I have a PYSPARK dataframe which is sorted ('timestamp' and 'ship' ascend):
+----------------------+------+
| timestamp | ship |
+----------------------+------+
| 2018-08-01 06:01:00 | 1 |
| 2018-08-01 06:01:30 | 1 |
| 2018-08-01 09:00:00 | 1 |
| 2018-08-01 09:00:00 | 2 |
| 2018-08-01 10:15:43 | 2 |
| 2018-08-01 11:00:01 | 3 |
| 2018-08-01 06:00:13 | 4 |
| 2018-08-01 13:00:00 | 4 |
| 2018-08-13 14:00:00 | 5 |
| 2018-08-13 14:15:03 | 5 |
| 2018-08-13 14:45:08 | 5 |
| 2018-08-13 14:50:00 | 5 |
+-----------------------------+
I want to add a new column to the dataframe called 'trip'. A trip is defined as a ship-number which sails within 2 hours from the start of the ship-record in the dataframe. If within the two hours the ship number changes, a new trip number should be added to the dataframe column 'trip'.
Desired output looks like:
+----------------------+------+-------+
| timestamp | ship | trip |
+----------------------+------+-------+
| 2018-08-01 06:01:00 | 1 | 1 | # start new ship number
| 2018-08-01 06:01:30 | 1 | 1 | # still within 2 hours of same ship number
| 2018-08-01 09:00:00 | 1 | 2 | # more than 2 hours of same ship number = new trip
| 2018-08-01 09:00:00 | 2 | 3 | # new ship number = new trip
| 2018-08-01 10:15:43 | 2 | 3 | # still within 2 hours of same ship number
| 2018-08-01 11:00:01 | 3 | 4 | # new ship number = new trip
| 2018-08-01 06:00:13 | 4 | 5 | # new ship number = new trip
| 2018-08-01 13:00:00 | 4 | 6 | # more than 2 hours of same ship number = new trip
| 2018-08-13 14:00:00 | 5 | 7 | # new ship number = new trip
| 2018-08-13 14:15:03 | 5 | 7 | # still within 2 hours of same ship number
| 2018-08-13 14:45:08 | 5 | 7 | # still within 2 hours of same ship number
| 2018-08-13 14:50:00 | 5 | 7 | # still within 2 hours of same ship number
+-----------------------------+-------+
In Pandas it would be done as such:
dt_trip = 2 # time duration trip per ship (in hours)
total_time = df['timestamp'] - df.groupby('name')['timestamp'].transform('min')
trips = total_time.dt.total_seconds().fillna(0)//(dt_trip*3600)
df['trip'] = df.groupby(['name', trips]).ngroup()+1
How would this be done in PYSPARK?
Try this using window functions, row_number(), collect_list(), and an incremental sum over conditions created.
from pyspark.sql import functions as F
from pyspark.sql.window import Window
w1=Window().partitionBy("ship").orderBy(F.unix_timestamp("timestamp")).rangeBetween(-7199, Window.currentRow)
w2=Window().partitionBy("ship").orderBy("timestamp")
w3=Window().orderBy("ship","timestamp")
df.withColumn("trip", F.sum(F.when(F.row_number().over(w2)==1, F.lit(1))\
.when(F.size(F.collect_list("ship").over(w1))==1, F.lit(1))\
.otherwise(F.lit(0))).over(w3)).orderBy("ship","timestamp").show()
#+-------------------+----+----+
#| timestamp|ship|trip|
#+-------------------+----+----+
#|2018-08-01 06:01:00| 1| 1|
#|2018-08-01 06:01:30| 1| 1|
#|2018-08-01 09:00:00| 1| 2|
#|2018-08-01 09:00:00| 2| 3|
#|2018-08-01 10:15:43| 2| 3|
#|2018-08-01 11:00:01| 3| 4|
#|2018-08-01 06:00:13| 4| 5|
#|2018-08-01 13:00:00| 4| 6|
#|2018-08-13 14:00:00| 5| 7|
#|2018-08-13 14:15:03| 5| 7|
#|2018-08-13 14:45:08| 5| 7|
#|2018-08-13 14:50:00| 5| 7|
#+-------------------+----+----+