query_string and simple_query_string cannot apply minimum should match consistently

[jimczi]

We apply the requested minimum_should_match on parsed queries produced by query_string and simple_query_string. Since these queries can target multiple fields we used to rely on the disableCoord property of the Boolean query. Now that coords are gone in Lucene 7 we are not able to distinguish queries with multiple positions like:

"query": {
	"simple_query_string": {
		"fields": ["field1", "field2"],
		"query": "foo bar",
		"minimum_should_match": 1
	}
}

.... with single position queries on multiple fields:

"query": {
	"simple_query_string": {
		"fields": ["field1", "field2", "field3],
		"query": "foo",
		"minimum_should_match": 2
	}
}

In 5.x we were able to detect that the later query produces a single position and that minimum_should_matchshould not be applied.

In 6.0 this query would be rewritten in:
foo must appear in 2 out of the 3 provided fields

This is problematic but I can't think of a good fix for this.
Furthermore applying minimum_should_match on query_string and simple_query_string is difficult to understand for users (it depends on settings like split_on_whitespace, tokenization, ...) so I wonder if we should just remove the support for this option in these query parsers.
The match and multi_match_query are not impacted since we apply minimum_should_match per analyzer group.

[s1monw]

do we have access to the place where we create the boolean queries so we can maybe wrap them or somehow mark them as eligible for minimum must match? Anyhow I removed the discuss label since we need to fix it either way ....

[jimczi]

I think we can just replace the boolean queries with DisjunctionMaxQuery. We already have an option for that in query_string called use_dis_max. When set to false (the default) we create boolean queries to merge queries targeting different fields and that makes minimum_should_match harder to apply. Instead we could still use a DisjunctionMaxQuery but with the tiebreaker set to 1.
The scoring should be equivalent to the boolean query and this would differentiate cross field queries from position queries and solve this particular problem.
We could also consider adding use_dismax in simple_query_string but this is not related to this issue.