Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by ajross on Monday March 26, 2012 @03:26PM
(#3758241)
Most of this is relatively straightforward and unsurprising. But the one part that grabbed me is about "jittering". They insert random delays into timed events (the example given is cache expiration) to prevent a thundering herd problem when all the parts of the distributed system see the event at the same time (and for popular content, presumably repopulate the cache from the backend simulteously).
This is simple enough when described, but is not a technique I've seen applied much in practice or discussed in the community. I'm wondering if it's something that gets reinvented for all the projects that need it or if it's secret sauce known only in youtube. Regardless, I thought it was pretty insightful.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by cpeterso on Monday March 26, 2012 @07:46PM
(#3759196)
The Adblock Plus blog details the thundering herd problems they faced. Their ad blocking lists checked for updates every 5 days. Eventually, many users' update schedules would converge on Mondays because office computers did not run over the weekend. Updates that had been scheduled for Saturday or Sunday would spill over to Monday.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by sk5t on Monday March 26, 2012 @08:36PM
(#3759367)
Windows does much the same thing w.r.t. policy refresh (which sucks down files from domain controllers) and update of "occasionally updated" metadata like last logon timestamp.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by toddh on Monday March 26, 2012 @05:28PM
(#3758752)
As an interesting counter point, Jeff Dean at Google says they prefer to have a known hit every once in a while, so they will have all the cron jobs go off at the same time versus introducing jitter. It perhaps has to do with their emphasis on reducing variability to control the size of the long tail distribution.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by jamii on Monday March 26, 2012 @04:57PM
(#3758609)
A nice extension of this is to use exponentially distributed delays. Then reoccurring events form poisson processes which are really easy to reason about when composed eg if you have N servers firing off events at the same exponentially distributed rate the times are distributed the same as if you had one server firing events at N x the rate and distributing jobs uniformly at random to the other servers.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by cpeterso on Monday March 26, 2012 @07:53PM
(#3759222)
A counterexample is that the Linux kernel tries to schedule timer events for the same deadline time. That allows the processor to sleep longer because the kernel doesn't need to wake up as often just to handle 1 or 2 timer events.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by cbsmith on Monday March 26, 2012 @04:52PM
(#3758584)
It's kind of a weird way to describe it. The better way is that you want your caching to be probabilistic rather than deterministic. In general, you want to avoid anything that would get the nodes in a distributed system to harmonize.
The other way to solve the problem though would be to handle that circumstance cleanly. There are ways to resolve a thundering herd without creating a scalability problem.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by davnola on Monday March 26, 2012 @03:46PM
(#3758326)
I've heard it called "splay" before e.g. `chef-client` will take an interval option and a splay option. Splay is a random amount of time added to the interval to prevent thundering herds.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by peterwwillis on Monday March 26, 2012 @04:19PM
(#3758455)
In the past i've seen jitter used in all sorts of applications, from cron jobs to configuration management to memcache key expiration. Any time you have a crapload of nodes that all need to do an operation at a specific time you rely on jitter to keep resources from bottlenecking. Probably used anywhere the systems or network has a miniscule amount of resource headroom (like cluster nodes that run at 93% utilization)
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by minimax on Monday March 26, 2012 @04:02PM
(#3758382)
I hadn't thought about it in the general case, but I do frequently find myself adding a sleep for PID mod some appropriate constant to the beginning of big distributed batch jobs in order to keep shared resources (NFS, database, etc) from getting hammered all at once.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by blhack on Monday March 26, 2012 @01:59PM
(#3757841)
What's this? Python? Apache? MySQL? But I thought you had to be running beta-release key value stores, esoteric web servers, and experimental programming languages if you wanted to scale!
