UPDATE:
I think we have identefied and solved the issue. It seems like using LAST with Databricks SQL requires to excplicitly be careful about setting the "ignoreNull" argument and also be careful about the correct datatype. I guess this is because of Databricks SQL using ANSI. Using LAST in notebook, all of this is taking care of by spark under the hood. 

With this in mind, we might just use notebooks instead of Databricks SQL solely as our got-to query UI, even for simple queries, so that we are 100% sure that we get correct results. @Retired_mod