Pyspark：如何遍历数据框列？

Question

I'm new to pyspark. I usually work with pandas. I to iterate through row by row using a column in pyspark. My dataset looks like:-

+-------------------+--------------------+--------+-----+
|           DateTime|           user_name|keyboard|mouse|
+-------------------+--------------------+--------+-----+
|2019-10-21 08:35:01|prathameshsalap@g...|   333.0|658.0|
|2019-10-21 08:35:01|vaishusawant143@g...|   447.5|  0.0|
|2019-10-21 08:35:01|     you@example.com|     0.5|  1.0|
|2019-10-21 08:40:01|     you@example.com|     0.0|  0.0|
|2019-10-21 08:40:01|prathameshsalap@g...|   227.0|366.0|
|2019-10-21 08:40:02|vaishusawant143@g...|   472.0|  0.0|
|2019-10-21 08:45:01|     you@example.com|     0.0|  0.0|
|2019-10-21 08:45:01|prathameshsalap@g...|    35.0|458.0|
|2019-10-21 08:45:01|vaishusawant143@g...|  1659.5|  0.0|
|2019-10-21 08:50:01|     you@example.com|     0.0|  0.0|
+-------------------+--------------------+--------+-----+

In the pandas data frame it also has a given index but in spark not. In pandas:-

## pandas
usr_log = pd.read_csv("data.csv")
unique_users = usr_log.user_name.unique()
usr_log.sort_values(by='DateTime', inplace=True)
users_new_data = dict()
users_new_data[user] = {'start_time': None}

for user in unique_users:
    count_idle = 0
    ## first part of the question
    for index in usr_log.index:
        if user == usr_log['user_name'][index]:
            if users_new_data[user]['start_time'] is None:
                users_new_data[user]['start_time'] = usr_log['DateTime'][index]

            ## Second part of the question

            if usr_log['keyboard'][index] == 0 and usr_log['mouse'][index] == 0:
                count_idle += 1
            else:
                count_idle = 0
            if count_idle >= 5:
                if count_idle == 5:
                    users_new_data[usr_log['user_name'][index]]['idle_time'] \
                        = users_new_data[usr_log['user_name'][index]].get('idle_time') \
                          + datetime.timedelta(0, 1500)
                else:
                    users_new_data[usr_log['user_name'][index]]['idle_time'] \
                        = users_new_data[usr_log['user_name'][index]].get('idle_time') \
                          + datetime.timedelta(0, 300)

Same thing how can do it in spark?

For each user data generated after 5 mins(Like if the user starts at 8:30:01 the next log generated at 8:35:01). In the second question in I want to find an idle hour for each user. The calculation of idle hours is if he does not move the mouse or use the keyboard the next 30 mins(1500) then I add in user idle hours.

After converting dictionary value into data frame my expected output like:-

+--------------------+-------------------+-------------------+
|           user_name|         start_time|          idle_time|
+--------------------+-------------------+-------------------+
|prathameshsalap@g...|2019-10-21 08:35:01|2019-10-21 05:05:00|
|vaishusawant143@g...|2019-10-21 08:35:01|2019-10-21 02:15:00|
|     you@example.com|2019-10-21 08:35:01|2019-10-21 01:30:00|
+--------------------+-------------------+-------------------+

Answer 1

If you want to find for each user the first timestamp that they have you can simplify it first in pandas, do this:

usr_log[['user_name','DateTime']].groupby(['user_name']).min()

And for spark will be very similar

urs_log = sparkSession.read.csv(...)
urs_log.groupBy("user_name").agg(min("DateTime"))

you only will have to rename DateTime column to the one you want, and try to not use for loops in pandas .

In spark, you have a distributed collection and it's impossible to do a for loop, you have to apply transformations to columns, never apply logic to a single row of data.

