We've seen cases where the default file handle (aka dhandle, aka data handle) parameters negatively impact performance and the workload has a large number of active tables.

SERVER-24949 lowered the default file handle timeout to 10 minutes (from 28 hours) and SERVER-56661 raised the default limit to 2000 handles (from 200).

Having too many open file handles concurrently can consume too much memory, but lowering the limits too far can negatively impact latencies due to WT schema lock contention. It's hard to choose one-size-fits-all policy, so we may want to consider tuning this dynamically based on feedback from the workload.

One note: in the cases we've seen, performance is most negatively affected when the workload involves tens of thousands of active handles (i.e. collections + indexes), which is a data model we discourage. However, our recommendations often involve raising one of these limits to regain performance.