Here are metrics from where the master does not accept read queries, you can see where the change was made and the uneven load of read query between the two replicas

To investigate this, I would need more details:

• MediaWiki version
• PHP version, and which db driver(s) are installed.
• Which of the MediaWiki configuration variables relating to DB and LB are set, and their values (all except specific hostnames/passwords).

Also, given way state is shared between web requests, it the object cache setup may also play a role here:

• Is APCu installed?
• What are $wgMainCacheType, $wgMainWANCache and $wgMainStash set to. And if their values are not a default key of $wgObjectCaches, then the value of that array as well.

• 1.34.0
• 7.2.24-0ubuntu0.18.04.2 (apache2handler)
• 10.1.34-MariaDB

<?php
# Single
#require_once( 'LocalSettings.DB-single.php' );
# Replica
require_once( 'LocalSettings.DB-replica2.php' );

<?php

$wgDBname = "ffxiv_wiki";

$wgDBservers = array(
         array(
                 'host' => "",
                 'dbname' => "",
                 'user' => "",
                 'password' => "",
                 'type' => "mysql",
                 'flags' => DBO_SSL,
                 'load' => 0,
         ),
         array(
                 'host' => "",
                 'dbname' => "",
                 'user' => "",
                 'password' => "",
                 'type' => "mysql",
                 'flags' => DBO_SSL,
                 'load' => 1,
         ),
         array(
                 'host' => "",
                 'dbname' => "",
                 'user' => "",
                 'password' => "",
                 'type' => "mysql",
                 'flags' => DBO_SSL,
                 'load' => 1,
         ),

);

Is APCu installed? Yes

<?php
$wgMainCacheType = CACHE_MEMCACHED;
$wgMainWANCache (not set, using default)
$wgMainStash (not set, using default)

Thanks, I'll let CPT triage this further.

There are a few things that might contribute here:

• With 'load' being 1, if anything happens that causes MW to try to scale back it can only scale back to 0. You might try setting them all to 100 instead.
• Servers that are lagged are avoided. When this happens, it looks like it gets logged at the "debug" level to the DBReplication channel.
• I've verified that the selection based on adjusted weights is fair: ArrayUtils::pickRandom() does not show bias.

Further testing on my part will have to wait until I have time to figure out how to set up replication locally to more directly test things. In the mean time, you might try repeatedly using maintenance/eval.php or maintenance/shell.php to execute wfGetDB( DB_REPLICA )->getLBInfo() to more directly measure the selection of connections.

Something something documentation.

https://www.mediawiki.org/wiki/Manual:$wgDBservers doesn't really go into detail how the values for load work other than an example of 0 and 1. So when you suggest something like setting values to 100 it doesn't really convert in my head well.

Yeah, documentation often is lacking. :( And the DB configuration in MediaWiki could probably really use an overhaul, I think there are two or three different systems of configuration thanks to people adding things over the years but never deprecating the old ways.

The real question is whether setting them to 100 makes a difference for you, or if it still behaves the same. That could help point us in the right direction.