I like how every rebuttal turns into "how fast can you compute a fibonacci number". I was looking for a function that burned a nontrivial amount of CPU, the choice of fibonacci was arbitrary. Let's move on from that.
What I was showing was that if your request handler does a nontrivial amount of CPU work, it will hold up the event loop and kill any "scalability" you think you're getting from Node.
If you Node guys were really that irritated by this, you're going to be super pissed when you learn how computers work.
I would argue that heavily CPU-bound stuff shouldn't be run in a web app. It's much better to offload the task to a worker system via redis/zeromq/kestrel/etc. The majority of activity you see in every web app I've ever designed has been almost exclusively I/O bound.
Don't you think there's at least some validity to Ted's argument that the statement "Because nothing blocks, less-than-expert programmers are able to develop fast systems." from the node.js homepage is somewhat misleading?
As Ted points out, there are things like Fugue and Nginx which people who are not "less-that-expert-programmers" do, "experts" will be fine whether they've got magical behind the scenes async stuff going on or not. The question as I see it is - are the node.js docs/homepage misleading about how easy is is to "develop fast systems"?
Wait, I think I've missed something because that response seems overly dismissive. How is offloading CPU-intensive stuff to a worker system not getting it done?
because you dont fix everything with another layer of indirection. add in some more queuing (which as we know never has a problem) just to ger around 'no threads' seems stupid.
In a classic thread-based system, it's okay if a single page takes a bit longer (e.g. >1 sec) to render if it's within your user's line of expectation.
You can't do that here because you'll block all the others.
But any decent developers knows that, so he takes the advantages Node.js offers and fixes the disadvantages that come along with it. Big deal.
Is this discussion really only about the scalability tagline? Some taglines are misleading, really?
That comment sort of destroyed all your credibility. Would you care to elaborate where in the "real world" there is ever a situation where a tight, asynchronous event-processing loop is required to do heavy CPU lifting?
The problem is that fibonacci was a bad choice in your example and it proved nothing.
It is fairly easy to scale node with multiple processes. As long as you don't have a long running (such as fibonacci) operation. If you do have tasks like that, process outside node and check for completion. Like how Tasks work in Google App Engine.
Also, most other web stacks will discourage you from running a 30s fib on a thread processing web requests. This isn't specific to node.
Node and coffeescript has worked really well for us. Product coming out later this month.
[EDIT: Just noticed that several other people pointed out the same thing. Looks like most node users are aware of potential problems, but I can see such issues being confusing for new users.]
> Also, most other web stacks will discourage you from running a 30s fib on a thread processing web requests. This isn't specific to node.
Difference being, with other stacks a request running for 30s will have little impact on the rest of the machine. With node, the whole server gets stuck, not just that precise request and the machine resources necessary to perform the computation (or whatever).
The fib example is extreme, but it's rooted into a real issue of cooperative multitasking: code does not always behave correctly and is not always perfect. You might have used a quadratic algorithm and it ran in 10ms on 10 items or so, but in production it happens a user is getting it to run on a hundred or a thousand items, and now other users are severely affected, in that their requests are completely frozen while the computation is going on. There are hundreds of other possibilities, small inefficiencies, shortcuts, plain bugs, etc... which are basically going to break your node application.
Even if you're not "crazy enough" to do what's prescribed, every user routed to the locked node instance will still be locked. You're just reducing the surface area of the freeze.
If your using node and doing work other than serving simple pages, then your probably sending that work off. Most production serious Node-ers know not to do heavy lifting within the system. I really don't know where anyone from the Node community has ever recommended stuffing their single threaded V8 backed event processor with cpu heavy tasks.
I still don't know why people are so maximist with their tools. NodeJS is a tool. It works well with your _other_ tools.
There - is - no - perfect - tool, no perfect programmer, not even perfect intent.
It is up to you to figure things out based on possible worst case runtime scenario coupled with your expected usage on _your_ hardware. You choose what work is defined 'trivial' (based on your resources). Trivial is always moving, and dependent on the scenario at hand. My trivial is not your trivial.
If the 'work' is too much you move it to another process. Either another NodeJS processor or some agnostic queue based managed worker. That worker could be anything.
True, and in fact, this is how nginx works: the process that owns the event loop is separate from the process(es) that does the work.
This is a decent way of mixing an event loop and multi-core processing, but with Node, you're forced to marry the HTTP server to the application, which is a dangerously tight coupling of responsibilities.
If you really want to do something silly like write your application in server-side JS because you're familiar with it, then it should be through some interface like WSGI in Python (or even CGI in days of yore), which properly separates HTTP connection handling from application serving.
