Javascript is single-threaded, so we normally have to do weird tricks to have Node.js fully utilize multicore CPUs. With worker threads, things have changed… Follow me on Twitter: / ryancodez
i think there is some confusion as to the actual meaning of async. Async does not really have anything to do with threads. They are closely related but fundamentally separate. You can have async execution within a single thread, sometimes this is even more performant than using a thread pool. Async execution just means that the event loop can be passed off to other tasks.
Your take on async and threads is completely wrong just saying. JavaScript has a asynchronous non-blocking I/O thread, a single thread that does not block execution. It’s more easy to explain in a metaphor. If Node was a server at a restaurant it would take that order. Give it to the chef come back out and take another order aka HTTP request. On the other hand PHP would grab the order. Wait for the Cook to complete that order before they ever took another order from another customer. Basic synchronous operations verse asynchronous operations. Research the JS event loop. What do you think callback hell is? Why did ES6 give us the Promise API? Just a fun fact synchronous languages don’t have a call stack. ✌️
@@thedelanyo Sorry you’re right I read it again. I just caught caught more with “threads” in the video. Maybe someone can explain to me how this is a “game changer” in 2023 being that the JS event loop has always been this way.
This isn't like the multithreaded nightmare that is Java. You explicitly need to define a chunk of memory or data to be shared and be mutable, it's not "anything goes".
A few of your videos have been dropped into my homepage, you look like have some big balls man. The videos you make are mostly things I already know but haven't used yet but it's like you're introducing them nicely.
People often confuse concurrency with parallelism, concurrent tasks run at the same time, but only one is executed at a given time, parallel tasks are executed at the same time, thus requiring synchronization
@@danielvega646 here's a easy way to understand. concurrent: cpu will spend x time on a task, but once the timer is up it goes to another; parallelism: cpu will process all the tasks at the same time. in this case, worker_threads execute parallel instead of concurrent. this is a big simplification so if you want to know more how concurrency and parallelism work you should look it up and also try to understand how the OS schedule these tasks to multiple cores at the same time.
People get confused by this because they use the phrase "at the same time." You used that phrase to describe both concepts. I feel like the industry chose poorly with the term concurrency, since the meaning of the English word is "at the same time," and more accurately describes parallelism. I'd have offered something like interspersed, interleaved, multitasking, or something like that. If computing concurrency is occuring, then the different tasks are very much not being performed "at the same time." I had been using the computing terms essentially interchangeably because of this misunderstanding, luckily without ever being in a context where it mattered.
i hit the subscribe button HAAARD im going to explore the rest of your videos and hope they are just as well done! thank you for the information and high quality production!
Context switching ! If you spawn more threads than the available number, they're going to take much longer, That's way you should create a thread pool first and assign tasks to it.
@@anothermouth7077 No one is going back to anything, worker threads are even present in thr rust + actix for example. You need to understand what you are trying to do, in this case if you have a long lock it's interesting to invest into a worker thread.
Glad you discovered it! ~3 years ago I failed a junior interview because the interviewer asked me "why javascript was single threaded". I explained her I felt that the question was a "gotcha question" because the initial affirmation was false and depending on the enviroment you could make use of multiple processes or threads with workers or clusters to achieve multithreading. She laughed at me and told me I was wrong because javascript used the event loop and couldn't run on multiple threads. I don't think she was an engineer but to this day I'm not sure and I think she was following a script, a bad one as you can see... Anyways, thanks for the video! Keep pushing!
@@josedallasta JS used to be single threaded. However modern JS has features like SharedArrayBuffers and Atomics, which are definitely multitreading features.
Node isn't always single-threaded, as its core is written in C/C++ and certain modules will run in separate threads. But Node's single-threaded nature means we avoid having to deal with potentially difficult to debug problems related to memory sharing / data synchronization between threads. I/O, network, and CRUD operations against a database are usually the bottlenecks, not long-running code blocking the main thread.
Well, not quite. In Node, both IO operations and long-running code can be bottlenecks. Yes, if you don't have long-running code, then your application will run fine. But as soon as you have long-running code, that'll bottleneck you way faster than in other languagues because it's that much harder to execute that code in another thread. In that sense, long-running code is the Node.js-specific bottleneck and is what should be avoided if you want to have low latency.
