The number of results returned by a query may be far too many to display each geo-point individually on a map. The geohash_grid aggregation buckets nearby geo-points together by calculating the geohash for each point, at the level of precision that you define.

The result is a grid of cells—​one cell per geohash—​that can be displayed on a map. By changing the precision of the geohash, you can summarize information across the whole world, by country, or by city block.

The aggregation is sparse—it returns only cells that contain documents. If your geohashes are too precise and too many buckets are generated, it will return, by default, the 10,000 most populous cells—​those containing the most documents. However, it still needs to generate all the buckets in order to figure out which are the most populous 10,000. You need to control the number of buckets generated by doing the following:

GET /attractions/restaurant/_search
{
  "size" : 0,
  "query": {
    "constant_score": {
      "filter": {
        "geo_bounding_box": {
          "location": { 
            "top_left": {
              "lat":  40.8,
              "lon": -74.1
            },
            "bottom_right": {
              "lat":  40.4,
              "lon": -73.7
            }
          }
        }
      }
    }
  },
  "aggs": {
    "new_york": {
      "geohash_grid": { 
        "field":     "location",
        "precision": 5
      }
    }
  }
}

Geohashes with precision 5 measure about 25km² each, so 10,000 cells at this precision would cover 250,000km². The bounding box that we specified measures approximately 44km x 33km, or about 1,452km², so we are well within safe limits; we definitely won’t create too many buckets in memory.

The response from the preceding request looks like this:

...
"aggregations": {
  "new_york": {
     "buckets": [ 
        {
           "key": "dr5rs",
           "doc_count": 2
        },
        {
           "key": "dr5re",
           "doc_count": 1
        }
     ]
  }
}
...

Again, we didn’t specify any sub-aggregations, so all we got back was the document count. We could have asked for popular restaurant types, average price, or other details.