Most of our stack is built around gevent, it works great just about everywhere and requires little to no thought about the asynchronous nature of the code. Performance has been incredible compared to the old thread based code it replaced. Asyncio seems nice in theory, being baked in and all, but we tried it with some sample apps and it just requires too much thought to stay asynchronous, it's a takes a total paradigm shift like twisted. It's easy to slip up and have parts of the code become synchronous and the code gets hairy. We decided to just stick with what we know.
I've been using Python for more than 12 years now. Used threads, then Twisted, then eventlet, then gevent.
Switching to Twisted was hard but oh well, there was nothing else that did the job right for lots of concurrent requests. Unfortunately it also created a separate library system, fragmented and cumbersome. You'd want a library, you find it "oh but it doesn't return deferreds it blocks the current thread". Or vice versa when using threads, I'd find the perfect library to parse my protocol but it returned deferreds only. Then eventlet and gevent came about it was like a breath of fresh air. I remember refactoring a largish code base from Twisted to eventlet and it was something like a 50% reduction in code size. Then gevent came about with more improvements and so on. Then async came, and now it's like back to Twisted. Find a library but well, it return the async thing not a result, so it has to live in a separate ecosystem. You have write async this async that everywhere.
Also, maybe if async came 5-7 years ago it would have taken over but now once people have to rewrite everything with async they'd also start looking at other languages: like Elixir and Go which somehow managed to avoid "async everywhere" and still do the right thing. Python almost had that too with greenlet based libraries. Embracing that would have probably made more sense, but oh well...
Gevent is just a little too good at what it does; the fact that it's performant and lightweight and easy to pick up and start moving on but still scales well... just makes it hard for asyncio to disrupt, especially when greenlets beat them to the table by a pretty huge margin.
I'm assuming your stack looks like mine: gevent has monkey patched everything.
1. fork'ing just doesn't work properly; something about the event loop in the child hangs.
2. You can't easily spawn an honest to god hardware thread anymore. (Since the typical call is monkey patched!) Yes, you can get the original threading module by getting it out of gevents hands, but…
3. …it doesn't compose well. Libraries, in order for gevent to play nice, need special "if gevent then …" sections.
No need for forking or threads, that's the point of async. Yeah there's limitations when CPU bound, but that's python. Fire up a few processes and load balance.
#3 hasn't been an issue for us, monkey patching works well. We've switched to pure python versions of a few libraries. When not available use a pool of instances.
The problem is with some specific features like select.epoll. gevent doesn't support that and will break such code. This includes socketio libraries - we use third party tools like pusher.com and all of them break with python 3.
E.g. https://github.com/benoitc/gunicorn/issues/1494
Gevent explicitly says :
"Given that epoll is more sophisticated than select or poll (e.g., support for edge triggered and level triggered events), I don't know if we can emulate that close enough with the capabilities that libev exposes (we'd need to pass the stdlib tests when monkey-patched), and I don't know what the performance characteristics would be (and if the subtle differences would break apps) "
I don't see a path forward for gevent with these issues in place.
Of course you need threads even with async. When you code blocks for CPU reasons, like handling a heavy HTTP request with lots of calculations, you don't want to stop serving other requests.
It depends. If the calculation can be split into nice chunks you can throw in some gevent.sleep(0) to prevent them from hogging the process. We use python for handling web requests, anything heavy tends to be handed off to another process. It's rare for us but yeah, sometimes you need threads.
First, forking is "firing up a few processes". Yes, you can do that prior to launching your process, and then let gevent just have the process, but you have to be able to do that, and it just doesn't make sense in all contexts. This is what I meant by gevent fights composability.
> or threads
There's a limit to what a single CPU can do, and eventually multiple hardware threads can buy you actual tangible results.
> Yeah there's limitations when CPU bound, but that's python.
My point being that it's very difficult to get to CPU bound when your greenlet library prevents you from taking advantage of threads or processes. "but that's python" implies to me that you think Python is hamstrung to a single thread, and that's not the case: Python is quite capable of not only hardware threads, but performing computation on hardware threads in parallel, simply because not all of my code is in Python. Sometimes, a component knows it needs an honest-to-god HW thread, and the monkey patching fights that.
