Elasticsearch排序结果异常：float字段sort值与实际值不符求助

问题说明 #

number_2字段定义为float类型，执行降序排序查询时，发现源数据中number_2: 20164279的文档，返回结果的sort字段值为20164280，导致排序逻辑混乱。删除并重建索引后问题仍然存在。

查询DSL #

{
  "from": 0,
  "size": 50,
  "timeout": "60s",
  "query": {
    "bool": {
      "must": [
        {
          "match_all": {
            "boost": 1
          }
        },
        {
          "bool": {
            "must": [
              {
                "term": {
                  "item": {
                    "value": "project_online_1.1002",
                    "boost": 1
                  }
                }
              }
            ],
            "adjust_pure_negative": true,
            "boost": 1
          }
        }
      ],
      "adjust_pure_negative": true,
      "boost": 1
    }
  },
  "sort": [
    {
      "number_2": {
        "order": "desc"
      }
    }
  ]
}

异常结果示例 #

{
    "_index": "blue_0",
    "_type": "_doc",
    "_id": "project_online_1_1002_20164279",
    "_score": null,
    "_source": {
        "project": "project_online_1",
        "item": "project_online_1.1002",
        "text_0": "emp3",
        "bool_0": false,
        "number_2": 20164279,
        "bool_1": false,
        "bool_2": false,
        "text_4": "3xs7vovx",
        "bool_3": false,
        "number_4": 0,
        "datetime_6": "2023-08-09T10:15:30.281Z",
        "datetime_7": "2023-08-09T10:16:06.751Z",
        "number_8": null,
        "text_8": "KeDUkyqM0OjJ7ZlIdDNMe9.png",
        "number_12": 2026,
        "text_18": null,
        "number_30": 188,
        "text_22": "4f4becdd-88b1-4b60-ad44-de90510fcb54",
        "text_1": "free",
        "text_2": null,
        "bool_4": true,
        "bool_5": false,
        "bool_6": false,
        "bool_7": false,
        "bool_8": true,
        "number_31": null,
        "text_23": null
    },
    "sort": [
        20164280
    ]
},
{
    "_index": "blue_0",
    "_type": "_doc",
    "_id": "project_online_1_1002_20164278",
    "_score": null,
    "_source": {
        "project": "project_online_1",
        "item": "project_online_1.1002",
        "text_0": "0809_app_setting_test_test_2",
        "bool_0": false,
        "number_2": 20164278,
        "bool_1": false,
        "bool_2": false,
        "text_4": "rotcqizm",
        "bool_3": false,
        "number_4": 0,
        "datetime_6": "2023-08-09T10:15:15.459Z",
        "datetime_7": "2023-08-10T10:04:06.839Z",
        "number_8": null,
        "text_8": "vMx5AqClFbKPc0Nov0PHV.png",
        "number_12": 2026,
        "text_18": null,
        "number_30": 227,
        "text_22": "277a7ae7-9cd3-40a4-82f8-fda0ea500742",
        "text_1": "free",
        "text_2": null,
        "bool_4": true,
        "bool_5": false,
        "bool_6": false,
        "bool_7": false,
        "bool_8": true,
        "number_31": null,
        "text_23": null
    },
    "sort": [
        20164278
    ]
}

原因分析 #

这是单精度float类型的精度限制导致的：

• float类型的有效十进制精度仅为6-7位，而20164279是8位整数，超出了float的精确存储范围。
• Elasticsearch在索引时会将该值转换为float类型存储，此时无法精确表示20164279，会被舍入到最近的可表示的float值（即20164280）。
• _source中保留的是原始输入的数值，但排序是基于索引中存储的近似float值执行的，因此出现了sort值与源数据不一致的情况。

解决方案 #

1. 改用整数类型：如果number_2字段存储的都是整数，将字段类型修改为integer或long，这类类型可以精确存储该范围内的数值。
2. 改用double类型：如果必须使用浮点类型，替换为double类型，double的有效精度为15-17位，能够精确存储该范围内的整数。

内容的提问来源于stack exchange，提问作者freedom