A short development tale
Apr 4, 2012 - 11 p.m.

I missed posting last week, and this week I'm also feeling pretty overwhelmed. I'm working on finding a place to live and starting a job in a new city. So instead of trying to come up with something completely new to write about and missing a second week, I'm going to just relate a tale of code deployment gone wrong.

Roughly three years ago, I was working, and at the time, the company was in a bit of flux. The development team was understaffed, so with a ow level of experience, I inherited the job of deploying the website to production.

This meant taking machines in and out of service by logging into the load balancer, rsyncing new code to every machine, and then logging directly into the production servers to clear cache before making them live again. Clearing the cache was done by going into the cache directory and typing "/bin/rm -rf ." - it was necessary to do this because rm had been aliased to "rm -i", which prompted for every file removed, mainly for safety reasons.

This process was careful work, but at the same time, it was quite tedious. So while various tasks were running, I did what any good programmer does during downtime. I dorked around on the internet.

So one day, while doing a deployment, I was looking at reddit. When I looked up, my boss was standing over my shoulder - not just any boss, but the big boss. I'm not sure how many management layers existed between us at the time, but it was several. Needless to say, I felt a bit put off that I had been caught screwing around. He had a question. I answered quickly, spun in my chair, and executed the next command in my routine. "/bin/rm -rf ." Then I got up to get some coffee.

When I came back, the command was still running, which was unusual. I didn't think much of it until I saw a string of permission denied errors appear. That was very unusual. It had never happened before. That's when I realized I had never changed directories to where the cache was. I had just removed everything in the home directory. Which was also the web root.


And that's why you automate your deployments.
(I eventually did)
On the importance of backups
Mar 24, 2012 - 03 p.m.

I like to keep my stuff backed up. I've got a huge external drive next to my laptop that I use as a mac time machine. For stuff like tax backups, I also throw a copy in my dropbox. And another on my web server. I used to put a copy on a thumb drive. I wish I could find the source of a piece of advice I once read, but my google skill is lacking today. Basically, the jist was: "If you have n copies of something, you really only have n - 1 copies for all practical purposes." If you only have one copy of anything, you could find yourself without that thing any moment.

Making bunches of copies may also appeal to my inner hoarder sensibilities without taking up extra space in my house, but I'm going with the "just being safe" argument. At least publicly.

Either way, I realized just the other day that, to my horror, the only copy I have of my blog is the single database at my web host. I know for a fact that they do regular backups, but for my own piece of mind, I'd like to have copies myself.

So I wrote a little script that hooks into my django models, converts my posts to markdown formatting, and saves them to flat files. It also saves some meta data about each post in a related json file. At the moment it's nothing to write home about really, and it messes up the markup that I use for code blocks, but it's a good enough backup for most purposes. I plan on extending it a bit so that I can write posts as flat files and have them converted to html and synced back out to the live site, automatically push the copies to github, etc.

That's the other advantage. Now that my blog posts exist outside of a database, I can push the markdown files out to github. That way, if my grammar sucks or something, anyone can correct it and submit a pull request. That's assuming anyone cares to correct my grammar, but let's not get all academic here.

You can check out my backup script here, and the repository with copies of all my blog posts here.
A little more Quicksorting
Mar 15, 2012 - 08 p.m.

This has been quite a week. I've made the decision to take a new job and move my fiancee and myself to Buffalo from Brooklyn. We think that moving there will give us a better opportunity to live the lifestyle we want. A shorter commute, more space and proximity to family are among a whole host of reasons we've decided to do it. It was a difficult decision, and we'll both miss a lot of things about Brooklyn, but I think its for the best, even though I feel a little overwhelmed right now.

Anyway, I was poking through my work computer to see if there was any code that I didn't want to leave behind, and I found something that played off last week's interest in Quicksort. It was a piece of code that I had written when I first taught myself the algorithm. And, unlike most of your old code you run into, this was more nicely written than what I wrote just recently.
#what I wrote last week
def quicksortLoop(set):
    if len(set) <= 1:
        return set

    high = []
    low = []
    pivot = set.pop(len(set) / 2)

    for i in set:
        if i >= pivot:
    return quicksortLoop(low) + [pivot] + quicksortLoop(high)

#what I wrote a long time ago
def quicksort(list):
    if len(list) <= 1:
        return list;

    pivot = list.pop(len(list) / 2)
    less = [i for i in list if i < pivot]
    more = [i for i in list if i >= pivot]
    return quicksort(less) + [pivot] + quicksort(more)

Instead of looping through the set and putting values into two arrays, this implementation uses python's list comprehensions to build the more and less lists. I think this approach is much more elegant, yielding easier to read code.

