It should also be pointed out that the "rollback" process restores your HDFS cluster to the state it was at when it was upgraded by restoring a checkpoint -- it does not "downgrade" your current set of blocks. So if you've pulled new data into HDFS since the upgrade, that will be lost - Aaron
On Wed, Mar 18, 2009 at 11:47 AM, David Ritch <david.ri...@gmail.com> wrote: > Bill, > > Thank you. Unfortunately, that will not work for us, as our hadoop-19 has > long been finalized. > > dbr > > On Wed, Mar 18, 2009 at 2:35 PM, Bill Au <bill.w...@gmail.com> wrote: > > > As long as the version upgrade has not been finalized, here's the > procedure > > I use to downgrade: > > > > > > 1. make sure that the previous version upgrade has not been finalized. > A > > version upgrade cannot be rollback if it has been finalized. > > > > bin/hadoop dfsadmin -upgradeProgress status > > > > 2. stop map-reduce cluster. > > > > bin/stop-mapred.sh > > > > 3. stop all applications and make sure that there are no running tasks. > > 4. stop HDFS cluster. > > > > bin/stop-dfs.sh > > > > 5. rollback version of Hadoop. Install previous version of Hadoop if > > previous version of Hadoop has been removed from the system. > > 6. start HDFS with the rollback option. > > > > bin/start-dfs.sh -rollback > > > > 7. monitor the rollback until it is complete. > > > > bin/hadoop dfsadmin -upgradeProgress status > > > > 8. start map-reduce cluster. > > > > bin/start-mapred.sh > > > > > > > > Bill > > > > On Wed, Mar 18, 2009 at 1:08 PM, David Ritch <david.ri...@gmail.com> > > wrote: > > > > > There is an established procedure for upgrading from one release of > > Hadoop > > > to a newer release. Is there something similar to move back to an > > > lower-numered release? > > > > > > Specifically, we have data in a cloud running Hadoop-19.0. Because of > > > stability issues, we are wondering whether we should move back to 18, > but > > > we > > > don't want to lose our data. Is there a downward migration path? > > > > > > Thanks, > > > > > > David > > > > > >
