关于如何以编程方式从 json 文件开始创建自定义 org.apache.spark.sql.types.StructType 架构对象

我必须使用 json 文件中的信息创建一个自定义 org.apache.spark.sql.types.StructType 架构对象，json 文件可以是任何内容，所以我在属性文件中对其进行了参数化。

属性文件如下所示：

//ruta al esquema del fichero output (por defecto se infiere el esquema del Parquet destino). Si existe, el esquema será en formato JSON, aplicable a DataFrame (ver StructType.fromJson)
schema.parquet=/Users/XXXX/Desktop/generated_schema.json
writing.mode=overwrite
separator=;
header=false

文件 generated_schema.json 如下所示：

{"type" : "struct","fields" : [ {"name" : "codigo","type" : "string","nullable" : true}, {"name":"otro", "type":"string", "nullable":true}, {"name":"vacio", "type":"string", "nullable":true},{"name":"final","type":"string","nullable":true} ]}

所以，这就是我认为我可以解决它的方式：

val path: Path = new Path(mra_schema_parquet)
val fileSystem = path.getFileSystem(sc.hadoopConfiguration)
val inputStream: FSDataInputStream = fileSystem.open(path)
val schema_json = Stream.cons(inputStream.readLine(), Stream.continually( inputStream.readLine))

System.out.println("schema_json looks like "  + schema_json.head)

val mySchemaStructType :DataType = DataType.fromJson(schema_json.head)

/*
After this line, mySchemaStructType have four StructFields objects inside it, the same than appears at schema_json
*/
logger.info(mySchemaStructType)

val myStructType = new StructType()
myStructType.add("mySchemaStructType",mySchemaStructType)

/*

After this line, myStructType have zero StructFields! here must be the bug, myStructType should have the four StructFields that represents the loaded schema json! this must be the error! but how can i construct the necessary StructType object?

*/

myDF = loadCSV(sqlContext, path_input_csv,separator,myStructType,header)
System.out.println("myDF.schema.json looks like " + myDF.schema.json)
inputStream.close()

df.write
  .format("com.databricks.spark.csv")
  .option("header", header)
  .option("delimiter",delimiter)
  .option("nullValue","")
  .option("treatEmptyValuesAsNulls","true")
  .mode(saveMode)
  .parquet(pathParquet)

当代码运行最后一行 .parquet(pathParquet) 时，会发生异常：

**parquet.schema.InvalidSchemaException: Cannot write a schema with an empty group: message root {
}**

这段代码的输出是这样的：

16/11/11 13:57:04 INFO AnotherCSVtoParquet$: The job started using this propertie file: /Users/aisidoro/Desktop/mra-csv-converter/parametrizacion.properties
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: path_input_csv is /Users/aisidoro/Desktop/mra-csv-converter/cds_glcs.csv
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: path_output_parquet  is /Users/aisidoro/Desktop/output900000
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: mra_schema_parquet is /Users/aisidoro/Desktop/mra-csv-converter/generated_schema.json
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: writting_mode is overwrite
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: separator is ;
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: header is false
16/11/11 13:57:05 INFO AnotherCSVtoParquet$: ATTENTION! aplying mra_schema_parquet  /Users/aisidoro/Desktop/mra-csv-converter/generated_schema.json
schema_json looks like {"type" : "struct","fields" : [ {"name" : "codigo","type" : "string","nullable" : true}, {"name":"otro", "type":"string", "nullable":true}, {"name":"vacio", "type":"string", "nullable":true},{"name":"final","type":"string","nullable":true} ]}
16/11/11 13:57:12 INFO AnotherCSVtoParquet$: StructType(StructField(codigo,StringType,true), StructField(otro,StringType,true), StructField(vacio,StringType,true), StructField(final,StringType,true))
 16/11/11 13:57:13 INFO AnotherCSVtoParquet$: loadCSV. header is false, inferSchema is false pathCSV is /Users/aisidoro/Desktop/mra-csv-converter/cds_glcs.csv separator is ;
 myDF.schema.json looks like {"type":"struct","fields":[]}

schema_json 对象和 myDF.schema.json 对象应该具有相同的内容，不是吗？但它并没有发生。我认为这一定会引发错误。

最后，工作因以下例外而崩溃：

**parquet.schema.InvalidSchemaException: Cannot write a schema with an empty group: message root {
}**

事实是，如果我不提供任何 json 模式文件，作业执行得很好，但是使用这个模式......

有谁能够帮助我？我只想从 csv 文件和 json 模式文件开始创建一些镶木地板文件。

谢谢。

依赖项是：

    <spark.version>1.5.0-cdh5.5.2</spark.version>
    <databricks.version>1.5.0</databricks.version>

    <dependency>
        <groupId>org.apache.spark</groupId>
        <artifactId>spark-sql_2.10</artifactId>
        <version>${spark.version}</version>
        <scope>compile</scope>
    </dependency>
    <dependency>
        <groupId>com.databricks</groupId>
        <artifactId>spark-csv_2.10</artifactId>
        <version>${databricks.version}</version>
    </dependency>

UPDATE

我可以看到有一个悬而未决的问题，

https://github.com/databricks/spark-csv/issues/61 https://github.com/databricks/spark-csv/issues/61

既然你说Custom Schema，你可以做这样的事情。

val schema = (new StructType).add("field1", StringType).add("field2", StringType)
sqlContext.read.schema(schema).json("/json/file/path").show

另外，看看this https://stackoverflow.com/questions/33807145/evolving-a-schema-with-spark-dataframe and this http://blog.antlypls.com/blog/2016/01/30/processing-json-data-with-sparksql/

您可以创建嵌套的 JSON 架构，如下所示。

例如：

{
  "field1": {
    "field2": {
      "field3": "create",
      "field4": 1452121277
    }
  }
}

val schema = (new StructType)
  .add("field1", (new StructType)
    .add("field2", (new StructType)
      .add("field3", StringType)
      .add("field4", LongType)
    )
  )