I'm looking into some strange results from a "more like this" query involving search_type.
I switched to dfs_query_then_fetch from the default query_then_fetch because I wanted more accurate results and the overhead didn't seem to make much difference.
However I'm finding the results with the standard query_then_fetch appear to be better, the ones that come back from dfs_query_then_fetch look less relevant (to the human eye anyway).
So I'm trying to understand the difference, have checked the documentation and it says
Distributed term frequencies are calculated locally for each shard running the search.
I'm not sure what locally means in this sense, and what shard runs the search?
For context my more like this query runs against ten indices, some have >10k documents, some much smaller with only 100 documents.