[CASSANDRA-15834] Bloom filter false positive rate calculation does not take into account true negatives

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.0.25, 3.11.11, 4.0-rc2, 4.0
• Component/s: Observability/JMX
• Bug Category: Correctness
• Severity: Normal
• Complexity: Normal
• Discovered By: Code Inspection
• Platform: All
• Since Version: 3.0.0

Source Control Link

https://github.com/apache/cassandra/commit/4e5bd572f4cca2831b87d6aa1b322e8ad846d015

Test and Documentation Plan

https://github.com/apache/cassandra-dtest/pull/70
https://github.com/apache/cassandra/pull/600

• Assignee: Jaroslaw Grabowski
• Reporter: Jaroslaw Grabowski

Authors

Jaroslaw Grabowski

Reviewers

Brandon Williams

• Created: 26/May/20 10:31
• Updated: 16/Mar/22 13:32
• Resolved: 18/May/21 19:26

Description

The bloom filter false positive ratio is currently computed as:

bf_fp_ratio = false_positive_count / (false_positive_count + true_positive_count)

However, this calculation doesn't take into account true negatives (false negatives never happen on bloom filters).

In a situation where there are 1000 reads for non existing rows, and there are 10 false positives, the bloom filter false positive ratio will be wrongly calculated as 10/10 = 1.0, while it should be 10/1000 = 0.01.

We should update the calculation to:

bf_fp_ratio = false_positive_count / #bf_queries

Original jira by pauloricardomg