There's nothing keeping you from running a separate web server and then proxying requests to Node using HTTP. Why is it a problem that you do this using HTTP? What alternative protocol would be so much better? FastCGI isn't really that different (each have their pros and cons) and CGI has obvious problems...
WSGI applications can use a built-in HTTP server too (or FastCGI or ...). Node has an internal interface (similar to WSGI) for handing over requests to the web application and so on. That's not fundamentally different from WSGI. The main difference is that WSGI is a standard API, so there are several "WSGI servers" implementing the same thing (each of them similar to Node in some sense).
I'm not sure if many people are choosing Node just because they're familiar with JavaScript. Most people would probably rather be less familiar with JavaScript. More likely they're choosing it because the runtime is very fast and a lot of libraries are packaged for it.
The WSGI/CGI bit is in fact mentioned in the article. :)
Node is an asynchronous programming framework bundled with a largely async library. If you have 40 cores on your system, you would presumably run 40 instances of node for a CPU intensive webserver(using multinode etc). So the event handler won't get stuck as long as there are available cores.
What about a single core system? Well I guess a threaded/multi-process solution would time slice the fibonacci requests between two threads so that both requests are served in 10 seconds, instead of one request in 5 seconds and the next one in 10 seconds like the node.js solution. Does not sound much better.
If you have done any kind of systems programming, you would know that availability of asynchronous I/O is a life saver, and can simplify your locking model greatly. 90% of the issues you face when building such systems is that some module deep inside grabbed a lock and issued a blocking I/O request and now the rest of the system is bottle necked behind it. Node.js is basically trying to eliminate the possibility of the existence of such a module. This complicates the issue of I/O calls, but simplifies locking in the sense that you don't really need all those locks in your system. In node.js of course there are no locks. The complexity moves from reasoning about locks to reasoning about correctly handling I/O calls and responses. IMO, this is the correct place to move the complexity to, because locks are simply an abstraction the programmer built. When debugging the system, we have to deal with - "How to get rid of this monolithic lock", when the real problem is - "This IO is taking too long we shouldn't be blocking on it". An async programming framework tackles this problem head on.
If you use Python/Perl you will never really know the number of instances of the process to run, Too many and you time slice requests, slow down all of them, increase your queueing buffers instead of just dropping the extra requests. Too few processes and you start dropping requests that you could have served. With a framework like node.js the number of instances you want is equal to the number of cores on the server.
Of course node.js can be an inappropriate solution for a wide variety of reasons, but I could not find anything really relevant regarding that in your post. Alex Payne discusses some issues here. You may want to read it.
http://al3x.net/2010/07/27/node.html
The point of the original article is that there's no point in avoiding blocking on IO while allowing blocking on CPU.
Furthermore, since node.js is single threaded, what's wrong with blocking on IO in that single thread? The process hangs, is put to sleep by the OS and wakes when there's IO available. You gain a simpler model of programming than using callbacks/continuations
Most web applications are io bound not cpu bound as they spend most of their time talking to other systems across the network like your db or queue or some rest service. the cost of spinning up threads is memory and context switches by the os. Async is just a way to avoid this for io operations. Node.js is just an incredibly convenient way of doing this as there are less ways of shooting yourself in the foot than if you apply the async patterns to other tech platforms.
Cpu bound is still cpu bound on any tech platform and needs another strategy to deal with it. I think it's better to look at node.js as a tool in your toolbox that is well adapted to running high load web-services and apps that are io bound instead of a swiss army-knife that does everything great.
But as any toolbox you need more tools in the box to be a good carpenter. That includes picking tech that can do the cpu bound stuff in a appropriate manner, be it java, scala, c++, erlang whatever. You don't write a graph database in node.js unless you are masochistic. Just as you hopefully don't write an async server in assembly.
From my usage perspective node.js is great for my async needs and the low resource usage means I need less hardware to scale my typical web app.
Yes it would and in fact most language platforms have async io. It's just harder to keep your yourself from introducing blocking calls in your event loop as they come with mostly blocking libs :)
Well, there is plenty of point avoiding blocking on IO if blocking on IO is your bottleneck, and blocking on CPU is not. For some reason, a lot of people have focused on that case for many years.
How is running 40 instances of node processes different from running 40 instances of a single threaded web server? How is running 40 instances of single threaded event loops better than running 10 instances of a process with 4 threads each? If you had a choice would you not rather have light-weight processes (threads) rather than actual processes because of lighter memory requirements? Threads are too hard to program to? Try STM? I'm not buying the notion that node's event model has any advantage over anything whatsoever. Try Erlang if you want a properly engineered, event-driven development environment.
