Apache
/
hadoop
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

HDFS-10301. Interleaving processing of storages from repeated block reports causes false zombie storage detection, removes valid blocks. Contributed by Vinitha Gankidi.

This commit is contained in:
Vinitha Reddy Gankidi 2016-07-30 17:01:49 -07:00 committed by Konstantin V Shvachko
parent 379d587d23
commit a76eb7fa34
10 changed files with 264 additions and 93 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/protocol/BlockListAsLongs.java
View File

@ -57,6 +57,34 @@ public abstract class BlockListAsLongs implements Iterable<BlockReportReplica> {
public Iterator<BlockReportReplica> iterator() {
return Collections.emptyIterator();
}
@Override
public boolean isStorageReport() {
return false;
}
};
// STORAGE_REPORT is used to report all storages in the DN
public static final BlockListAsLongs STORAGE_REPORT = new BlockListAsLongs() {
@Override
public int getNumberOfBlocks() {
return -1;
}
@Override
public ByteString getBlocksBuffer() {
return ByteString.EMPTY;
}
@Override
public long[] getBlockListAsLongs() {
return EMPTY_LONGS;
}
@Override
public Iterator<BlockReportReplica> iterator() {
return Collections.emptyIterator();
}
@Override
public boolean isStorageReport() {
return true;
}
};
/**
@ -172,6 +200,13 @@ public abstract class BlockListAsLongs implements Iterable<BlockReportReplica> {
*/
abstract public long[] getBlockListAsLongs();
/**
* Return true for STORAGE_REPORT BlocksListsAsLongs.
* Otherwise return false.
* @return boolean
*/
abstract public boolean isStorageReport();
/**
* Returns a singleton iterator over blocks in the block report. Do not
* add the returned blocks to a collection.
@ -305,6 +340,11 @@ public abstract class BlockListAsLongs implements Iterable<BlockReportReplica> {
return longs;
}
@Override
public boolean isStorageReport() {
return false;
}
@Override
public Iterator<BlockReportReplica> iterator() {
return new Iterator<BlockReportReplica>() {
@ -380,6 +420,11 @@ public abstract class BlockListAsLongs implements Iterable<BlockReportReplica> {
return longs;
}
@Override
public boolean isStorageReport() {
return false;
}
@Override
public Iterator<BlockReportReplica> iterator() {
return new Iterator<BlockReportReplica>() {

