HDFS-5840. Follow-up to HDFS-5138 to improve error handling during partial upgrade failures. Contributed by Aaron T. Myers, Suresh Srinivas, and Jing Zhao.

Jing9 · Jing9 · commit 3edb8ad9c1c4 · 2014-03-25T06:48:15.000Z
git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1581260 13f79535-47bb-0310-9956-ffa450edef68
diff --git a/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt b/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
@@ -687,6 +687,9 @@ Release 2.4.0 - UNRELEASED
 
     HDFS-5846. Assigning DEFAULT_RACK in resolveNetworkLocation method can break
     data resiliency. (Nikola Vujic via cnauroth)
+    
+    HDFS-5840. Follow-up to HDFS-5138 to improve error handling during partial
+    upgrade failures. (atm, jing9 and suresh via jing9)
 
   BREAKDOWN OF HDFS-5698 SUBTASKS AND RELATED JIRAS
 
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/DFSUtil.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/DFSUtil.java
@@ -1680,23 +1680,19 @@ public static HttpServer2.Builder httpServerTemplateForNNAndJN(
    * 
    * @param objects the collection of objects to check for equality.
    */
-  public static void assertAllResultsEqual(Collection<?> objects) {
-    Object[] resultsArray = objects.toArray();
-    
-    if (resultsArray.length == 0)
+  public static void assertAllResultsEqual(Collection<?> objects)
+      throws AssertionError {
+    if (objects.size() == 0 || objects.size() == 1)
       return;
     
-    for (int i = 0; i < resultsArray.length; i++) {
-      if (i == 0)
-        continue;
-      else {
-        Object currElement = resultsArray[i];
-        Object lastElement = resultsArray[i - 1];
-        if ((currElement == null && currElement != lastElement) ||
-            (currElement != null && !currElement.equals(lastElement))) {
-          throw new AssertionError("Not all elements match in results: " +
-            Arrays.toString(resultsArray));
-        }
+    Object[] resultsArray = objects.toArray();
+    for (int i = 1; i < resultsArray.length; i++) {
+      Object currElement = resultsArray[i];
+      Object lastElement = resultsArray[i - 1];
+      if ((currElement == null && currElement != lastElement) ||
+          (currElement != null && !currElement.equals(lastElement))) {
+        throw new AssertionError("Not all elements match in results: " +
+          Arrays.toString(resultsArray));
       }
     }
   }
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/client/QuorumJournalManager.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/client/QuorumJournalManager.java
@@ -571,7 +571,11 @@ public boolean canRollBack(StorageInfo storage, StorageInfo prevStorage,
       
       // Either they all return the same thing or this call fails, so we can
       // just return the first result.
-      DFSUtil.assertAllResultsEqual(call.getResults().values());
+      try {
+        DFSUtil.assertAllResultsEqual(call.getResults().values());
+      } catch (AssertionError ae) {
+        throw new IOException("Results differed for canRollBack", ae);
+      }
       for (Boolean result : call.getResults().values()) {
         return result;
       }
@@ -636,7 +640,11 @@ public long getJournalCTime() throws IOException {
       
       // Either they all return the same thing or this call fails, so we can
       // just return the first result.
-      DFSUtil.assertAllResultsEqual(call.getResults().values());
+      try {
+        DFSUtil.assertAllResultsEqual(call.getResults().values());
+      } catch (AssertionError ae) {
+        throw new IOException("Results differed for getJournalCTime", ae);
+      }
       for (Long result : call.getResults().values()) {
         return result;
       }
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/JNStorage.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/JNStorage.java
@@ -65,15 +65,15 @@ class JNStorage extends Storage {
    * @param errorReporter a callback to report errors
    * @throws IOException 
    */
-  protected JNStorage(Configuration conf, File logDir,
+  protected JNStorage(Configuration conf, File logDir, StartupOption startOpt,
       StorageErrorReporter errorReporter) throws IOException {
     super(NodeType.JOURNAL_NODE);
     
     sd = new StorageDirectory(logDir);
     this.addStorageDir(sd);
     this.fjm = new FileJournalManager(conf, sd, errorReporter);
-    
-    analyzeStorage();
+
+    analyzeAndRecoverStorage(startOpt);
   }
   