HW threads additionally lets you run multiple-long running computations; even if they're mutually bound to a single CPU, preemptive multitasking at the OS level (and Python's interpreter occasionally releasing the GIL) will ensure they all get time. In a cooperative greenlet, if you do not yield, you starve the other threads. A library making a long running computation again needs intimate knowledge that at the higher level, greenlets are in use: this breaks layers and mixes responsibilities (greenlets) in where it doesn't belong (some library calculating a thing). Greenlets do not compose.
> We've switched to pure python versions of a few libraries.
And you pretty much have to; if the library called out to C, it would block. But this fights good engineering practices: you can't just implement your library once in the language of choice, and then bind it into another language.
> When not available use a pool of instances.
This depends on what the library does; a "pool" does not necessarily save you. For example, if the library blocks, you need a threadpool (a real one) to process inputs to it.
Libraries exposing an event loop, perhaps with some basic primitives, have a hope of communicating across things like language barriers.
Excellent points, thanks for the thoughtful response. I suppose it depends on the use case and certainly there are limits to gevent, but for our use case--serving and routing web requests, they rarely come up and are easily worked around. In some cases these pain points can be avoided by only patching the libraries you need rather than monkey.patch_all(), other times we can just spawn another process to do the work. I can appreciate that there are workloads where asyncio is better suited than gevent, but gevent is just so easy.
We communicate across languages quite a bit using sockets. A given system may have 20 processes that communicate with each other and the outside world. Most but not all of the python processes use gevent. Composing software with sockets scales well and avoids compatibility issues. It's rare for our python processes to be cpu (or core) bound, but it's easy to be IO bound and that's what gevent solves.
Question from a mediocre programmer: what are the advantages of this over, say, a messaging queue and microservice architecture for an application (i.e. Web service built on flask, but the details don't really matter)
For example, would it be a bad idea to put my email module on an async function as opposed to in a different application?
They aren't really meant to accomplish the same goals, async and microservices. What I mean is, these are architectural choices. While at some level they may look interchangeable, they really aren't.
There's no 100% correct way to say when to use one or the other, but here's one way I would think of them.
Microservices are used to help define the architecture of your system overall. This is more ... high level.
Async would be more at a process level.
So, for your example, you could have a flask microservice that exposed an interface that accepted a list of email addresses and a message to send to each. The request handler for that could then call a method using asyncio to send the emails. This would allow the microservice to remain responsive and wait for more requests.
You're asking how modular your system should be. That's like asking whether a 20-line function should be split in two. It depends. Is the code easy to understand?
Yes, it is easy to understand (and even more than that - yes it can be in its own module/blueprint anyway). My previous thinking has been that if I have a message queue and separate application then I get the benefits of failure reporting through the queue as opposed to potentially missing something if I get a failure through async. Although I guess they aren't insurmountable.
Then, inevitably, you have growth so that one email function it may begin with, but later you're doing other fancy things and even though you may just want to send a single email at the start, later you're making database calls and doing other bits and pieces to make the template more fancy and personalised etc. So, better to start the architecture from the start in a slightly more complicated fashion..?
No. Start with "bad" architecture and refactor as needed.
Another layer of abstraction solves any problem. If you don't have a problem yet, then abstraction is unnecessary complexity.
BTW, writing your own async code may actually be more complicated than just having a message queue and an email program. It depends on how much trouble your message queue is giving you. Can you use a Sqlite table instead of a "real" message queue?
I'd have the web-server append "messages" to a sqlite table and have the emailer read from the same table. You can use cron to run the emailer every X minutes. If the table gets huge, use cron to run an archiver that prunes old messages during low-usage hours. Sure, the pruning will temporarily lock the database. If that gets to be an issue, switch from SQLite to PostgreSQL. If that architecture starts to fail, that's a good problem to have.
You could argue that writing to the queue should be done asynchronously inside the web-server, but, again, wait until your customers complain that the server isn't responding quickly enough.
In some circles, this idea is known as "incremental architecture". The idea is that you should architect and add abstraction layers to solve problems you have /today/, rather than solving problems in the future.
