Edit this page | Blame

Hanging database

Mariadb occassionally stops responding. We think updating myisam to InnodB is the solution because it prevents full table locking.

Also it is better to use mydumper, mydumper locks MyISAM tables and does not lock InnoDB so the dump is consistent.

Tags

  • assigned: pjotrp, zsloan
  • type: bug
  • keywords: database
  • status: unclear
  • priority: high

Tasks

for Penguin2, Tux01 and Tux02

  • First on Penguin2
  • Update mariadb to latest
  • Convert fulltext tables (see below)

+ ProbeSet + GeneRIF_BASIC + pubmedsearch

  • Good candidates

+ 2.1G Dec 4 22:15 ProbeSetXRef.MYD + 2.3G Dec 18 14:56 ProbeSet.MYD + 2.6G Aug 27 2019 ProbeSE.MYD + 7.1G Nov 2 05:07 ProbeSetSE.MYD + 11G Aug 27 2019 ProbeData.MYD + 63G Dec 4 22:15 ProbeSetData.MYD

  • Create test for every table that is going to switch
  • Convert largest tables to innodb
  • After some testing do same for Tux01 and Tux02

Tux02


Info

Mariadb is 'hanging'

In the last 12 hours GN2 monitoring shows the website is responding intermittendly. A quick check shows the database is blocking. Rather than simply restarting the database - which is known to sort the issue - the timing is that the US is sleeping so I can do some checking. Let's take a look.

Mariadb is at 4x CPU

PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
 40514 mysql     20   0   37.3g   1.8g  18268 S 394.4   0.7   4855:33 mysqld

The ps table shows a backup is ongoing

root      57559  0.0  0.0   2388   756 ?        Ss   03:00   0:00 /bin/sh -c /bin/su mysql -c /export/backup/scripts/tux01/backup_mariadb.sh  >> ~/cron.log 2>&1
mysql     57588  0.2  0.0 200112 27292 ?        Sl   03:00   0:25 mariabackup --backup --target-dir=/home/backup/tux01_mariadb_new/latest/ --user=webqtlout --password=x xxxxxxx

Tales in use:

MariaDB [db_webqtl]> show open tables where in_use > 1;
+-----------+----------------+--------+-------------+
| Database  | Table          | In_use | Name_locked |
+-----------+----------------+--------+-------------+
| db_webqtl | InfoFiles      |      5 |           0 |
| db_webqtl | GeneRIF_BASIC  |      3 |           0 |
| db_webqtl | ProbeFreeze    |      4 |           0 |
| db_webqtl | ProbeSet       |      4 |           0 |
| db_webqtl | ProbeSetXRef   |      4 |           0 |
| db_webqtl | Species        |      4 |           0 |
| db_webqtl | Geno           |      4 |           0 |
| db_webqtl | Chr_Length     |      2 |           0 |
| db_webqtl | Tissue         |      4 |           0 |
| db_webqtl | ProbeSetFreeze |      4 |           0 |
| db_webqtl | InbredSet      |      4 |           0 |
+-----------+----------------+--------+-------------+
11 rows in set (0.001 sec)

In the error log we seeing a lot of

2021-12-21  6:17:57 256179 [Warning] Aborted connection 256179 to db: 'db_webqtl' user: 'webqtlout' host: '128.169.5.59' (Got timeout reading communication packets)
SHOW FULL PROCESSLIST;
458 rows in set (0.003 sec)

with entries

| 256363 | webqtlout   | 128.169.5.59:59120 | db_webqtl | Query      |    15 | Waiting for table flush  | SELECT Id, Name, FullName, ShortName, DataScale FROM ProbeSetFreeze  WHERE  public > 0 AND (Name = "CB_M_0305_R" OR FullName = "CB_M_0305_R" OR ShortName = "CB_M_0305_R")

waiting for tables to flush!

at the top of the process list we find

Id  User  Host  db  Command Time  State Info  Progress
1 system user   NULL  Daemon  NULL  InnoDB purge coordinator  NULL  0.000
2 system user   NULL  Daemon  NULL  InnoDB purge worker NULL  0.000
3 system user   NULL  Daemon  NULL  InnoDB purge worker NULL  0.000
4 system user   NULL  Daemon  NULL  InnoDB purge worker NULL  0.000
5 system user   NULL  Daemon  NULL  InnoDB shutdown handler NULL  0.000
227365  webqtlout 127.0.0.1:33950 db_webqtl Sleep 13015   NULL  0.000
245634  webqtlout 127.0.0.1:38098 db_webqtl Sleep 23180   NULL  0.000

This is quite informative:

it suggests that the backup can be the root of the problem.