   FileJournalManager getJournalManager() {
@@ -216,6 +216,18 @@ protected void setLayoutVersion(Properties props, StorageDirectory sd)
     layoutVersion = lv;
   }
 
+  void analyzeAndRecoverStorage(StartupOption startOpt) throws IOException {
+    this.state = sd.analyzeStorage(startOpt, this);
+    final boolean needRecover = state != StorageState.NORMAL
+        && state != StorageState.NON_EXISTENT
+        && state != StorageState.NOT_FORMATTED;
+    if (state == StorageState.NORMAL && startOpt != StartupOption.ROLLBACK) {
+      readProperties(sd);
+    } else if (needRecover) {
+      sd.doRecover(state);
+    }
+  }
+
   void checkConsistentNamespace(NamespaceInfo nsInfo)
       throws IOException {
     if (nsInfo.getNamespaceID() != getNamespaceID()) {
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/Journal.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/Journal.java
@@ -43,6 +43,7 @@
 import org.apache.hadoop.hdfs.qjournal.protocol.QJournalProtocolProtos.PrepareRecoveryResponseProto;
 import org.apache.hadoop.hdfs.qjournal.protocol.QJournalProtocolProtos.SegmentStateProto;
 import org.apache.hadoop.hdfs.qjournal.protocol.RequestInfo;
+import org.apache.hadoop.hdfs.server.common.HdfsServerConstants.StartupOption;
 import org.apache.hadoop.hdfs.server.common.StorageErrorReporter;
 import org.apache.hadoop.hdfs.server.common.StorageInfo;
 import org.apache.hadoop.hdfs.server.namenode.EditLogOutputStream;
@@ -138,8 +139,9 @@ public class Journal implements Closeable {
   private static final int WARN_SYNC_MILLIS_THRESHOLD = 1000;
 
   Journal(Configuration conf, File logDir, String journalId,
-      StorageErrorReporter errorReporter) throws IOException {
-    storage = new JNStorage(conf, logDir, errorReporter);
+      StartupOption startOpt, StorageErrorReporter errorReporter)
+      throws IOException {
+    storage = new JNStorage(conf, logDir, startOpt, errorReporter);
     this.journalId = journalId;
 
     refreshCachedData();
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/JournalNode.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/qjournal/server/JournalNode.java
@@ -34,6 +34,7 @@
 import org.apache.hadoop.hdfs.DFSConfigKeys;
 import org.apache.hadoop.hdfs.HdfsConfiguration;
 import org.apache.hadoop.hdfs.qjournal.client.QuorumJournalManager;
+import org.apache.hadoop.hdfs.server.common.HdfsServerConstants.StartupOption;
 import org.apache.hadoop.hdfs.server.common.StorageErrorReporter;
 import org.apache.hadoop.hdfs.server.common.StorageInfo;
 import org.apache.hadoop.io.IOUtils;
@@ -77,19 +78,24 @@ public class JournalNode implements Tool, Configurable, JournalNodeMXBean {
    */
   private int resultCode = 0;
 
-  synchronized Journal getOrCreateJournal(String jid) throws IOException {
+  synchronized Journal getOrCreateJournal(String jid, StartupOption startOpt)
+      throws IOException {
     QuorumJournalManager.checkJournalId(jid);
     
     Journal journal = journalsById.get(jid);
     if (journal == null) {
       File logDir = getLogDir(jid);
       LOG.info("Initializing journal in directory " + logDir);      
-      journal = new Journal(conf, logDir, jid, new ErrorReporter());
+      journal = new Journal(conf, logDir, jid, startOpt, new ErrorReporter());
       journalsById.put(jid, journal);
     }
     
     return journal;
   }
+  
+  Journal getOrCreateJournal(String jid) throws IOException {
+    return getOrCreateJournal(jid, StartupOption.REGULAR);
+  }
 
