7 ms·

Growler: Asyncio Micro-Framework in Python

116 points by akubera 10y ago

tcgarvin 10y ago
Does this work with uvloop?
- sametmax 10y ago
  Does it blend ?
  - hyperbovine 10y ago
    *Will
- akubera 10y ago
  Yes! An initial test last night of a trivial growler app (i.e. one method that replies with a simple string) saw a median response time drop from ~200ms to ~80ms. (using apachebenchmark, n=1000 c=100)
  - zeemonkee3 10y ago
    How well would it work with, say, PostgreSQL? Would you need to use async wrappers like aiopg in your views?
    - ssahoo 10y ago
      https://github.com/aio-libs/aiopg https://github.com/aio-libs/aiopg should work.
    - akubera 10y ago
      For optimal performance, yes use aiopg or equivalent (it looks like there is an asyncio sqlalchemy project which might be nice.) Any middleware may be a coroutine object, so it's as easy as `await res.render("whatever")` in an async function
  - xxpor 10y ago
    Wait, p50 using the built in asyncio loop is 200ms? That seems incredibly slow.
  - sandGorgon 10y ago
    do you have this in a github somewhere ? would love to check it out. Also would be nice to see this same thing with postgres DB access.
phantom_oracle 10y ago
How is this different to Twisted?
- tcgarvin 10y ago
  It's asyncio native. Also I think syntactically it's different.
  - akhilcacharya 10y ago
    Like Flask + Express, specifically.
- akubera 10y ago
  As I understand, the asyncio module was inspired by Twisted. Growler is, at its heart, a super minimal interface: passing (req, res) pairs through a pipeline of middleware (similar to node). Since this is built with the standard library, all kinds of features, such as async DB access, will become available as the community builds them; and these may be added to your pipeline super cheap. It's just an experiment.
avyfain 10y ago
Awesome work! I just watched your PyOhio talk[0], and seems really promising. I use Flask daily, and while performance has not been an issue yet, I know the key word in that phrase is yet. Would love to play with the code. Could you add a CONTRIBUTING.md? I could try to figure out how to get started on my own, but having a few pointers would be really helpful. I see there are no issues on GH. What are things you're hoping to add soon/are not implemented yet? [0] https://www.youtube.com/watch?v=vs7XOn1TXVQ https://www.youtube.com/watch?v=vs7XOn1TXVQ
- akubera 10y ago
  Thanks! The biggest missing piece is simple, well-tested middleware. My focus has been on very general middleware; a kind of 'meta-middleware' that actually spec out a common behavior to minimize surprises and boilerplate. For example, my Renderer & RenderEngine classes[0] don't do anything on their own, but are intended to be used by other, more specific renderers (eg MakoRenderer, JinjaRenderer). I think some common functionality really need to be standardized and not done ad-hoc, as conflicts could arise. (more description on how Renderers work below) The next one of these I'd like to solve is authentication: a standard interface for passwords/oauth/etc. I've kinda used http://passportjs.org http://passportjs.org before, might be a starting point; I don't know what a pythonic version would look like. ## Non-Middleware - Figure out equivalent of JS EventEmitter interface. Currently I have no good way to do something like res.on('send-headers', cb) ~ (thinking res.events['send-headers'].add(cb)) with events as EventManager object. Maybe decorators here? - ETAGs - HTTP/2 - pytest-growler for automated fixtures - I don't think an admin panel is necessary (can't beat django) but would like some CLI tools for site creation, testing, and deployment. Renderer Explaination: (too many 'render' words) The author creates a RenderEngine as the actual middleware: app.use(MakoRenderer("/path/to/view")). All this middleware does is attach a callable Renderer object to res at res.render if it doesn't exist, then add itself to res.render - this allows multiple RenderEngines in the same app (if the author desires) - and none of them will clobber one another. Upon calling the Renderer: res.render('foo', data_obj), the callable scans through the engines until one with 'foo' is found, then the rendered text is sent. It's non-(obvious|trivial) situations like these I want in Growler proper (some batteries included) - the rest can be extensions. [0] https://github.com/pyGrowler/Growler/blob/dev/growler/middleware/renderer.py https://github.com/pyGrowler/Growler/blob/dev/growler/middle...
