1.
Building Real-Time Web
http://tinyurl.com/realtime2012
http:// Timothy Fitz .com
CTO Canvas

2.
What is “Realtime web”

3.
What does “Realtime” look like?

4.
What does “Realtime” look like?

5.
What does “Realtime” look like?

6.
“Push, not pull.”
REALTIME WEB

7.
Talking to the browser
High concurrency
Scaling up
3 HARD PROBLEMS

8.
Talking to the browser
• Short Polling
• Long Polling
• WebSocket
• Flash Socket

9.
Short Polling

10.
Long Polling

11.
Flash Socket

12.
WebSocket

13.
High Concurrency
• Blocking I/O
– Thread per process
– Tops out at 200 to 1k connections
• Non-blocking I/O
– One process, one thread
– 10k to 100k connections

14.
Django

15.
Django
Apache

16.
There is no apache for realtime

17.
Non-blocking I/O Servers
• Python
– Twisted
– Tornado
– gevent
• Not python
– Node.js
– Erlang something

18.
Twisted
• Pro
– Can talk every protocol ever
– Oldest and most widely used in production
• Con
– Overkill for web-only tasks
– Not simple

19.
Tornado
• Pro
– Simple
– Does HTTP stuff simply
• Con
– Might not interface with what you need
• Confusing
– You can run Tornado (HTTP layer) on top of
Twisted (networking layer)

20.
gevent
• Pro
– Coroutines are a better model than callbacks
– As such, very easy to write complicated logic
• Con
– Least well documented
– Least consensus on best practices
– New, uncertain about production readiness

21.
Node.js
• Pro
– Best documentation by far
– Socket.IO abstracts away browser communication
• Con
– Can’t share logic between Django app
– New, but has fairly large install base

22.
Erlang
• Pro
– Hands down best for complex realtime tasks
– Forces you to think about concurrency/scale
– Abstracts away the network
– Old and reliable
• Con
– Forces you to think about concurrency/scale
– Can’t share logic between Django app
– High spin-up cost (functional, concurrency driven)

23.
Just one
Frontend nodes x Backend nodes
More architecture decisions!
SCALING UP!

24.
Just one
• Everything in memory
• Django nodes talk directly to box
• Spare for availability
• Failover = realtime data loss
– Make realtime 100% redundant

25.
Probably good enough!
– WARNING: NAPKIN MATH
– 10k daily visits * 10.0min avg visit
= 70 average concurrent users
– One box can easily be built out to handle 3-5k
= Roughly 450k-700k daily visits

26.
Frontend nodes x Backend nodes
• Frontend handle users / connections
• Backend handles channels

27.
More architecture decisions!
• In memory backend
– Redis Pub/Sub
– ZeroMQ
– Roll your own
• Persisted to Disk:
– ActiveMQ
– RabbitMQ
– Amazon SQS

28.
Redis Pub/Sub
• Simplest to setup
• Simplest model
• SUBSCRIBE channel_name
• PUBLISH channel_name “Hello World!”

29.
ZeroMQ
• Publish/Subscribe semantics
• Request/Response
• Push/Pull (round robin)
• Extremely fast

30.
Roll your own
• Same language as your frontend
– (Twisted/Node/Whatever)
• Only do this if you have per-channel business
logic
– You probably don’t.
• Erlang maps really really well to this domain.

31.
Full Stack Services
• REST APIs to push to the browser
• http://pusher.com
• http://beaconpush.com

32.
Canvas
Amazon ELB Nginx + Twisted Redis

33.
Final Recommendations
• Need python? Twisted
• Don’t? Node.js/SocketIO
• Need scale/reliability? Redis backend.
• Complex? Going big? Erlang all the way.

34.
Questions?

35.
Further Reading
• IMVU IMQ talk http://www.slideshare.net/JonWatte/message-queuing-
on-a-large-scale-imvus-stateful-realtime-message-queue
• Twilio talk on gevent + zeromq (given by Jeff Lindsay, highly recomended):
http://www.twilio.com/conference/video/distributed-systems-with-
gevent-and-zeromq
• Last.fm scaling Eralng/Mochiweb to 1 million concurrent connections on
one machine: http://www.metabrew.com/article/a-million-user-comet-
application-with-mochiweb-part-1
• The original Comet blog post: http://infrequently.org/2006/03/comet-low-
latency-data-for-the-browser/
• Django + Socket.IO + gevent:
http://codysoyland.com/2011/feb/6/evented-django-part-one-socketio-
and-gevent/