Question

我有一个像这样的数据框

| id | date      |  KPI_1 | ... | KPI_n
| 1  |2012-12-12 |   0.1  | ... |  0.5
| 2  |2012-12-12 |   0.2  | ... |  0.4
| 3  |2012-12-12 |   0.66 | ... |  0.66 
| 1  |2012-12-13 |   0.2  | ... |  0.46
| 4  |2012-12-14 |   0.2  | ... |  0.45 
| ...
| 55| 2013-03-15 |  0.5  | ... |  0.55

我们有

X个标识符
给定日期的每个标识符的行
n个关键绩效指标

我必须为每一行计算一些派生的KPI，并且此KPI取决于每个ID的先前值。假设我得出的KPI是一个差异，那就是：

| id | date      |  KPI_1 | ... | KPI_n | KPI_1_diff | KPI_n_diff
| 1  |2012-12-12 |   0.1  | ... |  0.5  |   0.1      | 0.5
| 2  |2012-12-12 |   0.2  | ... |  0.4  |   0.2      |0.4
| 3  |2012-12-12 |   0.66 | ... |  0.66 |   0.66     | 0.66 
| 1  |2012-12-13 |   0.2  | ... |  0.46 |   0.2-0.1  | 0.46 - 0.66
| 4  |2012-12-13 |   0.2  | ... |  0.45  ...
| ...
| 55| 2013-03-15 |  0.5  | ... |  0.55

现在：我要做的是：

val groupedDF = myDF.groupBy("id").agg(
    collect_list(struct(col("date",col("KPI_1"))).as("wrapped_KPI_1"),
    collect_list(struct(col("date",col("KPI_2"))).as("wrapped_KPI_2")
    // up until nth KPI
)

我会得到汇总数据，例如：

[("2012-12-12",0.1),("2012-12-12",0.2) ...

然后，我将对这些包装的数据进行排序，解包并使用一些UDF在这些汇总结果上进行映射，并产生输出（计算差异和其他统计信息）。

另一种方法是使用窗口功能，例如：

val window = Window.partitionBy(col("id")).orderBy(col("date")).rowsBetween(Window.unboundedPreceding,0L)

然后做：

val windowedDF = df.select (
  col("id"),
  col("date"),
  col("KPI_1"),
  collect_list(struct(col("date"),col("KPI_1"))).over(window),  
  collect_list(struct(col("date"),col("KPI_2"))).over(window)
   )

这样我可以得到：

[("2012-12-12",0.1)]
[("2012-12-12",0.1), ("2012-12-13",0.1)]
...

看起来更好处理，但是我怀疑重复窗口会为每个KPI产生不必要的分组和排序。

所以这是问题：

我想采用分组方式吗？
我要去窗户吗？如果是这样，最有效的方法是什么？

Answer 1

我相信窗口方法应该是更好的解决方案，但是在使用窗口功能之前，您应该根据ID重新划分数据帧。这将仅对数据进行一次随机播放，并且所有窗口函数都应使用已随机播放的数据帧执行。希望对您有所帮助。

代码应该是这样的。

val windowedDF = df.repartition(col("id"))
  .select (
  col("id"),
  col("date"),
  col("KPI_1"),
  col("KPI_2"),
  collect_list(struct(col("date"),col("KPI_1"))).over(window),
  collect_list(struct(col("date"),col("KPI_2"))).over(window)
)

