Snowflake 是否支持 select 子句中的关联子查询？

Question

我无法确定 Snowflake 是否支持

select

子句中的相关子查询，因为我遇到了相互矛盾的证据。

Snowflake querying-subqueries 文档说

where

子句支持相关子查询，似乎暗示

select

子句不支持它们。 here 和 here 等社区讨论似乎证实了这一点。

然而...

以Chinook 数据库为模型，这条在

select

子句中带有相关子查询的SQL 语句有效。

-- Three levels: artist, album, track
-- Correlated sub-query: level 2 -> level 3
-- Works: ✓
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track
                     where track.albumid = album.albumid) y
              ) tracks
              from album
             where album.artistid = 1) x
      ) albums
      from
        artist
     where artistid = 1) x;

这个也是……

-- Three levels: artist, album, track
-- Correlated sub-query: level 1 -> level 2
-- Works: ✓
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track
                     where track.albumid = 1) y
              ) tracks
              from album
             where album.artistid = artist.artistid) x
      ) albums
      from
        artist
     where artistid = 1) x;

然而，这并不...

-- Three levels: artist, album, track
-- Correlated sub-query: level 1 -> level 2 -> level 3
-- Doesn't work: ✗
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track
                     where track.albumid = album.albumid) y
              ) tracks
              from album
             where album.artistid = artist.artistid) x
      ) albums
      from
        artist
     where artistid = 1) x;

Error: SQL compilation error:
Unsupported subquery type cannot be evaluated

就好像相关子查询 are 在

select

子句中得到半途支持，只要引用不超过两个级别，它们就可以工作。然而，即使是这种有限的支持也与社区论坛和 StackOverflow 上的文档和传统智慧相矛盾。

我尝试了

select

子句中的相关子查询，涉及两层嵌套和三层嵌套。我期望的是它们都不起作用，尽管我也接受它们都起作用。我没想到的是其中一些（2 个级别）可以工作，而另一些（3 个级别）则不能。

附录

您有多个具有相同别名的子查询——我会尝试使用不同的名称，然后看看相关性是否更好。

这会引发上述错误：

-- Three levels: artist, album, track
-- Correlated sub-query: level 1 -> level 2 -> level 3
-- Doesn't work: ✗
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track t3
                     where t3.albumid = t2.albumid) y
              ) tracks
              from album t2
             where t2.artistid = t1.artistid) x
      ) albums
      from
        artist t1
     where artistid = 1);

你能为这两种情况发布 EXPLAIN USING TABULAR 吗？

-- Three levels: artist, album, track
-- Correlated sub-query: level 1 -> level 2
-- Works: ✓
explain using tabular
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track t3
                     where t3.albumid = 1) y
              ) tracks
              from album t2
             where t2.artistid = t1.artistid) x
      ) albums
      from
        artist t1
     where artistid = 1) x;

step | id | parent | operation     | objects               | alias | expressions                    | partitionstotal | partitionsassigned | bytesassigned
-----+----+--------+---------------+-----------------------+-------+--------------------------------+-----------------+--------------------+--------------
     |    |        | GlobalStats   |                       |       |                                | 3               | 3                  | 116224       
1    | 0  |        | Result        |                       |       | ARRAY_AGG(OBJECT_CONSTRUCT(... |                 |                    |              
1    | 1  | 0      | Aggregate     |                       |       | aggExprs: [ARRAY_AGG(OBJECT... |                 |                    |              
1    | 2  | 1      | Filter        |                       |       | T3.ALBUMID = 1                 |                 |                    |              
1    | 3  | 2      | TableScan     | CHINOOK.PUBLIC.TRACK  | T3    | TRACKID, NAME, ALBUMID, MED... | 1               | 1                  | 101888       
2    | 0  |        | Result        |                       |       | ARRAY_AGG(OBJECT_CONSTRUCT(... |                 |                    |              
2    | 1  | 0      | Aggregate     |                       |       | aggExprs: [ARRAY_AGG(OBJECT... |                 |                    |              
2    | 2  | 1      | LeftOuterJoin |                       |       | joinKey: (T2.ARTISTID = T1.... |                 |                    |              
2    | 3  | 2      | Filter        |                       |       | ARRAY_AGG(OBJECT_CONSTRUCT(... |                 |                    |              
2    | 4  | 3      | Aggregate     |                       |       | aggExprs: [ARRAY_AGG(OBJECT... |                 |                    |              
2    | 5  | 4      | Filter        |                       |       | T2.ARTISTID = 1                |                 |                    |              
2    | 6  | 5      | TableScan     | CHINOOK.PUBLIC.ALBUM  | T2    | ALBUMID, TITLE, ARTISTID       | 1               | 1                  | 8192         
2    | 7  | 2      | Filter        |                       |       | T1.ARTISTID = 1                |                 |                    |              
2    | 8  | 7      | TableScan     | CHINOOK.PUBLIC.ARTIST | T1    | ARTISTID, NAME                 | 1               | 1                  | 6144         

-- Three levels: artist, album, track
-- Correlated sub-query: level 1 -> level 2 -> level 3
-- Doesn't work: ✗

explain using tabular
select
  array_agg(object_construct(*))
  from (
    select
      *,
      (
        select
          array_agg(object_construct(*))
          from (
            select
              *,
              (
                select
                  array_agg(object_construct(*))
                  from (
                    select
                      *
                      from track t3
                     where t3.albumid = t2.albumid) y
              ) tracks
              from album t2
             where t2.artistid = t1.artistid) x
      ) albums
      from
        artist t1
     where artistid = 1);

同样的错误。

Answer 1

最简单的答案是“它不起作用”，因为从性能的角度来看，它们很恶心，永远不应该使用。

但是在一些简单的例子中，它确实有时会起作用。但是会爆炸成当前优化器无法解决的情况，因此会出现错误。

在一个层面上，您可以指向“其他某个数据库”并说他们这样做，是的，他们这样做了。那么下一个问题是，那你为什么不使用那个数据库……我会提出的一个原因是性能。

获得高性能结果的最好方法是理解你的数据，并编写 SQL 来解决你的数据集的每一个缺点，而不是更多。因此，根据定义，必须考虑“所有边缘情况”的通用解决方案性能不佳。