// Set up fault injection mock.
/// <summary>
/// Run through the creation of a log without any faults injected,
/// and count how many RPCs are made to each node.
/// </summary>
/// <remarks>
/// Run through the creation of a log without any faults injected,
/// and count how many RPCs are made to each node. This sets the
/// bounds for the other test cases, so they can exhaustively explore
/// the space of potential failures.
/// </remarks>
/// <exception cref="System.Exception"/>
private static long DetermineMaxIpcNumber()
{
Configuration conf = new Configuration();
MiniJournalCluster cluster = new MiniJournalCluster.Builder(conf).Build();
QuorumJournalManager qjm = null;
long ret;
try
{
qjm = CreateInjectableQJM(cluster);
qjm.Format(QJMTestUtil.FakeNsinfo);
DoWorkload(cluster, qjm);
ICollection <int> ipcCounts = Sets.NewTreeSet();
foreach (AsyncLogger l in qjm.GetLoggerSetForTests().GetLoggersForTests())
{
TestQJMWithFaults.InvocationCountingChannel ch = (TestQJMWithFaults.InvocationCountingChannel
)l;
ch.WaitForAllPendingCalls();
ipcCounts.AddItem(ch.GetRpcCount());
}
// All of the loggers should have sent the same number of RPCs, since there
// were no failures.
NUnit.Framework.Assert.AreEqual(1, ipcCounts.Count);
ret = ipcCounts.First();
Log.Info("Max IPC count = " + ret);
}
finally
{
IOUtils.CloseStream(qjm);
cluster.Shutdown();
}
return(ret);
}
public virtual void TestRollbackWithQJM()
{
Configuration conf = new HdfsConfiguration();
MiniJournalCluster mjc = null;
MiniDFSCluster cluster = null;
Path foo = new Path("/foo");
Path bar = new Path("/bar");
try
{
mjc = new MiniJournalCluster.Builder(conf).NumJournalNodes(NumJournalNodes).Build
();
conf.Set(DFSConfigKeys.DfsNamenodeEditsDirKey, mjc.GetQuorumJournalURI(JournalId)
.ToString());
cluster = new MiniDFSCluster.Builder(conf).NumDataNodes(0).Build();
cluster.WaitActive();
DistributedFileSystem dfs = cluster.GetFileSystem();
DFSAdmin dfsadmin = new DFSAdmin(conf);
dfs.Mkdirs(foo);
// start rolling upgrade
dfs.SetSafeMode(HdfsConstants.SafeModeAction.SafemodeEnter);
NUnit.Framework.Assert.AreEqual(0, dfsadmin.Run(new string[] { "-rollingUpgrade",
"prepare" }));
dfs.SetSafeMode(HdfsConstants.SafeModeAction.SafemodeLeave);
// create new directory
dfs.Mkdirs(bar);
dfs.Close();
// rollback
cluster.RestartNameNode("-rollingUpgrade", "rollback");
// make sure /foo is still there, but /bar is not
dfs = cluster.GetFileSystem();
NUnit.Framework.Assert.IsTrue(dfs.Exists(foo));
NUnit.Framework.Assert.IsFalse(dfs.Exists(bar));
// check storage in JNs
for (int i = 0; i < NumJournalNodes; i++)
{
FilePath dir = mjc.GetCurrentDir(0, JournalId);
// segments:(startSegment, mkdir, endSegment), (startSegment, upgrade
// marker, mkdir, endSegment)
CheckJNStorage(dir, 4, 7);
}
}
finally
{
if (cluster != null)
{
cluster.Shutdown();
}
if (mjc != null)
{
mjc.Shutdown();
}
}
}
public virtual void TestRecoverAfterDoubleFailures()
{
long MaxIpcNumber = DetermineMaxIpcNumber();
for (int failA = 1; failA <= MaxIpcNumber; failA++)
{
for (int failB = 1; failB <= MaxIpcNumber; failB++)
{
string injectionStr = "(" + failA + ", " + failB + ")";
Log.Info("\n\n-------------------------------------------\n" + "Beginning test, failing at "
+ injectionStr + "\n" + "-------------------------------------------\n\n");
MiniJournalCluster cluster = new MiniJournalCluster.Builder(conf).Build();
QuorumJournalManager qjm = null;
try
{
qjm = CreateInjectableQJM(cluster);
qjm.Format(QJMTestUtil.FakeNsinfo);
IList <AsyncLogger> loggers = qjm.GetLoggerSetForTests().GetLoggersForTests();
FailIpcNumber(loggers[0], failA);
FailIpcNumber(loggers[1], failB);
int lastAckedTxn = DoWorkload(cluster, qjm);
if (lastAckedTxn < 6)
{
Log.Info("Failed after injecting failures at " + injectionStr + ". This is expected since we injected a failure in the "
+ "majority.");
}
qjm.Close();
qjm = null;
// Now should be able to recover
qjm = CreateInjectableQJM(cluster);
long lastRecoveredTxn = QJMTestUtil.RecoverAndReturnLastTxn(qjm);
NUnit.Framework.Assert.IsTrue(lastRecoveredTxn >= lastAckedTxn);
QJMTestUtil.WriteSegment(cluster, qjm, lastRecoveredTxn + 1, 3, true);
}
catch (Exception t)
{
// Test failure! Rethrow with the test setup info so it can be
// easily triaged.
