janPasi: I love how you started posting "as soon as we get more than a handful of users" when you have a crazy huge instance ^^
«we had about 10% load» I don't have enough experience, but doesn't that mean that the workers weren't the issue? And that you might have perf issue again.
i/o delay: were the workers threads in the "D" (uninterruptible) state when running `htop` or `top` or `ps`? Which mean there were waiting on the disk or a network socket.
We could have a situation with 10% CPU load and workers most of the time waiting for IO (D state)