Elasticsearch: _source not returned in "explain": true with certain queries

[nerophon]

Elasticsearch version: 5.9.6 & 6.2.4

Plugins installed: []

JVM version: independent

OS version: independent

Description of the problem including expected versus actual behavior:
The _source is not being returned for hits when explain is set to true for a specific set of queries. The conditions for _source to be missing appear to be:

• A mapping with a nested field
• A document with that nested field populated
• A query which:
  
  
  
  • Matches that document
  
  
  • Has explain set to true
  
  
  • Uses rescore, with a nested function_score query matching the nested field
  
  
  • Adjusts the score using script_score
  
  

Steps to reproduce:

PUT /test
{
  "settings": {
    "number_of_shards": 1,
    "number_of_replicas": 1
  },
  "mappings": {
    "type1": {
      "properties": {
        "object1": {
          "type": "nested",
          "properties": {
            "field1": {
              "type": "text"
            }
          }
        }
      }
    }
  }
}

PUT /test/type1/1
{
  "object1": {
    "field1": "some text goes here"
  }
}



md5-484cc6ab1aeac3a136681ca80dcfc99b



GET /test/_search
{
  "from" : 0,
  "size" : 10,
  "query" : {
    "match_all": { }
  },
  "explain" : true,
  "rescore" : [
    {
      "window_size" : 200,
      "query" : {
        "rescore_query" : {
          "nested" : {
            "query" : {
              "function_score" : {
                "query" : {
                  "match_all": { }
                },
                "functions" : [
                  {
                    "script_score" : {
                      "script" : {
                        "source" : "1",
                        "lang" : "painless"
                      }
                    }
                  }
                ],
                "score_mode" : "multiply"
              }
            },
            "path" : "object1"
          }
        },
        "query_weight" : 1.0,
        "rescore_query_weight" : 1.0,
        "score_mode" : "total"
      }
    }
  ]
}

[elasticmachine]

Pinging @elastic/es-search-aggs

[javanna]

I finally got to doing some digging on this one. Indeed, it is a very peculiar bug, and I am not sure yet how to solve it, but I can now explain how it happens: this only happens when explaining a script score function that's part of a nested query inside a rescore section.

When trying to explain the score function, we run the script which ends up calling LeafSearchLookup#setDocument which calls SourceLookup#setSegmentAndDocument that ends up zeroing the source that was previously loaded, as we are going through a different docId compared to the previous run (nested doc rather than parent doc).

[fhaase2]

@javanna do you have a suggestion how to make it still working? Is there some hack?

[jtibshirani]

@fhaase2 unfortunately I can't think of a workaround that doesn't change meaning of the request. If you could rework the request to remove the 'script score' part, that would avoid this issue.

I just opened a refactor #60179 that would fix this issue properly.

[fhaase2]

@jtibshirani great, thanks for the answer.

As a workaround i thought about making a subsequent request after the explain request, just to retrieve the source of the documents using the Ids query. But this would be a hack, so ill better wait for the refactor #60179