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by zmj on Monday March 26, 2012 @02:23PM
(#3757956)
Got your experimental programming language:
> Vitess - a new project released by YouTube, written in Go, it’s a frontend to MySQL. It does a lot of optimization on the fly, it rewrites queries and acts as a proxy. Currently it serves every YouTube database request. It’s RPC based.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by ajross on Monday March 26, 2012 @03:15PM
(#3758204)
Remote Procedure Call is a design paradigm for synchronous call-and-response network communication. The Sun RPC protocol is an actual technology defined in RFC1057: http://www.ietf.org/rfc/rfc1057.txt
It's not insane, though not terribly relevant in the modern world. The only common technology still using it is NFS.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by m0th87 on Monday March 26, 2012 @03:39PM
(#3758294)
If you look at the original RPC work by Bruce Nelson [1], it's pretty clear that there's no strict definition of it. I think most would argue that SOAP would be included, which is still pretty common.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by tav on Monday March 26, 2012 @03:41PM
(#3758299)
When Youtube refers to Vitess as being RPC-based, they are not referring to Sun RPC (callrpc), but rather to the generic design pattern of exposing service calls over the network. In particular, Vitess makes services callable [-] using either BSON or JSON serialisation over HTTP CONNECT calls.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by j2labs on Monday March 26, 2012 @02:52PM
(#3758092)
I know the rage is async & nonblocking, but putting Python behind Apache is a good way to get predictable performance without concerns for calls that block, since they're mitigated via threads.
Also, it helps to have money since this approach requires more boxes. But as I said, it's very reliable.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by peterwwillis on Monday March 26, 2012 @04:42PM
(#3758547)
Fast, cheap, scalable: pick two.
The fastest, easiest, arguably most reliable way to scale is throwing money at it. And apparently it's easier to hire good people who know 20-year old tech rather than 3-year old.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by mattdeboard on Monday March 26, 2012 @02:08PM
(#3757884)
The first 10 minutes are about monetization from one of the Youtube dev advocates. Skip to 9:45 to get to the "good stuff".
As an aside, this fellow is probably one of the best presenters I've seen from the pycon videos for this year. Confident, smooth, not reading from a computer screen or sheet of paper, clearly smart and in firm command of the subject matter.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by supersillyus on Monday March 26, 2012 @08:00PM
(#3759252)
I believe it could be faster still. Their Itoa cache is a map[int]string, where it could be a []string.
Also, I suspect that a few more primitive type special cases in the first type switch in EncodeField could go a long way.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by pothibo on Monday March 26, 2012 @02:07PM
(#3757875)
I fully agree with Youtube faking data. However, I reckon they are faking a bit too much. Many times I would see 2000 likes and the video having 1700 views (Viral videos that is).
I knew the view counter wasn't propagated but the likes were and I was like: "Damn this is Youtube, kinda disappointing..."
I guess if both were propagated at the same time I wouldn't mind.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by cbsmith on Monday March 26, 2012 @02:33PM
(#3758006)
I honestly don't understand why they simply don't use out of sync data. You could have nodes periodically send aggregates of likes & views, and then add those in to the total ever N heartbeats. Why bother fudging the in-between.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by moe on Monday March 26, 2012 @08:25PM
(#3759329)
They're probably propagating the likes and views independently. Which still doesn't explain why they allow counter-intuitive gaps like that instead of fixing them up on the client-side in javascript.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by modeless on Monday March 26, 2012 @10:25PM
(#3759680)
Yeah, the view counter is a clear case where they've gone too far. People care about view counts and it seems like every other video I watch has people complaining in the comments about the view count being obviously wrong.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by batiudrami on Monday March 26, 2012 @11:51PM
(#3759925)
The only time I notice this is when the view counts freeze at 300-ish for a few hours. That's because they're checking if the views are legit or from a bot, IIRC.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by xentronium on Monday March 26, 2012 @04:59PM
(#3758617)
"Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it." -- Brian Kernighan
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by InclinedPlane on Monday March 26, 2012 @04:44PM
(#3758559)
Complicated code that isn't easy to maintain isn't smart, it's just complicated. One thing people screw up a lot is implementing some design pattern in a way that ends up shotgunning configuration or business logic over a wide area instead of keeping it in one location and using DRY.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by apu on Monday March 26, 2012 @11:41PM
(#3759912)
At 11:28 he says, "at last count, there was over a million lines of python running this thing"
Having never worked with code-bases larger than ~50kloc, I have a lot of trouble understanding what 1 million lines of code is needed for, especially considering that python is such a high-level language.
