This issue is fully resolved.

We have been stable since the last announcement. We have found the underlying issue and we are deploying a fix. We will continue monitoring the situation.

Issues continue. We're applying manual fixes to prevent 408s from happening and we're working on a long-term fix in the meantime.

Beginning at roughly 10 PM UTC yesterday, up until 4:45 PM UTC today, we were experiencing an elevated rate of timeout errors. This was caused by a bug in a refactor we shipped that only showed up under very high load. This issue has been resolved -- if you are still experiencing elevated 408 error rates, please contact us.