For the core application of writing pixels to the board, I would of used a more efficient monolithic design with less points of failure. One big server could ha

For the core application of writing pixels to the board, I would of used a more efficient monolithic design with less points of failure. One big server could handle 333 pixels/s(their expected volume). Not sure that using cassandra and redis is really necessary or makes things easier. The secondary goal of pushing pixel updates to lots of users would probably want separate hardware. I would probably fake it so its not really realtime. If you send the web client batches of updates to replay on the board it looks realtime and doesn't require persistent connections.