Scala Spark UDF 过滤 Array of Struct。

Question

我有一个具有模式的数据框架。

root
 |-- x: Long (nullable = false)
 |-- y: Long (nullable = false)
 |-- features: array (nullable = true)
 |    |-- element: struct (containsNull = true)
 |    |    |-- name: string (nullable = true)
 |    |    |-- score: double (nullable = true)

例如，我有数据。

+--------------------+--------------------+------------------------------------------+
|                x   |              y     |       features                           |
+--------------------+--------------------+------------------------------------------+
|10                  |          9         |[["f1", 5.9], ["ft2", 6.0], ["ft3", 10.9]]|
|11                  |          0         |[["f4", 0.9], ["ft1", 4.0], ["ft2", 0.9] ]|
|20                  |          9         |[["f5", 5.9], ["ft2", 6.4], ["ft3", 1.9] ]|
|18                  |          8         |[["f1", 5.9], ["ft4", 8.1], ["ft2", 18.9]]|
+--------------------+--------------------+------------------------------------------+

我想过滤具有特定前缀（例如“ft”）的功能。最终，我希望得到以下结果：

+--------------------+--------------------+-----------------------------+
|                x   |              y     |       features              |
+--------------------+--------------------+-----------------------------+
|10                  |          9         |[["ft2", 6.0], ["ft3", 10.9]]|
|11                  |          0         |[["ft1", 4.0], ["ft2", 0.9] ]|
|20                  |          9         |[["ft2", 6.4], ["ft3", 1.9] ]|
|18                  |          8         |[["ft4", 8.1], ["ft2", 18.9]]|
+--------------------+--------------------+-----------------------------+

我没有使用Spark 2.4+，所以无法使用这里提供的解决方案：Spark（Scala）过滤不需扩展的结构数组

我尝试使用UDF，但仍然无法生效。以下是我的尝试。定义了一个UDF：

def filterFeature: UserDefinedFunction = 
udf((features: Seq[Row]) =>
    features.filter{
        x.getString(0).startsWith("ft")
    }
)

但如果我应用这个UDF函数。

df.withColumn("filtered", filterFeature($"features"))

我遇到了错误 Schema for type org.apache.spark.sql.Row is not supported。我发现无法从UDF中返回Row。然后我尝试了：

def filterFeature: UserDefinedFunction = 
udf((features: Seq[Row]) =>
    features.filter{
        x.getString(0).startsWith("ft")
    }, (StringType, DoubleType)
)

然后我遇到了一个错误：

 error: type mismatch;
 found   : (org.apache.spark.sql.types.StringType.type, org.apache.spark.sql.types.DoubleType.type)
 required: org.apache.spark.sql.types.DataType
              }, (StringType, DoubleType)
                 ^

根据一些答案的建议，我也尝试了一个case class：

case class FilteredFeature(featureName: String, featureScore: Double)
def filterFeature: UserDefinedFunction = 
udf((features: Seq[Row]) =>
    features.filter{
        x.getString(0).startsWith("ft")
    }, FilteredFeature
)

但我得到了：

 error: type mismatch;
 found   : FilteredFeature.type
 required: org.apache.spark.sql.types.DataType
              }, FilteredFeature
                 ^

我尝试过：

case class FilteredFeature(featureName: String, featureScore: Double)
def filterFeature: UserDefinedFunction = 
udf((features: Seq[Row]) =>
    features.filter{
        x.getString(0).startsWith("ft")
    }, Seq[FilteredFeature]
)

我得到：

<console>:192: error: missing argument list for method apply in class GenericCompanion
Unapplied methods are only converted to functions when a function type is expected.
You can make this conversion explicit by writing `apply _` or `apply(_)` instead of `apply`.
              }, Seq[FilteredFeature]
                    ^

我尝试过：

case class FilteredFeature(featureName: String, featureScore: Double)
def filterFeature: UserDefinedFunction = 
udf((features: Seq[Row]) =>
    features.filter{
        x.getString(0).startsWith("ft")
    }, Seq[FilteredFeature](_)
)

I got:

<console>:201: error: type mismatch;
 found   : Seq[FilteredFeature]
 required: FilteredFeature
              }, Seq[FilteredFeature](_)
                          ^

在这种情况下，我该做什么？

- al3xtouch

Answer 1

你有两个选项：

a) 给UDF提供一个模式，这样你就可以返回Seq[Row] b) 将Seq[Row]转换为Tuple2序列或一个case class的Seq，然后就不需要提供模式（但如果使用元组，结构字段名称会丢失！）

对于你的情况，我更喜欢选项a）（对于具有许多字段的结构很好用）：

val schema = df.schema("features").dataType

val filterFeature = udf((features:Seq[Row]) => features.filter(_.getAs[String]("name").startsWith("ft")),schema)

- Raphael Roth

Answer 2

试试这个：

def filterFeature: UserDefinedFunction =
    udf((features: Row) => {
      features.getAs[Array[Array[Any]]]("features").filter(in => in(0).asInstanceOf[String].startsWith("ft"))
 })

- Emiliano Martinez

Answer 3

如果您没有使用Spark 2.4，那么这应该适用于您的情况。

case class FilteredFeature(featureName: String, featureScore: Double)

import org.apache.spark.sql.functions._  
def filterFeature: UserDefinedFunction = udf((feature: Seq[Row]) => {
  feature.filter(x => {
    x.getString(0).startsWith("ft")
  }).map(r => FilteredFeature(r.getString(0), r.getDouble(1)))
})

df.select($"x", $"y", filterFeature($"feature") as "filter").show(false)

输出：

+---+---+-----------------------+
|x  |y  |filter                 |
+---+---+-----------------------+
|10 |9  |[[ft2,6.0], [ft3,10.9]]|
|11 |0  |[[ft1,4.0], [ft2,0.9]] |
|20 |9  |[[ft2,6.4], [ft3,1.9]] |
|18 |8  |[[ft4,8.1], [ft2,18.9]]|
+---+---+-----------------------+

- koiralo

---

但是在过滤后，您的特征模式与此不同（结构名称应为name和score）。 - Raphael Roth