ES7.17版本terms查询性能问题

背景

1.对于7版本（大版本）集群希望只维护一个版本，最终选择7.17，对同大版本的7.5版本集群进行升级

2.根据官方描述，_id放到堆外性能损失非常小可以忽略，且对BKD进行了优化

3.升级完成，一段时间之后，收到用户报障

4.抽样检查了下部分升级的集群，其中部分受到影响，部分不受影响。且每个集群内存均有一定优化（预期内）

调查&分析

1.发现is_deleted文档特别多，怀疑是7.17版本对于碎片过于敏感。做forcemerge，没什么效果。

2.GET _nodes/hot_threads 查看耗时部分，结果展示笼统，没得到关键信息。

3.给语句加上profile，查看耗时部分。

GET index-v1/_search
{"profile":"true","query":{"bool":{"filter":[{"term":{"xid":{"value":"11111111","boost":1.0}}},{"terms":{"status":[2,3,4],"boost":1.0}},{"terms":{"platform":["aaa","bbb"],"boost":1.0}},{"terms":{"pId":[1,2],"boost":1.0}}],"adjust_pure_negative":true,"boost":1.0}},"sort":[{"time":{"order":"desc"}}]}

从脱敏的简化结果中可以看出来，主要是 status、pId 字段耗时高，这两个字段都是integer类型，并且使用了terms查询。

{
  "took": 554,
  "timed_out": false,
  "_shards": {
    "total": 3,
    "successful": 3,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 5,
      "relation": "eq"
    },
    "max_score": null,
    "hits": [
      ...
    ]
  },
  "profile": {
    "shards": [
      {
        "id": "[APxxxxxxxxxxxxxxQ][index-v1][0]",
        "searches": [
          {
            "query": [
              {
                "type": "BooleanQuery",
                "description": "#xid:111111111 #status:{2 3 4} #ConstantScore(platform:aaa platform:bbb) #pId:{1 2}",
                "time_in_nanos": 415205306,
                "breakdown": {
                  ...
                  "build_scorer": 415028271
                },
                "children": [
                  {
                    "type": "TermQuery",
                    "description": "xid:111111111",
                    "time_in_nanos": 102656,
                    "breakdown": {
                      .....
                      "build_scorer": 86264
                    }
                  },
                  {
                    "type": "PointInSetQuery",
                    "description": "status:{2 3 4}",
                    "time_in_nanos": 220394978,
                    "breakdown": {
                      ....
                      "build_scorer": 220385119
                    }
                  },
                  {
                    "type": "ConstantScoreQuery",
                    "description": "ConstantScore(platform:aaa platform:bbb)",
                    "time_in_nanos": 341845,
                    "breakdown": {
                      .....
                      "build_scorer": 282277
                    },
                    "children": [
                      {
                        "type": "BooleanQuery",
                        "description": "platform:aaa platform:bbb",
                        "time_in_nanos": 329042,
                        "breakdown": {
                          .....
                          "build_scorer": 277752
                        },
                        "children": [
                          {
                            "type": "TermQuery",
                            "description": "platform:aaa",
                            "time_in_nanos": 62446,
                            "breakdown": {
                              .....
                              "build_scorer": 37931
                            }
                          },
                          {
                            "type": "TermQuery",
                            "description": "platform:bbb",
                            "time_in_nanos": 15093,
                            "breakdown": {
                              .....
                              "build_scorer": 6981
                            }
                          }
                        ]
                      }
                    ]
                  },
                  {
                    "type": "PointInSetQuery",
                    "description": "pId:{1 2}",
                    "time_in_nanos": 194164297,
                    "breakdown": {
                      ....
                      "build_scorer": 194160452
                    }
                  }
                ]
              }
            ],
            "rewrite_time": 40044,
            "collector": [
              {
                "name": "SimpleFieldCollector",
                "reason": "search_top_hits",
                "time_in_nanos": 144012
              }
            ]
          }
        ]

4.单个的profile无法说明问题，进一步排查：使用arthas工具获取一段时间内的火焰图

可以看到主要就是BKD数据结构占用的CPU。

5.参考官方论坛相似问题：https://discuss.elastic.co/t/very-slow-search-performance-after-upgrade-to-7-16-1/296152/3

6.integer类型的terms查询性能较差，看起来官方描述的BKD相关优化指的是range

7.测试验证，将字段改成keyword，查看结果，CPU查询耗时恢复到正常范围