When a failure occurs while replicating a block from a source DataNode to a target DataNode, the target node keeps an incomplete on-disk copy of the block in its temp data directory and an in-memory copy of the block in ongoingCreates queue. This causes two problems:
1. Since this block is not (should not) be finalized, NameNode is not aware of the existence of this incomplete block. It may schedule replicating the same block to this node again, which will fail with a message: "Block XX has already been started (though not completed), and thus cannot be created."
2. Restarting the datanode moves the blocks under the temp data directory to be valid blocks, thus introduces corrupted blocks into HDFS. Sometimes those corrupted blocks stay in the system undetected if it happens that the partial block and its checksums match.
A failed block replication should clean up both the in-memory & on-disk copies of the incomplete block.