Showing posts with label GTID. Show all posts
Showing posts with label GTID. Show all posts

Saturday, April 4, 2020

Fun with Bugs #96 - On MySQL Bug Reports I am Subscribed to, Part XXX

My weekdays are still busy even during this lockdown period, but weekend is a perfect time for yet another blog post about MySQL bugs I considered interesting. Very few followers read my posts on other topics anyway, especially if they have the "M....DB" word...

So, here is a new list of optimizer, replication, InnoDB and few other bugs in GA versions of MySQL that I've subscribed to since the end of February, 2020:
  • Bug #98675 - "LIMIT clause in derived table garbles EXPLAIN Note". It's not a big deal and in specific case presented by Øystein Grøvlen optimizer works properly (note text aside), but some clarifications to the note and/or documentation would still help. Comments are useful to read and shows a great example of proper cooperation from Oracle side.
  • Bug #98719 - "Parallel CREATE TABLE statement binlogged in a wrong order". This bug report from Song Libing shows that Oracle engineers readily accept and verify bug reports where some code modification (adding debug sync point, for example) is needed to demonstrate the problem with a repeatable, deterministic test case. I've seen other bug reports where code modification was considered a problem and bug was not verified immediately as a result. Note that this bug seem to be not repeatable on MySQL 8.0.19 and this was explicitly tested, with test results shared in public. Good job, Umesh Shastry!
  • Bug #98734 - "Same digest for different queries if we use integer value in ORDER BY caluse". Different column numbers in ORDER BY often cause totally different execution plans, so assuming these queries are the same ("ORDER BY ?") for the purpose of digesting is misleading. Moreover, as noted by Lalit Choudhary from Percona, with column names instead of numbers such queries are not considered the same, so Performance Schema has to be more consistent.
  • Bug #98739 - "TempTable storage engine slow if mmap is required". Take care when large temporary tables are used for your queries in MySQL 8.0.x. By default (without temptable_use_mmap = OFF) when the table is larger than temptable_max_ram you may notice a very slow query execution. Nice finding by Marcelo Altmann. This is not the only related performance regression I've seen reported recently. Looks like TempTable storage engine problems is a real main topic of this post!
  • Bug #98750 - "SHOW FIELDS command regression on MySQL 8.0". This performance regression bug was reported by Juan Arruti. See also similar Bug #92195 - "checking permissions 90 time" that was wrongly(!) declared "Not a bug" back in 2018 just because of the way the problem was demonstrated. This time Przemyslaw Malkowski helped to make the point based on Performance Schema instrumentation, so there was no other option but to accept this as a real performance regression bug. Take care if you use SHOW statements with MySQL 8!
  • Bug #98782 - "Using TempTable engine for GROUP BY is slower than using MEMORY engine". In this bug report Øystein Grøvlen demonstrated that MEMORY engine is about 10% faster for temporary tables in MySQL 8. Make sure to use internal_tmp_mem_storage_engine=MEMORY if you care about performance.
  • Bug #98869 - "Temp ibt tablespace truncation at disconnection stuck InnoDB under large BP". Bug reporter, Fungo Wang, used different methods to show the performance impact of the regression probably introduced by WL#11613. From pt-pmp to perf and other OS level tools. Make sure to check all comments that point out to other bugs and problems.
  • Bug #98974 - "InnoDB temp table could hurt InnoDB perf badly". Yet another bug report by Fungo Wang. This time it took a lot of time and efforts from the bug reporter and many MySQL Community members (including me) to get this bug properly processed, even though it started with a detailed source code analysis, properly described test case, stack traces analysis and perf profiling details shared. The ideal bug report got far from ideal treatment, unfortunately.
  • Bug #98976 - "Case of referenced column of foreign key not corrected on import". Tim Düsterhus found that in versions with the fix for Bug #88718 - "Foreign key is always in lower case" the case is not always correct in the output of subsequent mysqldump.
  • Bug #98980 - "A state inside mysql_real_connect_nonblocking still blocks". This bug report by Jay Edgar was verified surprisingly fast. This is why it ended up in my list.
  • Bug #98990 - "avg_count_reset for monitor set owner is always NULL in I_S.INNODB_METRICS ". In this case Fungo Wang had not only found a bug, but also provided a patch that was accepted by Oracle after signing the OCA.
  • Bug #99006 - "GTID Synchronisation speed when doing START SLAVE". Simon Mudd noted that in some cases with GTIDs START SLAVE may take a lot of time:
    root@myhost [(none)]> start slave;
    Query OK, 0 rows affected (3 min 17.74 sec)
    and what's worse nothing is logged to the error log in the process to show the reason, progress or anything useful. Probably it's expected that with many binary logs finding proper GTID in them takes time, but some feedback would be useful. The bug does not have any clear public test case and is still under analysis.
  • Bug #99010 - "The mtr case --send command not use ps-protocol when run with --ps-protocol ". Does not sound like a big deal for anyone but developers who write MTR test cases, but in this report Ze Yang had provided source code analysis and quite detailed how to repeat steps and still got useless requests for additional feedback and nothing else till today. This is, again, unfortunate.
  • Bug #99039 - "PSI thread info not getting updated after statement execution". Who could imagine that regression bugs may be introduced even into Performance Schema? But Pranay Motupalli found one introduced into 8.0.18, 5.7.28+ and 5.6.47+ by this commit! I hope to see it fixed soon.
  • Bug #99051 - "XA commit may do engine commit before MYSQL_BIN_LOG::ordered_commit". XA transactions is one of my favorite weak areas of MySQL. In this bug report Dennis GAO described a case when XA COMMIT operation may do a engine commit before binlog flush on Ubuntu 18.04. He had contributed the fix based on the assumption that the best way is to ensure the sequence in the plugin_hash, so that the binlog plugin should be always before all the transnational engine plugins. I only hope that one day a long list of XA bugs will be fixed in one of forks if not in MySQL itself.
This is my favorite point of view this year.
To summarize:
  1. I see some really good examples of bugs verification by Oracle engineers recently. All supported versions are tested, nobody tries to argue against the bug reporter approach used to demonstrate the problem, even if it includes source code modifications. Looks really promising.
  2. Even ideal bug reports sometimes are not processed properly without extra efforts from MySQL Community, unfortunately.
  3. Percona engineers still contribute a lot to MySQL with both new bug reports and useful comments and clarifications. I hope Oracle appreciates that.
  4. Looks like code review and regression testing in Oracle still may benefit from some improvements, as we still see new regression bugs...

Saturday, August 10, 2019

Fun with Bugs #88 - On MySQL Bug Reports I am Subscribed to, Part XXII

It's Saturday night. I have a week of vacation ahead that I am going to spend at home, working on things I usually do not have time for. I already did something useful (created a couple of test cases for MariaDB bugs here and there, among other things), so why not to get some fun and continue my previous review of recent interesting MySQL bug report...

