Optimize the reading of numerical frame arrays in MSQ

processing/src/main/java/org/apache/druid/frame/field/DoubleArrayFieldReader.java

@@ -37,6 +37,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
   {
     return new NumericArrayFieldSelector<Double>(memory, fieldPointer)
     {
+      private static final int FIELD_SIZE = Byte.BYTES + Double.BYTES;
       final SettableFieldPointer fieldPointer = new SettableFieldPointer();
       final ColumnValueSelector<?> columnValueSelector =
           DoubleFieldReader.forArray().makeColumnValueSelector(memory, fieldPointer);
@@ -45,7 +46,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
       @Override
       public Double getIndividualValueAtMemory(long position)
       {
-        fieldPointer.setPosition(position);
+        fieldPointer.setPositionAndLength(position, FIELD_SIZE);
         if (columnValueSelector.isNull()) {
           return null;
         }
@@ -55,7 +56,7 @@ public Double getIndividualValueAtMemory(long position)
       @Override
       public int getIndividualFieldSize()
       {
-        return Byte.BYTES + Double.BYTES;
+        return FIELD_SIZE;
       }
     };
   }

processing/src/main/java/org/apache/druid/frame/field/FloatArrayFieldReader.java

@@ -37,6 +37,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
   {
     return new NumericArrayFieldSelector<Float>(memory, fieldPointer)
     {
+      private static final int FIELD_SIZE = Byte.BYTES + Float.BYTES;
       final SettableFieldPointer fieldPointer = new SettableFieldPointer();
       final ColumnValueSelector<?> columnValueSelector =
           FloatFieldReader.forArray().makeColumnValueSelector(memory, fieldPointer);
@@ -45,7 +46,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
       @Override
       public Float getIndividualValueAtMemory(long position)
       {
-        fieldPointer.setPosition(position);
+        fieldPointer.setPositionAndLength(position, FIELD_SIZE);
         if (columnValueSelector.isNull()) {
           return null;
         }
@@ -55,7 +56,7 @@ public Float getIndividualValueAtMemory(long position)
       @Override
       public int getIndividualFieldSize()
       {
-        return Byte.BYTES + Float.BYTES;
+        return FIELD_SIZE;
       }
     };
   }

processing/src/main/java/org/apache/druid/frame/field/LongArrayFieldReader.java

@@ -37,6 +37,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
   {
     return new NumericArrayFieldSelector<Long>(memory, fieldPointer)
     {
+      private static final int FIELD_SIZE = Byte.BYTES + Long.BYTES;
       final SettableFieldPointer fieldPointer = new SettableFieldPointer();
       final ColumnValueSelector<?> columnValueSelector =
           LongFieldReader.forArray().makeColumnValueSelector(memory, fieldPointer);
@@ -45,7 +46,7 @@ public ColumnValueSelector<?> makeColumnValueSelector(
       @Override
       public Long getIndividualValueAtMemory(long position)
       {
-        fieldPointer.setPosition(position);
+        fieldPointer.setPositionAndLength(position, FIELD_SIZE);
         if (columnValueSelector.isNull()) {
           return null;
         }
@@ -55,7 +56,7 @@ public Long getIndividualValueAtMemory(long position)
       @Override
       public int getIndividualFieldSize()
       {
-        return Byte.BYTES + Long.BYTES;
+        return FIELD_SIZE;
       }
     };
   }

processing/src/main/java/org/apache/druid/frame/field/NumericArrayFieldSelector.java

@@ -25,8 +25,6 @@
 import org.apache.druid.segment.ColumnValueSelector;
 
 import javax.annotation.Nullable;
-import java.util.ArrayList;
-import java.util.List;
 
 /**
  * Base implementation of the column value selector that the concrete numeric field reader implementations inherit from.
@@ -66,12 +64,8 @@ public abstract class NumericArrayFieldSelector<ElementType extends Number> impl
   /**
    * Value of the row at the location beginning at {@link #currentFieldPosition}
    */
-  private final List<ElementType> currentRow = new ArrayList<>();
-
-  /**
-   * Nullity of the row at the location beginning at {@link #currentFieldPosition}
-   */
-  private boolean currentRowIsNull;
+  @Nullable
+  private Number[] currentRow = null;
 
   public NumericArrayFieldSelector(final Memory memory, final ReadableFieldPointer fieldPointer)
   {
@@ -89,13 +83,7 @@ public void inspectRuntimeShape(RuntimeShapeInspector inspector)
   @Override
   public Object getObject()
   {
-    final List<ElementType> currentArray = computeCurrentArray();
-
-    if (currentArray == null) {
-      return null;
-    }
-
-    return currentArray.toArray();
+    return computeCurrentArray();
   }
 