throw new RuntimeException("Test failed with injection: " + injectionStr, t);
}
finally
{
cluster.Shutdown();
cluster = null;
IOUtils.CloseStream(qjm);
qjm = null;
}
}
}
}
public virtual void TestSingleThreaded()
{
Configuration conf = new Configuration();
MiniJournalCluster cluster = new MiniJournalCluster.Builder(conf).Build();
URI uri = cluster.GetQuorumJournalURI(Jid);
QuorumJournalManager qjm = new QuorumJournalManager(conf, uri, FakeNsinfo);
try
{
qjm.Format(FakeNsinfo);
}
finally
{
qjm.Close();
}
try
{
// With no failures or contention, epochs should increase one-by-one
for (int i = 0; i < 5; i++)
{
qjm = new QuorumJournalManager(conf, uri, FakeNsinfo);
try
{
qjm.CreateNewUniqueEpoch();
NUnit.Framework.Assert.AreEqual(i + 1, qjm.GetLoggerSetForTests().GetEpoch());
}
finally
{
qjm.Close();
}
}
long prevEpoch = 5;
// With some failures injected, it should still always increase, perhaps
// skipping some
for (int i_1 = 0; i_1 < 20; i_1++)
{
long newEpoch = -1;
while (true)
{
qjm = new QuorumJournalManager(conf, uri, FakeNsinfo, new TestEpochsAreUnique.FaultyLoggerFactory
(this));
try
{
qjm.CreateNewUniqueEpoch();
newEpoch = qjm.GetLoggerSetForTests().GetEpoch();
break;
}
catch (IOException)
{
}
finally
{
// It's OK to fail to create an epoch, since we randomly inject
// faults. It's possible we'll inject faults in too many of the
// underlying nodes, and a failure is expected in that case
qjm.Close();
}
}
Log.Info("Created epoch " + newEpoch);
NUnit.Framework.Assert.IsTrue("New epoch " + newEpoch + " should be greater than previous "
+ prevEpoch, newEpoch > prevEpoch);
prevEpoch = newEpoch;
}
}
finally
{
cluster.Shutdown();
}
}
public virtual void TestRandomized()
{
long seed;
long userSpecifiedSeed = long.GetLong(RandSeedProperty);
if (userSpecifiedSeed != null)
{
Log.Info("Using seed specified in system property");
seed = userSpecifiedSeed;
// If the user specifies a seed, then we should gather all the
// IPC trace information so that debugging is easier. This makes
// the test run about 25% slower otherwise.
((Log4JLogger)ProtobufRpcEngine.Log).GetLogger().SetLevel(Level.All);
}
else
{
seed = new Random().NextLong();
}
Log.Info("Random seed: " + seed);
Random r = new Random(seed);
MiniJournalCluster cluster = new MiniJournalCluster.Builder(conf).Build();
// Format the cluster using a non-faulty QJM.
QuorumJournalManager qjmForInitialFormat = CreateInjectableQJM(cluster);
qjmForInitialFormat.Format(QJMTestUtil.FakeNsinfo);
qjmForInitialFormat.Close();
try
{
long txid = 0;
long lastAcked = 0;
for (int i = 0; i < NumWriterIters; i++)
{
Log.Info("Starting writer " + i + "\n-------------------");
QuorumJournalManager qjm = CreateRandomFaultyQJM(cluster, r);
try
{
long recovered;
try
{
recovered = QJMTestUtil.RecoverAndReturnLastTxn(qjm);
}
catch (Exception t)
{
Log.Info("Failed recovery", t);
CheckException(t);
continue;
}
NUnit.Framework.Assert.IsTrue("Recovered only up to txnid " + recovered + " but had gotten an ack for "
+ lastAcked, recovered >= lastAcked);
txid = recovered + 1;
// Periodically purge old data on disk so it's easier to look
// at failure cases.
if (txid > 100 && i % 10 == 1)
{
qjm.PurgeLogsOlderThan(txid - 100);
}
Holder <Exception> thrown = new Holder <Exception>(null);
for (int j = 0; j < SegmentsPerWriter; j++)
{
lastAcked = WriteSegmentUntilCrash(cluster, qjm, txid, 4, thrown);
if (thrown.held != null)
{
Log.Info("Failed write", thrown.held);
CheckException(thrown.held);
break;
}
txid += 4;
}
}
finally
{
qjm.Close();
}
}
}
finally
{
cluster.Shutdown();
}
}