问题
We can find the rolling/moving average of a time series data using window function in pyspark.
The data I am dealing with doesn't have any timestamp column but it does have a strictly increasing column frame_number. Data looks like this.
d = [
{'session_id': 1, 'frame_number': 1, 'rtd': 11.0, 'rtd2': 11.0,},
{'session_id': 1, 'frame_number': 2, 'rtd': 12.0, 'rtd2': 6.0},
{'session_id': 1, 'frame_number': 3, 'rtd': 4.0, 'rtd2': 233.0},
{'session_id': 1, 'frame_number': 4, 'rtd': 110.0, 'rtd2': 111.0,},
{'session_id': 1, 'frame_number': 5, 'rtd': 13.0, 'rtd2': 6.0},
{'session_id': 1, 'frame_number': 6, 'rtd': 43.0, 'rtd2': 233.0},
{'session_id': 1, 'frame_number': 7, 'rtd': 11.0, 'rtd2': 111.0,}]
df = spark.createDataFrame(d)
+------------+-----+-----+----------+
|frame_number| rtd| rtd2|session_id|
+------------+-----+-----+----------+
| 1| 11.0| 11.0| 1|
| 2| 12.0| 6.0| 1|
| 3| 4.0|233.0| 1|
| 4|110.0|111.0| 1|
| 5| 13.0| 6.0| 1|
| 6| 43.0|233.0| 1|
| 7| 11.0|111.0| 1|
+------------+-----+-----+----------+
I want to find the rolling average of the column rtd on the strictly increasing column frame_number.
I am trying something like this (using collect_list).
window_size=2
w = Window.partitionBy("session_id").orderBy("frame_number").rowsBetween(0, window_size)
df_lists = df.withColumn('rtd_list', F.collect_list('rtd').over(w))
+------------+-----+-----+----------+-------------------+
|frame_number| rtd| rtd2|session_id| rtd_list|
+------------+-----+-----+----------+-------------------+
| 1| 11.0| 11.0| 1| [11.0, 12.0, 4.0]|
| 2| 12.0| 6.0| 1| [12.0, 4.0, 110.0]|
| 3| 4.0|233.0| 1| [4.0, 110.0, 13.0]|
| 4|110.0|111.0| 1|[110.0, 13.0, 43.0]|
| 5| 13.0| 6.0| 1| [13.0, 43.0, 11.0]|
| 6| 43.0|233.0| 1| [43.0, 11.0]|
| 7| 11.0|111.0| 1| [11.0]|
+------------+-----+-----+----------+-------------------+
And then applying a UDF to get moving average.
windudf = F.udf( lambda v: str(np.nanmean(v)), StringType())
out = df_lists.withColumn("moving_average", windudf("rtd_list"))
+------------+-----+-----+----------+-------------------+------------------+
|frame_number| rtd| rtd2|session_id| rtd_list| moving_average|
+------------+-----+-----+----------+-------------------+------------------+
| 1| 11.0| 11.0| 1| [11.0, 12.0, 4.0]| 9.0|
| 2| 12.0| 6.0| 1| [12.0, 4.0, 110.0]| 42.0|
| 3| 4.0|233.0| 1| [4.0, 110.0, 13.0]|42.333333333333336|
| 4|110.0|111.0| 1|[110.0, 13.0, 43.0]|55.333333333333336|
| 5| 13.0| 6.0| 1| [13.0, 43.0, 11.0]|22.333333333333332|
| 6| 43.0|233.0| 1| [43.0, 11.0]| 27.0|
| 7| 11.0|111.0| 1| [11.0]| 11.0|
+------------+-----+-----+----------+-------------------+------------------+
Issue with above method is that it cannot define a slide duration for a window. Above method calculates moving average for evrey frame. I wnt to move my window by some amount before finding the average. Any ways to achieve this?
回答1:
Define window:
from pyspark.sql import functions as F
w = F.window(
F.col("frame_number").cast("timestamp"),
# Just example
windowDuration="10 seconds",
slideDuration="5 seconds",
).alias("window")
(df
.groupBy(w, F.col("session_id"))
.avg("rtd", "rtd2")
.withColumn("window", F.col("window").cast("struct<start:long,end:long>"))
.orderBy("window.start")
.show())
# +------+----------+------------------+------------------+
# |window|session_id| avg(rtd)| avg(rtd2)|
# +------+----------+------------------+------------------+
# |[-5,5]| 1| 34.25| 90.25|
# |[0,10]| 1|29.142857142857142|101.57142857142857|
# |[5,15]| 1|22.333333333333332|116.66666666666667|
# +------+----------+------------------+------------------+
Also please don't use collect_list with udf to compute average. It give no benefits and has severe performance implications.
来源：https://stackoverflow.com/questions/49807909/rolling-average-without-timestamp-in-pyspark