[CASSANDRA-11933] Cache local ranges when calculating repair neighbors

• Type: Improvement
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 2.1.15, 2.2.7, 3.0.8, 3.8
• Component/s: Legacy/Core
• Assignee: Mahdi Mohammadinasab
• Reporter: Cyril Scetbon

Authors

Mahdi Mohammadinasab

Reviewers

Paulo Motta

• Created: 31/May/16 23:08
• Updated: 16/Apr/19 09:30
• Resolved: 15/Jun/16 09:50

Description

During a full repair on a ~ 60 nodes cluster, I've been able to see that this stage can be significant (up to 60 percent of the whole time) :

https://github.com/apache/cassandra/blob/cassandra-2.1/src/java/org/apache/cassandra/service/StorageService.java#L2983-L2997

It's merely caused by the fact that https://github.com/apache/cassandra/blob/cassandra-2.1/src/java/org/apache/cassandra/service/ActiveRepairService.java#L189 calls

ss.getLocalRanges(keyspaceName)

everytime and that it takes more than 99% of the time. This call takes 600ms when there is no load on the cluster and more if there is. So for 10k ranges, you can imagine that it takes at least 1.5 hours just to compute ranges.

Underneath it calls ReplicationStrategy.getAddressRanges which can get pretty inefficient (jbellis's words)

ss.getLocalRanges(keyspaceName) should be cached to avoid having to spend hours on it.

Issue Links

• relates to: CASSANDRA-3912 repair user provided custom token range (support incremental repair controlled by external agent)