Question

我的数据框类似于：

df = spark.createDataFrame([(0, "departmentcode__50~#~p99189h8pk0__10483~#~prod_productcolor__Dustysalmon Pink","departmentcode__50~#~p99189h8pk0__10483~#~prod_productcolor__Dustysalmon Blue"), (1, "departmentcode__10~#~p99189h8pk0__10484~#~prod_productcolor__Dustysalmon Black","departmentcode__50~#~p99189h8pk0__10483~#~prod_productcolor__Dustysalmon Blue"), (2, "departmentcode__60~#~p99189h8pk0__10485~#~prod_productcolor__Dustysalmon White","departmentcode__50~#~p99189h8pk0__10483~#~prod_productcolor__Dustysalmon Blue")], ["id", "items_base", "item_target"])

我需要一个类似于followng的新数据框：

+---+-----------------+----------------+--------+------+
|id |dept0            |att0            |position|flag  |
+---+-----------------+----------------+--------+------+
|0  |departmentcode   |50              |1       |Base  |
|0  |p99189h8pk0      |10483           |2       |Base  |
|0  |prod_productcolor|Dustysalmon Pink|3       |Base  |
|0  |departmentcode   |50              |1       |Target|
|0  |p99189h8pk0      |10483           |2       |Target|
|0  |prod_productcolor|Dustysalmon Blue|3       |Target|
|1  |departmentcode   |10              |1       |Base  |
...
...
+---+-----------------+----------------+--------+------+

我将items_base和item_target拆分为〜＃〜和__并创建新的6行。 items_base中有3行，item_target中有3行（其中，position是分割操作后dept0的位置，flag表示是item_base还是items_target）

Answer 1

你必须做很多步骤来实现结果，但它们并不是很复杂。

base_df = df.select(
    'id',
    F.split('items_base', '~#~').alias('items_base')
).select(
    'id',
    F.posexplode('items_base')
).select(
    'id',
    F.split('col', '__').alias('items_base'),
    (F.col('pos')+1).alias('position'),
    F.lit('Base').alias('flag')
).select(
    'id',
    F.col('items_base').getItem(0).alias('dept0'),
    F.col('items_base').getItem(1).alias('att0'),
    'position',
    'flag',
)


target_df = df.select(
    'id',
    F.split('item_target', '~#~').alias('item_target')
).select(
    'id',
    F.posexplode('item_target')
).select(
    'id',
    F.split('col', '__').alias('item_target'),
    (F.col('pos')+1).alias('position'),
    F.lit('Target').alias('flag')
).select(
    'id',
    F.col('item_target').getItem(0).alias('dept0'),
    F.col('item_target').getItem(1).alias('att0'),
    'position',
    'flag',
)

base_df.union(target_df).show()

+---+-----------------+-----------------+--------+------+
| id|            dept0|             att0|position|  flag|
+---+-----------------+-----------------+--------+------+
|  0|   departmentcode|               50|       1|  Base|
|  0|      p99189h8pk0|            10483|       2|  Base|
|  0|prod_productcolor| Dustysalmon Pink|       3|  Base|
|  1|   departmentcode|               10|       1|  Base|
|  1|      p99189h8pk0|            10484|       2|  Base|
|  1|prod_productcolor|Dustysalmon Black|       3|  Base|
|  2|   departmentcode|               60|       1|  Base|
|  2|      p99189h8pk0|            10485|       2|  Base|
|  2|prod_productcolor|Dustysalmon White|       3|  Base|
|  0|   departmentcode|               50|       1|Target|
|  0|      p99189h8pk0|            10483|       2|Target|
|  0|prod_productcolor| Dustysalmon Blue|       3|Target|
|  1|   departmentcode|               50|       1|Target|
|  1|      p99189h8pk0|            10483|       2|Target|
|  1|prod_productcolor| Dustysalmon Blue|       3|Target|
|  2|   departmentcode|               50|       1|Target|
|  2|      p99189h8pk0|            10483|       2|Target|
|  2|prod_productcolor| Dustysalmon Blue|       3|Target|
+---+-----------------+-----------------+--------+------+

