Page MenuHomePhabricator

Possible to run writes (e.g. UPDATE) on Beta Cluster replica
Open, NormalPublic


I assume this is the same in prod, but I'm not sure.

I accidentally did a write operation on:

sql enwiki

on the Beta Cluster (see T109982), because I thought that still provided master. legoktm told me it now provides slave, but the UPDATE did not fail.

Not sure if this also applies to stat1003.

(I'm going to try to correct this case by re-running in master.)

Event Timeline

Mattflaschen-WMF updated the task description. (Show Details)
Mattflaschen-WMF raised the priority of this task from to Needs Triage.
Restricted Application added subscribers: Luke081515, Matanya, Aklapper. · View Herald TranscriptAug 24 2015, 10:17 PM
Mattflaschen-WMF set Security to None.
Krenair added a subscriber: Krenair.EditedAug 24 2015, 10:25 PM

All the sql command was changed to do was connect to a server returned by wfGetLB()->getServerName(wfGetLB()->getReaderIndex()) by default instead of wfGetLB()->getServerName(0), see T105046

However, it does appear that read_only is OFF on both deployment-db1 and deployment-db2.

Not sure if this also applies to stat1003.

Isn't that a production analytics host? What does it have to do with anything?

Isn't that a production analytics host? What does it have to do with anything?

There are also slaves of the wiki databases accessible from there:

mysql:research@s3-analytics-slave [mediawikiwiki]> SELECT * FROM page ORDER BY page_id DESC LIMIT 1;

Wondering if those slaves would also allow accidental UPDATEs.

jcrespo moved this task from Triage to Backlog on the DBA board.Sep 8 2015, 9:18 AM
jcrespo triaged this task as Normal priority.Sep 8 2015, 9:46 AM
jcrespo added a subscriber: jcrespo.

Analytics stores (dbstoreX) are all with read_only=0 (aka s3-analytics-slave). I do not know yet, I suppose that because they have their own databases that they have to write to.

Slaves on beta should be fixed, but I will need help for that.

I have also created T111766.

hashar added a subscriber: hashar.Sep 8 2015, 9:53 AM

From wmf-config/db-labs.php the Beta-Cluster-Infrastructure has two SQL servers, a master and a slave:

        'hostsByName' => array(
            'deployment-db1'  => '', # deployment-db1.eqiad.wmflabs
            'deployment-db2'  => '', # deployment-db2.eqiad.wmflabs
        'sectionLoads' => array(
            'DEFAULT' => array(
                'deployment-db1'     => 0,
                'deployment-db2'     => 400,

I think at one point some database had db2 has the master and db1 has the slave but that is no more the case apparently.

So it seems it is all about ™ setting the read-only flag on deployment-db2. They both have the puppet class `role::mariadb::beta which apply the template templates/mariadb/` that comes with:

read_only = 0

Maybe we can pass a type master/slave to vary that setting? Could be done via hiera maybe.

This caused complications when trying to fix T148111: DBQueryError when creating account on Beta, or logging into newly-created accounts too. @bd808 accidentally ran the ALTER on the slave, I then ran it on master, but that stopped replication, so we had to figure out how to restart it (it's fixed now).

Mattflaschen-WMF renamed this task from Possible to run writes (e.g. UPDATE) on slave to Possible to run writes (e.g. UPDATE) on Beta Cluster replica.Oct 14 2016, 9:37 PM

Just judging from the task title, this and T183245: Ensure replica DB in labs is read-only look like being duplicates?

hashar removed a subscriber: hashar.Jul 31 2018, 11:11 PM