I would do significant load testing for each use case before using it (or socket.io for that matter). I needed to push 30-50 messages per second to each connected client and faye started choking as soon as there were more than 20-30 clients. socket.io would choke at around 50-100 connected clients. Raw websockets were able to reach closer to 100 clients but performed more or less similar to socket.io.
We kept the design simple, otherwise a "batching" mechanism could be introduced that would replay a single batch of 50 messages but that would make everything a second delayed. However, part of the map allows you to login and see your messages live on the screen as you're sending them to your partner and for that the real time streaming is pretty critical.
A scalability question:
You note "Turn on sticky load balancing (for example by origin IP address). This ensures that long-polling connections for example always route requests to the same node where buffers of messages could be stored."
I read this to mean that we are responsible (in our load balancer/proxy/etc) to keep connections from clients returning to the same server. This is OK, but what about nodeJS clusters? How should I ensure that client A always connects to cluster-node member 3?
Related section of the blog post: http://socket.io/blog/introducing-socket-io-1-0/#scalability