An Asynchronous, Scalable Django with Twisted (PyCon TW 2016 Keynote)

Transcript

1. An Asynchronous, Scalable Django with Twisted
2. Hello, I’m Amber Brown (HawkOwl)
3. Twitter: @hawkieowl Pronouns: she/her
4. I live in Perth, Western Australia
6. Core Developer Release Manager Ported 40KLoC+ to Python 3
8. Binary release management across 3 distros Ported Autobahn|Python (Tx) and Crossbar.io to Python 3 Web API/REST integration in CB
9. Scaling Django Applications
10. Django serves one request at a time
11. gunicorn, mod_wsgi, etc run multiple copies in threads + processes
12. Concurrent Requests == processes x threadpool size
13. nginx gunicorn worker thread thread thread thread gunicorn worker thread thread thread thread Example server: two workers with four threads each
14. Need more requests? Add more web servers!
15. nginx gunicorn worker thread thread thread thread gunicorn worker thread thread thread thread nginx gunicorn worker thread thread thread thread gunicorn worker thread thread thread thread HAProxy Server 2 Server 3 Server 1
16. Scaling has required adding a new piece
17. Higher scale means higher complexity
18. Is there a better way to handle many requests?
19. Problem Domain
20. Modern web applications have two things that take a long time to do
21. CPU-bound work Math, natural language processing, other data processing
22. On most Python interpreters, Python threads are unsuitable for dispatching CPU-heavy work
23. Of N Python threads only 1 may run Python code because of the Global Interpreter Lock
24. Of N Python threads only N may run C code, since the Global Interpreter Lock is released
25. I/O-bound work Database requests, web requests, other network I/O
26. Threads work better for I/O-bound work
27. Thread switching overhead is expensive Rapidly acquiring/releasing the GIL is expensive
28. First, let's focus on I/O-bound applications.
29. Asynchronous I/O & Event-Driven Programming
30. Your code is triggered on events
31. Events can be: incoming data on the network some computation is finished a subprocess has ended etc, etc
32. How do we know when events have occurred?
33. All events begin from some form of I/O, so we just wait for that!
34. Event-driven programming frameworks
35. Twisted (the project I work on!)
36. (of SVN history)
37. asyncio was introduced much later
39. Same at their core, using "selector functions"
40. select() and friends (poll, epoll, kqueue)
41. Selector functions take a list of file descriptors (e.g. sockets, open files) and tell you what is ready for reading or writing
42. Selector loops can handle thousands of open sockets and events
43. Data is channeled through a transport to a protocol (e.g. HTTP, IMAP, SSH)
44. Sending data is queued until the network is ready
45. Nothing blocks, it simply gives control to the next event to be processed
46. No blocking means no threads
47. “I/O loops” or “reactors” (as it "reacts" to I/O)
48. Higher density per core No threads required! Concurrency, not parallelism
49. Best case: high I/O throughput, high-latency clients, low CPU processing
50. But what if we need to process CPU bound tasks?
51. Event Driven Programming with Work Queues
52. CPU bound tasks are added to a queue, rather than being ran directly
53. Web Server Task Queue Worker Worker Worker
54. We have made the CPU-bound task an I/O-bound one for our web server
55. We have also made the scaling characteristics horizontal
56. Web Server Task Queue Worker Server 2 CPU3 Worker Server 2 CPU2 Worker Server2 CPU1 Worker Server 1 CPU2 Worker Server 1 CPU1 Worker Server 2 CPU4
57. Putting tasks on the queue and removing them is cheap
58. Task queues scale rather well
59. Add more workers to scale!
60. Do we have an implementation of this?
61. The Architecture of Django Channels
62. Project to make an "asynchronous Django"
63. Authored by Andrew Godwin (behind South, Migrations)
64. Interface Server Channel Queue Worker Worker Worker Worker Worker Worker Server 1 Server 2 Server 3 Server 4
65. Interface server accepts requests, puts them on the Channel (task queue)
66. Workers take requests off the Channel and process them
67. Results from processed requests are written back to the Channel
68. The interface server picks up these responses and writes it back out to the HTTP request
69. The interface server is only I/O bound and does no "work" of its own
70. Perfect application for asynchronous I/O!
71. Daphne, the reference interface server implementation, is written in Twisted
72. Daphne is capable of handling thousands of requests a second on modest hardware
73. The channel layer can be sharded
74. Channel Queue Server 2 Interface Server Worker Worker Worker Worker Worker Worker Server 1 Server 4 Server 5 Channel Queue Server 3 (Sharding)
75. Workers do not need to be on the web server... but you can put them there if you want!
76. For small sites, the channel layer can simply be an Inter- Process-Communication bus
77. Channel Queue (Shared Memory) Interface Server Worker Worker Worker Server 1
78. And Twisted understands WebSockets... so can Channels too?
79. Yep!
80. How Channels Works
81. A Channel is where requests are put to be serviced
82. What is a request? - incoming HTTP requests - connected WebSocket connection - data on a WebSocket
83. http.request http.disconnect websocket.connect websocket.receive websocket.disconnect
84. Your worker listens on these channel names
85. Information about the request (e.g. a body and headers), and a "reply channel" identifier
86. http.response!<client> http.request.body!<client> websocket.send!<client>
87. http.response!c134x7y http.request.body!c134x7y websocket.send!c134x7y
88. Reply channels are connection specific so that the correct response gets to the correct connection
89. In handling a request, your code calls send() on a response channel
90. But because Channels is event-driven, you can't get a "response" from the event
91. The workers themselves do not use asynchronous I/O by default!
92. Under Channels, you write synchronous code, but smaller synchronous code
93. @receiver(post_save, sender=BlogUpdate) def send_update(sender, instance, **kwargs): Group("liveblog").send({ "id": instance.id, "content": instance.content, })
94. Group? What's a group?
95. Pool of request-specific channels for efficiently sending one-to-many messages
96. e.g: add all open WebSocket connections to a group that is notified when your model is saved
97. Handling different kinds of requests
98. Workers can listen on specific channels, they don't have to listen to all of them!
99. Interface Server Channel Queue Worker Worker Server 1 Server 2 Server 3 (high performance) Server 4 (standard) http.request bigdata.process
100. Because you can create and listen for arbitrary channels, you can funnel certain kinds of work into different workers
101. my_data_set = request.body Channel("bigdata.process").send( {"mydata": my_data_set})
102. How do we support sending that data down the current request when it's done?
103. my_data_set = request.body Channel("bigdata.process").send({ "mydata": my_data_set, "reply_channel": message.reply_channel})
104. All our big data worker needs to do then is send the response on the reply channel!
105. Channels as a bridge to an asynchronous future
106. A channel doesn't care if you are synchronous or asynchronous
107. ...or written in Django or even Python!
108. Channels implements "Asynchronous Server Gateway Interface"
109. The path to a hybrid future Go, Django, Twisted, etc etc
110. Channels is due to land in Django 1.11/2.0
111. Try it out! channels.readthedocs.io
112. Questions? (pls no statements, save them for after)