So, here is the list of 15 more MySQL community bug reports that I've subscribed to back in May and June 2019:
  • Bug #95491 - "The unused wake up in simulator AIO of function reserve_slot". I am impressed by the fact that these days many people just study InnoDB code, find problematic parts, improve them, test and prove the improvement, and submit patches to Oracle. Chen Zongzhi, who reported this bug, is one of such great people.
  • Bug #95496 - "Replication of memory tables". As pointed out by Iwo P, MySQL manual describes what happens to MEMORY tables replicated from master to slave in details, including the case of master restart. But one detail is missing - when slave with MEMORY tables restarts, even with super_read_only set, it also generates even to delete data from the table into its own binary log, with GTID, and this leads to problems in case of failover.
  • Bug #95507 - "innodb_stats_method is not honored when innodb_stats_persistent=ON". This bug was reported by my colleague Sergei Petrunia. See also related MDEV-19574  (still "Open").
  • Bug #95547 - "MySQL restarts automatically after ABORT_SERVER due to disk full". I am not sure this is a replication bug, more a problem of too primitive mysqld_safe or systemd startup scripts that do not check free space on disk or reason why mysqld process committed suicide. Anyway, this report by Przemyslaw Malkowski from Percona was accepted as a replication bug.
  • Bug #95582 - "DDL using bulk load is very slow under long flush_list". As found and proved by Fungo Wang, due to optimized away redo log writing for bulk load, to make sure such DDL is crash safe, dirtied pages must be flushed to disk synchronously. Currently this is done inefficiently, as shown in this bug report. See also a year old (and still "Open") Bug #92099 - "provide an option to record redo log for bulk loading" from the same reporter, where he asks for a way to get redo logging back (but still benefit from other "online" ALTER features). MariaDB has that option after the fix of MDEV-16809.
  • Bug #95612 - "replication breakage due to Can not lock user management caches for processing". Simon Mudd noted that in some cases simple DROP USER IF EXISTS ... breaks parallel replication in MySQL 8.0.16. Good to know that START SLAVE allows to continue :)
  • Bug #95698 - "JSON columns are returned with binary charset". This bug was reported by Daniel Book. Note that according to the original issue, DBD::MariaDB (I had no idea this exists!) works properly with MariaDB 10.3 or 10.4. I had not checked.
  • Bug #95734 - "Failing SET statement in a stored procedure changes variables". This really weird and unexpected bug was reported by Przemysław Skibiński.
  • Bug #95863 - "MTS ordered commit may cause deadlock ". Great finding by Lalit Choudhary from Percona!
  • Bug #95895 - "Shutdown takes long time when table uses AUTOINC and FOREIGN KEY constraints". It was reported by Satya Bodapati from Percona, who later contributed as patch. See also his related Bug #95898 - "InnoDB releases memory of table only on shutdown if table has FK constraint". I truly hope to see these fixed in next minor releases of MySQL. Note that originally bug (PS-5639) was reported by Sveta Smirnova.
  • Bug #95928 - "InnoDB crashes on compressed and encrypted table when changing FS block size". Make sure to check and write original filesystem block size somewhere... This bug was reported by Sergei Glushchenko from Percona.
  • Bug #95929 - "Duplicate entry for key 'PRIMARY' when querying information_schema.TABLES". This nice feature of new data dictionary of MySQL 8.0.16 under concurrent load was found by Manuel Rigger.
  • Bug #95934 - "innodb_flush_log_at_trx_commit = 0 get performance regression on 8 core machine". If I had access to 8 cores box, I'd try to check this sysbench-based complete test case immediately. I should do this on my quadcore box at least, next week. But here we see some 12 days just wasted and then bug report from Chen Zongzh ended up "Analyzing", till today, with no further feedback.
  • Bug #95954 - "CAST of negative function return value to UNSIGNED malfunctions with BIGINT". Yet another weird bug found by Manuel Rigger. There are patches (separate for 5.6, 5.7 and 8.0) contributed by Oleksandr Peresypkin.
Montjuïc is a place to enjoy nice view of Barcelona and forget about MySQL bugs. So, I could miss some interesting ones at the end of May, 2019...
To summarize:
  1. Percona engineers still contribute a lot of useful and interesting MySQL bug reports, often - with patches.
  2. I am impressed by quality of MySQL bug reports in 2019. Often we see detailed analysis, proper complete test cases and patches.
  3. There is a recent fork of DBD::mysql for MariaDB (and MySQL), DBD::MariaDB!
  4. I badly need some new hardware for running benchmarks and tests from recent bug reports...
  5. I should spend more time in May in Barcelona :)

Saturday, July 27, 2019

Fun with Bugs #86 - On Some Public Bugs Fixed in MySQL 5.7.27

This week new minor releases of MySQL server were announced. Of them all I am mostly interested in MySQL 5.7.27. I plan to concentrate on InnoDB, replication and optimizer bugs reported in public MySQL bugs database and fixed in MySQL 5.7.27. As usual I name bug reporters explicitly and give links to their remaining currently active bug reports, if any.

Let me start with InnoDB bug fixes:
  • Bug #94699 - "Mysql deadlock and bugcheck on aarch64 under stress test". This bug report with a fix for insufficient memory barriers in the rw-lock implementation was contributed by Cai Yibo.
  • Bug #93708 - "Page Cleaner will sleep for long time if clock changes". This bug caused long delays on shutdown. It was reported by Marcelo Altmann. It took some efforts from MySQL Community to have it accepted as a real bug, but now it's fixed!
  • Bug #67526 - "Duplicate key error on auto-inc PK with mixed auto_increment_increment clients". This regression bug was reported back in 2012 by Rob Lineweaver and affected all MySQL versions starting from 5.5.28. Release notes states that the fix was to revert the patch that fixed Bug #65225 - "InnoDB miscalculates auto-increment after changing auto_increment_increment" reported by  Elena Stepanova. I am not sure what had really happened without checking the source code or at least running test cases from both bugs on 5.7.27.
  • Bug #94383 - "simple ALTER cause unnecessary InnoDB index rebuilds, 5.7.23 or later 5.7 rlses". This bug was reported by Mikhail Izioumtchenko, who had contributed diagnostics patch.
  • Bug #93670 - "virtual generated column index data inconsistency". I list this bug that happens when foreign keys are involved as InnoDB one, as no other engine in MySQL 5.7.27 supports foreign keys anyway. The bug was reported by Rui Xu.
Now let me continue with replication bugs:
  • Bug #93771 - "Contribution: Add log_bin_implicit_delete setting". This report was created for the patch contributed by Daniël van Eeden (who had also got a public credit for this contribution in a blog post) and is now closed and documented as fixed. But only part of the original patch was used, adding a comment to the binary log. It's strange to see this happening without any comments on why the rest of the patch was not used.
  • Bug #93440 - "Noop UPDATE query is logged to binlog after read_only flag is set". The problem is actually in a new GTID being generated in this case. The bug was reported by Artem Danilov.
  • Bug #92398 - "point in time recovery using mysqlbinlog broken with temporary table -> errors". This bug was reported by Shane Bester himself. I do not see any complete test case in the public bug report, for some reason. But if you search for private bug number (28642318) at GitHub you can easily identify related commit and changes in the mysql-test/extra/binlog_tests/drop_temp_table.test.
  • Bug #89987 - "super_read_only with master_info_repository=TABLE breaks replication on restart". It was reported by Clemens Weiß and fixed in 5.7.27, but it took some efforts and public comment by Jean-François Gagné to get this clarified.
  • Bug #93395 - "ALTER USER succeeds on master but fails on slave". This bug was reported by Jean-François Gagné and is still marked as "Verified" and is NOT explicitly listed in the release notes. But read this:
    "CREATE USER and ALTER USER did not check the validity of a hashed authentication string when used with IDENTIFIED WITH auth_plugin AS 'hash_string' syntax. (Bug #29395944)"
    The description of the bug sounds very similar, and according to the release notes it's fixed in 5.7.27. How comes the bug above is NOT closed and listed as at least related? It seems somebody does NOT look for duplicates, either when copying public bugs into the internal bugs database, or when closing the bug as fixed. What we have as a result looks like a lack of proper documenting. As bug reporter noted in the comment:
    "This is the 2nd attribution omission I am finding in 5.7.27 release notes (the other one was on Bug#95484). Is this a new policy not to update public bugs with fixed version and not to mention the public bugs in the release notes ?"
    This situation is wrong. It looks awkward like this "house" in Brighton:

and attracts attention. I hope Oracle engineers will care to add public comments to these two bugs to clarify what really happened and why.

