Pyspark - how to backfill a DataFrame?

问题

How can you do the same thing as df.fillna(method='bfill') for a pandas dataframe with a pyspark.sql.DataFrame?

The pyspark dataframe has the pyspark.sql.DataFrame.fillna method, however there is no support for a method parameter.

In pandas you can use the following to backfill a time series:

Create data

import pandas as pd

index = pd.date_range('2017-01-01', '2017-01-05')
data = [1, 2, 3, None, 5]

df = pd.DataFrame({'data': data}, index=index)

Giving

Out[1]:
            data
2017-01-01  1.0
2017-01-02  2.0
2017-01-03  3.0
2017-01-04  NaN
2017-01-05  5.0

Backfill the dataframe

df = df.fillna(method='bfill')

Produces the backfilled frame

Out[2]:
            data
2017-01-01  1.0
2017-01-02  2.0
2017-01-03  3.0
2017-01-04  5.0
2017-01-05  5.0

How can the same thing be done for a pyspark.sql.DataFrame?

回答1:

Actually backfill on distributed dataset is not as easy task as in pandas (local) dataframe - you cannot be sure that value to fill exists in the same partition. I would use crossJoin with windowing, for example fo DF:

df = spark.createDataFrame([
    ('2017-01-01', None), 
    ('2017-01-02', 'B'), 
    ('2017-01-03', None), 
    ('2017-01-04', None), 
    ('2017-01-05', 'E'), 
    ('2017-01-06', None), 
    ('2017-01-07', 'G')], ['date', 'value'])
df.show()

+----------+-----+
|      date|value|
+----------+-----+
|2017-01-01| null|
|2017-01-02|    B|
|2017-01-03| null|
|2017-01-04| null|
|2017-01-05|    E|
|2017-01-06| null|
|2017-01-07|    G|
+----------+-----+

The code would be:

from pyspark.sql.window import Window

df.alias('a').crossJoin(df.alias('b')) \
    .where((col('b.date') >= col('a.date')) & (col('a.value').isNotNull() | col('b.value').isNotNull())) \
    .withColumn('rn', row_number().over(Window.partitionBy('a.date').orderBy('b.date'))) \
    .where(col('rn') == 1) \
    .select('a.date', coalesce('a.value', 'b.value').alias('value')) \
    .orderBy('a.date') \
    .show()

+----------+-----+
|      date|value|
+----------+-----+
|2017-01-01|    B|
|2017-01-02|    B|
|2017-01-03|    E|
|2017-01-04|    E|
|2017-01-05|    E|
|2017-01-06|    G|
|2017-01-07|    G|
+----------+-----+

来源：https://stackoverflow.com/questions/43773520/pyspark-how-to-backfill-a-dataframe

标签

pandas

pyspark

spark-dataframe

pyspark-sql