Answer 2

您可以使用flatMap将长度为N的RDD转换为N个集合的集合：

from pyspark.sql import Row

def etl(row) :
  list_row = []
  items_base = row.items_base.split('~#~')
  for item in items_base:
      row_items_base = Row(id = row.id, dept0 = item.split('__')[0], att0 = item.split('__')[1],  position = items_base.index(item) + 1, flag = 'Base')
      list_row.append(row_items_base)

  item_target = row.item_target.split('~#~')
  for item in item_target:
      row_items_base = Row(id = row.id, dept0 = item.split('__')[0], att0 = item.split('__')[1],  position = item_target.index(item) + 1, flag = 'Target')
      list_row.append(row_items_base)

  return list_row 


df.rdd.flatMap(etl).toDF().show()

输出：

Answer 3

您可以使用udf函数拆分和合并拆分字符串，最后使用explode和select函数获取最终数据框

from pyspark.sql import functions as f
from pyspark.sql import types as t
@f.udf(t.ArrayType(t.ArrayType(t.StringType())))
def splitUdf(base, target):
    return [s.split("__") + [str(index+1), 'base'] for index, s in enumerate(base.split("~#~"))] + [s.split("__") + [str(index+1), 'target'] for index, s in enumerate(target.split("~#~"))]

df.withColumn('exploded', f.explode(splitUdf(f.col('items_base'), f.col('item_target'))))\
    .select(f.col('id'), f.col('exploded')[0].alias('dept0'), f.col('exploded')[1].alias('att0'), f.col('exploded')[2].alias('position'), f.col('exploded')[3].alias('flag'))\
    .show(truncate=False)

应该给你

+---+-----------------+-----------------+--------+------+
|id |dept0            |att0             |position|flag  |
+---+-----------------+-----------------+--------+------+
|0  |departmentcode   |50               |1       |base  |
|0  |p99189h8pk0      |10483            |2       |base  |
|0  |prod_productcolor|Dustysalmon Pink |3       |base  |
|0  |departmentcode   |50               |1       |target|
|0  |p99189h8pk0      |10483            |2       |target|
|0  |prod_productcolor|Dustysalmon Blue |3       |target|
|1  |departmentcode   |10               |1       |base  |
|1  |p99189h8pk0      |10484            |2       |base  |
|1  |prod_productcolor|Dustysalmon Black|3       |base  |
|1  |departmentcode   |50               |1       |target|
|1  |p99189h8pk0      |10483            |2       |target|
|1  |prod_productcolor|Dustysalmon Blue |3       |target|
|2  |departmentcode   |60               |1       |base  |
|2  |p99189h8pk0      |10485            |2       |base  |
|2  |prod_productcolor|Dustysalmon White|3       |base  |
|2  |departmentcode   |50               |1       |target|
|2  |p99189h8pk0      |10483            |2       |target|
|2  |prod_productcolor|Dustysalmon Blue |3       |target|
+---+-----------------+-----------------+--------+------+

我希望答案很有帮助

<强>更新

如果从udf函数返回结构类型

，甚至更好

@f.udf(t.ArrayType(t.StructType([t.StructField('dept0', t.StringType(), True), t.StructField('att0', t.StringType(), True), t.StructField('position', t.IntegerType(), True), t.StructField('flag', t.StringType(), True)])))
def splitUdf(base, target):
    return [(s.split("__")[0], s.split("__")[1], index+1, 'base') for index, s in enumerate(base.split("~#~"))] + [(s.split("__")[0], s.split("__")[1], index+1, 'target') for index, s in enumerate(target.split("~#~"))]

df.withColumn('exploded', f.explode(splitUdf(f.col('items_base'), f.col('item_target'))))\
    .select(f.col('id'), f.col('exploded.*'))\
    .show(truncate=False)

应该给你相同的结果

将一行拆分为多行pyspark

3 个答案: