[HDFS-14768] EC : Busy DN replica should be consider in live replica check.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.0.2
• Fix Version/s: 3.3.0
• Component/s: datanode, erasure-coding, hdfs, namenode
• Labels: patch
• Target Version/s: 3.3.0
• Assignee: guojh
• Reporter: guojh
• Created: 23/Aug/19 05:15
• Updated: 09/Mar/22 07:37
• Resolved: 01/Nov/19 16:56

Description

Policy is RS-6-3-1024K, version is hadoop 3.0.2;

We suppose a file's block Index is [0,1,2,3,4,5,6,7,8], And decommission index[3,4], increase the index 6 datanode's

pendingReplicationWithoutTargets  that make it large than replicationStreamsHardLimit(we set 14). Then, After the method chooseSourceDatanodes of BlockMananger, the liveBlockIndices is [0,1,2,3,4,5,7,8], Block Counter is, Live:7, Decommission:2. 

In method scheduleReconstruction of BlockManager, the additionalReplRequired is 9 - 7 = 2. After Namenode choose two target Datanode, will assign a erasureCode task to target datanode.

When datanode get the task will build  targetIndices from liveBlockIndices and target length. the code is blow.

// code placeholder
targetIndices = new short[targets.length];
private void initTargetIndices() { 
  BitSet bitset = reconstructor.getLiveBitSet();
  int m = 0; hasValidTargets = false; 
  for (int i = 0; i < dataBlkNum + parityBlkNum; i++) {  
    if (!bitset.get) {    
      if (reconstructor.getBlockLen > 0) {
       if (m < targets.length) {
         targetIndices[m++] = (short)i;
         hasValidTargets = true;
        }
      }
    }
 }

targetIndices[0]=6, and targetIndices[1] is aways 0 from initial value.

The StripedReader is  aways create reader from first 6 index block, and is [0,1,2,3,4,5]

Use the index [0,1,2,3,4,5] to build target index[6,0] will trigger the isal bug. the block index6's data is corruption(all data is zero).

I write a unit test can stabilize repreduce.

// code placeholder
private int replicationStreamsHardLimit = DFSConfigKeys.DFS_NAMENODE_REPLICATION_STREAMS_HARD_LIMIT_DEFAULT;

numDNs = dataBlocks + parityBlocks + 10;

@Test(timeout = 240000)
public void testFileDecommission() throws Exception {
  LOG.info("Starting test testFileDecommission");
  final Path ecFile = new Path(ecDir, "testFileDecommission");
  int writeBytes = cellSize * dataBlocks;
  writeStripedFile(dfs, ecFile, writeBytes);
  Assert.assertEquals(0, bm.numOfUnderReplicatedBlocks());
  FileChecksum fileChecksum1 = dfs.getFileChecksum(ecFile, writeBytes);

  final INodeFile fileNode = cluster.getNamesystem().getFSDirectory()
      .getINode4Write(ecFile.toString()).asFile();

  LocatedBlocks locatedBlocks =
      StripedFileTestUtil.getLocatedBlocks(ecFile, dfs);

  LocatedBlock lb = dfs.getClient().getLocatedBlocks(ecFile.toString(), 0)
      .get(0);

  DatanodeInfo[] dnLocs = lb.getLocations();

  LocatedStripedBlock lastBlock =
      (LocatedStripedBlock)locatedBlocks.getLastLocatedBlock();

  DatanodeInfo[] storageInfos = lastBlock.getLocations();

  //
  DatanodeDescriptor datanodeDescriptor = cluster.getNameNode().getNamesystem()
      .getBlockManager().getDatanodeManager().getDatanode(storageInfos[6].getDatanodeUuid());

  BlockInfo firstBlock = fileNode.getBlocks()[0];
  DatanodeStorageInfo[] dStorageInfos = bm.getStorages(firstBlock);

  // the first heartbeat will consume 3 replica tasks
  for (int i = 0; i <= replicationStreamsHardLimit + 3; i++) {
    BlockManagerTestUtil.addBlockToBeReplicated(datanodeDescriptor, new Block(i),
        new DatanodeStorageInfo[]{dStorageInfos[0]});
  }
  assertEquals(dataBlocks + parityBlocks, dnLocs.length);
  int[] decommNodeIndex = {3, 4};

  final List<DatanodeInfo> decommisionNodes = new ArrayList<DatanodeInfo>();
  // add the node which will be decommissioning
  decommisionNodes.add(dnLocs[decommNodeIndex[0]]);
  decommisionNodes.add(dnLocs[decommNodeIndex[1]]);
  decommissionNode(0, decommisionNodes, AdminStates.DECOMMISSIONED);

  assertEquals(decommisionNodes.size(), fsn.getNumDecomLiveDataNodes());
  bm.getDatanodeManager().removeDatanode(datanodeDescriptor);
  //assertNull(checkFile(dfs, ecFile, 9, decommisionNodes, numDNs));

  // Ensure decommissioned datanode is not automatically shutdown
  DFSClient client = getDfsClient(cluster.getNameNode(0), conf);
  assertEquals("All datanodes must be alive", numDNs,
      client.datanodeReport(DatanodeReportType.LIVE).length);

  FileChecksum fileChecksum2 = dfs.getFileChecksum(ecFile, writeBytes);
  Assert.assertTrue("Checksum mismatches!",
      fileChecksum1.equals(fileChecksum2));

  StripedFileTestUtil.checkData(dfs, ecFile, writeBytes, decommisionNodes,
      null, blockGroupSize);
}

 

Issue Links

• is related to: HDFS-15186 Erasure Coding: Decommission may generate the parity block's content with all 0 in some case
• is related to: HDFS-16479 EC: NameNode should not send a reconstruction work when the source datanodes are insufficient
• relates to: HDFS-16497 EC: Add param comment for liveBusyBlockIndices with HDFS-14768