I know memory is fairly cheap but e.g. there are millions of new videos on youtube everyday, each probably few hundred MBs to few GBs. It all has to take enormous amount of space. Not to mention backups.
I know memory is fairly cheap but e.g. there are millions of new videos on youtube everyday, each probably few hundred MBs to few GBs. It all has to take enormous amount of space. Not to mention backups.
Not only that but for each video on YouTube there are different versions for each resolution. So if you upload a 1080p video, it gets converted to 1080p AVC/VP9, 720p AVC/VP9, 480p… also for the audio.
If you run
youtube-dl -F <youtube url>you will see different formats.Does youtube actually store copies of each one? Or does it store 1 master copy and downsaple as required in real time. Probably stores it since storage is cheaper than cpu time
If it converts every video in realtime it will require a lot of CPU per server, it’s cheaper to store multiple copies. Also the average video isn’t more than some 300MB, less if it’s lower quality.
Anyone with Plex or Jellyfin knows that it’s better to have the same movie in both qualities (1080,720) the transconding to avoid CPU usage.
It’s possible to have fast transconding with GPUs, but with high so many users on youtube that will require a lots of power and high energy prices, store is cheaper.
I believe they store and that’s why it processes lowest res first and works up
It’s transposed on the fly, this is a fairly simple lambda function in AWS so whatever the GCP equivalent is. You can’t up sample potato spec, the reason it looks like shit is due to bandwidth and the service determining a lower speed than is available.
Are you suggesting they don’t store different versions? This (speculative ik) suggests they do.
That response is almost 10 years old and completely outdated. I’ve designed and maintained a national media service and can confirm that on the fly transcoding is both cheaper and easier. It does make sense to store different formats of videos that are popular at the minute but in the medium to long term streams are transcoded.
Sure it’s old but the stats I posted in a lower comment show that at YouTube’s scale, it makes sense to store.
At a certain point the cost of compute is going to be cheaper than the cost of storage.
Do you have a source? My instinct is the opposite. Compute scales with users but storage scales with videos
No source but I imagine the amount of videos must be outpacing the amount of users. Users come and go but every uploaded video stays forever.
I think you might be underestimating how many users YouTube has! According to this, 720,000 hours per day are uploaded versus 1,000,000,000 hours are watched per day!
Consider two cases:
Design a system that optimizes for total cost.
It probably depends on how popular the video is anticipated to be.
I remember hearing that something like 80% of uploads to YouTube are never watched. 80% of the remaining 20% are watched only a handful of times. It’s only a tiny fraction that are popular, and the most popular are watched millions of times.
I’d guess that they don’t transcode the 80% that nobody ever watches. They definitely transcode and cache the popular 4%, but who knows what they do with the 16% in the middle that are watched a few times, but not more than 10x.
In real time would mean more cpu usage every time someone plays it. If converted in advance, they only need to do it once with the most effective codecs.