The recent eviction and checkpoint enhancements have led to a performance degradation in a particular YCSB workload. We need to investigate improving the performance for that workload.
The workload in question is a 50% read 50% update workload. The workload fits entirely in cache, and no eviction is needed in the steady state. When a checkpoint starts, the checkpoint scrub phase causes application threads to become involved in eviction which leads to a drop in throughput until the checkpoint dirty target is reached.