“N+1选择问题”在对象关系映射(ORM)讨论中通常被称为一个问题,我理解这与必须为对象世界中看似简单的东西进行大量数据库查询有关。

有人对这个问题有更详细的解释吗?


当前回答

SELECT 
table1.*
, table2.*
INNER JOIN table2 ON table2.SomeFkId = table1.SomeId

这将获得一个结果集,其中表2中的子行通过返回表2中每个子行的表1结果而导致重复。O/R映射器应根据唯一的键字段区分表1实例,然后使用所有表2列填充子实例。

SELECT table1.*

SELECT table2.* WHERE SomeFkId = #

N+1是第一个查询填充主对象,第二个查询填充返回的每个唯一主对象的所有子对象的位置。

考虑:

class House
{
    int Id { get; set; }
    string Address { get; set; }
    Person[] Inhabitants { get; set; }
}

class Person
{
    string Name { get; set; }
    int HouseId { get; set; }
}

以及具有类似结构的表格。对地址“22 Valley St”的单个查询可能返回:

Id Address      Name HouseId
1  22 Valley St Dave 1
1  22 Valley St John 1
1  22 Valley St Mike 1

O/RM应该用ID=1,Address=“22 Valley St”填充Home的实例,然后用Dave、John和Mike的People实例填充Inhabitants数组,只需一个查询。

对上述相同地址的N+1查询将导致:

Id Address
1  22 Valley St

使用单独的查询,如

SELECT * FROM Person WHERE HouseId = 1

并产生单独的数据集,如

Name    HouseId
Dave    1
John    1
Mike    1

并且最终结果与上述单个查询相同。

单一选择的优点是您可以提前获得所有数据,这可能是您最终想要的。N+1的优点是减少了查询复杂性,并且可以使用延迟加载,其中子结果集仅在第一次请求时加载。

其他回答

N+1 SELECT问题真的很难发现,尤其是在具有大型域的项目中,当它开始降低性能时。即使问题得到解决,即通过添加紧急加载,进一步的开发可能会破坏解决方案和/或在其他地方再次引入N+1 SELECT问题。

我创建了开源库jplusone来解决基于JPA的Spring Boot Java应用程序中的这些问题。该库提供两个主要功能:

生成将SQL语句与触发它们的JPA操作的执行相关联的报告,并将其放置在应用程序的源代码中

2020-10-22 18:41:43.236 DEBUG 14913 --- [           main] c.a.j.core.report.ReportGenerator        :
    ROOT
        com.adgadev.jplusone.test.domain.bookshop.BookshopControllerTest.shouldGetBookDetailsLazily(BookshopControllerTest.java:65)
        com.adgadev.jplusone.test.domain.bookshop.BookshopController.getSampleBookUsingLazyLoading(BookshopController.java:31)
        com.adgadev.jplusone.test.domain.bookshop.BookshopService.getSampleBookDetailsUsingLazyLoading [PROXY]
            SESSION BOUNDARY
                OPERATION [IMPLICIT]
                    com.adgadev.jplusone.test.domain.bookshop.BookshopService.getSampleBookDetailsUsingLazyLoading(BookshopService.java:35)
                    com.adgadev.jplusone.test.domain.bookshop.Author.getName [PROXY]
                    com.adgadev.jplusone.test.domain.bookshop.Author [FETCHING ENTITY]
                        STATEMENT [READ]
                            select [...] from
                                author author0_
                                left outer join genre genre1_ on author0_.genre_id=genre1_.id
                            where
                                author0_.id=1
                OPERATION [IMPLICIT]
                    com.adgadev.jplusone.test.domain.bookshop.BookshopService.getSampleBookDetailsUsingLazyLoading(BookshopService.java:36)
                    com.adgadev.jplusone.test.domain.bookshop.Author.countWrittenBooks(Author.java:53)
                    com.adgadev.jplusone.test.domain.bookshop.Author.books [FETCHING COLLECTION]
                        STATEMENT [READ]
                            select [...] from
                                book books0_
                            where
                                books0_.author_id=1

提供API,允许编写测试,检查应用程序使用JPA的效率(即断言延迟加载操作的数量)

@SpringBootTest
class LazyLoadingTest {

    @Autowired
    private JPlusOneAssertionContext assertionContext;

    @Autowired
    private SampleService sampleService;

    @Test
    public void shouldBusinessCheckOperationAgainstJPlusOneAssertionRule() {
        JPlusOneAssertionRule rule = JPlusOneAssertionRule
                .within().lastSession()
                .shouldBe().noImplicitOperations().exceptAnyOf(exclusions -> exclusions
                        .loadingEntity(Author.class).times(atMost(2))
                        .loadingCollection(Author.class, "books")
                );

