Well, I finally "fixed" this issue. The only thing I can figure is that the data processing requirements were somehow too great, and it was throwing errors as a result.
What finally fixed the issue was removing some complex calculated columns. I had three of them that were using CALCULATE with ALL and two sets of EARLIER filters each, and once they were removed, the refresh worked just fine.
Like I mentioned, the other thing that was fixing the issue was if my 10,000-row source dataset was smaller, like 9,700 rows. Which seemed strange because at 9 columns, that's not huge.
But it loaded at the smaller size no matter which 300 rows I removed, so it wasn't a matter of irregular data in those extra rows.
What it added up to was that Power BI was choking on the size of the calculations. Those EARLIER filters can't be cheap given the way they have to reference every row in the table for each row that it calculates. So 10,000 rows * 10,000 rows * 3 fields * 2 uses of EARLIER instances in each = 600,000,000 rows to scan every time it loads.
But still, it should be built to handle it, and not error out. Especially with misleading errors.
Edit:
I found I was able to replace my many CALCULATE + EARLIER columns with a single INDEX column that uses two EARLIER clauses, and for everything else, I'm using LOOKUPVALUE and referencing the index to find what I need. And where my data refreshes before would take multiple minutes or error out, they now update in a flash 🙂
