This is dangerously wrong and I would suggest that you reconsider the steps that got you to this understanding because something really important has been lost. It is absolutely critical to understand that deadlocks are not why you have locks. Correctness during concurrent operation is why you have locks. Deadlocks are a failure state when you do not have correctness during concurrent operation. So are things like double-increment and double-create.

Parallelism does not imply deadlocking, concurrency implies deadlocking, and both NodeJS and Python are concurrent runtime environments. And I can guarantee you that, with a little skull sweat, you can write a deadlock in NodeJS or Python. It is very easy. If you need some help, here's a trivial example (and ordinarily I wouldn't use a link shortener here but this one is hefty, it just goes to the Typescript playground): https://bit.ly/2Tvjyze

Also, as a concrete, real-world, yes-it-happens-here example of where locking is important, consider that I've recently built a dependency injection framework in NodeJS--tried to use others' first, but my situation isn't covered by existing ones--and had to resort to a mutex to avoid double creation of objects within a single lifecycle. Creation of objects within this lifecycle happens asynchronously--it has to, as the act of creating the objects might itself rely on asynchronous operations. So, if I were to have a diamond-dependency (A deps B and C, B and C dep D), I will non-deterministically, and based on the creation time of B and C, create either one or two instances of D. I rely upon a mutex, keyed upon the dependency being created, to ensure that this does not happen.

.

I would also submit that perhaps you should adopt a principle of charity and think real hard about whether your priors are correct before you start talking about what she "fails" to see. Rachel is one of those people who has Been Around and while I also have Been Around, I understand that Rachel has Been Around More and I probably should be listening more than I should be smarming at her.

Just a thought.

I have personally made this mistake on two or three occasions where I conflated file/module behavior with class behavior because I wanted a python file to act like it was a bit more declarative. Unfortunately this leads to a world of eternal pain. You can work around it, but you should have made everything a python class and pretended like the files/modules don't exist or at least have staggeringly different semantics hiding behind that innocent little `.` operator. Python simply cannot support the desire to solve a problem in a certain way because of the structure of the problem and forces you into using its happy path patterns if you want it to work in slightly different run time contexts.

Two relevant posts from instagram engineering on the subject which suggest that best practices for avoiding these kinds of issues are non-obvious and easy to miss.

https://news.ycombinator.com/item?id=20708889 https://news.ycombinator.com/item?id=21284669

Rachel says that a thread is only ever doing one thing at a time - it is handling one request, not many. But that's only true when you do CPU bound work. There is no way to write code using blocking IO-style code without using some form of event loop (gevent, async/await). You cannot spin up 100K native threads to handle 100K requests that are IO bound (which is very common in a microservice architecture, since requests will very quickly block on requests to other services). Or well, you can, but the native thread context switch overhead is very quickly going to grind the machine to a halt as you grow.

I'm a big fan of gevent, and while it does have these shortcomings - they are there because it's all on top of Python, a language which started out with the classic async model (native threads), rather than this model.

Golang, on the other hand, doesn't suffer from them as it was designed from the get-go with this threading model in mind. So it allows you to write blocking style code and get the benefits of an event loop (you never have to think about whether you need to await this operation). And on the other hand, goroutines can be preempted if they spend too long doing CPU work, just like normal threads.

You can just pass `--preload` to have gunicorn load the application once. If you're using a standard framework like Django or Flask and not doing anything obviously insane then this works really well and without much effort. Yeah I'm sure some dumb libraries do some dumb things, but that's on them, and you for using those libraries. Same as any language.

If you want to stick your nose up at Python and state outright "I will not write a service in it" then that's up to you, it just comes across as your loss rather than a damning condemnation of the language and it's ecosystem from Rachel By The Bay, an all-knowing and experienced higher power. I guess everyone else will keep quickly shipping value to customers with it while you worry about five processes waking up from a system call at once or an extra 150mb of memory usage.

The author seems to be saying something about how if a worker is busy doing CPU intensive work (is decoding JSON really that intensive?) then other requests accepted by that worker have to wait for that work to complete before they can respond, and the client might timeout while waiting?

If that's the case:

1. Wouldn't this affect any language/framework that uses a cooperative concurrency model, including node.js and ASP.NET or even Python's async/await based frameworks? How is this problem specific to Python/Gunicorn/Gevent?

2. What would be a better alternative? The author says something about using actual OS-level threads but I thought the whole point of green threads was that they are cheaper than thread switching?

On the first point, yeah Rachel’s posts are kinda snarky sometimes, but some of us find that entertaining particularly when they are highly detailed and thoroughly researched. I’ve worked with Rachel and she’s among the best “deep-dive” userspace-to-network driver problem solvers around. She knows her shit and we’re lucky she takes the time to put hard-earned lessons on the net for others to benefit from.

As for “microservices written in Python trading a bunch of sloppy JSON around via HTTP” is bad engineering: it is bad engineering, sometimes the flavor of the month is rancid (CORBA, multiple implementation inheritance, XSLT, I could go on). Introducing network boundaries where function calls would work is a bad idea, as anyone who’s dealt seriously with distributed systems for a living knows. JSON-over-HTTP for RPC is lazy, inefficient in machine time and engineering effort, and trivially obsolete in a world where Protocol Buffers/gRPC or Thrift and their ilk are so mature.

Now none of this is to say you should rewrite your system if it’s built that way, legacy stuff is a thing. But Rachel wrote a detailed piece on why you are asking for trouble if you build new stuff like this and people are, in my humble opinion, shooting the messenger.

As the maintainer of about 5 little services with this structure I have vowed never to write another one. The memory overhead alone is a source of eternal irritation ("Surely there must be a better way....").

Echoing other commenters here, the real cost isn't actually discussed. Namely that there is a solution to some of these problems (re long running tasks?), but it carries with it a major increase in complexity. Its name is Celery and oh boy have fun with the ops overhead that that is going to induce.

A while back I did some unscientific benchmarking of the various worker classes for python3.6 and pypy3 (7.0 at the time I think?). Quoting my summary notes: 1. "pypy3 with sync worker has roughly the same performance, gevent is monstrously slow gthread is about 20 rps slower than sync (1s over 1k requests), sync can get up to ~150rps" 2. "pypy3 clearly faster with tornado than anything running 3.6" 3. "pypy3 is also about 4x faster when dumping nt straight from the database, peaking at about 80MBps to disk on the same computer while python3.6 hits ~20MBps"

I won't mention the workload because it was the same for both implementations and would only confuse the point, which is that there are better solutions out there in python land if you are stuck with one of these systems.

One thing I would love to hear from others is how other runtimes do this in a sane and performant way. What is the better solution left implicit in this post?

I'm surprised how often devs treat this distinction as architecturally meaningful. Web requests are just RPCs with some of the parameters standardized and multiple surfaces for parameters and return values - query string, headers, body. This is completely orthogonal to the strategy used to schedule IO, concurrency, etc.

Because of the GIL, you can't make predictions at the same time you're processing network IO, which means that you need multiple processes to respond to clients quickly and keep the CPU busy. But models use a lot of memory and so you can't run all THAT many processes.

I actually did get the load-then-fork, copy-on-write thing to work, but Python's garbage collections cause things to get moved around in memory and triggers copying and makes the processes gradually consume more and more memory as the model becomes less and less shared. Ok, so then you can terminate and re-fork the processes periodically, and avoid OOM errors, but there's still a lot of memory overhead and CPU usage is pretty low even when there are lots of clients waiting and...

You know I hear Julia is pretty mature these days and hey didn't Google release this nifty C++ library for ML and notebooks aren't THAT much easier. Between the GIL and the complete insanity that is python packaging, I think it's actually the worst possible language to use for ML.

> So how do you keep this kind of monster running? First, you make sure you never allow it to use too much of the CPU, because empirically, it'll mean that you're getting distracted too much and are timing out some requests while chasing down others. You set your system to "elastically scale up" at some pitiful utilization level, like 25-30% of the entire machine.

