With a series of announcements made for R inclusion in Hadoop distribution and relational database software packages, suddenly R is becoming the flavor of predictive analytics. It’s time to preview R offerings for Hadoop world and where the things seemed to be heading.
Hadoop distribution with R
R was developed around 1993 by Ross Ihaka, Robert Gentleman as a statistical computing language and environment. Open sourced in 1996, the free software is available under GNU licence from r-project.org/CRAN (Comprehensive R Archive Network) servers. Most of the commercial Hadoop distributions today have a R package inclusion or have similar plans. Cloudera and Hortonworks have recently announced partnership with Revolution Analytics, a popular vendor offering enterprise edition. IBM has been offering R as statistical analysis package as part of Infosphere BigInsights for couple of years now. Also, IBM Netezza/Pure Data hardware systems offer R package as a software inclusion. While IBM SPSS is more of a bete noir of R, it also offers R integration. SAS, not a Hadoop focused company, but a prominent player in predictive analytics also offers R integration. Intel which ventured into Hadoop distributions also has recently announced a partnership with Revolution Analytics. MapR, another leading Hadoop distribution company has a partnership with Revelytix which offers Loom software with R integration. Greenplum/Pivotal also offer R integration as part of its database and Hadoop offering.
R on Hadoop integration hacks
Among the non Hadoop distribution options, RHIPE (pronounced hree-pay') is the R and Hadoop Integrated Programming Environment. RHive is another tool from a Korean company(NexR) which offers a way to integrate R with Hive for querying HDFS. Oracle among the big players offers it own distribution of R software. Further, it provides a connector for integrating R and Hadoop. Amazon EMR also offers you an option of running R programs now. If you are adventurous yourself and would like to hack your own R with Hadoop nodes you could either copy libraries to each node or use a JDBC connection approach from R with Hive or HBase. With SQL on Hadoop tools offering JDBC support, hadoopsphere.com believes there will be more than a likely chance of integrating R with Cloudera Impala, Apache Tez, IBM BigSQL and the likes in future.
The following extract from a Google paper lists down one of the approach to use R with MapReduce.
1.First, the user's code calls google.apply() with a list of inputs and a provided function, FUN. An archive is dynamically created on the client including R and all of the needed libraries and then staged to the cluster management system in a datacenter with available resources. FUN and its environment are serialized and written out to a Bigtable in that datacenter.
2.Second, workers tasks are spawned using the dynamically generated virtual machines providing access to all of the R packages that were loaded in the calling instance's R session. These workers read in the serialized environment from the Bigtable, execute the provided function over a unique element of the input list, and write out the serialized results to a different column of the Bigtable.
3.Third, and finally, the calling R instance reads back in the serialized return values from each worker task, performs the necessary error handling, and returns the computed list to the google.apply() caller
R on an unprecedented journey
It is no secret that the predictive modeling market is going to get hotter and hotter in the days to come. Inclusion of R in technology stacks offers one of the convenient ways of doing ensemble modeling. The amount of buzz suddenly around R is unprecedented in its lifetime. While there are a good number of folks having basic education on R, inclusion of another language in stacks is bound to complicate the skilled resource requirement. The commercial stack for R has been progressing rapidly but the open source counterpart has not benefited much from this pace. For instance, the open source R continues to rely on in-memory, single threaded model. There are limited enterprise R options and dependence on one particular commercial vendor leads to fears of vendor lock-in besides the fear of vendor acquisition considering its size and history. However, with ex-SPSS and ex-SAS folks forming the top management in Revolution Analytics, there is lesser doubt that R’s enterprise show is being run by experienced hands. Price point still remains one of the main winners for R vis-à-vis SPSS and SAS. We would not be surprised if more commercial vendors for R crop up which would be a good thing rather than a bad thing as long as the open source version continues to benefit from the commercial run.