49
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java
View File

@ -77,6 +77,7 @@ import org.apache.hadoop.hdfs.server.protocol.BlockReportContext;
import org.apache.hadoop.hdfs.server.protocol.BlocksWithLocations;
import org.apache.hadoop.hdfs.server.protocol.BlocksWithLocations.BlockWithLocations;
import org.apache.hadoop.hdfs.server.protocol.DatanodeCommand;
import org.apache.hadoop.hdfs.server.protocol.DatanodeRegistration;
import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage;
import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage.State;
import org.apache.hadoop.hdfs.server.protocol.KeyUpdateCommand;
@ -1800,8 +1801,8 @@ public class BlockManager {
*/
public boolean processReport(final DatanodeID nodeID,
final DatanodeStorage storage,
final BlockListAsLongs newReport, BlockReportContext context,
boolean lastStorageInRpc) throws IOException {
final BlockListAsLongs newReport,
BlockReportContext context) throws IOException {
namesystem.writeLock();
final long startTime = Time.monotonicNow(); //after acquiring write lock
final long endTime;
@ -1841,27 +1842,9 @@ public class BlockManager {
storageInfo.receivedBlockReport();
if (context != null) {
storageInfo.setLastBlockReportId(context.getReportId());
if (lastStorageInRpc) {
int rpcsSeen = node.updateBlockReportContext(context);
if (rpcsSeen >= context.getTotalRpcs()) {
List<DatanodeStorageInfo> zombies = node.removeZombieStorages();
if (zombies.isEmpty()) {
LOG.debug("processReport 0x{}: no zombie storages found.",
Long.toHexString(context.getReportId()));
} else {
for (DatanodeStorageInfo zombie : zombies) {
removeZombieReplicas(context, zombie);
}
}
node.clearBlockReportContext();
} else {
LOG.debug("processReport 0x{}: {} more RPCs remaining in this " +
"report.", Long.toHexString(context.getReportId()),
(context.getTotalRpcs() - rpcsSeen)
);
}
}
LOG.debug("Processing RPC with index {} out of total {} RPCs in "
+ "processReport 0x{}", context.getCurRpc(),
context.getTotalRpcs(), Long.toHexString(context.getReportId()));
}
} finally {
endTime = Time.monotonicNow();
@ -1887,6 +1870,26 @@ public class BlockManager {
return !node.hasStaleStorages();
}
public void removeZombieStorages(DatanodeRegistration nodeReg,
BlockReportContext context, Set<String> storageIDsInBlockReport)
throws UnregisteredNodeException {
namesystem.writeLock();
DatanodeDescriptor node = this.getDatanodeManager().getDatanode(nodeReg);
if (node != null) {
List<DatanodeStorageInfo> zombies =
node.removeZombieStorages(storageIDsInBlockReport);
if (zombies.isEmpty()) {
LOG.debug("processReport 0x{}: no zombie storages found.",
Long.toHexString(context.getReportId()));
} else {
for (DatanodeStorageInfo zombie : zombies) {
this.removeZombieReplicas(context, zombie);
}
}
}
namesystem.writeUnlock();
}
private void removeZombieReplicas(BlockReportContext context,
DatanodeStorageInfo zombie) {
LOG.warn("processReport 0x{}: removing zombie storage {}, which no " +

30
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/DatanodeDescriptor.java
View File

@ -18,7 +18,6 @@
package org.apache.hadoop.hdfs.server.blockmanagement;
import java.util.ArrayList;
import java.util.BitSet;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
@ -42,7 +41,6 @@ import org.apache.hadoop.hdfs.protocol.Block;
import org.apache.hadoop.hdfs.protocol.DatanodeID;
import org.apache.hadoop.hdfs.protocol.DatanodeInfo;
import org.apache.hadoop.hdfs.server.namenode.CachedBlock;
import org.apache.hadoop.hdfs.server.protocol.BlockReportContext;
import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage;
import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage.State;
import org.apache.hadoop.hdfs.server.protocol.StorageReport;
@ -67,24 +65,6 @@ public class DatanodeDescriptor extends DatanodeInfo {
// If node is not decommissioning, do not use this object for anything.
public final DecommissioningStatus decommissioningStatus = new DecommissioningStatus();
private long curBlockReportId = 0;
private BitSet curBlockReportRpcsSeen = null;
public int updateBlockReportContext(BlockReportContext context) {
if (curBlockReportId != context.getReportId()) {
curBlockReportId = context.getReportId();
curBlockReportRpcsSeen = new BitSet(context.getTotalRpcs());
}
curBlockReportRpcsSeen.set(context.getCurRpc());
return curBlockReportRpcsSeen.cardinality();
}
public void clearBlockReportContext() {
curBlockReportId = 0;
curBlockReportRpcsSeen = null;
}
/** Block and targets pair */
@InterfaceAudience.Private
@InterfaceStability.Evolving
@ -309,7 +289,8 @@ public class DatanodeDescriptor extends DatanodeInfo {
static final private List<DatanodeStorageInfo> EMPTY_STORAGE_INFO_LIST =
ImmutableList.of();
List<DatanodeStorageInfo> removeZombieStorages() {
List<DatanodeStorageInfo>
removeZombieStorages(Set<String> storageIDsInBlockReport) {
List<DatanodeStorageInfo> zombies = null;
synchronized (storageMap) {
Iterator<Map.Entry<String, DatanodeStorageInfo>> iter =
@ -317,18 +298,13 @@ public class DatanodeDescriptor extends DatanodeInfo {
while (iter.hasNext()) {
Map.Entry<String, DatanodeStorageInfo> entry = iter.next();
DatanodeStorageInfo storageInfo = entry.getValue();
if (storageInfo.getLastBlockReportId() != curBlockReportId) {
LOG.info(storageInfo.getStorageID() + " had lastBlockReportId 0x" +
Long.toHexString(storageInfo.getLastBlockReportId()) +
", but curBlockReportId = 0x" +
Long.toHexString(curBlockReportId));
if (!storageIDsInBlockReport.contains(storageInfo.getStorageID())) {
iter.remove();
if (zombies == null) {
zombies = new LinkedList<DatanodeStorageInfo>();
}
zombies.add(storageInfo);
}
storageInfo.setLastBlockReportId(0);
}
}
return zombies == null ? EMPTY_STORAGE_INFO_LIST : zombies;