Does anyone have any idea why there would be this much code?
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by apu on Tuesday March 27, 2012 @04:14PM
(#3763381)
Oh, it wasn't meant to be snarky or a slam -- I'm honestly just curious what kinds of things require so much code? E.g., is it one or two things that dominate usually in codebases this size, or is it just a LOT of components, each of which is tens of thousands of lines long? Do these kind of counts usually include auto-generated code?
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by cwbrandsma on Monday March 26, 2012 @01:47PM
(#3757783)
I love the part on faking data. I take the viewpoint that only software testers care that the comment count is exactly correct in the majority of system. Users don't care.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by RandallBrown on Monday March 26, 2012 @02:07PM
(#3757870)
I always thought it was a bug that view counts on YouTube varied depending on what page you saw them on. It can be different on the video view page, the channel page, and in your dashboard.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by leeskye on Monday March 26, 2012 @07:52PM
(#3759217)
The publisher who uploaded the video might care. Especially so, if they are selling to advertisers. Comments, ratings, and favorites are three variables we calculate to reflect how engaging a video is.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by ckg on Tuesday March 27, 2012 @12:44PM
(#3762460)
If they feel strongly enough about it to leave a comment then I think it's safe to say it does hinder their enjoyment - in the same way that obviously broken things distract and displeasure in any medium.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by mbell on Monday March 26, 2012 @09:10PM
(#3759456)
My biggest gripe with youtube: why are comment almost always repeated? Yea i realize that most you tube comments are relatively worthless but I do tend to speed through them to get a feel for what the response is to a particular video. Inevitably I get through 20 comments and then the same 20 are repeated over again, often they are repeated several times. Perhaps they are trying to give the illusion of lots of comments or assuming the comments don't matter. Personally I find it extremely annoying, I'd rather them block to load more than repeat.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by jcromartie on Monday March 26, 2012 @10:00PM
(#3759594)
What about this:
> The number of videos has gone up 9 orders of magnitude and the number of developers has only gone up two orders of magnitude.
2 orders of magnitude means at the very least, going from 9 to 100 developers, which is a huge increase, but it could mean way more. I wonder how big the team really is, and what the changing team dynamics are like on that scale at that pace.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by irahul on Tuesday March 27, 2012 @03:10AM
(#3760419)
I am not. From what I have seen of YouTube, comments are vile, and mostly there are two strangers posting pointless arguments about something equally pointless.
For me, YouTube is good for watching videos. If I want to discuss it, I post it on FB.
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by tbsdy on Tuesday March 27, 2012 @02:06AM
(#3760266)
Uh, cheating?
"Cheating - Know How to Fake Data
Awesome technique. The fastest function call is the one that doesn’t happen. When you have a monotonically increasing counter, like movie view counts or profile view counts, you could do a transaction every update. Or you could do a transaction every once in awhile and update by a random amount and as long as it changes from odd to even people would probably believe it’s real. Know how to fake data."
So all those people who buy views are kinda screwed now :-) I suspect this is a bad example. I HOPE this is a bad example, if only for the KONY2012 campaign :P
Re:7 Years Of YouTube Scalability Lessons In 30 Minutes
by henrikschroder on Tuesday March 27, 2012 @03:43AM
(#3760499)
No no, the correct amount of views will be recorded for a specific video, it's just that each webserver doesn't know the exact number all the time. You make each webserver fetch the correct value perhaps every hour, and fake it inbetween. You'll get an ok approximation, users can't tell the difference, and you don't have to fetch the actual number every single pageview.
