Cloudera vs Hortonworks vs MapR – Has MapR Already Won This Contest?

If you're part of the decision making process in how to get started with Big Data then you've probably already discovered that open source Apache Hadoop has a big lead. You also know that unless you've got a particularly deep NoSQL team or want to write your own proprietary code that the free open source version of Hadoop is not for the great majority of businesses.  You want the benefits of open source but with some support and some capabilities that tie the open source components together in an enterprise quality database system.  You want a Hadoop Distribution.

There are many companies providing this middleman distribution service that take on the load of managing updates, providing support, training, and consulting, and even adding some innovations of their own that make Hadoop reasonable for an enterprise to handle.  While there are plenty of these Hadoop distributors to choose from, it's most likely that your first choice will emerge from among the current big three, Cloudera, Hortonworks, and MapR.

Let's step back a minute though and talk about the decision tree that led up to this point.  Unless you were building an app that clearly needed a NoSQL graph database you were basically selecting from among the three remaining NoSQL options, Key-Value, Document Oriented, and Columnar databases.  (For more on these differentiations see our earlier blog series "9 Lessons for Starting a Big Data Initiative and Selecting the Right NoSQL Tools").

Keep in mind that most enterprises that are using Big Data have implemented three or four of these for different uses.  Looking back to 2013, Disney uses Cassandra, Hadoop, and Mongo.  Netflix uses Cassandra, Hbase, and SimpleDB.  Twitter uses Cassandra, FlockDB, Hbase, and MYSQL.  Mendeley uses Hbase, Mongo, Solr, and Voldemort.  The term for this is "Polyglot Persistence" meaning using the best database type for each circumstance.

I argued in an earlier blog that folks starting out with Big Data were better off with the flexibility of a document-oriented NoSQL database like Mongo.  Mongo had an early lead in the marketplace and still has a lead in installed base.  But while Mongo does offer all the appeal of NoSQL distributed processing across inexpensive resources it is not open-source and not technically HDFS.  Then there were the knocks on basic Key-Value Hadoop like it could only be run in batch, wasn't suited for on-line real time apps, and offered only eventual consistency.  Well it's a fast changing world in Big Data and here at the start of 2015 pretty much all of those restrictions on Hadoop have gone away.  Frequently that's thanks to the add-ons and additional developments offered by the Hadoop distributors.

For example, a Hadoop distribution includes both the traditional key-value store as well as the H-base version which is columnar and well suited to OLAP.  Keep in mind that these are two separate databases requiring two separate implementations but you can get them both from one vendor in an integrated package.  Many of the other impediments have also been effectively removed so that Hadoop can be used in on-line streaming applications, can offer the double level indexing of document-oriented databases, and has largely resolved the issue of eventual consistency among many others.

All of this being true, Hadoop is clearly the horse to ride if you're getting started with Big Data and one of the big three Hadoop distribution companies is most likely to be your provider. So which one to choose?  Like any three companies competing in a relatively narrow marketplace, each has staked out a somewhat different positioning, each with its benefits and limitations.

Cloudera: Formed in 2008 it was the first of the Hadoop distribution companies and still has the largest user base with the most number of clients.  Its core distribution is based on 100% open source Apache Hadoop and it has added its own proprietary Cloudera Management Suite to automate installation and provide other administrative tools and conveniences.  Cloudera users operate under a commercial license and Cloudera is moving to position itself as the ‘enterprise data hub' reducing or eliminating the need for traditional data warehouses.

Hortonworks: Growing quickly from its start in 2011, Hortonworks is the only commercial vendor positioning itself to provide complete open source Hadoop.  They have intentionally not developed proprietary software and use Ambari for management, Stinger for queries, and Solr for searches.  The Hortonworks distribution remains completely free to download and is easy to install.  Users operate under an open source license.

MapR:  MapR has chosen a slightly different path.  It offers many proprietary features including adding administrative and management layers on top of Hadoop and seeks to be the most enterprise ready in terms of improving reliability, providing technical assistance, and expanding capabilities.  It has customized underlying HDFS into its own proprietary version MapRFS that is intended to improve efficient management of data, reliability, and ease of use.  It is available in a free version but the upgrade versions contain many of the desirable proprietary features under commercial license.

So how has all this worked out in terms of features and capabilities? This 2014 table by Robert Schneider, in the Hadoop Buyer's Guide provides a good summary.

It doesn't take much analysis to see that MapR is clearly in the lead based on features and capabilities.  All of this led us to wonder whether this race is essentially over and that MapR is clearly the winner.  And it may well be.  But are there really any reservations one should consider before picking MapR?  Perhaps a few.

Proprietary File System: The principal objection quoted by some non-selectors is the proprietary file system which is not pure HDFS.  The issue according to this group is to avoid becoming locked in to one provider's proprietary system so that later switching will be less painful.  If you are a firmly committed open source shop this may be a legitimate reason.  If not, you will need to weigh the many advantages of MapR against its competitors.

Not Actually Free: The advanced proprietary modules and capabilities all have a cost associated.  This is true.  It's frankly difficult to build a Hadoop Distribution company on just training, consulting, and technical support.  Hadoop may have been the proverbial free lunch in the past but this is likely to fade away as more clients find the cost/value proposition of ‘enterprise ready' Hadoop to be financially attractive.

Mutable Keys: The MapR file system allows mutable keys while HDFS does not.  Without getting too technical, the idea of being able to change an established key (mutability) is risky and potentially endangers the ability of all previously built applications to use the data if the keys were inadvertently changed.  MapR argues that there are actually advantages to mutability and that good data management practices eliminate this risk.

We leave it to the reader to evaluate these strengths and weaknesses.  Surely if you are making this evaluation you will need to revisit all of these capabilities since the market and Hadoop itself is changing so fast.  But for now it certainly looks like MapR is well out in front of the pack if it hasn't already won the race for enterprise users.

• PRIVATE WiFi Launches Video Challenge Facebook Contest, Invites Creative Filmmakers to Win $500 Prize
• BlueData Grows Big Data Partner Ecosystem
• Datameer and Hortonworks Partner to Advance Apache Hadoop Solutions