[CASSANDRA-15600] Algorithmic token allocation does not handle the racks = RF case well

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 4.0, 4.0-alpha4
• Component/s: Feature/Virtual Nodes
• Bug Category: Correctness
• Severity: Normal
• Complexity: Normal
• Discovered By: User Report
• Platform: All
• Impacts: Docs
• Since Version: 4.0-alpha

Source Control Link

https://github.com/apache/cassandra/commit/c9c022e07fd4fc6afc59cd1149fd18bc84d3fb39

Test and Documentation Plan

In the cases where the keyspace replication factor is ONE, or if the keyspace replication factor is equal to the number of racks in the datacenter, the workload variance drops from ~10 percent to ~6 percent due to algorithmic optimizations.

 

• Assignee: Ekaterina Dimitrova
• Reporter: Branimir Lambov

Authors

Ekaterina Dimitrova

Reviewers

Brandon Williams

• Created: 25/Feb/20 16:01
• Updated: 21/Dec/20 08:07
• Resolved: 03/Mar/20 16:04

Description

When the number of racks in a data centre is equal to the replication factor, the node allocation algorithm may calculate the target ownership incorrectly and allocate unsuitable tokens. This causes some inefficiency when the number of nodes in the DC is small, and significant problems if the number of nodes in each rack is different.

In this case (racks count equal to replication factor) the load within each rack is effectively distributed independently. The allocation algorithm should reflect that, restricting the allocation space to the rack (instead of the DC) and using the RF=1 solution.