generated."
    He had also suggested a workaround to disable persistent statistics (SET GLOBAL innodb_stats_auto_recalc=0; SET GLOBAL innodb_stats_persistent=0;) before importing a dump. The problem here is a race condition between the importing of mysql database and the background statistics thread that can insert rows into the table between the CREATE and LOCK TABLE in the dump. See Bug #80705, "Mysqlpump in default configuration does not work", also, with a clear request: "Do not dump innodb_index_stats or innodb_table_stats". Something to think about.
  • Bug #84654, "Cardinality reset to 0 with persistent statistics when AUTO_INCREMENT touched".  Still "Verified".
  • Bug #84287, "row inserts, statement updates, persistent stats lead to table scans+ lag slaves". It seems automatic recalculation of statistics on slave is not triggered by inserting more rows via row-based replication events. Still "Verified".
  • Bug #82969 , "InnoDB statistics update may temporarily cause wrong index cardinalities". This bug (still "Verified") is my all times favorite race condition in the implementation of persistent statistics by Oracle (well, this one and Bug #82968 that is fixed at least in recent 5.7.x and in MariaDB).
 There are also bugs related to other details of InnoDB persistent statistics implementation:
  • Bug #78401, "ANALYZE TABLE" may assign temporary values to table stats during its execution". Statistics is not updated atomically, it is first reset and then recaclulated. Still "Verified".
  • Bug #86926, "The field table_name (varchar(64)) from mysql.innodb_table_stats can overflow." - this may be not enough for partitioned table, as partition names may be longer. Still "Verified".
  • Bug #67179, "mysql system tables innodb_table_stats,slave_master_info not accessible on clean". This was a famous bug during early days of MySQL 5.6 that affected many users who tried to upgrade. You may still need this file from it one day, to re-create missing InnoDB tables in the mysql database.
  • Bug #80986, "innodb complains about innodb_table_stats even if persistent stats disabled". Still "Verified".
  • Bug #86927, "Renaming a partitioned table does not update mysql.innodb_table_stats.". Fixed recently in MySQL 5.7.21 and 8.0.4.
  • Bug #84455 - the topic of this bug report is different and not relevant, but Shane Bester noted the following in the error log uploaded:
    [Warning] InnoDB: A transaction id in a record of table `mysql`.`innodb_table_stats` 
    is newer than the system-wide maximum.
    
    This is both suspicious and scary. May be related to the way background thread works.
  • Bug #74747, "Failing assertion: index->id == btr_page_get_index_id(page) btr0cur.cc line 899". Yes, this is a debug assertion only provoked explicitly, but note what is written in the error log before it happens:
    InnoDB: Cannot save table statistics for table "db1"."t1": Too many concurrent transactions
    It means background thread opens a separate transaction (no surprise, but still worth to note).
  • Bug #86702, "please disable persistent stats on the mysql.gtid_executed table". This is a valid and verified request to remove the related overhead for this "system" InnoDB table (as it was correctly done for several others).
So, the implementation of InnoDB's Persistent Optimizer Statistics is far from perfect or well documented. One may ask what I'd suggest instead? I often think and state that only engine-independent persistent statistics (in MariaDB style) should exist, and this should be recalculated only by explicit ANALYZE TABLE statement, maybe with more options to set sample size and other details than we have now. No background threads, nothing automatic until automated by the DBA (for this I'd appreciate a package like Oracle's dbms_stats).

This kind of idea is usually not well accepted. One of recent (valid) complains by Domas here were "No, thanks, don't need more MDL holders." and "I prefer lockless versioned stats, when it gets to what I prefer."

Some of the problems mentioned above may be resolved in MySQL 8 (or not) with its atomic data dictionary operations. Other idea presented in MDEV-15020 is to store statistics with data in the same .ibd file. We shell see what may happen, but current implementation, even though it improved a lot since early MySQL 5.6 days, is hardly long term acceptable.

No comments:

Post a Comment