Finally a couple of optimizer bugs were also fixed:
  • Bug #92809 - "Inconsistent ResultSet for different Execution Plans". The bug was reported by Juan Arruti. It took a lot of efforts from several Percona engineers to force it to be verified. Complete analysis (by Yura Sorokin) identified the root cause and the fact that the problem is already fixed in MySQL 8.0.x while MySQL 5.6.x is also affected (and seems NOT to be fixed). Good job by MySQL Community, somewhat questionable assistance from Oracle's engineer involved.
  • Bug #90398 - "Duplicate entry for key '<group_key>' error". It was reported by Yoseph Phillips. The fix is actually a more useful error message, nothing more.
That's all I have to say about the release. To summarize:
  1. MySQL 5.7.27 includes fixes to several serious InnoDB and replication bugs, so consider upgrade.
  2. Percona engineers contribute a lot to MySQL, both in form of bug reports, patches and by helping other community users to make their point and get their bugs fixed fast.
  3. There are things to improve in a way Oracle engineers handle bugs processing (verification, checking for duplicates, proper documenting of public bug reports that are fixed).
Some items above are the same as in my summary for the previous 5.7.26 release, isn't it?

Sunday, May 26, 2019

MySQL Support Engineer's Chronicles, Issue #10

As promised, I am trying to write one blog post in this series per week. So, even though writing about InnoDB row formats took a lot of time and efforts this weekend, I still plan to summarize my findings, questions, discussions, bugs and links I've collected over this week.

I've shared two links this week on Facebook that got a lot of comments (unlike links to my typical blog posts). The first one was to Marko Mäkelä's blog post at MariaDB.com, "InnoDB Quality Improvements in MariaDB Server". I do not see any comments (or any obvious way to comment) there, but the comments I've got at Facebook were mostly related to the statement that  
"We are no longer merging new MySQL features with MariaDB..."
noted in the text by Mark Callaghan and to the idea that "InnoDB" is a trademark of Oracle, so using it to refer to a fork (that is incompatible with the "upstream" InnoDB in too many ways since MariaDB 10.1 probably) is wrong, as stated by Matt Lord and Sunny Bains. People in the comments mostly agree that a new name makes sense (there are more reasons to give it now anyway than in the case of XtraDB by Percona), and we had a lot of nice and funny suggestions on Slack internally (FudDB was not among them, this is a registered trademark of Miguel Angel Nieto for many years already). We shell see how this may end up, but I would not be surprised by a new name announced soon. I suggest you to read comments in any case if you have a Facebook account, many of them are interesting.

The second link was to Kaj Arnö's post at mariadb.org, "On Contributions, Pride and Cockiness". It's worth checking just because of Monty's photo there. Laurynas Biveinis stated in the comments that any comparison of number of pull requests (open and processed) is meaningless when development model used by other parties is different (closed, with contributions coming mostly via bug reports in case of Oracle, or all changes, external and internal, coming via pull requests in case of Percona). MariaDB uses a mix of a kind, where some contributions from contractors come via pull requests, while engineers from MariaDB Corporation work on GitHub sources of MariaDB Server directly. Anyway (meaningless statistics aside), MariaDB seems to be the easiest target for contributions from Community at the moment, and nobody argued against that. My followers also agreed that the same workflow for internal and external contributions is a preferred development model in ideal world.

This kind of public discussions of (serious and funny) MySQL-related matters on Facebook (along with public discussions on MySQL bugs) make me think the way I use my Facebook page is proper and good for the mankind.

Now back to notes made while working on Support issues. This week I had to explain one case when MariaDB server was shut down normally (but unexpectedly for DBA):
2019-05-22 10:37:55 0 [Note] /usr/libexec/mysqld (initiated by: unknown): Normal shutdown
This Percona blog post summarizes different ways to find a process which sent a HUP/KILL/TERM or other signal to the mysqld process. I've used SystemTap-based solution like suggested in that blog post in the past successfully. In this context I find this summary of the ways to force MySQL to fail useful. for all kinds of testing. SELinux manual is also useful to re-read at times.

This week I've spent a lot of time and some efforts trying to reproduce the error (1942 and/or 1940 if anyone cares) on Galera node acting as an async replication slave. These efforts ended up with a bug report, MDEV-19572. Surely the idea to replicate MyISAM tables outside of mysql database to Galera cluster is bad at multiple levels, but why the error after running for a long time normally? In the process of testing I was reading various remotely related posts, so checked this and that... I also hit other problems in the process. Like this crash that happened probably while sending some signal to the node unintentionally:
190523 17:19:46 [ERROR] mysqld got signal 11 ;
This could be because you hit a bug. It is also possible that this binary
or one of the libraries it was linked against is corrupt, improperly built,
or misconfigured. This error can also be caused by malfunctioning hardware.

To report this bug, see https://mariadb.com/kb/en/reporting-bugs

We will try our best to scrape up some info that will hopefully help
diagnose the problem, but since we have already crashed,
something is definitely wrong and this may fail.

Server version: 10.2.23-MariaDB-log
key_buffer_size=134217728
read_buffer_size=131072
max_used_connections=3
max_threads=153
thread_count=65544
It is possible that mysqld could use up to
key_buffer_size + (read_buffer_size + sort_buffer_size)*max_threads = 467240 K  bytes of memory
Hope that's ok; if not, decrease some variables in the equation.

Thread pointer: 0x0
Attempting backtrace. You can use the following information to find out
where mysqld died. If you see no messages after this, something went
terribly wrong...
stack_bottom = 0x0 thread_stack 0x49000
/home/openxs/dbs/maria10.2/bin/mysqld(my_print_stacktrace+0x29)[0x7f6475eb5b49]
/home/openxs/dbs/maria10.2/bin/mysqld(handle_fatal_signal+0x33d)[0x7f64759d50fd]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x10330)[0x7f6473887330]
/home/openxs/dbs/maria10.2/bin/mysqld(+0xb3b817)[0x7f6475ebc817]
/home/openxs/dbs/maria10.2/bin/mysqld(+0xb3b9e6)[0x7f6475ebc9e6]
/home/openxs/dbs/maria10.2/bin/mysqld(+0xb3bb8a)[0x7f6475ebcb8a]
/home/openxs/dbs/maria10.2/bin/mysqld(lf_hash_delete+0x61)[0x7f6475ebcfa1]
/home/openxs/dbs/maria10.2/bin/mysqld(+0x601eed)[0x7f6475982eed]
include/my_atomic.h:298(my_atomic_storeptr)[0x7f6475983464]
sql/table_cache.cc:534(tdc_delete_share_from_hash)[0x7f6475811f17]
sql/table_cache.cc:708(tdc_purge(bool))[0x7f64759351ea]
sql/sql_base.cc:376(close_cached_tables(THD*, TABLE_LIST*, bool, unsigned long))[0x7f64757c9ec7]
nptl/pthread_create.c:312(start_thread)[0x7f647387f184]
/lib/x86_64-linux-gnu/libc.so.6(clone+0x6d)[0x7f6472d8c03d]
The manual page at http://dev.mysql.com/doc/mysql/en/crashing.html contains
information that should help you find out what is causing the crash.
I was not able so far top find the exact some backtrace in any known MariaDB bug, so one day I'll have to try to reproduce this crash as well.

I try to check some MariaDB ColumnStore issues from time ot time, for a change, and this week I ended up reading this KB page while trying to understand how much we can control placement of data there.

