DiskBlockManager can't clear all of the tmp dirs in k8s

I have a test for spark on k8s  that my job had completed, but some tmp dirs are still remained. Those dirs are start with "blockmgr". My "spark.local.dirs" is mount on local NVMe SSD.

The DiskBlockManager may be clear the dirs  just one iteration ， then the process be killed.

The shuffle data will be getting larger and   fill the disk space.