5 ms·
The gear isn't the problem: the dependency on a single data center is the problem. It requires a lot of software engineering effort to maintain a service that
by jrockway 14y ago
The gear isn't the problem: the dependency on a single data center is the problem. It requires a lot of software engineering effort to maintain a service that works when a data center suddenly goes away. To be fault-tolerant, Netflix has to do this engineering regardless of whether or not they own the servers. But if they use Amazon, they don't have to actually fix the servers when they break, freeing up engineering resources to fix the software.
(Why doesn't Amazon offer transparent replication? Because the price of replication is unbelievably high and most people can't believe how high it is. If you want to write to 12 replicas around the world, budget five seconds for your transaction to complete. Compare this to a MySQL database on an SSD that can do millions of writes on the same entity group in the same amount of time.
This doesn't even include the cost: 12x the storage cost, and 12x the bandwidth cost for data going from your frontend to all of the backends.)
- moe 14y agoYou're blowing this a little out of proportion. Replicating a few petabytes of static videos is not rocket science nor cost prohibitive for a company the size of Netflix. Nor is engineering a system that can withstand a datacenter outage. Especially one as trivial as Netflix which is largely read-only. Thousands of systems of higher complexity are engineered to that standard, many of them are much larger than netflix. You rarely hear of Google outages, or iTunes, or Youtube, or [insert six dozen other popular brands], do you? Yes it can happen to the best of them, but the EC2 outages are really piling up lately.
- pixl97 14y agoYou rarely hear of Google outages http://techcrunch.com/2012/12/10/gmail-experiences-a-widespread-outage-most-users-affected/ http://techcrunch.com/2012/12/10/gmail-experiences-a-widespr... ,or iTunes http://appleinsider.com/articles/12/11/19/itunes-match-down-as-apples-icloud-issues-continue http://appleinsider.com/articles/12/11/19/itunes-match-down-... ,or Youtube http://abcnews.go.com/blogs/technology/2012/10/youtube-goes-down-not-googles-best-day/ http://abcnews.go.com/blogs/technology/2012/10/youtube-goes-... Everyone goes down, but EC2 has been buggy as hell from what I can tell. And before everyone says Netflix should `just` move or add more bandwidth... http://www.nbcnews.com/technology/technolog/netflix-uses-32-7-percent-internet-bandwidth-119517 http://www.nbcnews.com/technology/technolog/netflix-uses-32-... They represent a huuuge amount of bandwidth usage!
- toomuchtodo 14y agoSo work with eyeball networks and move your gear to the edge; dumping it all in AWS and hoping for the best is failing (at best). Do all the auth/billing/recommendations/transcoding in AWS, and push all the video content out to boxes at the ISPs.
- mh- 14y agoThat bandwidth for delivering video is being streamed from a CDN outside AWS.
- res0nat0r 14y ago>Especially one as trivial as Netflix which is largely read-only. I stopped reading after this, since you obviously dont understand how complex their architecture really is.
- sbov 14y agoI'm never one to trivialize another company's problems so would hesitate to say Netflix's case is easy, but you cannot counter this statement by saying their architecture is complex. I can build a complex architecture for "Hello World" if I wanted to. Complex architecture != complex problem.
- moe 14y agoI work for a media streaming site (rank <30k on alexa) and I assure you the architecture doesn't change much when you scale it up to rank 99 (netflix). "Trivial" may sound harsh but in terms of large scale systems it really doesn't get much simpler than serving static files. Obviously their auxiliary services (billing, content ingestion, social etc.) are nowhere near trivial. But there's no reason for the core services that are required to get the catalog and video to most[1] devices to not be damn near 100% available. I'm not talking down netflix for having an outage anyway. Shit happens and afaik their overall track record is not bad at all. I only replied to jrockway's claim that multi-datacenter redundancy would require inconceivable engineering effort or amounts of money - neither is true. [1] Excluding those that need realtime transcoding, absurd DRM schemes or similar.
- jrockway 14y agoYou rarely hear of Google outages, or iTunes, or Youtube, or [insert six dozen other popular brands], do you? Yes it can happen to the best of them, but the EC2 outages are really piling up lately. To me, this makes perfect sense. Google products are almost always designed to tolerate data center failures, and so you don't hear about data center failures affecting Google products. On the other hand, Joe's Random EC2 app is not designed for the same high availability, and so it's down whenever Amazon is. (This, incidentally, is a reasonable trade-off for most people. A few hours of downtime a year is often much cheaper than paying engineers to ensure that those few hours are five minutes or less.) So the flaw here is not Amazon. The flaw is relying on one cluster of computers to handle all your computing needs. Amazon could make their product high-replication by default, but it would be so slow and expensive that nobody would use it. So the task is kicked to the application developers instead of the platform developers, and you should consider holding them responsible for the downtime. (Incidentally, you can compare EC2 to AppEngine here. How often is your favorite AppEngine app down? Less than your favorite EC2 app, probably, because AppEngine pretty much forces the high-replication datastore, even though the semantics are strange for developers used to traditional single-homed or master/slave architectures. And if you read the AppEngine discussion groups, you'll see that users raise most of the concerns that I describe here: "Why is AppEngine more expensive than EC2?" "Why is AppEgnine slower than EC2?", and so on.) Now, why does Netflix's video serving go down when an EC2 cell does? I have absolutely no idea. I do know that they do more than just stream bytes, though; they have to authenticate users, track what they're watching, and apply DRM to the streams. So there's that. If you want fault-tolerant video distribution, look no further than the Pirate Bay. Although their servers are frequently seized, the video bits keep flowing. That's by design, not by accident.
- toomuchtodo 14y agoNetflix: Spitting out 1 of 120 different encoded version of a video to a client device after its been authenticated. So, you've got an authentication layer, which can pass a token to the browser or physical hardware client access device. This should be very lightweight, no? Even with millions of users, this database shouldn't be enormous. And then you've got your encoded libraries. This should be data being spit off platers on CDN networks. CDNs aren't hard anymore. Cloudfront is cheap, or you can go with Akamai, Limelight, Level3, etc. Properly built, you should never NOT be able to serve from somewhere in the CDN, even if the serving location isn't optimal. Yes, this is HEAVILY simplified. I've left out recommendations, their encoding process/infrastructure, etc. This is not atomic data that is being replicated in realtime; these are videos that are lazily encoded, stored, and then served on demand. It should not be this hard for this use case.
- 1qaz2wsx3edc 14y agoJust poking in on this, does anyone have a graph of outages over time? I'm wondering if it has been increasing or decreasing; and therefore if amazon is getting better at managing it's demand. Bonus points if it compares the size of the customer base over time in comparison or relation.
- xugle 14y agojrockway, we know that google does things a certain way that blows up the costs of simple storage. It's not necessary to make 12 copies worldwide to get bare minimum redundancy required for this case. The data that needs to be replicated (video files) and account info is not realtime transactions. The video files can be transferred by fedexing a bunch or hard disks on a weekly basis (take a reasonable guess at Netflix catalog growth and compute the per week growth of data... it's indeed trivial).