[Feature] [Vectorized] Some Join opt in vec exec engine

Search before asking

• I had searched in the issues and found no similar issues.

Description

1. shoule adjust the out join column to nullable select * from t1 left join t2 on t1.a = t2.a all slots of the t2 shoule be nullable change in join node.

2. Adjust the join column of non nullaware join to non nullable select * from t1 join t2 on t1.a = t2.a if a is a nullable column shoule change to not nullable, because of join column = will skip the null.

3. Useless columns do not need to be output， here need a projection to reduce the column output to speed up the join operator.

Use case

No response

Related issues

No response

Are you willing to submit PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project’s Code of Conduct

[Gabriel39]

还有个场景我觉得可以做类似优化，就是比如agg+having过滤，having中的列不管之后有没有用到都会带下去，事实上如果之后用不到的话，在agg这里就可以裁剪掉了

[EmmyMiao87]

Join 性能优化

减少不必要的内存拷贝

Join Node 的输出 schema 与 Join Node 的输入 schema 不同。但当前 Doris 的 Join Node 算子在构造结果行时，直接将左右孩子的 tuple 进行拼接。 而实际上结果行的列可能是输入行中列的子集。这导致了很多无用的内存拷贝。

举例说明

select a.k1 from a, b where a.k1=b.k1;

输入 schema : a.k1, b.k1 输出 schema ：a.k1, b.k1 优化后输出 schema ：a.k1

MySQL [ssb]> select count(d_datekey) from lineorder inner join date on lo_orderdate = d_datekey;
+--------------------+
| count(`d_datekey`) |
+--------------------+
|          600037902 |
+--------------------+
1 row in set (10.555 sec)

打印 perf 发现，主要耗时函数在：

1. replicate 负责非 Join 列的结果填写函数。占用约 10%

测试

After PR: #8618

下面查询主要在 join 后可以裁剪列 lo_orderdate。效果如下：

MySQL [ssb]> select count(d_datekey) from lineorder inner join date on lo_orderdate = d_datekey;
+--------------------+
| count(`d_datekey`) |
+--------------------+
|          600037902 |
+--------------------+
1 row in set (7.286 sec)

MySQL [ssb]> set enable_hash_project=true;
Query OK, 0 rows affected (0.001 sec)

MySQL [ssb]> select count(d_datekey) from lineorder inner join date on lo_orderdate = d_datekey;
+--------------------+
| count(`d_datekey`) |
+--------------------+
|          600037902 |
+--------------------+
1 row in set (5.479 sec)

从测试结果看，开启裁剪后，性能提升 10% 符合预期。