Finally, for the records, this is the way to "fix" InnoDB statistics if needed (and the need is real as you can find out from Bug #95507 - "innodb_stats_method is not honored when innodb_stats_persistent=ON" reported by my colleague Sergei Petrunia):
update mysql.innodb_index_stats set last_update=now(), stat_value=445000000 where database_name='test' and table_name='t1' and index_name='i1' and stat_name='n_diff_pfx01';
I like to return to familiar nice places and topics, like Regent's Canal or MySQL bugs...
The last bug not the least, MySQL bugs. This week I've subscribed to the following (already "Verified") interesting bug reports (besides the one mentioned above):
  • Bug #95484 - "EXCHANGE PARTITION works wrong/weird with different ROW_FORMAT.". Jean-François Gagné found out that there is a way to have partitions with different row_format values in the same InnoDB table, at least in MySQL 5.7. So why is this not supported officially? See also his Bug #95478 - "CREATE TABLE LIKE does not honour ROW_FORMAT.". It's a week of ROW_FORMAT studies for me, for sure!
  • Bug #95462 - "Data comparison broke in MySQL 8.0.16". It's common knowledge how much I like regression bugs. MySQL 8.0.16 introduced a new one, reported by
    Raman Haran, probably based on some good and valid intentions. But undocumented changes in behavior in GA versions are hardly acceptable, no matter what are the intentions.
That's all for now. Some more links to MySQL bugs from me are always available on Twitter.

Saturday, January 26, 2019

Fun with Bugs #78 - On Some Public Bugs Fixed in MySQL 5.7.25

Today I'd like to continue my tradition of ignoring MySQL 8 (after all, I can not even build 8.0.14 any more on my Ubuntu 14.04, it's not supported suddenly because of old gcc version) and, of all MySQL server versions released by Oracle this week, concentrate on bugs reported in public bugs database and fixed in the latest minor release of MySQL 5.7 branch, 5.7.25.

This time there is only one InnoDB community-reported bug fixed, Buig #87423 - "os0file.cc assertion failed 'offset > 0' in os_file_io_complete", from Vasily Nemkov. See also it's duplicate, Bug #88956, by Aidan Diffey. Assertion failure seems to happen with 32-bit binaries running on 64-bit OS (both bugs are reported on Linux) and is related to infamous ulint type usage in the code, but for some reason release notes and closing comment mention Windows:
"An assertion was raised when attempting to write to a tablespace file greater than 4GB in size on a 64-bit Windows system. The failure was due to a narrowing cast."
Usually there are some bugs around flowers. But sometimes we can enjoy flowers without bugs...
There is a long enough list of replication bugs fixed:
  • Bug #92132 - "secure-file-priv breaks LOAD DATA INFILE replication in statement mode on 5.7.23". This regression bug (that had not got "regression" tag!) was reported by Nicolai Plum.
  • Bug #91941 - "Deadlock during purge_logs_before_date". Great bug report from Nikolai Ikhalainen. Note also that some hints by Jean-François Gagné and proper gdb backtrace analysis by Oracle engineers, Shane Bester and Dmitry Lenev, were needed to force proper processing of this bug. See also related Bug #92108 - "Deadlock by concurrent show binlogs, pfs session_variables table & binlog purge", from Shashank Sahni.
  • Bug #91548 - "LOCK_grant and LOCK_open can deadlock on a gtid slave". Great example of Shane Bester's regular work on bugs.
  • Bug #90640 - "`head->variables.gtid_next.type != UNDEFINED_GTID' ". One of those cases when assertion failure in debug build highlights real problem in the code. Great finding by Roel Van de Paar.
  • Bug #87832 - "Relay_Log_Space is inaccurate and leaks". Nice bug report by Manuel Ung. Related variable could be changed concurrently without any locking and get randomly wrong values.
  • Bug #84752 - "Multi-Slave Replication Fail: bogus data in log event". This bug was reported by Gonzalo Miguel Arruti . Eventually it seems internally reported Bug#22252394 - SLAVE I/O THREAD MAY STOP WHEN BINLOG ROTATES highlighted the real problem. Good to have this case fixed!
  • Bug #83003 - "Using temporary tables on slaves increases GTID sequence number". I am so happy to this bug reported by my former colleague Ovais Tariq fixed in 5.7! Patches provided by Laurynas Biveinis from Percona are finally backported. It's so sad to see slaves out of sync with master when GTIDs are used... For some reason this bug is not listed as replication-related and patches contributed are not mentioned in the release notes.
Some other fixes are also interesting to check:
  • Bug #92131  - "ASan: Direct leak of 272 byte(s) in main.mysqlpump_partial_bkp MTR test case". I mentioned this bug reported by Yura Sorokin from Percona in the past. Good to see it fixed. I truly hope Oracle is testing ASan builds regularly with MTR and we'll not see such reports from community any more. For now efforts of Percona engineers help, as we can see also from Bug #90238 - "Comparison of uninitailized memory in log_in_use" with patches from Zsolt Parragi and Laurynas Biveinis.
  • Bug #92049 - "bogus data when ordering results from variables_by_thread". Thanks to Shane Bester from Oracle (who still reports MySQL bugs in public), we have one bug less in otherwise near perfect Performance Schema.
  • Bug #90742 = "SIGHUP cause mysql server crash.". It was reported by Seunguck Lee. Good to see S1 crashing bug not blindly classified as "security" and hidden forever. Note also great analysis there provided by Jean-François Gagné! It would be impossible for any hidden bug to get this kind of useful feedback.
  • Bug #89214 - "The SELECT will deadlock in the stored procedure, if the result set is empty.". It took some time and efforts for bug reporter (two pg) to prove the point, confirm that the problem happens only with prepared statements and force proper processing. This regression bug (MySQL 5.6 was not affected) still does not have "regression" tag.
That's all bugs I wanted to mention today. To summarize:
  1. I'd surely consider immediate upgrade to 5.7.25 in any environment where replication is used.
  2. It seems Oracle engineers who process bugs still have no habit of adding "regression" tag when obviously needed. I'll ask bug reporters to do this themselves during my FOSDEM talk about MySQL bugs.
  3. Percona still helps Oracle to make MySQL better with proper QA efforts and patches.
  4. Most of the bugs mentioned here affected MySQL 8 as well, so from this post you know about many important fixes that happened in MySQL 8.0.14 :) 

Tuesday, January 15, 2019

Using dbdeployer With MariaDB Server

Some time ago I've noted that one of the tools I use for testing various MySQL and MariaDB cases and to reproduce potential bugs, MySQL-Sandbox, is not updated any more. It turned out that active development switched to its port in Go called dbdeployer. You can find detailed information about dbdeployer and reasons behind developing it provided by its author, Giuseppe Maxia, here and there. See also this post at Percona blog for some quick review of its main features. One of the points of dbdeployer (and reasons to use Go) is that it is built once (per platform supported) somewhere and then binaries are downloaded from GitHub and used everywhere, without any problems with dependencies etc.

I've added checking dbdeployer to my long ToDo list, as I planned to use it (if not MySQL Sandbox) for some tests and posts related to resolving typical practical problems with MariaDB GTID-based replication. Yesterday I've allocated some time to finally try it and, as usual, I've started with building it from source (as for me databases-related software that I can not build from source on my test systems is not any attractive as something new to study and use). I was immediately surprised by the lack of instructions on how to do this at GitHub, no Makefile of any kind etc. All I was able to find is build.sh script. Correction: just check README.md on how to build it properly, as Giuseppe Maxia explained in the comment.

