Lin Yiqun created HDFS-10227: -------------------------------- Summary: BlockManager should decrease blocksScheduled count for timeout replication Key: HDFS-10227 URL: https://issues.apache.org/jira/browse/HDFS-10227 Project: Hadoop HDFS Issue Type: Bug Reporter: Lin Yiqun Assignee: Lin Yiqun
In {{BlockManager#processPendingReplications}}, it suggests that we could invoke decBlocksScheduled() for timeout replication. {code} /** * If there were any replication requests that timed out, reap them * and put them back into the neededReplication queue */ private void processPendingReplications() { BlockInfo[] timedOutItems = pendingReplications.getTimedOutBlocks(); if (timedOutItems != null) { namesystem.writeLock(); try { for (int i = 0; i < timedOutItems.length; i++) { /* * Use the blockinfo from the blocksmap to be certain we're working * with the most up-to-date block information (e.g. genstamp). */ BlockInfo bi = blocksMap.getStoredBlock(timedOutItems[i]); if (bi == null) { continue; } NumberReplicas num = countNodes(timedOutItems[i]); if (isNeededReconstruction(bi, num.liveReplicas())) { neededReconstruction.add(bi, num.liveReplicas(), num.readOnlyReplicas(), num.decommissionedAndDecommissioning(), getReplication(bi)); } } } finally { namesystem.writeUnlock(); } /* If we know the target datanodes where the replication timedout, * we could invoke decBlocksScheduled() on it. Its ok for now. */ } } {code} The comment seems right. After the timeout items are added to {{neededReplications}}, the blocksScheduled count will be repeated increased when these timeout replications removed from {{neededReplications}} to {{pendingReplications}}. -- This message was sent by Atlassian JIRA (v6.3.4#6332)