I know the title of this post is a mouthful, but it’s the fun of pushing envelope of existing technologies. What I am looking to do is take my log data stored on S3 (which is in compressed JSON format) and run queries against it. In order to not have to learn everything about setting up Hadoop and still have the ability to leverage the power of Hadoop’s distributed data processing framework and not have to learn how to write map reduce jobs and … (this could go on for a while so I’ll just stop here). For all these reasons, I choose to use Amazon’s Elastic Map infrastructure and Pig.
Describing all these technologies is beyond the scope of this article. I will talk you through how I was able to do all this with a little help from the Pig community and a lot of late nights. I will also provide an example Pig script detailing a little about how I deal with my logs (which are admittedly slightly abnormal). I will also be making some assumptions here. Each time I make a large assumption, I will let you know.
First off, I am going to assume that you have an Amazon Web Services account (AWS) and you have also signed up for Elastic Map Reduce (EMR). For all this, I followed the instructions in this video by Ian @ AWS to get me going. Now SSH into the machine so we can get started.
As of the time of this writing, EMR is using Hadoop 0.20 and Pig 0.6. Everything I am going to talk about is for Pig 0.6. With any luck, upgrades to Pig will have taken a lot of this into account. Once you are on the EMR master host, type the following commands to get elephant-bird downloaded. We are going to use it to build a jar that will parse our JSON (big thanks to Dmitriy for all the help here). Note: The reason we are pulling with wget as opposed to git directly is that we want the jsonloader branch and this is just easier.
$ mkdir git && mkdir pig-jars
$ cd git && wget --no-check-certificate https://github.com/kevinweil/elephant-bird/tarball/eb1.2.1_with_jsonloader
$ tar xzf eb.1.2.1_with_jsonloader && cd elephant-bird
$ cp lib/google-collect-1.0.jar ~/pig-jars && cp lib/json-simple-1.1.jar ~/pig-jars
$ ant nonothing
$ cd build/classes
$ jar -cf ../elephant-bird-1.2.1-SNAPSHOT.jar com
$ cp ../*.jar ~/pig-jars
At this point we should have 3 jars in the pig-jars directory. I created an S3 bucket for myself and put the jars in there so I only have to do the compilation once. From here on in, I will be referencing those jars using my s3 bucket. If you like how my logs are organized, then I strongly recommend checking out Cloudera Flume for log aggregation. <shameless plug>I also wrote a blog post here on getting it going.</shameless plug>
Another item worthy of note is that I store all my logs in gzip format. Although this isn’t the best format for Hadoop in the long run because it can’t be split into chunks, it’s what I used. I had trouble getting everything going because Pig doesn’t decompress files when running in local mode. Please learn from that mistake of mine.
Now let’s get into the code a little. First thing we do is register all the jars necessary to parse JSON. This is done using our jars that we put directly on S3. Then we load up the JSON into maps. This is done using the JsonLoader(). You have to use the full path to it (which is listed in the code sample). Now the “interesting” thing about my log files is that they have 3 distinct types of log lines in them. Types i,b, and c. Each log line type has a different meaning so I sort them into 3 groups with the SPLIT command and some conditionals.
Now that I have my data broken out into the 3 buckets, I can start doing what I want with them. Let’s say that in log type i, there is a widget_value. And that widget_value is a string of any number. To show the top 5 values of that widget_value has, I just pull out all instances of widget_value in i. Then I iterate over those values and group them together by type (thus getting aggregate values). And finally I sort them in descending order and show only the top 5.
GeSHi Error: GeSHi could not find the language pig (using path /nas/content/live/elubow/wp-content/plugins/codecolorer/lib/geshi/) (code 2)
Last thing I want to share is some tips on getting everything going with Pig:
- Start small and continue small until everything is working
- Use subsets of your data that you have a good idea of what the results are going to be before you run your queries
- Step through you queries to ensure each step is doing what you think it’s doing
- Cast your data types to avoid weird behaviors. Map doesn’t always leave your variables in the type you want/expect
And I can’t forget to say thanks for all the help to the people who hang out in #hadoop-pig on irc.freenode.net.