Good, regular structure is important for deployment
Fortunately this is not the first project written in Go that I try to build (or change somehow and then build). The first one was this replication manager (that has proper build instructions in docs). So, I though I knew what to do. I've installed missing golang package on my netbook with Ubuntu 14.04 that I had at hand and tried the following typical steps:
openxs@ao756:~/go$ export GOPATH=$HOME/go
openxs@ao756:~/go$ echo $GOPATH
/home/openxs/go
openxs@ao756:~/go$ go get github.com/datacharmer/dbdeployer
# github.com/datacharmer/dbdeployer/common
src/github.com/datacharmer/dbdeployer/common/strutils.go:170: undefined: sort.Slice
...
That was a bit surprising, but quick Google search shown that this could be caused by outdated (pre-1.8) version of golang package. So, dbdeployer requires golang 1.8 or newer and there was no such package for my good old Ubuntu (it has some 1.2.x only). One day I'll upgrade it, but so far I am OK with 14.04 for all other testing purposes, so I had to give up on the idea to build from source temporary. 

Today during few free minutes I've retried on my good old desktop box with Fedora 27 (where I surely built some Go project(s) successfully):
[openxs@fc23 go]$ uname -a
Linux fc23 4.18.19-100.fc27.x86_64 #1 SMP Wed Nov 14 22:04:34 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
[openxs@fc23 ~]$ ls go
pkg  src
[openxs@fc23 ~]$ echo $GOPATH

[openxs@fc23 ~]$ export GOPATH=$HOME/go
[openxs@fc23 ~]$ cd go
[openxs@fc23 go]$ go versiongo version go1.9.7 linux/amd64
This environment should work for build, so I've proceeded with:
[openxs@fc23 go]$ go get github.com/datacharmer/dbdeployer
[openxs@fc23 go]$ ls src/github.com/datacharmer/   jmoiron/       nsf/           tanji/
go-sql-driver/ mattn/         ogier/
[openxs@fc23 go]$ ls src/github.com/datacharmer/dbdeployer/
abbreviations/ compare/       docs/          mkreadme/      test/
.build/        concurrent/    .git/          rest/          unpack/
cmd/           cookbook/      .github/       sandbox/       vendor/
common/        defaults/      globals/       scripts/
Now let's try that scripts/build.sh with linux as a parameter, as it's a way to build Linux binaries based on what I found:
[openxs@fc23 go]$ MKDOCS=1 src/github.com/datacharmer/dbdeployer/scripts/build.sh linux
+ env GOOS=linux GOARCH=386 go build --tags docs -o dbdeployer-1.17.0-docs.linux .
+ env GOOS=linux GOARCH=amd64 go build -o sort_versions.linux sort_versions.go
/home/openxs/go/src/github.com/datacharmer/dbdeployer
-rwxrwxr-x. 1 openxs openxs 8.1M Jan 14 10:27 dbdeployer-1.17.0-docs.linux
-rw-rw-r--. 1 openxs openxs 3.0M Jan 14 10:27 dbdeployer-1.17.0-docs.linux.tar.gz
[openxs@fc23 go]$ ls
bin  pkg  src
[openxs@fc23 go]$ ls bin
dbdeployer
[openxs@fc23 go]$ bin/dbdeployer --version
dbdeployer version 1.17.0
Now we know how to build dbdeployer from source, if needed. If some dependencies are missing you'll be informed and similar go get ... command should allow to install it.

I was somewhat surprised to see MariaDB NOT mentioned at all in README.md. It says:
"DBdeployer is a tool that deploys MySQL database servers easily."
while good old MySQL-Sandbox also mentions MariaDB explicitly:
"This package is a sandbox for testing features under any version of MySQL from 3.23 to 8.0 (and any version of MariaDB.)"
So, my idea was to double check that dbdeployer is both MySQL-Sandbox compatible and MariaDB compatible (it is). I have several sandboxes already created in the past. I also have MariaDB 10.2.21 .tar.gz binaries that I want to use with dbdeployer for further testing:
[openxs@fc23 go]$ ls ~/sandboxes/
clear_all        rsandbox_mariadb-10_0_19  send_kill_all  test_replication
plugin.conf      rsandbox_mariadb-10_1_12  start_all      use_all
restart_all      rsandbox_mysql-8_0_12     status_all
rsandbox_8_0_12  sandbox_action            stop_all
[openxs@fc23 go]$ ls ~/*.tar.gz
/home/openxs/galera-25.3.22-x86_64.tar.gz
/home/openxs/galera-25.3.24-x86_64.tar.gz
/home/openxs/galera-25.3.25-glibc_214-x86_64.tar.gz
/home/openxs/mariadb-10.2.12-linux-x86_64.tar.gz
/home/openxs/mariadb-10.2.21-linux-x86_64.tar.gz
With dbdeployer one has to unpack .tar.gz first with dbdeployer unpack command. So, I tried it immediately:
[openxs@fc23 go]$ bin/dbdeployer unpack ~/mariadb-10.2.21-linux-x86_64.tar.gz
directory '/home/openxs/opt/mysql' not found
You should create it or provide an alternate base directory using --sandbox-binary
It seems the tool now wants to use ~/opt/mysql as a directory to unpack to, while MySQL_Sandbox silently used ~:
[openxs@fc23 go]$ ls ~ | grep 8.0
8.0.12
I made a lame try to force it to use ~, but failed for the reason I was too lazy to study:
[openxs@fc23 go]$ bin/dbdeployer --sandbox-binary=/home/openxs unpack /home/openxs/mariadb-10.2.21-linux-x86_64.tar.gzUnpacking tarball /home/openxs/mariadb-10.2.21-linux-x86_64.tar.gz to $HOME/10.2.21
.........100.........200....&tar.Header{Name:"mariadb-10.2.21-linux-x86_64/mysql-test/mysql-test-run", Mode:511, Uid:1021, Gid:1004, Size:0, ModTime:time.Time{wall:0x0, ext:63681810892, loc:(*time.Location)(0xa47aa0)}, Typeflag:0x32, Linkname:"./mysql-test-run.pl", Uname:"dbart", Gname:"my", Devmajor:0, Devminor:0, AccessTime:time.Time{wall:0x0, ext:0, loc:(*time.Location)(nil)}, ChangeTime:time.Time{wall:0x0, ext:0, loc:(*time.Location)(nil)}, Xattrs:map[string]string(nil)}
#ERROR: symlink ./mysql-test-run.pl mariadb-10.2.21-linux-x86_64/mysql-test/mysql-test-run: file exists
I just created ~/opt/mysql and proceeded with default configuration. After unpack step completed successfully I've proceeded with deploy step to create new replication sandbox:
[openxs@fc23 go]$ bin/dbdeployer unpack /home/openxs/mariadb-10.2.21-linux-x86_64.tar.gz
Unpacking tarball /home/openxs/mariadb-10.2.21-linux-x86_64.tar.gz to $HOME/opt/mysql/10.2.21
.........100.........200.........300.........400.........500.........600.........700.........800... ...
Renaming directory /home/openxs/opt/mysql/mariadb-10.2.21-linux-x86_64 to /home/openxs/opt/mysql/10.2.21

[openxs@fc23 go]$ bin/dbdeployer deploy replication 10.2.21
Installing and starting master
. sandbox server started
Installing and starting slave1
. sandbox server started
Installing and starting slave2
. sandbox server started
$HOME/sandboxes/rsandbox_10_2_21/initialize_slaves
initializing slave 1
initializing slave 2
Replication directory installed in $HOME/sandboxes/rsandbox_10_2_21
run 'dbdeployer usage multiple' for basic instructions'
We have access to nice enough documentation:
[openxs@fc23 go]$ bin/dbdeployer usage multiple
 USING MULTIPLE SERVER SANDBOX
On a replication sandbox, you have the same commands (run "dbdeployer usage single"),
with an "_all" suffix, meaning that you propagate the command to all the members.
Then you have "./m" as a shortcut to use the master, "./s1" and "./s2" to access
the slaves (and "s3", "s4" ... if you define more).