NodeJS is single-threaded (CPU handles tasks sequentially), but the I/O is asynchronous. So, the I/O devices - like networks, storage, and other connected devices, do not have to hold on to the CPU. Other languages are multi-threaded, but their I/O were mostly synchronous. This is why you get HTTP timeouts with Java and Python, but you never get it with Node.
@@josephmariealba8483 both Java and Python has asynchronous. HTTP timeouts are caused by many reasons (and Node application can return HTTP timeout too), sometimes its due to bad code implement, sometime it's security mechanism (to avoid some kind of attacking method - which try to hold connection as long as possible).
@@DavidsKanal > that'll bottleneck you way faster than in other languagues because it's that much harder to execute that code in another thread How easy it is in other langs?
@@hl7297 for c#, more or less: new Thread(fib).Start() new Thread(fib).Start(singleParamPassedAsObject) new Thread(() => fib(param1, param2, param3)).Start() to create a completely new thread or Task.Run(fib) Task.Run(async () => { fib(param1, param2, param3); }) to run your code on one of pooled threads (generally better idea, unless the code is long running or does synchronous io) you also have PLINQ where you can for example do ' var x = someArray.AsParallel().Select(i => i*2 + 2).ToArray();' which will distribute the calculations over all cores and merge the result automagically. Note that while it's easy to start code on new thread, it's not that easy to make sure the code runs correctly in parallel (think locks, deadlocks, race conditions, compiler and cpu optimizitations causing writes to var on thread 2 not visible to thread , multiple threads accessing same var under coarse-grained lock effectively making your multithreading slower than single thread and many many more)
It is also possible to use the cluster built-in module is well-suited for scaling network applications, the cluster module is more suitable for scaling network applications, while the worker_threads module is more suitable for parallelizing CPU-bound tasks within a single process.
Hey, I had some experience with this feature while doing PWA application. And sadly I should say, that Workers aren't "magic stick", that can fix any performance issue by just wrapping a code into it. One big limitation is that data, transferred between worker and main thread should be serialized. Serialization can cost more than performance you potentially got by splitting algorithm into threads.
@@EuSouAnonimoCara Yes and no :D. Anyway you can only serialized data into shared buffer, so performance issue about serialization doesn't went away. It can be managed in early stage of developing module, but sadly you can't just wrap smth into web-worker to make it faster. My point was about it :D
You can use transfer lists to move memory to the other thread downside is you can’t use it in the thread that sent it but you can have the worker thread re transfer the same buffer back after doing whatever it needed to do. Nice thing is you can still use postMessage but can avoid the overhead of deep cloning
@@sawinjerwhere do you learn all this stuff? I'm taking a node js course and I'm not sure if I'd be able to keep up with the people in the comments when discussing this stuff. I think I've heard a few speakers talk about this, but that was a long time ago and I can't remember who or what they said. (Kyle Simpson maybe)? Did you study CS?
Nodejs is outdated BE framework. Why would anybody except for some "fullstack" devs use it nowadays? Libraries are absolute mess with tons of vulnerabilities and close to zero support, let alone half of them dies over 1-2 years. Why would you use it over python, c#, kotlin, go, java and etc... Those languages have frameworks that are way more mature with way more reliable libraries and way less security concerns. People claim that starting with nodejs is simple, while it is not true. It is infinitely easier to build a dotnet7 api for a newcomer. All the tooling works out of the box, no mess, actual code standards, no clickbait content, no confusion with libraries(what orm should I use? MikroORM, Objection, TypeORM, sequelize etc...). You just download SDK and IDE. No need to download nvm. No need to fight over npm, pnpm, yarn. No need for nvm either. Node is just a mess. Absolute nightmare to work with 3+ people on the same project. It is also absolute nightmare if you need to maintain the project for 5+ years. If I learned anything in the past decade, staying away from node is the best thing...
wow. just impressive how direct and well executed this video is. And it takes time and work. and you make it looks like simple. Just top of the line Job. Congratulations. The sad part is that I can only give a like once.
YOOOOO, your intro was amazinggg I thought whole video was like that for a sec hahaha tho if it was ngl it would be hella fun and exciting to watch it sure will take ton of time tho-;''
@3:02 That is why you don't do CPU intensive execution in Node.JS. But for input-output operation it is very optimum. It is still faster than other non-async language backend runtime.
Excellent vid as always... (and you should probably do voice-over work too, lol) - Have you had any experience with Caddyserver which is written in Go? - I am thinking of using it primarily for its reverse proxy features and it would be interesting to get your take on it...