        // trigger business operation which you wish to be asserted against the rule,
        // i.e. calling a service or sending request to your API controller
        sampleService.executeBusinessOperation();

        rule.check(assertionContext);
    }
}
SELECT 
table1.*
, table2.*
INNER JOIN table2 ON table2.SomeFkId = table1.SomeId

这将获得一个结果集,其中表2中的子行通过返回表2中每个子行的表1结果而导致重复。O/R映射器应根据唯一的键字段区分表1实例,然后使用所有表2列填充子实例。

SELECT table1.*

SELECT table2.* WHERE SomeFkId = #

N+1是第一个查询填充主对象,第二个查询填充返回的每个唯一主对象的所有子对象的位置。

考虑:

class House
{
    int Id { get; set; }
    string Address { get; set; }
    Person[] Inhabitants { get; set; }
}

class Person
{
    string Name { get; set; }
    int HouseId { get; set; }
}

以及具有类似结构的表格。对地址“22 Valley St”的单个查询可能返回:

Id Address      Name HouseId
1  22 Valley St Dave 1
1  22 Valley St John 1
1  22 Valley St Mike 1

O/RM应该用ID=1,Address=“22 Valley St”填充Home的实例,然后用Dave、John和Mike的People实例填充Inhabitants数组,只需一个查询。

对上述相同地址的N+1查询将导致:

Id Address
1  22 Valley St

使用单独的查询,如

SELECT * FROM Person WHERE HouseId = 1

并产生单独的数据集,如

Name    HouseId
Dave    1
John    1
Mike    1

并且最终结果与上述单个查询相同。

单一选择的优点是您可以提前获得所有数据,这可能是您最终想要的。N+1的优点是减少了查询复杂性,并且可以使用延迟加载,其中子结果集仅在第一次请求时加载。

N+1查询问题是什么

当数据访问框架执行N个额外的SQL语句以获取执行主SQL查询时可能检索到的相同数据时,就会出现N+1查询问题。

N值越大,执行的查询越多,性能影响越大。而且,与可以帮助您查找运行缓慢的查询的慢速查询日志不同,N+1问题不会出现,因为每个单独的附加查询运行速度都足够快,不会触发慢速查询日志。

问题是执行大量额外的查询,总的来说,这些查询需要足够的时间来降低响应时间。

让我们考虑以下post和post_comments数据库表,它们形成了一对多的表关系:

我们将创建以下4个柱行:

INSERT INTO post (title, id)
VALUES ('High-Performance Java Persistence - Part 1', 1)
 
INSERT INTO post (title, id)
VALUES ('High-Performance Java Persistence - Part 2', 2)
 
INSERT INTO post (title, id)
VALUES ('High-Performance Java Persistence - Part 3', 3)
 
INSERT INTO post (title, id)
VALUES ('High-Performance Java Persistence - Part 4', 4)

此外,我们还将创建4个post_comment子记录:

INSERT INTO post_comment (post_id, review, id)
VALUES (1, 'Excellent book to understand Java Persistence', 1)
 
INSERT INTO post_comment (post_id, review, id)
VALUES (2, 'Must-read for Java developers', 2)
 
INSERT INTO post_comment (post_id, review, id)
VALUES (3, 'Five Stars', 3)
 
INSERT INTO post_comment (post_id, review, id)
VALUES (4, 'A great reference book', 4)

普通SQL的N+1查询问题

如果使用此SQL查询选择post_comments:

List<Tuple> comments = entityManager.createNativeQuery("""
    SELECT
        pc.id AS id,
        pc.review AS review,
        pc.post_id AS postId
    FROM post_comment pc
    """, Tuple.class)
.getResultList();

稍后,您决定获取每个post_comment的相关文章标题:

for (Tuple comment : comments) {
    String review = (String) comment.get("review");
    Long postId = ((Number) comment.get("postId")).longValue();
 
    String postTitle = (String) entityManager.createNativeQuery("""
        SELECT
            p.title
        FROM post p
        WHERE p.id = :postId
        """)
    .setParameter("postId", postId)
    .getSingleResult();
 
    LOGGER.info(
        "The Post '{}' got this review '{}'",
        postTitle,
        review
    );
}

您将触发N+1查询问题,因为您执行了5(1+4)而不是一个SQL查询:

SELECT
    pc.id AS id,
    pc.review AS review,
    pc.post_id AS postId
FROM post_comment pc
 
SELECT p.title FROM post p WHERE p.id = 1
-- The Post 'High-Performance Java Persistence - Part 1' got this review
-- 'Excellent book to understand Java Persistence'
    
SELECT p.title FROM post p WHERE p.id = 2
-- The Post 'High-Performance Java Persistence - Part 2' got this review
-- 'Must-read for Java developers'
     
