[BUG] Missing inner hits in top hits of an aggregation results since upgrade to 2.13.0

[martijnbolhuis]

Describe the bug

I have query on a nested field names.full_name. I have enabled inner hits on this query. Furthermore, I have added an aggregation on the (non nested) field list_id and I'm using the top hits function to include results per bucket of the aggregation.

In OpenSearch version 2.12.0, the top hits included the inner hits (on names.full_name) but in version 2.13.0 these inner hits are missing.

Related component

Other

To Reproduce

The following script reproduces the problem:

# Create an index and mapping
curl -X DELETE "http://localhost:9200/names-test?pretty"
curl -X PUT "http://localhost:9200/names-test?pretty"  -H 'Content-Type: application/json' -d'
{
  "mappings": {
    "properties": {
      "list_id": {
        "type": "integer"
      },
      "names": {
        "type": "nested",
        "properties": {
          "full_name": {
            "type": "text"
          }
        }
      }
    }
  }
}
'

# Insert documents into the index
curl -X PUT "http://localhost:9200/names-test/_doc/1?refresh&pretty"  -H 'Content-Type: application/json' -d'
{
  "list_id": 1,
  "names": [
    {
      "full_name": "John Doe"
    },
    {
      "full_name": "John Micheal Doe"
    }
  ]
}
'

curl -X PUT "http://localhost:9200/names-test/_doc/2?refresh&pretty"  -H 'Content-Type: application/json' -d'
{
  "list_id": 2,
  "names": [
    {
      "full_name": "Jane Doe"
    },
    {
      "full_name": "Jane Michelle Doe"
    }
  ]
}
'

# Perform a query
curl -X POST "http://localhost:9200/names-test/_search?pretty"  -H 'Content-Type: application/json' -d'
{
  "query": {
    "nested": {
      "path": "names",
      "query": {
        "match": { "names.full_name": "Doe" }
      },
      "inner_hits": {}
    }
  },
  "size": 0,
  "aggs": {
    "lists": {
      "terms": {
        "field": "list_id"
      },
      "aggs": {
        "top_result": {
          "top_hits": {
            "size": 10
          }
        }
      }
    }
  }
}
'

Expected behavior

The following is the expected result which OpenSearch 2.12.0 gives. This includes inner_hits.

{
  "took" : 20,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : null,
    "hits" : [ ]
  },
  "aggregations" : {
    "lists" : {
      "doc_count_error_upper_bound" : 0,
      "sum_other_doc_count" : 0,
      "buckets" : [
        {
          "key" : 1,
          "doc_count" : 1,
          "top_result" : {
            "hits" : {
              "total" : {
                "value" : 1,
                "relation" : "eq"
              },
              "max_score" : 0.10607058,
              "hits" : [
                {
                  "_index" : "names-test",
                  "_id" : "1",
                  "_score" : 0.10607058,
                  "_source" : {
                    "list_id" : 1,
                    "names" : [
                      {
                        "full_name" : "John Doe"
                      },
                      {
                        "full_name" : "John Micheal Doe"
                      }
                    ]
                  },
                  "inner_hits" : {
                    "names" : {
                      "hits" : {
                        "total" : {
                          "value" : 2,
                          "relation" : "eq"
                        },
                        "max_score" : 0.11474907,
                        "hits" : [
                          {
                            "_index" : "names-test",
                            "_id" : "1",
                            "_nested" : {
                              "field" : "names",
                              "offset" : 0
                            },
                            "_score" : 0.11474907,
                            "_source" : {
                              "full_name" : "John Doe"
                            }
                          },
                          {
                            "_index" : "names-test",
                            "_id" : "1",
                            "_nested" : {
                              "field" : "names",
                              "offset" : 1
                            },
                            "_score" : 0.09739208,
                            "_source" : {
                              "full_name" : "John Micheal Doe"
                            }
                          }
                        ]
                      }
                    }
                  }
                }
              ]
            }
          }
        },
        {
          "key" : 2,
          "doc_count" : 1,
          "top_result" : {
            "hits" : {
              "total" : {
                "value" : 1,
                "relation" : "eq"
              },
              "max_score" : 0.10607058,
              "hits" : [
                {
                  "_index" : "names-test",
                  "_id" : "2",
                  "_score" : 0.10607058,
                  "_source" : {
                    "list_id" : 2,
                    "names" : [
                      {
                        "full_name" : "Jane Doe"
                      },
                      {
                        "full_name" : "Jane Michelle Doe"
                      }
                    ]
                  },
                  "inner_hits" : {
                    "names" : {
                      "hits" : {
                        "total" : {
                          "value" : 2,
                          "relation" : "eq"
                        },
                        "max_score" : 0.11474907,
                        "hits" : [
                          {
                            "_index" : "names-test",
                            "_id" : "2",
                            "_nested" : {
                              "field" : "names",
                              "offset" : 0
                            },
                            "_score" : 0.11474907,
                            "_source" : {
                              "full_name" : "Jane Doe"
                            }
                          },
                          {
                            "_index" : "names-test",
                            "_id" : "2",
                            "_nested" : {
                              "field" : "names",
                              "offset" : 1
                            },
                            "_score" : 0.09739208,
                            "_source" : {
                              "full_name" : "Jane Michelle Doe"
                            }
                          }
                        ]
                      }
                    }
                  }
                }
              ]
            }
          }
        }
      ]
    }
  }
}

The actual result in 2.13.0 is the following which is missing the inner_hits:

