[CASSANDRA-17877] Speculative execution threshold unit mismatch

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 4.1-beta1, 5.0-alpha1, 5.0
• Component/s: Consistency/Coordination
• Bug Category: Degradation - Resource Management
• Severity: Normal
• Complexity: Normal
• Discovered By: User Report
• Platform: All
• Since Version: 4.1-alpha1

Source Control Link

https://github.com/apache/cassandra/commit/6748b8b7ead6fff7820045ebbfe23f6c050f2efe

Test and Documentation Plan

Run existing tests, added an updated check for the speculation thresholds being in the expected range.

• Assignee: Jon Meredith
• Reporter: Jon Meredith

Authors

Jon Meredith

Reviewers

Caleb Rackliffe, Yifan Cai, Yifan Cai

• Created: 02/Sep/22 14:49
• Updated: 12/Sep/23 13:02
• Resolved: 06/Sep/22 16:10

Description

CASSANDRA-16760 changed the ColumnFamilyStore.coordinatorRead/WriteLatency to be in microseconds rather than nanoseconds but did not change the calculation for ColumnFamilyStore.sampleReadLatencyNanos/additionalWriteLatency which is used as the threshold for speculative read execution / issuing additional writes.

The impact of this bug is that the threshold is 1000x smaller than intended when using percentile based speculation (tables default to 99p), so effectively speculation happens for ALL reads configured to speculate on a percentile, including min hybrid policies that use percentiles.