Hi HDFS devs, Would you mind taking a look at HDFS-2691? This JIRA deals with the case where the NN fails over, and then a lease needs to be recovered. Since we don't persist any locations to the edit log, the SBN will have no record of where the under-construction block is, so the recovery will accidentally truncate the file unless we do something about it. I've implemented one approach there, but would like your opinions if you have any better ideas. I have been testing this approach with HBase on a small cluster running with replication=1, and haven't seen any issues yet.
Konstantin - I think the BN-based approach is probably susceptible to the same issue, so curious if you've solved it in your implementation. Thanks -Todd -- Todd Lipcon Software Engineer, Cloudera