有了这个模式:
root
|-- Elems: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- Elem: integer (nullable = true)
| | |-- Desc: string (nullable = true)
我们如何像这样添加一个新字段?
root
|-- Elems: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- New_field: integer (nullable = true)
| | |-- Elem: integer (nullable = true)
| | |-- Desc: string (nullable = true)
我已经用一个简单的结构体实现了这个功能(详见本帖底部),但我无法使用结构体数组来完成它。
以下是测试代码:
val schema = new StructType()
.add("Elems", ArrayType(new StructType()
.add("Elem", IntegerType)
.add("Desc", StringType)
))
val dataDS = Seq("""
{
"Elems": [ {"Elem":1, "Desc": "d1"}, {"Elem":2, "Desc": "d2"}, {"Elem":3, "Desc": "d3"} ]
}
""").toDS()
val df = spark.read.schema(schema).json(dataDS.rdd)
df.show(false)
+---------------------------+
|Elems |
+---------------------------+
|[[1, d1], [2, d2], [3, d3]]|
+---------------------------+
一旦我们有了DF,我建议的最佳方法是为每个元素创建一个结构数组:
val mod_df = df.withColumn("modif_elems",
struct(
array(lit("")).as("New_field"),
col("Elems.Elem"),
col("Elems.Desc")
))
mod_df.show(false)
+---------------------------+-----------------------------+
|Elems |modif_elems |
+---------------------------+-----------------------------+
|[[1, d1], [2, d2], [3, d3]]|[[], [1, 2, 3], [d1, d2, d3]]|
+---------------------------+-----------------------------+
mod_df.printSchema
root
|-- Elems: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- Elem: integer (nullable = true)
| | |-- Desc: string (nullable = true)
|-- modif_elems: struct (nullable = false)
| |-- New_field: array (nullable = false)
| | |-- element: string (containsNull = false)
| |-- Elem: array (nullable = true)
| | |-- element: integer (containsNull = true)
| |-- Desc: array (nullable = true)
| | |-- element: string (containsNull = true)
我们没有失去任何数据,但这并不完全是我想要的。
更新: 在PD1中找到了解决方法。
额外内容: 修改一个结构体(不在数组中)
代码几乎相同,但现在我们没有一个结构体数组,所以修改结构体会更容易:
val schema = new StructType()
.add("Elems", new StructType()
.add("Elem", IntegerType)
.add("Desc", StringType)
)
val dataDS = Seq("""
{
"Elems": {"Elem":1, "Desc": "d1"}
}
""").toDS()
val df = spark.read.schema(schema).json(dataDS.rdd)
df.show(false)
+-------+
|Elems |
+-------+
|[1, d1]|
+-------+
df.printSchema
root
|-- Elems: struct (nullable = true)
| |-- Elem: integer (nullable = true)
| |-- Desc: string (nullable = true)
在这种情况下,为了添加该字段,我们需要创建另一个结构体:
val mod_df = df
.withColumn("modif_elems",
struct(
lit("").alias("New_field"),
col("Elems.Elem"),
col("Elems.Desc")
)
)
mod_df.show
+-------+-----------+
| Elems|modif_elems|
+-------+-----------+
|[1, d1]| [, 1, d1]|
+-------+-----------+
mod_df.printSchema
root
|-- Elems: struct (nullable = true)
| |-- Elem: integer (nullable = true)
| |-- Desc: string (nullable = true)
|-- modif_elems: struct (nullable = false)
| |-- New_field: string (nullable = false)
| |-- Elem: integer (nullable = true)
| |-- Desc: string (nullable = true)
PD1:
好的,我使用了Spark SQL函数arrays_zip(2.4.0版本中新增),它基本上是我想要的,但我不知道如何更改元素名称(as或alias在此处无效):
val mod_df = df.withColumn("modif_elems",
arrays_zip(
array(lit("")).as("New_field"),
col("Elems.Elem").as("Elem"),
col("Elems.Desc").alias("Desc")
)
)
mod_df.show(false)
+---------------------------+---------------------------------+
|Elems |modif_elems |
+---------------------------+---------------------------------+
|[[1, d1], [2, d2], [3, d3]]|[[, 1, d1], [, 2, d2], [, 3, d3]]|
+---------------------------+---------------------------------+
mod_df.printSchema
root
|-- Elems: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- Elem: integer (nullable = true)
| | |-- Desc: string (nullable = true)
|-- modif_elems: array (nullable = true)
| |-- element: struct (containsNull = false)
| | |-- 0: string (nullable = true)
| | |-- 1: integer (nullable = true)
| | |-- 2: string (nullable = true)
结构体 modif_elems 应该包含三个元素,名称分别为 New_field、Elem 和 Desc,而不是 0、1 和 2。
ArraysZip
是来自org/apache/spark/sql/catalyst/expressions/collectionOperations.scala
的一个case类。您可以使用您喜欢的代码编辑器并打开arrays_zip
函数源代码来查看它。 - rvillaelem_struct_recomposed
? - sohil