> How is running 40 instances of node processes different from running 40 instances of a single threaded web server?
40 Instances of a single threaded webserver can block for I/O. If your webserver is 50% CPU bound this means that your CPU utilization is lower than in the case of a perfectly async system. You will serve fewer requests per second than an async framework. This is where the rule of thumb "no of threads = 2X no of cores" originates. Of course this rule wont work well for heavily I/O bound servers with high latency I/O. With node.js latency/IO percentage etc won't matter.
> If you had a choice would you not rather have light-weight processes (threads) rather than actual processes because of lighter memory requirements?
It would be great to have a multithreaded async framework. However a multithreaded environment eventually ends up introducing several blocking I/O functions which Dahl wanted to avoid. Hence the choice of Javascript.
node.js is one of a 100 possible solutions. Nobody insists that you use it. In fact I haven't even written a single line of node code. However I have done enough systems work to know the benefits of async programming.
> Threads are too hard to program to? Try STM?
STM can only handle scenarios that do not involve I/O. One of my colleagues was in the group at Microsoft tried STM with I/O that fell hard on their faces. Sure there are plenty of approaches - threads, actors, STM. Async programming is one such approach. If you want to write an async web server, right now node.js is the only solution. I think it might be possible to do a pure async web server in Haskell, as any IO gets captured in the type signature but I don't know of any async Haskell webserver framework.
> I'm not buying the notion that node's event model has any advantage over anything whatsoever.
You are basically asserting that async programming has no advantage over any other approach whatsoever. Having dug into hard disk device drivers, filesystems and caching for the Windows CE kernel, I would have killed to have a proper async I/O framework in CE from the ground up. We had a gazillion locks in the kernel modules, for gazillion data structures when all we really wanted to do was perform I/O without grabbing a lock. The Linux epoll, BSD kqueue and Windows IO completion ports are all async APIs added for high performance systems. These APIs are, strictly speaking, not required if you have threads but when you get into sufficiently advanced systems programming you cannot live without these. Trying to say that async programming is useless is equivalent to claiming that APIs such as epoll/kqueues are useless.
@zohebv there are several errors in deduction here:
> a multithreaded environment eventually ends up introducing several blocking I/O functions
really? multi-threading and non blocking io are two separate tools and one doesn't have to choose one of them exclusively. they can be intertwined, boost::asio allows you to run a single async event processor on as many threads you want. all without doing any explicit locking. if you're averse to locking (which based on your experience seems to be the case) you're taking things too far by avoiding threads completely.
> You are basically asserting that async programming has no advantage over any other approach whatsoever
i don't think anyone's asserting that. the original argument made was - if all you have in one process is a tight loop dispatching non blocking io handlers then you can't handle computationally intensive tasks. of course, you can spawn 40 other process - but i don't like a model where that's your /only/ option. there are middle grounds and any system that discounts them is short sighted.
> Trying to say that async programming is useless
yet again. i don't see that being said anywhere.
> Node is an asynchronous programming framework bundled with a largely async library ... If you have done any kind of systems programming, you would know that availability of asynchronous I/O is a life saver
now, if you'd indulge me with my own escapades in logic and word play.
icing is sugar whipped in butter. if you have eaten any
desert, you'd know that sugar is nice and tastes very sweet. thus, conclusively, irrevocably, icing is good and we shall eat nothing else.
> @zohebv there are several errors in deduction here:
@ajd I would disagree with this statement very easily and very confidently.
> really
yes
> multi-threading and non blocking io are two separate tools
asio is a way of minimizing the number of threads in a system and extracting maximum performance.
> i don't think anyone's asserting that
It follows by logical deduction. There are only 3 possible criticisms of node.js
1. I need shared memory
2. I dislike Javascript
3. I dislike asio
Given that the cancer post doesn't make a big deal of 1 and 2. you are left with 3. And if you read his follow up post he is again complaining about "event-loop" programming. Yes he is indeed complaining about asio. asio necessarily introduces event loops.
There is only one reason to use multiple threads over multiple pre-started processes. You need the shared memory. You should still be able to pull off shared memory in node but it can be cumbersome. I am yet to see anyone complain about needing shared memory and hence disliking nodejs. Its either
1. "threads are robust for cpu bound workloads"
- which can be solved by simply launching as many nodejs processes as Python processes, though you only need to launch as many nodejs instances as cores if you are not interested in winning this argument.
2. "Don't force this programming model on me"
It is not forcing any programming model on you except that shared memory is harder. And yes you cannot acquire a lock and go off and do io. If you want to do that then you can, by piling up pending events in a queue, but the ugliness of the code will stick out easily. You might as well switch to Python/Java threads or whatsoever.
