C[omp]ute

From: andrew cooke <andrew@...>

Date: Wed, 23 Jan 2013 22:36:50 -0300

The last few weeks I've been entertaining myself by writing a small Python
server that generates "random" sentences.  The idea comes from this blog post
http://rachelbythebay.com/w/2012/08/29/info/ which suggests that such a
service could be used in a similar way to people taking photos of themselves
with the day's newspaper.

For such a service to be useful it has to be moderately reliable.  But what
does "reliable" mean?

 - persistent state, so that restarting continues correctly without
   losing or repeating values;

 - distributed state, so that people can access data even when one source is
   unavailable;

 - simple code, so that the data generated are correct;

 - good uptime.

I can't afford (well, don't want to spend) fancy hosting, so "good uptime" is
restricted by whatever my provider can do (actually, webfaction have been very
good), but within that I can still:

 - separate the generation and serving of sentences into decoupled processes;

 - reduce load on the server by using a cache/CDN;

Putting all that together, with an emphasis on simple code, I came up with
https://github.com/andrewcooke/pn-art-adj-adj-n-v-adv which is visible at
http://colorlessgreen.net/


Some notable features of the design are:

 - a single-threaded server.  This is very simple and serves all data from
   memory (so doesn't block on IO, apart from the request and response).  It
   sits behind a free CloudFire CDN which I hope will help handle peaks in
   demand.
   https://github.com/andrewcooke/pn-art-adj-adj-n-v-adv/blob/master/src/pnartadjadjnvadv/server.py

 - separate processes for sentence generation and web serving.  I wanted to
   the two so that an overloaded server wouldn't stop generation.  And in
   Python (with the fabled global interpreter lock) that means using two
   processes.  Communication is via a queue - the sentence generator sends new
   sentences across as they are made.
   https://github.com/andrewcooke/pn-art-adj-adj-n-v-adv/blob/master/src/pnartadjadjnvadv/process.py

 - changes in state (new sentences) are written immediately to disk.  The
   process restart (both) reads from this, and it is also pushed remotely
   via git.

 - I've tried to delegate as much as possible to third party services.  So
   Cloudflare handles high loads; github helps distributed state; Twitter acts
   as a message bus from which other people can consume values.

   Since I've not had a huge amount of load it's hard to tell how much
   Cloudflare helps (presumably it helps more when load is high, as it will
   repond to more requests beforce checking for updates).  And I am not
   providing any guidance on caching in headers, so it is probably being quite
   conservative (caching for short periods only).  But given all that, it
   currently seems to be reducing the load by half.


Andrew