public virtual void TestDnFencing()
{
// Create a file with replication level 3.
DFSTestUtil.CreateFile(fs, TestFilePath, 30 * SmallBlock, (short)3, 1L);
ExtendedBlock block = DFSTestUtil.GetFirstBlock(fs, TestFilePath);
// Drop its replication count to 1, so it becomes over-replicated.
// Then compute the invalidation of the extra blocks and trigger
// heartbeats so the invalidations are flushed to the DNs.
nn1.GetRpcServer().SetReplication(TestFile, (short)1);
BlockManagerTestUtil.ComputeInvalidationWork(nn1.GetNamesystem().GetBlockManager(
));
cluster.TriggerHeartbeats();
// 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);
cluster.TransitionToActive(1);
// Check that the standby picked up the replication change.
NUnit.Framework.Assert.AreEqual(1, nn2.GetRpcServer().GetFileInfo(TestFile).GetReplication
());
// Dump some info for debugging purposes.
Banner("NN2 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);
// Force a rescan of postponedMisreplicatedBlocks.
BlockManager nn2BM = nn2.GetNamesystem().GetBlockManager();
BlockManagerTestUtil.CheckHeartbeat(nn2BM);
BlockManagerTestUtil.RescanPostponedMisreplicatedBlocks(nn2BM);
// The blocks should no longer be postponed.
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetPostponedMisreplicatedBlocks
());
// Wait for NN2 to enact its deletions (replication monitor has to run, etc)
BlockManagerTestUtil.ComputeInvalidationWork(nn2.GetNamesystem().GetBlockManager(
));
cluster.TriggerHeartbeats();
HATestUtil.WaitForDNDeletions(cluster);
cluster.TriggerDeletionReports();
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetUnderReplicatedBlocks()
);
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetPendingReplicationBlocks
());
Banner("Making sure the file is still readable");
FileSystem fs2 = cluster.GetFileSystem(1);
DFSTestUtil.ReadFile(fs2, TestFilePath);
Banner("Waiting for the actual block files to get deleted from DNs.");
WaitForTrueReplication(cluster, block, 1);
}
public virtual void TestComplexFailoverIntoSafemode()
{
Banner("Starting with NN0 active and NN1 standby, creating some blocks");
DFSTestUtil.CreateFile(fs, new Path("/test"), 3 * BlockSize, (short)3, 1L);
// Roll edit log so that, when the SBN restarts, it will load
// the namespace during startup and enter safemode.
nn0.GetRpcServer().RollEditLog();
Banner("Creating some blocks that won't be in the edit log");
DFSTestUtil.CreateFile(fs, new Path("/test2"), 5 * BlockSize, (short)3, 1L);
Banner("Deleting the original blocks");
fs.Delete(new Path("/test"), true);
Banner("Restarting standby");
RestartStandby();
// We expect it to be on its way out of safemode, since all of the blocks
// from the edit log have been reported.
AssertSafeMode(nn1, 3, 3, 3, 0);
// Initiate a failover into it while it's in safemode
Banner("Initiating a failover into NN1 in safemode");
NameNodeAdapter.AbortEditLogs(nn0);
cluster.TransitionToActive(1);
AssertSafeMode(nn1, 5, 5, 3, 0);
}
public virtual void TestNNClearsCommandsOnFailoverWithReplChanges()
{
// Make lots of blocks to increase chances of triggering a bug.
DFSTestUtil.CreateFile(fs, TestFilePath, 30 * SmallBlock, (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(TestFile, (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();
Sharpen.Thread.Sleep(1000);
}
Banner("triggering BRs");
cluster.TriggerBlockReports();
nn1.GetRpcServer().SetReplication(TestFile, (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.
NUnit.Framework.Assert.AreEqual(1, nn2.GetRpcServer().GetFileInfo(TestFile).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);
// Force a rescan of postponedMisreplicatedBlocks.
BlockManager nn2BM = nn2.GetNamesystem().GetBlockManager();
BlockManagerTestUtil.CheckHeartbeat(nn2BM);
BlockManagerTestUtil.RescanPostponedMisreplicatedBlocks(nn2BM);
// The block should no longer be postponed.
NUnit.Framework.Assert.AreEqual(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();
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetUnderReplicatedBlocks()
);
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetPendingReplicationBlocks
());
Banner("Making sure the file is still readable");
FileSystem fs2 = cluster.GetFileSystem(1);
DFSTestUtil.ReadFile(fs2, TestFilePath);
}
public virtual void TestNNClearsCommandsOnFailoverAfterStartup()
{
// Make lots of blocks to increase chances of triggering a bug.
DFSTestUtil.CreateFile(fs, TestFilePath, 30 * SmallBlock, (short)3, 1L);
Banner("Shutting down NN2");
cluster.ShutdownNameNode(1);
Banner("Setting replication to 1, rolling edit log.");
nn1.GetRpcServer().SetReplication(TestFile, (short)1);
nn1.GetRpcServer().RollEditLog();
// Start NN2 again. When it starts up, it will see all of the
// blocks as over-replicated, since it has the metadata for
// replication=1, but the DNs haven't yet processed the deletions.
Banner("Starting NN2 again.");
cluster.RestartNameNode(1);
nn2 = cluster.GetNameNode(1);
Banner("triggering BRs");
cluster.TriggerBlockReports();
// We expect that both NN1 and NN2 will have some number of
// deletions queued up for the DNs.
Banner("computing invalidation on nn1");
BlockManagerTestUtil.ComputeInvalidationWork(nn1.GetNamesystem().GetBlockManager(
));
Banner("computing invalidation on nn2");
BlockManagerTestUtil.ComputeInvalidationWork(nn2.GetNamesystem().GetBlockManager(
));
// 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);
cluster.TransitionToActive(1);
// Check that the standby picked up the replication change.
NUnit.Framework.Assert.AreEqual(1, nn2.GetRpcServer().GetFileInfo(TestFile).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);
// Force a rescan of postponedMisreplicatedBlocks.
BlockManager nn2BM = nn2.GetNamesystem().GetBlockManager();
BlockManagerTestUtil.CheckHeartbeat(nn2BM);
BlockManagerTestUtil.RescanPostponedMisreplicatedBlocks(nn2BM);
// The block should no longer be postponed.
NUnit.Framework.Assert.AreEqual(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();
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetUnderReplicatedBlocks()
);
NUnit.Framework.Assert.AreEqual(0, nn2.GetNamesystem().GetPendingReplicationBlocks
());
Banner("Making sure the file is still readable");
FileSystem fs2 = cluster.GetFileSystem(1);
DFSTestUtil.ReadFile(fs2, TestFilePath);
}
