Facebook users upload so many videos, it had to invent a whole new system to deal with them
Luckily for Facebook's dedicated users, the Streaming Video Engine, as it's called, was able to be put in place behind the scenes, without anyone ever noticing.
All users know is that their videos are getting uploaded and ready for viewing up to 10 times faster than they were before.This new Streaming Video Engine was built during Facebook's ramp-up from 1 billion video views per day in January 2014 up to more than 8 billion a day now - a milestone it hit in late 2015.
The key concept is the engine's ability to break up a video into chunks, letting Facebook process and upload several pieces simultaneously. Before, Facebook had to process videos all in one go, giving the system one big chokepoint.
Now that it's in place, Facebook's new Streaming Video Engine handles all the video uploads from all of its apps and services, including the core social network, Facebook Messenger, and Instagram.
And it's just in time, too, Facebook says - since being put in place late last year, the Streaming Video Engine has already dealt with 3 times the amount of peak video traffic as the old system, Facebook says.
Preparing for the next video wave
Looking to the future, the Streaming Video Engine is a solid foundation for serving up the next wave of video, Facebook says. That includes recent features like Profile Videos, as well as 360-degree video content that's primed for virtual reality.
Indeed, virtual reality is also a big focus area for Facebook's video teams, given the fact that it owns the forthcoming and much-hyped Oculus Rift VR headset.To that end, Facebook also today announced new updates to its 360-degree video technology that reduces the file size of uploaded VR videos, while also dramatically cutting down on the buffering time to play them. That's important, given the sheer quantity of data that a 360-degree, high-resolution VR video has to get across.
Finally, Facebook announced today a new approach to its artificial intelligence strategy for chopping up and analyzing videos. Right now, most models for artificial intelligence use a so-called "supervised" model, where AI engineers have to help a computer understand what it's seeing in a video.
But with a new model discussed today, Facebook AI Research's Vision Understanding team thinks it has the foundation for letting the system learn what's in videos on its own, through a novel process where it analyzes each "voxel," or video pixel, and crunches it individually.
Processing, analyzing, and serving up videos is both a huge challenge and a huge opportunity, especially in the era of streaming video and virtual reality - just ask IBM, which just launched a big video services initiative earlier today.