We don't have much data on this one, or reproduction steps. However, we're looking for insight on where to look to debug. Maybe you also saw something similar before.

We had an issue with one of our primary nodes where a background index build caused mongod to run out of connections. In the log we saw lines like "refusing connection because we already have 20.000 connections". We couldn't even connect to mongod via a shell, which made it much harder to debug (feature request: make sure we can access mongod even when max number of connections is reached. maybe always allow an one extra connection from localhost?).

Here's an interesting part from the mongodb.log: https://gist.github.com/igorcanadi/9182e6f1a49af989642f

It looks like bunch of operations spent a lot of times waiting for an exclusive write lock. Is this expected?

Any insight would be helpful.