Caddy is nice! But nginx and apache are pretty good too though. If you're not familiar with any of them and you don't care about battle testedness then caddy for sure
Thanks for this video! I am running a nextjs app, where each request takes around 10-20 seconds to complete (some GPU related stuff). The whole app used to freeze when more than 4 users used to connect at a time. It was hard to debug because I thought it should be async, turns out as you said if a blocking operation is performed in an async context, the whole event loop freezes and the users see a 502. I created a redis queue and a separate service that does the blocking requests and everything is working fine now. Would love to try out worker threads, looks promising.
I wish to make a nextjs app too but after watching this video and comments, I think I have not fully understood nodejs yet... I also want to use nodejs but ofcourse, I don't want frozen app with just a few users! Any tips on learning and understanding it in the real world sense?
@@ritik-patel05 nextjs api backend puts json stringified data on redis queue. Node script conitinuously looks for new value in queue.. When it gets a value from it, the value is parsed and gets sent to the GPU server. After, the GPU server is done fulfilling the request the node script gets the data and this data is again stringified and put in Redis . While this was happening, the client polls the nextjs app every few seconds to see the status of the job in queue. After processing of the job, nextjs app gets the output from Redis and sends it back to the client.
I already have a difficult enough time keeping node workers alive for more than a few months without PM2. Can't imagine how bad the reliability would be with that responsibility also being pushed directly into 1 node process, lol.
Nice and accurate video my friend! I wonder if backend frameworks like ExpressJS or Nest take advantage of this technique. Doing some processing on an image or something is a perfect use case to use worker threads.
2:08 Small tip: You may use "console.time" which will log time automatically in ms. Like this: console.time("Fib"); const result = fibonacci(iterations); console.timeEnd("Fib");
It is so easier to do in Go though. It is essentially the same amount of code as in the first (not working) example, but it works asynchronously as you would expect. You can create async functions just by adding "go" keyword to a function declaration (or right beside the execution also, if you want). Use a waitgroup instead of Promise.All and it just magically works as you would expect. Crazy, right?
@Name I think you don't know basic English. The other guy was talking about go being better than JavaScript and you said that go is a bad language, meaning in context that is no better than JavaScript, V doesn't have absolutely anything to do with this. Plus you look like a complete moron if you go around shoving your favourite language down everyones throats when its not asked for.
@@twothreeoneoneseventwoonefour5 it all good until that GC cycle kick in and grab your ball and you have nothing to do to even reduce it a bit. Go is as good as simple but when you start adding more memory related to it, then GC will happily blow all your work. i/e: personally i prefer Jai more, but since god know when it got release, Zig is fine
I don't know the performance difference between fork and worker thread, but i wanted to point to the fact that worker thread create v8 instance which have the overheard of allocating memory and all that stuff because it's fundamentally creating new process
At 3:05 I guess you meant 'sequentially' instead of 'synchronously'. In your fibo code case Concurrent = Parallel = Synchronous (concurrent not necessarily sync. Async can be concurrent too). Good tips btw.
This example seems to be about parallelization rather than concurrency. The first time the calculations ran on a single core and it wouldn't matter if they ran asynchronously or blocking, because the rest of the cores were idle. The second time the job finished faster because of mutithreading, but the result would be the same even it would run in a blocking fashion.
Ryan, that's a new thing for me either . Theoretically, I though it's enough to wrap sync func in Promise API, however we both were wrong with this understanding . 😮 Also just wondering how about multi-threaded (child) processes to employ Golang goroutines . Communication-wise between Node and Go use standard streams ? 🤔✌️
You can split sync func into asynchronously executed pieces, but it's not always possible, and requires completely different approach. Meaning that functions like fibonacci cannot be executed neither asynchronously nor in parallel, as every next step requires data from previous one (adding previous numbers). However you can Promise.resolve(returns of fibonacci), so the fibonacci function itself won't execute faster, but now event loop is able to execute some of the tasks in-between recursive fibonacci calls. In this case you'll have to await every return from fib fn, but now it won't be as blocking, but in reality it won't make any difference, and most certainly would still block the event loop on high numbers. Essentially you can't make anything faster with async in javascript, with exception of I/O tasks, which are actually handled by separate thread/threads, so that event loop while it's waiting for I/O to respond - is able to perform something in-between.
