/**
* Regression test for HDFS-2742. The issue in this bug was: - DN does a block report while file
* is open. This BR contains the block in RBW state. - Standby queues the RBW state in
* PendingDatanodeMessages - Standby processes edit logs during failover. Before fixing this bug,
* it was mistakenly applying the RBW reported state after the block had been completed, causing
* the block to get marked corrupt. Instead, we should now be applying the RBW message on OP_ADD,
* and then the FINALIZED message on OP_CLOSE.
*/
@Test
public void testBlockReportsWhileFileBeingWritten() throws Exception {
FSDataOutputStream out = fs.create(TEST_FILE_PATH);
try {
AppendTestUtil.write(out, 0, 10);
out.hflush();
// Block report will include the RBW replica, but will be
// queued on the StandbyNode.
cluster.triggerBlockReports();
} finally {
IOUtils.closeStream(out);
}
cluster.transitionToStandby(0);
cluster.transitionToActive(1);
// Verify that no replicas are marked corrupt, and that the
// file is readable from the failed-over standby.
BlockManagerTestUtil.updateState(nn1.getNamesystem().getBlockManager());
BlockManagerTestUtil.updateState(nn2.getNamesystem().getBlockManager());
assertEquals(0, nn1.getNamesystem().getCorruptReplicaBlocks());
assertEquals(0, nn2.getNamesystem().getCorruptReplicaBlocks());
DFSTestUtil.readFile(fs, TEST_FILE_PATH);
}
/**
* Another regression test for HDFS-2742. This tests the following sequence: - DN does a block
* report while file is open. This BR contains the block in RBW state. - The block report is
* delayed in reaching the standby. - The file is closed. - The standby processes the OP_ADD and
* OP_CLOSE operations before the RBW block report arrives. - The standby should not mark the
* block as corrupt.
*/
@Test
public void testRBWReportArrivesAfterEdits() throws Exception {
final CountDownLatch brFinished = new CountDownLatch(1);
DelayAnswer delayer =
new GenericTestUtils.DelayAnswer(LOG) {
@Override
protected Object passThrough(InvocationOnMock invocation) throws Throwable {
try {
return super.passThrough(invocation);
} finally {
// inform the test that our block report went through.
brFinished.countDown();
}
}
};
FSDataOutputStream out = fs.create(TEST_FILE_PATH);
try {
AppendTestUtil.write(out, 0, 10);
out.hflush();
DataNode dn = cluster.getDataNodes().get(0);
DatanodeProtocolClientSideTranslatorPB spy = DataNodeTestUtils.spyOnBposToNN(dn, nn2);
Mockito.doAnswer(delayer)
.when(spy)
.blockReport(
Mockito.<DatanodeRegistration>anyObject(),
Mockito.anyString(),
Mockito.<StorageBlockReport[]>anyObject());
dn.scheduleAllBlockReport(0);
delayer.waitForCall();
} finally {
IOUtils.closeStream(out);
}
cluster.transitionToStandby(0);
cluster.transitionToActive(1);
delayer.proceed();
brFinished.await();
// Verify that no replicas are marked corrupt, and that the
// file is readable from the failed-over standby.
BlockManagerTestUtil.updateState(nn1.getNamesystem().getBlockManager());
BlockManagerTestUtil.updateState(nn2.getNamesystem().getBlockManager());
assertEquals(0, nn1.getNamesystem().getCorruptReplicaBlocks());
assertEquals(0, nn2.getNamesystem().getCorruptReplicaBlocks());
DFSTestUtil.readFile(fs, TEST_FILE_PATH);
}
/**
* Test that, when a block is re-opened for append, the related datanode messages are correctly
* queued by the SBN because they have future states and genstamps.
*/
@Test
public void testQueueingWithAppend() throws Exception {
int numQueued = 0;
int numDN = cluster.getDataNodes().size();
FSDataOutputStream out = fs.create(TEST_FILE_PATH);
try {
AppendTestUtil.write(out, 0, 10);
out.hflush();
// Opening the file will report RBW replicas, but will be
// queued on the StandbyNode.
numQueued += numDN; // RBW messages
} finally {
IOUtils.closeStream(out);
numQueued += numDN; // blockReceived messages
}
cluster.triggerBlockReports();
numQueued += numDN;
try {
out = fs.append(TEST_FILE_PATH);
AppendTestUtil.write(out, 10, 10);
// RBW replicas once it's opened for append
numQueued += numDN;
} finally {
IOUtils.closeStream(out);
numQueued += numDN; // blockReceived
}
cluster.triggerBlockReports();
numQueued += numDN;
assertEquals(
numQueued, cluster.getNameNode(1).getNamesystem().getPendingDataNodeMessageCount());
cluster.transitionToStandby(0);
cluster.transitionToActive(1);
// Verify that no replicas are marked corrupt, and that the
// file is readable from the failed-over standby.
BlockManagerTestUtil.updateState(nn1.getNamesystem().getBlockManager());
BlockManagerTestUtil.updateState(nn2.getNamesystem().getBlockManager());
assertEquals(0, nn1.getNamesystem().getCorruptReplicaBlocks());
assertEquals(0, nn2.getNamesystem().getCorruptReplicaBlocks());
AppendTestUtil.check(fs, TEST_FILE_PATH, 20);
}
/**
* Verify that the NameNode is able to still use <tt>READ_ONLY_SHARED</tt> replicas even when the
* single NORMAL replica is offline (and the effective replication count is 0).