11
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/DatanodeStorageInfo.java
View File

@ -115,9 +115,6 @@ public class DatanodeStorageInfo {
private volatile BlockInfoContiguous blockList = null;
private int numBlocks = 0;
// The ID of the last full block report which updated this storage.
private long lastBlockReportId = 0;
/** The number of block reports received */
private int blockReportCount = 0;
@ -182,14 +179,6 @@ public class DatanodeStorageInfo {
this.blockPoolUsed = blockPoolUsed;
}
long getLastBlockReportId() {
return lastBlockReportId;
}
void setLastBlockReportId(long lastBlockReportId) {
this.lastBlockReportId = lastBlockReportId;
}
State getState() {
return this.state;
}

31
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/datanode/BPServiceActor.java
View File

@ -471,10 +471,35 @@ class BPServiceActor implements Runnable {
} else {
// Send one block report per message.
for (int r = 0; r < reports.length; r++) {
StorageBlockReport singleReport[] = { reports[r] };
DatanodeCommand cmd = bpNamenode.blockReport(
bpRegistration, bpos.getBlockPoolId(), singleReport,
StorageBlockReport[] singleReport = {reports[r]};
DatanodeCommand cmd;
if (r != reports.length - 1) {
cmd = bpNamenode.blockReport(bpRegistration, bpos.getBlockPoolId(),
singleReport, new BlockReportContext(reports.length, r,
reportId));
} else {
StorageBlockReport[] lastSplitReport =
new StorageBlockReport[perVolumeBlockLists.size()];
// When block reports are split, the last RPC in the block report
// has the information about all storages in the block report.
// See HDFS-10301 for more details. To achieve this, the last RPC
// has 'n' storage reports, where 'n' is the number of storages in
// a DN. The actual block replicas are reported only for the
// last/n-th storage.
i = 0;
for(Map.Entry<DatanodeStorage, BlockListAsLongs> kvPair :
perVolumeBlockLists.entrySet()) {
lastSplitReport[i++] = new StorageBlockReport(
kvPair.getKey(), BlockListAsLongs.STORAGE_REPORT);
if (i == r) {
lastSplitReport[i] = reports[r];
break;
}
}
cmd = bpNamenode.blockReport(
bpRegistration, bpos.getBlockPoolId(), lastSplitReport,
new BlockReportContext(reports.length, r, reportId));
}
numReportsSent++;
numRPCs++;
if (cmd != null) {

16
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/NameNodeRpcServer.java
View File