This lead me to the question of whether or not the list comprehensions incurred a signifigant amount of overhead. To find out, I stripped away all the parts that distiguished the two different implementations from one another with the exception of the loop vs list comprehension pieces. Then I wrote some code to the run the two different function through the same paces.

from time import time
from random import randint

#function definitions here

if __name__ == "__main__":

    loopTimes = []
    compTimes = []

    for i in range(1, 1000):
        testSet = [randint(0, 100) for x in range(1000)]

        cSet = list(testSet)
        loopSet = list(testSet)

        compTime = time()
        compTime = time() - compTime

        loopTime = time()
        loopTime = time() - loopTime

    print "Avg with loop: ", sum(loopTimes)/1000
    print "Avg with list comprehension: ", sum(compTimes)/1000
What this is does is sort a thousand lists, each containing a thousand numbers between one and one hundred, using both sorting functions. It counts the time each function takes on each list, and then calculates an average.

The results are sort of interesting.
[snpxw@PWadeiMAC:random-bits ]$ python quicksortTimer.py 
Avg with loop:  0.00623641085625
Avg with list comprehension:  0.0061008477211
List comprehensions beat the loop every time, which is the opposite of what I expected. I'll speculate that the difference is either some internal optimization python makes for list comprehensions, or the pre-allocation of the low and high arrays in the loop version. I'm going to have to do a bit more research about what happens inside python to figure it out.

Regardless, the difference is negligible. Even if the list being sorted was five thousand times larger (that is, five million elements), the difference in the two implementations would be about 0.5 seconds. Not really enough to bother most people.

(The full code from this post is here.)
Mar 9, 2012 - 05 p.m.

I've spent a bit of time this week thinking about QuickSort. More specifically, I've been thinking about the complexity of the Quicksort worst case. On an already sorted set, Quicksort has a complexity of O(n^2), which I guess I understand on a basic level. Generally speaking, when the set is already sorted, Quicksort needs to compare every element to every other element each time it partitions the set. Partition n times, doing n comparisons, then you have n^2 comparisons. Makes sense, but instead of just leaving it, I've been trying to grasp it more definetly. In my mind, the comparisons on subsequent sets should be getting smaller. ie n-1, n-2, n-3 comparisons and so on.

I thought perhaps coding up a measurement of the complexity would help me wrap my head around it.

class QuickSorter:

    def __init__(self):
        self.depth = 0
        self.complexity = 0

    def quickSort(self, set):
        self.depth += 1
        self.complexity += 1;
        if(len(set) == 1 or len(set) == 0):
            return set

        high = []
        low = []
        pivot = set[0]

        for i in set:
            self.complexity += 2
            if i > pivot:
        return self.quickSort(low) + [pivot] + self.quickSort(high)

if __name__ == "__main__":
    q = QuickSorter()
    runset = [1, 2, 3, 4, 5, 6, 7]
    print q.quickSort(runset)
    print q.complexity

After a bit of messing around trying to decide what should be part of the complexity measurement, you'll see I decided on counting the check on the length of the array, and the greater than and less than checks on values. It all seemed to make perfect sense, but my numbers kept coming out wrong.

On an already sorted array of length 3, instead of getting back 9 operations, I got 11. A set of 6? 41 operations. It was making me insane, but then I realized that the actual complexity of the worst case ended up being O(n^2 + (n - 1)). The n-1 comes from the length checks on all of the empty sets. This promptly made me remember the way big O notation is handled, where the slowest growing elements of the complexity are discarded. What is the growth of n-1 in comparison to the speed of n^2? Insignificant.

It was a nice bonus to figure that all out, but I still don't think I'm satisifed. The issue with n-1 should be a lesson that understanding the complexity of these algorithms is less about actual exact measurement, and more about mathematical understanding of the situation. So, while I think I "get" it enough to get by, I'm going to spend a bit of time re-familiarizing myself with the math.

Anyway, I put the code up here. There is also an implementation of the in-place version of Quicksort there as a bonus.

As a quick aside, I realize that picking the first element in the array as a pivot is bad practice, as it often triggers worst case behavior. That was my intention here.

Update: I've spent a bit more time reading after writing this post, and I've come to the following understanding of Quicksort: in the best case, we evenly divide the array into roughly half the size, so we recursively build log n levels before we reach our base case, a list of size one. On each level, there is O(n) amount of work done, which gives us the case, O(n log n). By the same token, when we encounter the worst case, we must recursively divide the set n times before we get to the base case, giving us n levels with O(n) complexity, coming out to the worst case O(n^2).

Whew. That feels much better.
Mar 3, 2012 - 06 p.m.

Instead of a blog post here, this week I opted to write about bread over at my (and Hillary's) cooking blog: Us Versus Dinner. If you're into that sort of thing, or just like looking at pictures of tasty bread, give it a look. I'll be back here next week with something more appropriately nerdy.