I think you have concurrency and parallelism mixed up. As well as the definition for asynchronous function. When you ran the worker threads they were running in parallel, that is, multiple threads running the same program on the same tick. When you ran asynchronous, those are concurrent.
I think you're right: "Concurrency is the task of running and managing the multiple computations at the same time. While parallelism is the task of running multiple computations simultaneously."
Thank you very much, I did lost an amazing job opportunity because I could not explain exactly it. You did it sou clearly that next time I will no fail 😅
Workers are really cool if you understand them. I use them for some front-end applications, happy to see they are in node now. They are expensive to create though. They're not equivalent to doing like a goroutine or something.
@@matth23e2 Yeah pretty much. I run a reasonably cpu intensive search algorithm in a worker, and it just returns the array of references. Keeps the main thread unblocked. If you give workers some thought you can use them for a lot of things though.
Technically fork is only slightly slower these days. One could man handle fork to essentially be threads. In fact, under Linux, threads are not actually "threads" but "light weight processes" and thats because forking is actually what happens when we create threads; only that when we create threads, we dont clone things like the file descriptor table and some other things, so thread creation becomes somewhat faster than a normal fork. That being said, I loved this video. New sub.
@@boohba yes, clone is the underlying syscall that it's all built upon. I figured keeping it in the fork/join terminology would be simpler in this case, as fork/join parallelism technically is the ancestor of multi threading. So forking, creating threads, its all built using the clone system call with different parameters.
Hey Ryan, while I understand this video is to demonstrate multicore execution I think it's important to point out that you can achieve asynchronous execution inside the event loop without workers. It would have been good to see you include the following async version of the code in your demonstration so that users understand there is a middle ground that allows asynchronous execution without requiring worker management. async function fibonacci(n) { return n < 1 ? 0 : n
Async and promises do the same thing. They are just syntactically different. The video is highlighting that you need multiple cores in order to execute things in parallel. You are not able to execute things in parallel in JS because it only has access to one CPU which can only handle one process at a time. JS Async is not handling tasks at the same time, it is scheduling tasks to be handled at a later time. This is concurrency. This video is about parallelism not concurrency.
if you've ever used assembly language before, Its like putting a jump statement at the end of a block of code to go to a previous (or future)spot in the code, the asynchronous portion of the code, then have another jump statement to jump to the next part of the code the dev wishes to execute when the asynchronous code finishes.
@@terrencemoore8739 If you read my comment you should see that I said that async/await is not multi core execution and my main point was that his example used promises in a way that didn't really make sense, he could have just put the calls to the functions as direct calls as he wasn't really making use of promises in a meaningful way. If the code is going to use promises it should demonstrate how they can execute code asynchronously.
@@John.Whitson The video is not about asynchronous code, its about parallelism. He acknowledges that JS is async but if you want to do two things at the same time, its impossible. CPUs handle one process at a time. If you want two things to run at once , you need two CPUs
worker threads are using new processes to accomplish theirs task on separate threads (one thread by process in a single thread fashion) you can get the PID of each worker_threads I guess the only multi-thread capabilities is with IO via LibUv using callbacks, promisses and async await on those...
Is this a right approach ? Correct me 1. Assign cpu intensive sync tasks to worker threads 2. Let node event loop/thread pool handle async ops 3. Spawn multiple processes with 1&2 to balance the traffic load I always thought this is the way, as I said open to corrections
That worker thread API is incredible. I had SO MUCH trouble making an efficient web crawler in NodeJS that I switched to Java and then Python. No joke, I ended up as a Python dev because I didn't understand NodeJS haha.
If you do understand NodeJS, you will eventually end up as Python and that's a good choice. NodeJS worker thread is not that great. Performance is bad and it takes a lot of time to transfer messages between the main thread to workers and vice versa. In this video example, you can see that each doFib() in the main thread is done in 480ms but when using worker thread, it needs around 970ms to finish. Let's choose a real multi-thread language when we need it for CPU intensive tasks. NodeJS event-loop is good for IO tasks only.
Node는 System call로 던질 수 있는 I/O 작업은 이를 통해 os로 던지고. Crypto나 기타 os로 못 던지는 작업들은 node의 스레드풀을 이용한다고 알고 있습니다. 결국 worker thread를 사용하지 않는다 하더라도 node는 멀티 스레딩을 활용하는 것이지요.