One big reason for this is because we're often wrong about what problems we'll have in the future, or in what way they will manifest. Moreover, often the "bad" solution will work for far longer than we expect, and the day that we're anticipating where we'll be having problem X might never even come.
What's worse is that in the meantime each layer of abstraction adds a tax to how easy it is to fully understand and work with the code, to change things and add new things.
What you do want to do is to write code that is /open to/ being abstracted. E.g. keep things structured nicely, start with a basic function (named similarly) or a few related functions before putting in a full blown class, name things well so it's easy to find them while refactoring, etc.
I've done this with iteration 2 (current version) or my backend but now I'm not a one man shop and have a team (well one other junior backend and a senior mobile). And we are architecting for some pretty major new features so it's interesting trying to strike a balance with both what I've learned along the way and what it seems 'best practice' is when it comes to python
Careful with where you read "best practice" from. Some folks like to brag about their new high-scalability architecture, but it's way over-engineered for most software.
The pattern I described of using SQLite tables to mimic a message queue will be very easy to switch to PostgreSQL or MySQL. If you hit serious scale you can switch to something like Kafka or Kinesis. That'll be a bit more involved, but hopefully your revenue and team scaled to match the compute needs. When you switch to Kafka, you can use Spark or Storm instead of cron. If you switch to Kinesis, you might choose AWS Lambda. So you're not writing yourself into a corner.
I've always wondered about this: doesn't the OS read buffer pretty much keep the whole thing in memory anyway, if you have memory to spare? Does it make that much of a difference?
In the beginning I also thought it would be about multiprocessing, and asynchronous execution. But actually it's just a programming style for linear, single thread execution.
Personal opinion: It's not needed if you don't want to code fancy. In some regards its even disappointing because the name promises to solve a generation old Python problem, which it actually doesn't even tackle.
There will be interesting community impacts from python's asyncio -- I'd venture its going to create a lot of fragmentation in the realtime application ecosystem. The issue seems to be that there isn't an agreed way to consume derivative apis as the flow control and interfaces around asyncio are super clunky. Further, the all-or-nothing nature of asyncio (don't block the thread) means a significant portion of applications will have to have some rewrite to support asyncio (webframeworks, sqlalchmy and such).
There are already two fairly large ecosystems around asnycio network libs + http frameworks which are large rewrites of existing libraries (https://github.com/klen/muffin and https://github.com/aio-libs). It seems like having a whole framework/ecosystem for loop lifecycle management, async libs, state/comm, etc will unfortunately be needed. I hope we find a way to work around this although I doubt there is a good way.
> It seems like having a whole framework/ecosystem for loop lifecycle management, async libs, state/comm, etc will unfortunately be needed.
... twisted has been around for ages for exactly this reason
as a 'framework' unfortunately this means it's not 100%
ported to python3 yet, but I suspect (not a dev, etc) that as
twisted becomes more mature and 2.x dies off, there might be more
overlap between these two things.. and perhaps other similar projects
Looking at the http://www.dabeaz.com/coroutines/Coroutines.pdf recommended in the article, coroutines remind me of partial functions in Python. Can folks who know more about Python than I do explain when to use coroutines vs partial functions? Is the difference only in the "generator" part of it?
A partial function in Python is just a wrapper around the underlying function, that automatically supplies whatever arguments you passed in to functools.partial. It doesn't change anything about how the function is executed.
A coroutine gets executed differently from an ordinary function (or an ordinary generator, for that matter); when you call it using the await syntax, the Python interpreter can choose to run some other code while it is waiting for your function to return a result from I/O (for example, a network request). A normal function call doesn't allow that. (A generator might, if you use yield from and do a number of other things, but you have to do them by hand, whereas the await syntax does it all for you automatically inside the interpreter.)
This is slightly misleading considering async funcions are just simply generators and nothing but parsing the keywords is actually implemented in the interpreter. The whole logic is in the asyncio library.
This is a great talk where the presenter goes into how asyncio works in detail and builds a mini version of the asyncio library:
> async funcions are just simply generators and nothing but parsing the keywords is actually implemented in the interpreter.
As I understand it, async functions have an extra flag that ordinary generators don't have, and the interpreter does some different things if that flag is present. But I have not dug deeply into the source code.
