|
zhar posted:isn't that just a function of financials? You can see all the datacenters google have and are building here, looks like they cost around $600m each to build and doubtless have some fairly large operating costs but they are used for all google services not just youtube. according to their latest earning report youtube ads brought in $8.633b last quarter alone and it doesn't look like that includes the premium stuff/superchat and so on, doesn't seem that unsustainable to me. A lot of these DC buildouts are for their cloud services as well, so to support more cloud customers.
|
# ? Feb 3, 2022 20:19 |
|
|
# ? Jun 8, 2024 07:58 |
|
Mr Interweb posted:okay well that would explain the financial part, i suppose. but what about building infrastructure (server farms, bandwidth etc.) ? Eh, there isn't really a "the server youtube runs on" or anything. I'm sure youtube is just a bunch of docker containers or something similar kinda running on an as needed basis in endless random server racks serving a bunch of files from a bunch of weird collocation endpoints. Which isn't to say that it is totally abstract and only exists mystically in "the cloud" but more like, I'm sure it's nothing special and just exists as some allocation on server racks in general more or less like all modern web stuff does. Like I'm sure most of the bare metal it runs on is just running various hypervisors and the hardware running youtube now will be running some business website in 10 minutes and a porn site 10 minutes after that and an MMO login server after that and so on and so on. Cloud stuff is mostly having tons of interchangeable hardware and running whatever whenever as it scales up and down. I'm sure youtube is just thrown in that sort of general population on google's cloud servers. I doubt they build much "YOUTUBE SERVER, HERE IS THE HARD DRIVE YOUTUBE IS SAVED ON, IT PLUGS IN RIGHT HERE" and more "On US east we'd like this function to scale from 800 to 800,000 servers, this function is serverless, we want to cache these videos here and those videos can go into cheaper deep storage"
|
# ? Feb 3, 2022 20:35 |
|
Arsenic Lupin posted:Somebody posted a Tweet the other day of her dishwasher demanding a firmware update on its screen, then showing a progress bar. Meanwhile, it wouldn't actually wash the dishes. Well what good is washing your dishes without a wifi connection?
|
# ? Feb 3, 2022 22:47 |
|
For another reference point, Google total data center expenses is (educated guess) around $50-60B per year. This would include hardware, bandwidth, energy, and staff for all core services, including YouTube. So the question is how much of that is YouTube-related vs. search, Maps and the other Google services. If YouTube is more than half then it's loss-making.
|
# ? Feb 3, 2022 23:02 |
|
https://twitter.com/justinlabel/status/1489317547686125570?s=20&t=SiZh2GDuDTrjHQvmBsOU4w https://twitter.com/justinlabel/status/1489317573556678657?s=20&t=SiZh2GDuDTrjHQvmBsOU4w
|
# ? Feb 3, 2022 23:17 |
|
Owlofcreamcheese posted:Eh, there isn't really a "the server youtube runs on" or anything. I'm sure youtube is just a bunch of docker containers or something similar kinda running on an as needed basis in endless random server racks serving a bunch of files from a bunch of weird collocation endpoints. so this isn't wholly correct. while google's storage teams are probably not saying "here is the specific youtube storage server" they are certainly aware of youtube's capacity projections and it likely defines the scale at which underlying services need to operate, similarly for compute. you also need a higher degree of coordination around "running whatever whenever" to ensure that you have enough spare capacity to serve peak requests for the things that matter, and youtube matters. one of the biggest benefits of the cloud is that what is being built to support the scale of youtube is also accessible to everything else that isn't unfathomably as large, and so they get the benefit of fitting into the noise and have the illusion of infinite scalability.
