On Fri, Dec 18, 2015 at 2:00 PM, Mark Lavin <[email protected]> wrote:
> Anssi criticisms are fair and I feel that some of these responses are > glossing over the details. > I'm sorry if it comes across like that - a lot of these are things I've been considering for a while and so I can forget to provide context. > You've claimed this is the same or equivalent to a forked worker model but > it isn't because there is no process management/link between the interface > and worker and because you've chosen to make this network transparent. As > much as you'd like to claim this isn't like Celery, the same issues that > exist when trying to (ab)use Celery for blocking RPC calls is what you will > have here. > I never claimed this isn't like Celery; it is quite a bit like it, but with specific changes (no guaranteed delivery, no single-response mechanism) that make it better at throughput. > > > Then the response is dropped, the same way a WSGI worker drops a > connection if it dies. > > It isn't entirely the same because in this channel case the worker doesn't > know the client/interface dropped the connection. It's still working hard > to generate a response which will sit in the response channel (Redis, > memory, etc) until it expires (assuming that all backends expire channel > messages). That doesn't happen in the current WSGI interface. > True, though from what I understand of the WSGI spec you also don't know the client has disconnected until you try to write out content to it. Most views would still run the entire thing and make a response, only to drop it. There is also a http.disconnect message type planned to be implemented for when a client disconnects before the response is entirely read, but that's more for long-poll usage where you want to keep track of who has an open connection. > > > All Django requests already involve multiple round trips to database > servers, and the Redis backend at least is much quicker on the processing > side. > > This isn't true. Database round trips are not a requirement for Django's > current architecture. I'll concede that many views touch the DB but that's > a choice of the developer, not Django's. When using channels the views will > still need to make the same DB calls, though that processing happens at the > worker. Putting Redis in between doesn't make it faster. > Right, and channels will not be a requirement for Django's future architecture. They'll just be a thing that I expect most people to turn on as they provide a feature set a lot of types of sites need. And I never said it would be faster - you'll see me repeatedly say channels provides no performance gain - at best, it might help smooth response times with the way the workers load balance jobs based on when they're free. > > > You could make the same argument about not having separate load > balancers and nginx serving static files - after all, you're just adding > another network roundtrip to send traffic onward from the other server to > Django. > > There are notable differences here. These are persistent HTTP connections > and don't suffer from the same problem of client/server drops previously > noted. They are also cacheable in a known way allowing round-trips or > bandwidth to be avoided. Many Django applications will continue to use load > balancers with channels. It can't be denied that channels introduce two > more network round trips that didn't exist before. Trying to paint this as > "we already talk over the network, so what's a couple more" is not a > compelling argument to me. > Sorry, I think my tone came across wrong there. I'm more just saying that it's equivalent to adding an extra layer of that kind of infrastructure, in that it also uses persistent connections and should only be a few milliseconds of delay, and for most Django sites a few milliseconds is perhaps a percentage point of their response time. Again, if someone wants higher performance, they don't have to use channels and can just connect directly. You seem to be assuming I'm here to foist a brand new middle layer on everyone; I'm not. I'm here to make one that fits neatly into Django, that I think most people will want to turn on, and that provides a lot of value in exchange for a slight round-trip performance hit - my goal is sub-5ms, and preferably sub-3. If it starts being 10/20/30 milliseconds of cost, then