In DSC & ASC we will evict clean pages with updates with the goal of lowering the updates quantity in cache, i.e. servicing the updates target / trigger.

This can create an undesirable behaviour where those pages will be immediately read back in, and the reading threads will need to wait some time for that to happen.

An analysis performed on the ycsb128thread-load workload suggests that threads are spending ~15 seconds per second waiting for these reads to complete. It is unclear if all that time is as a result of update eviction.

Scope:

• Create a new version of scrub eviction that can replace the page in place with it's disk image removing the updates as it does so
• Test and see FTDC implications
• If the change is promising, merge.