We are planning to setup a namenode failover. For that, we are not sure whether the attribute "dfs.name.dir" if given a comma-delimited set of directories, would write to all such directories synchronously ?
Because if it does so synchronously then we can't have any data loss in any case. There might be some latency due to this. But if it does so asynchronously, then we have to synch it up using any synchronisation tool such as rsynch etc. so as to make sure that no data loss occur. Raghu Angadi wrote: > > Ankur Goel wrote: >> What should happen ideally is when the namenode fails, secondary >> namenode should automatically pickup from the last checkpoint without >> hindering normal operations. The data-nodes should also detect this >> change and upgrade their configurations accordingly. Not sure about the >> behavior of JobTracker and TaskTracker in the middle of a running job. >> I think we should have a separate jira for this. > > I think a new feature to have a 'hot' or 'warm' replica of NameNode is > under the works. It does pretty much exactly what you want. Not sure > about more details or when the jira is going to be filed. > > Raghu. > > -- View this message in context: http://www.nabble.com/Automatic-recovery-Mechanism-for-namenode-failure...-tp18606291p18739599.html Sent from the Hadoop core-dev mailing list archive at Nabble.com.
