如何使用 window 而不是 Pyspark groupBy 进行聚合

Question

我在使用 window 函数而不是 GroupBy 来聚合每个用户时遇到问题，在我的例子中 110 和 220 用户 ID：

1- 计算每个 p_uuid

的行数

2- 为每个 p_uuid

创建带有最小和最大时间戳的新列

df = spark.createDataFrame([(1, 110, 'aaa', 'walk', 'work', '2019-09-28 13:40:19-04:00'),
                        (2, 110, 'aaa', 'walk', 'work', '2019-09-28 13:40:19-04:01'),
                        (3, 110, 'aaa', 'walk', 'work', '2019-09-28 13:40:19-04:02'),
                        (4, 110, 'aaa', 'metro', 'work', '2019-09-28 13:41:19-04:00'),
                        (5, 110, 'aaa', 'metro', 'work', '2019-09-28 13:41:19-04:01'),
                        (6, 110, 'aaa', 'walk', 'work', '2019-09-28 13:42:19-04:00'),
                        (7, 110, 'aaa', 'walk', 'work', '2019-09-28 13:42:19-04:01'),
                        (8, 110, 'bbb', 'bike', 'home', '2019-09-17 14:40:19-04:00'),
                        (9, 110, 'bbb', 'bus', 'home', '2019-09-17 14:41:19-04:00'),
                        (10, 110, 'bbb', 'walk', 'home', '2019-09-17 14:43:19-04:00'),
                        (16, 110, 'ooo', None, None, '2019-08-29 16:01:19-04:00'),
                        (17, 110, 'ooo', None, None, '2019-08-29 16:02:19-04:00'),
                        (18, 110, 'ooo', None, None, '2019-08-29 16:02:19-04:00'),
                        (19, 222, 'www', 'car', 'work', '2019-09-28 08:00:19-04:00'),
                        (20, 222, 'www', 'metro', 'work', '2019-09-28 08:01:19-04:00'),
                        (21, 222, 'www', 'walk', 'work', '2019-09-28 08:02:19-04:00'),
                        (22, 222, 'xxx', 'walk', 'friend', '2019-09-17 08:40:19-04:00'),
                        (23, 222, 'xxx', 'bike', 'friend', '2019-09-17 08:42:19-04:00'),
                        (24, 222, 'xxx', 'bus', 'friend', '2019-09-17 08:43:19-04:00'),
                        (30, 222, 'ooo', None, None, '2019-08-29 10:00:19-04:00'),
                        (31, 222, 'ooo', None, None, '2019-08-29 10:01:19-04:00'),
                        (32, 222, 'ooo', None, None, '2019-08-29 10:02:19-04:00')],
                    ['idx', 'u_uuid', 'p_uuid', 'mode', 'place', 'timestamp']
                )
 df.show(30, False)

我用过

 win = Window.partitionBy("u_uuid", "p_uuid").orderBy("timestamp")
 df.withColumn("count_", F.count('p_uuid').over(win))
 df.withColumn("max_timestamp", F.max("timestamp").over(win))
 df.withColumn("min_timestamp", F.min("timestamp").over(win))

它似乎不起作用（例如：获取 max_）

remarque：忘记 trip_id、subtrip_id 和 track_id 列

Answer 1

您需要使用 unboundedPreceding,unboundedFollowing 和 .partitionBy 子句，如果我们提供 orderBy clause，则默认值为 unboundedPreceding,currentRow。

在您的 window 规范中添加 .rowsBetween，然后再次添加运行。

win = Window.partitionBy("u_uuid", "p_uuid").orderBy("timestamp").rowsBetween(Window.unboundedPreceding, Window.unboundedFollowing)

Example:

df.withColumn("max_timestamp", max("timestamp").over(win)).show(10,False)
+---+------+------+-----+------+-------------------------+-------------------------+
|idx|u_uuid|p_uuid|mode |place |timestamp                |max_timestamp            |
+---+------+------+-----+------+-------------------------+-------------------------+
|8  |110   |bbb   |bike |home  |2019-09-17 14:40:19-04:00|2019-09-17 14:43:19-04:00|
|9  |110   |bbb   |bus  |home  |2019-09-17 14:41:19-04:00|2019-09-17 14:43:19-04:00|
|10 |110   |bbb   |walk |home  |2019-09-17 14:43:19-04:00|2019-09-17 14:43:19-04:00|
|16 |110   |ooo   |null |null  |2019-08-29 16:01:19-04:00|2019-08-29 16:02:19-04:00|
|17 |110   |ooo   |null |null  |2019-08-29 16:02:19-04:00|2019-08-29 16:02:19-04:00|
|18 |110   |ooo   |null |null  |2019-08-29 16:02:19-04:00|2019-08-29 16:02:19-04:00|
+---+------+------+-----+------+-------------------------+-------------------------+

Answer 2

您必须使用 rowsBetween 将 window 扩展到整个框架：

win = Window.partitionBy("u_uuid", "p_uuid").orderBy("timestamp").rowsBetween(Window.unboundedPreceding, Window.unboundedFollowing)

如何使用 window 而不是 Pyspark groupBy 进行聚合

How to aggregate using window instead of Pyspark groupBy

apache-spark

apache-spark-sql

pyspark

pyspark-dataframes