The RFC meeting today covered the multi-datacenter work (https://www.mediawiki.org/wiki/Requests_for_comment/Master_%26_slave_datacen...).

The basic summary is that there will be one datacenter used for all POSTs, and that GETs will go to the closest data center.

This means GET should mutate state as little as possible (since otherwise the far-off GET center will have to write to the main data center in Virginia), delaying the user. There are special provisions for avoiding this in the cases where writes are needed.

It's general practice that GET should never mutate state anyway, but there exceptions to this in MediaWiki (though some, e.g. rollback, will be fixed as part of this).

For example, when a GET needs to queue a job, it uses EnqueueJob, which adds the job to a local job queue. It is then copied to the real job queue without delaying the user.

I know they've already worked with us on how this affects Flow, and the RFC commits to documenting the new patterns further.

Matt Flaschen