[SPARK-19500] [SQL] Fix off-by-one bug in BytesToBytesMap

## What changes were proposed in this pull request?

Radix sort require that half of array as free (as temporary space), so we use 0.5 as the scale factor to make sure that BytesToBytesMap will not have more items than 1/2 of capacity. Turned out this is not true, the current implementation of append() could leave 1 more item than the threshold (1/2 of capacity) in the array, which break the requirement of radix sort (fail the assert in 2.2, or fail to insert into InMemorySorter in 2.1).

This PR fix the off-by-one bug in BytesToBytesMap.

This PR also fix a bug that the array will never grow if it fail to grow once (stay as initial capacity), introduced by #15722 .

## How was this patch tested?

Added regression test.

Author: Davies Liu <[email protected]>

Closes #16844 from davies/off_by_one.

Author: Davies Liu

core/src/main/java/org/apache/spark/unsafe/map/BytesToBytesMap.java

@@ -698,7 +698,7 @@ public boolean append(Object kbase, long koff, int klen, Object vbase, long voff
       if (numKeys == MAX_CAPACITY
         // The map could be reused from last spill (because of no enough memory to grow),
         // then we don't try to grow again if hit the `growthThreshold`.
-        || !canGrowArray && numKeys > growthThreshold) {
+        || !canGrowArray && numKeys >= growthThreshold) {
         return false;
       }
 
@@ -742,7 +742,7 @@ public boolean append(Object kbase, long koff, int klen, Object vbase, long voff
         longArray.set(pos * 2 + 1, keyHashcode);
         isDefined = true;
 
-        if (numKeys > growthThreshold && longArray.size() < MAX_CAPACITY) {
+        if (numKeys >= growthThreshold && longArray.size() < MAX_CAPACITY) {
           try {
             growAndRehash();
           } catch (OutOfMemoryError oom) {
@@ -911,6 +911,7 @@ public void reset() {
       freePage(dataPage);
     }
     allocate(initialCapacity);
+    canGrowArray = true;
     currentPage = null;
     pageCursor = 0;
   }

sql/core/src/test/scala/org/apache/spark/sql/execution/UnsafeFixedWidthAggregationMapSuite.scala

@@ -342,4 +342,44 @@ class UnsafeFixedWidthAggregationMapSuite
     }
   }
 
+  testWithMemoryLeakDetection("convert to external sorter after fail to grow (SPARK-19500)") {
+    val pageSize = 4096000
+    val map = new UnsafeFixedWidthAggregationMap(
+      emptyAggregationBuffer,
+      aggBufferSchema,
+      groupKeySchema,
+      taskMemoryManager,
+      128, // initial capacity
+      pageSize,
+      false // disable perf metrics
+    )
+
+    val rand = new Random(42)
+    for (i <- 1 to 63) {
+      val str = rand.nextString(1024)
+      val buf = map.getAggregationBuffer(InternalRow(UTF8String.fromString(str)))
+      buf.setInt(0, str.length)
+    }
+    // Simulate running out of space
+    memoryManager.limit(0)
+    var str = rand.nextString(1024)
+    var buf = map.getAggregationBuffer(InternalRow(UTF8String.fromString(str)))
+    assert(buf != null)
+    str = rand.nextString(1024)
+    buf = map.getAggregationBuffer(InternalRow(UTF8String.fromString(str)))
+    assert(buf == null)
+
+    // Convert the map into a sorter. This used to fail before the fix for SPARK-10474
+    // because we would try to acquire space for the in-memory sorter pointer array before
+    // actually releasing the pages despite having spilled all of them.
+    var sorter: UnsafeKVExternalSorter = null
+    try {
+      sorter = map.destructAndCreateExternalSorter()
+      map.free()
+    } finally {
+      if (sorter != null) {
+        sorter.cleanupResources()
+      }
+    }
+  }
 }