   @Override
@@ -143,34 +131,29 @@ public boolean isNull()
   public abstract int getIndividualFieldSize();
 
   @Nullable
-  private List<ElementType> computeCurrentArray()
+  private Number[] computeCurrentArray()
   {
     final long fieldPosition = fieldPointer.position();
+    final long fieldLength = fieldPointer.length();
 
     if (fieldPosition != currentFieldPosition) {
-      updateCurrentArray(fieldPosition);
+      updateCurrentArray(fieldPosition, fieldLength);
     }
 
     this.currentFieldPosition = fieldPosition;
-
-    if (currentRowIsNull) {
-      return null;
-    }
     return currentRow;
-
   }
 
-  private void updateCurrentArray(final long fieldPosition)
+  private void updateCurrentArray(final long fieldPosition, final long fieldLength)
   {
-    currentRow.clear();
-    currentRowIsNull = false;
+    currentRow = null;
 
     long position = fieldPosition;
     long limit = memory.getCapacity();
 
     // Check the first byte, and if it is null, update the current value to null and return
     if (isNull()) {
-      currentRowIsNull = true;
+      // Already set the currentRow to null
       return;
     }
 
@@ -179,9 +162,13 @@ private void updateCurrentArray(final long fieldPosition)
       position++;
     }
 
+    int numElements = numElements(fieldLength);
+    currentRow = new Number[numElements];
+
     // Sanity check, to make sure that we see the rowTerminator at the end
     boolean rowTerminatorSeen = false;
 
+    int curElement = 0;
     while (position < limit) {
       final byte kind = memory.getByte(position);
 
@@ -193,12 +180,26 @@ private void updateCurrentArray(final long fieldPosition)
 
       // If terminator not seen, then read the field at that location, and increment the position by the element's field
       // size to read the next element.
-      currentRow.add(getIndividualValueAtMemory(position));
+      currentRow[curElement] = getIndividualValueAtMemory(position);
       position += getIndividualFieldSize();
+      curElement++;
     }
 
-    if (!rowTerminatorSeen) {
+    if (!rowTerminatorSeen || curElement != numElements) {
       throw DruidException.defensive("Unexpected end of field");
     }
   }
+
+  int numElements(long fieldSize)
+  {
+    if (fieldSize <= 1) {
+      throw DruidException.defensive("fieldSize should be greater than 1 for non null array elements");
+    }
+    // Remove one byte for the nullity byte, and one for the array terminator
+    long cumulativeFieldSize = fieldSize - Byte.BYTES - Byte.BYTES;
+    if (cumulativeFieldSize % getIndividualFieldSize() != 0) {
+      throw DruidException.defensive("cumulativeFieldSize should be a multiple of the individual fieldSize");
+    }
+    return Math.toIntExact(cumulativeFieldSize / getIndividualFieldSize());
+  }
 }

processing/src/main/java/org/apache/druid/frame/field/ReadableFieldPointer.java

@@ -31,4 +31,9 @@ public interface ReadableFieldPointer
    * Starting position of the field.
    */
   long position();
+
+  /**
+   * Length of the field.
+   */
+  long length();
 }

processing/src/main/java/org/apache/druid/frame/field/RowMemoryFieldPointer.java

@@ -24,6 +24,10 @@
 
 /**
  * A {@link ReadableFieldPointer} that is derived from a row-based frame.
+ *
+ * Returns the position and the length of a field at a particular position for the row that the rowPointer is pointing
+ * to at the time. It caches the values of the position and the length based on position of the rowPointer.
+ * This method is not thread-safe
  */
 public class RowMemoryFieldPointer implements ReadableFieldPointer
 {
@@ -32,6 +36,16 @@ public class RowMemoryFieldPointer implements ReadableFieldPointer
   private final int fieldNumber;
   private final int fieldCount;
 
+  // Caching of position() calls
+  private long rowWithPositionCached = -1L;
+  private long cachedPosition = -1L;
+
+  // Caching of length() calls
+  // We cache the length() calls separately, because not all field types call length(), therefore it's wasteful to
+  // read length() if we are not reading it
+  private long rowWithLengthCached = -1L;
+  private long cachedLength = -1L;
+
   public RowMemoryFieldPointer(
       final Memory memory,
       final ReadableFrameRowPointer rowPointer,
@@ -47,6 +61,56 @@ public RowMemoryFieldPointer(
 