Letting a Python web service, written in your framework of choice, perform CPU-bound work is just bad design. A Python web service should essentially be router for data, controlling authentication/authorization, I/O formatting, and not much else. CPU intensive tasks should be submitted to a worker queue and handled out of process. Since this is Python we don't have the luxury of using threads to perform CPU-bound work (because of the Global Interpreter Lock).

The question, however, is why one would use gevent at this point in Python's evolution. There's async await now, and things like FastAPI. If you want to use, say, the Django ecosystem, use Nginx and uWSGI and be done with it. Maybe you need to spend some more resources to deploy your Python. Okay. Is that a problem? Why are you using Python? Is it because it's quick to use and helps you solve problems faster with its gigantic, mature ecosystem that lets you focus on your business logic? Then this, while admittedly not great, is going to be a rounding error. Is it because you began using it in the aforementioned case and now you're boxed into an expensive corner and you need to figure out how to scale parts of your presumably useful production architecture serving a Very Useful Application?

Maybe you need to start splitting up your architecture into separate services, so that you can use Python for the things that it does well and use some other technology for the parts that aren't I/O bound and could benefit from that. But that's not this article is about. This article is about someone making the wrong choices when better choices existed and then making a categorical decision against using Python for a service. I'd say that's what "we have to talk about" if you ask me.

Let's say your server has 4 CPUs. The conservative option is to limit yourself to 4 requests at a time. But for most web applications, requests use tiny bursts of CPU in between longer spans of I/O, so your CPUs will be mostly idle.

Let's say we want to make better use of our CPUs and accept 40 requests at a time. Some environments (Java, Go, etc) allow any of the 40 requests to run on any of the CPUs. A request will have to wait only if 4+ of the 40 requests currently need to do CPU work.

Some environments (Node, Python, Ruby) allow a process to only use a single CPU at a time (roughly). You could run 40 processes, but that uses a lot of memory. The standard alternative is to do process-per-CPU; for this example we might run 4 processes and give each process 10 concurrent requests.

But now requests will have to wait if more than 1 of the 10 requests in its process needs to do CPU work. This has a higher probability of happening than "4+ out of 40". That's why this setup will result in higher latency.

And there's a bunch more to it. For example, it's slightly more expensive (for cache/NUMA reasons) for a request to switch from one CPU to another, so some high-performance frameworks intentionally pin requests to CPUs, e.g. Nginx, Seastar. A "work-stealing" scheduler tries to strike a balance: requests are pinned to CPUs, but if a CPU is idle it can "steal" a request from another CPU.

The starvation/timeout problem described in the post is strictly more likely to happen in process-per-CPU, sure. But for a ton of web app workloads, the odds of it happening are low, and there are things you can do to improve the situation.

The post also talks about Gunicorn accepting connections inefficiently and that should probably be fixed, but that space has very similar tradeoffs <https://blog.cloudflare.com/the-sad-state-of-linux-socket-ba....

- https://www.uvicorn.org/

- https://www.starlette.io/

Those below who complain about the complaints are missing the point.

We (computer programmers as a general class) have not learnt from history. We keep reinventing wheels and each time they are heavier and clunkier.

What we used to do in 40K of scripts now takes two gigabytes in python/django/whateverthehellelse. E.g. mail list servers. Mailman3 hang your head in shame!

> "Why in the hell would you fork then load, instead of load then fork?"

In python it often seems to make little difference. The continual refcount incrementing and decrementing sooner or later touches most everything and causes the copy to happen whether you're mutating an object or not.

I've had some broad thoughts about how one would give cpython the ability to "turn off" gc and refcounting for some "forever" objects which you know you're never going to want to free, but it wouldn't be pretty as it would require segregating these objects into their own arenas to prevent neighbour writes dirtying the whole page anyway...

> A connection arrives on the socket. Linux runs a pass down the list of listeners doing the epoll thing -- all of them! -- and tells every single one of them that something's waiting out there. They each wake up, one after another, a few nanoseconds apart.

