Each time a handle has a checkpoint started, it waits for the eviction generation to become obsolete (i.e. it waits for all active evictions to drain). This happens in wt_sync_file, where it calls wt_gen_next_drain.

We should have a statistic exposing when this is happening - measuring the time spent waiting would be acceptable, but sub-optimal (since it can only report once the drain is complete). Perhaps the checkpoint_state statistic could be utilised in conjunction with a time measurement.