spark scala比较具有时间戳列的数据帧

Question

我正在尝试比较 2 组数据。一个是数据帧，一组静态数据并以 Avro 格式写入。现在，这种比较从 Avro 读取并检查哪个具有时间戳列，并且比较失败，因为 Avro 将数据存储为 Long 并且 sql 类型转换给出了不同的值

**--CREATE THE DATAFRAME**
val data = Seq(Row("1",java.sql.Timestamp.valueOf("2019-03-15 18:20:06.456")))
val schemaOrig = List( StructField("rowkey",StringType,true)
,StructField("txn_ts",TimestampType,true))

val sourceDf =  spark.createDataFrame(spark.sparkContext.parallelize(data),StructType(schemaOrig))
sourceDf.write.avro("test")
sourceDf.printSchema
root
 |-- rowkey: string (nullable = true)
 |-- txn_ts: timestamp (nullable = true)
sourceDf.show(false)


+----------------+-----------------------+
|rowkey          |txn_ts                 |
+----------------+-----------------------+
|1               |2019-03-15 18:20:06.456|
+----------------+-----------------------+

--As shown above the avro file has the expected schema specified ie String and Timestamp
--Now Read the data back from Avro 
val avroDf=spark.read.avro("test")
avroDf.printSchema

root
 |-- rowkey: string (nullable = true)
 |-- txn_ts: long (nullable = true)

avroDf.show(false)
--Avro Df schema is printing the timestamp field as long and data showing epoch time 

+----------------+-------------+
|rowkey          |txn_ts       |
+----------------+-------------+
|1               |1552688406456|
+----------------+-------------+
compare the 2 Df 
sourceDf.except(avroDf).show(false)
--Gives error due to datatype mismatch 
org.apache.spark.sql.AnalysisException: Except can only be performed on tables with the compatible column types. bigint <> timestamp at the second column of the second table;;
'Except
:- AnalysisBarrier

CAST the avro data long field back to time
stamp 
val modifiedAvroDf=avroDf.withColumn("txn_ts", col("txn_ts").cast(TimestampType))
modifiedAvroDf.printSchema

 |-- rowkey: string (nullable = true)
  |-- txn_ts: timestamp (nullable = true)
      modifiedAvroDf.show(false)
--Showing wrong timestamp value 
+----------------+-----------------------+
|rowkey          |txn_ts                 |
+----------------+-----------------------+
|1               |51172-09-26 11:07:366.0|
+----------------+-----------------------+

--Now Try to cast the source column to long 
val sourceModDf=sourceDf.withColumn("txn_ts",col("txn_ts").cast(LongType))
sourceModDf.printSchema

 |-- rowkey: string (nullable = true)
 |-- txn_ts: long (nullable = true)
sourceModDf.show(false)
sourceModDf.except(modifiedAvroDf).show(false)

Answer 1

创建 UDF 以将 long 转换为时间戳字符串。请检查以下代码。

scala> val df = Seq(1552688406456L).toDF
df: org.apache.spark.sql.DataFrame = [value: bigint]

scala> import org.joda.time.DateTime
import org.joda.time.DateTime

scala> import org.joda.time.DateTimeZone
import org.joda.time.DateTimeZone

scala> val datetime = udf((date: Long) => new DateTime(date, DateTimeZone.UTC).toString.replace("Z","").replace("T"," "))
datetime: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,StringType,Some(List(LongType)))

scala> df.select(datetime($"value").as("dt")).show(false)
+------------------------+
|dt                      |
+------------------------+
|2019-03-15 22:20:06.456 |
+------------------------+

scala> df.select(datetime($"value").as("dt").cast("timestamp")).show(false)
+-----------------------+
|dt                     |
+-----------------------+
|2019-03-15 22:20:06.456|
+-----------------------+


scala> df.select(datetime($"value").as("dt").cast("timestamp")).printSchema
root
 |-- dt: timestamp (nullable = true)