@ -1314,6 +1314,7 @@ class NameNodeRpcServer implements NamenodeProtocols {
boolean noStaleStorages = false;
for (int r = 0; r < reports.length; r++) {
final BlockListAsLongs blocks = reports[r].getBlocks();
if (!blocks.isStorageReport()) {
//
// BlockManager.processReport accumulates information of prior calls
// for the same node and storage, so the value returned by the last
@ -1322,13 +1323,24 @@ class NameNodeRpcServer implements NamenodeProtocols {
final int index = r;
noStaleStorages = bm.runBlockOp(new Callable<Boolean>() {
@Override
public Boolean call() throws IOException {
public Boolean call()
throws IOException {
return bm.processReport(nodeReg, reports[index].getStorage(),
blocks, context, (index == reports.length - 1));
blocks, context);
}
});
metrics.incrStorageBlockReportOps();
}
}
if (nn.getFSImage().isUpgradeFinalized() &&
context.getTotalRpcs() == context.getCurRpc() + 1) {
Set<String> storageIDsInBlockReport = new HashSet<>();
for (StorageBlockReport report : reports) {
storageIDsInBlockReport.add(report.getStorage().getStorageID());
}
bm.removeZombieStorages(nodeReg, context, storageIDsInBlockReport);
}
if (nn.getFSImage().isUpgradeFinalized() &&
!namesystem.isRollingUpgrade() &&

10
hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestBlockManager.java
View File

@ -678,12 +678,12 @@ public class TestBlockManager {
reset(node);
bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
BlockListAsLongs.EMPTY, null, false);
BlockListAsLongs.EMPTY, null);
assertEquals(1, ds.getBlockReportCount());
// send block report again, should NOT be processed
reset(node);
bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
BlockListAsLongs.EMPTY, null, false);
BlockListAsLongs.EMPTY, null);
assertEquals(1, ds.getBlockReportCount());
// re-register as if node restarted, should update existing node
@ -694,7 +694,7 @@ public class TestBlockManager {
// send block report, should be processed after restart
reset(node);
bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
BlockListAsLongs.EMPTY, null, false);
BlockListAsLongs.EMPTY, null);
// Reinitialize as registration with empty storage list pruned
// node.storageMap.
ds = node.getStorageInfos()[0];
@ -723,7 +723,7 @@ public class TestBlockManager {
reset(node);
doReturn(1).when(node).numBlocks();
bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
BlockListAsLongs.EMPTY, null, false);
BlockListAsLongs.EMPTY, null);
assertEquals(1, ds.getBlockReportCount());
}
@ -796,7 +796,7 @@ public class TestBlockManager {
// Make sure it's the first full report
assertEquals(0, ds.getBlockReportCount());
bm.processReport(node, new DatanodeStorage(ds.getStorageID()),
builder.build(), null, false);
builder.build(), null);
assertEquals(1, ds.getBlockReportCount());
// verify the storage info is correct

74
hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestNameNodePrunesMissingStorages.java
View File

