dlsecyx pgroammr (dyslexic programmer - cool stuff for scaling)
TRANSCRIPT
Dlsiecyx Pgroeammr
The cool stuff for scaling
Required Listening
Disclaimer about me
- I like both Ruby and Python, but I dislike the generator/deferred model they implement for non-blocking I/O
- I think node.js beats them because its better wrapped, as EM synchrony does to EM and gevent for python.
- I don't think this model is mandatory, not even migrate from Blocking to non-blocking I/O is needed for most cases. The best solution may not lie in code complexity. Maybe the right architecture.
- I like Erlang, the actor/message passing model and pattern matching. Way more. And I like message queues. Specially this one: http://www.restmq.com
Disclaimer about dyslexia
From wikipedia (http://en.wikipedia.org/wiki/Dyslexia):
"... Many students reach higher education before they encounter the threshold at which they are no longer able to compensate for their learning weaknesses.
One common misconception about dyslexia is that dyslexic readers write words backwards or move letters around when reading. In fact, this only occurs in a very small population of dyslexic readers. Dyslexic people are better identified by writing that does not seem to match their level of intelligence from prior observations. Additionally, dyslexic people often substitute similar-looking, but unrelated, words in place of the ones intended (what/want, say/saw, help/held, run/fun, fell/fall, to/too, etc.)"
Basics
Asynchronous I/O - Posix aio_*Non-Blocking I/OSelect, Poll, EPoll, KQueueEvented I/O - Abstraction over any I/O ModelReactor - Pattern for multiplexing I/OFuture ProgrammingGenerator Deferred Event Emitters The c10k problem - http://www.kegel.com/c10k.html
Scalability, not Performance
Scalability, not Performance
1. Performance: same work with less effort2. Scalability: more work with more resources3. Scientific measures for performance: How much and How fast4. Tricks to improve performance can hurt scalability. e.g:
Map/Reduce is scalable, not fast (not suited for real time queries)
5. Non-Blocking I/O is not an optimization for Performance6. The time spent waiting for an external resource can be used
to other tasks, given a known introduced complexity in the code.
7. Most of time architectural changes can avoid deeper code changes
(1,2,3,4 - Goetz et al, Java Concurrency in Practice, M/R example is mine) (5,6,7 - Lucindo, R. http://slidesha.re/aYz4Mk, wording on 6 is mine)
Wait, Wat ?
Sequential Blocking Example
Ruby
Python
URL Fetch Benchmark
URL Fetch Benchmark
2 urlsSequential blocking I/O using ruby: 0m3.061sNon-blocking I/O using ruby and Eventmachine: 0m0.795s
28 urlsSequential blocking I/O using ruby: 0m56.014sNon-blocking I/O using ruby and Eventmachine: 0m24.519s
Python and Twisted
Python/Twisted
Python/Twisted - Verbose and Inlined
Python/Twisted - Generators
Python/Twisted - Deferreds
Python/Twisted - Callbacks
Python/Twisted - Inline Callbacks
Python and GEvent
Python/gevent
Python/gevent - monkey patch
Python/gevent - join the greenlets
Ruby and EventMachine
Ruby/EventMachine
Ruby/EventMachine - Generators
Ruby/EventMachine - Deferreds
Ruby/EventMachine - Callbacks
Ruby/EventMachine/EM-Synchrony - Fiber based
Node.js
Node.js
Node.js - Creating and ending the Request
Node.js - Event handlers
Intermission: Erlang
Erlang
Erlang - Spawning a process to each URL
Erlang - Single URL download
Reminder of what we wanted to do
Ruby
Python
Architectural building blocks
Decoupling db writes with Message Queues
Coupled comment
Uncoupled comment - producer
Uncoupled comment - consumer
Cache
HTML processing - no cache
http://github.com/gleicon/vuvuzelr/proxy_no_cache.rb
HTML processing - Cached
http://github.com/gleicon/vuvuzelr/proxy.rb
