[FLINK-15512] Refactor the mechanism of how to constructure the cache and write buffer manager shared across RocksDB instances

• Type: Sub-task
• Status: Closed
• Priority: Blocker
• Resolution: Implemented
• Fix Version/s: 1.10.0
• Component/s: Runtime / State Backends
• Labels: pull-request-available
• Assignee: Yun Tang
• Reporter: Yun Tang
• Created: 08/Jan/20 07:32
• Updated: 19/Oct/20 02:49
• Resolved: 13/Jan/20 07:39

Description

FLINK-14484 introduce a LRUCache to share among RocksDB instances, so that the memory usage by RocksDB could be controlled well. However, due to the implementation and some bugs in RocksDB (issue-6247), we cannot limit the memory strictly.

The way to walk around this issue is to consider the buffer which memtable would overuse (1/2 write buffer manager size). By introducing this, the actual cache size for user to share is not the same as the managed off-heap memory or user configured memory.

Issue Links

• relates to: FLINK-15368 Add end-to-end test for controlling RocksDB memory usage
• relates to: FLINK-15532 Enable strict capacity limit for memory usage for RocksDB
• links to: GitHub Pull Request #10820