[swift-server-dev] Prototype of the discussed HTTP API Spec

[Johannes Weiss]

> On 1 Jun 2017, at 6:02 pm, Paulo Faria <paulo at zewo.io> wrote:
> 
> Johannes.
> 
> > But we're trying to design a HTTP API that is implementable with reasonable performance and that I believe should be done by only offering async APIs.
> 
> We can definitely do both. We provide a base of concrete types and protocols like Open Swift to allow other frameworks to build upon it *AND* have a default official implementation using libdispatch which is Apple's official library.

Totally agreed, we can do both. The currently proposed API has already been implemented based on synchronous and asynchronous IO.
That is however only possible because the API offered to the web app is asynchronous only. If there were any synchronous calls in there it wouldn't work anymore (unless you play coroutine tricks).

And that's exactly the main reason why I think we should offer only async APIs to the user of the HTTP API: Then it can be implemented on top of blocking and non-blocking IO libraries (and also libdill/mill/venice). The reverse is not true.

-- Johannes


> 
> On 1 June 2017 at 13:53, Johannes Weiss <johannesweiss at apple.com> wrote:
> Hi Michael,
> 
> > On 1 Jun 2017, at 5:08 pm, Michael Chiu <hatsuneyuji at icloud.com> wrote:
> >
> > Hi Johannes
> >
> >>>
> >>> I think i need to clarify something: I’m ok with a asynchronous api that executes synchronously, for example if the api is something like [[ a. {  b() } ; c() ]], executes as [[ a(); b(); c() ]], it is totally fine since it’s just synchronous api with syntactic sugar.
> >>
> >> We actually have a synchronous implementation of the proposed API next to the DispatchIO one that we normally use. The synchronous one uses problem system calls and only services one request per thread. It's handy for unit testing and for specialised use-cases. The synchronous implementation only uses the following syscalls: open, close, read and write, that's it so nothing fancy.
> >
> > I think even exposing these apis to user will be good. No need for fancy support just include it and it will be good enough.
> >
> >>
> >> I think i need to clarify something: I’m ok with a asynchronous api that executes synchronously, for example if the api is something like [[ a. {  b() } ; c() ]], executes as [[ a(); b(); c() ]], it is totally fine since it’s just synchronous api
> >>
> >> ie. you use write as a blocking system call because the file descriptor isn't set to be non-blocking.
> >>
> >> Just as a side note: You won't be able to repro this issue by replacing the macOS `telnet` with the macOS `nc` (netcat) as netcat will only read more to the socket after it was able to write it. Ie. the implementation of standard macOS `nc` happens to make your implementation appear non-blocking. But the macOS provided telnet seems to do the right thing. You can use pbjnc (http://www.chiark.greenend.org.uk/~peterb/linux/pjbnc/) if you prefer which also doesn't have the same bug as `nc`.
> >
> > As I said both snippet of code are just sketches only for proof of concept. But I do missed on the kevent write one that’s for sure.
> >
> >
> >>
> >>>> I'd guess that most programmers prefer an asynchronous API with callback (akin to Node.js/DispatchIO) to using the eventing mechanism directly and I was therefore assuming you wanted to build that from kevent() (which is what they're often used for). Nevertheless, kevent() won't make your programming model any nicer than asynchronous APIs and as I mentioned before you can build one from the other in a quite straightforward way. What we don't get from that is ordinary synchronous APIs that don't block kernel threads and that happens to be what most people would prefer eventually. Hence libdill/mill/venice and Zewo :).
> >>>
> >>> Johannes, I totally agree with you. A asynchronous API is more intuitive and I agree with that. But since we are providing low level API for ppl like Zewo, Prefect, and Kitura, it is not right for us to assume their model of programming.
> >>>
> >>> For libdill/mill/venice, even with green threads they will block when there’s nothing to do,
> >>
> >> If you read in libdill/mill/venice, it will switch the user-level thread to _not_ block a kernel thread. That's the difference and that's what we can't achieve with Swift today (without using UB).
> >
> > I’m quite confused on this one, since a green thread, if that’s what we think we were referring to, can not enter kernel (It can, but when it enters what happened is that the kernel thread associated with enters kernel).
> > So you can’t switch to another user-level thread to not block a kernel thread.
