HBASE-18471 The DeleteFamily cell is skipped when StoreScanner seeks to next column

hbase-common/src/main/java/org/apache/hadoop/hbase/CellUtil.java

@@ -3135,4 +3135,57 @@ public final class CellUtil {
       return Type.DeleteFamily.getCode();
     }
   }
+
+  /**
+   * @return An new cell is located following input cell. If both of type and timestamp are
+   *         minimum, the input cell will be returned directly.
+   */
+  @InterfaceAudience.Private
+  public static Cell createNextOnRowCol(Cell cell) {
+    long ts = cell.getTimestamp();
+    byte type = cell.getTypeByte();
+    if (type != Type.Minimum.getCode()) {
+      type = KeyValue.Type.values()[KeyValue.Type.codeToType(type).ordinal() - 1].getCode();
+    } else if (ts != HConstants.OLDEST_TIMESTAMP) {
+      ts = ts - 1;
+      type = Type.Maximum.getCode();
+    } else {
+      return cell;
+    }
+    return createNextOnRowCol(cell, ts, type);
+  }
+
+  private static Cell createNextOnRowCol(Cell cell, long ts, byte type) {
+    if (cell instanceof ByteBufferCell) {
+      return new LastOnRowColByteBufferCell(((ByteBufferCell) cell).getRowByteBuffer(),
+              ((ByteBufferCell) cell).getRowPosition(), cell.getRowLength(),
+              ((ByteBufferCell) cell).getFamilyByteBuffer(),
+              ((ByteBufferCell) cell).getFamilyPosition(), cell.getFamilyLength(),
+              ((ByteBufferCell) cell).getQualifierByteBuffer(),
+              ((ByteBufferCell) cell).getQualifierPosition(), cell.getQualifierLength()) {
+        @Override
+        public long getTimestamp() {
+          return ts;
+        }
+
+        @Override
+        public byte getTypeByte() {
+          return type;
+        }
+      };
+    }
+    return new LastOnRowColCell(cell.getRowArray(), cell.getRowOffset(), cell.getRowLength(),
+            cell.getFamilyArray(), cell.getFamilyOffset(), cell.getFamilyLength(),
+            cell.getQualifierArray(), cell.getQualifierOffset(), cell.getQualifierLength()) {
+      @Override
+      public long getTimestamp() {
+        return ts;
+      }
+
+      @Override
+      public byte getTypeByte() {
+        return type;
+      }
+    };
+  }
 }

hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/querymatcher/ScanQueryMatcher.java

@@ -290,6 +290,19 @@ public abstract class ScanQueryMatcher implements ShipperListener {
   public abstract boolean moreRowsMayExistAfter(Cell cell);
 
   public Cell getKeyForNextColumn(Cell cell) {
+    // We aren't sure whether any DeleteFamily cells exist, so we can't skip to next column.
+    // TODO: Current way disable us to seek to next column quickly. Is there any better solution?
+    // see HBASE-18471 for more details
+    // see TestFromClientSide3#testScanAfterDeletingSpecifiedRow
+    // see TestFromClientSide3#testScanAfterDeletingSpecifiedRowV2
+    if (cell.getQualifierLength() == 0) {
+      Cell nextKey = CellUtil.createNextOnRowCol(cell);
+      if (nextKey != cell) {
+        return nextKey;
+      }
+      // The cell is at the end of row/family/qualifier, so it is impossible to find any DeleteFamily cells.
+      // Let us seek to next column.
+    }
     ColumnCount nextColumn = columns.getColumnHint();
     if (nextColumn == null) {
       return CellUtil.createLastOnRowCol(cell);

hbase-server/src/test/java/org/apache/hadoop/hbase/client/TestFromClientSide3.java

@@ -34,6 +34,7 @@ import org.apache.commons.logging.Log;
 import org.apache.commons.logging.LogFactory;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.hbase.Cell;
+import org.apache.hadoop.hbase.CellUtil;
 import org.apache.hadoop.hbase.Coprocessor;
 import org.apache.hadoop.hbase.coprocessor.ObserverContext;
 import org.apache.hadoop.hbase.coprocessor.RegionCoprocessorEnvironment;
@@ -159,6 +160,106 @@ public class TestFromClientSide3 {
     }
   }
 
