相关度评分 TF&IDF算法
Elasticsearch的相关度评分（relevance score）算法采用的是term frequency/inverse document frequency算法，简称为TF/IDF算法。
算法介绍：
- relevance score算法：简单来说就是，就是计算出一个索引中的文本，与搜索文本，它们之间的关联匹配程度。
- TF/IDF算法：分为两个部分，IF 和IDF
-
Term Frequency(TF)： 搜索文本中的各个词条在field文本中出现了多少次，出现的次数越多，就越相关
例如：
搜索请求：hello world
doc1: hello you, and world is very good
doc2: hello, how are you
那么此时根据TF算法，doc1的相关度要比doc2的要高 -
Inverse Document Frequency(IDF)：搜索文本中的各个词条在整个索引的所有文档中出现的次数，出现的次数越多，就越不相关。
搜索请求： hello world
doc1: hello, today is very good.
doc2: hi world, how are you.
比如在index中有1万条document, hello这个单词在所有的document中，一共出现了1000次，world这个单词在所有的document中一共出现100次。那么根据IDF算法此时doc2的相关度要比doc1要高。 -
field-length norm：field-length norm就是field长度越长，相关度就越弱
搜索请求：hello world
doc1: {"title": "hello article", "content": "1万个单词"}
doc2: {"title": "my article", "content": "1万个单词， hi world"}
此时hello world在整个index中出现的次数是一样多的。但是根据Field-length norm此时doc1比doc2相关度要高。因为title字段更短。
_score是如何被计算出来的
GET /test_index/test_type/_search?explain
{
"query": {
"match": {
"test_field": "test hello"
}
}
}
{
"took": ,
"timed_out": false,
"_shards": {
"total": ,
"successful": ,
"failed":
},
"hits": {
"total": ,
"max_score": 0.843298,
"hits": [
{
"_shard": "[test_index][2]",
"_node": "1LdqLFqxQQq4xg2MphI_gw",
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": 0.843298,
"_source": {
"test_field": "test test"
},
"_explanation": {
"value": 0.843298,
"description": "sum of:",
"details": [
{
"value": 0.843298,
"description": "sum of:",
"details": [
{
"value": 0.843298,
"description": "weight(test_field:test in 0) [PerFieldSimilarity], result of:",
"details": [
{
"value": 0.843298,
"description": "score(doc=0,freq=2.0 = termFreq=2.0\n), product of:",
"details": [
{
"value": 0.6931472,
"description": "idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:",
"details": [
{
"value": ,
"description": "docFreq",
"details": []
},
{
"value": ,
"description": "docCount",
"details": []
}
]
},
{
"value": 1.2166219,
"description": "tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:",
"details": [
{
"value": ,
"description": "termFreq=2.0",
"details": []
},
{
"value": 1.2,
"description": "parameter k1",
"details": []
},
{
"value": 0.75,
"description": "parameter b",
"details": []
},
{
"value": 1.75,
"description": "avgFieldLength",
"details": []
},
{
"value": 2.56,
"description": "fieldLength",
"details": []
}
]
}
]
}
]
}
]
},
{
"value": ,
"description": "match on required clause, product of:",
"details": [
{
"value": ,
"description": "# clause",
"details": []
},
{
"value": ,
"description": "_type:test_type, product of:",
"details": [
{
"value": ,
"description": "boost",
"details": []
},
{
"value": ,
"description": "queryNorm",
"details": []
}
]
}
]
}
]
}
},
{
"_shard": "[test_index][1]",
"_node": "1LdqLFqxQQq4xg2MphI_gw",
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": 0.43445712,
"_source": {
"test_field": "test client 2"
},
"_explanation": {
"value": 0.43445715,
"description": "sum of:",
"details": [
{
"value": 0.43445715,
"description": "sum of:",
"details": [
{
"value": 0.43445715,
"description": "weight(test_field:test in 0) [PerFieldSimilarity], result of:",
"details": [
{
"value": 0.43445715,
"description": "score(doc=0,freq=1.0 = termFreq=1.0\n), product of:",
"details": [
{
"value": 0.47000363,
"description": "idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:",
"details": [
{
"value": ,
"description": "docFreq",
"details": []
},
{
"value": ,
"description": "docCount",
"details": []
}
]
},
{
"value": 0.92436975,
