[SPARK-25091] UNCACHE TABLE, CLEAR CACHE, rdd.unpersist() does not clean up executor memory

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Duplicate
• Affects Version/s: 2.3.1
• Component/s: SQL
• Assignee: Unassigned
• Reporter: Yunling Cai
• Created: 11/Aug/18 17:50
• Updated: 10/Sep/18 17:48
• Resolved: 10/Sep/18 17:48

Description

UNCACHE TABLE and CLEAR CACHE does not clean up executor memory.

Through Spark UI, although in Storage, we see the cached table removed. In Executor, the executors continue to hold the RDD and the memory is not cleared. This results in huge waste in executor memory usage. As we call CACHE TABLE, we run into issues where the cached tables are spilled to disk instead of reclaiming the memory storage. 

Steps to reproduce:

CACHE TABLE test.test_cache;

UNCACHE TABLE test.test_cache;

== Storage shows table is not cached; Executor shows the executor storage memory does not change == 

CACHE TABLE test.test_cache;

CLEAR CACHE;

== Storage shows table is not cached; Executor shows the executor storage memory does not change == 

Similar behavior when using pyspark df.unpersist().

Issue Links

• duplicates: SPARK-24889 dataset.unpersist() doesn't update storage memory stats
• links to: [Github] Pull Request #22335 (cfangplus)