SELECT p.title FROM post p WHERE p.id = 3
-- The Post 'High-Performance Java Persistence - Part 3' got this review
-- 'Five Stars'
     
SELECT p.title FROM post p WHERE p.id = 4
-- The Post 'High-Performance Java Persistence - Part 4' got this review
-- 'A great reference book'

修复N+1查询问题非常简单。您只需提取原始SQL查询中所需的所有数据,如下所示:

List<Tuple> comments = entityManager.createNativeQuery("""
    SELECT
        pc.id AS id,
        pc.review AS review,
        p.title AS postTitle
    FROM post_comment pc
    JOIN post p ON pc.post_id = p.id
    """, Tuple.class)
.getResultList();
 
for (Tuple comment : comments) {
    String review = (String) comment.get("review");
    String postTitle = (String) comment.get("postTitle");
 
    LOGGER.info(
        "The Post '{}' got this review '{}'",
        postTitle,
        review
    );
}

这次,只执行一个SQL查询来获取我们进一步感兴趣的所有数据。

JPA和Hibernate的N+1查询问题

在使用JPA和Hibernate时,有几种方法可以触发N+1查询问题,因此了解如何避免这些情况非常重要。

对于下一个示例,考虑我们将post和post_comments表映射到以下实体:

JPA映射如下所示:

@Entity(name = "Post")
@Table(name = "post")
public class Post {
 
    @Id
    private Long id;
 
    private String title;
 
    //Getters and setters omitted for brevity
}
 
@Entity(name = "PostComment")
@Table(name = "post_comment")
public class PostComment {
 
    @Id
    private Long id;
 
    @ManyToOne
    private Post post;
 
    private String review;
 
    //Getters and setters omitted for brevity
}

获取类型.EAGER

隐式或显式地为JPA关联使用FetchType.EAGER是一个坏主意,因为您将获取更多所需的数据。此外,FetchType.EAGER策略还容易出现N+1个查询问题。

不幸的是,@ManyToOne和@OneToOne关联默认使用FetchType.EAGER,因此如果映射如下所示:

@ManyToOne
private Post post;

您使用的是FetchType.EAGER策略,每当您在使用JPQL或Criteria API查询加载某些PostComment实体时忘记使用JOIN FETCH时:

List<PostComment> comments = entityManager
.createQuery("""
    select pc
    from PostComment pc
    """, PostComment.class)
.getResultList();

您将触发N+1查询问题:

SELECT 
    pc.id AS id1_1_, 
    pc.post_id AS post_id3_1_, 
    pc.review AS review2_1_ 
FROM 
    post_comment pc

SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 1
SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 2
SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 3
SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 4

请注意执行的其他SELECT语句,因为在返回PostComment实体列表之前必须获取post关联。

与调用EntityManager的find方法时使用的默认获取计划不同,JPQL或Criteria API查询定义了Hibernate无法通过自动注入JOIN fetch来更改的显式计划。因此,您需要手动执行。

如果你根本不需要post关联,那么你在使用FetchType.EAGER时就不走运了,因为无法避免获取它。这就是为什么默认情况下最好使用FetchType.LAZY。

但是,如果您想使用后关联,那么可以使用JOIN FETCH来避免N+1查询问题:

List<PostComment> comments = entityManager.createQuery("""
    select pc
    from PostComment pc
    join fetch pc.post p
    """, PostComment.class)
.getResultList();

for(PostComment comment : comments) {
    LOGGER.info(
        "The Post '{}' got this review '{}'", 
        comment.getPost().getTitle(), 
        comment.getReview()
    );
}

这次,Hibernate将执行一条SQL语句:

SELECT 
    pc.id as id1_1_0_, 
    pc.post_id as post_id3_1_0_, 
    pc.review as review2_1_0_, 
    p.id as id1_0_1_, 
    p.title as title2_0_1_ 
FROM 
    post_comment pc 
INNER JOIN 
    post p ON pc.post_id = p.id
    
-- The Post 'High-Performance Java Persistence - Part 1' got this review 
-- 'Excellent book to understand Java Persistence'

-- The Post 'High-Performance Java Persistence - Part 2' got this review 
-- 'Must-read for Java developers'

-- The Post 'High-Performance Java Persistence - Part 3' got this review 
-- 'Five Stars'

-- The Post 'High-Performance Java Persistence - Part 4' got this review 
-- 'A great reference book'

获取类型.LAZY

即使您切换到对所有关联显式使用FetchType.LAZY,您仍然会遇到N+1问题。

这一次,后关联映射如下:

@ManyToOne(fetch = FetchType.LAZY)
private Post post;

现在,当您获取PostComment实体时:

List<PostComment> comments = entityManager
.createQuery("""
    select pc
    from PostComment pc
    """, PostComment.class)
.getResultList();