   @Override
   public void setConf(Configuration conf) {
@@ -301,12 +307,12 @@ public void doFinalize(String journalId) throws IOException {
 
   public Boolean canRollBack(String journalId, StorageInfo storage,
       StorageInfo prevStorage, int targetLayoutVersion) throws IOException {
-    return getOrCreateJournal(journalId).canRollBack(storage, prevStorage,
-        targetLayoutVersion);
+    return getOrCreateJournal(journalId, StartupOption.ROLLBACK).canRollBack(
+        storage, prevStorage, targetLayoutVersion);
   }
 
   public void doRollback(String journalId) throws IOException {
-    getOrCreateJournal(journalId).doRollback();
+    getOrCreateJournal(journalId, StartupOption.ROLLBACK).doRollback();
   }
 
   public void discardSegments(String journalId, long startTxId)
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSEditLog.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSEditLog.java
@@ -44,9 +44,9 @@
 import org.apache.hadoop.hdfs.security.token.delegation.DelegationTokenIdentifier;
 import org.apache.hadoop.hdfs.server.blockmanagement.BlockInfo;
 import org.apache.hadoop.hdfs.server.common.HdfsServerConstants.NamenodeRole;
-import org.apache.hadoop.hdfs.server.common.Storage;
 import org.apache.hadoop.hdfs.server.common.Storage.FormatConfirmable;
 import org.apache.hadoop.hdfs.server.common.Storage.StorageDirectory;
+import org.apache.hadoop.hdfs.server.common.StorageInfo;
 import org.apache.hadoop.hdfs.server.namenode.FSEditLogOp.AddBlockOp;
 import org.apache.hadoop.hdfs.server.namenode.FSEditLogOp.AddCacheDirectiveInfoOp;
 import org.apache.hadoop.hdfs.server.namenode.FSEditLogOp.AddCachePoolOp;
@@ -1411,7 +1411,7 @@ public synchronized void doFinalizeOfSharedLog() throws IOException {
     }
   }
   
-  public synchronized boolean canRollBackSharedLog(Storage prevStorage,
+  public synchronized boolean canRollBackSharedLog(StorageInfo prevStorage,
       int targetLayoutVersion) throws IOException {
     for (JournalAndStream jas : journalSet.getAllJournalStreams()) {
       if (jas.isShared()) {
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSImage.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSImage.java
@@ -393,6 +393,10 @@ void doUpgrade(FSNamesystem target) throws IOException {
 
     saveFSImageInAllDirs(target, editLog.getLastWrittenTxId());
 
+    // upgrade shared edit storage first
+    if (target.isHaEnabled()) {
+      editLog.doUpgradeOfSharedLog();
+    }
     for (Iterator<StorageDirectory> it = storage.dirIterator(false); it.hasNext();) {
       StorageDirectory sd = it.next();
       try {
@@ -402,9 +406,6 @@ void doUpgrade(FSNamesystem target) throws IOException {
         continue;
       }
     }
-    if (target.isHaEnabled()) {
-      editLog.doUpgradeOfSharedLog();
-    }
     storage.reportErrorsOnDirectories(errorSDs);
     
     isUpgradeFinalized = false;
@@ -430,14 +431,19 @@ void doRollback(FSNamesystem fsns) throws IOException {
             HdfsConstants.NAMENODE_LAYOUT_VERSION)) {
           continue;
         }
+        LOG.info("Can perform rollback for " + sd);
         canRollback = true;
       }
       
       if (fsns.isHaEnabled()) {
         // If HA is enabled, check if the shared log can be rolled back as well.
         editLog.initJournalsForWrite();
-        canRollback |= editLog.canRollBackSharedLog(prevState.getStorage(),
-            HdfsConstants.NAMENODE_LAYOUT_VERSION);
+        boolean canRollBackSharedEditLog = editLog.canRollBackSharedLog(
+            prevState.getStorage(), HdfsConstants.NAMENODE_LAYOUT_VERSION);
+        if (canRollBackSharedEditLog) {
+          LOG.info("Can perform rollback for shared edit log.");
+          canRollback = true;
+        }
       }
       
       if (!canRollback)
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/NNUpgradeUtil.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/NNUpgradeUtil.java
@@ -26,6 +26,8 @@
 import org.apache.hadoop.hdfs.server.common.Storage.StorageDirectory;
 import org.apache.hadoop.hdfs.server.common.StorageInfo;
 
