Elasticsearch Terms or Cardinality Aggregation – Order by number of distinct values

aggregationcardinalityelasticsearch

Friends,

I am doing some analysis to find unique pairs from 100s of millions of documents. The mock example is as shown below:

doc field1 field2

  1. AAA : BBB
  2. AAA : CCC
  3. PPP : QQQ
  4. PPP : QQQ
  5. XXX : YYY
  6. XXX : YYY
  7. MMM : NNN

90% of the document contains an unique pair as shown above in doc 3, 4, 5, 6 and 7 which I am not interested on my aggregation result. I am interested to aggregate doc 1 and 2.

Terms Aggregation Query:



    "aggs": {
        "f1": {
            "terms": {
                "field": "FIELD1",
                "min_doc_count": 2
              },
              "aggs": {
                "f2": {
                      "terms": {
                    "field": "FIELD2"
                      }
                }
              }
        }
    }

Term Aggregation Result



    "aggregations": {
         "f1": {
               "buckets": [
                 {
                    "key": "PPP",
                    "doc_count": 2,
                    "f2": {
                        "buckets": [
                              {
                               "key": "QQQ",
                            "doc_count": 2
                              }
                         ]
                    }
                  },
                  {
                    "key": "XXX",
                    "doc_count": 2,
                    "f2": {
                        "buckets": [
                              {
                               "key": "YYY",
                            "doc_count": 2
                              }
                         ]
                    }
                  },
                  {
                    "key": "AAA",
                    "doc_count": 2,
                    "f2": {
                        "buckets": [
                              {
                                   "key": "BBB",
                                "doc_count": 1
                              },
                              {
                               "key": "CCC",
                               "doc_count": 1
                              }
                         ]
                    }
                  }
            ]
         }
    }

I am interested only on key AAA to be in the aggregation result. What is the best way to filter the aggregation result containing distinct pairs?

I tried with cardinality aggregation which result unque value count. However I am not able to filter out what I am not interested from the aggregation results.

Cardinality Aggregation Query



    "aggs": {
        "f1": {
            "terms": {
                "field": "FIELD1",
                "min_doc_count": 2
              },
            "aggs": {
                "f2": {
                      "cardinality": {
                        "field": "FIELD2"
                      }
                }
              }
        }
    }

Cardinality Aggregation Result



    "aggregations": {
        "f1": {
               "buckets": [
                 {
                    "key": "PPP",
                    "doc_count": 2,
                    "f2": {
                          "value" : 1
                    }
                  },
                  {
                    "key": "XXX",
                    "doc_count": 2,
                    "f2": {
                          "value" : 1
                    }
                  },
                  {
                    "key": "AAA",
                    "doc_count": 2,
                    "f2": {
                          "value" : 2
                    }
                  }
            ]
         }
    }

Atleast if I could sort by cardinal value, that would be help me to find some workarounds. Please help me in this regard.

P.S: Writing a spark/mapreduce program to post process/filter the aggregation result is not expected solution for this issue.

Best Answer

I suggest to use filter query along with aggregations, since you are only interested in field1=AAA.

I have a similar example here.

For example, I have an index of all patients in my hospital. I store their drug use in a nested object DRUG. Each patient could take different drugs, and each could take a single drug for multiple times.

Now if I wanted to find the number of patients who took aspirin at least once, the query could be:

{
  "size": 0,
  "_source": false,
  "query": {
    "filtered": {
      "query": {
        "match_all": {}
      },
      "filter": {
        "nested": {
          "path": "DRUG",
          "filter": {
            "bool": {
              "must": [{ "term": { "DRUG.NAME": "aspirin" } }]
  }}}}}},
  "aggs": {
    "DRUG_FACETS": {
      "nested": {
        "path": "DRUG"
      },
      "aggs": {
        "DRUG_NAME_FACETS": {
          "terms": { "field": "DRUG.NAME", "size": 0 },
          "aggs": {
            "DISTINCT": { "cardinality": { "field": "DRUG.PATIENT" } }
          }
  }}}}
}

Sample result:

{
  "took": 6,
  "timed_out": false,
  "_shards": {
    "total": 5,
    "successful": 5,
    "failed": 0
  },
  "hits": {
    "total": 6,
    "max_score": 0,
    "hits": []
  },
  "aggregations": {
    "DRUG_FACETS": {
      "doc_count": 11,
      "DRUG_NAME_FACETS": {
        "buckets": [
          {
            "key": "aspirin",
            "doc_count": 6,
            "DISTINCT": {
              "value": 6
            }
          },
          {
            "key": "vitamin-b",
            "doc_count": 3,
            "DISTINCT": {
              "value": 2
            }
          },
          {
            "key": "vitamin-c",
            "doc_count": 2,
            "DISTINCT": {
              "value": 2
            }
          }
        ]
      }
    }
  }
}

The first one in the buckets would be aspirin. But you can see other 2 patients had also taken vitamin-b when they took aspirin.

If you change the field value of DRUG.NAME to another drug name for example "vitamin-b", I suppose you would get vitamin-b in the first position of the buckets.

Hopefully this is helpful to your question.

Related Topic