we'll have to change our approach until it's acceptable. If you don't want the new features and the resulting change in stack, Django as it is now will be there for you, but if you're that sensitive to performance then Django maybe isn't for you already unless you're heavily modifying it. Andrew > > On Fri, Dec 18, 2015 at 5:01 AM, Andrew Godwin <[email protected]> > wrote: > >> >> >> On Fri, Dec 18, 2015 at 7:34 AM, Anssi Kääriäinen <[email protected]> >> wrote: >> >>> I have a gut feeling this isn't going to work that well. The reasons >>> include: >>> - Backwards compatibility: how is a large site going to upgrade from >>> 1.9 to 1.10? >>> >> >> None of the core view API will change. A 1.9 codebase will boot and work >> on 1.10/channels with no code changes. >> >> >>> - Complexity of setup. >>> >> >> Could you elaborate? If you don't want it, you don't need to configure >> anything, and if you do, for most people it's just getting a Redis server >> running and pointing a setting at it. >> >> >>> - Error conditions: for example, what happens when an interface server >>> sends a request to worker, and then dies (that is, the response channel has >>> zero listeners). Similarly for chunked messages. >>> >> >> Then the response is dropped, the same way a WSGI worker drops a >> connection if it dies. Chunked responses will time out after a certain >> period and be dropped too, in the same way that a deadlocked normal server >> would. >> >> >>> - Does the architecture really scale enough? The channel backend is >>> going to be a bottleneck, it needs the ability to handle a huge amount of >>> data and a huge amount of individual messages. In particular, the request >>> channel is going to be contested. We know classic http scales, but is the >>> same true for interface server architecture? >>> >> >> I believe it will - have you read through the sharding and scaling plan >> in the docs? Channels is carefully designed to have no state in anything >> but interface servers, all workers handling all message types and queuing >> of messages exactly so you can scale horizontally; you can divide a very >> large site into several clusters of interfaces and workers fronted by load >> balancers, and each cluster would have multiple Redis (e.g.) backends with >> requests equally sharded across them using consistent hashing. >> >> >>> - Performance. Each request and response needs two additional network >>> roundtrips. One to save to the channel server, one to fetch from the >>> channel server. If the messages are large, this adds a lot of latency. >>> >> >> All Django requests already involve multiple round trips to database >> servers, and the Redis backend at least is much quicker on the processing >> side. You could make the same argument about not having separate load >> balancers and nginx serving static files - after all, you're just adding >> another network roundtrip to send traffic onward from the other server to >> Django. >> >> >>> - Untested architecture: does any big site use this kind of >>> architecture for all http handling? >>> >> >> I agree with you here - see below for my justification. I've seen it used >> for other things (data update networks, service calls), but not direct HTTP. >> >> >>> >>> A realistic test for this is to push a scalable amount of scalable sized >>> requests through the stack. The stack should recover even if you shut down >>> parts of the network, any single interface server, channel backend server >>> or worker server. Of course, when a server or a part of the network >>> recovers, the stack would need to recover from that. Compare the >>> performance, simplicity of setup and ability to recover from error >>> conditions to a setup with only classic Django http servers. >>> >>> I'm sorry if this feels negative. But, you are planning to change the >>> very core of what a Django server is, and I feel we need to know with >>> certainty that the new architecture really works. And not only that, it >>> needs to be at least as good as classic http handling for existing users. >>> >> >> That's why a decent part of my proposal to Mozilla for funding was to >> help us fund hardware and time for extensive performance and scale testing. >> I've seen this architecture work at scale before for non-HTTP traffic, and >> I believe that it will work as well for HTTP and WebSockets. >> >> Don't