Oh sure. I was just discussing in the context of the example given in David Beazly's slides.
def grep(pattern):
print "Looking for %s" % pattern
while True:
line = (yield)
if pattern in line:
print line,
You initialize the coroutine with g = grep("python"). Isn't this step same as having a partial function with it's `pattern` argument with value "python"?
def grep(pattern, line):
if pattern in line:
print line,
And then:
g = partial(grep, pattern="python")
g(line='python in a line')
I'm sure this is a very simple example of coroutines, but if this is all I want to do (setting aside async vs sync for a moment), why would I use a coroutine here instead of a partial function?
EDIT: As I see more examples of coroutines especially the pipelining use cases, I was wrong and agree that it's completely different from partial functions.
I think I see where the confusion is coming from. A coroutine can be nexted repeatedly. A partial function can only be called once. The coroutine in the example is just a toy, so it's effectively the same as calling a partial many times in a loop. A more interesting coroutine might encapsulate state that changes across next'ing. Yes, other kinds of functions can change state across calls, too... It turns out many techniques can accomplish the same thing.
For me, I didn't understand async/await until I saw it in Python. The C# explanations never worked for me, and it wasn't until after I saw the Python version that I realized why: C# uses an implicit thread pool so you can run async functions directly. With Python, you have to manually add your awaitables to an event loop and then start the loop, but C# just magicks all the glue into place.
This is one of the places where the Zen of Python comes to mind:
> Explicit is better than implicit.
Edit: To see what I mean, compare Microsoft's async/await example to Python's:
The Python example tells you how to set up an event loop and add awaitables to it, while Microsoft's "complete example" just has you add async event handlers to a window without showing you anything about setting up an event loop. It wasn't until I saw the event loop being set up in Python that I finally understood what async/await did.
The most confusing thing in Python is that there is more than one way to do it. :)
Python asyncio library supports coroutines, callbacks, futures, thread-pool executors... All have similar APIs and dissimilar semantics. For instance, canceling a coroutine is different than canceling a future and different than canceling a thread. (I've heard asynchronous exceptions are bad, but this is just an example.)
Error handling is difficult. You can't observe an error if you don't remember to write code to observe an error.
Also confusing is that coroutines in Python can be driven by any event loop. Which is great! You can write your own. But that will be even more confusing. Erlang has one OTP, Python will have many.
The most confusing thing in Python is that there is more than one way to do it. :)
Exactly. Python had a multiple thread and blocking I/O model. That's well understood. Javascript had a single thread callback-based model. That's well understood. When you have both concepts in the same system, things get much more complicated. It's even worse when it's a retrofit.
Consider Python's Queue module. If you call Queue.get(), your thread can block. But the async system doesn't understand that kind of block. (Is that right?) So you've blocked all the async stuff running off that thread.
Go has lightweight threads too, called goroutines. But because it wasn't a retrofit, all lock mechanisms are compatible with lightweight threads. When a goroutine blocks on a lock, its underlying thread is re-dispatched. Python 3.6 isn't set up that way, because async operations were retrofitted in parallel with the existing machinery.
As someone who has NOT gone low level into the theory of programming languages (bailed out on CS a decade ago), I wonder what the practical difference is between this and threading or multiprocessing.dummy.
I understand the idea of blocking I/O and skipping that until a function/thread returns.
The practical difference is that with async I/O, there's less overhead because you don't need a separate thread for each blocking I/O operation. It's the difference between using a single-threaded event loop to demultiplex between 10,000 sockets, and using 10,000 threads to read one socket each. The former strategy has way less overhead.
Of course, many high-performance, I/O intensive applications like web-servers can effectively make use of a combination of multiple threads and I/O demultiplexing as well.
"what the practical difference is between this and threading"
The biggest practical difference is synchronization-free programming. In other words event loops are already always synchronized and you can do whatever you want with shared memory knowing that no other code is going to interfere until you explicitly stop and give up control. This allows for more reliable and more complex concurrent applications with less effort.
> I understand the idea of blocking I/O and skipping that until a function/thread returns.
You understand it better than you realize. The idea of a "thread" is that it's a sequentially run context for computation with a bunch of context carried with it (a stack, references to a heap). This sequence of computations can be pre-empted.
