function_score scores are rounded

Question

It seems that elasticsearch is rounding scores that comes from function_score. I have a complex query that use function_score and has_child to sort parent depending on a creation_date date of a child (therefore I can't use sort). The return scores are rounded even in a very simple scenario:

{
  "query": {
    "function_score": {
      "boost_mode": "replace",
      "functions": [
        {
          "script_score": {
            "script": "doc[\"creation_date\"].value"
          }
        }
      ]
    }
  },
  "script_fields": {
    "date": {
      "script": "doc['creation_date'].date"
    },
    "value": {
      "script": "doc['creation_date'].value"
    }
  }
}

I tried with different script_score :

doc["creation_date"].date.getMillis()

or (creation_date_ts is the same date as a unixtimestamp)

doc["creation_date_ts"].value

or even a literal int:

1416398851930

But whatever I use I end up with rounded scores.

Here is an example of the results :

{
   "_id": "nyrhhh5z",
   "_index": "xxx-1416487038",
   "_score": 1416398900000,
   "_type": "xxx",
   "fields": {
      "date": [
         "2014-11-19T12:07:31.930Z"
      ],
      "value": [
         1416398851930
      ]
   }
}

The _score should be 1416398851930 (value) be gets rounded to 1416398900000. It is a normal behavior, is it documented somewhere (I've searched)? and it there a way to force elasticsearch not to round the scores.

Version of elasticsearch is 1.3.2 (ubuntu 14.04).

Answer 1

The _score value is internally hard coded as a float which can only accurately represent integers up to the value 134217728. Therefore, if you want to make use, in the scoring function, of a field stored as a number larger than that, it will overflow the buffer and be truncated. See this github issue