Yesterday, we had multiple slowdown on our primary server:

From our metrics, we can see that the cache evictions threads were not able to evict cache fast enough until the cache used reached 95% and that applications threads started to help the dedicated threads (source: http://source.wiredtiger.com/mongodb-3.4/tune_cache.html):

We already encountered this issue with mongo 3.2 two years ago (on the secondary server though) and the only thing that fixed the issue was to upgrade to mongo 3.4: https://jira.mongodb.org/browse/SERVER-27700

We plan to upgrade to mongo 3.6 in the following months but can you check that a regression has not be introduced in our current mongo version ?

Another related ticket: https://jira.mongodb.org/browse/SERVER-27753

I can provide you diagnostic.data and slow log files. Thx for you help.