@Håkon Åmdal​ to be clear, I think what you're suggesting is in fact the only approach that might work. i.e. an upsert for each data source, sequentially.

This being the only way seems to be a spark limitation; as the answer at apache spark - Structured Streaming extract most recent values for each id - Stack Overflow discusses. There are multiple discussed solutions for a joint-approach online but none seemed to work for the answerer there, nor do they for me, meaning it needs to be done sequentially instead of jointly.

I may still switch to that approach, but before I do I'd like to understand better why the databricks engineer I talked to recommended against it and what the alternatives are. The one the engineer suggested seemed to suggest going to a more noSQL-like style of tables and processing updates via aggregate functions on that result; I'm still working through understanding that.