Why does Elasticsearch sum score become inaccurate with big numbers?

Question

If I have an index that looks something like this:

[{:index=>
   {:_index=>"candidates",
    :_id=>"a1786607-e095-4621-bdf9-de2706475614",
    :data=>
     {:name=>"Carli Stark",
      :is_verified=>true, :has_work_permit=>true}}},
 {:index=>
   {:_index=>"candidates",
    :_id=>"57f78d3f-392e-4cdf-a5ff-6d10e7c89d5b",
    :data=>
     {:name=>"Gayla Keeling",
      :is_verified=>false, :has_work_permit=>true}}}]

And I make a query with score_mode sum and boost_mode replace (as I want only my relevance score to be considered):

GET candidates/_search
{
  "query": {
    "function_score": {
      "query": {
        "match_all": {}
      },
      "functions": [
        {
          "filter": {
            "term": {
              "is_verified": true
            }
          },
          "weight": 1000
        },
        {
          "filter": {
            "term": {
              "has_work_permit": true
            }
          },
          "weight": 100000000000
        }
      ],
      "score_mode": "sum",
      "boost_mode": "replace"
    }
  },
  "_source": ["is_verified"],
  "size": 50
}

Then why does Elasticsearch return exactly the same score for both documents? (Also notice that the order is wrong)

{
  "took" : 0,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : 9.9999998E10,
    "hits" : [
      {
        "_index" : "candidates_production_20240828114811152",
        "_type" : "_doc",
        "_id" : "cbd1b70b-f889-4136-a43e-f6782955f58e",
        "_score" : 9.9999998E10,
        "_source" : {
          "is_verified" : false,
          "has_work_permit" : true
        }
      },
      {
        "_index" : "candidates_production_20240828114811152",
        "_type" : "_doc",
        "_id" : "d644a5e5-09e0-496e-8830-c1a772c46611",
        "_score" : 9.9999998E10,
        "_source" : {
          "is_verified" : true
          "has_work_permit" : true
        }
      }
    ]
  }
}

If I use a bigger weight (such as 10000 instead of 1000) then the scores are different as expected:

{
  "took" : 1,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : 1.00000006E11,
    "hits" : [
      {
        "_index" : "candidates_production_20240828114811152",
        "_type" : "_doc",
        "_id" : "d644a5e5-09e0-496e-8830-c1a772c46611",
        "_score" : 1.00000006E11,
        "_source" : {
          "is_verified" : true,
          "has_work_permit" : true
        }
      },
      {
        "_index" : "candidates_production_20240828114811152",
        "_type" : "_doc",
        "_id" : "cbd1b70b-f889-4136-a43e-f6782955f58e",
        "_score" : 9.9999998E10,
        "_source" : {
          "is_verified" : false
          "has_work_permit" : true
        }
      }
    ]
  }
}

But how do I make it accurate? I need the smaller weights to be also considered in the scores regardless of how big the score is.

My Elasticsearch version is 7.10.1 (AWS ES)

Answer 1

This is due to the fact that the score is stored into a float, which can only represent integers accurately up to 2^24.

Relevant issue here.

My solution was to apply the sqrt function to the weights to keep them with smaller numbers but similar distributions.