@Raphael Roth

在这里，我们正在汇总一个窗口。这就是为什么您可能会看到相同的执行计划的原因。请参阅下面的示例，其中只能从一个分区完成多个窗口的聚合。

val list = Seq(( "2", null, 1, 11, 1, 1 ),
  ( "2", null, 1, 22, 2, 2 ),
  ( "2", null, 1, 11, 1, 3 ),
  ( "2", null, 1, 22, 2, 1 ),
  ( "2", null, 1, 33, 1, 2 ),
  ( null, "3", 3, 33, 1, 2 ),
  ( null, "3", 3, 33, 2, 3 ),
  ( null, "3", 3, 11, 1, 1 ),
  ( null, "3", 3, 22, 2, 2 ),
  ( null, "3", 3, 11, 1, 3 )
)

val df = spark.sparkContext.parallelize(list).toDF("c1","c2","batchDate","id", "pv" , "vv")

val c1Window = Window.partitionBy("batchDate", "c1")
val c2Window = Window.partitionBy("batchDate", "c2")

val agg1df = df.withColumn("c1List",collect_list("pv").over(c1Window))
  .withColumn("c2List", collect_list("pv").over(c2Window))

val agg2df = df.repartition($"batchDate")
  .withColumn("c1List",collect_list("pv").over(c1Window))
  .withColumn("c2List", collect_list("pv").over(c2Window))


agg1df.explain()
== Physical Plan ==
Window [collect_list(pv#18, 0, 0) windowspecdefinition(batchDate#16, c2#15, ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS c2List#38], [batchDate#16, c2#15]
+- *Sort [batchDate#16 ASC NULLS FIRST, c2#15 ASC NULLS FIRST], false, 0
   +- Exchange hashpartitioning(batchDate#16, c2#15, 1)
      +- Window [collect_list(pv#18, 0, 0) windowspecdefinition(batchDate#16, c1#14, ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS c1List#28], [batchDate#16, c1#14]
         +- *Sort [batchDate#16 ASC NULLS FIRST, c1#14 ASC NULLS FIRST], false, 0
            +- Exchange hashpartitioning(batchDate#16, c1#14, 1)
               +- *Project [_1#7 AS c1#14, _2#8 AS c2#15, _3#9 AS batchDate#16, _4#10 AS id#17, _5#11 AS pv#18, _6#12 AS vv#19]
                  +- *SerializeFromObject [staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, assertnotnull(input[0, scala.Tuple6, true])._1, true) AS _1#7, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, assertnotnull(input[0, scala.Tuple6, true])._2, true) AS _2#8, assertnotnull(input[0, scala.Tuple6, true])._3 AS _3#9, assertnotnull(input[0, scala.Tuple6, true])._4 AS _4#10, assertnotnull(input[0, scala.Tuple6, true])._5 AS _5#11, assertnotnull(input[0, scala.Tuple6, true])._6 AS _6#12]
                     +- Scan ExternalRDDScan[obj#6]

agg2df.explain()
== Physical Plan ==
Window [collect_list(pv#18, 0, 0) windowspecdefinition(batchDate#16, c2#15, ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS c2List#60], [batchDate#16, c2#15]
+- *Sort [batchDate#16 ASC NULLS FIRST, c2#15 ASC NULLS FIRST], false, 0
   +- Window [collect_list(pv#18, 0, 0) windowspecdefinition(batchDate#16, c1#14, ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS c1List#50], [batchDate#16, c1#14]
      +- *Sort [batchDate#16 ASC NULLS FIRST, c1#14 ASC NULLS FIRST], false, 0
         +- Exchange hashpartitioning(batchDate#16, 1)
            +- *Project [_1#7 AS c1#14, _2#8 AS c2#15, _3#9 AS batchDate#16, _4#10 AS id#17, _5#11 AS pv#18, _6#12 AS vv#19]
               +- *SerializeFromObject [staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, assertnotnull(input[0, scala.Tuple6, true])._1, true) AS _1#7, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, assertnotnull(input[0, scala.Tuple6, true])._2, true) AS _2#8, assertnotnull(input[0, scala.Tuple6, true])._3 AS _3#9, assertnotnull(input[0, scala.Tuple6, true])._4 AS _4#10, assertnotnull(input[0, scala.Tuple6, true])._5 AS _5#11, assertnotnull(input[0, scala.Tuple6, true])._6 AS _6#12]
                  +- Scan ExternalRDDScan[obj#6]

火花窗口功能VS按性能分组

1 个答案: