I apologize for the buzzword heavy title but it was the best I could do. I couldn’t find a good quick start explaining how to get started with Hive so I thought I’d share my experiences.
Anyway, a client of ours came to us needing to analyze a dataset that was about ~200 million rows over 6 months and is currently growing at about 10 million rows a week and increasing. From a reporting standpoint, they were looking to run aggregate counts and group bys over the data and then display the results on charts. Additionally, they were also looking to select subsets of the data and use them later – basically SELECT * FROM table WHERE x AND y AND z.
Obviously, doing the calculations in real time was out of the question so we knew we were looking for a solution that would be easy to use, support the necessary requirements and that would predictably scale with the increasing generation rate of data.
On the surface, MySQL looks like a decent approach but it presents a couple of issues pretty quickly:
- In order for the SUM, GROUP BY, and COUNT queries to be at all useful the MySQL tables would have to be heavily indexed. Unfortunately, due to the write heavy workload of the app this would mean having to copy data into an indexed MySQL database before running any reports.
- Even with indexes, MySQL was pretty awful at selecting subsets of the data from a performance perspective.
- And probably the biggest issue with MySQL is that it doesn’t scale linearly in the sense that if the data is growing at 500 million rows a week you can’t simply “throw more hardware” at it and be done with it.
With requirements in hand we hit the Internet and finally arrived at Hive running on top of Hadoop. Per Wikipedia,
From our perspective, this stack fits our requirements nicely since it doesn’t rely on keeping a second “reporting” MySQL database available, it will handle both sum/count/group by and selecting subets, and probably most importantly it will allow us at least in the near term to scale with the increasing rate of data generation.
“Apache Hadoop is a software framework that supports data-intensive distributed applications under a free license. It enables applications to work with thousands of nodes and petabytes of data. Hadoop was inspired by Google’s MapReduce and Google File System (GFS) papers.”
To grossly over simplify, Hadoop provides a framework that allows you to break up a data intensive task into discrete pieces, run the pieces in a distributed fashion, and then combine the results giving you the results of the completed task. The quintessential example of a task that can be parallelized in this fashion is sorting a *really* big list since the list can be sorted in pieces and then the results can be combined at the end. See Merge Sort
The second piece of the tool chain is Hive. Again via Wikipedia,
“Apache Hive is a data warehouse infrastructure built on top of Hadoop for providing data summarization, query, and analysis.Apache Hive is a data warehouse infrastructure built on top of Hadoop for providing data summarization, query, and analysis.”
Basically, Hive is a tool that leverages the Hadoop framework to provide reporting and query capabilities using a syntax similar to SQL.
That just leaves Sqoop, the app with a funny name and no Wikipedia entry. Sqoop was originally developed by Cloudera and basically serves as an import tool for Hadoop. For my purposes, it allowed me to easily import the data from my MySQL database into Hadoop’s HDFS so I could use it in Hive.
The rest of this post walks you through setting up Hadoop+Hive and analyzing some MySQL data.
Now that you know the players, lets figure out what we’re actually trying to do.
- We want to start a Hadoop cluster to use Hive on.
- Load our data from a MySQL database into this Hadoop cluster.
- Use Hive to run some reports on this data.
- Warehouse the results of this data in MySQL so we can graph it (not that exciting).
Starting the cluster
Theres actually one more tool you’ll need to get this to work – Apache Whirr. Whirr is actually really cool, it lets you automatically start cluster services (Hadoop, Voldermont, etc.) at a handful of cloud platforms (AWS, Rackspace, etc.)
NOTE: We exclusively use AWS for our hosting so everything described here is specific to AWS.
Fisrt, download the latest copy of Whirr – http://www.fightrice.com/mirrors/apache//incubator/whirr/ to your local machine. Whirr should work everywhere but these directions will match up against Linux/OSX the best.