In group sandboxes without a master slave relationship (group replication and
multiple sandboxes) the nodes can be accessed by ./n1, ./n2, ./n3, and so on.

start_all    [options] > starts all nodes
status_all             > get the status of all nodes
restart_all  [options] > restarts all nodes
stop_all               > stops all nodes
use_all         "SQL"  > runs a SQL statement in all nodes
use_all_masters "SQL"  > runs a SQL statement in all masters
use_all_slaves "SQL"   > runs a SQL statement in all slaves
clear_all              > stops all nodes and removes all data
m                      > invokes MySQL client in the master
s1, s2, n1, n2         > invokes MySQL client in slave 1, 2, node 1, 2


The scripts "check_slaves" or "check_nodes" give the status of replication in the sandbox.
Typical sandbox directory (with some differences like use_all_slaves etc) is created in ~/sandboxes/ and shortcut commands work as expected:
[openxs@fc23 go]$ cd ~/sandboxes/rsandbox_10_2_21/
[openxs@fc23 rsandbox_10_2_21]$ ls
check_slaves       n2           sbdescription.json  test_sb_all
clear_all          node1        send_kill_all       use_all
initialize_slaves  node2        start_all           use_all_masters
m                  restart_all  status_all          use_all_slaves
master             s1           stop_all
n1                 s2           test_replication

[openxs@fc23 rsandbox_10_2_21]$ ls master/
add_option    init_db         restart             show_binlog    status   use
clear         load_grants     sbdescription.json  show_log       stop
data          my              sb_include          show_relaylog  test_sb
grants.mysql  my.sandbox.cnf  send_kill           start          tmp

[openxs@fc23 rsandbox_10_2_21]$ ls ../rsandbox_mariadb-10_1_12/
check_slaves             m       node1        s2             test_replication
clear_all                master  node2        send_kill_all  use_all
connection.json          n1      README       start_all
default_connection.json  n2      restart_all  status_all
initialize_slaves        n3      s1           stop_all

[openxs@fc23 rsandbox_10_2_21]$ ls ../rsandbox_mariadb-10_1_12/master/
add_option       default_connection.json  my              send_kill      tmp
change_paths     grants_5_7_6.mysql       mycli           show_binlog    use
change_ports     grants.mysql             my.sandbox.cnf  show_relaylog  USING
clear            json_in_db               proxy_start     start
connection.json  load_grants              README          status
data             msb                      restart         stop

[openxs@fc23 rsandbox_10_2_21]$ ./status_all
REPLICATION  /home/openxs/sandboxes/rsandbox_10_2_21
master : master on  -  port     23322 (23322)
node1 : node1 on  -  port       23323 (23323)
node2 : node2 on  -  port       23324 (23324)
[openxs@fc23 rsandbox_10_2_21]$ ./use_all "show variables like 'gtid%'"
# master
Variable_name   Value
gtid_binlog_pos 0-100-12
gtid_binlog_state       0-100-12
gtid_current_pos        0-100-12
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos
gtid_strict_mode        OFF
# server: 1
Variable_name   Value
gtid_binlog_pos
gtid_binlog_state
gtid_current_pos        0-100-12
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos  0-100-12
gtid_strict_mode        OFF
# server: 2
Variable_name   Value
gtid_binlog_pos
gtid_binlog_state
gtid_current_pos        0-100-12
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos  0-100-12
gtid_strict_mode        OFF
[openxs@fc23 rsandbox_10_2_21]$ ./m
Welcome to the MariaDB monitor.  Commands end with ; or \g.
Your MariaDB connection id is 14
Server version: 10.2.21-MariaDB-log MariaDB Server

Copyright (c) 2000, 2018, Oracle, MariaDB Corporation Ab and others.

Type 'help;' or '\h' for help. Type '\c' to clear the current input statement.

master [localhost:23322] {msandbox} ((none)) > show master status;
+------------------+----------+--------------+------------------+
| File             | Position | Binlog_Do_DB | Binlog_Ignore_DB |
+------------------+----------+--------------+------------------+
| mysql-bin.000001 |     2835 |              |                  |
+------------------+----------+--------------+------------------+
1 row in set (0.00 sec)

master [localhost:23322] {msandbox} ((none)) > show variables like 'gtid%';
+------------------------+----------+
| Variable_name          | Value    |
+------------------------+----------+
| gtid_binlog_pos        | 0-100-12 |
| gtid_binlog_state      | 0-100-12 |
| gtid_current_pos       | 0-100-12 |
| gtid_domain_id         | 0        |
| gtid_ignore_duplicates | OFF      |
| gtid_seq_no            | 0        |
| gtid_slave_pos         |          |
| gtid_strict_mode       | OFF      |
+------------------------+----------+
8 rows in set (0.00 sec)

master [localhost:23322] {msandbox} ((none)) > exit
Bye
For my further tests I needed slaves to have log_slave_updates enabled and gtid_strict_mode=ON. So, I've added these settings to my.sandbox.cnf in node1 and node2 subdirectories for both slaves and restarted them:
[openxs@fc23 rsandbox_10_2_21]$ ./restart_all# executing 'stop' on /home/openxs/sandboxes/rsandbox_10_2_21
stop /home/openxs/sandboxes/rsandbox_10_2_21/node1
stop /home/openxs/sandboxes/rsandbox_10_2_21/node2
stop /home/openxs/sandboxes/rsandbox_10_2_21/master
# executing 'start' on /home/openxs/sandboxes/rsandbox_10_2_21
executing 'start' on master
. sandbox server started
executing 'start' on slave 1
. sandbox server started
executing 'start' on slave 2
. sandbox server started
I need a table to play with and I want to check that slaves are in sync:
[openxs@fc23 rsandbox_10_2_21]$ ./m
Welcome to the MariaDB monitor.  Commands end with ; or \g.
Your MariaDB connection id is 12
Server version: 10.2.21-MariaDB-log MariaDB Server

Copyright (c) 2000, 2018, Oracle, MariaDB Corporation Ab and others.

Type 'help;' or '\h' for help. Type '\c' to clear the current input statement.

master [localhost:23322] {msandbox} ((none)) > use test
Database changed
master [localhost:23322] {msandbox} (test) > create table t1(id int primary key, c1 int);
Query OK, 0 rows affected (0.17 sec)

master [localhost:23322] {msandbox} (test) > exit
Bye
[openxs@fc23 rsandbox_10_2_21]$ ./use_all "show variables like 'gtid%'"
# master
Variable_name   Value
gtid_binlog_pos 0-100-13
gtid_binlog_state       0-100-13
gtid_current_pos        0-100-13
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos
gtid_strict_mode        OFF
# server: 1
Variable_name   Value
gtid_binlog_pos 0-100-13
gtid_binlog_state       0-100-13
gtid_current_pos        0-100-13
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos  0-100-13
gtid_strict_mode        ON
# server: 2
Variable_name   Value
gtid_binlog_pos 0-100-13
gtid_binlog_state       0-100-13
gtid_current_pos        0-100-13
gtid_domain_id  0
gtid_ignore_duplicates  OFF
gtid_seq_no     0
gtid_slave_pos  0-100-13
gtid_strict_mode        ON
[openxs@fc23 rsandbox_10_2_21]$
Note the value of gtid_current_pos on master and gtid_slave_pos on each slave. They are the same and slaves are in sync. If you want to find out more about the format of GTIDs in MariaDB or all that gtid% server variables, please, check this KB article.

* * *

To summarize, dbdeployer is a nice port of MySQL-Sandbox into Go, with some additional features. It can be easily built from source if you have golang version 1.8 or newer (or just downloaded if you have not). Sandboxes created with dbdeployer may co-exists with older sandboxes in the same default directory (but .tar.gz files are unpacked into different directory by default). It still works well with MariaDB. I am going to use replication sandboxes built with it for some further testing of various real life use cases and problems of MariaDB's GTIDs implementation (that may be presented in further posts).