Linux is known to have poor fairness with multiple processes listening to the same socket. For most setups that require forking a process, you run a local loadbalancer on box, whether it's haproxy or something else, and have each process listen on its own port. This not only allows you to ensure fairness by whatever load balance policy you want, but also lets you have healthchecks, queueing, etc.

>Meanwhile, that original request is getting old. The request it made has since received a response, but since there's not been an opportunity to flip back to it, the new request is still cooking. Eventually, that new request's computations are done, and it sends back a reply: 200 HTTP/1.1 OK, blah blah blah.

This can happen whether it's an os threaded design or a userspace green-thread runtime. If a process is overloaded, clients can and will timeout on the request. The main difference is in a green-thread runtime it's about overloading the process vs. utilizing all threads. Can make this better by using a local load balancer on box and spreading load evenly. It's also best practice to minimize "blocking" in the application that causes these pauses to happen.

>That's why they fork-then-load. That's why it takes up so much memory, and that's why you can't just have a bunch of these stupid things hanging around, each handling one request at a time and not pulling a "SHINYTHING!" and ignoring one just because another came in. There's just not enough RAM on the machine to let you do this. So, num_cpus + 1 it is.

Delayed imports (because of cyclical dependencies) is bad practice. That being said, forking N processes is standard for languages/runtimes that can only utilize a single core (python, ruby, javascript, etc.).

This is not to say that this solution is ideal -- just that with a small bit of work you can improve the scalability/reliability/behavior under load of these systems by quite a bit.

We have money, let’s just blow it. /s

I noticed in the logs that I am getting a lot of Critical Worker Timeouts and I am wondering if this has anything to do with it.

Is the message that epoll and co are lot efficient enough? That’s also true. Api Problems and thundering here are known. And not only limited to Python applications as users. io completion based models (eg throuh uring) solve some of the issues.

Or is this mainly about Python and/or Gevent? If yes, then I don’t understand it, since the described issues can be found in the same way in libuv, node.js, Rust, Netty, etc

Definitely, I am smarter than the guy who wrote this because then I wouldn't have these problems(Or He is smarter and I just didn't ask him about his rationale).

What I design wouldn't run into these BS problems that I have to fix, It just wouldn't run into problems generally. (Or It would have more problems than this one)

I had these conversations with myself at least a thousand times, and then it was just the case in the parentheses.

If I use uWSGI is problem gone?

Every time you access an object, its reference count is mutated, which will the memory page to be copied.

There are workarounds if you're willing to mess with the Python interpreter: https://instagram-engineering.com/copy-on-write-friendly-pyt...

I dug into it ~year ago, Django loads almost everything lazily, so simple --preload did next to nothing. I had to write code to load app for real at import time, exact thing article and common wisdom tells us not to do.

It's not always trivial to ensure none of your dependencies have import-time side effects. Sometimes the productivity/business benefit provided by the depedendency outweighs the pain introduced by the side effects.

With the current state of ecosystems, this quality-vs-quantity mutual exclusivity is much less pronounced. These days, you can fire up these services as quick or quicker than in Python with better performance and resource usage that is also more maintainable. Unless you speak of highly ecosystem-dependant libraries (e.g. ML), Python defenses that rely on time to market say more about the author's narrow comfort than general expediency.

If I were her, I’d put that on my CV. :D

2. Green threads have lower overhead, but it's a false economy if it causes you to do needlessly redo work because of timeouts that could have been avoided.

Which it seems it must because the kernel doesn't have the insight to know whether a green thread is doing that epoll because it's ACTUALLY idle or because it's not but it's willing to try to juggle a second (or third) thing while it has something on the back burner. So the kernel indiscriminately assigns work to threads without regard for whether they are juggling a lot or nothing.

Whereas with native threads, they never ask the kernel for more work while they're blocked on something else because they are literally blocked and thus won't be making that epoll system call.

(The article also mentions something about LIFO policy, which exacerbates the problem because it favors assigning work to the process which is likely to already have most of it.)

https://docs.python.org/3/library/asyncio-eventloop.html#asy...

