What changes were proposed in this pull request?

Here is a case, let see how it goes in different SQL engines.

select cast('1 ' as int) as v1, '1 ' = 1 as v2

spark 1.6

NULL true

spark 2.1

NULL true

spark 2.2

NULL NULL

spark 2.3

NULL NULL

spark 2.4

NULL NULL

hive

NULL true

PostgreSQL

postgres=# select cast('1 ' as int) as v1, '1 ' = 1 as v2;
 v1 | v2
----+----
  1 | t
(1 row)

presto

presto> select cast('1 ' as int) as v1, '1 ' = 1 as v2;
Query 20191120_060530_00002_f5kcs failed: line 1:38: '=' cannot be applied to varchar(2), integer
select cast('1 ' as int) as v1, '1 ' = 1 as v2

presto> select cast('1 ' as int) as v1, '1 ' = '1 ' as v2;
Query 20191120_060545_00003_f5kcs failed: Cannot cast '1 ' to INT

Our behavior is unstable because type coercion changed since 2.2.
Personally, I think what PostgreSQL and Presto does here is more reasonable and consistent

Currently, this pull request obeys PostgreSQL, might need further discussion against this behavior change.

Why are the changes needed?

For better dirty data auto handling, keep consistency with older version sparks

Does this PR introduce any user-facing change?

ad ut