在 Elasticsearch 5.x 有一个字段折叠（Field Collapsing，#22337）的功能非常有意思，在这里分享一下，
字段折叠是一个很有历史的需求了，可以看这个 issue，编号#256，最初是2010年7月提的issue，也是讨论最多的帖子之一（240+评论），熬了6年才支持的特性，你说牛不牛，哈哈。
目测该特性将于5.3发布，尝鲜地址：Elasticsearch-5.3.0-SNAPSHOT，文档地址：search-request-collapse。
So，什么是字段折叠，可以理解就是按特定字段进行合并去重，比如我们有一个菜谱搜索，我希望按菜谱的“菜系”字段进行折叠，即返回结果每个菜系都返回一个结果，也就是按菜系去重，我搜索关键字“鱼”，要去返回的结果里面各种菜系都有，有湘菜，有粤菜，有中餐，有西餐，别全是湘菜，就是这个意思，通过按特定字段折叠之后，来丰富搜索结果的多样性。
说到这里，有人肯定会想到，使用 term agg+ top hits agg 来实现啊，这种组合两种聚和的方式可以实现上面的功能，不过也有一些局限性，比如，不能分页，#4915；结果不够精确（top term+top hits，es 的聚合实现选择了牺牲精度来提高速度）；数据量大的情况下，聚合比较慢，影响搜索体验。
而新的的字段折叠的方式是怎么实现的的呢，有这些要点：
- 折叠+取 inner_hits 分两阶段执行（组合聚合的方式只有一个阶段），所以 top hits 永远是精确的。
- 字段折叠只在 top hits 层执行，不需要每次都在完整的结果集上对为每个折叠主键计算实际的 doc values 值，只对 top hits 这小部分数据操作就可以，和 term agg 相比要节省很多内存。
- 因为只在 top hits 上进行折叠，所以相比组合聚合的方式，速度要快很多。
- 折叠 top docs 不需要使用全局序列（global ordinals）来转换 string，相比 agg 这也节省了很多内存。
- 分页成为可能，和常规搜索一样，具有相同的局限，先获取 from+size 的内容，再合并。
- search_after 和 scroll 暂未实现，不过具备可行性。
- 折叠只影响搜索结果，不影响聚合，搜索结果的 total 是所有的命中纪录数，去重的结果数未知（无法计算）。
下面来看看具体的例子，就知道怎么回事了，使用起来很简单。
- 先准备索引和数据，这里以菜谱为例，name：菜谱名，type 为菜系，rating 为用户的累积平均评分
DELETE recipes
PUT recipes
POST recipes/type/_mapping
{
"properties": {
"name":{
"type": "text"
},
"rating":{
"type": "float"
},"type":{
"type": "keyword"
}
}
}
POST recipes/type/
{
"name":"清蒸鱼头","rating":1,"type":"湘菜"
}
POST recipes/type/
{
"name":"剁椒鱼头","rating":2,"type":"湘菜"
}
POST recipes/type/
{
"name":"红烧鲫鱼","rating":3,"type":"湘菜"
}
POST recipes/type/
{
"name":"鲫鱼汤（辣）","rating":3,"type":"湘菜"
}
POST recipes/type/
{
"name":"鲫鱼汤（微辣）","rating":4,"type":"湘菜"
}
POST recipes/type/
{
"name":"鲫鱼汤（变态辣）","rating":5,"type":"湘菜"
}
POST recipes/type/
{
"name":"广式鲫鱼汤","rating":5,"type":"粤菜"
}
POST recipes/type/
{
"name":"鱼香肉丝","rating":2,"type":"川菜"
}
POST recipes/type/
{
"name":"奶油鲍鱼汤","rating":2,"type":"西菜"
} - 现在我们看看普通的查询效果是怎么样的，搜索关键字带“鱼”的菜，返回3条数据
POST recipes/type/_search
{
"query": {"match": {
"name": "鱼"
}},"size": 3
} {
"took": 2,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": 0.26742277,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYF_OA-dG63Txsd",
"_score": 0.26742277,
"_source": {
"name": "鲫鱼汤（变态辣）",
"rating": 5,
"type": "湘菜"
}
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHXO_OA-dG63Txsa",
"_score": 0.19100356,
"_source": {
"name": "红烧鲫鱼",
"rating": 3,
"type": "湘菜"
}
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHWy_OA-dG63TxsZ",
"_score": 0.19100356,
"_source": {
"name": "剁椒鱼头",
"rating": 2,
"type": "湘菜"
}
}
]
}
}POST recipes/type/_search
{
"query": {"match": {
"name": "鱼"
}},"sort": [
{
"rating": {
"order": "desc"
}
}
],"size": 3
} {
"took": 1,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYF_OA-dG63Txsd",
"_score": null,
"_source": {
"name": "鲫鱼汤（变态辣）",
"rating": 5,
"type": "湘菜"