> > AFAIK all majority OS(Liunx, FreeBSD, Solaris….) adopted 1:1 threading model instead of n:m, not sure about Darwin but I think it applies to Darwin as well according to an old WWDC video (I could be wrong), hence any user threads (except for green threads) are in fact kernel threads. Since kevent and epoll are designed to block when they should, I don’t think anyone could avoid blocking something.
> 
> Yes, Linux, macOS, FreeBSD and so on offer only a 1:1 threading model from the OS (Windows I think has Fibers). But libmill/dill/venice implement something akin to user-level threads themselves, you don't need kernel support for that at all (to schedule them cooperatively). Check out the man pages of setjmp and longjmp. Or check out Go (goroutines), Lua (coroutines), ... These are all basically cooperatively scheduled threads.
> 
> In other words: With setjmp() you can make a snapshot of the current environment and with longjmp() you can replace the current environment with one that you previously saved. That's like cooperatively switching something like a user-level thread/coroutine/green thread.
> 
> Run for example the code in this stackoverflow example: https://stackoverflow.com/a/14685524
> 
> This document explains it pretty well too: http://libdill.org/structured-concurrency.html
> 
> 
> >>> in fact all the example you listed above all uses events api internally. Hence I don’t think if an api will block a kernel thread is a good argument here.
> >>
> >> kernel threads are a finite resource and most modern networking APIs try hard to only spawn a finite number of kernel threads way smaller than the number of connections handled concurrently. If you use Dispatch as your concurrency mechanism, your thread pool will have a maximum size of 64 threads by default on Darwin. (Sure you can spawn more using (NS)Thread from Foundation or pthreads or so)
> >
> > Yes Kernel threads are finite resources especially in 1:1 model but I’m not sure how is it relevant. My concern on not include a synchronous API is that it make people impossible to write synchronous code, with server side swift tools, despite blocking or not, which they might want to. I’m not saying sync is better, I’m just saying we could give them a chance.
> 
> No one's taking anything away from you. Everything you have today will still be available. But I believe the APIs (which is what we're designing here) a web app uses should in today's world in Swift be asynchronous.
> 
> Of course to implement the asynchronous API, synchronous system calls will be used (eg. kevent/epoll). But the user-facing API that is currently proposed is async-only in order for it to be implementable in a performant way. If we were to put synchronous functions in the user-facing API, then we'll struggle to implement them in a performant way).
> 
> Imagine the function to write a HTTP body chunk looked like this:
> 
>   func writeBodyChunk(_ data: Data) throws -> Void
> 
> then the user can expect this to only return when the data has been written successfully and that the connection was dropped if it throws.
> But the implementation now has a problem: What to do if we can't write the bytes immediately? The only option we have is block this very thread and wait until we have written the bytes. Then we can return and let the user know if the write worked or not.
> 
> Comparing this to
> 
>   func writeBodyChunk(_ data: Data, completion: (Error?) -> Void) -> Void
> 
> we can now register the attempt to write the data and move on with the calling thread. When the data has been written we invoke the completion handler and everything's good.
> 
> 
> 
> >>> And even if such totally non-blocking programming model it will be expensive since the kernel is constantly scheduling a do-nothing-thread. (( if the io thread of a server-side application need to do something constantly despite there’s no user and no connection it sounds like a ghost story to me )).
> >>
> >> what is the do-nothing-thread? The IO thread will only be scheduled if there's something to do and then normally the processing starts on that very thread. In systems like Netty they try very hard to reduce hopping between different threads and to only spawn a few. Node.js is the extreme which handles everything on one thread. It will be able to do thousands of connections with only one thread.
> >>
> >
> > The kernel has no idea is a thread have anything to do unless it sleeps/enterKernel, unless a thread fits in these requirements, it will always scheduled by the kernel.
> 
> but that's exactly what epoll/kevent do. They enter the kernel and tell the kernel what the user-space needs next.
> 
> The thread is now scheduled only if an event that kevent/epoll are waiting for turns up. And when kevent/epoll then return, most of the time the user space handler is submitted as a callback.
