As always, your specific situation, hardware, risk profile, etc are unique to others, but let's revisit a couple of things first. HDFS metadata is physically persisted in "image" and "edits" files. In a HA configuration, the two NN processes are the ones who write out the image files and the JournalNodes (JN) are the processes that persist the edits files. Even without any kind of HBA/RAID configuration (not a bad place to be, but they aren't necessarily bad/wrong either) we get some pretty good spread of recording this information in multiple places (2x copies of image files and 3+x copies of the edits files). The historical rule of thumb was to make sure the NN data (especially before HA) was to write to two local disk and one soft-mounted NFS directory as we simply did not want to ever have a "bunker scene". I, again my strong personal believe, would still suggest that you record to at least two local disks (maybe you decide your RAID approach satisfies this) as well as NFS directory. I'd even follow this along with periodic backups of both the image and edits data files. My thinking on this topic is captured in a blog posting at https://martin.atlassian.net/wiki/x/EoC3Ag. Good luck!
... View more
