Shuffle data not always be cleaned

In our cluster, when we run a spark streaming job, after running for many hours, the shuffle data seems not all be cleaned, here is the shuffle data:
rw-r---- 1 tdwadmin users 23948 Sep 17 13:21 shuffle_132_34_0
rw-r---- 1 tdwadmin users 18237 Sep 17 13:32 shuffle_143_22_1
rw-r---- 1 tdwadmin users 22934 Sep 17 13:35 shuffle_146_15_0
rw-r---- 1 tdwadmin users 27666 Sep 17 13:35 shuffle_146_36_1
rw-r---- 1 tdwadmin users 12864 Sep 17 14:05 shuffle_176_12_0
rw-r---- 1 tdwadmin users 22115 Sep 17 14:05 shuffle_176_33_1
rw-r---- 1 tdwadmin users 15666 Sep 17 14:21 shuffle_192_0_1
rw-r---- 1 tdwadmin users 13916 Sep 17 14:38 shuffle_209_53_0
rw-r---- 1 tdwadmin users 20031 Sep 17 14:41 shuffle_212_26_0
rw-r---- 1 tdwadmin users 15158 Sep 17 14:41 shuffle_212_47_1
rw-r---- 1 tdwadmin users 42880 Sep 17 12:12 shuffle_63_1_1
rw-r---- 1 tdwadmin users 32030 Sep 17 12:14 shuffle_65_40_0
rw-r---- 1 tdwadmin users 34477 Sep 17 12:33 shuffle_84_2_1

The shuffle data of stage 63, 65, 84, 132... are not cleaned.

In ContextCleaner, it maintains a weak reference for each RDD, ShuffleDependency, and Broadcast of interest, to be processed when the associated object goes out of scope of the application. Actual cleanup is performed in a separate daemon thread.

There must be some reference for ShuffleDependency , and it's hard to find out.