On Mon, May 31, 2010 at 12:31 AM, 史英杰 <[email protected]> wrote:
> Hi, All > I learned from some forums that HBase focuses CA, which are two properties > in CAP. Does HBase support strong consistency in the data replicas, which > mechanism does it use to guarantee consistency? > Yes, HBase is strongly consistent. Each row is hosted by a single region server at a time, and a combination of row locks and multiversion concurrency control is used to provide consistency within a row. During failover, we take care to not allow writes from a new regionserver until the previous one has been blocked out. There are currently a couple of rare bugs where this doesn't work quite right (see HBASE-2231 for example) but they are all blockers for the next release. Replication is taken care of at the HDFS layer. In the next release of HBase it will also be completely durable - no edit will be acked to the client until it has been flushed to three HDFS replicas. -Todd -- Todd Lipcon Software Engineer, Cloudera
