[SPARK-32877][SQL] Fix Hive UDF not support decimal type in complex type by ulysses-you · Pull Request #29749 · apache/spark

ulysses-you · 2020-09-14T10:41:06Z

What changes were proposed in this pull request?

This pr aims to support Hive UDF when input type is not expected. We make HiveSimpleUDF extends ImplicitCastInputTypes then using Analyzer to implicit cast input data type.

Why are the changes needed?

It's a compatible issue. In normal case, data type is converted by Hive ObjectInspector at running time. But Hive not support input decimal type when method required double type. Unfortunately the default type of 1.1 is different between Spark and Hive, which are decimal and double. Then caused this issue.

Before this pr, we failed in this code.

class ArraySumUDF extends UDF {
  import scala.collection.JavaConverters._
  def evaluate(values: java.util.List[java.lang.Double]): java.lang.Double = {
    var r = 0d
    for (v <- values.asScala) {
      r += v
    }
    r
  }
}

sql(s"CREATE FUNCTION testArraySum AS '${classOf[ArraySumUDF].getName}'")
sql("SELECT testArraySum(array(1, 1.1, 1.2))")

-- error msg
Error in query: No handler for UDF/UDAF/UDTF 'ArraySumUDF': org.apache.hadoop.hive.ql.exec.NoMatchingMethodException: No matching method for class ArraySumUDF with (array<decimal(11,1)>). Possible choices: _FUNC_(array<double>)  ; line 1 pos 7

Does this PR introduce any user-facing change?

Yes, user can get correct result.

How was this patch tested?

add test.

ulysses-you · 2020-09-14T10:44:05Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala

Generally, UDF just has one method of evaluate.

SparkQA · 2020-09-14T11:26:13Z

Test build #128642 has finished for PR 29749 at commit e692b9f.

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

SparkQA · 2020-09-15T01:38:49Z

Test build #128674 has finished for PR 29749 at commit e73ccbf.

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

SparkQA · 2020-09-15T07:05:01Z

Test build #128696 has finished for PR 29749 at commit f9e3d57.

This patch fails due to an unknown error code, -9.
This patch merges cleanly.
This patch adds no public classes.

maropu · 2020-09-15T08:01:19Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala

Why do we need special handling for decimal types?

It's a compatible issue. In normal case, data type is converted by Hive ObjectInspector at running time. But Hive not support input decimal type when method required double type. Unfortunately the default type of 1.1 is different between Spark and Hive, which are decimal and double. Then caused this issue.

Ah, I see. Could you describe it in the PR description?

Sorry I didn't get it. ImplicitCastInputTypes can implicit cast decimal to double, doesn't it?

In first commit I did it for all types but test not passed. The reason is a UDF required an Object type.

We convert Java Object Type to NullType in HiveInspectors.javaTypeToDataType (seems this can be changed ?)

// Hive seems to return this for struct types? case c: Class[_] if c == classOf[java.lang.Object] => NullType

So we can't reflect the UDF method using a NullType, the error msg is:

in query: cannot resolve 'example_format('%o', 93)' due to data type mismatch: argument 2 requires array<null> type, however, '93' is of int type.; line 1 pos 7;

if a hive udf requires Object, I think it means AnyDataType. We should only special case it.

How about this ?

val expectTypes = method.getGenericParameterTypes.map(javaTypeToDataType) if (expectTypes.exists(_.existsRecursively(_.isInstanceOf[NullType]))) { children.map(_.dataType) } else { expectTypes }

method.getGenericParameterTypes.map(javaTypeToDataType).map { dt => if (dt.existsRecursively(_.isInstanceOf[NullType])) AnyDataType else dt }

Seems we should check data type and replace NullType to AnyDataType on by one to avoid such case Map<Double, Object> input.

I will do an another check if we can change HiveInspectors.javaTypeToDataType that using AnyDataType directly.

ulysses-you · 2020-09-17T04:39:09Z

also cc @cloud-fan @dongjoon-hyun the similar issue with #13930

cloud-fan · 2020-09-17T06:26:13Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala

  with Logging
-  with UserDefinedExpression {
+  with UserDefinedExpression
+  with ImplicitCastInputTypes {


ScalaUDF doesn't extend ImplicitCastInputTypes either. Does it have the same problem?

No, ImplicitTypeCasts has checked ScalaUDF as case udf: ScalaUDF if udf.inputTypes.nonEmpty =>

cloud-fan · 2020-09-17T11:19:58Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala

I'm confused. The expected type should be defined by the function signature, but not the actual function inputs. What are we doing here?

As an example, ScalaUDF.inputTypes is derived from the function signature (captured by encoders).

The ScalaUDF also implicit convert input type to expected type at ImplicitTypeCasts.

Let's say we have a udf Array[Double] => Double = { data => data.sum } and we run spark.sql("select udf(array(1.0, 1.1, 1.2))"). Then ImplicitTypeCasts will cast array<decimal> to array<double>.

But Hive udf can't enjoy it, now we only use Hive ObjectInspector to convert data type at running time. It's the difference.

