Geohash_grid aggregation with unique and max field doesn't respect unique results

[ink4]

Let's have a records where we have some records for few devices, each record contains information about position, timestamp and deviceId:

[
    {
        deviceId: 1,
        geoPoint: {
            lat: 1, lon: 2
        },
        timestamp: '2018-01-01 12:00:00'
    },
    {
        deviceId: 1,
        geoPoint: {
            lat: 3, lon: 4
        },
        timestamp: '2018-01-01 12:01:00'
    },
    {
        deviceId: 1,
        geoPoint: {
            lat: 5, lon: 6
        },
        timestamp: '2018-01-01 12:02:00'
    },
    {
        deviceId: 2,
        geoPoint: {
            lat: 1, lon: 2
        },
        timestamp: '2018-01-01 12:01:00'
    },
    {
        deviceId: 2,
        geoPoint: {
            lat: 3, lon: 4
        },
        timestamp: '2018-01-01 12:02:00'
    },
    {
        deviceId: 2,
        geoPoint: {
            lat: 5, lon: 6
        },
        timestamp: '2018-01-01 12:03:00'
    }
];

Mapping is quite simple:

    deviceId: {type: 'keyword'},
    geoPoint: {
        type: 'geo_point'
    },
    timestamp: {type: 'date', format: DateUtil.ES_FORMAT},

I would like to fetch clusters based on geohash_grid for last positions for distinct devices, quite simple, however query doesn't return correct results - instead it looks like doesn't take into account sub-aggregations.

Query:

    {
           size: 0,
            aggs: {
            clustering: {
                geohash_grid: {
                    field: 'geoPoint',
                    precision: 2
                },
                aggs: {
                    last_device_position: {
                        terms: {
                            field: 'deviceId',
                            size: 1,
                            order: {
                                last_one: 'desc'
                            }
                        },
                        aggs: {
                            last_one: {
                                max: {
                                    field: 'timestamp'
                                }
                            }
                        }
                    }
                }
            }
        }
    };

Results:

   {
   "took": 3,
    "timed_out": false,
    "_shards": {
      "total": 5,
      "successful": 5,
      "skipped": 0,
      "failed": 0
    },
    "hits": {
      "total": 6,
      "max_score": 0,
      "hits": []
    },
    "aggregations": {
      "clustering": {
        "buckets": [
    {
      "key": "u2",
      "doc_count": 4,
      "last_device_position": {
        "doc_count_error_upper_bound": 0,
        "sum_other_doc_count": 2,
        "buckets": [
          {
            "key": "1",
            "doc_count": 2,
            "last_one": {
              "value": 1514764860000,
              "value_as_string": "2018-01-01 00:01:00"
            }
          }
        ]
      }
    },
    {
      "key": "t5",
      "doc_count": 2,
      "last_device_position": {
        "doc_count_error_upper_bound": 0,
        "sum_other_doc_count": 1,
        "buckets": [
          {
            "key": "1",
            "doc_count": 1,
            "last_one": {
              "value": 1514764920000,
              "value_as_string": "2018-01-01 00:02:00"
            }
          }
        ]
      }
    }
  ]
}
    }
  }

Which gives incorrect results, only records with timestamp: 2018-01-01 00:02:00 should be taken into account.

I tried as well with:

geohash_grid: {
                    field: 'last_device_position > geoPoint',
                    precision: 2
                },

or

geohash_grid: {
                    field: 'last_one > geoPoint',
                    precision: 2
                }

But no results.

What is the right way to solve this issue ?

[ink4]

any clues ?