elasticsearch - filter by percentile

问题

Say if I want to filter documents by some field within 10th to 20th percentile. I'm wondering if it's possible by some simple query, something like {"fieldName":{"percentile": [0.1, 0.2]}}.

Say I have these documents:

[{"a":1,"b":101},{"a":2,"b":102},{"a":3,"b":103}, ..., {"a":100,"b":200}]

I need to filter the top 10th of them by a (with ascending order), that would be a from 1 to 10. Then I need to sort those results by b with descending order, then take the paginated result (like page No.2, with 10 items every page).

One solution in mind would be:

get the total count of the documents.
sort the documents by a, take the corresponding _id with limit 0.1 * total_count
write the final query, something like id in (...) order by b

But the shortcomings are pretty obvious too:

seems not effecient if we're talking about subsecond latency
the second query might not work if we have too many _id returned in the first query (ES only allows 1000 by default. I can change the config of course, but there's always a limit).

回答1:

I doubt that there is a way to do this in one query if the exact values of a are not known beforehand, although I think one pretty efficient approach is feasible.

I would suggest to do a percentiles aggregation as first query and range query as second.

In my sample index I have only 14 documents, so for explanatory reasons I will try to find those documents that are from 30% to 60% of field a and sort them by field b in inverse order (so to be sure that sort worked).

Here are the docs I inserted:

{"a":1,"b":101}
{"a":5,"b":105}
{"a":10,"b":110}
{"a":2,"b":102}
{"a":6,"b":106}
{"a":7,"b":107}
{"a":9,"b":109}
{"a":4,"b":104}
{"a":8,"b":108}
{"a":12,"b":256}
{"a":13,"b":230}
{"a":14,"b":215}
{"a":3,"b":103}
{"a":11,"b":205}

Let's find out which are the bounds for field a between 30% and 60% percentiles:

POST my_percent/doc/_search
{
    "size": 0,
    "aggs" : {
        "percentiles" : {
            "percentiles" : {
                "field" : "a",
                "percents": [ 30, 60, 90 ]
            }
        }
    }
}

With my sample index it looks like this:

{
...
  "hits": {
    "total": 14,
    "max_score": 0,
    "hits": []
  },
  "aggregations": {
    "percentiles": {
      "values": {
        "30.0": 4.9,
        "60.0": 8.8,
        "90.0": 12.700000000000001
      }
    }
  }
}

Now we can use the boundaries to do the range query:

POST my_percent/doc/_search
{
    "query": {
      "range": {
            "a" : {
                "gte" : 4.9,
                "lte" : 8.8
            }
        }
    },
    "sort": {
      "b": "desc"
    }
}

And the result is:

{
  "took": 5,
  "timed_out": false,
  "_shards": {
    "total": 5,
    "successful": 5,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": 4,
    "max_score": null,
    "hits": [
      {
        "_index": "my_percent",
        "_type": "doc",
        "_id": "vkFvYGMB_zM1P5OLcYkS",
        "_score": null,
        "_source": {
          "a": 8,
          "b": 108
        },
        "sort": [
          108
        ]
      },
      {
        "_index": "my_percent",
        "_type": "doc",
        "_id": "vUFvYGMB_zM1P5OLWYkM",
        "_score": null,
        "_source": {
          "a": 7,
          "b": 107
        },
        "sort": [
          107
        ]
      },
      {
        "_index": "my_percent",
        "_type": "doc",
        "_id": "vEFvYGMB_zM1P5OLRok1",
        "_score": null,
        "_source": {
          "a": 6,
          "b": 106
        },
        "sort": [
          106
        ]
      },
      {
        "_index": "my_percent",
        "_type": "doc",
        "_id": "u0FvYGMB_zM1P5OLJImy",
        "_score": null,
        "_source": {
          "a": 5,
          "b": 105
        },
        "sort": [
          105
        ]
      }
    ]
  }
}

Note that the results of percentiles aggregation are approximate.

In general, this looks like a task better solved by pandas or a Spark job.

Hope that helps!

来源：https://stackoverflow.com/questions/50166949/elasticsearch-filter-by-percentile

标签

ElasticSearch

percentile