|
# ? Feb 3, 2022 23:44 |
|
FamDav posted:so this isn't wholly correct. while google's storage teams are probably not saying "here is the specific youtube storage server" they are certainly aware of youtube's capacity projections and it likely defines the scale at which underlying services need to operate, similarly for compute. you also need a higher degree of coordination around "running whatever whenever" to ensure that you have enough spare capacity to serve peak requests for the things that matter, and youtube matters. I mean it is unlikely much dedicated youtube hardware even exists. I'm sure it's a modern cloud application. I doubt they just run 'youtube.exe" on some windows server somewhere in a room, I'm sure they have a billion kubernetes containers and a bunch of whatever google calls lamda functions. I am sure they bill themself for the resources they use in some form but I'm sure it just exists the way all cloud stuff does. Abstracted from specific hardware. The data center upgrades to run it in that it is generally upgrading to generally run things. No one is building tons of youtube hardware compared to just "cloud hardware" that will run many things including youtube.
|
# ? Feb 4, 2022 00:16 |
|
Owlofcreamcheese posted:I mean it is unlikely much dedicated youtube hardware even exists. I'm sure it's a modern cloud application. I doubt they just run 'youtube.exe" on some windows server somewhere in a room, I'm sure they have a billion kubernetes containers and a bunch of whatever google calls lamda functions. I am sure they bill themself for the resources they use in some form but I'm sure it just exists the way all cloud stuff does. Abstracted from specific hardware. The data center upgrades to run it in that it is generally upgrading to generally run things. No one is building tons of youtube hardware compared to just "cloud hardware" that will run many things including youtube. Eh, no I suspect a large portion of Youtube is dedicated hardware largely because the amount of rendering and hosting required. Ironically, not a lot of heavy workloads are cloud native, even cloud vendoers like AWS and Google have openly said most of their workloads remains as VMs or even bare metal. https://www.quora.com/Where-is-the-server-of-YouTube Specifically in Youtubes case there's caching servers cohosted in a lot of ISPs datacenters and routing centers. Its far from some sort of serverless/containers cloud native app. From 2008: http://highscalability.com/youtube-architecture Some portions of it may be cloud native or containers, but the core stuff is no doubt more traditional. CommieGIR fucked around with this message at 00:32 on Feb 4, 2022 |
# ? Feb 4, 2022 00:20 |
|
CommieGIR posted:Ironically, not a lot of heavy workloads are cloud native, even cloud vendoers like AWS and Google have openly said most of their workloads remains as VMs or even bare metal. They don't like to say that too loudly because they want YOU (your company) to use their cloud services for everything even when it doesn't make any sense. Too many companies have fallen for this trap. Including the last few I've worked for.
|
# ? Feb 4, 2022 00:32 |
|
Motronic posted:They don't like to say that too loudly because they want YOU (your company) to use their cloud services for everything even when it doesn't make any sense. "You can lift and shift to the cloud, it'll be cheap!" Nope.
|
# ? Feb 4, 2022 00:35 |
|
CommieGIR posted:"You can lift and shift to the cloud, it'll be cheap!" yeah it's cheap if you use 100% of the cloud poo poo you prepaid for but you're also 100% likely to not actually use all of it so you overpaid
|
# ? Feb 4, 2022 02:13 |
|
CommieGIR posted:"You can lift and shift to the cloud, it'll be cheap!"
|
# ? Feb 4, 2022 02:25 |
|
cat botherer posted:It's really easy to underestimate how many/much services you wind up needing, and thus the cost and time. It's wild how many resources go towards dealing with the massive deficiencies in functionality, but also deciphering terrible/wrong documentation. About half of my time is spent dealing with this kind of bullshit instead of actual problems. Why yes, I'd like to learn an entire new language to configure the thing that was supposed to make my life easier. Yup. The reality is companies view cloud as some way to eliminate legacy ops style IT. Except you can't. You still need Ops and Systems Engineers/Network guys to do cloud. Its just as bad, even worse sometimes because most of the cloud stuff has taken fairly simple architecture issues and compartmentalized it to bill you more for each and every component, I/O, gb, connection, etc.