},
"sort": [
5
]
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYW_OA-dG63Txse",
"_score": null,
"_source": {
"name": "广式鲫鱼汤",
"rating": 5,
"type": "粤菜"
},
"sort": [
5
]
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHX7_OA-dG63Txsc",
"_score": null,
"_source": {
"name": "鲫鱼汤（微辣）",
"rating": 4,
"type": "湘菜"
},
"sort": [
4
]
}
]
}
}
GET recipes/type/_search
{
"query": {
"match": {
"name": "鱼"
}
},
"sort": [
{
"rating": {
"order": "desc"
}
}
],"aggs": {
"type": {
"terms": {
"field": "type",
"size": 10
},"aggs": {
"rated": {
"top_hits": {
"sort": [{
"rating": {"order": "desc"}
}],
"size": 1
}
}
}
}
},
"size": 0,
"from": 0
} {
"took": 4,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": 0,
"hits": []
},
"aggregations": {
"type": {
"doc_count_error_upper_bound": 0,
"sum_other_doc_count": 0,
"buckets": [
{
"key": "湘菜",
"doc_count": 6,
"rated": {
"hits": {
"total": 6,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYF_OA-dG63Txsd",
"_score": null,
"_source": {
"name": "鲫鱼汤（变态辣）",
"rating": 5,
"type": "湘菜"
},
"sort": [
5
]
}
]
}
}
},
{
"key": "川菜",
"doc_count": 1,
"rated": {
"hits": {
"total": 1,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYr_OA-dG63Txsf",
"_score": null,
"_source": {
"name": "鱼香肉丝",
"rating": 2,
"type": "川菜"
},
"sort": [
2
]
}
]
}
}
},
{
"key": "粤菜",
"doc_count": 1,
"rated": {
"hits": {
"total": 1,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYW_OA-dG63Txse",
"_score": null,
"_source": {
"name": "广式鲫鱼汤",
"rating": 5,
"type": "粤菜"
},
"sort": [
5
]
}
]
}
}
},
{
"key": "西菜",
"doc_count": 1,
"rated": {
"hits": {
"total": 1,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHY3_OA-dG63Txsg",
"_score": null,
"_source": {
"name": "奶油鲍鱼汤",
"rating": 2,
"type": "西菜"
},
"sort": [
2
]
}
]
}
}
}
]
}
}
}
GET recipes/type/_search
{
"query": {
"match": {
"name": "鱼"
}
},
"collapse": {
"field": "type"
},
"size": 3,
"from": 0
}{
"took": 1,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoDNlRJ_OA-dG63TxpW",
"_score": 0.018980097,
"_source": {
"name": "鲫鱼汤（微辣）",
"rating": 4,
"type": "湘菜"
},
"fields": {
"type": [
"湘菜"
]
}
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoDNlRk_OA-dG63TxpZ",
"_score": 0.013813315,
"_source": {
"name": "鱼香肉丝",
"rating": 2,
"type": "川菜"
},
"fields": {
"type": [
"川菜"
]
}
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoDNlRb_OA-dG63TxpY",
"_score": 0.0125863515,
"_source": {
"name": "广式鲫鱼汤",
"rating": 5,
"type": "粤菜"
},
"fields": {
"type": [
"粤菜"
]
}
}
]
}
}{
"took": 1,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoDNlRw_OA-dG63Txpa",
"_score": 0.012546891,
"_source": {
"name": "奶油鲍鱼汤",
"rating": 2,
"type": "西菜"
},
"fields": {
"type": [
"西菜"
]
}
}
]
}
}GET recipes/type/_search
{
"query": {
"match": {
"name": "鱼"