Saturday, November 24, 2018

Fun with Bugs #73 - On MySQL Bug Reports I am Subscribed to, Part X

It's time to continue my review of MySQL bug reports that I considered interesting for some reason recently. I had not got any notable reaction from Oracle engineers to my previous post about recent regression bugs in MySQL 8.0.13, so probably this topic is not really that hot. In this boring post I'll just review some bugs I've subscribed to since August that are still not closed, starting from the oldest.

Let me start with a couple of bug reports that remain "Open":
  • Bug #91959 - "UBSAN: signed integer overflow in lock_update_trx_age". It's really unusual to see bug reported by Shane Bester himself just "Open" for months. As he noted, it's really strange to see age defined as 32-bit value here when age_updated two lines later is 64-bit. On the other hand, from comments in the code it seems the age value is supposed to be small enough and if it grows too much this is a problem.
  • Bug #92964 - "Slave performance degrades over time". As description says:
    "When mysql slave is running gtid based asynchronous multi threaded replication, performance degrades over time in significant way when session_track_gtids is set to OWN_GTID; "
    The bug is assigned, but there are no public comments, while the way to reproduce seem to be defined clearly.
Same as the weather at Seven Sisters Cliffs back on September 14, 2018, some recent MySQL bug reports do not add confidence to those few visitors who explore them...

Now let me continue with some "Verified" bugs:
  • Bug #91981 - "Inconsistent user@host definitions for definer/grantor/grantee columns". It seems consistency is not a high priority these days.
  • Bug #92020 - "Introduce new SQL mode rejecting queries with results depending on query plan". I already mentioned this report while reviewing recent MySQL optimizer bugs. I have nothing to add to my conclusion there:
    "So, current MySQL considers different results when different execution plans are used normal!"
  • Bug #92032 - "Restrict usage of session foreign_key_checks". It's a great request from Federico Razzoli and I am happy to see it verified as a bug, fast.
  • Bug #92093 - "Replication crash safety needs relay_log_recovery even with GTID." Recently Jean-François Gagné pays special attention to MySQL replication crash safety in his bug reports, talks and blog posts. This specific report ended up as a request for a more clear documentation that should not provoke (false) safety feelings. See also his Bug #92109 - "Please make replication crash safe with GTID and less durable setting (bis)." for the request to improve MySQL in this regard one day.
  • Bug #92131 - "ASan: Direct leak of 272 byte(s) in main.mysqlpump_partial_bkp MTR test case". This kind of reports make me nervous. They make me think that still there is no regular testing of ASan-enabled builds in Oracle. Lucky we are, Percona engineers (like Yura Sorokin) do this for Oracle.
  • Bug #92133 - "DICT_SYS mutex contention causes complete stall when running with 40 mill tables". Proper data dictionary in MySQL 8.0 was supposed to solve some performance problems for instances with a lot of tables. As this perfect bug report from Alexander Rubin shows, this is not yet the case. We can still push MySQL instance over the limits and single instance for thousands of different small databases is hardly usable.
  • Bug #92209 - "AVG(YEAR(datetime_field)) makes an error result because of overflow". Perfect analysis from Pin Lin. All recent MySQL (and MariaDB) versions are affected.
  • Bug #92252 - "Mysql generates gtid gaps because of option slave-skip-errors". Yet another bug report from Pin Lin
  • Bug #92364 - "events_transactions_summary_global_by_event_name not working as expected". This bug was reported by Przemyslaw Malkowski from Percona. We all know Performance Schema is near perfect, but it turned out that "...performance_schema.events_transactions_summary_global_by_event_name behavior for instrumenting transactions seems completely broken.".
  • Bug #92398 - "point in time recovery using mysqlbinlog broken with temporary table -> errors". Let me quote Shane Bester:
    "Running a multi-threaded workload involving temporary and non-temporary tables leads to binary log playback giving errors."
  • Bug #92421 - "Queries with views and operations over local variables don't use indexes". This is a kind of regression in MySQL 5.7. MySQL 8 allows to workaround the problem properly using ROW_NUMBER() function.
  • Bug #92540 - "Comparison between DATE and DECIMAL does not round nanoseconds". This bug was reported by Alexander Barkov from MariaDB. MariaDB 10.3.x is not affected based on my tests.
To summarize:
  1. It seems ASan-enabled builds are not tested by Oracle engineers on a regular basis.
  2. Percona engineers help a lot with regular MySQL QA and testing for various extreme use cases.
  3. There is a lot to do to make GTID-based replication crash-safe and working fast in common use cases.
 More bug review are coming soon, so stay tuned.



Sunday, September 23, 2018

On MySQL Replication Bugs

While writing about problematic Oracle MySQL features previously I concentrated mostly on InnoDB problems that I have to fight with really often and deliberately skipped replication from even the preliminary list of features to study in details for that blog post. First of all, I mostly work with MariaDB users now, and implementation of many replication features in MariaDB is notably different already (and has its own list of known problems). But this happened also because (asynchronous) replication plays a key role in most MySQL environments and deserves a detailed study in a separate post.

It took me some time to prepare lists of recent bugs reported for MySQL replication (row-based and statement-based, GTIDs, multi-threaded and crash safe, hopefully) and then try to summarize my findings from them. Looks like I have something related to share today, finally.