|
# ? Feb 4, 2022 02:31 |
|
knox_harrington posted:I keep getting emails telling me to update the firmware on my bike jacket I can't tell if this is a joke or not e: apparently this is a thing: https://www.bksleather.co.uk/tech-air/tech-air-race-vest Mr Chips fucked around with this message at 02:39 on Feb 4, 2022 |
# ? Feb 4, 2022 02:36 |
|
Mr Chips posted:I can't tell if this is a joke or not Gotta update your firmware and renew your subscription
|
# ? Feb 4, 2022 02:41 |
|
Don't forget the $400 safety vest that won't activate if you don't also pay a monthly subscription: https://www.msn.com/en-us/money/news/this-motorcycle-airbag-vest-would-stop-working-if-you-dont-pay-a-subscription/ar-BB1gtSxh Turning EVERYTHING into a subscription service is the biggest tech scam. efb
|
# ? Feb 4, 2022 02:45 |
|
cat botherer posted:It's really easy to underestimate how many/much services you wind up needing, and thus the cost and time. It's wild how many resources go towards dealing with the massive deficiencies in functionality, but also deciphering terrible/wrong documentation. About half of my time is spent dealing with this kind of bullshit instead of actual problems. Why yes, I'd like to learn an entire new language to configure the thing that was supposed to make my life easier. It's so much worse, as you know. "I want to put one server in the cloud". Well, ackshually you need to run that across 3 availability zones in at least two regions which means you need to rearchitect everything. Also, since you now have 6 boxes you can buy this HA/proxy product from us to make this actually work again. Whoops.....what happened to my cheap cloud hosting? There is absolutely a place and a time for it. It's super useful for spike loads and many other things. But "everything, all the time" is just stupid.
|
# ? Feb 4, 2022 03:05 |
|
Tuxedo Gin posted:Don't forget the $400 safety vest that won't activate if you don't also pay a monthly subscription: Scam as a Service
|
# ? Feb 4, 2022 04:23 |
|
so to summarize: there is no worry on youtube/google's part that they'll be hardware constrained any time soon?
|
# ? Feb 4, 2022 04:44 |
|
Motronic posted:It's so much worse, as you know. "I want to put one server in the cloud". Well, ackshually you need to run that across 3 availability zones in at least two regions which means you need to rearchitect everything. Also, since you now have 6 boxes you can buy this HA/proxy product from us to make this actually work again. I mean, they don’t force you to do that stuff, you do that stuff because it’s advantageous to do that when running a large site and is a great advantage over just running from a server on your desk.
|
# ? Feb 4, 2022 04:48 |
|
Tuxedo Gin posted:Don't forget the $400 safety vest that won't activate if you don't also pay a monthly subscription: What happens if their auth servers go down or your phone is out of batteries?
|
# ? Feb 4, 2022 06:29 |
|
I'm pretty sure YouTube just runs on borg (the internal inspiration for kubernetes) like 90% of everything at Google. The other 10% is Android and a few things that actually do run on GCP.
|
# ? Feb 4, 2022 06:46 |
|
CommieGIR posted:Eh, no I suspect a large portion of Youtube is dedicated hardware largely because the amount of rendering and hosting required. Ironically YT is exactly the kind of service that could use custom hardware for parts of it, because it has both huge scale and some fairly fixed workloads. e.g. I would be very surprised if there wasn't at least an exploratory doc on issues/benefits on shifting most of the video transcoding onto FPGAs.
|
# ? Feb 4, 2022 08:52 |
|
Mr Interweb posted:so to summarize: there is no worry on youtube/google's part that they'll be hardware constrained any time soon? FWIW early last year MS had a capacity crunch for Azure, because Covid both delayed bringing up moar capacity and increased cloud capacity demand due to wfh. No idea how it evolved afterwards, I quit after April.
|
# ? Feb 4, 2022 08:54 |
|
holy poo poo, my assumption about google's revenue streams were waaaaaay off. i thought YT was its biggest money maker. turns out that's hilariously wrong:quote:YouTube is the smallest of Google’s three main advertising revenue sources, accounting for nearly $20 billion in revenue in 2020 — about 13% of Google’s total ad revenues. seems that the thing that brings in the big bucks is what started it all: quote:Search is Google’s most lucrative unit. In 2020, the company generated $104 billion in “search and other” revenues, making up 71% of Google’s ad revenue and 57% of Alphabet’s total revenue. i mean, thinking about it, i suppose it makes sense, but still. didn't think YT was that low
|
# ? Feb 4, 2022 09:22 |
|
Motronic posted:It's so much worse, as you know. "I want to put one server in the cloud". Well, ackshually you need to run that across 3 availability zones in at least two regions which means you need to rearchitect everything. Also, since you now have 6 boxes you can buy this HA/proxy product from us to make this actually work again.
