Range Aggregation is not working properly for float fields

[ofirEdi]

Elasticsearch version (bin/elasticsearch --version): 7.12.1

Plugins installed: []

JVM version (java -version):1.8.0_292

OS version (uname -a if on a Unix-like system): Ubuntu 18.04.1

Description of the problem including expected versus actual behavior:
According to the documentation: https://www.elastic.co/guide/en/elasticsearch/reference/7.12/search-aggregations-bucket-range-aggregation.html range aggreagation creates buckets in a way that from is inclusive and to is exclusive. When working with float fields Elasticsearch seems to work vise versa and classify to the wrong bucket.

Steps to reproduce:

1. create index with float field:

PUT test1
{
  "settings": {
    "number_of_replicas": 0
  },
  "mappings": {
    "properties": {
      "amount": {
        "type": "float"
      }
    }
  }
}

2. insert a doc:

PUT test1/_doc/1
{
  "amount": 0.04
}

3. make range aggregation with limits of 0.04

POST test1/_search
{
  "size": 0,
  "aggs": {
    "testRange": {
      "range": {
        "field": "amount",
        "ranges": [
          {
            "from": 0.01,
            "to": 0.04
          },
          {
            "from": 0.04,
            "to": 0.06
          }
        ]
      }
    }
  }
}

the result i'm getting is that the doc belongs to 0.01-0.04 bucket and not to 0.04-0.06 bucket:

{
  "took" : 2,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 1,
      "relation" : "eq"
    },
    "max_score" : null,
    "hits" : [ ]
  },
  "aggregations" : {
    "testRange" : {
      "buckets" : [
        {
          "key" : "0.01-0.04",
          "from" : 0.01,
          "to" : 0.04,
          "doc_count" : 1
        },
        {
          "key" : "0.04-0.06",
          "from" : 0.04,
          "to" : 0.06,
          "doc_count" : 0
        }
      ]
    }
  }
}

I tested with Integers and it seems to be fine but for floats the aggregation not behaving like in the documentation.

[elasticmachine]

Pinging @elastic/es-analytics-geo (Team:Analytics)

[not-napoleon]

Hi @ofirEdi ,

Thanks for submitting this. I ran your repro steps and it also fails for me. Feels like it might be a rounding bug. I'll look into it and see if I can track down a fix.

[ofirEdi]

Thank you @not-napoleon . I'll keep watching.

[ofirEdi]

HI @not-napoleon any update?

[not-napoleon]

Hi @ofirEdi, I actually haven't had a chance to start on this. I've had a few other big things come up, unfortunately. It is next on my list though. Thank you for your patience.

[ofirEdi]

Hi @not-napoleon, thank you for the update.

[not-napoleon]

Okay, so I've done some digging, and the issue here is (basically) that we always parse the range end points as doubles, but the field value gets parsed (and stored) as a float. Then we do our comparison, and 0.04f < 0.04d == true. This happens because 0.04 isn't exactly representable as a double or float, and they cut off the approximation at different points.

I'm looking into a solution, but it's a little tricky. I think the right thing to do is parse the range end points with the type of the field being aggregated on, if we can. Problem is we don't know the field type until after parse time. Anyway, I'm working on it. Hopefully will have something soon.

[ofirEdi]

@not-napoleon i will keep following on the pull request. Thanks for taking the time and address this issue.