Ah I misread the code. We get children data type only to skip decimal.

SparkQA · 2020-09-18T16:39:21Z

Test build #128879 has finished for PR 29749 at commit e13f2b9.

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

cloud-fan · 2020-09-21T07:10:48Z

My last concern is to check ScalaUDF and see if the type coercion rules are the same with HiveUDF. ScalaUDF doesn't extend ImplicitCastInputTypes.

ulysses-you · 2020-09-21T12:50:33Z

ScalaUDF has been checked in ImplicitTypeCasts just like extends ImplicitCastInputTypes. Here is some code from ImplicitTypeCasts.

      case e: ImplicitCastInputTypes if e.inputTypes.nonEmpty =>
        val children: Seq[Expression] = e.children.zip(e.inputTypes).map { case (in, expected) =>
          // If we cannot do the implicit cast, just use the original input.
          implicitCast(in, expected).getOrElse(in)
        }
        e.withNewChildren(children)

      case udf: ScalaUDF if udf.inputTypes.nonEmpty =>
        val children = udf.children.zip(udf.inputTypes).map { case (in, expected) =>
          // Currently Scala UDF will only expect `AnyDataType` at top level, so this trick works.
          // In the future we should create types like `AbstractArrayType`, so that Scala UDF can
          // accept inputs of array type of arbitrary element type.
          if (expected == AnyDataType) {
            in
          } else {
            implicitCast(
              in,
              udfInputToCastType(in.dataType, expected.asInstanceOf[DataType])
            ).getOrElse(in)
          }

        }

ulysses-you · 2020-09-21T14:28:06Z

Seems there is somethings wrong, HiveSimpleUDF.method is called before ImplicitCastInputTypes that will throw exception before we cast input types.

ulysses-you · 2020-09-22T01:47:04Z

After checked Hive code, it's a Hive version issue. In Hive 2.3.x this pr will not happen since the HIVE-13380 merged. And what I said is based on Hive 1.2.x.

An other thing is we cannot implicit cast input type because we need use children data type to reflect UDF method first and get exception if it fails. So we have no chance to go in ImplicitTypeCasts. We should skip method input data type check like my first commit 89a2966.

Sorry for the mistake, seems it's not need to fix the issue.

cloud-fan · 2020-09-22T08:30:35Z

Thanks for the investigation! You can probably open a new PR to add comments for HiveSimpleUDF, explaining why it can't extend ImplicitTypeCasts

ulysses-you · 2020-09-22T12:52:56Z

thanks @cloud-fan @maropu

dongjoon-hyun · 2020-09-22T17:18:33Z

Hi, @ulysses-you and @cloud-fan and @maropu ?
Can we have this test case in our Apache Spark repository?

dongjoon-hyun · 2020-09-22T17:19:50Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveUDFSuite.scala

    }
  }

+  test("SPARK-32877: Fix Hive UDF not support decimal type in complex type") {


If this is not covered in any other test case, it looks worth of having this.

dongjoon-hyun · 2020-09-22T17:40:51Z

I'm wondering if this test fails with '-Phive1.2' still.

ulysses-you · 2020-09-23T08:56:53Z

@dongjoon-hyun I will open a new PR that only add test and some comment.

ulysses-you added 2 commits September 14, 2020 18:35

init

89a2966

add line

e692b9f

probot-autolabeler bot added the SQL label Sep 14, 2020

ulysses-you commented Sep 14, 2020

View reviewed changes

reduce influence

e73ccbf

fix

f9e3d57

maropu reviewed Sep 15, 2020

View reviewed changes

cloud-fan reviewed Sep 17, 2020

View reviewed changes

anydatatype

e13f2b9

ulysses-you closed this Sep 22, 2020

dongjoon-hyun reviewed Sep 22, 2020

View reviewed changes

Conversation

ulysses-you commented Sep 14, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

Why are the changes needed?

Does this PR introduce any user-facing change?

How was this patch tested?

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Sep 14, 2020

Uh oh!

SparkQA commented Sep 15, 2020

Uh oh!

SparkQA commented Sep 15, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cloud-fan Sep 18, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ulysses-you commented Sep 17, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ulysses-you Sep 17, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Sep 18, 2020

Uh oh!

cloud-fan commented Sep 21, 2020

Uh oh!

ulysses-you commented Sep 21, 2020

Uh oh!

ulysses-you commented Sep 21, 2020

Uh oh!

ulysses-you commented Sep 22, 2020

Uh oh!

cloud-fan commented Sep 22, 2020

Uh oh!

ulysses-you commented Sep 22, 2020

Uh oh!

dongjoon-hyun commented Sep 22, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dongjoon-hyun commented Sep 22, 2020

Uh oh!

ulysses-you commented Sep 23, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

ulysses-you commented Sep 14, 2020 •

edited

Loading

cloud-fan Sep 18, 2020 •

edited

Loading

ulysses-you Sep 17, 2020 •

edited

Loading

dongjoon-hyun commented Sep 22, 2020 •

edited

Loading