Supports casting between ANSI interval types and integral types

[res-life]

Contributes #5113
Closes #5111
Supports:

• Cast year-month interval type to (byte | short | int | long) types
• Cast (byte | short | int | long) types to year-month interval type
• Cast day-time interval type to (byte | short | int | long) types
• Cast (byte | short | int | long) types to day-time interval type

Signed-off-by: Chong Gao [email protected]

[res-life]

build

[res-life]

Depends on this: #5352.
Cherry-picked the #5352 PR

Spark change is:
apache/spark@9553ed7072

Cpu throws a exception for the following scenario because startField != endField

spark.sql("select cast(1 as interval year to month)").show()
===== CPU output: =====
pyspark.sql.utils.AnalysisException: cannot resolve 'CAST(1 AS INTERVAL YEAR TO MONTH)' due to data type mismatch: cannot cast int to interval year to month; line 1 pos 7;
'Project [unresolvedalias(cast(1 as interval year to month), None)]
+- OneRowRelation

// cast(1 as interval year) is OK.
// cast(1 as interval year to month) is NOK.

// this is the Spark code  check `startField` == `endField`
case (_: IntegralType, DayTimeIntervalType(s, e)) if s == e => true

Gpu do not need to do this. CPU already checked this in the analysis phase.

[res-life]

build

[revans2]

But input is not guaranteed to be a ColumnVector for nested types. Like casting an array of DayTimeInterval to an Array of longs. These need to be ColumnView, but you should be able to treat it exactly the same as a ColumnVector.

[res-life]

Done

[revans2]

nit: If we want to have these all as separate queries, then lets have a separate test for each one so we can parallelize the execution and not have one test failure keep another test from running. If we are okay with them being a single test where one failure can mask another, like it is here, then can we combine them into a single query so it can run faster?

[res-life]

Done, combined into a single query to run faster.

[res-life]

build

[res-life]

No need to update AnsiCast I think, please double-check. @revans2
Spark does not allow inserting the interval column into the integral column.

spark.sql("create table t2(c1 interval year, c2 long) using parquet")
spark.sql("insert into t2 values (interval 1 year, 9223372036854775807)")
spark.sql("insert into t2 select c2, c1 from t2")
pyspark.sql.utils.AnalysisException: Cannot write incompatible data to table '`default`.`t2`':
- Cannot safely cast 'c1': bigint to interval year
- Cannot safely cast 'c2': interval year to bigint

[res-life]

@revans2 Help review, thanks.