Comment by mochomocha

Yep in Netflix case they pack bare-metal instances with a very large amount of containers and oversubscribe them (similar to what Borg reports: hundreds of containers per VM is common), so there are always more runnable threads than CPUs and your runqueues fill up.