Video Conferencing with Durable Streams
12 points by infiniteregrets
12 points by infiniteregrets
it's a neat demo showing just how much utility you get "for free" when using a fast durable stream backend, but uh, shoving JPEGs one at a time is about the worst possible way to do video streaming in almost all aspects (encode/decode performance, amount of data being sent, latency due to that) with few benefits (no PLIs because everything is a keyframe 🙃). Same criticism applies to their predecessors. again, neat demo, very bad architecture for anything real.
A long time ago there was an article published on the Amazon Engineering blog about improving a video processing pipeline performance by 100x. It turned out that in drinking the cloud Koolaid they had done exactly this, using a separate Lambda call and S3 object for each frame of video processing. It apparently worked fine for a while until, I presume, someone noticed the cost. They removed the blog a while ago, at the time I believed it was because they were embarrassed, but it was probably just bitrot.