Millions of user_newtalk requests, conspicuously absent from the typical dump. Many of them are doing "statistics", whatever that means, and the rest are locked. One side of me is curious as to what "statistics" means and why it produces this behaviour. The other side of me says
Could these hanging selects be just a part of the phenomenon? I mean, assume, the mysql has some reason not to respond, but still accepts new requests. Then all anonymous users reading a page will add a new process looking for a newtalk flag. They don't cause the slowdown, they are its victims.
So the question is: Why does it slow down every 15 minutes? Do we have some cron jobs running every quarter of an hour?
I think you might be right that it's a symptom not a cause. I tried disabling user_newtalk requests for anonymous users, but the problem still occurred. But instead of large numbers of user_newtalk requests blocking, it was article text requests. Common to both was that the blocked queries were all in the "statistics" state.
Is it a cron job? Well if so the period should be perfectly regular. In the captured output previously attached, there are 3 events, with the separations between the events being 23'20" and 19'25". Another file I have here shows a period of 22'05" between two events. Performance tends to slow down before it actually stops, so I'd put error bars on those times of about 20 seconds.
Currently I'm blaming the statistics state. Conventional wisdom is that a long statistics state is caused by a complex join with no specified index.
-- Tim Starling.