|
# ? Feb 4, 2022 09:25 |
|
Owlofcreamcheese posted:Eh, there isn't really a "the server youtube runs on" or anything. I'm sure youtube is just a bunch of docker containers or something similar kinda running on an as needed basis in endless random server racks serving a bunch of files from a bunch of weird collocation endpoints. It's not that simple. Youtube handles livestreams and the automatic archival thereof and live video delivery has much stricter timing requirements than regular content. A stream with 100k CCV and rewind on has to be transported to 100k people but some of those are actually watching the vod which needs to be generated and delivered on the spot and actually it's not one stream but rather 6 to 9 different ones 'cause the source's being transcoded into all available resolutions and the vods too. Which reminds me I can actually tell they've got lag in disseminating vods across their CDN because a newly made vod loads much slower (like, "buffers every 5s" slow) than one that's about half a day old. Interestingly livestreams work fine which suggest they use a different (and quicker) channel. Motronic posted:It's so much worse, as you know. "I want to put one server in the cloud". Well, ackshually you need to run that across 3 availability zones in at least two regions which means you need to rearchitect everything. Also, since you now have 6 boxes you can buy this HA/proxy product from us to make this actually work again. How is having one cloud instance different from hosting in your one local machine/datacenter? Everything you mention is a service level upgrade that clearly you weren't planning for 'cause that's why you only had one server to begin with. Kyte fucked around with this message at 09:41 on Feb 4, 2022 |
# ? Feb 4, 2022 09:34 |
|
Arsenic Lupin posted:Somebody posted a Tweet the other day of her dishwasher demanding a firmware update on its screen, then showing a progress bar. Meanwhile, it wouldn't actually wash the dishes. I think that's the big thing pissing people off the most about internet of things stuff. They never seem to update on their own, but wait until someone goes to use them. Then lock themselves out until the update is done. Which is two different bits of bad design, right there. Like that guy last year who had his IoT BBQ decide to do a massive update on some national beer drinking holiday.
|
# ? Feb 4, 2022 12:36 |
|
Whenever I run the numbers (as pure due dilligence) of putting our storage into "the cloud", I arrive at numbers where "the cloud" is more expensive per quarter than our entire budget over 3-5 years (depending on grant). Not counting compute, transfer (!!!), etc. And then it gets fun when I consider typical usage scenarios, like a masters student doing some work with, say, a 50TB ground penetrating radar data set. They get access to a workstation, read permissions to the dataset, and they can get to work right there. In any cloud scenario, I don't even know where to start. An entire lecture series on how to get poo poo to run up there, plus a horrific bill for all the compute they end up wasting while loving around?
|
# ? Feb 4, 2022 14:13 |
|
Megillah Gorilla posted:I think that's the big thing pissing people off the most about internet of things stuff. Like if someone was smart in designing they would have it update AFTER it is used and just inform users, when you turn this device off this time, it is going to update instead and THEN turn off the power completely
|
# ? Feb 4, 2022 14:31 |
|
Smart people don't make IoT devices.
|
# ? Feb 4, 2022 14:39 |
|
Kyte posted:How is having one cloud instance different from hosting in your one local machine/datacenter? Everything you mention is a service level upgrade that clearly you weren't planning for 'cause that's why you only had one server to begin with. Most local datacenters don't charge you per GB transferred, per I/O, per session. And most cloud providers want to lock you in so putting data in is "free" but taking it out you get charged for.
