依靠Expand实现aggr distinct(1)

现在的aggr distinct支持

现在IMCI依靠Join + HashGroupby实现对aggr distinct的支持，在这里分类列出各个情况下的改写策略

不含grouping clause的情况

仅含一个agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1) 
FROM t1;

会被改写为

SELECT 
  COUNT(tmp.col1) 
FROM 
  (
     SELECT
       col1
     FROM
       t1
     GROUP BY col1);

仅含有多个agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1), COUNT(DISTINCT col2)
FROM t1;

会被改写为

CREATE VIEW v1 AS 
  SELECT 
    COUNT(tmp.col1)
  FROM 
    (
       SELECT
         col1
       FROM
         t1
       GROUP BY col1);

CREATE VIEW v2 AS 
  SELECT 
    COUNT(tmp.col2)
  FROM 
    (
       SELECT
         col2
       FROM
         t1
       GROUP BY col2);
       
SELECT * FROM v1, v2;

同时含有普通agg和agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1), COUNT(DISTINCT col2), SUM(col3)
FROM t1;

会被改写为

CREATE VIEW v1 AS 
  SELECT 
    COUNT(tmp.col1)
  FROM 
    (
       SELECT
         col1
       FROM
         t1
       GROUP BY col1);

CREATE VIEW v2 AS 
  SELECT 
    COUNT(tmp.col2)
  FROM 
    (
       SELECT
         col2
       FROM
         t1
       GROUP BY col2);

CREATE VIEW v3 AS 
  SELECT SUM(col3) FROM t1;
       
SELECT * FROM v1, v2, v3;

含grouping clause的情况

仅含一个agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1) 
FROM t1
GROUP BY col2;

会被改写为

SELECT 
  COUNT(tmp.col1) 
FROM 
  (
     SELECT
       col1, col2
     FROM
       t1
     GROUP BY col1, col2)
GROUP BY col2;

仅含有多个agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1), COUNT(DISTINCT col2)
FROM t1
GROUP BY col3;

会被改写为

 CREATE VIEW v1 AS 
  SELECT 
    col3, COUNT(tmp.col1)
  FROM 
    (
       SELECT
         col1, col3
       FROM
         t1
       GROUP BY col1, col3
    )
  GROUP BY col3;

CREATE VIEW v2 AS 
  SELECT 
    col3, COUNT(tmp.col2)
  FROM 
    (
       SELECT
         col2, col3
       FROM
         t1
       GROUP BY col2, col3
     )
   GROUP BY col3;
       
SELECT * FROM v1 JOIN v2 ON v1.col3 = v2.col3;

同时含有普通agg和agg discinct

形如以下的SQL

SELECT 
  COUNT(DISTINCT col1), COUNT(DISTINCT col2), SUM(col3)
FROM t1
GROUP BY col4;

会被改写为

CREATE VIEW v1 AS 
  SELECT 
    col4, COUNT(tmp.col1)
  FROM 
    (
       SELECT
         col1, col4
       FROM
         t1
       GROUP BY col1, col4
    )
  GROUP BY col4;

CREATE VIEW v2 AS 
  SELECT 
    col4, COUNT(tmp.col2)
  FROM 
    (
       SELECT
         col2, col4
       FROM
         t1
       GROUP BY col2, col4
     )
   GROUP BY col4;

CREATE VIEW v3 AS 
  SELECT SUM(col3) FROM t1 GROUP BY col4;
       
SELECT * FROM v1, v2, v3 WHERE v1.col4 = v2.col4 and v2.col4 = v3.col4;

可以看出，大体的思路是：

1. 对于单个agg(distinct col), 使用叠加一个group by的方式使y = distinct col，然后计算agg(y)
2. 对于含有多个agg的情形，先拆分为多个group by，每个group by只含有一个agg，因此每个含有agg distinct的group by都可以依赖1来解决

但是，这样的作法在OSS外表的查询执行下有一些问题，因为这样的改写会生成大量的TableScan，在单机引擎中，我们认为TableScan算子是一个开销很小的算子，因为在单机上我们可以利用lru cache, pruner, prefetch等优化，同时读local disk和polar store相对较快，但是OSS上的TableScan和单机引擎有一些不同

• 以上的lru cache, pruner, prefetch等优化均不存在
• 目前的过滤条件使通过Scan + Filter两个算子实现，Filter没有下推到scan，scan输出的数据量很大

这直接导致了OSS上含有多个agg, distinct的query查询效果很差，因此，可能需要一个不含有join的处理agg distinct的方式