HDFS-630 is kinda critical to us over in hbase. We'd like to get it into
0.21 (Its been committed to TRUNK). Its probably hard to argue its a
blocker for 0.21. We could run a vote. Or should we just file it against
0.21.1 hdfs and commit it after 0.21 goes out? What would folks suggest?
Without it, a node crash (datanode+regionserver) will bring down a second
regionserver, particularly if the cluster is small (See HBASE-1876 for
description of the play-by-play where NN keeps giving out dead DN as place
to locate new blocks). Since the bulk of hbase clusters are small --
whether evaluations, test, or just small productions -- this issue is an
important fix for us. If the cluster is of 5 or less nodes, we'll probably
recover but there'll be a period of churn. At a minimum mapreduce jobs
running against the cluster will fail (usually some kind of bullk upload).