   @Override
   public long position()
+  {
+    updatePosition();
+    return cachedPosition;
+  }
+
+  @Override
+  public long length()
+  {
+    updatePositionAndLength();
+    return cachedLength;
+  }
+
+  private void updatePosition()
+  {
+    long rowPointerPosition = rowPointer.position();
+    if (rowPointerPosition == rowWithPositionCached) {
+      return;
+    }
+    // Update the cached position for position()
+    rowWithPositionCached = rowPointerPosition;
+
+    // Update the start position
+    cachedPosition = startPosition(fieldNumber);
+  }
+
+  // Not all field types call length(), and therefore there's no need to cache the length of the field
+  private void updatePositionAndLength()
+  {
+    updatePosition();
+
+    long rowPointerPosition = rowPointer.position();
+    if (rowPointerPosition == rowWithLengthCached) {
+      return;
+    }
+    // Update the cached position for length()
+    rowWithLengthCached = rowPointerPosition;
+
+    if (fieldNumber == fieldCount - 1) {
+      // If the field is the last field in the row, then the length of the field would be the length of the row minus the
+      // start position of the row
+      cachedLength = (rowPointerPosition + rowPointer.length()) - cachedPosition;
+    } else {
+      // Else the length of the field would be the difference between the start position of the given field and
+      // the subsequent field
+      cachedLength = startPosition(fieldNumber + 1) - cachedPosition;
+    }
+  }
+
+
+  private long startPosition(int fieldNumber)
   {
     if (fieldNumber == 0) {
       // First field starts after the field end pointers -- one integer per field.

processing/src/main/java/org/apache/druid/frame/field/SettableFieldPointer.java

@@ -20,21 +20,31 @@
 package org.apache.druid.frame.field;
 
 /**
- * A simple {@link ReadableFieldPointer} that returns the position that was set on its object.
+ * A simple {@link ReadableFieldPointer} that returns the position and the length that was set on its object.
  */
 public class SettableFieldPointer implements ReadableFieldPointer
 {
-
   long position = 0;
+  long length = -1;
 
-  public void setPosition(long position)
+  /**
+   * Sets the position and the length to be returned when interface's methods are called.
+   */
+  public void setPositionAndLength(long position, long length)
   {
     this.position = position;
+    this.length = length;
   }
 
   @Override
   public long position()
   {
     return position;
   }
+
+  @Override
+  public long length()
+  {
+    return length;
+  }
 }

processing/src/main/java/org/apache/druid/frame/segment/row/ReadableFrameRowPointer.java

@@ -30,7 +30,13 @@
  */
 public interface ReadableFrameRowPointer
 {
+  /**
+   * Position of the start of the row relative to the start of the Frame
+   */
   long position();
 
+  /**
+   * Length of the row (in bytes)
+   */
   long length();
 }

processing/src/test/java/org/apache/druid/frame/field/ComplexFieldReaderTest.java

@@ -123,7 +123,7 @@ public void test_makeColumnValueSelector_null()
     writeToMemory(null);
 
     final ColumnValueSelector<?> readSelector =
-        new ComplexFieldReader(SERDE).makeColumnValueSelector(memory, new ConstantFieldPointer(MEMORY_POSITION));
+        new ComplexFieldReader(SERDE).makeColumnValueSelector(memory, new ConstantFieldPointer(MEMORY_POSITION, -1));
 
     Assert.assertNull(readSelector.getObject());
   }
@@ -134,7 +134,7 @@ public void test_makeColumnValueSelector_aValue()
     writeToMemory("foo");
 
     final ColumnValueSelector<?> readSelector =
-        new ComplexFieldReader(SERDE).makeColumnValueSelector(memory, new ConstantFieldPointer(MEMORY_POSITION));
+        new ComplexFieldReader(SERDE).makeColumnValueSelector(memory, new ConstantFieldPointer(MEMORY_POSITION, -1));
 
     Assert.assertEquals("foo", readSelector.getObject());
   }

processing/src/main/java/org/apache/druid/frame/field/ConstantFieldPointer.java → processing/src/test/java/org/apache/druid/frame/field/ConstantFieldPointer.java

@@ -22,15 +22,23 @@
 public class ConstantFieldPointer implements ReadableFieldPointer
 {
   private final long position;
+  private final long length;
 
-  public ConstantFieldPointer(long position)
+  public ConstantFieldPointer(long position, long length)
   {
     this.position = position;
+    this.length = length;
   }
 
   @Override
   public long position()
   {
     return position;
   }
+
+  @Override
+  public long length()
+  {
+    return length;
+  }
 }