Hi!
Thank you very much for your response @MaBeuLux88_xxx !
Your Python example looks functionally identical to what I made, kudos for getting the intent from Scala code 
The tests I had made (four tests on two different days) were always within the same AWS region, Oregon in our case. Today I tested a setup similar to yours, setting my M10 Atlas cluster to US east coast, with the application remaining on west coast. No write concerns happened and failover test was successful with the same code.
Amount of increments performed during this timeframe when within the same Oregon region was around 120000. When done from west coast to east coast, 6000 were made. This leads me to believe the reason why the same region test results in the write concern error is related to the latency. A remote connection spends more time “on the way” than doing operations on Mongo side and the problematic period is probably really short.
Does this sound sensible to you?
My main concern is ensuring that our application can handle Atlas maintenance events without running into non-retryable errors. We can handle using only supported operations, but a write concern error seems difficult to cope with.
Valtteri