[ https://issues.apache.org/jira/browse/HDFS-6791?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14080580#comment-14080580 ]
Ming Ma commented on HDFS-6791: ------------------------------- Use the following steps to repro: 1. Create a situation where node A has several blocks with replication factor equal to 1. 2. Start to decommission node A. Right after the decommission process starts, kill the DN JVM on node A. 3. Wait until NN marks node A dead. After that, NN will mark the node as decommissioned. That is because when there is no block left for the DN, decommission is considered done. Given node A hasn't finished copying its blocks, there will be missing blocks at this point. {noformat} BlockManager.java boolean isReplicationInProgress(DatanodeDescriptor srcNode) { boolean status = false; ... final Iterator<? extends Block> it = srcNode.getBlockIterator(); while(it.hasNext()) { ... // set status if there is block under replication } ... return status; } {noformat} 4. Restart the node A. Upon datanode registration, given the node is already in decommissioned state, no decommission is performed. So node A will be in decommissioned state and its blocks aren't copied to other nodes. Some ideas on how to fix it, 1. When a DN becomes dead during decommission, NN can continue to mark the DN "decommission-in-progress". That will allow the DN to resume the decommission process when it rejoins the cluster. 2. Another approach could be to relax the definition of "decommissioned" state so that when BlockManager choose source datanode for replication, it could choose "decommissioned" under special condition, e.g., there is no other datanode available. Suggestions? > A block could remain under replicated if all of its replicas are on > decommissioned nodes > ---------------------------------------------------------------------------------------- > > Key: HDFS-6791 > URL: https://issues.apache.org/jira/browse/HDFS-6791 > Project: Hadoop HDFS > Issue Type: Bug > Reporter: Ming Ma > > Here is the scenario. > 1. Normally before NN transitions a DN to decommissioned state, enough > replicas have been copied to other "in service" DNs. However, in some rare > situations, the cluster got into a state where a DN is in decommissioned > state and a block's only replica is on that DN. In such state, the number of > replication reported by fsck is 1; the block just stays in under replicated > state; applications can still read the data, given decommissioned node can > served read traffic. > This can happen in some error situations such DN failure or NN failover. For > example > a) a block's only replica is node A temporarily. > b) Start decommission process on node A. > c) When node A is in "decommission-in-progress" state, node A crashed. NN > will mark node A as dead. > d) After node A rejoins the cluster, NN will mark node A as decommissioned. > 2. In theory, NN should take care of under replicated blocks. But it doesn't > for this special case where the only replica is on decommissioned node. That > is because NN has the policy of "decommissioned node can't be picked the > source node for replication". > {noformat} > BlockManager.java > chooseSourceDatanode > // never use already decommissioned nodes > if(node.isDecommissioned()) > continue; > {noformat} > 3. Given NN marks the node as decommissioned, admins will shutdown the > datanode. Under replicated blocks turn into missing blocks. > 4. The workaround is to recommission the node so that NN can start the > replication from the node. -- This message was sent by Atlassian JIRA (v6.2#6252)