For Spark >=1.5, you are provided with numbers of date processing functions and you can use these functions in your case.
pyspark.sql.functions.year
pyspark.sql.functions.month
pyspark.sql.functions.dayofmonth
pyspark.sql.functions.dayofweek()
pyspark.sql.functions.dayofyear
pyspark.sql.functions.weekofyear()
import datetime
from pyspark.sql.functions import year, month, dayofmonth
elevDF = sc.parallelize([
(datetime.datetime(1984, 1, 1, 0, 0), 1, 638.55),
(datetime.datetime(1984, 1, 1, 0, 0), 2, 638.55),
(datetime.datetime(1984, 1, 1, 0, 0), 3, 638.55),
(datetime.datetime(1984, 1, 1, 0, 0), 4, 638.55),
(datetime.datetime(1984, 1, 1, 0, 0), 5, 638.55)
]).toDF(["date", "hour", "value"])
elevDF.select(
year("date").alias('year'),
month("date").alias('month'),
dayofmonth("date").alias('day')
).show()
# +----+-----+---+
# |year|month|day|
# +----+-----+---+
# |1984| 1| 1|
# |1984| 1| 1|
# |1984| 1| 1|
# |1984| 1| 1|
# |1984| 1| 1|
# +----+-----+---+
Now, I would simply suggest you to use map with any other RDD:
elevDF = sqlContext.createDataFrame(sc.parallelize([
Row(date=datetime.datetime(1984, 1, 1, 0, 0), hour=1, value=638.55),
Row(date=datetime.datetime(1984, 1, 1, 0, 0), hour=2, value=638.55),
Row(date=datetime.datetime(1984, 1, 1, 0, 0), hour=3, value=638.55),
Row(date=datetime.datetime(1984, 1, 1, 0, 0), hour=4, value=638.55),
Row(date=datetime.datetime(1984, 1, 1, 0, 0), hour=5, value=638.55)]))
(elevDF
.map(lambda (date, hour, value): (date.year, date.month, date.day))
.collect())
and the result is:
[(1984, 1, 1), (1984, 1, 1), (1984, 1, 1), (1984, 1, 1), (1984, 1, 1)]
Note: datetime.datetime stores an hour anyway so keeping it separately seems to be a waste of memory.
If you wish to learn Spark visit this Spark Tutorial.