[SPARK-15776][SQL] Type coercion incorrect

[Sephiroth-Lin]

What changes were proposed in this pull request?

CREATE TABLE cdr (
  debet_dt          int      ,
  srv_typ_cd        string   ,
  b_brnd_cd         smallint ,
  call_dur          int
)
ROW FORMAT delimited fields terminated by ','
STORED AS TEXTFILE;

SELECT debet_dt,
       SUM(CASE WHEN srv_typ_cd LIKE '0%' THEN call_dur / 60 ELSE 0 END)
FROM cdr
GROUP BY debet_dt
ORDER BY debet_dt;

== Analyzed Logical Plan ==
debet_dt: int, sum(CASE WHEN srv_typ_cd LIKE 0% THEN (call_dur / 60) ELSE 0 END): bigint
Project [debet_dt#16, sum(CASE WHEN srv_typ_cd LIKE 0% THEN (call_dur / 60) ELSE 0 END)#27L]
+- Sort [debet_dt#16 ASC], true
   +- Aggregate [debet_dt#16], [debet_dt#16, sum(cast(CASE WHEN srv_typ_cd#18 LIKE 0% THEN (cast(call_dur#21 as double) / cast(60 as double)) ELSE cast(0 as double) END as bigint)) AS sum(CASE WHEN srv_typ_cd LIKE 0% THEN (call_dur / 60) ELSE 0 END)#27L]
      +- MetastoreRelation default, cdr

SELECT debet_dt,
       SUM(CASE WHEN b_brnd_cd IN(1) THEN call_dur / 60 ELSE 0 END)
FROM cdr
GROUP BY debet_dt
ORDER BY debet_dt;

== Analyzed Logical Plan ==
debet_dt: int, sum(CASE WHEN (CAST(b_brnd_cd AS INT) IN (CAST(1 AS INT))) THEN (CAST(call_dur AS DOUBLE) / CAST(60 AS DOUBLE)) ELSE CAST(0 AS DOUBLE) END): double
Project [debet_dt#76, sum(CASE WHEN (CAST(b_brnd_cd AS INT) IN (CAST(1 AS INT))) THEN (CAST(call_dur AS DOUBLE) / CAST(60 AS DOUBLE)) ELSE CAST(0 AS DOUBLE) END)#87]
+- Sort [debet_dt#76 ASC], true
   +- Aggregate [debet_dt#76], [debet_dt#76, sum(CASE WHEN cast(b_brnd_cd#80 as int) IN (cast(1 as int)) THEN (cast(call_dur#81 as double) / cast(60 as double)) ELSE cast(0 as double) END) AS sum(CASE WHEN (CAST(b_brnd_cd AS INT) IN (CAST(1 AS INT))) THEN (CAST(call_dur AS DOUBLE) / CAST(60 AS DOUBLE)) ELSE CAST(0 AS DOUBLE) END)#87]
      +- MetastoreRelation default, cdr

The only difference is WHEN condition, but will result different output column type(one is bigint, one is double) , so update type coercion order.

How was this patch tested?

Add new UT

[SparkQA]

Test build #60051 has finished for PR 13524 at commit 10b906a.

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

[SparkQA]

Test build #60295 has finished for PR 13524 at commit e396735.

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

[watermen]

@davies @rxin Can you review the code for this?

[rxin]

Can you put the description in the pull request description? it becomes part of the commit log and would be great to avoid requiring looking up jira.

[Sephiroth-Lin]

@rxin Done. Pleas help review, thank you.

[hvanhovell]

Just to be sure: are you moving the Division rule above the FunctionArgumentConversion rule in order to pass correct double arguments to functions? If so, could you add a test for this?

[clockfly]

@Sephiroth-Lin

I think you can use a simpler case to reproduce this in the description of this PR.

Such as:

select sum(4/3)

The expected result is:

1.33333..

The actual result is:

1

[clockfly]

This seems is a blocking issue.
I have created a PR at #13651, please help to review.

[rxin]

@Sephiroth-Lin can you close this pull request? Thanks.