Fun with Bugs #97 - On MySQL Bug Reports I am Subscribed to, Part XXXI

Time for the next blog post about MySQL bugs! The previous one covered some bugs I considered interesting in March, this one will be about community bug reports that I've picked up since March 26, 2020. I'd better review bug fixed in MySQL 5.7.30 instead of this, but it is still not released, even though we know it must get some important security fixes based on "Oracle Critical Patch Update Advisory - April 2020" recently published.

As usual, I am mostly interested in replication, InnoDB, optimizer and few more categories. Here is the list:

• Bug #99082 - "Problem with replication: XA transaction, temp tables and row based binlog". As noted by Libor Laichmann, creating temporary tables inside XA transaction leads to broken replication in MySQL 5.7.29, with somewhat misleading error message. This is unfortunate. We do not see any evidence, yet, if 5.6 and 8.0 branches are similarly affected.
• Bug #99094 - "coredump when install information schema plugin". Bug reporter, Lou Shuai, tried to create simple plugin for the information_schema and it crashed MySQL server 8.0.19 upon installation attempt. Bot the plugin source code and patch for the bug were contributed.
• Bug #99100 - "GROUP BY will throw table is full when temptable memory allocation exceed limit". More fun with TempTable storage engine that I blamed previously. This bug was reported by Dean Zhou, who had performed detailed analysis in gdb and suggested a fix. It took some efforts for the bug reporter to get it "Verified" as a regression bug in MySQL 8.0.19 (without a "regression" tag, surely).
• Bug #99101 - "SELECT .. LOCK IN SHARE MODE requires LOCK TABLES grant". Simple regression in MySQL 8.0.11+ (or incompatible change in behavior, if you prefer) vs MySQL 5.7 was found and reported by Matthew Boehm.
• Bug #99136 - "TempTable wastes 1MB for each connection in thread cache". Nikolai Ikhalainen demonstrated this additional memory usage in MySQL 8.0.16 comparing to 5.7 quite clearly.
• Bug #99174 - "Prepare transaction won't rollback if server crash after online ddl prepare stage". This bug was reported by Zhang Xiaojian. Additional debugging code (one DBUG_EXECUTE_IF statement) was added to the source to demonstrate the problem easily and it caused some questionable arguments of a kind:
  

  "First of all, changing our server code in order to cause a bug, can not be considered a repeatable test case."

  But the bug reporter provided more details and clear test case, and had not agreed with the above. The bug was soon "Verified", even though no attempts to check (or explain) if 5.7 may be similarly affected were made in public, so we still have to wonder if this is a regression.
• Bug #99180 - "Accessing freed memory in perfschema when aggregating status vars". Let me just quote Manuel Ung:
  

  "When aggregate_thread_status is called for other threads, it's possible for that thread to have exited and freed the THD between the time we check that the thread was valid, until the time we call get_thd_status_var."

  Ironically, he had to add some conditional my_sleep() call to server code to get a repeatable test case, and this was NOT a problem for a good Oracle engineer to verify the bug immediately. There are still bugs in my beloved Performance Schema. Who could imagine that?
• Bug #99200 - "CREATE USER get stuck on SHOW PROCESSLIST and ps.threads on slave". So, slave may disclose some sensitive information. As demonstrated by Marcelo Altmann, if a query had been rewritten by the parser because it contains sensitive information, it won't be cleaned up when slave's SQL thread applies it, making it visible in SHOW PROCESSLIST and performance_schema.threads. Both 5.7.29 and 8.0.19 are affected. Bug reporter had contributed fixes for both versions. See also another bug in Performance Schema that he reported and contributed a fix for, Bug #99204 - "performance_schema threads table PROCESSLIST_INFO incorrect behaviour".
• Bug #99205 - "xa prepare write binlog while execute failed". Then XA PREPARE from the binary log is executed on slave and... we are in troubles. The bug was reported by Phoenix Zhang. Unfortunately it is still not clear from this verified bug report if MySQL 5.6 and 5.7 are similarly affected (probably they are).
• Bug #99206 - "lock_wait_timeout is taking twice time while adding new partition in table". Nice finding by Lalit Choudhary. MySQL 8.0.19 is not affected by this bug.
• Bug #99244 - "bad performance in hash join because join with no join condition between tables". This optimizer bug (wrong join order when new hash join optimization is used) was reported by Chunyang Xu. But note also a simple test case and perf outputs contributed later by Shane Bester.
• Bug #99257 - "Inconsistent output and wrong ORDER BY Sorting for query with LIMIT". Yet another 5.7-only optimizer bug found by Lalit Choudhary. My quick test presented in a comment shows that MySQL 5.6.27 produced correct results, so this is a regression bug.
• Bug #99273 - "Item_ref in Having Clause points to invalid outer table field". This may lead to wrong results for simple enough queries. This regression bug that affects both MySQL 5.7.29 and 8.0.19 was reported by Shanshan Ying.
• Bug #99286 - "Concurrent update cause crash in row_search_mvcc". This great bug report with code analysis, fix suggested and a test case with detailed instructions was created by Zkong Kong. It was promptly verified, but I still miss any documented attempt to check (by running the test case or by code analysis if the bug applies only to 5.7.29 or MySQL 8.0.x is also potentially affected. For now I'll try to remember this assertion line:
  

  InnoDB: Assertion failure in thread 47758491551488 in file rem0rec.cc line 586

  and "crash with row_search_mvcc in backtrace". Who know when I hit something similar and what fork/version it will be...

Rainy birthday at the seaside in Cap-d'Ail. Almost 15 years of my 50 were spent checking new MySQL bug reports almost every day.

To summarize:

1. We still see many regression bugs in recent versions of MySQL 5.7.x and 8.0.x, often without "regression" tag.
2. There are cases when the bug is verified, but there are no clearly documented checks if all GA versions are affected.
3. XA transactions are (and had always been) a real disaster for modern MySQL versions in replication environments - all kinds of replication breakage and inconsistencies are possible.
4. Check "MySQL Bug Reporter Hall of Fame" if you want to know who from MySQL Community had contributed a lot of bug reports over last 10 years.