HDFS-16420. Avoid deleting unique data blocks when deleting redundancy striped blocks. (#3880)

Reviewed-by: litao <tomleescut@gmail.com>
Signed-off-by: Takanobu Asanuma <tasanuma@apache.org>
(cherry picked from commit d8862822d22dbba8facb94e80333cea5ffeed5ba)
(cherry picked from commit 74d3e83be012a3201d0a19124bee59d1e6d93b47)

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockManager.java

@@ -746,6 +746,11 @@ public class BlockManager implements BlockStatsMXBean {
     return placementPolicies.getPolicy(CONTIGUOUS);
   }
 
+  @VisibleForTesting
+  public BlockPlacementPolicy getStriptedBlockPlacementPolicy() {
+    return placementPolicies.getPolicy(STRIPED);
+  }
+
   /** Dump meta data to out. */
   public void metaSave(PrintWriter out) {
     assert namesystem.hasReadLock(); // TODO: block manager read lock and NS write lock

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockPlacementPolicy.java

@@ -197,8 +197,9 @@ public abstract class BlockPlacementPolicy {
     if (moreThanOne.remove(cur)) {
       if (storages.size() == 1) {
         final DatanodeStorageInfo remaining = storages.get(0);
-        moreThanOne.remove(remaining);
-        exactlyOne.add(remaining);
+        if (moreThanOne.remove(remaining)) {
+          exactlyOne.add(remaining);
+        }
       }
     } else {
       exactlyOne.remove(cur);

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/BaseReplicationPolicyTest.java

@@ -50,6 +50,7 @@ abstract public class BaseReplicationPolicyTest {
   protected NameNode namenode;
   protected DatanodeManager dnManager;
   protected BlockPlacementPolicy replicator;
+  private BlockPlacementPolicy striptedPolicy;
   protected final String filename = "/dummyfile.txt";
   protected DatanodeStorageInfo[] storages;
   protected String blockPlacementPolicy;
@@ -90,6 +91,7 @@ abstract public class BaseReplicationPolicyTest {
 
     final BlockManager bm = namenode.getNamesystem().getBlockManager();
     replicator = bm.getBlockPlacementPolicy();
+    striptedPolicy = bm.getStriptedBlockPlacementPolicy();
     cluster = bm.getDatanodeManager().getNetworkTopology();
     dnManager = bm.getDatanodeManager();
     // construct network topology
@@ -111,6 +113,10 @@ abstract public class BaseReplicationPolicyTest {
     }
   }
 
+  public BlockPlacementPolicy getStriptedPolicy() {
+    return striptedPolicy;
+  }
+
   @After
   public void tearDown() throws Exception {
     namenode.stop();

hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/server/blockmanagement/TestReplicationPolicy.java

@@ -1017,6 +1017,64 @@ public class TestReplicationPolicy extends BaseReplicationPolicyTest {
     assertEquals(chosen, storages[1]);
   }
 
+  /**
+   * Test for the chooseReplicaToDelete are processed based on
+   * EC and STRIPED Policy.
+   */
+  @Test
+  public void testStripedChooseReplicaToDelete() throws Exception {
+    List<DatanodeStorageInfo> replicaList = new ArrayList<>();
+    List<DatanodeStorageInfo> candidate = new ArrayList<>();
+    final Map<String, List<DatanodeStorageInfo>> rackMap
+        = new HashMap<String, List<DatanodeStorageInfo>>();
+
+    replicaList.add(storages[0]);
+    replicaList.add(storages[1]);
+    replicaList.add(storages[2]);
+    replicaList.add(storages[4]);
+
+    candidate.add(storages[0]);
+    candidate.add(storages[2]);
+    candidate.add(storages[4]);
+
+    // Refresh the last update time for all the datanodes
+    for (int i = 0; i < dataNodes.length; i++) {
+      DFSTestUtil.resetLastUpdatesWithOffset(dataNodes[i], 0);
+    }
+
+    List<DatanodeStorageInfo> first = new ArrayList<>();
+    List<DatanodeStorageInfo> second = new ArrayList<>();
+    BlockPlacementPolicy policy = getStriptedPolicy();
+    policy.splitNodesWithRack(replicaList, candidate, rackMap, first,
+        second);
+    // storages[0] is in first set as its rack has two replica nodes,
+    // while storages[2] and dataNodes[4] are in second set.
+    assertEquals(1, first.size());
+    assertEquals(2, second.size());
+    List<StorageType> excessTypes = new ArrayList<>();
+    excessTypes.add(StorageType.DEFAULT);
+    DatanodeStorageInfo chosen = ((BlockPlacementPolicyDefault) policy)
+        .chooseReplicaToDelete(first, second, excessTypes, rackMap);
+    // Within all storages, storages[0] is in the rack that has two replica blocks
+    assertEquals(chosen, storages[0]);
+    policy.adjustSetsWithChosenReplica(rackMap, first, second, chosen);
+    assertEquals(0, first.size());
+    assertEquals(2, second.size());
+
+    // Within second set, storages[2] should be next to be deleted in order.
+    excessTypes.add(StorageType.DEFAULT);
+    chosen = ((BlockPlacementPolicyDefault) policy).chooseReplicaToDelete(
+        first, second, excessTypes, rackMap);
+    assertEquals(chosen, storages[2]);
+    policy.adjustSetsWithChosenReplica(rackMap, first, second, chosen);
+    assertEquals(0, first.size());
+    assertEquals(1, second.size());
+
+    chosen = ((BlockPlacementPolicyDefault) policy).chooseReplicaToDelete(
+        first, second, excessTypes, rackMap);
+    assertEquals(chosen, null);
+  }
+
   private long calculateRemaining(DatanodeDescriptor dataNode) {
     long sum = 0;
     for (DatanodeStorageInfo storageInfo: dataNode.getStorageInfos()){