This is simple enough when described, but is not a technique I've seen applied much in practice or discussed in the community. I'm wondering if it's something that gets reinvented for all the projects that need it or if it's secret sauce known only in youtube. Regardless, I thought it was pretty insightful.
https://adblockplus.org/blog/downloading-a-file-regularly-ho...
http://www.stdlib.net/~colmmacc/2009/11/26/period-pain-3/
(http://www.stdlib.net/~colmmacc/2009/09/27/period-pain-part-... http://www.stdlib.net/~colmmacc/2009/09/14/period-pain/ for the background material).
The reasons primes are optimal should be obvious :)
http://en.wikipedia.org/wiki/Poisson_process
The other way to solve the problem though would be to handle that circumstance cleanly. There are ways to resolve a thundering herd without creating a scalability problem.
http://ee.lbl.gov/papers/sync_94.pdf
http://files.multicastdns.org/draft-cheshire-dnsext-multicas...
Search for "random". Cool spec written by Stuart Cheshire of Apple.
/s
> Vitess - a new project released by YouTube, written in Go, it’s a frontend to MySQL. It does a lot of optimization on the fly, it rewrites queries and acts as a proxy. Currently it serves every YouTube database request. It’s RPC based.
It's not insane, though not terribly relevant in the modern world. The only common technology still using it is NFS.
1: http://nd.edu/~dthain/courses/cse598z/fall2004/papers/birrel...
[-] http://code.google.com/p/vitess/source/browse/go/rpcwrap/
Also, it helps to have money since this approach requires more boxes. But as I said, it's very reliable.
The fastest, easiest, arguably most reliable way to scale is throwing money at it. And apparently it's easier to hire good people who know 20-year old tech rather than 3-year old.
As an aside, this fellow is probably one of the best presenters I've seen from the pycon videos for this year. Confident, smooth, not reading from a computer screen or sheet of paper, clearly smart and in firm command of the subject matter.
I'd love to see more talks from him.
Curious to hear more about that one. If true, I hope they open source it, because that could potentially make MongoDB a lot faster for everyone.
EDIT: It's apparently in their vitess code. Relevant code: http://code.google.com/p/vitess/source/browse/#hg%2Fgo%2Fbso...
I knew the view counter wasn't propagated but the likes were and I was like: "Damn this is Youtube, kinda disappointing..."
I guess if both were propagated at the same time I wouldn't mind.
This has to go down in history as one of the best pivot decisions ever made.
The more magical the code is the harder is to figure out how it works."
A nice formulation of the kind of advice I keep reading here in HN.
Having never worked with code-bases larger than ~50kloc, I have a lot of trouble understanding what 1 million lines of code is needed for, especially considering that python is such a high-level language.
Does anyone have any idea why there would be this much code?
It's the world's 3rd biggest website with hundreds of billions of views and dozens of millions of users, so maybe that's why.
I think it's a credit to Python that a website that does that and has grown in a fairly haphazard fashion only has about 1000k SLOCs.
Turns out they're just making that shit up.
Since this was just a field in a database, it involved some simple update code.
The results? More people are interested in anything that they think are popular (or they are curious as to why so many people viewed it).
My client got more actual hits overall on these topics.
> The number of videos has gone up 9 orders of magnitude and the number of developers has only gone up two orders of magnitude.
2 orders of magnitude means at the very least, going from 9 to 100 developers, which is a huge increase, but it could mean way more. I wonder how big the team really is, and what the changing team dynamics are like on that scale at that pace.
For me, YouTube is good for watching videos. If I want to discuss it, I post it on FB.
"Cheating - Know How to Fake Data
Awesome technique. The fastest function call is the one that doesn’t happen. When you have a monotonically increasing counter, like movie view counts or profile view counts, you could do a transaction every update. Or you could do a transaction every once in awhile and update by a random amount and as long as it changes from odd to even people would probably believe it’s real. Know how to fake data."
So all those people who buy views are kinda screwed now :-) I suspect this is a bad example. I HOPE this is a bad example, if only for the KONY2012 campaign :P