diff --git a/core/src/main/scala/org/apache/spark/shuffle/sort/SortShuffleManager.scala b/core/src/main/scala/org/apache/spark/shuffle/sort/SortShuffleManager.scala
index 66b6bbc61fe8..ad26eef50830 100644
--- a/core/src/main/scala/org/apache/spark/shuffle/sort/SortShuffleManager.scala
+++ b/core/src/main/scala/org/apache/spark/shuffle/sort/SortShuffleManager.scala
@@ -88,7 +88,8 @@ private[spark] class SortShuffleManager(conf: SparkConf) extends ShuffleManager
       shuffleId: Int,
       numMaps: Int,
       dependency: ShuffleDependency[K, V, C]): ShuffleHandle = {
-    if (SortShuffleWriter.shouldBypassMergeSort(SparkEnv.get.conf, dependency)) {
+    if (SparkEnv.get != null &&
+      SortShuffleWriter.shouldBypassMergeSort(SparkEnv.get.conf, dependency)) {
       // If there are fewer than spark.shuffle.sort.bypassMergeThreshold partitions and we don't
       // need map-side aggregation, then write numPartitions files directly and just concatenate
       // them at the end. This avoids doing serialization and deserialization twice to merge