get me wrong - I don't believe this is a magical panacea to solve >> all problems, and we're going to have to do plenty of testing and >> development work to get the solution to the level of existing HTTP >> handling, but remember, it also brings positive results: >> >> - Downtime-less code deploys (if you stop workers, requests will just >> wait for new ones to appear until they hit timeout) >> - Ability to add and remove processing capacity live without >> loadbalancer reconfiguration >> - Lets you run different parts of the site on different Python runtimes, >> if you want (e.g. one part on PyPy, one part on CPython 2, one part on >> CPython 3) >> - Background task processing >> - And, of course, WebSockets/HTTP2/long-poll HTTP/other >> non-request-response protocol support >> >> It's never going to be a solution that works for everyone, but the whole >> nice part about channels is that, like all the best parts of Django, you >> can just ignore it and not use it if you don't want it; we're not going to >> get rid of WSGI support, and the default shipping version in 1.10 isn't >> going to make you find a Redis server before you can even boot it up; it'll >> just work like it does now, with extra flexibility there if you want to go >> turn it on and read through the next part of the tutorial/docs. >> >> I don't expect to get this past the community, core and technical board >> and approved into a release until it's proven itself to run and work at >> scale, and I already have several offers of testbeds to help prove this out >> with realistic web loads, which we can combine with synthetic load tests. >> >> Put it this way - I do not see any other way to handle WebSockets that is >> as feasible as this. Most solutions either require us to run the whole of >> Django in an async Python environment, which comes with its own set of >> issues, or they're more stateful proxy servers or run-alongside-servers >> that don't seem to have a story for scaling them to hundreds of thousands >> of connections without blowing up every packet received into HTTP requests. >> >> I think Django absolutely has to adapt to the modern web environment and >> move away from just rendering templates when browsers request them, and >> this to me is part of that. If there are other solutions to the same >> problems I think we should consider them as well; I've just not run across >> any that work as well in the two years I've been planning this out before I >> brought it out to be talked about. >> >> Andrew >> >> >>> On Thursday, December 17, 2015, Andrew Godwin <[email protected]> >>> wrote: >>> >>>> Yes, that is the idea. While it obviously adds overhead (a millisecond >>>> or two in my first tests), it also adds natural load balancing between >>>> workers and then lets us have the same architecture for websockets and >>>> normal HTTP. >>>> >>>> (The interface server does do all the HTTP parsing, so what gets sent >>>> over is slightly less verbose than normal HTTP and needs less work to use, >>>> but it's not a big saving) >>>> >>>> Andrew >>>> >>>> On Thu, Dec 17, 2015 at 9:01 PM, Anssi Kääriäinen <[email protected]> >>>> wrote: >>>> >>>>> Is the idea a large site using classic request-response architecture >>>>> would get the requests at interface servers, these would then push the >>>>> HTTP >>>>> requests through channels to worker processes, which process the message >>>>> and push the response through the channel backend back to the interface >>>>> server and from there back to the client? >>>>> >>>>> - Anssi >>>>> >>>>> On Thursday, December 17, 2015, Andrew Godwin <[email protected]> >>>>> wrote: >>>>> >>>>>> To address the points so far: >>>>>> >>>>>> - I'm not yet sure whether "traditional" WSGI mode would actually >>>>>> run through the in memory backend or just be plugged in directly to the >>>>>> existing code path; it really depends on how much code would need to be >>>>>> moved around in either case. I'm pretty keen on keeping a raw-WSGI path >>>>>> around for performance/compatability reasons, and so we can hard fail if >>>>>> you try *any* channels use (right now the failure mode for trying to use >>>>>> channels with the wsgi emulation is silent failure) >>>>>> >>>>>> - Streaming HTTP responses are already in the channels spec as >>>>>> chunked messages; you just keep sending response-style