Hibernate将执行一条SQL语句:

SELECT 
    pc.id AS id1_1_, 
    pc.post_id AS post_id3_1_, 
    pc.review AS review2_1_ 
FROM 
    post_comment pc

但是,如果之后,您将引用延迟加载的post关联:

for(PostComment comment : comments) {
    LOGGER.info(
        "The Post '{}' got this review '{}'", 
        comment.getPost().getTitle(), 
        comment.getReview()
    );
}

您将获得N+1查询问题:

SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 1
-- The Post 'High-Performance Java Persistence - Part 1' got this review 
-- 'Excellent book to understand Java Persistence'

SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 2
-- The Post 'High-Performance Java Persistence - Part 2' got this review 
-- 'Must-read for Java developers'

SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 3
-- The Post 'High-Performance Java Persistence - Part 3' got this review 
-- 'Five Stars'

SELECT p.id AS id1_0_0_, p.title AS title2_0_0_ FROM post p WHERE p.id = 4
-- The Post 'High-Performance Java Persistence - Part 4' got this review 
-- 'A great reference book'

由于后期关联是延迟获取的,因此在访问延迟关联时将执行一个辅助SQL语句,以便生成日志消息。

同样,修复方法包括在JPQL查询中添加JOIN FETCH子句:

List<PostComment> comments = entityManager.createQuery("""
    select pc
    from PostComment pc
    join fetch pc.post p
    """, PostComment.class)
.getResultList();

for(PostComment comment : comments) {
    LOGGER.info(
        "The Post '{}' got this review '{}'", 
        comment.getPost().getTitle(), 
        comment.getReview()
    );
}

而且,就像FetchType.EAGER示例中一样,这个JPQL查询将生成一个SQL语句。

即使您正在使用FetchType.LAZY,并且没有引用双向@OneToOne JPA关系的子关联,您仍然可以触发N+1查询问题。

如何自动检测N+1查询问题

如果您想在数据访问层中自动检测N+1查询问题,可以使用db-util开源项目。

首先,您需要添加以下Maven依赖项:

<dependency>
    <groupId>com.vladmihalcea</groupId>
    <artifactId>db-util</artifactId>
    <version>${db-util.version}</version>
</dependency>

之后,您只需使用SQLStatementCountValidator实用程序来断言生成的底层SQL语句:

SQLStatementCountValidator.reset();

List<PostComment> comments = entityManager.createQuery("""
    select pc
    from PostComment pc
    """, PostComment.class)
.getResultList();

SQLStatementCountValidator.assertSelectCount(1);

如果您正在使用FetchType.EAGER并运行上述测试用例,则会出现以下测试用例失败:

SELECT 
    pc.id as id1_1_, 
    pc.post_id as post_id3_1_, 
    pc.review as review2_1_ 
FROM 
    post_comment pc

SELECT p.id as id1_0_0_, p.title as title2_0_0_ FROM post p WHERE p.id = 1

SELECT p.id as id1_0_0_, p.title as title2_0_0_ FROM post p WHERE p.id = 2


-- SQLStatementCountMismatchException: Expected 1 statement(s) but recorded 3 instead!

提供的链接有一个非常简单的n+1问题示例。如果你将它应用于Hibernate,它基本上是在谈论相同的事情。查询对象时,实体将被加载,但任何关联(除非另有配置)都将被延迟加载。因此,一个查询用于根对象,另一个查询加载每个根对象的关联。返回的100个对象意味着一个初始查询,然后是100个附加查询,以获得每个对象的关联,n+1。

http://pramatr.com/2009/02/05/sql-n-1-selects-explained/

因为这个问题,我们离开了Django的ORM。基本上,如果你尝试

for p in person:
    print p.car.colour

ORM将很高兴地返回所有人(通常作为Person对象的实例),但随后需要为每个Person查询car表。

一种简单且非常有效的方法是我称之为“扇形折叠”的方法,它避免了来自关系数据库的查询结果应该映射回组成查询的原始表的荒谬想法。

步骤1:宽选择

  select * from people_car_colour; # this is a view or sql function

这将返回类似

  p.id | p.name | p.telno | car.id | car.type | car.colour
  -----+--------+---------+--------+----------+-----------
  2    | jones  | 2145    | 77     | ford     | red
  2    | jones  | 2145    | 1012   | toyota   | blue
  16   | ashby  | 124     | 99     | bmw      | yellow

第2步:客观化

将结果吸入通用对象创建器中,并在第三项之后添加一个要拆分的参数。这意味着“jones”对象不会被制作多次。

步骤3:渲染

for p in people:
    print p.car.colour # no more car queries

有关python的扇形折叠的实现,请参阅此网页。