We have code that maintains a write generation on each page. Every time an update is added to a page the write generation is incremented. When the page is marked clean by reconciliation the write generation is reset to 0.
It is possible to construct a workload where checkpoints don't mark the page clean (there is always an update in front of the checkpoint), and the page is never evicted from cache either (never split). A case where this might be the case is a single document is being updated very regularly.
We should fix the write generation code to be resilient to the write generation wrapping, or implement a clean/dirty tracking mechanism that isn't susceptible to wrapping.