+import com.google.common.base.Preconditions;
+
 abstract class NNUpgradeUtil {
   
   private static final Log LOG = LogFactory.getLog(NNUpgradeUtil.class);
@@ -82,7 +84,8 @@ static void doFinalize(StorageDirectory sd) throws IOException {
       return;
     }
     LOG.info("Finalizing upgrade of storage directory " + sd.getRoot());
-    assert sd.getCurrentDir().exists() : "Current directory must exist.";
+    Preconditions.checkState(sd.getCurrentDir().exists(),
+        "Current directory must exist.");
     final File tmpDir = sd.getFinalizedTmp();
     // rename previous to tmp and remove
     NNStorage.rename(prevDir, tmpDir);
@@ -105,9 +108,14 @@ static void doPreUpgrade(StorageDirectory sd) throws IOException {
     File curDir = sd.getCurrentDir();
     File prevDir = sd.getPreviousDir();
     File tmpDir = sd.getPreviousTmp();
-    assert curDir.exists() : "Current directory must exist.";
-    assert !prevDir.exists() : "previous directory must not exist.";
-    assert !tmpDir.exists() : "previous.tmp directory must not exist.";
+
+    Preconditions.checkState(curDir.exists(),
+        "Current directory must exist for preupgrade.");
+    Preconditions.checkState(!prevDir.exists(),
+        "Previous directory must not exist for preupgrade.");
+    Preconditions.checkState(!tmpDir.exists(),
+        "Previous.tmp directory must not exist for preupgrade."
+            + "Consider restarting for recovery.");
 
     // rename current to tmp
     NNStorage.rename(curDir, tmpDir);
@@ -136,6 +144,11 @@ static void doUpgrade(StorageDirectory sd, Storage storage) throws
       
       File prevDir = sd.getPreviousDir();
       File tmpDir = sd.getPreviousTmp();
+      Preconditions.checkState(!prevDir.exists(),
+          "previous directory must not exist for upgrade.");
+      Preconditions.checkState(tmpDir.exists(),
+          "previous.tmp directory must exist for upgrade.");
+
       // rename tmp to previous
       NNStorage.rename(tmpDir, prevDir);
     } catch (IOException ioe) {
@@ -154,14 +167,19 @@ static void doUpgrade(StorageDirectory sd, Storage storage) throws
   static void doRollBack(StorageDirectory sd)
       throws IOException {
     File prevDir = sd.getPreviousDir();
-    if (!prevDir.exists())
+    if (!prevDir.exists()) {
       return;
+    }
 
     File tmpDir = sd.getRemovedTmp();
-    assert !tmpDir.exists() : "removed.tmp directory must not exist.";
+    Preconditions.checkState(!tmpDir.exists(),
+        "removed.tmp directory must not exist for rollback."
+            + "Consider restarting for recovery.");
     // rename current to tmp
     File curDir = sd.getCurrentDir();
-    assert curDir.exists() : "Current directory must exist.";
+    Preconditions.checkState(curDir.exists(),
+        "Current directory must exist for rollback.");
+
     NNStorage.rename(curDir, tmpDir);
     // rename previous to current
     NNStorage.rename(prevDir, curDir);
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/site/apt/HDFSHighAvailabilityWithQJM.apt.vm b/hadoop-hdfs-project/hadoop-hdfs/src/site/apt/HDFSHighAvailabilityWithQJM.apt.vm
@@ -780,14 +780,19 @@ digest:hdfs-zkfcs:vlUvLnd8MlacsE80rDuu6ONESbM=:rwcda
 
     [[1]] Shut down all of the NNs as normal, and install the newer software.
 
-    [[2]] Start one of the NNs with the <<<'-upgrade'>>> flag.
+    [[2]] Start up all of the JNs. Note that it is <<critical>> that all the
+    JNs be running when performing the upgrade, rollback, or finalization
+    operations. If any of the JNs are down at the time of running any of these
+    operations, the operation will fail.
+
+    [[3]] Start one of the NNs with the <<<'-upgrade'>>> flag.
   
