PySpark group by with rolling window

Question

Suppose I have a table with three columns: dt, id and value.

df_tmp = spark.createDataFrame([('2023-01-01', 1001, 5),
                                ('2023-01-15', 1001, 3),
                                ('2023-02-10', 1001, 1),
                                ('2023-02-20', 1001, 2),
                                ('2023-01-02', 1002, 7),
                                ('2023-01-02', 1002, 6),
                                ('2023-01-03', 1002, 1)],
                               ["date", "id", "value"])
df.show()
# +----------+----+-----+
# |      date|  id|value|
# +----------+----+-----+
# |2023-01-01|1001|    5|
# |2023-01-15|1001|    3|
# |2023-02-10|1001|    1|
# |2023-02-20|1001|    2|
# |2023-01-02|1002|    7|
# |2023-01-02|1002|    6|
# |2023-01-03|1002|    1|
# +----------+----+-----+

I would like to compute the 30-day rolling sum of value grouped by id for every date, and additionally, a number of distinct dates that the id was seen. Something that would look like this:

+----------+----+-----+----------------+-------------------------+
|      date|  id|value|30_day_value_sum|days_seen_in_past_30_days|
+----------+----+-----+----------------+-------------------------+
|2023-01-01|1001|    5|               0|                        0|
|2023-01-15|1001|    3|               0|                        1|
|2023-02-10|1001|    1|               3|                        1|
|2023-02-20|1001|    2|               1|                        2|
|2023-01-02|1002|    7|               0|                        0|
|2023-01-02|1002|    6|               7|                        1|
|2023-01-03|1002|    1|              13|                        2|
+----------+----+-----+----------------+-------------------------+

I suspect one could do it using Window but am not clear about the explicit details.

I believe the sum is incorrect for id=1001. Also, why is it that id=1002 has two identic dates 2023-01-02, but calculations are different in those rows? — ZygD, Aug 04 '23 at 15:42

ZygD · Accepted Answer · 2023-08-04T16:51:04.347

I assume that originally you have date data type in your dataframe, so I used a bit modified input than yours:

from pyspark.sql import functions as F, Window as W
df_tmp = spark.createDataFrame(
    [('2023-01-01', 1001, 5),
     ('2023-01-15', 1001, 3),
     ('2023-02-10', 1001, 1),
     ('2023-02-20', 1001, 2),
     ('2023-01-01', 1002, 7),
     ('2023-01-02', 1002, 6),
     ('2023-01-03', 1002, 1)],
    ["date", "id", "value"]
).withColumn('date', F.col('date').cast('date'))

In this case, the following window should work:

w = W.partitionBy('id').orderBy(F.expr("unix_date(date)")).rangeBetween(-30, -1)
df = df_tmp.withColumn('30_day_value_sum', F.sum('value').over(w)) \
           .withColumn('days_seen_in_past_30_days', F.count('id').over(w))
df = df.fillna(0, subset=['30_day_value_sum'])

df.show()
# +----------+----+-----+----------------+-------------------------+
# |      date|  id|value|30_day_value_sum|days_seen_in_past_30_days|
# +----------+----+-----+----------------+-------------------------+
# |2023-01-01|1001|    5|               0|                        0|
# |2023-01-15|1001|    3|               5|                        1|
# |2023-02-10|1001|    1|               3|                        1|
# |2023-02-20|1001|    2|               1|                        1|
# |2023-01-01|1002|    7|               0|                        0|
# |2023-01-02|1002|    6|               7|                        1|
# |2023-01-03|1002|    1|              13|                        2|
# +----------+----+-----+----------------+-------------------------+

If you have string type in the "date" column, then you should use the following .orderBy` clause:

.orderBy(F.expr("unix_date(to_date(date))")).rangeBetween(-30, -1)

More options and details are here.

PySpark group by with rolling window

1 Answers1