Yes; but it's for the duration of work on that completion, it's not like sockets are owned by the thread, so it doesn't have the same misbalancing effect, where

Yes; but it's for the duration of work on that completion, it's not like sockets are owned by the thread, so it doesn't have the same misbalancing effect, where if the original thread is not available completions get starved. The LIFO nature is a side effect of the user mode selection of next completion to run; something round robin would be slower with kernel transitions.