A couple of weeks back I read a couple of papers (Who Says What to Whom on Twitter and What is Twitter, a Social Network or a News Media?) that cited data collected by researchers for the latter paper.
This 5 gigabyte compressed (26 gigabyte uncompressed) dataset makes for a good excuse to use MapReduce and MrJob for processing. MrJob makes it easy to test MapReduce jobs locally as well as run them on a local Hadoop cluster or on Amazon’s Elastic MapReduce.
Designing MapReduce Jobs
I usually find myself going through the same basic tasks when writing MapReduce tasks:
- Examine the data input format and the data that you have to play with. This is sometimes explained in a metadata document or you may have to use a utility such as head if you’re trying to look at the very beginning of a text file.
- Create a small amount of synthetic data for use while designing your job. It should be obvious to determine if the output of your job is correct or not based on this data. This data is also useful when writing unit tests.
- Write your job, using synthetic data as test input.
- Create sample data based on your real dataset and continue testing your job with that data. This can be done via reservoir sampling to create a more representative sample or it could be as simple as head -1000000 on a very large file.
- Run your job against the sample data and make sure the results look sane.
- Configure MrJob to run using Elastic MapReduce. An example configuration can be found in conf/mrjob-emr.conf but will require you to update it with your credentials and S3 bucket information before it will work.
- Run your sample data using Elastic MapReduce and a small number of low-cost instances. It’s a lot cheaper to fix configuration problem when you’re just
running two cheap instances.
- Once you’re comfortable with everything, run your job against the full dataset on Elastic MapReduce.
Analyzing the data
This project contains two MapReduce jobs:
- A simple single-stage MapReduce job that reads the data in and sums the number of followers each user has.
- This is a two-phase MapReduce job that first sums the number of followers a each user has then for each follower count sums the number of users that have that number of followers. This is one of many interesting ways at looking at this raw data.
Running the jobs
The following assumes you have a modern Python and have already installed MrJob (pip install MrJob or easy_install MrJob or install it from source).
To run the sample data locally:
$ python jobs/follower_count.py data/twitter_synthetic.txt
This should print out a summary of how many followers each user (represented by id) has:
5 2 6 1 7 3 8 2 9 1
You can also run a larger sample (the first 10 million rows of the full dataset mentioned above) locally though it will likely take several minutes to process:
$ python jobs/follower_count.py data/twitter_sample.txt.gz
After editing conf/mrjob-emr.conf you can also run the sample on Elastic MapReduce:
$ python jobs/follower_count.py -c conf/mrjob-emr.conf -r emr \ -o s3://your-bucket/your-output-location --no-output data/twitter_sample.txt.gz
You can also upload data to an S3 bucket and reference it that way:
$ python jobs/follower_count.py -c conf/mrjob-emr.conf -r emr \ -o s3://your-bucket/your-output-location --no-output s3://your-bucket/twitter_sample.txt.gz
You may also download the full dataset and run either the follower count or the histogram job. The following general steps are required:
- Download the whole data file from Kwak, Haewoon and Lee, Changhyun and Park, Hosung and Moon, Sue via bittorrent. I did this on a small EC2 instance in order to make uploading to S3 easier.
- To make processing faster, decompress it, split it in to lots of smaller files (split -l 10000000
- Upload to an S3 bucket.
- Run the full job (it took a little over 15 minutes to read through 1.47 billion relationships and took just over an hour to complete).
python jobs/follower_histogram.py -c conf/mrjob-emr.conf -r emr \ -o s3://your-bucket/your-output-location --no-output s3://your-split-input-bucket/
Speeding things up with PyPy
While there are lots of other things to explore in the data, I also wanted to be able to run PyPy on Elastic MapReduce. Through the use of bootstrap actions, we can prepare our environment to use PyPy and tell MrJob to execute jobs with PyPy instead of system Python. The following need to be added to your configuration file (and vary between 32 and 64 bit):
# Use PyPy instead of system Python bootstrap_scripts: - bootstrap-pypy-64bit.sh python_bin: /home/hadoop/bin/pypy
This configuration change (available in conf/mrjob-emr-pypy-32bit.conf and conf/mrjob-emr-pypy-64bit.conf) also makes use of a custom bootstrap script found in conf/bootstrap-pypy-32bit.sh and conf/bootstrap-pypy-64bit.sh).
A single run of “follower_histogram.py“ with 8 “c1.xlarge“ instances took approximately 66 minutes using Elastic MapReduce’s system Python. A single run with PyPy in the same configuration took approximately 44 minutes. While not a scientific comparison, that’s a pretty impressive speedup for such a simple task. PyPy should speed things up even more for more complex tasks.
Thoughts on Elastic MapReduce
It’s been great to be able to temporarily rent my own Hadoop cluster for short periods of time, but Elastic MapReduce definitely has some downsides. For starters, the standard way to read and persist data during jobs is via S3 instead of HDFS which you would most likely be using if you were running your own Hadoop cluster. This means that you spend a lot of time (and money) transferring data between S3 and nodes. You’re not bringing the data to computing resources like a dedicated Hadoop cluster running HDFS might.
All in all though it’s a great tool for the toolbox, particularly if you don’t have the need for a full-time Hadoop cluster.
Play along at home
All of the source code and configuration mentioned in this post can be found at social-graph-analysis and is released under the BSD license.
6 responses to “Social Graph Analysis using Elastic MapReduce and PyPy”
I would love to see a cost comparison for this approach vs. using your own Hadoop and HDFS on Cluster Compute instances as the size of the job grows. Is there a point at which Elastic Map Reduce is less cost effective?
To be clear, The S3 transfer charges are to get your data initially onto S3. There are no transfer charges between nodes on the cluster.
[…] Delicious popular This entry was posted in yahoo rss and tagged yahoo. Bookmark the permalink. ← 10 […]
Interesting article, I went to grab the dataset out of curiosity but it would seem Twitter have changed their Terms of Service and such data is no longer allowed to be distributed 🙁
Very nice. However, aren’t these tasks mostly IO-bound anyway? I guess that’s why PyPy only gives this small speedup 🙂
The tweets themselves aren’t available, but the social graph itself in the Data section is. Here’s the link: http://an.kaist.ac.kr/traces/WWW2010.html
Most researchers seem to collect their own data with the Twitter Streaming API (http://dev.twitter.com/pages/streaming_api). There are some commercial sources as well such as Gnip (http://gnip.com/) and DataSift (http://datasift.net/).