Neo4j Cypher：当基数很高时如何优化 NOT EXISTS 查询

Question

当发帖的基数 b/w 用户约为 8000（一个用户查看约 8000 个帖子）时，以下查询占用 1 秒以上的时间和消费者约 7 MB。由于高且线性增长的延迟和内存消耗，很难对其进行扩展。是否有可能对此进行不同的建模and/or 优化查询？

查询

PROFILE MATCH (u:User)-[:CREATED]->(p:Post) WHERE NOT (:User{ID: 2})-[:VIEWED]->(p) RETURN p.ID

计划

| Plan      | Statement   | Version      | Planner | Runtime       | Time | DbHits  | Rows | Memory (Bytes) |
+-----------------------------------------------------------------------------------------------------------+
| "PROFILE" | "READ_ONLY" | "CYPHER 4.1" | "COST"  | "INTERPRETED" | 1033 | 3721750 | 10   | 6696240        |
+-----------------------------------------------------------------------------------------------------------+


+------------------------------+-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| Operator                     | Details                                       | Estimated Rows | Rows | DB Hits | Cache H/M | Memory (Bytes) | Ordered by     |
+------------------------------+-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +ProduceResults@neo4j        | `p.ID`                                        |           2158 |   10 |       0 |       0/0 |                |                |
| |                            +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +Projection@neo4j            | p.ID AS `p.ID`                                |           2158 |   10 |      10 |       0/0 |                |                |
| |                            +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +Filter@neo4j                | u:User                                        |           2158 |   10 |      10 |       0/0 |                |                |
| |                            +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +Expand(All)@neo4j           | (p)<-[anon_15:CREATED]-(u)                    |           2158 |   10 |      20 |       0/0 |                |                |
| |                            +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +AntiSemiApply@neo4j         |                                               |           2158 |   10 |       0 |       0/0 |                |                |
| |\                           +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| | +Expand(Into)@neo4j        | (anon_47)-[anon_61:VIEWED]->(p)               |            233 |    0 | 3695819 |       0/0 |        6696240 | anon_47.ID ASC |
| | |                          +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| | +NodeUniqueIndexSeek@neo4j | UNIQUE anon_47:User(ID) WHERE ID = $autoint_0 |           8630 | 8630 |   17260 |       0/0 |                | anon_47.ID ASC |
| |                            +-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+
| +NodeByLabelScan@neo4j       | p:Post                                        |           8630 | 8630 |    8631 |       0/0 |                |                |
+------------------------------+-----------------------------------------------+----------------+------+---------+-----------+----------------+----------------+

Answer 1

是的，这可以改进。

首先，让我们了解一下这是做什么的。

首先，它从 NodeByLabelScan 开始。有道理，没法回避。

但是，对于标签的每个节点（以下执行 PER ROW！），它与用户 2 匹配，并展开所有 :VIEWED 来自用户 2 的关系，以查看它们是否是 post 对于那个特定的行。

你能看出为什么这样效率低下吗？按照PROFILE计划有8630个post节点，所以用户2被索引查找了8630次，他们的:VIEWED关系被展开了8630次。为什么是8630次？因为这是每个 :Post 节点发生的。

相反，试试这个：

MATCH (:User{ID: 2})-[:VIEWED]->(viewedPost)
WITH collect(viewedPost) as viewedPosts
MATCH (:User)-[:CREATED]->(p:Post) 
WHERE NOT p IN viewedPosts 
RETURN p.ID

这会稍微改变一下。

首先它与用户 2 的查看 post 匹配（查找和扩展只执行一次），然后收集那些查看的 post。

然后它将进行标签扫描，并进行过滤，使 post 不在已查看 post 的集合中。

Neo4j Cypher：当基数很高时如何优化 NOT EXISTS 查询

Neo4j Cypher: How to optimize a NOT EXISTS Query when cardinality is high

neo4j

cypher