저도 그렇게 알고 있습니다만, 그 멀티 스레드 활용이 제한적이라는 게 여기서의 논쟁거리인 것 같습니다. Ryan 의 피보나치 테스트에서 수치가 나왔으니까요... 만약, Ryan 의 코드에서 피보나치 연산을 외부의 C 스크립트로 처리했고, Node 는 그 연산결과를 받아오는 방식이었다면, Node 의 멀티스레딩을 신뢰하는 사람들의 기대에 부합하는 결과가 나왔을 수도 있겠죠. 적어도, C/Python 으로 처리해야 할 작업과, Node 가 직접 연산해야하는 작업의 판단은 있어야 하고, worker 로 스레드를 분산해 준다면, 그에 따르는 위험도 관리할 수 있어야 한다는 정도로 저는 정리해봅니다.
actually, everything that's being called with async is in fact separate thread but it's doing some OS stuff in a background, what you also shown us is concurrency not the parallel execution so in fact it's not fast, it's just handling executions frequently between them, but it's not increasing speed, concurrency only increases speed when it comes to IO bound operations, but this is true for different techs like in Python or Ruby, Java can tho truly run in parallel and utilise all cores, so there are many misconceptions in net about this.
Single thread is the beauty of nodejs because context switching is expensive and you don't have to worry about variable locking and complicated race condition. Node has an event loop and execution stack to handle async operation in single thread. That's proven to be very suitable in server environment in last 10 years. If you have long execution need to pull off from the execution stack, you can simply use setTimeout(() => {}, 0) to allow other execution in the queue to jump in. That's what non-blocking IO means.
Your concern is valid, but i have to say that all program will crash anyway. Thats just the nature of software development. I have never seen a program that doesnt crash in production. Its a matter of how you manage it and avoid it.
Take something like image processing though, a very common and costly need. How do you call setTimeout in the middle of that. Afaik it doesn't matter if you've called it async, your thread is still blocked once it begins.
@@KerryOConnor1 I understand your point. Serious image processing shouldn't be done in node level, and this is not what Node is good at. You can manage a child process to do that. Use the right tool for the right thing. To answer your question, calling setTimeout will put your stack to the end of the queue again, so other waiting processes can be handled first. Then, your turn comes again, and you setTimeout again after taking some time to process it. Don't block your thead is the key. That's how non-blocking scenario is done to handle async IO usage. If you are using Node to purely do intensive CPU calculation like what this video is doing, I am sorry to say that this is not the way to use Node. My conclusion is, nothing is perfect, most important is to understand the strength of Node and use it right.
Context switching isnt as expensive as it used to be for multiple reasons. Mostly due to optimizations of cpu cores / cache Creating new threads isnt the end of the world like it was before. A lot of that stuff was written before multi core processing lmao
NodeJS (or JS) doesn't have a lock for volatile variables, so how can we ensure that two (or more) processes are not changing the same buffer in the memory at the same time?
For using other cpus/cores you still need to use cluster module, This feature only lets you spawn threads in the current process, if thats fits your requirement, yeah you should definitely try Worker threads .but if you ask me, you should use the combination of both. Simply put, you can assign cpu intensive sync tasks to workers , let node handle async tasks and do multiple forks of this combination to handle incoming traffic
does it share the loaded lib at runtime ? What if a package has a singleton and it is imported by both workers and master ?? Do they share a ref ? What s the overhead using it ? Besides the sharedbuffer interface, it looks likes a neat improvement of past ipc systems. looks good overall.
Worker threads are not a new thing in nodejs. It even works in browsers lol. JS ecosystem is still slow and no one writing fast apps in it. Nodejs designed for IO, not cpu-bound tasks
This is the best video ever. First of all it finally explained to me what nodejs is, and secondly such a clear explanation of worker threads and multi threading. 🔥🔥🔥
I had to stop watching after around 1:17. The Javascript always run on a single thread, but all of the IO, file handling, etc isn't happening in Javascript. All of those things are managed in a thread pool. By default, NodeJS maintains a thread pool of 4. This can be overridden by setting the UV_THREADPOOL_SIZE env variable.
Make sure you have the cores to support those threads. If you’re writing multithreaded code and then deploying as a 1-cpu container, you won’t see any benefit.
Hi Ryan, how then do we decide how many workers to spawn? My basic understanding is a cpu can have multiple cores, and a core can handle multiple threads.
