Simple random sampling is a technique often used in data analysis to (substantially) reduce the amount of data to be processed. In this post, I’ll take a stab at explaining how this can be done with Dumbo.

Consider the file commentcounts.txt, derived as follows from the comments.txt file that was generated as part of an earlier post:

$ sort comments.txt | uniq -c | \ sed 's/^[^0-9]*\([0-9]*\) \(.*\)*/\2\t\1/' > commentcounts.txt

which leads to lines of the form:

<Hadoop JIRA issue number>\t<comment author>\t<number of comments>

By means of the following Dumbo program, some statistics can be computed for each of the comment authors that occur in this file:

def mapper(key, value): issuenr, commenter, count = value.split("\t") yield commenter, int(count) if __name__ == "__main__": from dumbo import run, statsreducer, statscombiner run(mapper, statsreducer, statscombiner)

The following three lines are part of the output obtained by running this program on commentcounts.txt:

Doug Cutting 1487 2.28984532616 2.94902633612 1 52 Owen O'Malley 1062 1.90301318267 1.87736300761 1 21 Tom White 268 2.26492537313 2.45723899975 1 24

Each of these lines ends with a number sequence consisting of the total number of issues commented on, followed by the mean, standard deviation, minimum, and maximum, respectively, of the number of comments per issue.

An easy way to take a random sample from commentcounts.txt is by means of the following Dumbo program:

from random import random def mapper(key, value): if random() < 0.5: yield key, value if __name__ == "__main__": from dumbo import run run(mapper)

Running the statistics program on the output of this sampling program gave me the following lines for Doug, Owen, and Tom:

Doug Cutting 760 2.22105263158 2.94351722074 1 52 Owen O'Malley 506 1.84584980237 1.69629699266 1 19 Tom White 131 2.23664122137 2.0296732376 1 11

A quick comparison of these lines with the ones above reveals that:

- The number of issues about halved for each of the considered comment authors, which makes sense since Python’s
*random()*function generates uniformly distributed floats between*0.0*(inclusive) and*1.0*(exclusive). - The means and standard deviations computed from the sample are quite close to the ones obtained from the complete dataset.
- Two out of the three maximums computed from the sample aren’t very close to the real one. In Tom’s case it even isn’t close at all.

So, apparently, we can reliably compute means and standard deviations — but not maximums — from a sample, which shouldn’t come as a surprise to anyone who has some basic knowledge of statistics. Just like maximums, minimums can, in general, not be computed from samples by the way, even though this happens to work fine in the example above.

Even for means and standard deviations, some caution has to be taken when computing them from a random sample though. In my case, for instance, the mean computed from the sample is *4.0*, whereas the real mean is *7.0*:

Klaas Bosteels 4 7.0 6.0415229868 1 16 Klaas Bosteels 3 4.0 3.55902608401 1 9

The reason for this inaccuracy is that I only commented on four Hadoop issues so far. Instead of regarding all of this as taking one random sample from commentcounts.txt, it’s better to think of it as taking a random sample for each comment author mentioned in this file. The number of issues for each author then corresponds to the sample size, which is inversely proportional to the sampling error you can expect for many statistics. More precisely, the error will, on average, be inversely proportional to the square root of the sample size for many statistics, including the mean.

Now, you might wonder why random sampling is useful in a Hadoop context. After all, the point of Hadoop is to allow you to easily process very large datasets, and hence you usually shouldn’t have to revert to random sampling to compute the numbers you’re interested in. Nevertheless, random sampling does have some advantages though, the main one being that it can allow you to get a representative dataset that fits entirely into your desktop’s memory, which creates the possibility of using a convenient software package like Pylab, R, Matlab, or even Excel for your analysis. Furthermore, random sampling is often also a faster way of computing the statistics you’re after, especially when your Hadoop cluster is very busy, since the sampling doesn’t require any reduce task slots (which can be hard to get a hold of on a busy cluster).

[…] Dumbotics Pseudo-random ramblings about Dumbo and Hadoop « Random sampling […]

I’ve perused a few articles from your site and I’d like to say thank you. I like your form and I subscribed to your rss feed. You do update your rss right?

[…] the samples, build a model or models, and then test it/them on the other samples. Apparently, random sampling is possible in Hadoop. What this means is that if I get these random samples from the DBMS, I can just use the same […]