Node.js will have that issue and in fact the stdlib JSON encoding/decoding can't even be paused so once you start a processing something you're stuck until it's done. You could, however, write an incremental serializer/deserializer that could spread processing out across many event loop cycles to mitigate.

Go, ASP.NET, and others not so much, depending, because the schedulers can pause and resume the tasks(on top of being threaded).

I think she's against anything that has that problem. Not every green thread implementation has that problem. For example Go doesn't have that problem. Because there were 4 CPU threads (I think) and only 2 things needing to be done, with Go's M:N scheduling those 2 things would be sure to both be running.

Yes. Decoding JSON with Python is CPU-intensive.

This a very simple shell script around Python, that is designed from the get-go to crash with an exception. However, it may not be the exception you expect:

    n="$(python3 -c 'import math; import sys; sys.stdout.write(str(math.floor(sys.getrecursionlimit() - 4)))')"
    left="$(yes [ | head -n "$n" | tr -d '\n')"
    echo "$left" | python3 -c 'import json; print(json.loads(input()))'

You won't. You'll hit RecursionError.

Because, despite how badly Python deals with recursion, the JSON library depends on it extensively. Which means that there is a huge stack-tree being built every time you try and decode JSON in Python (Dictionaries, function call overhead, etc, etc.), only for it to be thrown away.

in Python, everything is generally CPU intensive compared to what it would be in compiled languages, even though things like JSON decoding are usually happening in a C library, Python programs that do close to nothing still use way more CPU than you would if you were running in the JVM, or Go, C, whatever.

> Wouldn't this affect any language/framework that uses a cooperative concurrency model, including node.js and ASP.NET or even Python's async/await based frameworks? How is this problem specific to Python/Gunicorn/Gevent?

CPU bound-ness affects all of these platforms, yes. It affects Python and other intepreted languages the most however because these platforms get the most CPU-bound the most quickly. Also, applications that are written in scripting languages tend to have a lot of business logic going on in the first place; after all, if you just wanted to serve static pages you could use Apache with the Event NPM and if you wanted to proxy HTTP requests you'd use HAProxy; both event-based systems that are very much not CPU bound.

But yes, most importantly, Python's asyncio system is completely impacted by these same issues and I would have preferred she address that, as asyncio is part of the standard library now and is way more popular than gevent.

> What would be a better alternative? The author says something about using actual OS-level threads but I thought the whole point of green threads was that they are cheaper than thread switching?

I will grant she lost me a bit with the "use a real RPC system with <feature> <feature> <feature>" thing, and additionally the "load the application in the child process" thing is pretty typical, a worker process should obviously have either threads or greenthreads in use so that each process can handle multiple concurrent requests, but only as many as you'd want handled effectively by one core since the GIL is going to enforce that (another thing you wouldn't have to deal with in other languages such as the above mentioned compiled languages), but it's typical that child processes are going to have a mostly original copy of things.

But as far as the "context switching" thing, I've yet to see benchmarks that show the overhead of OS-level context switching actually being more of a performance burden than the less frequent, but more work intensive context switching that user-space schemes like asyncio have to use. If you are writing a logic-heavy, or even a logic-just-a-bit service that receives Python requests you will also have to worry about CPU-bound issues all the time. Using regular threads with processes, like what you get using something like mod_wsgi, will allow individual processes to attend to web requests more evenly. With mod_wsgi you can configure worker daemons that run multiple OS level threads and you can also have multiple daemon processes.

I'm not sure if the multi-process model used by mod_wsgi has solved the accept() problem, however in my experience the bigger problem is when a service configures itself to allow for 1000 greenlets within each process, while each process is realistically capable from a CPU perspective of handling maybe 5 or 10 concurrent requests, there's no mechanism that ensures that each process gets an even balance of requests. That is, you might have all your requests waiting in one process, because you told them it can process 1000 at a time, while other processes are idle.

TL;DR I'm in the "event based programming is extremely overrated in Python" camp.