When OO systems offer "promises" what they're doing is essentially freeing a thread from the continuity of the computation and leveraging closures or object scope to hold the context. Instead of a thread representing the computation, now threads pick up computations and work on them until they stop for I/O, then resume them when the underlying async I/O layer has work for them.
I can tell you about it in terms of monadic computations or how it relates to other concurrency strategies if you'd like, but it will become more theory-dense rather quickly.
I don't think it confuses anyone that's worked on lower levels, particularly listening on sockets or kernel callbacks. It's a new concept to a lot of upper level programmers that used to be isolated from this stuff by threads.
Try writing code with JavaScript promises in the browser or on NodeJS. There is a ton of literature and work there to bring people on board. C# has some good docs too.
I think it's because async. means nothing if you don't specify how, where and when something is async. For starters you can't use languages that don't share memory between threads if you wish to see any improvements by adding async. threading to your non-blocking network code. So Python, Go, Ruby, Javascript, Erlang are all out of the argument when talking about async.
Only C (C++, etc.) and Java (C#, etc.) families of languages that share memory between threads can scale better with async. on multicore processors when it comes to network related performance.
Edit: Please comment if you downvote.
Edit: Now I can't comment, the site tells me I'm writing too fast.
Important point: Promises, ContT monadic constructions[1], Threads, and Actors all are just variations on how to structure sequential computation with asynchronous events.
You can prove this to yourself via construction: you can write an interpreter that simulates threads (and even pre-emption) with Promises, Actors, or a more formal monadic style. You can repeat this feat for any single member of the set and if you do it carefully that's a proof by construction.
Essentially what you choose to implement is arbitrary and should be informed by your underlying performance requirements coupled with your preferences.
And, as people have noted: your post seems confused. Everyone benefits somewhat uniformly from asynchronous operations. Java's implementations are often a little bit more tunable because the programmer can provide the threadpool for async workers.
Of all of the languages you named, Erlang is the most distinct as it offers an abstraction (and cost) of 1 heap per actor from the perspective of the garbage collector. This decision (made in a time when GC algorithms were quite a bit worse than they evolved to be) lead Erlang to something they want to keep: an abstraction that makes remote and local computation identical. Other languages that follow in Erlang's footsteps (Pony comes to mind, yay Pony!) do not do things this way.
[1]: Okay EitherT ContT. Sure. Get picky.
P.S., You cannot comment because your overall karma is too low, putting you in a rate-limited category. It generally clears up in a few hours.
However, I do not think there is much more of value to say other than to thank that one poster for a few pretty cool papers!
> you can't use languages that don't share memory between threads ... Python, Go, Ruby
> trust me, I'm not confused
Sorry but you really are confused about this. You said that Python, Go and Ruby don't have shared memory between threads, but they simply do. All of these languages's concurrency and parallelism models are fundamentally shared-memory, and all even allow completely unsynchronised access to shared memory if you wanted to do that.
If you won't believe me on that, here are five peer-reviewed papers from reputable venues on Ruby alone that talk about the shared memory model it has. Either all of these experts and all the reviewers are mistaken, or you are.
B. Daloze, S. Marr, D. Bonetta, H. Mössenböck. Efficient and Thread-Safe Objects for Dynamically-Typed Languages. In Proceedings of the ACM International Conference on Object Oriented Programming Systems Languages and Applications (OOPSLA), 2016.
C. Ding, B. Gernhardt, P. Li, and M. Hertz. Safe Parallel Programming in an Interpreted Language. In Proceedings of the First Workshop on the High Performance Scripting Languages, 2015.
L. Lu, W. Ji, and M. L. Scott. Dynamic Enforcement of Determinism in a Parallel Scripting Language. In Proceedings of the 35th Conference on Programming Language Design and Implementation (PLDI), 2014.
R. Odaira, J. G. Castanos, and H. Tomari. Eliminating Global Interpreter Locks in Ruby through Hardware Transactional Memory. In Proceedings of the 19th Symposium on Principles and Practice of Parallel Programming (PPoPP), 2014.
