function_score should throw exception if decay function is used with multi valued field

[brwe]

The decay functions for function_score do not handle fields with multiple values. Decay functions silently compute the distance from the first value. This can lead to confusion, as seen in issue #3926. Instead of using the first value in the field, an exception should be thrown if a field has multiple values.

[bobrik]

@brwe I'm wondering if "flattering" before applying function_score could work too. Like in #3926 each function could be applied to each value and score_mode could be used to calculate result from NxM values where N is functions count and M is values count.

[s1monw]

we have a bunch of places where we do similar things like in Sorting we use org.elasticsearch.index.fielddata.fieldcomparator.SortMode where users can simply pick if they want the min/max/sum/average of a multivalued field.

[brwe]

I see. Might make sense.
Do you think we should also take care of the following use case?

I have a multi valued field with N entries, but it actually represents an N-dim vector. Instead of computing avg, min or max distance, I actually want to give the origin as vector as well and then compute euclidean distance of origin and field vector.

In this case the number of values in the field should be the same as the number of values given in the origin parameter.

[brwe]

Closed this issue. Feel free to open a new issue if you think multiple values should be handled as described above.

[jpountz]

@brwe Unfortunately, the isMultiValued method might return true even if the field is single-valued. The reason is that our doc-values-based impls can't know whether they are single-valued. (This reminds me that we should randomly use doc values instead of uninverted field data in our tests /cc @javanna @s1monw)

[brwe]

Thanks! I reverted the commits for now.

[s1monw]

@brwe I pushed a commit on how we could solve this - just as an example to start the discussion

[brwe]

@s1monw That would work I think.

[brwe]

Opened pull request #5940

[brwe]

Closing, the issue is fixed by #5940