sandGorgon 10y ago
a quick request (and something similar was discussed here - https://news.ycombinator.com/item?id=11626454 https://news.ycombinator.com/item?id=11626454) could you add a representative example to show how we can use your framework with Postgres ? For example, psycopg vs psycopg2 vs psycogreen vs psycopg2-cffi vs aiopg ... I dont even know which one will work properly using your framework and how (and maintain async characteristics).
- Leynos 10y ago
  (Comment from a bystander). You probably want aiopg in order to maintain the async characteristics and use the "friendly" asyncio syntax. There is also peewee-async (http://peewee-async.readthedocs.io/en/latest/ http://peewee-async.readthedocs.io/en/latest/) if you want an ORM using asyncio. (I've been meaning to try this myself, but haven't had an opportunity).
  - sandGorgon 10y ago
    absolutely.. i am concerned about frameworks not fundamentally thinking about DB. With async, there is that additional problem of "this does not work in async mode". So I would hope the developers include database hookup as part of their unit test suite.
bullen 10y ago
If you want this feature, but multithreaded and with hot-deploy + filesystem database clustered you can try: http://github.com/tinspin/rupy http://github.com/tinspin/rupy
- andybak 10y ago
  Slightly disingenuous not to point out that's a Java app. I suspect the choice of language is a factor for a fair few people, maybe even the majority.
  - bullen 10y ago
    Yep, languages are like religion. But you can't hotdeploy any other language properly.
    - fleetfox 10y ago
      I'm not sure what exactly categorizes as "hot deploy" but how about erlang, elixir?
      - bullen 10y ago
        Rupy can switch code on the fly, meaning you can have users hitting your server and they will run the old code one request and new code the next. No other system I know of works like this, hot-deploying on a remote machine over HTTP seamlessly. Maybe erlang can do it but then again I can't write a 3D game in that language so Java still has the upper hand for my requirements.
        akubera 10y ago
        Off the top of my head, if you write a custom MiddlewareChain class that exposed its own management port, you could inspect and modify the state of the chain at any time, checkboxes to disable middleware, change endpoints, insert debug middleware, etc. I like this... definitely worth thinking about. Thanks, all!
gregn610 10y ago
Impressive stuff. But um, about that project name; ever watched "Bo Selecta" https://en.wikipedia.org/wiki/Bo%27_Selecta https://en.wikipedia.org/wiki/Bo%27_Selecta! ?
- akubera 10y ago
  No. But if it's related to the urbandictionary definition... I'm taking back the word!
cdnsteve 10y ago
Could you clarify this vs aiohttp (aiohttp.web)? http://aiohttp.readthedocs.io/en/stable/ http://aiohttp.readthedocs.io/en/stable/ Here's a good example of use: http://igordavydenko.com/talks/lvivpy-4/#slide-1 http://igordavydenko.com/talks/lvivpy-4/#slide-1
- akubera 10y ago
  I have not used aiohttp, but it looks like their application must be initialized with a list of middleware_factories. I do not see any logic in choosing middleware - just a straight shot. Growler apps just add functions to an internal list, and these may be 'mounted' on particular end points/http requests. Also, in aiohttp it is up to your endpoint to create the response object and return it. While in Growler it is created by the application (or rather the GrowlerHttpResponder) and passed around the middleware - getting enhancements (res.render('foo') is not standard) and having callbacks registered (to, say, log ALL headers sent - something you want done only at the time of sending headers). I think it would be possible (or at least I would like to get to the point) where you could actually mount other apps from other frameworks into a Growler app. So if there is blog software you like written in aiohttp, there is special middleware to transform growler requests into a aiohttp requests. import growler import aiohttp.web aio_blog_app = aiohttp.web.Application() ... app = growler.App("main site") app.use(GrowlerAiohttpMiddleware(aio_blog_app), path='/blog') ... THAT would be kinda cool.

