The Analytics Mysql Meta instance is a Mariadb database currently running on the Analytics Hadoop coordinator host (an-coord1001), that hosts some low volume db schemas related to Hadoop tools like Hive, Oozie, Druid, etc..
MariaDB [(none)]> show databases; +---------------------+ | Database | +---------------------+ | druid | | druid_public_eqiad | | hive_metastore | | hue | | | oozie | | search_airflow | | superset_production | | superset_staging | +---------------------+
The Mariadb instance collects metadata about the following systems:
- Both Druid clusters (Analytics/Systems/Druid)
- Superset staging and live hosts (Analytics/Systems/Superset)
- Hue (Analytics/Cluster/Hue)
- Oozie (Analytics/Systems/Cluster/Oozie)
- Hive Server2 and Metastore (Analytics/Systems/Cluster/Hive)
- Airflow for the Search instance
Please remember that any maintenance to the database will need to keep into consideration that we have replication to db1108.
This is a mistake that I made after replication was set up the first time:
- created a new db called hue_next (I was testing a new version of Hue)
- granted to the hue user some privileges for hue_next
- replication broken on db1108 + icinga alarms
The main problem was that at the time we replicated only the databases, not the users on db1108 (since we only needed the replication user and not the rest) so when the grant command was replicated, it failed. The fix was easy, basically adding users to the db on db1108, but please keep this in mind when operating on an-coord1001.
There are currently two backup workflows currently running:
- All the databases are periodically backed up in /srv/backup/mysql/analytics-meta on the Analytics Hadoop Master Standby (an-master1002). Daily snapshots are backed up into HDFS at /wmf/data/archive/backup/mysql/analytics-meta/.
- All the above databases are replicated to db1108, see Analytics/Systems/DB_Replica. db1108 dumps are taken periodically into the db provisioning/recovery hosts (dbprov1002), which are stored long term into Bacula
Restore a Backup
There are multiple ways to restore data:
- Analytics/Systems/DB_Replica shows that we have a replica on db1108 and bacula backups. MariaDB/Backups#Pre-requisites before recovering the backup and below shows how to do it from dbprov hosts, otherwise it is always possible to run mysqldump or xtrabackup from the replica on db1108.
- We keep 14 days of binlog, that means being able to recover to a specific transaction in time if needed. See https://dev.mysql.com/doc/refman/8.0/en/point-in-time-recovery-binlog.html. The main downside is that the binlog is global, so not specific to a certain database.
This procedure should be used only if an-coord1001 is not available for some reason, for example unrecoverable hardware failure. There is no automatic failover, only a manual one (described by this procedure). The high level idea is the following:
- Check what daemons are running on an-coord1001. The Hive metastore and oozie probably need to be moved elsewhere, check what host/vm can hold them temporarily.
- Ssh to db1108, and stop replication on the Analytics Meta's mysql instance (
sudo mysql -S /run/mysqld/mysqld.analytics_meta.sockand
- Stop bacula backups for the db1108's mysql instance, since we found out in the past that the backup locking overhead causes timeouts and problems to daemons connecting to the Analytics Meta DB.
- Set the db1108's meta replica in puppet to read/write (check
- Look for all occurrences of an-coord1001 in puppet and figure out the ones that need to be moved to db1108. Keep in mind that the port of the mysql instance on db1108 is not 3306, so replacing an-coord1001.eqiad.wmnet with db1108.eqiad.wmnet in puppet is not enough. A high level list is:
- hive metastore mysql configuration (hive-site.xml)
- oozie mysql configuration (oozie-site.xml)
- druid clusters configuration (druid common config, in hiera)
- superset configs (superset_config.py)
- hue config for oozie and hive
- airflow mysql config for Search (alert them when this happens to warn about airflow working in degraded mode).
- Merge the above change and run puppet across the hosts, you'll likely need to roll restart the impacted daemons.
- At this point, you should have recovery.
- Important: Bacula backups are happening weekly and they may interfere with daemons using the database, due to locking etc.. In order to temporary disable backups, comment the lines related to analytics meta in https://gerrit.wikimedia.org/r/c/operations/puppet/+/617650/1/modules/profile/templates/mariadb/backup_config/dbprov1002.cnf.erb
This procedure should be used to restore service after a failover from an-coord1001 to db1108 happened. Since the dbs on an-coord1001 need to be rebuild, this will need some downtime for all services.
The high level idea is the following:
- Set up some downtime for all the daemons using the analytics meta instance (see in the Failover section).
- Dump the databases on db1108, copy it to an-coord1001 and restore them (see Analytics/Systems/DB_Replica for info about how to do it). You may probably need to drop the existing databases/data before doing so.
- Move all daemons to an-coord1001 (read the Failover procedure for more info).
- Set db1108's instance to read-only and set up replication again.
- Re-enable Bacula backups.
- It is possible to reply events from a single database with -d