在上一篇文章 “Elasticsearch：运用scroll接口对大量数据实现更好的分页”，我们讲述了如何运用scroll接口来对大量数据来进行有效地分页。在那篇文章中，我们讲述了两种方法：

1. from加上size的方法来进行分页
2. 运用scroll接口来进行分页

对于大量的数据而言，我们尽量避免使用from+size这种方法。这里的原因是index.max_result_window的默认值是10K，也就是说from+size的最大值是1万。搜索请求占用堆内存和时间与from+size成比例，这限制了内存。假如你想hit从990到1000，那么每个shard至少需要1000个文档：

为了避免过度使得我们的cluster繁忙，通常Scroll接口被推荐作为深层次的scrolling，但是因为维护scroll上下文也是非常昂贵的，所以这种方法不推荐作为实时用户请求。search_after参数通过提供实时cursor来解决此问题。 我们的想法是使用上一页的结果来帮助检索下一页。

我们先输入如下的文档到twitter索引中：

POST _bulk
{ "index" : { "_index" : "twitter", "_id": 1} }
{"user":"双榆树-张三", "DOB":"1980-01-01", "message":"今儿天气不错啊，出去转转去","uid":2,"age":20,"city":"北京","province":"北京","country":"中国","address":"中国北京市海淀区","location":{"lat":"39.970718","lon":"116.325747"}}
{ "index" : { "_index" : "twitter", "_id": 2 }}
{"user":"东城区-老刘", "DOB":"1981-01-01", "message":"出发，下一站云南！","uid":3,"age":30,"city":"北京","province":"北京","country":"中国","address":"中国北京市东城区台基厂三条3号","location":{"lat":"39.904313","lon":"116.412754"}}
{ "index" : { "_index" : "twitter", "_id": 3} }
{"user":"东城区-李四", "DOB":"1982-01-01", "message":"happy birthday!","uid":4,"age":30,"city":"北京","province":"北京","country":"中国","address":"中国北京市东城区","location":{"lat":"39.893801","lon":"116.408986"}}
{ "index" : { "_index" : "twitter", "_id": 4} }
{"user":"朝阳区-老贾","DOB":"1983-01-01", "message":"123,gogogo","uid":5,"age":35,"city":"北京","province":"北京","country":"中国","address":"中国北京市朝阳区建国门","location":{"lat":"39.718256","lon":"116.367910"}}
{ "index" : { "_index" : "twitter", "_id": 5} }
{"user":"朝阳区-老王","DOB":"1984-01-01", "message":"Happy BirthDay My Friend!","uid":6,"age":50,"city":"北京","province":"北京","country":"中国","address":"中国北京市朝阳区国贸","location":{"lat":"39.918256","lon":"116.467910"}}
{ "index" : { "_index" : "twitter", "_id": 6} }
{"user":"虹桥-老吴", "DOB":"1985-01-01", "message":"好友来了都今天我生日，好友来了,什么 birthday happy 就成!","uid":7,"age":90,"city":"上海","province":"上海","country":"中国","address":"中国上海市闵行区","location":{"lat":"31.175927","lon":"121.383328"}}

这里共有6个文档。假设检索第一页的查询如下所示：

GET twitter/_search
{
  "size": 2,
  "query": {
    "match": {
      "city": "北京"
    }
  },
  "sort": [
    {
      "DOB": {
        "order": "asc"
      }
    },
    {
      "user.keyword": {
        "order": "asc"
      }
    }
  ]
}

显示的结果为：

{
  "took" : 29,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 5,
      "relation" : "eq"
    },
    "max_score" : null,
    "hits" : [
      {
        "_index" : "twitter",
        "_type" : "_doc",
        "_id" : "1",
        "_score" : null,
        "_source" : {
          "user" : "双榆树-张三",
          "DOB" : "1980-01-01",
          "message" : "今儿天气不错啊，出去转转去",
          "uid" : 2,
          "age" : 20,
          "city" : "北京",
          "province" : "北京",
          "country" : "中国",
          "address" : "中国北京市海淀区",
          "location" : {
            "lat" : "39.970718",
            "lon" : "116.325747"
          }
        },
        "sort" : [
          315532800000,
          "双榆树-张三"
        ]
      },
      {
        "_index" : "twitter",
        "_type" : "_doc",
        "_id" : "2",
        "_score" : null,
        "_source" : {
          "user" : "东城区-老刘",
          "DOB" : "1981-01-01",
          "message" : "出发，下一站云南！",
          "uid" : 3,
          "age" : 30,
          "city" : "北京",
          "province" : "北京",
          "country" : "中国",
          "address" : "中国北京市东城区台基厂三条3号",
          "location" : {
            "lat" : "39.904313",
            "lon" : "116.412754"
          }
        },
        "sort" : [
          347155200000,
          "东城区-老刘"
        ]
      }
    ]
  }
}

上述请求的结果包括每个文档的sort值数组。 这些sort值可以与search_after参数一起使用，以开始返回在这个结果列表之后的任何文档。 例如，我们可以使用上一个文档的sort值并将其传递给search_after以检索下一页结果：

GET twitter/_search
{
  "size": 2,
  "query": {
    "match": {
      "city": "北京"
    }
  },
  "search_after": [
    347155200000,
    "东城区-老刘"
  ],
  "sort": [
    {
      "DOB": {
        "order": "asc"
      }
    },
    {
      "user.keyword": {
        "order": "asc"
      }
    }
  ]
}

在这里在search_after中，我们把上一个搜索结果的sort值放进来。 显示的结果为：

{
  "took" : 47,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 5,
      "relation" : "eq"
    },
    "max_score" : null,
    "hits" : [
      {
        "_index" : "twitter",
        "_type" : "_doc",
        "_id" : "3",
        "_score" : null,
        "_source" : {
          "user" : "东城区-李四",
          "DOB" : "1982-01-01",
          "message" : "happy birthday!",
          "uid" : 4,
          "age" : 30,
          "city" : "北京",
          "province" : "北京",
          "country" : "中国",
          "address" : "中国北京市东城区",
          "location" : {
            "lat" : "39.893801",
            "lon" : "116.408986"
          }
        },
        "sort" : [
          378691200000,
          "东城区-李四"
        ]
      },
      {
        "_index" : "twitter",
        "_type" : "_doc",
        "_id" : "4",
        "_score" : null,
        "_source" : {
          "user" : "朝阳区-老贾",
          "DOB" : "1983-01-01",
          "message" : "123,gogogo",
          "uid" : 5,
          "age" : 35,
          "city" : "北京",
          "province" : "北京",
          "country" : "中国",
          "address" : "中国北京市朝阳区建国门",
          "location" : {
            "lat" : "39.718256",
            "lon" : "116.367910"
          }
        },
        "sort" : [
          410227200000,
          "朝阳区-老贾"
        ]
      }
    ]
  }
}

注意：当我们使用search_after时，from值必须设置为0或者-1。

search_after不是自由跳转到随机页面而是并行scroll多个查询的解决方案。 它与scroll API非常相似，但与它不同，search_after参数是无状态的，它始终针对最新版本的搜索器进行解析。 因此，排序顺序可能会在步行期间发生变化，具体取决于索引的更新和删除。