{
  "took" : 3,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : null,
    "hits" : [ ]
  },
  "aggregations" : {
    "lists" : {
      "doc_count_error_upper_bound" : 0,
      "sum_other_doc_count" : 0,
      "buckets" : [
        {
          "key" : 1,
          "doc_count" : 1,
          "top_result" : {
            "hits" : {
              "total" : {
                "value" : 1,
                "relation" : "eq"
              },
              "max_score" : 0.10607058,
              "hits" : [
                {
                  "_index" : "names-test",
                  "_id" : "1",
                  "_score" : 0.10607058,
                  "_source" : {
                    "list_id" : 1,
                    "names" : [
                      {
                        "full_name" : "John Doe"
                      },
                      {
                        "full_name" : "John Micheal Doe"
                      }
                    ]
                  }
                }
              ]
            }
          }
        },
        {
          "key" : 2,
          "doc_count" : 1,
          "top_result" : {
            "hits" : {
              "total" : {
                "value" : 1,
                "relation" : "eq"
              },
              "max_score" : 0.10607058,
              "hits" : [
                {
                  "_index" : "names-test",
                  "_id" : "2",
                  "_score" : 0.10607058,
                  "_source" : {
                    "list_id" : 2,
                    "names" : [
                      {
                        "full_name" : "Jane Doe"
                      },
                      {
                        "full_name" : "Jane Michelle Doe"
                      }
                    ]
                  }
                }
              ]
            }
          }
        }
      ]
    }
  }
}

Additional Details

Plugins
No plugins / default opensearch installation

Host/Environment (please complete the following information):

• OS: Arch Linux
• Version latest

I'm using opensearch from docker: https://hub.docker.com/layers/opensearchproject/opensearch/2.13.0/images/sha256-00f052502297cbc599af34b93605e1eb485438f0e9670dc8d82a4976da7d3feb?context=explore

Additional context

I set "size": 0 in the main query because I'm not interested in the "regular" hits but only in the aggregated top hits. If I change this to for example "size": 100, the "regular" hits will include the inner hits so there it works as expected. But I do this, the top hits still will not include the inner hits.

[dblock]

@martijnbolhuis Looks like a regression. Care to add a (failing) YAML REST test that reproduces this problem and try to bisect it to where the bug was introduced?

[martijnbolhuis]

@dblock Thanks for your response.

The following YAML REST test reproduces the problem. It seems that commit 965d85a introduces the problem. See also #12503

I've added the following file rest-api-spec/src/main/resources/rest-api-spec/test/search.aggregation/400_inner_hits.yml

setup:
  - do:
      indices.create:
          index: test_1
          body:
            settings:
              number_of_replicas: 0
            mappings:
              properties:
                list_id:
                  type: integer
                names:
                  type: nested
                  properties:
                    full_name:
                      type: text

  - do:
       bulk:
         refresh: true
         body:
           - index:
               _index: test_1
               _id:    1
           - list_id: 1
             names:
               - full_name: John Doe
               - full_name: John Micheal Doe
           - index:
               _index: test_1
               _id:    2
           - list_id: 2
             names:
               - full_name: Jane Doe
               - full_name: Jane Michelle Doe

---
"Include inner hits in top hits":
  - do:
      search:
        rest_total_hits_as_int: true
        body:
          query:
            nested:
              path: names
              query:
                match:
                  names.full_name: Doe
              inner_hits: { }
          size: 0
          aggs:
            lists:
              terms:
                field: list_id
              aggs:
                top_result:
                  top_hits:
                    size: 10

  - length: { hits.hits: 0 }
  - length: { aggregations.lists.buckets: 2 }
  - length: { aggregations.lists.buckets.0.top_result.hits.hits: 1 }
  - length: { aggregations.lists.buckets.0.top_result.hits.hits.0.inner_hits.names.hits.hits: 2 }
  - length: { aggregations.lists.buckets.1.top_result.hits.hits: 1 }
  - length: { aggregations.lists.buckets.1.top_result.hits.hits.0.inner_hits.names.hits.hits: 2 }

In the following my test fails:

git checkout 965d85aba69baf83f2f60649bc97e3eae44bff05
./gradlew --no-daemon --no-build-cache run
# In another shell
./gradlew ':rest-api-spec:yamlRestTest' --tests "org.opensearch.test.rest.ClientYamlTestSuiteIT" -Dtests.method="test {p0=search.aggregation/400_inner_hits/*}" -Dtests.cluster=localhost:9200 -Dtests.clustername=test -Dtests.rest.cluster=localhost:9200
...
Tests with failures:
 - org.opensearch.test.rest.ClientYamlTestSuiteIT.test {p0=search.aggregation/400_inner_hits/Include inner hits in top hits}

1 test completed, 1 failed

And in the following my test works:

git checkout 965d85aba69baf83f2f60649bc97e3eae44bff05~1
./gradlew --no-daemon --no-build-cache run
# In another shell
./gradlew ':rest-api-spec:yamlRestTest' --tests "org.opensearch.test.rest.ClientYamlTestSuiteIT" -Dtests.method="test {p0=search.aggregation/400_inner_hits/*}" -Dtests.cluster=localhost:9200 -Dtests.clustername=test -Dtests.rest.cluster=localhost:9200
...
BUILD SUCCESSFUL in 14s

[peternied]

[Triage - attendees 1 2 3 4 5 6 7 8]
@martijnbolhuis Thanks for creating the detailed issue, we'd welcome a pull request to address the issue.