Mike, I don't see how SPOF comes into the picture when HA is already present in the releases that also carry Federation and each NN (federated or non) can be assigned further Standby-NN roles. At this point we can stop using the word "SPOF" completely for HDFS. Would do great good for avoiding further FUD around this :)
The federated NameNodes are namespace-divided; and typically I think users may use it to divide applications into using different namespaces each, while using the same set of compute and data nodes. So if HA isn't deployed, an application may be lost but not "portion of the cluster", since DNs will remain unaffected and so would the other namespaces. On Sun, Feb 24, 2013 at 11:23 PM, Michel Segel <michael_se...@hotmail.com> wrote: > I think part of the confusion stems from the fact that federation of name > nodes only splits the very large cluster in to smaller portions of the same > cluster. If you lose a federated name node, you only lose a portion of the > cluster not the whole thing. So now instead of one SPOF, you have two SPOFs. > > The advantage of Federation is that you reduce the amount of memory required > for the NN. > > > Sent from a remote device. Please excuse any typos... > > Mike Segel > > On Feb 24, 2013, at 10:48 AM, Harsh J <ha...@cloudera.com> wrote: > >> Hi, >> >> Federated namenodes are independent of one another (except that they >> both get reports from all the/common DNs in the cluster). It is >> natural to see one roll its edit logs based on its own rate of >> metadata growth, as compared to the other. Their edits, image, etc. >> everything is independent - they also do not know about/talk to each >> other. >> >> Given the above, I'd say (1) is true and (2) is invalid, and (3) can >> be found in PDFs attached to >> https://issues.apache.org/jira/browse/HDFS-1052. >> >> On Sun, Feb 24, 2013 at 8:39 PM, YouPeng Yang <yypvsxf19870...@gmail.com> >> wrote: >>> Hi All >>> >>> I'm testing the HDFS Federation.I have 2 namenodes in my cluster. >>> I find that it is Rolling Edit Log continuously on one namenode,however >>> the other one changes nothing. >>> >>> My question : >>> 1. Is it the right situation. >>> 2. I have thought that the two namenodes shoud keep concurrency. >>> why they get differences. >>> 3. is there any advanced docs about HDFS Federation. >>> >>> thanks. >>> >>> Regards. >> >> >> >> -- >> Harsh J >> -- Harsh J