Groupby with when condition in Pyspark - apache-spark-sql

Question

My data frame looks like
id |reg_date | txn_date|
+----------+----------+--------------------+
|1 |2019-01-06| 2019-02-15 12:51:15|
|1 |2019-01-06| 2019-03-29 13:15:27|
|1 |2019-01-06| 2019-06-01 01:42:57|
|1 |2019-01-06| 2019-01-06 17:01:...|
|5 |2019-06-16| 2019-07-19 11:50:34|
|5 |2019-06-16| 2019-07-13 19:49:39|
|5 |2019-06-16| 2019-08-27 17:37:22|
|2 |2018-07-30| 2019-01-01 07:03:...|
|2 |2018-07-30| 2019-07-30 01:27:57|
|2 |2018-07-30| 2019-02-01 00:08:35
I want to pickup the 1st txn_date after reg_date , i.e. the first txn_date of reg_date >= txn_date.
Expected output
id |reg_date | txn_date|
+----------+----------+--------------------+
|1 |2019-01-06| 2019-01-06 17:01:...|
|5 |2019-06-16| 2019-07-13 19:49:39|
|2 |2018-07-30| 2019-07-30 01:27:57|
I have done so far,
df = df.withColumn('txn_date',to_date(unix_timestamp(F.col('txn_date'),'yyyy-MM-dd HH:mm:ss').cast("timestamp")))
df = df.withColumn('reg_date',to_date(unix_timestamp(F.col('reg_date'),'yyyy-MM-dd').cast("timestamp")))
gg = df.groupBy('id','reg_date').agg(min(F.col('txn_date')))
But getting wrong results.

Answer 1

The condition reg_date >= txn_date can be ambiguous.
Does 2019-01-06>=2019-01-06 17:01:30 mean 2019-01-06 00:00:00>=2019-01-06 17:01:30 or 2019-01-06 23:59:59>=2019-01-06 17:01:30?
In your example, 2019-01-06>=2019-01-06 17:01:30 is evaluated to be true, so I assume it is the latter case, i.e. the case with 23:59:59.
Proceeding with the assumption above, here is how I coded it.
import pyspark.sql.functions as F
#create a sample data frame
data = [('2019-01-06','2019-02-15 12:51:15'),('2019-01-06','2019-03-29 13:15:27'),('2019-01-06','2019-01-06 17:01:30'),\
('2019-07-30','2019-07-30 07:03:01'),('2019-07-30','2019-07-30 01:27:57'),('2019-07-30','2019-07-30 00:08:35')]
cols = ('reg_date', 'txn_date')
df = spark.DataFrame(data,cols)
#add 23:59:59 to reg_date as a dummy_date for a timestamp comparison later
df = df.withColumn('dummy_date', F.concat(F.col('reg_date'), F.lit(' 23:59:59')))
#convert columns to the appropriate time data types
df = df.select([F.to_date(F.col('reg_date'),'yyyy-MM-dd').alias('reg_date'),\
F.to_timestamp(F.col('txn_date'),'yyyy-MM-dd HH:mm:ss').alias('txn_date'),\
F.to_timestamp(F.col('dummy_date'),'yyyy-MM-dd HH:mm:ss').alias('dummy_date')])
#implementation part
(df.orderBy('reg_date')
.filter(F.col('dummy_date')>=F.col('txn_date'))
.groupBy('reg_date')
.agg(F.first('txn_date').alias('txn_date'))
.show())
#+----------+----------------------+
#| reg_date| txn_date|
#+----------+----------------------+
#|2019-01-06| 2019-01-06 17:01:30|
#|2019-07-30| 2019-07-30 07:03:01|
#+----------+----------------------+

Answer 2

You don't need to order. You can discard all smaller values with a filter, then aggregate by id and get the smaller timestamp, because the first timestamp will be the minimum. Something like:
df.filter(df.reg_date >= df.txn_date) \
.groupBy(df.reg_date) \
.agg(F.min(df.txn_date)) \
.show()