语境

在最近的SO-post中,我发现在处理堆栈/链列表达式以及不同的Windows规范时,使用withColumn可以改善DAG。但是,在此示例中,withColumn实际上使DAG变得更糟,并且不同于使用select的结果。

可重现的例子

首先,一些测试数据(PySpark 2.4.4独立版):

import pandas as pd
import numpy as np

from pyspark.sql import SparkSession, Window
from pyspark.sql import functions as F

spark = SparkSession.builder.getOrCreate()

dfp = pd.DataFrame(
    {
        "col1": np.random.randint(0, 5, size=100),
        "col2": np.random.randint(0, 5, size=100),
        "col3": np.random.randint(0, 5, size=100),
        "col4": np.random.randint(0, 5, size=100),
        "col5": np.random.randint(0, 5, size=100),

    }
)

df = spark.createDataFrame(dfp)
df.show(5)

+----+----+----+----+----+
|col1|col2|col3|col4|col5|
+----+----+----+----+----+
|   0|   3|   2|   2|   2|
|   1|   3|   3|   2|   4|
|   0|   0|   3|   3|   2|
|   3|   0|   1|   4|   4|
|   4|   0|   3|   3|   3|
+----+----+----+----+----+
only showing top 5 rows

这个例子很简单。中包含2个窗口规范和基于它们的4个独立列表达式:
w1 = Window.partitionBy("col1").orderBy("col2")
w2 = Window.partitionBy("col3").orderBy("col4")

col_w1_1 = F.max("col5").over(w1).alias("col_w1_1")
col_w1_2 = F.sum("col5").over(w1).alias("col_w1_2")
col_w2_1 = F.max("col5").over(w2).alias("col_w2_1")
col_w2_2 = F.sum("col5").over(w2).alias("col_w2_2")

expr = [col_w1_1, col_w1_2, col_w2_1, col_w2_2]

withColumn-4个随机播放

如果withColumn与交替的窗口规范一起使用,则DAG会创建不必要的混洗:
df.withColumn("col_w1_1", col_w1_1)\
  .withColumn("col_w2_1", col_w2_1)\
  .withColumn("col_w1_2", col_w1_2)\
  .withColumn("col_w2_2", col_w2_2)\
  .explain()

== Physical Plan ==
Window [sum(col5#92L) windowspecdefinition(col3#90L, col4#91L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w2_2#147L], [col3#90L], [col4#91L ASC NULLS FIRST]
+- *(4) Sort [col3#90L ASC NULLS FIRST, col4#91L ASC NULLS FIRST], false, 0
   +- Exchange hashpartitioning(col3#90L, 200)
      +- Window [sum(col5#92L) windowspecdefinition(col1#88L, col2#89L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w1_2#143L], [col1#88L], [col2#89L ASC NULLS FIRST]
         +- *(3) Sort [col1#88L ASC NULLS FIRST, col2#89L ASC NULLS FIRST], false, 0
            +- Exchange hashpartitioning(col1#88L, 200)
               +- Window [max(col5#92L) windowspecdefinition(col3#90L, col4#91L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w2_1#145L], [col3#90L], [col4#91L ASC NULLS FIRST]
                  +- *(2) Sort [col3#90L ASC NULLS FIRST, col4#91L ASC NULLS FIRST], false, 0
                     +- Exchange hashpartitioning(col3#90L, 200)
                        +- Window [max(col5#92L) windowspecdefinition(col1#88L, col2#89L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w1_1#141L], [col1#88L], [col2#89L ASC NULLS FIRST]
                           +- *(1) Sort [col1#88L ASC NULLS FIRST, col2#89L ASC NULLS FIRST], false, 0
                              +- Exchange hashpartitioning(col1#88L, 200)
                                 +- Scan ExistingRDD[col1#88L,col2#89L,col3#90L,col4#91L,col5#92L]

选择-2个随机播放

如果所有列都使用select传递,则DAG是正确的。
df.select("*", *expr).explain()

== Physical Plan ==
Window [max(col5#92L) windowspecdefinition(col3#90L, col4#91L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w2_1#119L, sum(col5#92L) windowspecdefinition(col3#90L, col4#91L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w2_2#121L], [col3#90L], [col4#91L ASC NULLS FIRST]
+- *(2) Sort [col3#90L ASC NULLS FIRST, col4#91L ASC NULLS FIRST], false, 0
   +- Exchange hashpartitioning(col3#90L, 200)
      +- Window [max(col5#92L) windowspecdefinition(col1#88L, col2#89L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w1_1#115L, sum(col5#92L) windowspecdefinition(col1#88L, col2#89L ASC NULLS FIRST, specifiedwindowframe(RangeFrame, unboundedpreceding$(), currentrow$())) AS col_w1_2#117L], [col1#88L], [col2#89L ASC NULLS FIRST]
         +- *(1) Sort [col1#88L ASC NULLS FIRST, col2#89L ASC NULLS FIRST], false, 0
            +- Exchange hashpartitioning(col1#88L, 200)
               +- Scan ExistingRDD[col1#88L,col2#89L,col3#90L,col4#91L,col5#92L]

问题

有一些关于为什么应该避免使用withColumn的现有信息,但是它们主要涉及许多次调用withColumn的问题,并且没有解决偏离DAG的问题(请参阅herehere)。有谁知道为什么DAG在withColumnselect之间有所不同? Spark的优化算法在任何情况下都应适用,并且不应依赖于表达完全相同的事物的不同方法。

提前致谢。

最佳答案

使用嵌套withColumns和窗口函数时?

假设我要执行以下操作:

w1 = ...rangeBetween(-300, 0)
w2 = ...rowsBetween(-1,0)

(df.withColumn("some1", col(f.max("original1").over(w1))
   .withColumn("some2", lag("some1")).over(w2)).show()


即使使用非常小的数据集,我也会遇到很多内存问题和大量溢出。如果我使用select而不是withColumn做同样的事情,它的执行速度会更快。
df.select(
    f.max(col("original1")).over(w1).alias("some1"),
    f.lag("some1")).over(w2)
).show()

09-30 15:14