带有 JOIN FETCH 性能问题的查询

Question

我无法弄清楚休眠查询性能的问题。在下面的代码 sn-p 中，我需要选择具有至少一个映射和过滤映射的实体。我为此使用 FETCH JOIN 来仅加载过滤后的映射。 但在那种情况下，我的查询存在性能问题。 Hibernate 说警告日志：

org.hibernate.hql.ast.QueryTranslatorImpl - 集合获取指定的 firstResult/maxResults；申请 记忆！

当我省略 FETCH JOIN 并只留下 JOIN 查询时，速度很快。但结果我将所有映射加载到实体，这对我来说是不可接受的状态。有没有办法提高查询性能？映射表有很多行。

HQL 查询：

select distinct e from Entity 
   join fetch e.mappings as mapping 
where e.deleted = 0 and e.mappings is not empty 
   and e = mapping.e and mapping.approval in (:approvals)

实体：

@Entity
@Table(name="entity")
class Entity {

   ...

   @OneToMany(mappedBy="entity", cascade=CascadeType.REMOVE, fetch=FetchType.LAZY)
   @OrderBy("created")
   private List<Mapping> mappings = new ArrayList<Mapping>();

   ...
}

@Entity
@Table(name="mapping")
class Mapping {

public static enum MappingApproval {
    WAITING, // mapping is waiting for approval
    APPROVED, // mapping was approved
    DECLINED; // mapping was declined
}

...

    @ManyToOne(fetch=FetchType.EAGER)
    @JoinColumn(name="entity_id", nullable=false)
    private Entity entity;

    @Enumerated(EnumType.STRING)
    @Column(name="approval", length=20)
    private MappingApproval approval;

...

}

谢谢

Answer 1

来自 JPA 规范

将 setMaxResults 或 setFirstResult 应用于查询的效果 涉及集合上的 fetch join 是未定义的。 （JPA“企业 JavaBeans 3.0，最终版本”，Kapitel 3.6.1 查询接口）

Hibernate 做了正确的事情，但是在内存中执行了一部分查询，这非常慢。在我的情况下，差异在 3-5 毫秒到 400-500 毫秒之间。

我的解决方案是在查询本身内实现分页。使用 JOIN FETCH 可以快速运行。

Answer 2

如果您需要带有“fetch”的 firstResult/maxResults，您可以将查询拆分为 2 个查询：

1. 使用 firstResult/maxResults 查询您的实体 ID，但不使用子表上的“获取”：

   select entity.id from entity (without fetch) where .... (with firstResult/maxResults)
   

2. 使用您的第一个查询返回的 id 上的“获取”查询您的实体：

   select entity from entity fetch ... where id in <previous ids>
   

Answer 3

之所以慢是因为Hibernate在执行SQL查询时根本没有分页，而且限制是在内存中完成的。

但是，如果连接必须扫描并获取 100k 条记录，而您只对 100 个结果感兴趣，那么提取器完成的 99.9% 的工作以及通过网络完成的所有 I/O 都只是浪费。

您可以轻松转换同时使用 JOIN FETCH 和分页的 JPQL 查询：

List<Post> posts = entityManager.createQuery("""
    select p
    from Post p
    left join fetch p.comments
    where p.title like :title
    order by p.id
    """, Post.class)
.setParameter("title", titlePattern)
.setMaxResults(maxResults)
.getResultList();

进入一个 SQL 查询，通过父标识符使用 DENSE_RANK 限制结果：

@NamedNativeQuery(
    name = "PostWithCommentByRank",
    query =
        "SELECT * " +
        "FROM (   " +
        "    SELECT *, dense_rank() OVER (ORDER BY \"p.created_on\", \"p.id\") rank " +
        "    FROM (   " +
        "        SELECT p.id AS \"p.id\", " +
        "               p.created_on AS \"p.created_on\", " +
        "               p.title AS \"p.title\", " +
        "               pc.id as \"pc.id\", " +
        "               pc.created_on AS \"pc.created_on\", " +
        "               pc.review AS \"pc.review\", " +
        "               pc.post_id AS \"pc.post_id\" " +
        "        FROM post p  " +
        "        LEFT JOIN post_comment pc ON p.id = pc.post_id " +
        "        WHERE p.title LIKE :titlePattern " +
        "        ORDER BY p.created_on " +
        "    ) p_pc " +
        ") p_pc_r " +
        "WHERE p_pc_r.rank <= :rank ",
    resultSetMapping = "PostWithCommentByRankMapping"
)
@SqlResultSetMapping(
    name = "PostWithCommentByRankMapping",
    entities = {
        @EntityResult(
            entityClass = Post.class,
            fields = {
                @FieldResult(name = "id", column = "p.id"),
                @FieldResult(name = "createdOn", column = "p.created_on"),
                @FieldResult(name = "title", column = "p.title"),
            }
        ),
        @EntityResult(
            entityClass = PostComment.class,
            fields = {
                @FieldResult(name = "id", column = "pc.id"),
                @FieldResult(name = "createdOn", column = "pc.created_on"),
                @FieldResult(name = "review", column = "pc.review"),
                @FieldResult(name = "post", column = "pc.post_id"),
            }
        )
    }
)

查询可以这样执行：

List<Post> posts = entityManager
.createNamedQuery("PostWithCommentByRank")
.setParameter(
    "titlePattern",
    "High-Performance Java Persistence %"
)
.setParameter(
    "rank",
    5
)
.unwrap(NativeQuery.class)
.setResultTransformer(
    new DistinctPostResultTransformer(entityManager)
)
.getResultList();

要将表格结果集转换回实体图，您需要一个ResultTransformer，如下所示：

public class DistinctPostResultTransformer
        extends BasicTransformerAdapter {
 
    private final EntityManager entityManager;
 
    public DistinctPostResultTransformer(
            EntityManager entityManager) {
        this.entityManager = entityManager;
    }
 
    @Override
    public List transformList(
            List list) {
             
        Map<Serializable, Identifiable> identifiableMap =
            new LinkedHashMap<>(list.size());
             
        for (Object entityArray : list) {
            if (Object[].class.isAssignableFrom(entityArray.getClass())) {
                Post post = null;
                PostComment comment = null;
 
                Object[] tuples = (Object[]) entityArray;
 
                for (Object tuple : tuples) {
                    if(tuple instanceof Identifiable) {
                        entityManager.detach(tuple);
 
                        if (tuple instanceof Post) {
                            post = (Post) tuple;
                        }
                        else if (tuple instanceof PostComment) {
                            comment = (PostComment) tuple;
                        }
                        else {
                            throw new UnsupportedOperationException(
                                "Tuple " + tuple.getClass() + " is not supported!"
                            );
                        }
                    }
                }
 
                if (post != null) {
                    if (!identifiableMap.containsKey(post.getId())) {
                        identifiableMap.put(post.getId(), post);
                        post.setComments(new ArrayList<>());
                    }
                    if (comment != null) {
                        post.addComment(comment);
                    }
                }
            }
        }
        return new ArrayList<>(identifiableMap.values());
    }
}

就是这样！

Answer 4

在为 JVM 增加内存之后，情况会好很多。毕竟我最终没有在查询中使用 FETCH。