Answer 2

Here is solution on same,

dataFrame = (spark.read.format("csv").option("sep", ",").option("header", "true").load("data.csv"))

df.show()
+-------------------+--------------------+--------+-----+
|           DateTime|           user_name|keyboard|mouse|
+-------------------+--------------------+--------+-----+
|2019-10-21 08:35:01|prathameshsalap@g...|   333.0|658.0|
|2019-10-21 08:35:01|vaishusawant143@g...|   447.5|  0.0|
|2019-10-21 08:35:01|     you@example.com|     0.5|  1.0|
|2019-10-21 08:40:01|prathameshsalap@g...|   227.0|366.0|
|2019-10-21 08:40:02|vaishusawant143@g...|   472.0|  0.0|
|2019-10-21 08:45:01|     you@example.com|     0.0|  0.0|
|2019-10-21 08:45:01|prathameshsalap@g...|    35.0|458.0|
|2019-10-21 08:45:01|vaishusawant143@g...|  1659.5|  0.0|
|2019-10-21 08:50:01|     you@example.com|     0.0|  0.0|
+-------------------+--------------------+--------+-----+
df1 = df.groupBy("user_name").agg(min("DateTime"))
df1.show()
+--------------------+-------------------+
|           user_name|      min(DateTime)|
+--------------------+-------------------+
|prathameshsalap@g...|2019-10-21 08:35:01|
|vaishusawant143@g...|2019-10-21 08:35:01|
|     you@example.com|2019-10-21 08:35:01|
+--------------------+-------------------+

Other Part -

df1 = df.withColumn("count",when(((col("keyboard")==0.0) & (col("mouse")==0.0)), count_idle+1).otherwise(0))

df2 = df1.withColumn("Idle_Sec",when((col("count")==0), 300).otherwise(1500))

df2.show()
+-------------------+--------------------+--------+-----+-----+--------+
|           DateTime|           user_name|keyboard|mouse|count|Idle_Sec|
+-------------------+--------------------+--------+-----+-----+--------+
|2019-10-21 08:35:01|prathameshsalap@g...|   333.0|658.0|    0|     300|
|2019-10-21 08:35:01|vaishusawant143@g...|   447.5|  0.0|    0|     300|
|2019-10-21 08:35:01|     you@example.com|     0.5|  1.0|    0|     300|
|2019-10-21 08:40:01|     you@example.com|     0.0|  0.0|    1|    1500|
|2019-10-21 08:40:01|prathameshsalap@g...|   227.0|366.0|    0|     300|
|2019-10-21 08:40:02|vaishusawant143@g...|   472.0|  0.0|    0|     300|
|2019-10-21 08:45:01|     you@example.com|     0.0|  0.0|    1|    1500|
|2019-10-21 08:45:01|prathameshsalap@g...|    35.0|458.0|    0|     300|
|2019-10-21 08:45:01|vaishusawant143@g...|  1659.5|  0.0|    0|     300|
|2019-10-21 08:50:01|     you@example.com|     0.0|  0.0|    1|    1500|
+-------------------+--------------------+--------+-----+-----+--------+

df3 = df2.groupBy("user_name").agg(min("DateTime").alias("start_time"),sum("Idle_Sec").alias("Sum_Idle_Sec"))

+--------------------+-------------------+------------+
|           user_name|         start_time|Sum_Idle_Sec|
+--------------------+-------------------+------------+
|prathameshsalap@g...|2019-10-21 08:35:01|         900|
|vaishusawant143@g...|2019-10-21 08:35:01|         900|
|     you@example.com|2019-10-21 08:35:01|        4800|
+--------------------+-------------------+------------+

df3.withColumn("Idle_time",(F.unix_timestamp("start_time") + col("Sum_Idle_Sec")).cast('timestamp')).show()
+--------------------+-------------------+---------+----------------------+
|           user_name|         start_time|Sum_Idle_Sec|          Idle_time|
+--------------------+-------------------+---------+----------------------+
|prathameshsalap@g...|2019-10-21 08:35:01|         900|2019-10-21 08:50:01|
|vaishusawant143@g...|2019-10-21 08:35:01|         900|2019-10-21 08:50:01|
|     you@example.com|2019-10-21 08:35:01|        4800|2019-10-21 09:55:01|
+--------------------+-------------------+---------+----------------------+

Answer 3

You should do as the following example:

• df.withColumn("user_name", do_something )

" do_something " can be any function that you define.