[CASSANDRA-16881] L0 compactions never recover when thousands unleveled

• Type: Improvement
• Status: Resolved
• Priority: Normal
• Resolution: Duplicate
• Component/s: Local/Compaction/STCS
• Change Category: Operability
• Complexity: Normal
• Platform: All
• Assignee: Josh McKenzie
• Reporter: Josh McKenzie

Authors

Josh McKenzie

• Created: 24/Aug/21 16:11
• Updated: 24/Aug/21 18:34
• Resolved: 24/Aug/21 18:34

Description

While this is somewhat mitigated by incremental repair and CASSANDRA-14388, we can still find ourselves in a scenario where we have thousands of small sstables living in L0 on STCS and get into a heavily degraded state.

We should add a JMX flag to allow operators to flip the behaviour when doing STCS in L0 from picking the ‘hottest’ bucket to picking the largest one - this should allow us to get down to reasonable numbers of sstables in L0 quicker in this pathological case. If this proves to work particularly well we could consider making this the default in a future version. 

Issue Links

• duplicates: CASSANDRA-16782 Improve the way we pick sstables for STCS-in-L0 and in TWCS 'current' window