EMR vs EC2/Hadoop on AWS

EMR does a lot of things for you that you won't find on standard Hadoop on EC2. Some particularly important ones include

• Copying Hadoop logs from your machines to S3. This is very useful for debugging errors after the cluster has been shut down.
• Running job flows of multiple MapReduce, Pig, or Hive jobs
• Setting sensible configuration defaults based on hardware size you choose
• Access to spot instances for cheaper compute
• Ability to resize clusters dynamically

You'll also find that the EMR S3 filesystem is faster and more reliable than the standard one packaged with Apache Hadoop. It supports Multipart upload, and streams writes directly to S3 rather than buffering to disk first. For a bit more on this, see Tip #5

Additionally, if you do decide to use EC2 directly, I'd recommend using instance-storage instead of EBS for your nodes. There's really no reason to pay the extra cost of EBS for Hadoop; you'll notice that EMR clusters all run on instance-storage nodes as well.

You are correct that EMR uses instance-store backed EC2 instances, rather than EBS. However, there's nothing stopping you from creating an instance-store based instance, packing an AMI and using it for your Hadoop cluster. Using EBS also might not represent a lot of additional costs, depending on your workload and frequency. Also, there's an added cost to the EC2 instance when using it through EMR.

I've been using EMR for two years now and I would highly recommend the service as you don't need to invest time in managing and updating your distribution. If your workload is compatible with EMR (getting data from DynamoDB or S3), I would go for EMR as opposed to EC2/Hadoop.