@ -19,6 +19,12 @@
package org.apache.hadoop.hdfs.server.blockmanagement;
import com.google.common.base.Supplier;
import java.util.ArrayList;
import java.util.Map;
import java.util.concurrent.Callable;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configuration;
@ -29,12 +35,17 @@ import org.apache.hadoop.hdfs.DFSTestUtil;
import org.apache.hadoop.hdfs.DistributedFileSystem;
import org.apache.hadoop.hdfs.HdfsConfiguration;
import org.apache.hadoop.hdfs.MiniDFSCluster;
import org.apache.hadoop.hdfs.protocol.BlockListAsLongs;
import org.apache.hadoop.hdfs.protocol.DatanodeID;
import org.apache.hadoop.hdfs.protocol.ExtendedBlock;
import org.apache.hadoop.hdfs.server.datanode.DataNode;
import org.apache.hadoop.hdfs.server.datanode.DataNodeTestUtils;
import org.apache.hadoop.hdfs.server.datanode.fsdataset.FsVolumeSpi;
import org.apache.hadoop.hdfs.server.protocol.BlockReportContext;
import org.apache.hadoop.hdfs.server.protocol.DatanodeCommand;
import org.apache.hadoop.hdfs.server.protocol.DatanodeRegistration;
import org.apache.hadoop.hdfs.server.protocol.DatanodeStorage;
import org.apache.hadoop.hdfs.server.protocol.StorageBlockReport;
import org.apache.hadoop.hdfs.server.protocol.StorageReport;
import org.apache.hadoop.test.GenericTestUtils;
import org.junit.Assert;
@ -249,4 +260,67 @@ public class TestNameNodePrunesMissingStorages {
}
}
}
@Test(timeout=300000)
public void testInterleavedFullBlockReports() throws Exception {
Configuration conf = new HdfsConfiguration();
conf.setLong(DFSConfigKeys.DFS_BLOCKREPORT_INTERVAL_MSEC_KEY,
36000000L);
int numStoragesPerDatanode = 6;
final MiniDFSCluster cluster = new MiniDFSCluster
.Builder(conf).numDataNodes(1)
.storagesPerDatanode(numStoragesPerDatanode)
.build();
try {
LOG.info("waiting for cluster to become active...");
cluster.waitActive();
// Get the datanode registration and the block reports
DataNode dn = cluster.getDataNodes().get(0);
final String blockPoolId = cluster.getNamesystem().getBlockPoolId();
LOG.info("Block pool id: " + blockPoolId);
final DatanodeRegistration dnR = dn.getDNRegistrationForBP(blockPoolId);
Map<DatanodeStorage, BlockListAsLongs> perVolumeBlockLists =
dn.getFSDataset().getBlockReports(blockPoolId);
final StorageBlockReport[] reports =
new StorageBlockReport[perVolumeBlockLists.size()];
int reportIndex = 0;
for(Map.Entry<DatanodeStorage, BlockListAsLongs> kvPair :
perVolumeBlockLists.entrySet()) {
DatanodeStorage dnStorage = kvPair.getKey();
BlockListAsLongs blockList = kvPair.getValue();
reports[reportIndex++] =
new StorageBlockReport(dnStorage, blockList);
}
// Get the list of storage ids associated with the datanode
// before the test
BlockManager bm =
cluster.getNameNode().getNamesystem().getBlockManager();
final DatanodeDescriptor dnDescriptor = bm.getDatanodeManager().
getDatanode(cluster.getDataNodes().get(0).getDatanodeUuid());
DatanodeStorageInfo[] storageInfos = dnDescriptor.getStorageInfos();
// Send the full block report concurrently using
// numThreads=numStoragesPerDatanode
ExecutorService executorService = Executors.
newFixedThreadPool(numStoragesPerDatanode);
List<Future<DatanodeCommand>> futureList =
new ArrayList<>(numStoragesPerDatanode);
for (int i = 0; i < numStoragesPerDatanode; i++) {
futureList.add(executorService.submit(new Callable<DatanodeCommand>() {
@Override
public DatanodeCommand call() throws IOException {
return cluster.getNameNodeRpc().blockReport(dnR, blockPoolId,
reports, new BlockReportContext(1, 0, System.nanoTime()));
}
}));
}
for (Future<DatanodeCommand> future: futureList) {
future.get();
}
executorService.shutdown();
// Verify that the storages match before and after the test
Assert.assertArrayEquals(storageInfos, dnDescriptor.getStorageInfos());
} finally {
cluster.shutdown();
}
}
}

33
hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/datanode/TestDnRespectsBlockReportSplitThreshold.java
View File

