Sunday, November 18, 2018

Fun with Bugs #72 - On MySQL Bug Reports I am Subscribed to, Part IX

I've subscribed to more than 60 new bug reports since my previous post in this series. It means that I'd need 4-5 posts to cover all new subscriptions and reasons behind them. I still plan to write about most of the bug reports I was interested in recently, but for this post I decided to pick up only MySQL 8.0 regression bugs and pay special attention to those that could be handled better or faster by Oracle engineers, as well as those handled perfectly.

The initial reason for this special attention originally was Bug #93085 - "Stall when concurrently execute create/alter user with flush privilege", that caused a lot of interesting Twitter discussions. It took some time, comments (in the bug report and in social media) and pressure from MySQL Community (including yours truly) to get it accepted as a real (regression!) bug to work on, and got "Verified". Unfortunately too often recently I see more time spent on arguing that something is not a bug or can not be reproduced, or is an example of improper usage of some MySQL feature etc instead of simply checking how things worked before MySQL 8.0 and how this changed, to worse.

Another example of "interesting" approach to bugs in MySQL 8.0 is Bug #93102 - "I_S queries don't return column names in the same case as requested.". It's indeed a duplicate of old and well known Bug #84456 - "column names in metadata appearing as uppercase when selecting from I_S" reported at early 8.0 development stage by Shane Bester from Oracle and community user (see Bug #85947). Still, it was decided NOT to fix it and tell users to rely on workaround, while this breaks application compatibility and is a regression.

Take a look at Bug #92998 - "After finishing the backup with mysqldump, mysql crashes and restarts" also. It ended up in "Unsupported" status, with some statements that "Dumping and restoring data between different 8.0 releases is not supported". This can be classified as a regression by itself. What I miss is a link to the manual saying it's not supported (was not able to find it in 5 minutes) any explanation of crash and restart - supported or not, running mysqldump should NOT cause server restarts in a general case. I think this bug report could end up in many statuses, but of them all "Unsupported" is hardly correct.

This my photo is far from ideal and can be criticized from different points of view, but there is no point to argue with the fact that it shows clouds in the sky. I wish the fact that MySQL 8.0 GA releases still have regression bugs is accepted with less arguing and more attention.

Now let me continue with a list of recently reported regression bugs in MySQL 8.0 that were handled mostly properly:

  • Bug #93215 - "Case statement use in conditional with sub_part index regression in 8.0". MySQL of versions < 8.0 (and MariaDB 10.3.7) work as expected also. The bug was verified fast, but it still misses explicit "regression" tag.
  • Bug #93214 - "LIMIT is applied before HAVING when you have a subquery". The bug was "Verified" quickly, but I still miss the exact 8.0.x version(s) affected and the results of checking with older versions. I strongly suspect it's a regression, as MariaDB 10.3.7 provides expected result:
MariaDB [test]> CREATE TABLE test (id INT PRIMARY KEY, value INT);
Query OK, 0 rows affected (0.510 sec)
MariaDB [test]> INSERT INTO test VALUES (1, 99), (2,98), (3, 97);
Query OK, 3 rows affected (0.057 sec)
Records: 3  Duplicates: 0  Warnings: 0
MariaDB [test]> SELECT, (SELECT t2.value FROM test t2 WHERE = AS sub_value FROM test t1 HAVING sub_value = 99 ORDER BY value LIMIT 1;
| id | sub_value |
|  1 |        99 |
1 row in set (0.116 sec)
            • Bug #93170 - "undo truncation in 8.0.13 is not crash safe". The bug was quickly verified (after all, it's a failure of existing innodb_undo.truncate_recover MTR test case), but had not got "regression" tag. I am still not sure how it was missed during regular testing and ended up in the MySQL 8.0.13 release.
            • Bug #93147 - "Upgrade to 8.0.13 from 8.0.11 fails". In pre-8.0 releases there was no strict need to update to every intermediate minor version, so it's also a regression of a kind for any production DBA.
            • Bug #92979 - "MySQL 8.0 performance degradation on INSERT with foreign_key_checks=0". This is a verified performance regression comparing to MySQL 5.7, but "regression" tag is still missing. 

            To summarize, there are some regressions noted by community users recently in MySQL 8.0 GA releases. Some of them were demonstrated with simple test cases, so it's strange they were not noted by Oracle's QA. What's worse, it seems some of Oracle engineers are not ready to accept the fact that the best ever MySQL 8.0 GA release they worked on may get some things done incorrectly and worse than before, so they seem to waste time on useless discussions that everything is OK, work as expected and nothing can be done differently.  I also see some processed and verified bug reports without detailed check for regressions presented to users or even with "regression" tag NOT added when needed.

            I hope this is not going to become a new trend. I wish all community bug reports and features of MySQL get as much attention and detailed study from Oracle employees as (far from perfect) JSON support in MariaDB...

            No comments:

            Post a Comment