Trouble with has_parent query containing scripted function_score

Question

I have two document types, in a parent-child relationship:

"myParent" : {
  "properties" : {
    "weight" : {
      "type" : "double"
    }
  }
}

"myChild" : {
  "_parent" : {
    "type" : "myParent"
  },
  "_routing" : {
    "required" : true
  }
}

The weight field is to be used for custom scoring/sorting. This query directly against the parent documents works as intended:

{
  "query" : {
    "function_score" : {
      "script_score" : {
        "script" : "_score * doc['weight'].value"
      }                 
    }                                                                       
  }    
}

However, when trying to do similar scoring for the child documents with a has_parent query, I get an error:

{
  "query" : {
    "has_parent" : {
      "query" : {
        "function_score" : {                                                    
          "script_score" : {
            "script" : "_score * doc['weight'].value"
          }
        }
      },
      "parent_type" : "myParent",
      "score_type" : "score"
    }
  }
}

The error is:

QueryPhaseExecutionException[[myIndex][3]: query[filtered(ParentQuery[myParent](filtered(function score (ConstantScore(:),function=script[_score * doc['weight'].value], params [null]))->cache(_type:myParent)))->cache(_type:myChild)],from[0],size[10]: Query Failed [failed to execute context rewrite]]; nested: ElasticSearchIllegalArgumentException[No field found for [weight] in mapping with types [myChild]];

It seems like instead of applying the scoring function to the parent and then passing its result to the child, ES is trying to apply the scoring function itself to the child, causing the error.

If I don't use score for score_type, the error doesn't occur, although the results scores are then all 1.0, as documented.

What am I missing here? How can I query these child documents with custom scoring based on a parent field?

Answer 1

This I would say is a bug: it is using the myChild mapping as the default context, even though you are inside a has_parent query. But I'm not sure how easy the bug would be to fix. properly.

However, you can work around it by including the type name in the full field name:

curl -XGET "http://localhost:9200/t/myChild/_search" -d'
{
  "query": {
    "has_parent": {
      "query": {
        "function_score": {
          "script_score": {
            "script": "_score * doc[\"myParent.weight\"].value"
          }
        }
      },
      "parent_type": "myParent",
      "score_type": "score"
    }
  }
}'

I've opened an issue to see if we can get this fixed #4914

Answer 2

I think the problem is that you are trying to score child documents based on a field in the parent document and that the function score should really be the other way round.

To solve the problem my idea would be to store the parent/child relation and the score with the child documents. Then you would filter for child documents and score them according to the weight in the child document.

An example:

"myParent" : {
    "properties" : {
        "name" : {
            "type" : "string"
        }
    }
}

"myChild" : {
    "_parent" : {
        "type" : "myParent"
    },
    "_routing" : {
        "required" : true
    },
    "properties": {
        "weight" : {
            "type" : "double"
        }
    }
}

Now you could use a has_parent filter to select all child documents that have a certain parent and then score them using the function score:

{
    "query": {
        "filtered": {
            "query": {
                "function_score" : {
                    "script_score" : {
                        "script" : "_score * doc['weight'].value"
                    }
                }
            },
            "filter": {
                "has_parent": {
                    "parent_type": "myParent",
                    "query": {
                        "term": {
                            "name": "something"
                        }
                    }
                }
            }
        }
    }
}

So if parent documents were blog posts and child comments, then you could filter all posts and score the comments based on weight. I doubt that scoring childs based on parents is possible though I might be wrong :)

Disclaimer: 1st post to stack overflow...