-    [[3]] On start, this NN will not enter the standby state as usual in an HA
+    [[4]] On start, this NN will not enter the standby state as usual in an HA
     setup. Rather, this NN will immediately enter the active state, perform an
     upgrade of its local storage dirs, and also perform an upgrade of the shared
     edit log.
   
-    [[4]] At this point the other NN in the HA pair will be out of sync with
+    [[5]] At this point the other NN in the HA pair will be out of sync with
     the upgraded NN. In order to bring it back in sync and once again have a highly
     available setup, you should re-bootstrap this NameNode by running the NN with
     the <<<'-bootstrapStandby'>>> flag. It is an error to start this second NN with
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/qjournal/server/TestJournal.java b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/qjournal/server/TestJournal.java
@@ -34,6 +34,7 @@
 import org.apache.hadoop.hdfs.qjournal.protocol.QJournalProtocolProtos.NewEpochResponseProtoOrBuilder;
 import org.apache.hadoop.hdfs.qjournal.protocol.QJournalProtocolProtos.SegmentStateProto;
 import org.apache.hadoop.hdfs.qjournal.protocol.RequestInfo;
+import org.apache.hadoop.hdfs.server.common.HdfsServerConstants.StartupOption;
 import org.apache.hadoop.hdfs.server.common.Storage;
 import org.apache.hadoop.hdfs.server.common.Storage.StorageDirectory;
 import org.apache.hadoop.hdfs.server.common.StorageErrorReporter;
@@ -70,7 +71,7 @@ public class TestJournal {
   public void setup() throws Exception {
     FileUtil.fullyDelete(TEST_LOG_DIR);
     conf = new Configuration();
-    journal = new Journal(conf, TEST_LOG_DIR, JID,
+    journal = new Journal(conf, TEST_LOG_DIR, JID, StartupOption.REGULAR,
       mockErrorReporter);
     journal.format(FAKE_NSINFO);
   }
@@ -179,7 +180,8 @@ public void testRestartJournal() throws Exception {
     journal.close(); // close to unlock the storage dir
     
     // Now re-instantiate, make sure history is still there
-    journal = new Journal(conf, TEST_LOG_DIR, JID, mockErrorReporter);
+    journal = new Journal(conf, TEST_LOG_DIR, JID, StartupOption.REGULAR,
+        mockErrorReporter);
     
     // The storage info should be read, even if no writer has taken over.
     assertEquals(storageString,
@@ -239,7 +241,8 @@ public void testJournalLocking() throws Exception {
 
     journal.newEpoch(FAKE_NSINFO,  1);
     try {
-      new Journal(conf, TEST_LOG_DIR, JID, mockErrorReporter);
+      new Journal(conf, TEST_LOG_DIR, JID, StartupOption.REGULAR,
+          mockErrorReporter);
       fail("Did not fail to create another journal in same dir");
     } catch (IOException ioe) {
       GenericTestUtils.assertExceptionContains(
@@ -250,7 +253,8 @@ public void testJournalLocking() throws Exception {
     
     // Journal should no longer be locked after the close() call.
     // Hence, should be able to create a new Journal in the same dir.
-    Journal journal2 = new Journal(conf, TEST_LOG_DIR, JID, mockErrorReporter);
+    Journal journal2 = new Journal(conf, TEST_LOG_DIR, JID,
+        StartupOption.REGULAR, mockErrorReporter);
     journal2.newEpoch(FAKE_NSINFO, 2);
     journal2.close();
   }
@@ -279,7 +283,8 @@ public void testFinalizeWhenEditsAreMissed() throws Exception {
     // Check that, even if we re-construct the journal by scanning the
     // disk, we don't allow finalizing incorrectly.
     journal.close();
-    journal = new Journal(conf, TEST_LOG_DIR, JID, mockErrorReporter);
+    journal = new Journal(conf, TEST_LOG_DIR, JID, StartupOption.REGULAR,
+        mockErrorReporter);
     
     try {
       journal.finalizeLogSegment(makeRI(4), 1, 6);
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/namenode/ha/TestHAStateTransitions.java b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/namenode/ha/TestHAStateTransitions.java
