The nice thing (at least from my admittedly-fairly-ad-hoc experience) about video processing is that it scales fairly linearly, at least for the basic features.
Things like building a scalable transcoding cluster can be made fairly simple ("just add more boxes") using a decent message-queue service. From there, everything is pretty simple HTTP.
Of course, this is only the very barebones stuff, and covering all of Youtube's features exhaustively would be quite difficult, and requires some good Machine Learning work.
Things like building a scalable transcoding cluster can be made fairly simple ("just add more boxes") using a decent message-queue service. From there, everything is pretty simple HTTP.
Of course, this is only the very barebones stuff, and covering all of Youtube's features exhaustively would be quite difficult, and requires some good Machine Learning work.