*/
@Test
public void testNormalReplicaOffline() throws Exception {
// Stop the datanode hosting the NORMAL replica
cluster.stopDataNode(normalDataNode.getXferAddr());
// Force NameNode to detect that the datanode is down
BlockManagerTestUtil.noticeDeadDatanode(cluster.getNameNode(), normalDataNode.getXferAddr());
// The live replica count should now be zero (since the NORMAL replica is offline)
NumberReplicas numberReplicas = blockManager.countNodes(block);
assertThat(numberReplicas.liveReplicas(), is(0));
// The block should be reported as under-replicated
BlockManagerTestUtil.updateState(blockManager);
assertThat(blockManager.getUnderReplicatedBlocksCount(), is(1L));
// The BlockManager should be able to heal the replication count back to 1
// by triggering an inter-datanode replication from one of the READ_ONLY_SHARED replicas
BlockManagerTestUtil.computeAllPendingWork(blockManager);
DFSTestUtil.waitForReplication(cluster, extendedBlock, 1, 1, 0);
// There should now be 2 *locations* for the block, and 1 *replica*
assertThat(getLocatedBlock().getLocations().length, is(2));
validateNumberReplicas(1);
}
private void validateNumberReplicas(int expectedReplicas) throws IOException {
NumberReplicas numberReplicas = blockManager.countNodes(block);
assertThat(numberReplicas.liveReplicas(), is(expectedReplicas));
assertThat(numberReplicas.excessReplicas(), is(0));
assertThat(numberReplicas.corruptReplicas(), is(0));
assertThat(numberReplicas.decommissionedReplicas(), is(0));
assertThat(numberReplicas.replicasOnStaleNodes(), is(0));
BlockManagerTestUtil.updateState(blockManager);
assertThat(blockManager.getUnderReplicatedBlocksCount(), is(0L));
assertThat(blockManager.getExcessBlocksCount(), is(0L));
}
/**
* Test case that reduces replication of a file with a lot of blocks and then fails over right
* after those blocks enter the DN invalidation queues on the active. Ensures that fencing is
* correct and no replicas are lost.
*/
@Test
public void testNNClearsCommandsOnFailoverWithReplChanges() throws Exception {
// Make lots of blocks to increase chances of triggering a bug.
DFSTestUtil.createFile(fs, TEST_FILE_PATH, 30 * SMALL_BLOCK, (short) 1, 1L);
banner("rolling NN1's edit log, forcing catch-up");
HATestUtil.waitForStandbyToCatchUp(nn1, nn2);
// Get some new replicas reported so that NN2 now considers
// them over-replicated and schedules some more deletions
nn1.getRpcServer().setReplication(TEST_FILE, (short) 2);
while (BlockManagerTestUtil.getComputedDatanodeWork(nn1.getNamesystem().getBlockManager())
> 0) {
LOG.info("Getting more replication work computed");
}
BlockManager bm1 = nn1.getNamesystem().getBlockManager();
while (bm1.getPendingReplicationBlocksCount() > 0) {
BlockManagerTestUtil.updateState(bm1);
cluster.triggerHeartbeats();
Thread.sleep(1000);
}
banner("triggering BRs");
cluster.triggerBlockReports();
nn1.getRpcServer().setReplication(TEST_FILE, (short) 1);
banner("computing invalidation on nn1");
BlockManagerTestUtil.computeInvalidationWork(nn1.getNamesystem().getBlockManager());
doMetasave(nn1);
banner("computing invalidation on nn2");
BlockManagerTestUtil.computeInvalidationWork(nn2.getNamesystem().getBlockManager());
doMetasave(nn2);
// Dump some info for debugging purposes.
banner("Metadata immediately before failover");
doMetasave(nn2);
// Transition nn2 to active even though nn1 still thinks it's active
banner("Failing to NN2 but let NN1 continue to think it's active");
NameNodeAdapter.abortEditLogs(nn1);
NameNodeAdapter.enterSafeMode(nn1, false);
BlockManagerTestUtil.computeInvalidationWork(nn2.getNamesystem().getBlockManager());
cluster.transitionToActive(1);
// Check that the standby picked up the replication change.
assertEquals(1, nn2.getRpcServer().getFileInfo(TEST_FILE).getReplication());
// Dump some info for debugging purposes.
banner("Metadata immediately after failover");
doMetasave(nn2);
banner("Triggering heartbeats and block reports so that fencing is completed");
cluster.triggerHeartbeats();
cluster.triggerBlockReports();
banner("Metadata after nodes have all block-reported");
doMetasave(nn2);
// The block should no longer be postponed.
assertEquals(0, nn2.getNamesystem().getPostponedMisreplicatedBlocks());
// Wait for NN2 to enact its deletions (replication monitor has to run, etc)
BlockManagerTestUtil.computeInvalidationWork(nn2.getNamesystem().getBlockManager());
HATestUtil.waitForNNToIssueDeletions(nn2);
cluster.triggerHeartbeats();
HATestUtil.waitForDNDeletions(cluster);
cluster.triggerDeletionReports();
assertEquals(0, nn2.getNamesystem().getUnderReplicatedBlocks());
assertEquals(0, nn2.getNamesystem().getPendingReplicationBlocks());
banner("Making sure the file is still readable");
FileSystem fs2 = cluster.getFileSystem(1);
DFSTestUtil.readFile(fs2, TEST_FILE_PATH);
}
