What kind of documentation can we put in the user manual about this?
Recommend to only decom one rack at a time until we get the issue sorted
out in Hadoop-land?
dlmar...@comcast.net wrote:
BLUF: There exists the possibility of data loss when performing DataNode
decommissioning with Accumulo running. This note applies to installations of
Accumulo 1.5.0+ and Hadoop 2.5.0+.
DETAILS: During DataNode decommissioning it is possible for the NameNode to
report stale block locations (HDFS-8208). If Accumulo is running during this
process then it is possible that files currently being written will not close
properly. Accumulo is affected in two ways:
1. During compactions temporary rfiles are created, then closed, and renamed.
If a failure happens during the close, the compaction will fail.
2. Write ahead log files are created, written to, and then closed. If a failure
happens during the close, then the NameNode will have a walog file with no
finalized blocks.
If either of these cases happen, decommissioning of the DataNode could hang
(HDFS-3599, HDFS-5579) because the files are left in an open for write state.
If Accumulo needs the write ahead log for recovery it will be unable to read
the file and will not recover.
RECOMMENDATION: Assuming that the replication pipeline for the write ahead log
is working properly, then you should not run into this issue if you only
decommission one rack at a time.