@ -41,6 +41,7 @@ import org.mockito.ArgumentCaptor;
import org.mockito.Mockito;
import static org.hamcrest.core.Is.is;
import static org.junit.Assert.assertEquals;
import static org.junit.Assert.assertThat;
import static org.mockito.Matchers.*;
import static org.mockito.Mockito.times;
@ -88,6 +89,34 @@ public class TestDnRespectsBlockReportSplitThreshold {
blockCount * BLOCK_SIZE, BLOCK_SIZE, REPL_FACTOR, seed);
}
private void verifyCapturedArgumentsSplit(
ArgumentCaptor<StorageBlockReport[]> captor,
int expectedReportsPerCall,
int expectedTotalBlockCount) {
List<StorageBlockReport[]> listOfReports = captor.getAllValues();
int numBlocksReported = 0;
int storageIndex = 0;
int listOfReportsSize = listOfReports.size();
for (StorageBlockReport[] reports : listOfReports) {
if (storageIndex < (listOfReportsSize - 1)) {
assertThat(reports.length, is(expectedReportsPerCall));
} else {
assertThat(reports.length, is(listOfReportsSize));
}
for (StorageBlockReport report : reports) {
BlockListAsLongs blockList = report.getBlocks();
if (!blockList.isStorageReport()) {
numBlocksReported += blockList.getNumberOfBlocks();
} else {
assertEquals(blockList.getNumberOfBlocks(), -1);
}
}
storageIndex++;
}
assert(numBlocksReported >= expectedTotalBlockCount);
}
private void verifyCapturedArguments(
ArgumentCaptor<StorageBlockReport[]> captor,
int expectedReportsPerCall,
@ -136,7 +165,7 @@ public class TestDnRespectsBlockReportSplitThreshold {
anyString(),
captor.capture(), Mockito.<BlockReportContext>anyObject());
verifyCapturedArguments(captor, 1, BLOCKS_IN_FILE);
verifyCapturedArgumentsSplit(captor, 1, BLOCKS_IN_FILE);
}
/**
@ -200,7 +229,7 @@ public class TestDnRespectsBlockReportSplitThreshold {
anyString(),
captor.capture(), Mockito.<BlockReportContext>anyObject());
verifyCapturedArguments(captor, 1, BLOCKS_IN_FILE);
verifyCapturedArgumentsSplit(captor, 1, BLOCKS_IN_FILE);
}
}

30
hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/datanode/TestNNHandlesBlockReportPerStorage.java
View File

@ -20,6 +20,7 @@ package org.apache.hadoop.hdfs.server.datanode;
import java.io.IOException;
import org.apache.hadoop.hdfs.protocol.BlockListAsLongs;
import org.apache.hadoop.hdfs.server.protocol.BlockReportContext;
import org.apache.hadoop.hdfs.server.protocol.DatanodeRegistration;
import org.apache.hadoop.hdfs.server.protocol.StorageBlockReport;
@ -35,13 +36,30 @@ public class TestNNHandlesBlockReportPerStorage extends BlockReportTestBase {
@Override
protected void sendBlockReports(DatanodeRegistration dnR, String poolId,
StorageBlockReport[] reports) throws IOException {
int i = 0;
for (StorageBlockReport report : reports) {
LOG.info("Sending block report for storage " + report.getStorage().getStorageID());
StorageBlockReport[] singletonReport = { report };
for (int r = 0; r < reports.length; r++) {
LOG.info("Sending block report for storage " +
reports[r].getStorage().getStorageID());
StorageBlockReport[] singletonReport = {reports[r]};
if (r != reports.length - 1) {
cluster.getNameNodeRpc().blockReport(dnR, poolId, singletonReport,
new BlockReportContext(reports.length, i, System.nanoTime()));
i++;
new BlockReportContext(reports.length, r, System.nanoTime()));
} else {
StorageBlockReport[] lastSplitReport =
new StorageBlockReport[reports.length];
// When block reports are split, send a dummy storage report for all
// other storages in the blockreport along with the last storage report
for (int i = 0; i <= r; i++) {
if (i == r) {
lastSplitReport[i] = reports[r];
break;
}
lastSplitReport[i] =
new StorageBlockReport(reports[i].getStorage(),
BlockListAsLongs.STORAGE_REPORT);
}
cluster.getNameNodeRpc().blockReport(dnR, poolId, lastSplitReport,
new BlockReportContext(reports.length, r, System.nanoTime()));
}
}
}
}