We have seen a few statistics logs from WiredTiger users that indicate our eviction queueing algorithm is inefficient. The particular metrics that point to the problem tend to be that pages seen by eviction walk is an order of magnitude higher than pages queued for eviction, while at the same time files with new eviction walks started is surprisingly high.

Pages generally aren't queued either because they are dirty and the cache only needs to make clean space available or they are internal pages. The workloads I've seen don't have a lot of dirty content, so it's likely (possible at least) that the eviction server is seeing many internal pages.

The other shared characteristic of the workloads is that there are many tables. Usually in excess of 1000 active tables, with data size and activity distributed unevenly amongst the tables.