"description": "tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:",
"details": [
{
"value": ,
"description": "termFreq=1.0",
"details": []
},
{
"value": 1.2,
"description": "parameter k1",
"details": []
},
{
"value": 0.75,
"description": "parameter b",
"details": []
},
{
"value": 3.3333333,
"description": "avgFieldLength",
"details": []
},
{
"value": ,
"description": "fieldLength",
"details": []
}
]
}
]
}
]
}
]
},
{
"value": ,
"description": "match on required clause, product of:",
"details": [
{
"value": ,
"description": "# clause",
"details": []
},
{
"value": ,
"description": "_type:test_type, product of:",
"details": [
{
"value": ,
"description": "boost",
"details": []
},
{
"value": ,
"description": "queryNorm",
"details": []
}
]
}
]
}
]
}
},
{
"_shard": "[test_index][3]",
"_node": "1LdqLFqxQQq4xg2MphI_gw",
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": 0.25316024,
"_source": {
"test_field": "test client 1"
},
"_explanation": {
"value": 0.25316024,
"description": "sum of:",
"details": [
{
"value": 0.25316024,
"description": "sum of:",
"details": [
{
"value": 0.25316024,
"description": "weight(test_field:test in 0) [PerFieldSimilarity], result of:",
"details": [
{
"value": 0.25316024,
"description": "score(doc=0,freq=1.0 = termFreq=1.0\n), product of:",
"details": [
{
"value": 0.2876821,
"description": "idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:",
"details": [
{
"value": ,
"description": "docFreq",
"details": []
},
{
"value": ,
"description": "docCount",
"details": []
}
]
},
{
"value": 0.88,
"description": "tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:",
"details": [
{
"value": ,
"description": "termFreq=1.0",
"details": []
},
{
"value": 1.2,
"description": "parameter k1",
"details": []
},
{
"value": 0.75,
"description": "parameter b",
"details": []
},
{
"value": ,
"description": "avgFieldLength",
"details": []
},
{
"value": ,
"description": "fieldLength",
"details": []
}
]
}
]
}
]
}
]
},
{
"value": ,
"description": "match on required clause, product of:",
"details": [
{
"value": ,
"description": "# clause",
"details": []
},
{
"value": ,
"description": "*:*, product of:",
"details": [
{
"value": ,
"description": "boost",
"details": []
},
{
"value": ,
"description": "queryNorm",
"details": []
}
]
}
]
}
]
}
}
]
}
}
doc value正排索引
在我们搜索的时候，要依靠倒排索引，但是当我们排序的时候，需要依靠正排索引。通过倒排索引锁定文档document之后，看到每个document的每个field，然后进行排序，所谓的正排索引就是doc values。
对于ES而言，在建立索引的时候，一方面会建立倒排索引，以供搜索使用；一方面会建立正排索引，也就是doc values,以供排序，聚合，过滤等使用。
doc values是被保存在磁盘上的，此时如果内存足够，OS操作系统会自动将其缓存在内存中，性能还是会很高的，如果内存不够用，OS操作系统会将其写入磁盘。
下面举个例子描述正排索引和倒排索引
假设某个index有两个doc
doc1 : hello world you and me
doc2 : hi world, how are you
建立倒排索引
word doc1 doc2
hello *
world * *
you * *
and *
me *
hi *
how *
are *
假设某个index有两个doc
doc1: {"name": "jack", "age": }
doc2: {"name": "tom", "age": }
建立正排索引
document name age
doc1 jack
doc2 tom
解密query、fetch phrase原理
query pharse
基本原理：
- 搜索请求发送到某一个coordinate node协调节点，会构建一个priority queue，长度以paging操作from和size为准，默认是10
- coordinate node将请求转发到所有的shard,每个shard本地搜索，并构建一个本地的priority queue
- 各个shard将自己的priority queue返回给coordinate node,并构建一个全局的priority queue
fetch phrase
基本原理：
- coordinate node协调节点构建完priority queue之后，就发送mget请求去所有shard上获取对应的document
- 各个shard将document返回给coordinate node
- coordinate node将合并后的document结果返回给客户端。
也就是ES的query pharse是根据priority queue去构建搜索结果的
示例
比如总共有60000条数据，三个primary shard,每个shard上分了20000条数据，每页是10条数据，这个时候，你要搜索到第1000页，实际上要拿到的是10001~10010，也就是会构建一个10010大小的priority queue。
注意这里千万不要理解成每个shard都是返回10条数据。这样理解是错误的！
下面做一下详细的分析：