}
},
"collapse": {
"field": "type",
"inner_hits": {
"name": "top_rated",
"size": 2,
"sort": [
{
"rating": "desc"
}
]
}
},
"sort": [
{
"rating": {
"order": "desc"
}
}
],
"size": 2,
"from": 0
}{
"took": 1,
"timed_out": false,
"_shards": {
"total": 5,
"successful": 5,
"failed": 0
},
"hits": {
"total": 9,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYF_OA-dG63Txsd",
"_score": null,
"_source": {
"name": "鲫鱼汤（变态辣）",
"rating": 5,
"type": "湘菜"
},
"fields": {
"type": [
"湘菜"
]
},
"sort": [
5
],
"inner_hits": {
"top_rated": {
"hits": {
"total": 6,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYF_OA-dG63Txsd",
"_score": null,
"_source": {
"name": "鲫鱼汤（变态辣）",
"rating": 5,
"type": "湘菜"
},
"sort": [
5
]
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHX7_OA-dG63Txsc",
"_score": null,
"_source": {
"name": "鲫鱼汤（微辣）",
"rating": 4,
"type": "湘菜"
},
"sort": [
4
]
}
]
}
}
}
},
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYW_OA-dG63Txse",
"_score": null,
"_source": {
"name": "广式鲫鱼汤",
"rating": 5,
"type": "粤菜"
},
"fields": {
"type": [
"粤菜"
]
},
"sort": [
5
],
"inner_hits": {
"top_rated": {
"hits": {
"total": 1,
"max_score": null,
"hits": [
{
"_index": "recipes",
"_type": "type",
"_id": "AVoESHYW_OA-dG63Txse",
"_score": null,
"_source": {
"name": "广式鲫鱼汤",
"rating": 5,
"type": "粤菜"
},
"sort": [
5
]
}
]
}
}
}
}
]
}
}
[尊重社区原创，转载请保留或注明出处]
本文地址：http://elasticsearch.cn/article/132
本文地址：http://elasticsearch.cn/article/132
27 个评论
这个功能真心实用
这个功能对性能有影响吗 不过真的很实用 赞
千呼万唤始出来！good！
老司机，5.3加了这个collapse之后。如何统计出折叠之后的总数呢？例如你这个例子里如何知道折叠后共有4个菜系的。
根据github上的issue，现在还不支持条数统计：
For the record, the number of groups after collapsing is not computed since it would be too costly. You can run a cardinality aggregation to get this information.
不过就算通过聚合，得到的结果也是不准确的，因为不知道每个菜系匹配了多少道菜
For the record, the number of groups after collapsing is not computed since it would be too costly. You can run a cardinality aggregation to get this information.
不过就算通过聚合，得到的结果也是不准确的，因为不知道每个菜系匹配了多少道菜
刚好要做去重，特意升级到5.3.1试了下，效果很棒
collapse非常好，不过针对text的类型就无法适用。只能是keywords、numbers类型字段。像搜索商品后，对商品的名称去重这类的场景就不是特别适用。
有点像数据库里面的row_number() over()，这个特性还是很赞的
Field Collapsing只支持单个字段字段，如果能支持多个字段就非常完美了
这个功能很实用！
想问下有对应的java示例么？感觉CollapseBuilder比较难用
Hi 大神，目前我的index里面的document长成这样： {building : {id: 1, address: "xxx", offices: [{officeId: "1", officeType: "L"}, {officeId: "2", officeType: "S"}]}. 现在我们想先按照officeType来进行group，得出的结果像： {building : {id: 2, address: "xxx", offices: [{officeId: "3", officeType: "M", price: 100}, {officeId: "5", officeType: "M", price: 1000}]},然后再在每个group里面给office按照price从大到小排序。 这种情况我可以怎么做啊？
hits的total仍然是总数啊，不是去重后的总数，怎么做分页？提前用cardinality再查一次吗？还是有更好的方法
用下拉分页的方式可以不考虑总数,考虑总数,速度就慢了.
哈哈哈这个感觉有点意思
感谢大佬科普，目前刚好有需求
collapse怎么按多个字段折叠呢？想实现按三个字段的groupby效果