And then

suggests

  • Wait for the long-running queries which are blocking the FLUSH TABLE to complete;
  • Identify the long-running queries and kill them;
  • Restart the server

and that is somewhat amusing.

Stripping out all reqular queries we get:

grep localhost test.out |grep -vi probesetfreeze|grep -vi species
255559  webqtlout       localhost       NULL    Query   13092   Waiting for table flush FLUSH NO_WRITE_TO_BINLOG TABLES0.000
256351  webqtlout       localhost       db_webqtl       Field List      1588    Waiting for table flush NULL    0.000
256383  webqtlout       localhost       db_webqtl       Query   0       Init    SHOW FULL PROCESSLIST   0.000

and it appears everyone is waiting for id 255559. Let's kill that.

kill 255559;

and inspect

mysql -u webqtlout -pwebqtlout db_webqtl -A -e "show processlist;"|less

it started processing again. To speed up recovery back to:

systemctl restart mysql

of course that stopped the running backup. But processing is back in business.

My first conclusion is that this problem was triggered by the backup procedure. Interestingly, it happens irregularly. We also have seen this issue before this backup procedure was instated, so I figure it has to do with Mariadb.

The version on production is from 2017 - we should update that soon:

Server version: 10.3.27-MariaDB-0+deb10u1-log Debian 10

we have been running a more recent version of mariadb on luna. Still, that is unlikely to fix this issue because I think it really has to do with myisam and locking of large tables. Switching to innodb does away with global locks and is the default on mariadb (there are less and less people using myisam).

Mariadb table locked

Arthur reports: MariaDB is not responding Saturday, July 24 2021 at 10:48 pm. I tried to enter data to the table ProbeSetXRef.pValue and when normally takes few seconds, now is more than 10 minutes without completion/responding.

Zach: before restarting the table can you check the status first?

some ideas here

we are still using MyISAM for these tables: a switch to InnoDB may help.

Another round of mariadb

Arthur complained again that the DB is slow. The standard performance tests are not failing. The slow log shows slow queries:

# Thread_id: 1715564  Schema: db_webqtl  QC_hit: No
# Query_time: 399.159339  Lock_time: 0.000000  Rows_sent: 0  Rows_examined: 0
SET timestamp=1647055599;
UPDATE ProbeSet SET description = REPLACE(description, ";", "") WHERE ChipId=11;

# Query_time: 2006.780492  Lock_time: 0.000000  Rows_sent: 0  Rows_examined: 0
SET timestamp=1647057611;
update ProbeSetXRef set mean = (select AVG(value) from ProbeSetData where ProbeSetData.Id = ProbeSetXRef.DataId) where ProbeSetXRef.ProbeSetFreezeId = 385;

# Query_time: 1158.804376  Lock_time: 0.000000  Rows_sent: 0  Rows_examined: 0
use db_webqtl;
SET timestamp=1647142312;
update ProbeSet set description="ArfGAP with SH3 domain, ankyrin repeat and PH domain 3" where Id=1719426 AND ChipId=11;

so we should be able to reproduce that.

  • mariadb was restarted 1 week ago
  • mariadb is not eating CPU or RAM, so the general state looks healthy.
  • Disk state looks healty - but there are files in /tmp which do not belong there.
  • The ps table shows a backup is ongoing (after 3 hours).
  • There are no tables in use according to `show open tables where in_use > 1;`
  • There are no blocking prcesses according to `SHOW FULL PROCESSLIST;`
  • Number of threads is a healthy 4 with `show status where `variable_name` = 'Threads_connected';`

So, no issues. And this query runs at

MariaDB [db_webqtl]> update ProbeSet set description="ArfGAP with SH3 domain, ankyrin repeat and PH domain 3" where Id=1719426 AND ChipId=11;
Query OK, 1 row affected (0.210 sec)
Rows matched: 1  Changed: 1  Warnings: 0

There must have been an update block. Best to do diagnostics when the system is blocking.


(made with skribilo)