Cleversafe cuddles up to MapReduce, kicks HDFS out of bed
Hey Hadoop, we think your filesystem sucks
Object storage specialist Cleversafe is after a piece of Big Data analytics action, and has wheeled in MapReduce to make it happen. In the same stroke it rejected HDFS as vulnerable and wasteful of storage capacity.
MapReduce is the analysis part of Hadoop, the flavour-of-the-month open-source data hoarding software. It processes information stored in Hadoop's Distributed Filesystem (HDFS), which employs a single metadata server and makes three copies of files to protect against drive failures and similar upsets.
Object storage systems such as Cleversafe's dsNet are designed to store masses of data on masses of connected nodes. It can self-heal by distributing metadata (and some kind of content hash) around the system of Slicestor nodes. Instead of adding HDFS capabilities alongside dsNet it's adding MapReduce to each Slicestor node.
Cleversafe says it stores only one copy of the MapReduce data instead of three, which is cheaper as capacities grow from terabytes to exabytes and on to petabytes. It isn't vulnerable to a catastrophic loss of data access caused by a metadata server failure because the metadata is distributed and protected - which is better than Hadoop according to Cleversafe.
There is no need for separate MapReduce servers and most Slicestor MapReduce apps will access data in their node rather than going out across the network so the system is more efficient.
Adding Cleversafe nodes adds capacity, both for raw data and metadata, and increases overall performance, the company claimed. The dsNet system can scale up into the petabyte range and beyond that into exabytes, it added.
Lockheed Martin is said to be working with Cleversafe to develop a suitable version of dsNet for US federal government agencies.
The initial MapReduce version of dsNet should be ready by the end of the year in a Cleversafe 3.0 software release. Support for third-party management tools - think Cloudera and HortonWorks - is expected to be added in the first half of 2013 in a 3.1 release. Conveniently, these builds may need Slicestore node hardware upgrades to cope with the added processing, memory and network loads. ®