|
# ? Feb 4, 2022 14:39 |
|
Owlofcreamcheese posted:I mean, they don’t force you to do that stuff, you do that stuff because it’s advantageous to do that when running a large site and is a great advantage over just running from a server on your desk. Another great uninformed take. It's literally baked into every major cloud hosting provider's best practices and if you don't do that your uptime will be worse than the server sitting under your desk that the cleaning staff unplugs every weekend to use the outlet for their vacuum. Sagacity posted:So you're saying that it's weird that having a massively more resilient solution somehow costs more than than just hosting it in a single rack in a single datacenter? I'm saying that it's not equivalent yet it is being advertised/sold as a cost savings measure which is absolutely disingenuous. A single server in a rack at a datacenter is much more reliable than a single instance at one of the major cloud service providers. Anyone who has managed large scale deployments of both knows this very well. And that is something solidly in my experience and skillset. Kyte posted:How is having one cloud instance different from hosting in your one local machine/datacenter? Everything you mention is a service level upgrade that clearly you weren't planning for 'cause that's why you only had one server to begin with. The former is much, much less reliable. You have no choice but to do all of the massively distributed things just for reasonably resiliency. And at any sort of larger-than-startup-scale it's still cheaper to do it on your own hardware at multiple datacenters. It's also more flexible. Also there's aren't huge variable costs due to things like this: CommieGIR posted:most cloud providers want to lock you in so putting data in is "free" but taking it out you get charged for. .....and "oh, you need more IOPS/PPS/whatever? "Well just upgrade to the next instance size. Yeah, I know you don't need all those CPUs, but that's the only type of instance where we offer this much IO so you have to pay for it anyway. Yeah, we know your Dell box at the datacenter had much lower specs than this and could handle more of your workload, but it's not the same in the cloud." Motronic fucked around with this message at 17:17 on Feb 4, 2022 |
# ? Feb 4, 2022 17:15 |
|
Motronic posted:A single server in a rack at a datacenter is much more reliable than a single instance at one of the major cloud service providers. Anyone who has managed large scale deployments of both knows this very well. And that is something solidly in my experience and skillset. I can only speak from my own experience but both the companies Ive worked at that migrated from racks to cloud (both rackspace -> Azure) the procedure was certainly not as easy as its advertised and there were certainly hidden costs but the cost WAS lower and the uptime WAS higher. Maybe this is a result of rackspace being dreadful, I'm not sure (well they are dreadful) but I'm not sure i agree with the idea that cloud is always worse.
|
# ? Feb 4, 2022 17:27 |
|
IMO IT guy debates are the real tech nightmare.
|
# ? Feb 4, 2022 17:29 |
|
Mega Comrade posted:I can only speak from my own experience but both the companies Ive worked at that migrated from racks to cloud (both rackspace -> Azure) the procedure was certainly not as easy as its advertised and there were certainly hidden costs but the cost WAS lower and the uptime WAS higher. Its probably their architecture was just that bad when it was on-prem.
|
# ? Feb 4, 2022 17:43 |
|
withak posted:IMO IT guy debates are the real tech nightmare. Eh, the start of this IT rabbit hole began with folks who absolutely had no understanding of modern cloud tech .(no offense, its evolved FAST). It's good to hear some shop talked every now and again. Every deployment style has it's advantages, based on budget, location, user base, available expertise, application complexity, and risk factor.
|
# ? Feb 4, 2022 17:46 |
|
Motronic posted:Another great uninformed take. Yeah, because it's best practices to do that. You can not do it locally cheaper because you are doing something that sucks. You want things like cross region replication. The fact your local data center is cheaper because it doesn't do that is a bad thing. It's like how you can get cheaper data storage by keeping it on a hard disk from walmart with no backup. Just being worse makes things cheaper.
|
# ? Feb 4, 2022 17:54 |
|
AtomikKrab posted:Like if someone was smart in designing they would have it update AFTER it is used and just inform users, when you turn this device off this time, it is going to update instead and THEN turn off the power completely
|
# ? Feb 4, 2022 17:55 |
|
|
# ? Jun 8, 2024 07:58 |
|
CommieGIR posted:Its probably their architecture was just that bad when it was on-prem. This is very much a possibility.
|
# ? Feb 4, 2022 17:58 |