HDFS-12217. HDFS snapshots doesn't capture all open files when one of the open files is deleted.

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/protocol/SnapshotException.java

@@ -30,4 +30,8 @@ public class SnapshotException extends IOException {
   public SnapshotException(final Throwable cause) {
     super(cause);
   }
+
+  public SnapshotException(final String message, final Throwable cause) {
+    super(message, cause);
+  }
 }

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

@@ -4971,7 +4971,7 @@ public class FSNamesystem implements Namesystem, FSNamesystemMBean,
     return blockId;
   }
 
-  private boolean isFileDeleted(INodeFile file) {
+  boolean isFileDeleted(INodeFile file) {
     // Not in the inodeMap or in the snapshot but marked deleted.
     if (dir.getInode(file.getId()) == null) {
       return true;

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

@@ -162,18 +162,25 @@ public class LeaseManager {
    *
    * @return Set<INodesInPath>
    */
-  public Set<INodesInPath> getINodeWithLeases() {
+  @VisibleForTesting
+  Set<INodesInPath> getINodeWithLeases() throws IOException {
     return getINodeWithLeases(null);
   }
 
   private synchronized INode[] getINodesWithLease() {
-    int inodeCount = 0;
-    INode[] inodes = new INode[leasesById.size()];
+    List<INode> inodes = new ArrayList<>(leasesById.size());
+    INode currentINode;
     for (long inodeId : leasesById.keySet()) {
-      inodes[inodeCount] = fsnamesystem.getFSDirectory().getInode(inodeId);
-      inodeCount++;
+      currentINode = fsnamesystem.getFSDirectory().getInode(inodeId);
+      // A file with an active lease could get deleted, or its
+      // parent directories could get recursively deleted.
+      if (currentINode != null &&
+          currentINode.isFile() &&
+          !fsnamesystem.isFileDeleted(currentINode.asFile())) {
+        inodes.add(currentINode);
+      }
     }
-    return inodes;
+    return inodes.toArray(new INode[0]);
   }
 
   /**
@@ -186,7 +193,7 @@ public class LeaseManager {
    * @return Set<INodesInPath>
    */
   public Set<INodesInPath> getINodeWithLeases(final INodeDirectory
-      ancestorDir) {
+      ancestorDir) throws IOException {
     assert fsnamesystem.hasReadLock();
     final long startTimeMs = Time.monotonicNow();
     Set<INodesInPath> iipSet = new HashSet<>();
@@ -233,7 +240,7 @@ public class LeaseManager {
       try {
         iipSet.addAll(f.get());
       } catch (Exception e) {
-        LOG.warn("INode filter task encountered exception: ", e);
+        throw new IOException("Failed to get files with active leases", e);
       }
     }
     final long endTimeMs = Time.monotonicNow();

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

@@ -195,11 +195,17 @@ public class DirectorySnapshottableFeature extends DirectoryWithSnapshotFeature
     s.getRoot().setModificationTime(now, Snapshot.CURRENT_STATE_ID);
 
     if (captureOpenFiles) {
-      Set<INodesInPath> openFilesIIP =
-          leaseManager.getINodeWithLeases(snapshotRoot);
-      for (INodesInPath openFileIIP : openFilesIIP)  {
-        INodeFile openFile = openFileIIP.getLastINode().asFile();
-        openFile.recordModification(openFileIIP.getLatestSnapshotId());
+      try {
+        Set<INodesInPath> openFilesIIP =
+            leaseManager.getINodeWithLeases(snapshotRoot);
+        for (INodesInPath openFileIIP : openFilesIIP) {
+          INodeFile openFile = openFileIIP.getLastINode().asFile();
+          openFile.recordModification(openFileIIP.getLatestSnapshotId());
+        }
+      } catch (Exception e) {
+        throw new SnapshotException("Failed to add snapshot: Unable to " +
+            "capture all open files under the snapshot dir " +
+            snapshotRoot.getFullPathName() + " for snapshot '" + name + "'", e);
       }
     }
     return s;

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/namenode/snapshot/TestOpenFilesWithSnapshot.java

@@ -496,6 +496,132 @@ public class TestOpenFilesWithSnapshot {
     flumeOutputStream.close();
   }
 