Gondolas in Venice often look very similar to each other. They all are probably replicas of few different original designs. But small differences exist. This should not happen to your data in a replication setup...
Let me start with a summary of problems I found and then proceed with the lists of bug reports I based my conclusions on:
  1. Both minor problems with mysqlbinlog outputs, small regressions, missing useful features and real bugs (that may lead to data differences on slaves or break replication), hangs (including MTS deadlocks), crashes and performance problems with replication that can affect production setups do not seem to get high priority over last 2 years. Even those reported in public bugs database by famous MySQL Community members or current and former Oracle MySQL engineers.
  2. GTIDs and replication crash safety features were supposed to make it easier to use replication in production. In reality there are some GTID-specific ways to break replication, and it is still not as crash safe as expected in many use cases.
  3. Multi-threaded slave (a.k.a MTS or parallel replication) bugs, even serious and verified, are ignored for way more than last 2 years! The feature is not completed and may not help as much as one may expect.
  4. Manual is still wrong or missing when describing some details of current replication implementation.
  5. Nobody cares much about debug assertions found by Percona QA. We may have some bombs in the replication code that may explode in a weird way any time. 
  6. It seems group replication bugs often get proper attention fast, but nobody in Oracle has enough time to deal with asynchronous replication bugs properly. Some of them hang around without even proper verification for months...
Now let's get into details. Here is the list of some recent row-based replication specific bugs and feature requests:
  • Bug #91822 - "incorrect datatype in RBR event when column is NULL and not explicit in query". The column missing in the INSERT statement is given the datatype of the column immediately preceding it. This bug was reported by Ernie Souhrada back in July.
  • Bug #89339 - "RBR slave index choosing logic is too simple". As Fungo Wang pointed out based on source code review, it is based on rules, not considered data distribution at all. Slave uses primary key if it exists, otherwise tries unique key that does not allow NULLs (if any), then tries other keys and finally falls back to full table scan if there are no indexes. When there is no primary or unique key on columns not allowing NULL values, no statistics is taken into account when picking up indexes vs full table scan, and this easily leads to bad plans. Surely, one has to have primary keys in all tables, especially those involved in replication, but it would be great if slave code was a bit smarter in corner cases (as MariaDB's one).
    By the way, setting slave_rows_search_algorithms to HASH_SCAN may not help, as it may be too slow in practice. See Bug #84887 by Dennis Gao for more details and an example of perf and gdb usage for the analysis.
  • Bug #89297 - "please allow number to varchar slave_type_conversions if target is large enough". Nice feature request by Shane Bester.
  • Bug #88791 - "Binary log for generated column contains new value as WHERE clause, not old value". This bug was reported by MC Brown. Surely, mysqlbinlog output for row-based events is not supposed to be 100% valid SQL for copy-paste, but it should not be misleading.
  • Bug #88595 - "Row-based master-master replication broken by add column or drop column". This is more a documentation issue. It's strange to see this limitation not documented here.
  • Bug #88223 - "Replication with no tmpdir space and InnoDB as tmp_storage_engine can break". It's really weird to see this bug reported by Sveta Smirnova hanging in "Analyzing" status for almost a year!
  • Bug #88057 - "Intermediary slave does not log master changes with binlog_rows_query_log_events". Yet another bug report from Sveta, this time "Verified". Row query log events for UPDATEs and DELETEs are NOT recorded in the binary log by the intermediate masters...
  • Bug #85292 - "unreachable statement in function process_event with event type QUERY_EVENT". Shane Bester found a part of the code that probably should be rewritten. No further public attention to this report for 1.5 years...
  • Bug #84287 - "row inserts, statement updates, persistent stats lead to table scans+ lag slaves". Yet another great finding by Shane Bester. As often happens, another problematic feature I truly hate, InnoDB's persistent statistics, is also involved in this case, leading to bad performance on slave.
  • Bug #83431 - "RBR slave's "BEGIN" binlog query event doesn't carry a session database name". It was reported almost 2 years ago by Teng Li.
  • Bug #83418 - "When binlog_row_image is MINIMAL, some updates queries logs redundant columns". This was reported by Abhinav Sharma also almost two years ago.
  • Bug #83090 - "AUTO_INCREMENT not updated when using INSERT ON DUPLICATE KEY UPDATE". This bug was reported by Monty Solomon two years ago. See also his related Bug #83030 - "AUTO_INCREMENT not updated when using REPLACE INTO".
  • Bug #80821 - "Replication breaks if multi-table DELETE is used in conjunction with Foreign Key". I do not understand why this perfect (regression?) bug report by Philip Stoev, with a clear MTR test case was immediately verified and then just hangs around for 2.5 years...
  • Bug #80770 - "Compression of row-based event". Great feature contributed by Davi Arnaut.
  • Bug #79831- "Unexpected error message on crash-safe slave with max_relay_log_size set". Let me just conclude this with a 5.6-spcific (probably) bug report by yours truly...
I have to stop at this point, as other RBR-related bug reports are known for way more time and probably are just forgotten for good already by MySQL development. They hardly have a chance for the fix any time soon.

Let's continue with replication bugs that are not closely related to row-based binlog format:
  • Bug #92398 - "point in time recovery using mysqlbinlog broken with temporary table -> errors". Shane Bester found yet another case when statement-based binlog format may lead to problems for point in time recovery.
  • Bug #92252 - "Mysql generates gtid gaps because of option slave-skip-errors". This bug was reported by Pin Lin. We get gaps in GTID values on slave when errors are skipped. If binlog files on master include transactions which were applied on slave and hit errors in slave-skip-errors were purged, slave can't reconnect to master and reports error 1236. Features that were supposed to make replication more robust lead to more breaks in replication.
  • Bug #92132 - "secure-file-priv breaks LOAD DATA INFILE replication in statement mode on 5.7.23". This regression bug was reported by Nicolai Plum. Recent MySQL 5.7 and 8.0 versions are affected.
  • Bug #92109 - "Please make replication crash safe with GTID and less durable setting (bis)." Jean-François Gagné keeps looking for options to make replication really crash safe in all practically important cases.
  • Bug #92073 - "int change_master(): Assertion `mi->rli->inited' failed." Roel Van de Paar continues his QA efforts in a hope to make MySQL better one day and reports all debug assertions in public. In this case we seem to have a kind of regression in 8.0 vs 5.7. Check also these his bug reports: Bug #91980 - "Assertion `head->variables.gtid_next.type == ANONYMOUS_GTID' failed." and Bug #90946 - "Assertion `get_state() >= 0' failed on RESET MASTER while tx_read_only". Debug assertions are in the code for a reason usually, so in a latter bug report Roel recently asked for some update...
  • Bug #91766 - "Log of replication position when running CHANGE MASTER TO is wrong.". Jean-François Gagné found yet another annoying detail - one would expect the error log to have data about the SQL Thread position, but it has data about the IO Thread position.
  • Bug #91744 - "START SLAVE UNTIL going further than it should." Now this bug reported by  Jean-François Gagné is really serious. Still it hangs around "Open" since July 20...
  • Bug #91633 - "Replication failure (errno 1399) on update in XA tx after deadlock". XA is yet another MySQL feature that is doomed to stay half-baked it seems... More on this you should expect in a separate blog post one day. The bug that causes replication break was reported by Lukas Sydorowski.
  • The last but not the least bug in this list (that is actually long if we try to check all active bugs reported over last 2 years) that I want to include is Bug #90448 - "INSERT IGNORE not ignored on SQL thread", by Simon Mudd. Restarting slave helps and it seems MySQL 8.0.x is not affected. See also his Bug #89876 - "mysqladmin flush-hosts is not safe in GTID mode". 
I have to stop at some point. In conclusion let me review a list of (less) recent bugs in a multi-threaded slave (MTS) implementation specifically:
  • Bug #91477 - "Slave writes less when connected intermediary master with blackhole engine". Just 2 threads are used in this case as found by Sveta Smirnova.
  • Bug #90342 - "MySQL parallel replication automatically restarts". Nobody cares about this crashing bug for 5+ months. It's still "Open".
  • Bug #89247 - "Deadlock with MTS when slave_preserve_commit_order = ON.Jean-François Gagné reported this 9 months ago, the bug was verifvied and based on last comments and pt-pmp outputs the hang really happen for many users. Still nobody cares in any obvious way.
  • Bug #89229 - "FLUSH PRIVILEGES may cause MTS deadlock". This bug was reported by Libing Song 9 months ago.
  • Bug #87796 - "Commit_order_manager can't terminate MTS worker properly when deadlock happens". This bug was reported by Seunguck Lee a year ago!
  • Bug #85747 - "Documentation on MTS statistics". Yet another case of missing manual reported by Juan Arruti:
    "I didn't find information in MySQL reference manual about MTS statistics, I had to look into source code in order to get more information about them."
  • Bug #85142 - "reducing MTS checkpointing causes high IO load". The bug reported by Trey Raymond is "Verified" for 1.5 years. Check also his Bug #81386 - "MTS hangs on file related *_EVENT forced checkpoint" and Bug #81385 - "relay log info not updated after using sql_after_mts_gaps" that are "Open" since May 2016!
  • Bug #73066 - "Replication stall with multi-threaded replication". Ovais Tariq reported this bug more than 4 years ago, when both still worked in Percona. Nobody cares to fix it.
  • Bug #71495 - "Logging of MTS breakage is insufficient." One of Oracle customers who use MTS in production, Simon Mudd had reported this practical problem 4.5+ years ago. No fix, no comments from Oracle engineers from more than 3 years.
I have to stop at this stage, as checking more/older replication related bugs makes me even more sad than I planned or want to be. We see fixes in every release, but the backlog seems huge. Dear Oracle MySQL engineers, please, review public MySQL replication bug reports collected over last 4-5 years and spend some more time and resources on processing and fixing them! Otherwise this Acqua alta of bugs will make this great legacy feature obsolete one sad day.

I had not planned to study semi-sync replication bugs in this post (maybe, some day). If you are interested in more details about this feature, crash safety of replication, parallel replication and replication in general, I suggest you to check J-F Gagné's MySQL Blog.