[SPARK-28429][SQL] Support implicit cast string type to interval type in PromoteStrings

[wangyum]

What changes were proposed in this pull request?

PostgreSQL support implicit cast string type to interval type when adding timestamp type with string type:

postgres=# explain verbose select now() + '100 days', '100 days' + now();
                                QUERY PLAN
--------------------------------------------------------------------------
 Result  (cost=0.00..0.02 rows=1 width=16)
   Output: (now() + '100 days'::interval), (now() + '100 days'::interval)
(2 rows)

This pr add this feature.

Teradata, Oracle, DB2, Presto, MariaDB and MySQL does not support this syntax, PostgreSQL and Vertica support this syntax:

dbadmin=> select version(), current_timestamp + cast(interval '1' year as varchar(255)), cast(interval '2' year as varchar(255)) + current_timestamp, current_timestamp - cast(interval '3' year as varchar(255));
              version               |           ?column?            |           ?column?            |           ?column?
------------------------------------+-------------------------------+-------------------------------+-------------------------------
 Vertica Analytic Database v9.1.1-0 | 2019-07-24 13:14:03.003878+00 | 2019-07-25 13:14:03.003878+00 | 2019-07-20 13:14:03.003878+00
(1 row)

How was this patch tested?

unit tests

[SparkQA]

Test build #107836 has finished for PR 25190 at commit c5d751a.

• This patch fails Spark unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #107844 has finished for PR 25190 at commit 073c113.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[maropu]

In postgresql, is this handled as implicit casts? I checked a cast list in postgresql though, I couldn't find an entry, text -> interval, for implicit casts;

postgres=# \dC text
                               List of casts
    Source type    |    Target type    |      Function      |   Implicit?   
-------------------+-------------------+--------------------+---------------
 "char"            | text              | text               | yes
 boolean           | text              | text               | in assignment
 character         | text              | text               | yes
 character varying | text              | (binary coercible) | yes
 cidr              | text              | text               | in assignment
 inet              | text              | text               | in assignment
 name              | text              | text               | yes
 pg_dependencies   | text              | (with inout)       | yes
 pg_ndistinct      | text              | (with inout)       | yes
 pg_node_tree      | text              | (binary coercible) | yes
 text              | "char"            | char               | in assignment
 text              | character         | (binary coercible) | yes
 text              | character varying | (binary coercible) | yes
 text              | name              | name               | yes
 text              | regclass          | regclass           | yes
 text              | xml               | xml                | no
 xml               | text              | (binary coercible) | in assignment
(17 rows)

In postgresql, it seems select now() + '100 days' is hanldes as select now() + interval '100 days' in the parser? (I'm not 100% sure though)

[dongjoon-hyun]

Hi, @wangyum .

• Could you rebase this PR once more since CalendarInterval is changed to Interval?
• Could you answer to the above @maropu 's question?
• I merged your [SPARK-28424][SQL] Support typed interval expression. Do we still need this, too?

[wangyum]

Yes. We still need this fix. But I'm not sure how PostgreSQL support this feature.

[SparkQA]

Test build #108328 has finished for PR 25190 at commit a227143.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[maropu]

PostgreSQL supports conversions from a string literal to any data type (e.g., text->interval), but IIUC it doesn't support implicit casts from text-typed data in a relation;

postgres=# create table t2 (t timestamp, i text);
postgres=# insert into t2 values(now(), '1 hour');
postgres=# select * from t2;
             t              |   i    
----------------------------+--------
 2019-07-30 09:24:07.728206 | 1 hour
(1 row)

postgres=# select t + i from t2;
ERROR:  operator does not exist: timestamp without time zone + text
LINE 1: select t + i from t2;
                 ^
HINT:  No operator matches the given name and argument types. You might need to add explicit type casts.