For high update rate workloads, it is possible (common in some benchmarks like YCSB) for the checkpoint to cause cache utilization patterns that lead to application operation throughput stalls.

The reason that happens is that checkpoint consumes I/O that is used for reclaiming dirty space in cache, and also because there are also restrictions on what dirty content can be evicted while a checkpoint is being created. Those restrictions are greater when enabling precise checkpoints to have checkpoint include exactly a point-in-time worth of data.

We should experiment with dynamically adjusting dirty and updates cache thresholds while checkpoints are being created. This should only be configured when precise checkpoints are enabled.

Care will need to be taken when reducing thresholds once checkpoint has completed, to avoid extended stalls for application operations as eviction works to reduce utilization below new (lower) thresholds.