请求首先可能是打到一个不包含这个index的shard的node上去，这个node就是一个协调节点coordinate node，那么这个coordinate node就会将搜索请求转发到index的三个shard所在的node上去。比如说我们之前说的情况下，要搜索60000条数据中的第1000页，实际上每个shard都要将内部的20000条数据中的第10001~10010条数据，拿出来，不是才10条，是10010条数据。3个shard的每个shard都返回10010条数据给协调节点coordinate node，coordinate node会收到总共30030条数据,此时会构建一个30030大小的priority queue，然后在这些数据中进行排序，根据_score相关度分数，然后取到10001~10010这10条数据，就是我们要的第1000页的10条数据。
如下图所示：
Bouncing Results问题
想象一下有两个文档有同样值的时间戳字段，搜索结果用 timestamp 字段来排序。 由于搜索请求是在所有有效的分片副本间轮询的，那就有可能发生主分片处理请求时，这两个文档是一种顺序， 而副本分片处理请求时又是另一种顺序。
-
bouncing results 问题:：每次用户刷新页面，搜索结果表现是不同的顺序。 让同一个用户始终使用同一个分片，这样可以避免这种问题， 可以设置 preference 参数为一个特定的任意值比如用户会话ID来解决。
偏好这个参数 preference 允许 用来控制由哪些分片或节点来处理搜索请求。 它接受像 _primary, _primary_first, _local, _only_node:xyz, _prefer_node:xyz, 和 _shards:2,3 这样的值, 这些值在 search preference 文档页面被详细解释。
但是最有用的值是某些随机字符串，它可以避免 bouncing results 问题。 - timeout：已经讲解过原理了，主要就是限定在一定时间内，将部分获取到的数据直接返回，避免查询耗时过长
- routing：document文档路由，_id路由，routing=user_id，这样的话可以让同一个user对应的数据到一个shard上去
- search_type：默认default：query_then_fetch，dfs_query_then_fetch可以提升revelance sort精准度
基于scoll技术滚动搜索大量数据
在实际应用中，通过from+size不可避免会出现深分页的瓶颈，那么通过scoll技术就是一个很好的解决深分页的方法。比如如果我们一次性要查出10万条数据，那么使用from+size很显然性能会非常的差，priority queue会非常的大。此时如果采用scroll滚动查询，就可以一批一批的查，直到所有数据都查询完。
scroll原理
scoll搜索会在第一次搜索的时候，保存一个当时的视图快照，之后只会基于该旧的视图快照提供数据搜索，如果这个期间数据变更，是不会让用户看到的。而且ES内部是基于_doc进行排序的方式，性能较高。
示例：
# 使用scroll
POST /test_index/_search?scroll=1m
{
"query": {
"match_all": {}
},
"sort": [
"_doc"
],
"size":
}
获取到scroll_id
{
"_scroll_id": "DnF1ZXJ5VGhlbkZldGNoBQAAAAAAAI-sFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACPqxYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3AAAAAAAAj68WMUxkcUxGcXhRUXE0eGcyTXBoSV9ndwAAAAAAAI-tFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACPrhYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3",
"took": ,
"timed_out": false,
"_shards": {
"total": ,
"successful": ,
"failed":
},
"hits": {
"total": ,
"max_score": null,
"hits": [
{
"_index": "test_index",
"_type": "test_type",
"_id": "AWypxxLYFCl_S-ox4wvd",
"_score": null,
"_source": {
"test_content": "my test"
},
"sort": [
]
},
{
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": null,
"_source": {
"test_field": "test test"
},
"sort": [
]
},
{
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": null,
"_source": {
"test_field": "test client 1"
},
"sort": [
]
}
]
}
}
滚动搜索
# 滚动搜索
POST _search/scroll
{
"scroll":"1m",
"scroll_id":"DnF1ZXJ5VGhlbkZldGNoBQAAAAAAAJDMFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACQzRYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3AAAAAAAAkM8WMUxkcUxGcXhRUXE0eGcyTXBoSV9ndwAAAAAAAJDOFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACQ0BYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3"
}
搜索结果
{
"_scroll_id": "DnF1ZXJ5VGhlbkZldGNoBQAAAAAAAJDMFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACQzRYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3AAAAAAAAkM8WMUxkcUxGcXhRUXE0eGcyTXBoSV9ndwAAAAAAAJDOFjFMZHFMRnF4UVFxNHhnMk1waElfZ3cAAAAAAACQ0BYxTGRxTEZxeFFRcTR4ZzJNcGhJX2d3",
"took": ,
"timed_out": false,
"terminated_early": true,
"_shards": {
"total": ,
"successful": ,
"failed":
},
"hits": {
"total": ,
"max_score": null,
"hits": [
{
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": null,
"_source": {
"num": ,
"tags": []
},
"sort": [
]
},
{
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": null,
"_source": {
"test_field": "test client 2"
},
"sort": [
]
},
{
"_index": "test_index",
"_type": "test_type",
"_id": "",
"_score": null,
"_source": {
"test_field": "test4"
},
"sort": [
]
}
]
}
}