The first thing you’ll need is a Whirr configuration file describing the cluster you want to build. Create a file called hadoop.properties and paste in the following:
whirr.instance-templates=1 hadoop-namenode+hadoop-jobtracker,2 hadoop-datanode+hadoop-tasktracker
whirr.identity=[REPLACE THIS WITH YOUR AWS ID]
whirr.credential=[REPLACE THIS WITH YOUR AWS SECRET]
There isn’t a ton going on in the file but you’ll need to switch out the credential lines for your AWS credentials. Also, you’ll need to double check that the ssh paths are accurate for your account.
The next step, is to actually launch the cluster. To do this run this command – double check the path to your hadoop.properties file is accurate:
./bin/whirr launch-cluster --config hadoop.properties
Just give it a few minutes, you’ll see a bunch of debug info scrolling across your terminal and hopefully a success message once its done. At this point, you’ll have a fully built Hadoop cluster with 3 nodes as described in your properties file ( 1 hadoop-namenode+hadoop-jobtracker,2 hadoop-datanode+hadoop-tasktracker ).
You can see all your nodes by checking out your Whirr cluster directory.
Prepping and loading the cluster
Now that the cluster is up, you’ll need to prep it and then load your data with Sqoop.
One of the most irritating “gotchas” I stumbled across was that Whirr adds the firewall rules necessary for Hadoop to its AWS security group.
Before you do anything, open your EC2 control panel and modify the new Whirr security group (#jcloud-something) so that all of your nodes can connect to each other on port 3306 (MySQL)
The next step is to install mysql-client across the entire cluster since Sqoop uses mysqldump to get at your data. You could manually ssh into every machine but Whirr provides a convenient “run-script” command to do just that.
Create a file called “prepCluster.sh” and put “sudo apt-get -q -y install mysql-client” in it. Then make sure the paths are right and run,
./bin/whirr run-script --script prepCluster.sh --config hadoop.properties
Once its done, you’ll see the aptitude output from all your nodes as they downloaded the MySQL client.
The next step is to install mysql-server, hive, and sqoop on the jobtracker. Doing this is pretty straightforward, look at the .whirr/hadoop/instances file from above and copy the namenode hostname.
Next, ssh in to that machine using your current username as the username. Once you’re in, just run the following to install everything:
sudo apt-get -q -y install sqoop
sudo apt-get -q -y install hadoop-hive
sudo apt-get -q -y install screen
NOTE: You’ll also need the MySQL ConnectorJ library so that Sqoop can connect to MySQL. Download it here and place it in “/usr/lib/sqoop/lib/”
Once everything is done installing, you’ll most likely want to move your MySQL data directories from their default location onto the /mnt partition since it’s much larger. Check out this article for a good walk through. Don’t forget to update AppArmour or MySQL won’t start. Once MySQL is setup, load the data you want to crunch.
Now, you’ll need to use Sqoop to load the data from the MySQL database into Hadoop’s HDFS. While logged into the jobtracker node you can just run the following to do that. You’ll need to swap out the placeholders in the command and change the u/p.
sqoop-import-all-tables --connect jdbc:mysql://[IP of your jobtracker]/[your db_name] --username root --password root --verbose --hive-import
Once it completes, Sqoop will have copied all your MySQL data into Hadoop’s HDFS file system and initialized Hive for you.
Crunching the data
Run “hive” on the jobtracker and you’ll be ready to start crunching your data.
Check out the Hive language manual for more info on exactly what queries you can write.
Once you’ve narrowed down how to write your queries, you can use Hive’s “INSERT OVERWRITE LOCAL DIRECTORY” command to output the results of your query into a local directory.
Then, the next step would be to TAR up these results and use scp to copy the results back to your local machine to analyze or warehouse.
Shutting it down
The final thing you’ll need to do is shut down the cluster. Whirr makes it pretty easy:
./bin/whirr destroy-cluster --config hadoop.properties
Give it a few minutes and Whirr will shutdown the cluster and clean up the EC2 security group as well.
Anyway, hope this walk through proves useful for someone. As always, feedback, questions, and comments are all more than welcome.