> of course, you can spawn 40 other process - but i don't like a model where that's your /only/ option
Can you describe the other options you want to try?
Check this out
http://teddziuba.com/2011/10/straight-talk-on-event-loops.ht...
He has broken down his own defense with his fancifully named "Theorem 2". If you do more IO than CPU then "use more threads". Except he doesn't give you the number of threads because he doesn't know how many. In fact, he cannot know. And this is why asio is a win.
Fact is on an n-core system if you launch n nodejs processes you are guaranteed one of the 4 hold true
1. You service all requests thrown at the system
2. You maximize the CPU utilization
3. You maximize I/O utilization
4. You maximize RAM utilization - "4. is some what pedantic"
i.e. it will extract the maximum possible performance from the hardware you throw at it. It is just a consequence of making sure all io is non blocking. With a threaded solution you will never get the number of threads right and you will end up with a server that cannot serve all requests even when it has spare CPU, spare I/O capacity and spare RAM. Maximizing CPU utilization implicitly assumes the absence of locks. If you use asio as well as locks CPU utilization will not be maximized. This is something that most of the "nodejs" critics don't understand or fail to appreciate.
Yes shared memory is harder to do in node, but erlang doesn't do shared memory, Python cannot do threads sensibly, Java cannot do coroutines, why dump hate on nodejs because it isn't an ideal environment for a solution that requires shared memory? Sure there are many problems that require shared memory. However 95% of webservers don't fall into this category.
Lastly ted's inexperience really shows here, he is only 27 years old and has a lot left to learn. To start with he can stop trying to school Ryan Dahl, who is someone who certainly knows his Computer Science and is making a valuable contribution to the community.
Lastly as a practical exercise, try to build (at least as a thought experiment) some web service that outperforms node using respected platforms such as Python and Ruby. asio is a technique for IO bound loads but node will do better than Ruby/Python even on CPU bound loads thanks to V8. An order of magnitude faster. I have said that you need to have n processes for n cores, but in practice it seems just one process turns out to be enough as loads tend to be io bound and V8 is typically 10 times faster than Ruby. And if you have coroutines event based programming is not hard at all. So yes people have discovered that one nodejs process has replaced their two dozen Ruby processes and is serving out twice as many requests from the same box and they are impressed. And they don't care about what Ted thinks.
As a footnote, of course, you can do shared memory within a single nodejs process(trivially true), but for pure CPU workloads Java/C++ would probably be a better option.
my issue with node is independent of the programming language. i came here neither to bury ted nor
to praise him. i don't care what he has written now. i found his original article funny and thought
that it had some truth in it. all the other "logical deductions" you're making, require axioms that
you have and i don't. i don't know this ryan guy either but if he's made a system so many people
(including yourself, who i admire sincerely) think so strongly about, he must be an all around great
guy. good for him.
as for me and this discussion, if node=asio and threads=avoid in your world, so be it, i'm ok with
that :)
> one nodejs process has replaced their two dozen Ruby processes and is serving out twice as many
> requests from the same box and they are impressed
that says a lot about Ruby (with which i have absolutely no experience)
As it so happens, I am now implementing a load balancing solution on the JVM using the hybrid technique you recommend, lots of threads and some asio coming soon. However asio support on the JVM is poor. You are correct, in that using asio does not preclude you from using threads and viceversa; while node.js does not let you use threads. Your argument is valid, but this is very different from the argument ted is making viz. a CPU bound task will make the server useless. This is demonstrably false, and ignores the fact that typical web workloads are io bound.
The C++ solution you posted is probably a hybrid solution as you described, but it has its own set of restrictions. The environment does not guarantee run time safety, functional programming support is poor and coroutine support is non-existent or weak(I noticed there is an unfinished Boost.coroutine library that must heavily depend on #include <functional>). While node shuts out threads it enables other solutions more suitable for a heavily I/O bound server. I only want to draw attention to the fact that almost all solutions available today involve some kind of compromise and hybrid solutions are possible in almost all of them.
As for the "deductions", I thought that they were obvious. You need threads rather than processes, so that you can share memory and avoid the serialization/deserialization complexity when communicating between threads. Or less often, you need a more elaborate synchronization mechanism such as reader/writer locks. As for asio extracting maximum performance, I think I will be better off writing a blog post about it.
What I was showing was that if your request handler does a nontrivial amount of CPU work, it will hold up the event loop and kill any "scalability" you think you're getting from Node.
If you Node guys were really that irritated by this, you're going to be super pissed when you learn how computers work.
I ain't even mad.