Pretty proud of this. We are testing distributed processes synchronized via pipes which makes it deterministic. Pretty cool.
We boot a listener in the parent process we then fork. Before we touch the file we verify the fork is booted using pipes. Then the parent process will touch the file while the fork waits on a pipe. Once the parent process signals that the file has been touched we continue inside of the fork.
We need one file checker booted per process as talked about in #24990. Before we do a check to see if any updates have been registered by the listener we first check to make sure that the current process has booted a listener.
We are intentionally not starting a listener when the checker is created. This way we can avoid #25259 in which puma warns of multiple threads created before fork. As written the listener for each process will be invoked by the `ActionDispatch::Executor` middleware when the `updated?` method is called. This is the first middleware on the stack and will be invoked before application code is read into memory.
The downside of this approach is that the API is a little less obvious. I.e. that you have to call `updated?` to get the listener to start is not intuitive. We could make `boot!` not private if we want to make the API a little nicer. Alternatively we could boot when the checker is initialized however this reintroduces the puma threads warning, and also means that in cases of `rails server` or when using `preload!` that we have extra threads notifying of changes on a process that we don't care about.
[close#24990] [close#25259]