116 points by akubera 10y ago

tcgarvin 10y ago

Does this work with uvloop?

sametmax 10y ago
Does it blend ?
- hyperbovine 10y ago
  *Will
akubera 10y ago
Yes! An initial test last night of a trivial growler app (i.e. one method that replies with a simple string) saw a median response time drop from ~200ms to ~80ms. (using apachebenchmark, n=1000 c=100)
- zeemonkee3 10y ago
  How well would it work with, say, PostgreSQL? Would you need to use async wrappers like aiopg in your views?
  - ssahoo 10y ago
    https://github.com/aio-libs/aiopg https://github.com/aio-libs/aiopg should work.
  - akubera 10y ago
    For optimal performance, yes use aiopg or equivalent (it looks like there is an asyncio sqlalchemy project which might be nice.) Any middleware may be a coroutine object, so it's as easy as `await res.render("whatever")` in an async function
- xxpor 10y ago
  Wait, p50 using the built in asyncio loop is 200ms? That seems incredibly slow.
- sandGorgon 10y ago
  do you have this in a github somewhere ? would love to check it out. Also would be nice to see this same thing with postgres DB access.

phantom_oracle 10y ago

How is this different to Twisted?

tcgarvin 10y ago
It's asyncio native. Also I think syntactically it's different.
- akhilcacharya 10y ago
  Like Flask + Express, specifically.
akubera 10y ago
As I understand, the asyncio module was inspired by Twisted. Growler is, at its heart, a super minimal interface: passing (req, res) pairs through a pipeline of middleware (similar to node). Since this is built with the standard library, all kinds of features, such as async DB access, will become available as the community builds them; and these may be added to your pipeline super cheap. It's just an experiment.

avyfain 10y ago

Awesome work! I just watched your PyOhio talk[0], and seems really promising. I use Flask daily, and while performance has not been an issue yet, I know the key word in that phrase is yet. Would love to play with the code. Could you add a CONTRIBUTING.md? I could try to figure out how to get started on my own, but having a few pointers would be really helpful. I see there are no issues on GH. What are things you're hoping to add soon/are not implemented yet? [0] https://www.youtube.com/watch?v=vs7XOn1TXVQ https://www.youtube.com/watch?v=vs7XOn1TXVQ

akubera 10y ago
Thanks! The biggest missing piece is simple, well-tested middleware. My focus has been on very general middleware; a kind of 'meta-middleware' that actually spec out a common behavior to minimize surprises and boilerplate. For example, my Renderer & RenderEngine classes[0] don't do anything on their own, but are intended to be used by other, more specific renderers (eg MakoRenderer, JinjaRenderer). I think some common functionality really need to be standardized and not done ad-hoc, as conflicts could arise. (more description on how Renderers work below) The next one of these I'd like to solve is authentication: a standard interface for passwords/oauth/etc. I've kinda used http://passportjs.org http://passportjs.org before, might be a starting point; I don't know what a pythonic version would look like. ## Non-Middleware - Figure out equivalent of JS EventEmitter interface. Currently I have no good way to do something like res.on('send-headers', cb) ~ (thinking res.events['send-headers'].add(cb)) with events as EventManager object. Maybe decorators here? - ETAGs - HTTP/2 - pytest-growler for automated fixtures - I don't think an admin panel is necessary (can't beat django) but would like some CLI tools for site creation, testing, and deployment. Renderer Explaination: (too many 'render' words) The author creates a RenderEngine as the actual middleware: app.use(MakoRenderer("/path/to/view")). All this middleware does is attach a callable Renderer object to res at res.render if it doesn't exist, then add itself to res.render - this allows multiple RenderEngines in the same app (if the author desires) - and none of them will clobber one another. Upon calling the Renderer: res.render('foo', data_obj), the callable scans through the engines until one with 'foo' is found, then the rendered text is sent. It's non-(obvious|trivial) situations like these I want in Growler proper (some batteries included) - the rest can be extensions. [0] https://github.com/pyGrowler/Growler/blob/dev/growler/middleware/renderer.py https://github.com/pyGrowler/Growler/blob/dev/growler/middle...

sandGorgon 10y ago

a quick request (and something similar was discussed here - https://news.ycombinator.com/item?id=11626454 https://news.ycombinator.com/item?id=11626454) could you add a representative example to show how we can use your framework with Postgres ? For example, psycopg vs psycopg2 vs psycogreen vs psycopg2-cffi vs aiopg ... I dont even know which one will work properly using your framework and how (and maintain async characteristics).