> 
> 
> > I’m saying, if there exists a real non-blocking programming model, defined that by “never call any ‘wait’ system calls’, than any IO threads of that model must constantly poll the kernel, hence such thread _cannot_be_scheduled_on_demand since the thread itself has no idea if it has anything to do. The only way to have an IO thread to do know they have to do something, they will either need
> >
> > 1) An external listener call blocking event api and poke the IO thread on demand
> > 2) The IO thread has to constantly poll the kernel
> > 3) An external listener polls the kernel constantly and poke the IO thread when ready.
> >
> > 2 and 3 are the do-nothing-thread I’m referring to, they are running, polling, wasting kernel resources but not actually being productive (when there’s no connection).
> 
> Nothing is in a tight polling loop. We run handlers as long as we can and when all handlers have run, kevent/epoll is entered again, done. Often you just spawn as many threads as you have CPUs and all is good. These threads are mostly sitting in kevent/epoll and as soon as some file descriptor becomes readable/writable the respective handler is invoked.
> 
> That's what DispatchSources do, what Node.js does, what Netty does, ...
> 
> 
> >> You should definitely check the return value of write(), it's very important. Even if positive you need to handle the case that it's less bytes than you wanted it to write. And if negative, the bytes are _lost_ which happens all the time with the current implementation.
> >>
> >> Anyway, to fix the EAGAIN you'll need to ask kevent() when you can write next.
> >
> > It was suppose to be a proof of concept sketch work. As mentioned in the comments of the code it was assuming to satisfy one single client. Now I’ve improved it so it handles multiple clients while remain synchronous and non blocking. EAGAIN is the only “error” will raise if you consider it as error but for me it’s part of the non blocking IO.
> 
> well, it means the write hasn't happened. You will need to do the write again and that is normally done when kevent/epoll tell you to. And that's what I mean by inversion of control.
> 
> 
> >> Foundation/Cocoa is I guess the Swift standard library and they abandon synchronous&blocking APIs completely. I don't think we should create something different (without it being better) than what people are used to.
> >>
> >> Again, there are two options for IO at the moment:
> >> 1) synchronous & blocking kernel threads
> >> 2) asynchronous/inversion of control & not blocking kernel threads
> >>
> >> Even though I would love a synchronous programming model, I'd chose option (2) because the drawbacks of (1) are just too big. The designers of Foundation/Cocoa/Netty/Node.js/many more have made the same decision. Not saying all other options aren't useful but I'd like the API to be implementable with high-performance and not requiring the implementors to block a kernel thread per connection.
> >
> > To be honest I will choose 2 as well. But we are in not a 2 choose 1 situation. The main difference between we and netty/node.js is that ppl use them to, write a server, what we do is, writing something ppl use to write something like netty and node.js. So it is reasonable to think there’s demand on a lower-level, synchronous api, despite the possible “drawbacks” they might encounter.
> 
> This is the HTTP group so people will only write web servers with it, the API that was proposed it definitely not meant to implement anything netty or node like. It's to implement web apps in Swift.
> 
> There is however also a Networking/Transport group which will be more low-level than this (I assume) and there we do need to consider the lower-level APIs. And those will contain blocking system calls, namely kevent/epoll (if it won't be based on top of DispatchSources/DispatchIO which do the eventing out of the box, obviously also implemented with kevent/epoll).
> 
> 
> > Maybe we have some misunderstanding here. I’m not saying a synchronous api that happens to be able to handle a vector of sockets in single call without blocking anything, I’m saying a synchronous api that can just do one simple thing, which is, read/write in a synchronous way despite block or not, if it will block, just let them know by throwing an exception, the api call itself, will not block anything that way.
> 
> there may well be a misunderstanding here. No one wants to take all synchronous APIs away from you. They are available in Swift today and will remain there tomorrow.
> 
> But we're trying to design a HTTP API that is implementable with reasonable performance and that I believe should be done by only offering async APIs.
> 
> 
> Thanks,
>   Johannes
>