3 ms·
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.
by tombert 10y ago
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.
- vishalzone2002 10y agoI work on data science. I also had an idea around youtube. Want to have a chat?
- tombert 10y agoSure, you can email me at thomas@gebert.sexy (yes, really).
- vram22 10y agoPartly related anecdote: I did some contract work some years ago for a startup (Tellytopia, later acquired by Concurrent Systems) that was into the video over IP space. Video transcoding was a part of the project, but I didn't work on that, which was in C++. I worked on some of the surrounding parts, in Python with XML-RPC, talking to Java, across machines. I remember that all or most functions in the app were XML-RPC-enabled, so that they could be located in different machines, which must have helped somewhat with scaling.