While helping @mperham in #20989 I realized that the current code (under concurrent load) is broken.

The problem arises if there are multiple threads in app space region of Dependencies.interlock, then an .rb file is changed. Next thread arrives at ActionDispatch::Reloader middleware, it invokes reloaders to check if any of the files have changed, and indeed there is a change, the thread attempts to grab an exclusive Dependencies.interlock (via poll) does not succeed (because of those other threads in app space) and proceeds on.

However reloaders report a change in .rb file only once, subsequent calls to updated? return false, this is because since the last time updated? has been invoked nothing has changed, so if an exclusive grab of Dependencies.interlock was not successful a change to .rb file is "lost".

Additionally, if a typical use case of sidekiq is somewhat similar to @mperham's example in #20928 (comment), I'm fairly confident that even if the "lost" .rb change issue is fixed, unless unloading is blocking (and an attempt to obtain an exclusive lock blocks out new readers), then update is going to be probably "starved" anyways.

/cc @matthewd