On Mon, Feb 7, 2011 at 7:01 AM, Ted Dunning <tdunn...@maprtech.com> wrote: > HDFS does not provide for keyed access to data, nor column oriented access > when only a subset of related data is needed. Also, HDFS is a write-once > file system while hbase provides random updates.
Note that the write-once aspect is a peculiarity of HDFS, GFS doesn't have this limitation. Bigtable was originally invented to provide low-latency key-value oriented serving. Without Bigtable, an application had to load its data directly from GFS (which is a high throughput but high latency distributed file system) to serve it out of memory. For large data sets that can't fit on a single machine, this is actually hard to do, especially if you want to be able to change individual data items. HBase is solving the same problem as Bigtable. The main goal is low latency serving for a very large number of fairly small data items. -- Benoit "tsuna" Sigoure Software Engineer @ www.StumbleUpon.com