plan: fix predicate push down for UnionScan

[coocood]

What problem does this PR solve?

mysql> create table ta (a varchar(20));
Query OK, 0 rows affected (0.08 sec)

mysql> insert ta values ('1'), ('2');
Query OK, 2 rows affected (0.11 sec)

mysql> create table tb (a int);
Query OK, 0 rows affected (0.09 sec)

mysql> begin;
Query OK, 0 rows affected (0.00 sec)

mysql> insert tb values (0);
Query OK, 1 row affected (0.00 sec)

mysql> select * from ta where a = 1;
2 rows in set (0.00 sec)

Expected result is

1

Got

1
2

What is changed and how it works?

The predicate push down for UnionScan plan returns retained predicates instead of nil.

If we return nil predicates, the parent Selection will be removed, but conditions in UnionScan is only used to evaluate dirty table added rows, the conditions that cannot be pushed down to coprocessor will be lost.

Check List

Tests

• Unit test

Related changes

• Need to cherry-pick to the release branch

[coocood]

@jackysp @zz-jason @shenli PTAL

[coocood]

/run-all-tests

[zz-jason]

@winoros @eurekaka @XuHuaiyu PTAL

[zz-jason]

It's better to add another explain test.

[winoros]

Eval it in getSnapshotRow?

[coocood]

@winoros
I think do filter in Selection is better.

[coocood]

@zz-jason PTAL

[zz-jason]

why not Eval it in getSnapshotRow?

[coocood]

It's out of the scope of UnionScan's responsibility

[zz-jason]

Then why not removing all the filters inside the UnionScan?

[coocood]

It can be refactored later.

[zz-jason]

The filter a = 1, which is Selection_6, should be added as a parent of TableReader_9.

[coocood]

This is better, we can do it in the next PR.

[eurekaka]

LGTM

[XuHuaiyu]

LGMT

[shenli]

/run-all-tests