Another option to add to the list would be EMC's Isilon, though that is a storage hardware appliance. It's a scale-out NAS that has a number of protocols on top, HDFS being one of the many. It doesn't actually run anything that looks like HDFS architecturally, but fakes the NameNode and DataNode RPC calls so it looks like real HDFS to a client. http://www.emc.com/big-data/scale-out-storage-hadoop.htm
I used it a few times while I was working for EMC and it was fantastic, but I never had the chance to try out Accumulo (or HBase). I imagine it might be good and bad in different ways. I saw some "interesting" performance profiles for MapReduce where it performed better in some cases and worse in others and I would expect the same for BigTable access patterns. I think in Accumulo you might see things like compactions speed up significantly (if they don't happen all at once), as single file throughput of a beefy isilon is significantly better than a single drive, and also behind the scenes Isilon doesn't do 3x replication (it does as raid-5-like striping across nodes). -Don On Mon, Nov 18, 2013 at 8:19 PM, <[email protected]> wrote: > MapR works. > > I have a three node cluster with MapR running on my laptop and larger ones > at work. > > Phil > > > I tried Accumulo on QFS today. It can be a drop-in replacement for HDFS. > > > > http://quantcast.github.io/qfs/ > > > > > > I sent a note to the QFS dev team on my results. > > > > https://groups.google.com/forum/#!topic/qfs-devel/VT9ROYrn1tg > > > > > > QFS was very easy to start and play with. Hopefully someone will answer > > my > > questions and we can have another file system option (MapR should already > > work). > > > > -Eric > > > > >
