Reimagining video infrastructure to empower YouTube

https://blog.youtube/inside-youtube/new-era-video-infrastructure/

by prostoalex • 5 years ago

70 30 5 years ago

blog.youtube

axiosgunnar• 5 years ago

Reimagining video infrastructure to empower YouTube

What level of corporate buzzword bingo is this on?

wmf2• 5 years ago

Webscale level.

singhkays• 5 years ago

So they ended up adopting MongoDB? :)

https://www.youtube.com/watch?v=b2F-DItXtZs

pimlottc• 5 years ago

I get that this a engineering article but wow, what a tone deaf title. Nobody outside of Google wants to “empower” YouTube, they have plenty of power already. If anyone needs empowering, it’s the users (aka actual people).

jchw2• 5 years ago

The good faith interpretation would probably be “YouTube [Creators]” which seems to be what was meant.

ffhhj• 5 years ago

Quantum supremacy level.

2OEH8eoCRo0• 5 years ago

Fits perfectly imo. Did you read that article?

1_player• 5 years ago

"Empower" is one of those stupidly meaningless words corporate drones like to use in their slides to impress upper management.

What happened to you, Google? You used to be cool.

handrous• 5 years ago

TL;DR they started using dedicated and purposed-designed video encoding/transcoding chips.

sheikheddy• 5 years ago

Any implications for end-users, or notable improvements from an engineering standpoint?

noir_lord• 5 years ago

Any implications for end-users

For end users competing against them yes, they and the other "We run the internet's biggest sites and are also a cloud provider" providers are leveraging dominance in one area to fund and build technology that you simply can't compete with.

From their point of view an entirely rational thing to do (and hardly new) but it does create an un-level playing field when you use money from one area to enter and dominate a market in another.

For anyone familiar this has all the hallmarks of the standard oil days when an oil company took over large swathes of the transport industry and charged differential rates based on local competition, selling at a loss where their was competition.

xnx• 5 years ago

Previously: https://news.ycombinator.com/item?id=27034627

xmprt• 5 years ago

Anecdotally, whatever YouTube has done in recently years has completely ruined a lot of 480p and 360p videos. I uploaded a native 360p video about 10 years ago and in the end card, I added small red text that was visible. And when I uploaded it at the time, it was visible on YouTube as well. Now that section just looks like a bunch of red blurry lines.

If you ever wondered why old videos look so bad, it's not just because you're used to 1080p and so 360p looks terrible in comparison. YouTube is actively reducing old video quality through whatever reencoding they're doing.

mdonahoe• 5 years ago

They probably lowered the max bitrate for 360p videos.

Wowfunhappy• 5 years ago

Which might make sense for variable bitrate content, but makes no sense for videos that were only ever 360p in the first place! Youtube basically made a lot of old videos unwatchable.

denimnerd42• 5 years ago

Which would destroy any 360p video so they should have known better than that. Transcoding is pretty bad stuff.

giantrobot• 5 years ago

If your original content wasn't one of YouTube's current formats it was likely reencoded. If the original was MPEG-4 part 2 or Sorensen Spark that a decade ago YouTube might just passed through, today it would be reencoded before delivery.

Their pipeline taking 1080p or 4K content to a 360p stream will produce decent quality. The encoding configuration for today's use also might be very different from the original. If their 360p profile used today ups the GOP length and drops the video bitrate the video content can definitely end up bit starved. Readable text in a 360p video would be difficult to keep readable reencoding without manual tweaking of the encoder settings.

I also wonder if old content with low viewership is just encoded on the fly with a hardware encoder and thrown into a big YouTube-wide LRU cache. Hardware encoders can have some really coarse grained settings and make low resolution/low bitrate content look pretty horrible.

Hubi• 5 years ago

I have noticed that as well. I recently stumbled across a movie clip in 240p and the thing that was most notable was that during a monologue, the actor's lips did not move at all. He turned his head a little, but his expression seemed to be frozen. After a very noticeable key frame it returned to normal shortly, but his face ended up getting stuck again. I don't recall ever seeing something like this on YouTube, not even in ultra-LQ clips from 2007.

ChrisArchitect• 5 years ago

bad re-invite, this is what this one is about/being discussed already:

Google Replaces Millions of Intel's CPUs with Its Own Homegrown Chips

https://news.ycombinator.com/item?id=27390398

ExtraE• 5 years ago

When did google buy the .youtube TLD? Anyone know how to view a list of corporate TLDs? (Yes I know they’re mostly owned by ICAN anymore and all therefore corporate in some sense).

wmf2• 5 years ago

Wow, there are 1502 TLDs. https://en.wikipedia.org/wiki/List_of_Internet_top-level_dom...

HWR_14• 5 years ago

.youtube isn't shocking. Google also owns .dev A lot of people used to use .dev as a reserved, localhost style dns lookup.

totony• 5 years ago

Fun fact, .dev forces https on modern browsers. Caused a bit of trouble when a colleague of mine tried setting up a new local test server.

dr_hooo• 5 years ago

Wow, this entire interview basically contained no information what so ever, except YouTube is awesome.

yellow_lead2• 5 years ago

Jeff: When I interview candidates for jobs here, I always mention how more than 500 hours of video content on average is uploaded every minute to the platform - that always resonates with them.

Jeff seems like pretentious upper management. Not sure about others, but when I interview at companies and they brag about their scale, it's really off-putting.

Mehdi2277• 5 years ago

I personally like scale as I work in ml and that helps in making datasets large. You can do a lot of ml on small datasets but life is much easier when you get to work with what feels like infinite data. It was nice having models train on trillionish data points at a prior company. The opposite end I worked at a startup that had a couple thousand data points and dataset size was quite annoying especially in not having enough variety.

yellow_lead2• 5 years ago

That's fair but it always comes off as bragging. Typically the candidate will realize something like YouTube is a big scale site before even applying.

nojito• 5 years ago

Scale problems are very fun to attack and solve.

Reimagining video infrastructure to empower YouTube

Related Stories