Leynos 10y ago
(Comment from a bystander). You probably want aiopg in order to maintain the async characteristics and use the "friendly" asyncio syntax. There is also peewee-async (http://peewee-async.readthedocs.io/en/latest/ http://peewee-async.readthedocs.io/en/latest/) if you want an ORM using asyncio. (I've been meaning to try this myself, but haven't had an opportunity).
- sandGorgon 10y ago
  absolutely.. i am concerned about frameworks not fundamentally thinking about DB. With async, there is that additional problem of "this does not work in async mode". So I would hope the developers include database hookup as part of their unit test suite.

bullen 10y ago

If you want this feature, but multithreaded and with hot-deploy + filesystem database clustered you can try: http://github.com/tinspin/rupy http://github.com/tinspin/rupy

andybak 10y ago
Slightly disingenuous not to point out that's a Java app. I suspect the choice of language is a factor for a fair few people, maybe even the majority.
- bullen 10y ago
  Yep, languages are like religion. But you can't hotdeploy any other language properly.
  - fleetfox 10y ago
    I'm not sure what exactly categorizes as "hot deploy" but how about erlang, elixir?
    - bullen 10y ago
      Rupy can switch code on the fly, meaning you can have users hitting your server and they will run the old code one request and new code the next. No other system I know of works like this, hot-deploying on a remote machine over HTTP seamlessly. Maybe erlang can do it but then again I can't write a 3D game in that language so Java still has the upper hand for my requirements.
      - akubera 10y ago
        Off the top of my head, if you write a custom MiddlewareChain class that exposed its own management port, you could inspect and modify the state of the chain at any time, checkboxes to disable middleware, change endpoints, insert debug middleware, etc. I like this... definitely worth thinking about. Thanks, all!

gregn610 10y ago

Impressive stuff. But um, about that project name; ever watched "Bo Selecta" https://en.wikipedia.org/wiki/Bo%27_Selecta https://en.wikipedia.org/wiki/Bo%27_Selecta! ?

akubera 10y ago
No. But if it's related to the urbandictionary definition... I'm taking back the word!

cdnsteve 10y ago

Could you clarify this vs aiohttp (aiohttp.web)? http://aiohttp.readthedocs.io/en/stable/ http://aiohttp.readthedocs.io/en/stable/ Here's a good example of use: http://igordavydenko.com/talks/lvivpy-4/#slide-1 http://igordavydenko.com/talks/lvivpy-4/#slide-1

akubera 10y ago
I have not used aiohttp, but it looks like their application must be initialized with a list of middleware_factories. I do not see any logic in choosing middleware - just a straight shot. Growler apps just add functions to an internal list, and these may be 'mounted' on particular end points/http requests. Also, in aiohttp it is up to your endpoint to create the response object and return it. While in Growler it is created by the application (or rather the GrowlerHttpResponder) and passed around the middleware - getting enhancements (res.render('foo') is not standard) and having callbacks registered (to, say, log ALL headers sent - something you want done only at the time of sending headers). I think it would be possible (or at least I would like to get to the point) where you could actually mount other apps from other frameworks into a Growler app. So if there is blog software you like written in aiohttp, there is special middleware to transform growler requests into a aiohttp requests. import growler import aiohttp.web aio_blog_app = aiohttp.web.Application() ... app = growler.App("main site") app.use(GrowlerAiohttpMiddleware(aio_blog_app), path='/blog') ... THAT would be kinda cool.