Thanks Lee. My reading of the page you linked to shows it as a performance feature - which is nice and often downright essential.
I'd like to make a case for a choice to override this on this an index-by-index granularity. The thing is, while I generate most of the Visualizations, some of the consumers (who are often not savvy) also create their own visualizations - and they see this unexpected behaviour and are put off by the tool as unreliable. Having an index level option would let me ensure that they see what they expect.
EDIT:
Rerunning the dataset with 1 shard seems to make the bug go away (I used 4 shards until now) - so it seems likely that what you said is indeed the reason for the bug.
But I assume I am leaving a lot of performance on the table since my system has 4 threads available and a shard would use only 1? Would increasing copies help?