Running Hadoop in the Cloud

Running Hadoop in the Cloud

With the growing popularity of cloud computing enterprises are seriously looking at moving workloads to the cloud. There are issues around  multi-tenancy, data security, software  license, data integration etc that have to be considered before enterprises cam make this shift.  Even then, not all workloads can be easily moved to the  cloud. In recent years, hadoop has gained a lot of interest as a big data technology that can help enterprises, cost effectively  store and analyze massive amounts of data. As enterprises start evaluating hadoop one of the questions frequently asked is “Can we run hadoop in the cloud?”.

To answer this, the following key aspects of the hadoop infrastructure is important to understand:

1. Hadoop best runs on physical servers. A hadoop cluster comprises of a master node called the Name Node and multiple child nodes called Data Nodes. These data nodes are separate physical servers with dedicated storage ( like your PC hard drive), instead of  a common shared storage.

2. Hadoop is “Rack Aware” – Hadoop data node (servers) are installed in racks. Each rack typically contains multiple data node servers with a top of rack switch for network communication. “Rack awareness” means that the Name Node, knows where each data node server is and in which rack.  This  ensures that hadoop can write data to 3 (default) different data nodes, that are not on the same physical rack (prevents data loss due to data node and rack failure). When a Map Reduce job needs access to data blocks, the name node ensures that the job is assigned to the closest data node that contains the data thereby reducing the network traffic.  The hadoop system admin manually maintains this rack awareness information for the cluster. Since the hadoop cluster has a  lot of network traffic it is recommended that the they be isolated into their own network, instead of using VLAN (refer: Brad hedlund’s article on hadoop rack awareness –

Options for running Hadoop in the Cloud

  • Hadoop as a Service in the Public Cloud : –  Hadoop distributions (Cloudera CDH, IBM BigInsights, Map-R, Hortonworks ) can be launched and run on the public clouds like AWS, Rackspace, MS Azure, IBM SmartCloud etc .. which offer Infrastructure as a Service (IAAS) . In a public cloud you are sharing the infrastructure with other customers. As a result, you have very limited control over which server the VM is being spun up and what other VM’s ( yours or other customers) are running on the same physical server. There is no “rack awareness”  that  you have access to and can configure in the Name Node. The performance and availability of the cluster may be affected as you running  on VM. Enterprises can use and pay for these hadoop clusters on demand.  There are options for creating your own private network using VLAN, but hadoop cluster performance recommendation is to have a separate isolated network because of high network traffic between nodes. In all the cases with the exception of the AWS EMR, you have to install and configure the hadoop cluster on the cloud. 
    • Map-Reduce as a Service – Amazon’s EMR (Elastic Map Reduce) provides a quick and easy way to run Map-Reduce jobs without having to install a hadoop cluster on its cloud.  This can be a good way to develop hadoop programming  expertise internally within your organization or if you only want to run map reduce jobs in your workloads.
    • Hadoop on S3 : You can run Hadoop using Amazon’s S3  instead of HDFS to store data. Performance of S3 is slower than HDFS, but it provides other features like bucket versioning and elasticity as well as its own data loss protection  schemes. This may be an option if your data is already being stored in S3 for your business (e.g. Netflix uses a hadoop cluster using S3)
    • Hadoop in private Cloud:  We have the same set of considerations for a private cloud deployment for hadoop as well. However in case of a private cloud, you may have more control over your infrastructure that will enable you to provision bare metal servers or create a separate isolated network for your hadoop clusters.  Some of these private cloud solutions also provide a Paas layer that offers pre build patterns for deploying hadoop clusters easily (e.g. IBM offers patterns for deploying BigInsights on their SmartCloud Enterprise ). In addition you also have an options of deploying  a “Cloud in a Box”  like the IBM Pure Data/Pure Apps, which is hadoop ready in your own data center. The big reason for private cloud deployment, would be around data security and access control for your data as well better visibility and control of your hadoop infrastructure.

Key things to consider before deploying hadoop cluster in the cloud:

  • Enterprise should evaluate the security criteria for deploying workloads in public cloud, before moving any data into the hadoop cluster.  Hadoop cluster security is very limited. There is no native security for data that will satisfy enterprise data security requirements around SOX, PII, HIPPA etc.
  • Evaluate hadoop distributions that you would want to use and the operating system standards of your enterprise. Preferably go with distributions that are close to the open source apache distributions. All hadoop distributions typically run on Linux.  Hortonworks provides a hadoop distribution for Windows that is currently available on MS Azure cloud.
  • When using AWS, be aware that using hadoop with S3 would tie you to the Amazon’s cloud. For open standards , look at OpenStack based cloud providers like Rackspace, IBM smart cloud, HP etc.
  • Look at the entire hadoop ecosystem and not just the basic hadoop cluster. The value from hadoop is the analytics and data visualization that can be applied on a large data sets. Ensure that the tools you want to use for analytics (e.g. Tableau, R, SPSS etc) are available for use on the cloud provider.
  • Get an understanding on where the data to be loaded into hadoop comes from. Are you going to load data from your internal systems that are not on the cloud or if the data is already in the cloud. Most public cloud provides charge for data transmission fees  if you are moving data back and forth.
  • Hadoop clusters on VM will be slow. You may be able to use these for Dev and test clusters. VMware’s project Serengeti is trying to address the  deployment of hadoop clusters on virtual machines without taking a performance hit. However with this approach you will be  tied to VMware’s hypervisor which should be a criteria to consider when selecting a cloud provider.