Summary:
Saw an issue with a build because of an ENONT error: https://fb.facebook.com/groups/716936458354972/permalink/923628747685741/
My hypothesis:
1. We issue a ping to the socket (in SocketInterface/index.js) a decides if the available socket is alive
2. We see that it's alive but by the time we actually connect to it the server would've died
Solution:
1. The server shouldn't die as long as there are clients connected to it (currently it only stay alive as long as there are jobs)
2. The "ping" should only disconnect once the client is connected
3. Finally, have a better error message than ENOENT
Summary:
A few potential races to fix:
1. Multiple clients maybe racing to delete a zombie socket
2. Servers who should die because other servers are already listening are taking the socket with them (move `process.on('exit'` code to after the server is listening
3. Servers which are redundant should immediatly die
Summary:
Buck (our build system) currently starts multiple packager instances for each target and may build multiple targets in parallel. This means we're paying startup costs and are duplicating the work. This enables us to start one instance of the packager and connect to it via socket to do all the work that needs to be done.
The way this is structured:
1. SocketServer: A server that listens on a socket path that is generated based on the server options
2. SocketClient: Interfaces with the server and exposes the operations that we support as methods
3. SocketInterface: Integration point and responsible for forking off the server
Martín Bigio
Amjad Masad