datetime range filter in PySpark SQL

Question

What is the correct way to filter data frame by timestamp field?

I have tried different date formats and forms of filtering, nothing helps: either pyspark returns 0 objects, or throws an error that it doesn't understand datetime format

Here is what i got so far:

from pyspark import SparkContext
from pyspark.sql import SQLContext

from django.utils import timezone
from django.conf import settings

from myapp.models import Collection

sc = SparkContext("local", "DjangoApp")
sqlc = SQLContext(sc)
url = "jdbc:postgresql://%(HOST)s/%(NAME)s?user=%(USER)s&password=%(PASSWORD)s" % settings.DATABASES['default']
sf = sqlc.load(source="jdbc", url=url, dbtable='myapp_collection')

range for timestamp field:

system_tz = timezone.pytz.timezone(settings.TIME_ZONE)
date_from = datetime.datetime(2014, 4, 16, 18, 30, 0, 0, tzinfo=system_tz)
date_to = datetime.datetime(2015, 6, 15, 18, 11, 59, 999999, tzinfo=system_tz)

attempt 1

date_filter = "my_col >= '%s' AND my_col <= '%s'" % (
    date_from.isoformat(), date_to.isoformat()
)
sf = sf.filter(date_filter)
sf.count()

Out[12]: 0

attempt 2

sf = sf.filter(sf.my_col >= date_from).filter(sf.my_col <= date_to)
sf.count()

---------------------------------------------------------------------------
Py4JJavaError: An error occurred while calling o63.count.
: org.apache.spark.SparkException: Job aborted due to stage failure:
Task 0 in stage 4.0 failed 1 times, most recent failure: 
Lost task 0.0 in stage 4.0 (TID 3, localhost): org.postgresql.util.PSQLException: 
ERROR: syntax error at or near "18"
# 
# ups.. JDBC doesn't understand 24h time format??

attempt 3

sf = sf.filter("my_col BETWEEN '%s' AND '%s'" % \
     (date_from.isoformat(), date_to.isoformat())
     )
---------------------------------------------------------------------------
Py4JJavaError: An error occurred while calling o97.count.
: org.apache.spark.SparkException: Job aborted due to stage failure:
Task 0 in stage 17.0 failed 1 times, most recent failure:
Lost task 0.0 in stage 17.0 (TID 13, localhost): org.postgresql.util.PSQLException:
ERROR: syntax error at or near "18"

the data do exist in the table, though:

django_filters = {
    'my_col__gte': date_from,
    'my_col__lte': date_to
    }
Collection.objects.filter(**django_filters).count()

Out[17]: 1093436

Or this way

django_range_filter = {'my_col__range': (date_from, date_to)}
Collection.objects.filter(**django_range_filter).count()

Out[19]: 1093436

Answer 1

The following seems to be working for me (someone let me know if this is bad form or inaccurate though)...

First, create a new column for each end of the window (in this example, it's 100 days to 200 days after the date in column: column_name.

from pyspark.sql import functions as F
new_df = new_df.withColumn('After100Days', F.lit(F.date_add(new_df['column_name'], 100)))
new_df = new_df.withColumn('After200Days', F.lit(F.date_add(new_df['column_name'], 200)))

Filter as follows...

For filtering dates inside a particular range:

result= df.where((df.col1> df.col2) & (df.col1 < df.col3))

For filtering dates outside a particular range:

result= df.where((df.col1 < df.col2) | (df.col1 > df.col3))

Answer 2

How about something like this:

import pyspark.sql.functions as func

df = df.select(func.to_date(df.my_col).alias("time"))
sf = df.filter(df.time > date_from).filter(df.time < date_to)

Answer 3

Lets assume your data frame looks as follows:

sf = sqlContext.createDataFrame([
    [datetime.datetime(2013, 6, 29, 11, 34, 29)],
    [datetime.datetime(2015, 7, 14, 11, 34, 27)],
    [datetime.datetime(2012, 3, 10, 19, 00, 11)],
    [datetime.datetime(2016, 2, 8, 12, 21)],
    [datetime.datetime(2014, 4, 4, 11, 28, 29)]
], ('my_col', ))

with schema:

root
 |-- my_col: timestamp (nullable = true)

and you want to find dates in a following range:

import datetime, time 
dates = ("2013-01-01 00:00:00",  "2015-07-01 00:00:00")

timestamps = (
    time.mktime(datetime.datetime.strptime(s, "%Y-%m-%d %H:%M:%S").timetuple())
    for s in dates)

It is possible to query using timestamps either computed on a driver side:

q1 = "CAST(my_col AS INT) BETWEEN {0} AND {1}".format(*timestamps)
sf.where(q1).show()

or using unix_timestamp function:

q2 = """CAST(my_col AS INT)
        BETWEEN unix_timestamp('{0}', 'yyyy-MM-dd HH:mm:ss')
        AND unix_timestamp('{1}', 'yyyy-MM-dd HH:mm:ss')""".format(*dates)

sf.where(q2).show()

It is also possible to use udf in a similar way I described in an another answer.

If you use raw SQL it is possible to extract different elements of timestamp using year, date, etc.

sqlContext.sql("""SELECT * FROM sf
    WHERE YEAR(my_col) BETWEEN 2014 AND 2015").show()

EDIT:

Since Spark 1.5 you can use built-in functions:

dates = ("2013-01-01",  "2015-07-01")
date_from, date_to = [to_date(lit(s)).cast(TimestampType()) for s in dates]

sf.where((sf.my_col > date_from) & (sf.my_col < date_to))

You can also use pyspark.sql.Column.between, which is inclusive of the bounds:

from pyspark.sql.functions import col
sf.where(col('my_col').between(*dates)).show(truncate=False)
#+---------------------+
#|my_col               |
#+---------------------+
#|2013-06-29 11:34:29.0|
#|2014-04-04 11:28:29.0|
#+---------------------+