W. Ji, L. Lu, and M. L. Scott. TARDIS: Task-level Access Race Detection by Intersecting Sets. In Proceedings of the 4th Workshop on Determinism and Correctness in Parallel Programming (WoDet), 2013.
> R. Odaira, J. G. Castanos, and H. Tomari. Eliminating Global Interpreter Locks in Ruby through Hardware Transactional Memory. In Proceedings of the 19th Symposium on Principles and Practice of Parallel Programming (PPoPP), 2014.
Well thanks for this, can't wait to sit down & read it.
Python, Go and Ruby do not have the ability to run multiple real OS threads that can run on separate cores and share process memory at the same time.
They might try to work around that, but they will never be able to run one common task to/from many sockets on all cores of one machine at the same time.
Yes, they do. Python does. I'm sorry but your wires are getting crossed somewhere. You also completely ignored the comment you where replying to and the sources he cites for you and repeated essentially the same thing you said before. This is quite rude.
Python has a GIL, so only one thread can interact with the Python interpreter at one time. You can have any number of threads running at one time if they release the GIL and don't touch the interpreter (i.e if they are doing IO or calling a C function).
They, however, are still real OS threads, running (waiting for a lock) on separate cores and sharing process memory. That's the definition of a thread. They are not fake threads. They share memory. They run on separate cores (perhaps).
This is not technically true, but often is practically so. No matter, it has no bearing at all on the concept of "async" as modeled or in contexts you've discussed.
I'm not sure why you're trying to salvage this or what your endgame is here but please reconsider.
Would you prefer I use the phrase "completely incorrect about the memory models of the majority of languages you named and the erroneous conclusions you reached from them?"
That is certainly more accurate, but seemed needlessly confrontational in my first draft, so I opted for the more generic phrase "confused post", which implied maybe the post itself was simply jumbled up and didn't clearly offer your intent to the audience. You did ask after all.
I was trying to give a non-confrontational way to explain why I almost downvoted your post for being very misleading, and opted instead to respond with a slightly more correct post.
HN needs less fighting and more knowledge.
I thought what you were trying to say was, "parallelism is not the same as concurrency" but that it got jumbled up. That is a correct statement.
Ok, didn't know that, can you point to some pages that explain how you can reference the same memory from two threads in the same process with these languages?
Here's a snippet of Go code, with comments, which spins up different goroutines (lightweight thread primitives) to read and write values in a shared map: https://gobyexample.com/mutexes
Can goroutines run on separate cores? Because I'm pretty sure go explicitly decided to not share memory between threads because google thinks async. is too hard to understand for programmers...
You probably mean channels - yes they let you avoid dealing with synchronization of shared memory access explicitly, but they still use mutexes underneath, so you can use those directly if you want (instead of/along with channels).
Yes, goroutines run on separate cores. Go uses an M:N threading model. goroutines (M) which are ready to run are scheduled on a thread pool of POSIX threads (N) which I believe is sized to match the number of cores in the processor. Nothing stops those goroutines from racing on shared memory. However, there is a convention that before accessing some shared memory, goroutines synchronize themselves using channels. ("Do not communicate by sharing memory; instead, share memory by communicating.")
Are you sure? I thought nim could use shared memory semantics if you wanted to? My understanding was it (cleverly) makes a subset of the language for concurrent work that requires you to be explicit about that activity.
Sure, but then you are not having all cores contributing to the same solution. Each nginx request is then distinct from every other request. And really then it's easier to scale with separate machines like google does in it's completely synchronous environment.
So you start a number of workers and distribute the load among them. At least that's one way to solve the problem in Perl/Python/Ruby. You can contribute to the same solution by passing data to a master process, or using shared memory.
In Erlang and Scala you can solve it even more ellegantly using actors and chanels. Go has goroutines and channels. All of these can span multiple cores.
what is good about asyncio, it allows to replace default event loop. I am working on asyncio event loop based on tokio-rs (rust). it could bridge asyncio with code written in rust. so potentially, it should be possible to mix python and rust code on higher level than just simple c-extension.
Apologies for digressing a bit, but is there something like Asyncio for PHP? I've been looking into this for quite some time, but the standard answer I get is using queues. Would love to hear about any latest projects related to this.
