We have oscillated in reconciliation between doing update-restore reconciliations, not choosing to reconcile pages where we can't save much space, and using the lookaside file instead of the reconciliation.

I believe we oscillate for several reasons:

• We don't run workloads that stress the aspects of reconciliation, to allow us to make decisions empirically.
• We don't keep and share enough context between eviction and reconciliation to make the right decision in the right scenario.
• Many of the cases that lead to tweaking the behavior are generated by test/format, and aren't representative of end-user workloads. i.e: there is a tension between making corner cases work, and making the best choice in more common scenarios.

The particular work here starts with creating and running several workloads that exercise update/restore and/or lookaside eviction in a controlled way. Then we can use those workloads (or that workload) to make informed optimizations.