+  /**
+   * Test snapshot capturing open files when an open file with active lease
+   * is deleted by the client.
+   */
+  @Test (timeout = 120000)
+  public void testSnapshotsForOpenFilesAndDeletion() throws Exception {
+    // Construct the directory tree
+    final Path snapRootDir = new Path("/level_0_A");
+    final String flumeFileName = "flume.log";
+    final String hbaseFileName = "hbase.log";
+    final String snap1Name = "snap_1";
+    final String snap2Name = "snap_2";
+    final String snap3Name = "snap_3";
+
+    // Create files and open streams
+    final Path flumeFile = new Path(snapRootDir, flumeFileName);
+    createFile(flumeFile);
+    final Path hbaseFile = new Path(snapRootDir, hbaseFileName);
+    createFile(hbaseFile);
+    FSDataOutputStream flumeOutputStream = fs.append(flumeFile);
+    FSDataOutputStream hbaseOutputStream = fs.append(hbaseFile);
+
+    // Create Snapshot S1
+    final Path snap1Dir = SnapshotTestHelper.createSnapshot(
+        fs, snapRootDir, snap1Name);
+    final Path flumeS1Path = new Path(snap1Dir, flumeFileName);
+    final long flumeFileLengthAfterS1 = fs.getFileStatus(flumeFile).getLen();
+    final Path hbaseS1Path = new Path(snap1Dir, hbaseFileName);
+    final long hbaseFileLengthAfterS1 = fs.getFileStatus(hbaseFile).getLen();
+
+    // Verify if Snap S1 file length is same as the the current versions
+    Assert.assertEquals(flumeFileLengthAfterS1,
+        fs.getFileStatus(flumeS1Path).getLen());
+    Assert.assertEquals(hbaseFileLengthAfterS1,
+        fs.getFileStatus(hbaseS1Path).getLen());
+
+    long flumeFileWrittenDataLength = flumeFileLengthAfterS1;
+    long hbaseFileWrittenDataLength = hbaseFileLengthAfterS1;
+    int newWriteLength = (int) (BLOCKSIZE * 1.5);
+    byte[] buf = new byte[newWriteLength];
+    Random random = new Random();
+    random.nextBytes(buf);
+
+    // Write more data to files
+    flumeFileWrittenDataLength += writeToStream(flumeOutputStream, buf);
+    hbaseFileWrittenDataLength += writeToStream(hbaseOutputStream, buf);
+
+    // Create Snapshot S2
+    final Path snap2Dir = SnapshotTestHelper.createSnapshot(
+        fs, snapRootDir, snap2Name);
+    final Path flumeS2Path = new Path(snap2Dir, flumeFileName);
+    final Path hbaseS2Path = new Path(snap2Dir, hbaseFileName);
+
+    // Verify current files length are same as all data written till now
+    final long flumeFileLengthAfterS2 = fs.getFileStatus(flumeFile).getLen();
+    Assert.assertEquals(flumeFileWrittenDataLength, flumeFileLengthAfterS2);
+    final long hbaseFileLengthAfterS2 = fs.getFileStatus(hbaseFile).getLen();
+    Assert.assertEquals(hbaseFileWrittenDataLength, hbaseFileLengthAfterS2);
+
+    // Verify if Snap S2 file length is same as the current versions
+    Assert.assertEquals(flumeFileLengthAfterS2,
+        fs.getFileStatus(flumeS2Path).getLen());
+    Assert.assertEquals(hbaseFileLengthAfterS2,
+        fs.getFileStatus(hbaseS2Path).getLen());
+
+    // Write more data to open files
+    writeToStream(flumeOutputStream, buf);
+    hbaseFileWrittenDataLength += writeToStream(hbaseOutputStream, buf);
+
+    // Verify old snapshots have point-in-time/frozen file
+    // lengths even after the current versions have moved forward.
+    Assert.assertEquals(flumeFileLengthAfterS1,
+        fs.getFileStatus(flumeS1Path).getLen());
+    Assert.assertEquals(flumeFileLengthAfterS2,
+        fs.getFileStatus(flumeS2Path).getLen());
+    Assert.assertEquals(hbaseFileLengthAfterS1,
+        fs.getFileStatus(hbaseS1Path).getLen());
+    Assert.assertEquals(hbaseFileLengthAfterS2,
+        fs.getFileStatus(hbaseS2Path).getLen());
+
+    // Delete flume current file. Snapshots should
+    // still have references to flume file.
+    boolean flumeFileDeleted = fs.delete(flumeFile, true);
+    Assert.assertTrue(flumeFileDeleted);
+    Assert.assertFalse(fs.exists(flumeFile));
+    Assert.assertTrue(fs.exists(flumeS1Path));
+    Assert.assertTrue(fs.exists(flumeS2Path));
+
+    SnapshotTestHelper.createSnapshot(fs, snapRootDir, "tmp_snap");
+    fs.deleteSnapshot(snapRootDir, "tmp_snap");
+
+    // Delete snap_2. snap_1 still has reference to
+    // the flume file.
+    fs.deleteSnapshot(snapRootDir, snap2Name);
+    Assert.assertFalse(fs.exists(flumeS2Path));
+    Assert.assertTrue(fs.exists(flumeS1Path));
+
+    // Delete snap_1. Now all traces of flume file
+    // is gone.
+    fs.deleteSnapshot(snapRootDir, snap1Name);
+    Assert.assertFalse(fs.exists(flumeS2Path));
+    Assert.assertFalse(fs.exists(flumeS1Path));
+
+    // Create Snapshot S3
+    final Path snap3Dir = SnapshotTestHelper.createSnapshot(
+        fs, snapRootDir, snap3Name);
+    final Path hbaseS3Path = new Path(snap3Dir, hbaseFileName);
+
+    // Verify live files length is same as all data written till now
+    final long hbaseFileLengthAfterS3 = fs.getFileStatus(hbaseFile).getLen();
+    Assert.assertEquals(hbaseFileWrittenDataLength, hbaseFileLengthAfterS3);
+
+    // Write more data to open files
+    hbaseFileWrittenDataLength += writeToStream(hbaseOutputStream, buf);
+
+    // Verify old snapshots have point-in-time/frozen file
+    // lengths even after the flume open file is deleted and
+    // the hbase live file has moved forward.
+    Assert.assertEquals(hbaseFileLengthAfterS3,
+        fs.getFileStatus(hbaseS3Path).getLen());
+    Assert.assertEquals(hbaseFileWrittenDataLength,
+        fs.getFileStatus(hbaseFile).getLen());
+
+    hbaseOutputStream.close();
+  }
+
   private void restartNameNode() throws Exception {
     cluster.triggerBlockReports();
     NameNode nameNode = cluster.getNameNode();