[SPARK-41339] RocksDB state store WriteBatch doesn't clean up native memory

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.3.1
• Fix Version/s: 3.3.2, 3.4.0
• Component/s: SQL, Structured Streaming
• Assignee: Adam Binford
• Reporter: Adam Binford
• Created: 30/Nov/22 16:40
• Updated: 01/Dec/22 06:50
• Resolved: 01/Dec/22 06:50

Description

The RocksDB state store uses a WriteBatch to hold updates that get written in a single transaction to commit. Somewhat indirectly abort is called after a successful task which calls writeBatch.clear(), but the data for a writeBatch is stored in a std::string in the native code. Not sure why it's stored as a string, but it is. rocksdb/write_batch.h at main · facebook/rocksdb · GitHub

writeBatch.clear simply calls rep_.clear() and rep._resize() (rocksdb/write_batch.cc at main · facebook/rocksdb · GitHub), neither of which actually releases the memory built up by a std::string instance. The only way to actually release this memory is to delete the WriteBatch object itself.

Currently, all memory taken by all write batches will remain until the RocksDB state store instance is closed, which never happens during the normal course of operation as all partitions remain loaded on an executor after a task completes.

Issue Links

• links to: [Github] Pull Request #38853 (Kimahriman)