Laziness is a virtue in our profession :) huge "citation required" on it being more efficient in engineering effort and most importantly it's ubiquitous and interoperable with any language without having to rely on grpc or thrift implementation and tools for that language/platform.

Re machine time efficiency - to the extent I understood what she was actually complaining about simply none of those issues are attributable to json-over-http being used.

I genuinely don't see much of a lesson to learn from this particular blogpost, and it appears neither did many others in HN. If there is one, beyond "don't use x", it's hard to find it.

I get the impression that this particular post is being upvoted to the top of HN because of who the author is, not necessarily because this post itself has value. This results in a whole bunch of others reading it, wondering why they're wasting their time with such a rambling post.

Also, I feel like she (intentionally maybe) is conflating concurrency and parallelism. These "green thread" systems provide concurrency, but not parallelism. That should be something people are aware of when they use them.

[1] https://capnproto.org/faq.html#how-do-i-make-a-field-require...

These systems tend to take care of a bunch of easy to mess up logic which tends to be accumulate around any system that wants to send something and not mess up. Any sufficiently old web request system tends to look identical to a high level RPC system designed badly.

So choosing an RPC system should give you all the features you’d eventually end up building around using web requests without spending your time rewriting it.

Also, TLS is not really ideally lightweight for RPCs, but you should absolutely encrypt your RPC traffic (imo.) So I really think the whole stack is out.

(P.S.: If you are wondering what kinds of 'lightweight' replacements for TLS exist, I think my personal favorite attempt is CurveCP, although it is a bit dated nowadays. I wouldn't often recommend people roll their own, but you could certainly do something simple with NaCl/libsodium directly. Maybe QUIC also fits the bill?)

The problems alluded to are, yes, solvable in python. But they also seem endemic in python systems.

When everyone who uses the tool uses it wrong, maybe it's not the user's fault.

(That said, I generally do think there's a time and place for python systems or web apps. That time is generally when speed and maintainability is significantly less important than flexibility)

gevent and gunicorn were good attempts to remedy a bad situation. async/await is the solution that the Python community is coalescing around. Even with Django, there are active efforts to support ASGI. [1]

[1]: https://docs.djangoproject.com/en/3.0/howto/deployment/asgi/...

I wasn't aware of this particular inefficiency, but gevent is still fulfilling its purpose for me very well, and I see no reason to change. I like lightweight threads and thinking in terms of background jobs and dividing up work instead of remembering what things to annotate and when. I use locks if I need predictability. I like Python because I can develop quickly with it, and I can do so even faster with gevent while still getting more than enough performance.

This is not trivial to set up correctly and other ecosystems already have sane, scalable, secure-out-of-the-box deployment schemes figured out.

It's an amazing library and a very unique way to write cooperatively-scheduled applications. Best of all it works with existing libraries and doesn't require special "asyncio" implementations from top to bottom. It's not a silver bullet, but don't fool yourself that asyncio is because it's been blessed.

This is not the whole story of the value of those languages. As the article even alludes to, at small loads or with lots of care this can be made to "work". But it is something that an engineer should know about them before picking them up and using a tool for something it isn't really good at.

[1]: I add this caveat because I don't think there's anything about dynamic scripting languages that makes them intrinsically difficult to thread any moreso than any other category of language, it's just that by an accident of history, they all come to us from the 1990s personal computer world, and they all spent at least a decade cooking and setting and building libraries and communities and developer skillsets before a serious need for threading was even on the horizon.

You don't necessarily need one; it depends on what kind of application you have and what its bottleneck is. If your application is network I/O bound, event-driven asynchronous I/O, which is what this article is describing, works fine, and a single process/thread even in a dynamic language like Python can handle a large volume of requests.

The specific issue this article is describing is due to a particular poor implementation of event-driven asynchronous I/O, not a general problem with the entire concept.

If your application is CPU bound, then yes, you need to use threads (or multiple processes), and you shouldn't be trying to mix event-driven asynchronous I/O with that.