03 July 2014

Python micro benchmarks for fun and profit.

There comes a certain point in a developer’s life where they have the insatiable desire to do a benchmark. And not just any bechmark but a micro benchmark of something like string manipulation or data lookups. Whatever.

Let’s set the mood. You get a lot of http requests. A ton. Metric tonne. And based on either one of two headers you need to compare it to a list and decide if this request is allowed or not. To be more specific you will have:

header1 = "value:junk"
header2 = "junk=value"

And also you will have a precomputed list with all the naugty headers. The no-no list. And now the questions is, how do you do it? You also have python. Beause it’s pretty much a good all around language.

Trial by fire

import timeit

# This was was my initial guess. Trying to be clever with sets and not using lists. And I predicted that 
# string mangling is not great.

# Simple set intersection
denormalized_set_intersect = '{header1, header2} & denormalized_bad_list' 

denormalized_setup = '''\
 'a:something', 'something=a',
 'b:something', 'something=b',
 'c:something', 'something=c'
 'd:something', 'something=d',
 'e:something', 'something=e',
 'f:something', 'something=f'

Ok that might work, but is it worth it? No spoilers!! But anyway, these test programs are in string because they will be benchmarked in a bit.

# But then I was wondering if it was really better than trimming the strings..

normalized_set_intersect = '''\
{header1.split(':')[0], header2.split('=')[1]} & normalized_bad_list  

normalized_setup = '''\
normalized_bad_list={'a', 'b', 'c', 'd', 'e', 'f'}

Now I’m on a roll! What else is there…? Derp. The obvious, booleans!

# And then I was thinking that I was over thinking it and tried an "or" on a whim.
normalized_boolean = '''\
(header1.split(':')[0] in normalized_bad_list) or (header2.split('=')[1] in normalized_bad_list)

# Then just to make it full circle, it came back to the denormilzed with an "or"
denormalized_boolean = '''\
(header1 in denormalized_bad_list) or (header2 in denormalized_bad_list)


Let the games begin

# Welcome to the ultimate showdown ladies and gentlemen!
    ('Denormalized Set Sammy', denormalized_set_intersect, denormalized_setup),
    ('Normalized Set Jen', normalized_set_intersect, normalized_setup),
    ('Normalized Boolean Bob', normalized_boolean, normalized_setup),
    ('Denormalized Boolean Alice', denormalized_boolean, denormalized_setup),

for name, prog, setup in BENCHMARKS:
    print "{0:-<27} {1} sec".format(name, timeit.timeit(prog, setup=setup, number=1000000))


Denormalized Set Sammy----- 0.195351839066 sec
Normalized Set Jen--------- 0.832314968109 sec
Normalized Boolean Bob----- 0.677690982819 sec
Denormalized Boolean Alice- 0.0748100280762 sec   <--- Our lightweight champion!

So this is what I get for trying to be smart. A simple double boolean is all it takes. Some of our more clever readers will have noticed, “Hey, what you are searching for is not actually in the set!” Yes, my astute friend, and this is because there is a relatively tiny number of headers that could be matched. And the most common case is a miss. So the goal here is to ensure that all normal usage is not affected too much by this filter. I hope this make sense.

Until next time, take care of yourself, and each other.