Testing has uncovered a situation where a checkpoint can discard history required for snapshot readers. The problematic mode of eviction helps avoid stalls due to checkpoints filling the cache with dirty data during replication lag.
Disable this eviction for now when history can be lost, to make sure snapshot transactions get correct results. We can re-enable it and fix history retention in a future change (see
WT-4961 for the long-term fix).
- is related to
WT-4961 Checkpoints with cache overflow must keep history for reads