HN Explorer

21 comments

[ 3.4 ms ] story [ 65.8 ms ] thread
spenrose
He doesn't use the cPickle binary format, which is significantly faster: http://docs.python.org/library/pickle.html#pickle.HIGHEST_PR...
j2d2j2d2
Thanks for pointing this out.

I am going to do some follow up tests to cover the advice I got here. I will include this, tnetstrings and decodes too.

thezilch
Any interests in also including Google's Protocol Buffers and Facebook's Thrift since it would appear your intentions are to use these blobs in an RPC?
llambda
I'm confused, is cjson still recommended? Is it still maintained? Its status seems ambiguous based on the PyPi discussion. Furthermore it seems there's a number of issues with cjson that need to be resolved...
ericflo
My understanding is that cjson is not recommended or maintained.
bdarnell
When I tested cjson (~6 months ago?) it had bugs with escaping - strings containing quotes and/or backslashes would be encoded/decoded incorrectly. Not recommended, unless there have been fixes since then.
llambda
Based on the PyPi page for cjson, the comments seem to indicate many problems still persist, notwithstanding new updates that have yet to be pushed
jchrisa
sounds like the kind of project a decent coder could make big improvements on fast
davvid
Maybe our information is out of date, but here's what we learned:

https://github.com/jsonpickle/jsonpickle/commit/0b97652e4102...

jsonpickle: Remove cjson support

"First, please don't use cjson for anything. It's got multiple bugs and misfeatures, and is generally unsuited for anything except impressive benchmarks. It was easier to write my own library, from scratch, than try to fix cjson."

-- John Millikin, author of jsonlib

d0mine
Python 2.7 on Ubuntu:

  $ python -mtimeit -s "from json import dumps; 
  d = {
      'foo': 'bar',
      'food': 'barf',
      'good': 'bars',
      'dood': 'wheres your car?',
      'wheres your car': 'dude?',
  }
  " "dumps(d)"
  100000 loops, best of 3: 6.89 usec per loop

  $ python -mtimeit -s "from cPickle import dumps; 
  d = {
      'foo': 'bar',
      'food': 'barf',
      'good': 'bars',
      'dood': 'wheres your car?',
      'wheres your car': 'dude?',
  }
  " "dumps(d)"
  100000 loops, best of 3: 7.78 usec per loop
So `json` seems faster than `cPickle`. Right? Wrong!:

  $ python -mtimeit -s "from cPickle import dumps; 
  d = {
      'foo': 'bar',
      'food': 'barf',
      'good': 'bars',
      'dood': 'wheres your car?',
      'wheres your car': 'dude?',
  }
  " "dumps(d, -1)"
  100000 loops, best of 3: 3.59 usec per loop
trafficlight
What does the "-1" dump argument mean in 3rd example?
kingkilr
It says to use the highest optimization level.
d0mine
No. It says use the highest available protocol.

The highest available protocol is not necessarily the fastest. There might be other than pure speed reasons to introduce the next Pickle protocol.

thezilch
The equivalent of using pickle.HIGHEST_PROTOCOL or pickle.dumps(obj, protocol=pickle.HIGHEST_PROTOCOL) -- any negative value will force highest protocol.
thezilch
And because caching systems are typically a means for speeding up reads, where the article only focuses on writes.

Python 2.5 -- including pickle's (faster) bin mode

    Starting simplejson dumps...
    done: 31.7636611462
    Starting cjson encode...
    done: 3.80062890053
    Starting pickle dumps...
    done: 10.4873199463
    Starting pickle dumps (protocol=-1 -- force HIGHEST_PROTOCOL)...
    done: 5.98110699654
    Starting simplejson loads...
    done: 138.709590197
    Starting cjson decode...
    done: 1.85300803185
    Starting pickle loads...
    done: 5.24735999107
    Starting pickle loads (protocol=-1 -- force HIGHEST_PROTOCOL)...
    done: 3.66428518295
michaelfairley
Apple and oranges (even though they're mostly used for the same thing). pickle is a Turing complete language used to recreate python objects, while json is only used for serialization. pickle is slightly more powerful, but should also not be used to load untrusted data. See http://nadiana.com/python-pickle-insecure