为什么Window函数失败并且“Window function X不采用帧规范”？

Question

我正在尝试在pyspark 1.4.1中使用Spark 1.4 window functions

但主要是错误或意外结果。这是一个我认为应该工作的非常简单的例子：

from pyspark.sql.window import Window
import pyspark.sql.functions as func

l = [(1,101),(2,202),(3,303),(4,404),(5,505)]
df = sqlContext.createDataFrame(l,["a","b"])

wSpec = Window.orderBy(df.a).rowsBetween(-1,1)

df.select(df.a, func.rank().over(wSpec).alias("rank"))  
    ==> Failure org.apache.spark.sql.AnalysisException: Window function rank does not take a frame specification.

df.select(df.a, func.lag(df.b,1).over(wSpec).alias("prev"), df.b, func.lead(df.b,1).over(wSpec).alias("next"))  
    ===>  org.apache.spark.sql.AnalysisException: Window function lag does not take a frame specification.;


wSpec = Window.orderBy(df.a)

df.select(df.a, func.rank().over(wSpec).alias("rank"))
    ===> org.apache.hadoop.hive.ql.exec.UDFArgumentTypeException: One or more arguments are expected.

df.select(df.a, func.lag(df.b,1).over(wSpec).alias("prev"), df.b, func.lead(df.b,1).over(wSpec).alias("next")).collect()

    [Row(a=1, prev=None, b=101, next=None), Row(a=2, prev=None, b=202, next=None), Row(a=3, prev=None, b=303, next=None)]

如您所见，如果我添加rowsBetween帧规范，rank()和lag/lead()窗口函数都不会识别它：“窗口函数不采用帧规范”。

如果我在leas rowsBetween省略lag/lead()框架规范，不要抛出异常但返回意外（对我而言）结果：总是None。并且rank()仍然不能用于不同的例外。

任何人都可以帮助我使我的窗口功能正常吗？

UPDATE

好吧，开始看起来像一个pyspark错误。我在纯Spark（Scala，spark-shell）中准备了相同的测试：

import sqlContext.implicits._
import org.apache.spark.sql._
import org.apache.spark.sql.types._

val l: List[Tuple2[Int,Int]] = List((1,101),(2,202),(3,303),(4,404),(5,505))
val rdd = sc.parallelize(l).map(i => Row(i._1,i._2))
val schemaString = "a b"
val schema = StructType(schemaString.split(" ").map(fieldName => StructField(fieldName, IntegerType, true)))
val df = sqlContext.createDataFrame(rdd, schema)

import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._

val wSpec = Window.orderBy("a").rowsBetween(-1,1)
df.select(df("a"), rank().over(wSpec).alias("rank"))
    ==> org.apache.spark.sql.AnalysisException: Window function rank does not take a frame specification.;

df.select(df("a"), lag(df("b"),1).over(wSpec).alias("prev"), df("b"), lead(df("b"),1).over(wSpec).alias("next"))
    ===> org.apache.spark.sql.AnalysisException: Window function lag does not take a frame specification.;


val wSpec = Window.orderBy("a")
df.select(df("a"), rank().over(wSpec).alias("rank")).collect()
    ====> res10: Array[org.apache.spark.sql.Row] = Array([1,1], [2,2], [3,3], [4,4], [5,5])

df.select(df("a"), lag(df("b"),1).over(wSpec).alias("prev"), df("b"), lead(df("b"),1).over(wSpec).alias("next"))
    ====> res12: Array[org.apache.spark.sql.Row] = Array([1,null,101,202], [2,101,202,303], [3,202,303,404], [4,303,404,505], [5,404,505,null])

即使rowsBetween不能应用于Scala，rank()和lag()/lead()都可以正常工作，因为我省略了rowsBetween。

Answer 1

据我所知，有两个不同的问题。 Hive GenericUDAFRank，GenericUDAFLag和GenericUDAFLead根本不支持窗口框架定义，因此您看到的错误是预期的行为。

关于以下PySpark代码的问题

wSpec = Window.orderBy(df.a)
df.select(df.a, func.rank().over(wSpec).alias("rank"))

看起来它与我的问题https://stackoverflow.com/q/31948194/1560062有关，应该由SPARK-9978解决。到目前为止，您可以通过将窗口定义更改为此来使其工作：

wSpec = Window.partitionBy().orderBy(df.a)