Fix the SNN repeatedly checkpoint after fsimage transfer failure on one of the multiple NNs

Author: eddy.cao

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/FSNamesystem.java

@@ -7174,6 +7174,11 @@ public synchronized void verifyToken(DelegationTokenIdentifier identifier,
   public EditLogTailer getEditLogTailer() {
     return editLogTailer;
   }
+
+  @VisibleForTesting
+  public long getStandbyLastCheckpointTime() {
+    return standbyCheckpointer.getLastCheckpointTime();
+  }
 
   @VisibleForTesting
   public void setEditLogTailerForTests(EditLogTailer tailer) {

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/namenode/ha/StandbyCheckpointer.java

@@ -342,7 +342,7 @@ public TransferFsImage.TransferResult call()
       throw ie;
     }
 
-    if (!ioes.isEmpty()) {
+    if (ioes.size() > activeNNAddresses.size() / 2) {
       throw MultipleIOException.createIOException(ioes);
     }
   }
@@ -375,6 +375,11 @@ static int getCanceledCount() {
     return canceledCount;
   }
 
+  @VisibleForTesting
+  public long getLastCheckpointTime() {
+    return lastCheckpointTime;
+  }
+
   private long countUncheckpointedTxns() {
     FSImage img = namesystem.getFSImage();
     return img.getCorrectLastAppliedOrWrittenTxId() -

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/namenode/ha/TestStandbyCheckpoints.java

@@ -663,7 +663,46 @@ private void doCreate() throws IOException {
     out.write(42);
     out.close();
   }
-  
+
+  @Test(timeout = 300000)
+  public void testPutFsimagePartFailed() throws Exception {
+    for (int i = 1; i < NUM_NNS; i++) {
+      cluster.shutdownNameNode(i);
+
+      // Make true checkpoint for DFS_NAMENODE_CHECKPOINT_PERIOD_KEY
+      cluster.getConfiguration(i).setInt(DFSConfigKeys.DFS_NAMENODE_CHECKPOINT_PERIOD_KEY, 3);
+      cluster.getConfiguration(i).setInt(DFSConfigKeys.DFS_NAMENODE_CHECKPOINT_TXNS_KEY, 1000);
+    }
+    doEdits(0, 10);
+    cluster.transitionToStandby(0);
+
+    for (int i = 1; i < NUM_NNS; i++) {
+      cluster.restartNameNode(i, false);
+    }
+    cluster.waitClusterUp();
+    setNNs();
+
+    for (int i = 0; i < NUM_NNS; i++) {
+      // Once the standby catches up, it should do a checkpoint
+      // and save to local directories.
+      HATestUtil.waitForCheckpoint(cluster, i, ImmutableList.of(12));
+    }
+
+    long snnCheckpointTime1 = nns[1].getNamesystem().getStandbyLastCheckpointTime();
+    cluster.transitionToActive(0);
+    cluster.transitionToObserver(2);
+    cluster.shutdownNameNode(2);
+
+    doEdits(11, 20);
+    nns[0].getRpcServer().rollEditLog();
+    HATestUtil.waitForCheckpoint(cluster, 0, ImmutableList.of(23));
+
+    long snnCheckpointTime2 = nns[1].getNamesystem().getStandbyLastCheckpointTime();
+
+    // Make sure that standby namenode checkpoint success and update the lastCheckpointTime
+    // even though it send fsimage to nn2 failed because nn2 is shut down.
+    assertTrue(snnCheckpointTime2 > snnCheckpointTime1);
+  }
 
   /**
    * A codec which just slows down the saving of the image significantly