Summary

We lack repeatable measurements showing how Atlas Infinite planned failovers perform against availability targets under sustained high read/write load. This leaves a gap in our understanding of the client-visible availability experience as load increases.

Current behaviour

HELP-101625 reports majority-write availability gaps of 4.9–6.3 seconds during vertical scaling at approximately 90% CPU utilization, exceeding the reported 3-second write-availability target.

These results cover one workload. We do not have a repeatable Kudzu workload that measures planned-failover availability across load levels and compares the results against the applicable read/write availability targets.

Proposed solution

Create configurable workloads in Kudzu that maintain read/write load across planned failovers, using the attached scalebench-workload-20261012.zip as a starting point.

Measure primary and secondary read availability and w:1 and w:majority write availability with non-retrying probes. Run repeated measurements at different load levels, including sustained high load, and compare the observed interruption durations against the applicable targets.

Record the workload configuration, achieved load, client measurements, logs, and FTDC for each run.

Definition of Done

• Reusable Kudzu workloads sustain configurable read/write load across planned failovers.
• Repeated runs characterize client-visible availability at different load levels, including high load.
• Results compare measured availability against the applicable targets and identify where gaps remain.
• Workload configuration, run instructions, and measurement artifacts are available for repeatable testing.