messages with a >>>>>> flag >>>>>> saying "there's more". >>>>>> >>>>>> - File uploads are more difficult, due to the nature of the worker >>>>>> model (you can't guarantee all the messages will go to the same worker). >>>>>> My >>>>>> current plan here is to revise the message spec to allow infinite size >>>>>> messages and make the channel backend handle chunking in the best way >>>>>> (write to shared disk, use lots of keys, etc), but if there are other >>>>>> suggestions I'm open. This would also let people return large http >>>>>> responses without having to worry about size limits. >>>>>> >>>>>> - Alternative serialisation formats will be looked into; it's up to >>>>>> the channel backend what to use, I just chose JSON as our previous >>>>>> research >>>>>> into this at work showed that it was actually the fastest overall due to >>>>>> the fact it has a pure C implementation, but that's a year or two old. >>>>>> Whatever is chosen needs large support and forwards compatability, >>>>>> however. >>>>>> The message format is deliberately specified as JSON-capable structures >>>>>> (dicts, lists, strings) as it's assumed any serialisation format can >>>>>> handle >>>>>> this, and so it can be portable across backends. >>>>>> >>>>>> - I thought SCRIPT_NAME was basically unused by anyone these days, >>>>>> but hey, happy to be proved wrong. Do we have any usage numbers on it to >>>>>> know if we'd need it for a new standalone server to implement? It's >>>>>> really >>>>>> not hard to add it into the request format, just thought it was one of >>>>>> those CGI remnants we might finally be able to kill. >>>>>> >>>>>> Andrew >>>>>> >>>>>> On Thu, Dec 17, 2015 at 6:32 PM, Anssi Kääriäinen <[email protected] >>>>>> > wrote: >>>>>> >>>>>>> On Thursday, December 17, 2015, Carl Meyer <[email protected]> wrote: >>>>>>> >>>>>>>> Hi Andrew, >>>>>>> >>>>>>> >>>>>>>> - I share Mark's concern about the performance (latency, >>>>>>>> specifically) >>>>>>>> implications for projects that want to keep deploying old-style, >>>>>>>> given >>>>>>>> all the new serialization that would now be in the request path. I >>>>>>>> think >>>>>>>> some further discussion of this, with real benchmark numbers to >>>>>>>> refer >>>>>>>> to, is a prerequisite to considering Channels as a candidate for >>>>>>>> Django >>>>>>>> 1.10. To take a parallel from Python, Guido has always said that he >>>>>>>> won't consider removing the GIL unless it can be done without >>>>>>>> penalizing >>>>>>>> single-threaded code. If you think a different approach makes sense >>>>>>>> here >>>>>>>> (that is, that it's OK to penalize the simple cases in order to >>>>>>>> facilitate the less-simple ones), can you explain your reasons for >>>>>>>> that >>>>>>>> position? >>>>>>>> >>>>>>> >>>>>>> We would also need some form of streamed messages for streamed http >>>>>>> responses. >>>>>>> >>>>>>> Is it possible to handle old-style http the way it has always been >>>>>>> handled? >>>>>>> >>>>>>> - Anssi >>>>>>> >>>>>>> -- >>>>>>> You received this message because you are subscribed to the Google >>>>>>> Groups "Django developers (Contributions to Django itself)" group. >>>>>>> To unsubscribe from this group and stop receiving emails from it, >>>>>>> send an email to [email protected]. >>>>>>> To post to this group, send email to >>>>>>> [email protected]. >>>>>>> Visit this group at >>>>>>> https://groups.google.com/group/django-developers. >>>>>>> To view this discussion on the web visit >>>>>>> https://groups.google.com/d/msgid/django-developers/CALMtK1Gz%3DaYMLyFW2da2C6Wo_-c_V2T_4p6K9eh0vwrKB91dKw%40mail.gmail.com >>>>>>> <https://groups.google.com/d/msgid/django-developers/CALMtK1Gz%3DaYMLyFW2da2C6Wo_-c_V2T_4p6K9eh0vwrKB91dKw%40mail.gmail.com?utm_medium=email&utm_source=footer> >>>>>>> . >>>>>>> >>>>>>> For more options, visit https://groups.google.com/d/optout. >>>>>>> >>>>>> >>>>>> -- >>>>>> You received this message because you are subscribed to the Google >>>>>> Groups "Django developers (Contributions to Django itself)" group. >>>>>> To unsubscribe from this group and stop receiving emails from it, >>>>>> send an email to [email protected]. >>>>>> To post to this group, send email to >>>>>> [email protected]. >>>>>> Visit this group at https://groups.google.com/group/django-developers >>>>>> . >>>>>> To view this discussion on the web visit >>>>>> https://groups.google.com/d/msgid/django-developers/CAFwN1upm8N8tOhCdAgPEbYbBO6MU%2BmnEQ%3D%3Dp6EmW75%3DbNXHkfg%40mail.gmail.com >>>>>> <https://groups.google.com/d/msgid/django-developers/CAFwN1upm8N8tOhCdAgPEbYbBO6MU%2BmnEQ%3D%3Dp6EmW75%3DbNXHkfg%40mail.gmail.com?utm_medium=email&utm_source=footer> >>>>>> . >>>>>> For more options, visit https://groups.google.com/d/optout. >>>>>> >>>>> -- >>>>> You received this message because you are subscribed to the Google >>>>> Groups "Django developers (Contributions to Django itself)" group. >>>>> To unsubscribe from this group and stop receiving emails from it, send >>>>> an email to [email protected]. >>>>> To post to this group, send email to >>>>> [email protected]. >>>>> Visit this group at https://groups.google.com/group/django-developers. >>>>> To view this discussion on the web visit >>>>> https://groups.google.com/d/msgid/django-developers/CALMtK1FOVa6K-MMsZ9vACfcw0w0KHwdCXJ2vxu7_Y5Q9PHJ6Gg%40mail.gmail.com >>>>> <https://groups.google.com/d/msgid/django-developers/CALMtK1FOVa6K-MMsZ9vACfcw0w0KHwdCXJ2vxu7_Y5Q9PHJ6Gg%40mail.gmail.com?utm_medium=email&utm_source=footer> >>>>> . >>>>> >>>>> For more options, visit https://groups.google.com/d/optout. >>>>> >>>> >>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "Django developers (Contributions to Django itself)" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to [email protected]. >>>> To post to this group, send email to [email protected] >>>> . >>>> Visit this group at https://groups.google.com/group/django-developers. >>>> To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/django-developers/CAFwN1uoCX67D4E5J8n_wOo3VvPVjZYVnOG%2BshX6oDQk3zgu3vw%40mail.gmail.com >>>> <https://groups.google.com/d/msgid/django-developers/CAFwN1uoCX67D4E5J8n_wOo3VvPVjZYVnOG%2BshX6oDQk3zgu3vw%40mail.gmail.com?utm_medium=email&utm_source=footer> >>>> . >>>> For more options, visit https://groups.google.com/d/optout. >>>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "Django developers (Contributions to Django itself)" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> To post to this group, send email to [email protected]. >>> Visit this group at https://groups.google.com/group/django-developers. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/django-developers/CALMtK1E4HLL5v-z4u7temEDh6WfBpa9mcOXOf%3D-nSpKWRV%3DJsA%40mail.gmail.com >>> <https://groups.google.com/d/msgid/django-developers/CALMtK1E4HLL5v-z4u7temEDh6WfBpa9mcOXOf%3D-nSpKWRV%3DJsA%40mail.gmail.com?utm_medium=email&utm_source=footer> >>> . >>> >>> For more options, visit https://groups.google.com/d/optout. >>> >> >> -- >> You received this message because you are subscribed to a topic in the >> Google Groups "Django developers (Contributions to Django itself)" group. >> To unsubscribe from this topic, visit >> https://groups.google.com/d/topic/django-developers/CZPvEE0WPi4/unsubscribe >> . >> To unsubscribe from this group and all its topics, send an email to >> [email protected]. >> To post to this group, send email to [email protected]. >> Visit this group at https://groups.google.com/group/django-developers. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/django-developers/CAFwN1uoUc7_-4XbterzhriSB9vLhJ3DLfqrSUH9iyMNyVwSD7A%40mail.gmail.com >> <https://groups.google.com/d/msgid/django-developers/CAFwN1uoUc7_-4XbterzhriSB9vLhJ3DLfqrSUH9iyMNyVwSD7A%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> >> For more options, visit https://groups.google.com/d/optout. >> > > -- > You received this message because you are subscribed to the Google Groups > "Django developers (Contributions to Django itself)" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at https://groups.google.com/group/django-developers. > To view this discussion on the web visit > https://groups.google.com/d/msgid/django-developers/CAG2P6nRSPW04heEaDj2rgNu77kgKNb9s1jP_kiL1M5P4XQ12jg%40mail.gmail.com > <https://groups.google.com/d/msgid/django-developers/CAG2P6nRSPW04heEaDj2rgNu77kgKNb9s1jP_kiL1M5P4XQ12jg%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > > For more options, visit https://groups.google.com/d/optout. > -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/django-developers. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/CAFwN1uo8p3R-8pZBk%3DdbhYrqiWdjzN8f%2BbTz4-r9yEZUb0KBeQ%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
