Structured Streaming中如何通過schema_of_json方法動態解析Kafka傳入的JSON數據的Schema

Structured Streaming中如何解析Kafka傳入的JSON數據的Schema

在實際生產中消息中的字段可能會發生變化，比如多加一個字段什麼的，但是Spark程序又不能停下來，所以考慮在程序中不是自定義好Schema，而是通過Kafka輸入消息中json串來infer Schema。當然，也可以通過廣播變量來更新配置文件，定期更新Schema，這也是一種寫法

在之前Spark Streaming中解析kafka的json格式數據時，採用的也是Schema infer來推斷，如下

dStream.map(_.value).foreachRDD(rdd=>{
  ...
  val spark = SparkSession.builder().config(rdd.sparkContext.getConf).getOrCreate()
  val df = spark.read.json(spark.createDataSet(rdd))
  ...
})

這樣通過解析json字符串，可以直接把json串的key作爲DataFrame的Columns列名

但是Structured Streaming中是直接生成DataFrame的，這樣做就不行。翻了下api發現了一個從json字符串推斷Schema的方法——schema_of_json

/**
 * Parses a JSON string and infers its schema in DDL format.
 *
 * @param json a JSON string.
 *
 * @group collection_funcs
 * @since 2.4.0
 */
def schema_of_json(json: String): Column = schema_of_json(lit(json))

/**
 * Parses a JSON string and infers its schema in DDL format.
 *
 * @param json a string literal containing a JSON string.
 *
 * @group collection_funcs
 * @since 2.4.0
 */
def schema_of_json(json: Column): Column = withExpr(new SchemaOfJson(json.expr))

絕對是第一手資料，我Google了一下這個方法的使用，除了Stack Overflow上有兩個帖子討論，但這個帖子給出的方法還是會報錯

How to query JSON data column using Spark DataFrames?

Implicit schema discovery on a JSON-formatted Spark DataFrame column?

這裏新建了個df測試下

嘗試使用第二個方法df.select(schema_of_json($“col”))時報錯

scala> df.select(schema_of_json($"col"))
org.apache.spark.sql.AnalysisException: cannot resolve 'schemaofjson(`col`)' due to data type mismatch: The input json should be a string literal and not null; however, got `col`.;;

看報錯信息是需要給一個字符串參數，所以做如下嘗試

也就是取出第一行Row的值，用來作推斷，發現是可以的

最後修改如下

scala> df.select(schema_of_json(df.take(1)(0).get(0).toString).alias("schema")).select(regexp_extract($"schema","struct<(.*?)>",1) as "schema").show(false)

+--------------------------------------------------------------------+
|schema                                                              |
+--------------------------------------------------------------------+
|cardno:string,cardtype:string,flag:string,times:string,userid:string|
|cardno:string,cardtype:string,flag:string,times:string,userid:string|
+--------------------------------------------------------------------+

最終寫法來創建一個Schema

scala> val str = df.select(schema_of_json(df.take(1)(0).get(0).toString).alias("schema")).select(regexp_extract($"schema","struct<(.*?)>",1)).take(1)(0).getAs[String](0)
str: String = cardno:string,cardtype:string,flag:string,times:string,userid:string

scala> val columns = str.split(",").map(x=>x.split(":")).map(x=>x(0))
columns: Array[String] = Array(cardno, cardtype, flag, times, userid)

scala> var schema = (new StructType)
schema: org.apache.spark.sql.types.StructType = StructType()

scala> columns.map(x=>{schema = schema.add(x,StringType,true)})
res154: Array[Unit] = Array((), (), (), (), ())

scala> schema
res159: org.apache.spark.sql.types.StructType = StructType(StructField(cardno,StringType,true), StructField(cardtype,StringType,true), StructField(flag,StringType,true), 
StructField(times,StringType,true), StructField(userid,StringType,true))

scala> schema.simpleString
res160: String = struct<cardno:string,cardtype:string,flag:string,times:string,userid:string>

就可以用推斷出的Schema來解析json了

df.select(from_json($"col",schema) as "json_value").select("json_value.*")

但這樣寫還是適合簡單的Schema結構，要是存在嵌套的結構，像Stack Overflow上有個問題，如何定義如下的Schema

struct<abc:struct<name:string>,pqr:struct<address:string>> 

參考：How to create schema (StructType) with one or more StructTypes?

import org.apache.spark.sql.types._
val name = new StructType().add($"name".string)
scala> println(name.simpleString)
struct<name:string>

val address = new StructType().add($"address".string)
scala> println(address.simpleString)
struct<address:string>

val schema = new StructType().add("abc", name).add("pqr", address)
scala> println(schema.simpleString)
struct<abc:struct<name:string>,pqr:struct<address:string>>

scala> schema.simpleString == "struct<abc:struct<name:string>,pqr:struct<address:string>>"
res4: Boolean = true

scala> schema.printTreeString
root
 |-- abc: struct (nullable = true)
 |    |-- name: string (nullable = true)
 |-- pqr: struct (nullable = true)
 |    |-- address: string (nullable = true)

這樣的話要寫個專門來解析的方法

---

其次厚臉皮推薦一個總結歸納的項目吧，面向Linux學習、大數據、機器學習、數據分析、算法等，歡迎前往推薦更多資源： Coding-Now

學習記錄的一些筆記，以及所看得一些電子書eBooks、視頻資源和平常收納的一些自己認爲比較好的博客、網站、工具
https://github.com/josonle/BigData-Learning