+  private static List<Cell> toList(ResultScanner scanner) {
+    try {
+      List<Cell> cells = new ArrayList<>();
+      for (Result r : scanner) {
+        cells.addAll(r.listCells());
+      }
+      return cells;
+    } finally {
+      scanner.close();
+    }
+  }
+
+  @Test
+  public void testScanAfterDeletingSpecifiedRow() throws IOException {
+    TableName tableName = TableName.valueOf(name.getMethodName());
+    TableDescriptor desc = TableDescriptorBuilder.newBuilder(tableName)
+            .addColumnFamily(ColumnFamilyDescriptorBuilder.of(FAMILY))
+            .build();
+    TEST_UTIL.getAdmin().createTable(desc);
+    byte[] row = Bytes.toBytes("SpecifiedRow");
+    byte[] value0 = Bytes.toBytes("value_0");
+    byte[] value1 = Bytes.toBytes("value_1");
+    try (Table t = TEST_UTIL.getConnection().getTable(tableName)) {
+      Put put = new Put(row);
+      put.addColumn(FAMILY, QUALIFIER, VALUE);
+      t.put(put);
+      Delete d = new Delete(row);
+      t.delete(d);
+      put = new Put(row);
+      put.addColumn(FAMILY, null, value0);
+      t.put(put);
+      put = new Put(row);
+      put.addColumn(FAMILY, null, value1);
+      t.put(put);
+      List<Cell> cells = toList(t.getScanner(new Scan()));
+      assertEquals(1, cells.size());
+      assertEquals("value_1", Bytes.toString(CellUtil.cloneValue(cells.get(0))));
+
+      cells = toList(t.getScanner(new Scan().addFamily(FAMILY)));
+      assertEquals(1, cells.size());
+      assertEquals("value_1", Bytes.toString(CellUtil.cloneValue(cells.get(0))));
+
+      cells = toList(t.getScanner(new Scan().addColumn(FAMILY, QUALIFIER)));
+      assertEquals(0, cells.size());
+
+      TEST_UTIL.getAdmin().flush(tableName);
+      cells = toList(t.getScanner(new Scan()));
+      assertEquals(1, cells.size());
+      assertEquals("value_1", Bytes.toString(CellUtil.cloneValue(cells.get(0))));
+
+      cells = toList(t.getScanner(new Scan().addFamily(FAMILY)));
+      assertEquals(1, cells.size());
+      assertEquals("value_1", Bytes.toString(CellUtil.cloneValue(cells.get(0))));
+
+      cells = toList(t.getScanner(new Scan().addColumn(FAMILY, QUALIFIER)));
+      assertEquals(0, cells.size());
+    }
+  }
+
+  @Test
+  public void testScanAfterDeletingSpecifiedRowV2() throws IOException {
+    TableName tableName = TableName.valueOf(name.getMethodName());
+    TableDescriptor desc = TableDescriptorBuilder.newBuilder(tableName)
+            .addColumnFamily(ColumnFamilyDescriptorBuilder.of(FAMILY))
+            .build();
+    TEST_UTIL.getAdmin().createTable(desc);
+    byte[] row = Bytes.toBytes("SpecifiedRow");
+    byte[] qual0 = Bytes.toBytes("qual0");
+    byte[] qual1 = Bytes.toBytes("qual1");
+    try (Table t = TEST_UTIL.getConnection().getTable(tableName)) {
+      Delete d = new Delete(row);
+      t.delete(d);
+
+      Put put = new Put(row);
+      put.addColumn(FAMILY, null, VALUE);
+      t.put(put);
+
+      put = new Put(row);
+      put.addColumn(FAMILY, qual1, qual1);
+      t.put(put);
+
+      put = new Put(row);
+      put.addColumn(FAMILY, qual0, qual0);
+      t.put(put);
+
+      Result r = t.get(new Get(row));
+      assertEquals(3, r.size());
+      assertEquals("testValue", Bytes.toString(CellUtil.cloneValue(r.rawCells()[0])));
+      assertEquals("qual0", Bytes.toString(CellUtil.cloneValue(r.rawCells()[1])));
+      assertEquals("qual1", Bytes.toString(CellUtil.cloneValue(r.rawCells()[2])));
+
+      TEST_UTIL.getAdmin().flush(tableName);
+      r = t.get(new Get(row));
+      assertEquals(3, r.size());
+      assertEquals("testValue", Bytes.toString(CellUtil.cloneValue(r.rawCells()[0])));
+      assertEquals("qual0", Bytes.toString(CellUtil.cloneValue(r.rawCells()[1])));
+      assertEquals("qual1", Bytes.toString(CellUtil.cloneValue(r.rawCells()[2])));
+    }
+  }
+
   // override the config settings at the CF level and ensure priority
   @Test(timeout = 60000)
   public void testAdvancedConfigOverride() throws Exception {