getting duplicate count but retaining duplicate rows in pyspark - pyspark

I am trying to find the duplicate count of rows in a pyspark dataframe. I found a similar answer here
but it only outputs a binary flag. I would like to have the actual count for each row.
To use the orignal post's example, if I have a dataframe like so:
+--+--+--+--+
|a |b |c |d |
+--+--+--+--+
|1 |0 |1 |2 |
|0 |2 |0 |1 |
|1 |0 |1 |2 |
|0 |4 |3 |1 |
|1 |0 |1 |2 |
+--+--+--+--+
I would like to result in something like:
+--+--+--+--+--+--+--+--+
|a |b |c |d |row_count |
+--+--+--+--+--+--+--+--+
|1 |0 |1 |2 |3 |
|0 |2 |0 |1 |0 |
|1 |0 |1 |2 |3 |
|0 |4 |3 |1 |0 |
|1 |0 |1 |2 |3 |
+--+--+--+--+--+--+--+--+
Is this possible?
Thank You

Assuming df is your input dataframe:
from pyspark.sql.window import Window
from pyspark.sql import functions as F
from pyspark.sql.functions import *
w = (Window.partitionBy([F.col("a"), F.col("b"), F.col("c"), F.col("D")]))
df=df.select(F.col("a"), F.col("b"), F.col("c"), F.col("D"), F.count(F.col("a")).over(w).alias("row_count"))
If, as per your example, you want to replace every count 1 with 0 do:
from pyspark.sql.window import Window
from pyspark.sql import functions as F
from pyspark.sql.functions import *
w = (Window.partitionBy([F.col("a"), F.col("b"), F.col("c"), F.col("D")]))
df=df.select(F.col("a"), F.col("b"), F.col("c"), F.col("D"), F.count(F.col("a")).over(w).alias("row_count")).select("a", "b", "c", "d", F.when(F.col("row_count")==F.lit(1), F.lit(0)). otherwise(F.col("row_count")).alias("row_count"))

Related

spark sql max function not producing right value

I'm trying to find the max of a column grouped by spark partition id. I'm getting the wrong value when applying the max function though. Here is the code:
val partitionCol = uuid()
val localRankCol = "test"
df = df.withColumn(partitionCol, spark_partition_id)
val windowSpec = WindowSpec.partitionBy(partitionCol).orderBy(sortExprs:_*)
val rankDF = df.withColumn(localRankCol, dense_rank().over(windowSpec))
val rankRangeDF = rankDF.agg(max(localRankCol))
rankRangeDF.show(false)
sortExprs is applying an ascending sort on sales.
And the result with some dummy data is (partitionCol is 5th column):
+--------------+------+-----+---------------------------------+--------------------------------+----+
|title |region|sales|r6bea781150fa46e3a0ed761758a50dea|5683151561af407282380e6cf25f87b5|test|
+--------------+------+-----+---------------------------------+--------------------------------+----+
|Die Hard |US |100.0|1 |0 |1 |
|Rambo |US |100.0|1 |0 |1 |
|Die Hard |AU |200.0|1 |0 |2 |
|House of Cards|EU |400.0|1 |0 |3 |
|Summer Break |US |400.0|1 |0 |3 |
|Rambo |EU |100.0|1 |1 |1 |
|Summer Break |APAC |200.0|1 |1 |2 |
|Rambo |APAC |300.0|1 |1 |3 |
|House of Cards|US |500.0|1 |1 |4 |
+--------------+------+-----+---------------------------------+--------------------------------+----+
+---------+
|max(test)|
+---------+
|5 |
+---------+
"test" column has a max value of 4 but 5 is being returned.

How can I make a unique match with join with two spark dataframes and different columns?

I have two dataframes spark(scala):
First:
+-------------------+------------------+-----------------+----------+-----------------+
|id |zone |zone_father |father_id |country |
+-------------------+------------------+-----------------+----------+-----------------+
|2 |1 |123 |1 |0 |
|2 |2 |123 |1 |0 |
|3 |3 |1 |2 |0 |
|2 |4 |123 |1 |0 |
|3 |5 |2 |2 |0 |
|3 |6 |4 |2 |0 |
|3 |7 |19 |2 |0 |
+-------------------+------------------+-----------------+----------+-----------------+
Second:
+-------------------+------------------+-----------------+-----------------+
|country |id |zone |zone_value |
+-------------------+------------------+-----------------+-----------------+
|0 |2 |1 |7 |
|0 |2 |2 |7 |
|0 |2 |4 |8 |
|0 |0 |0 |2 |
+-------------------+------------------+-----------------+-----------------+
Then I need following logic:
1 -> If => first.id = second.id && first.zone = second.zone
2 -> Else if => first.father_id = second.id && first.zone_father = second.zone
3 -> If neither the first nor the second is true, follow the latter => first.country = second.zone
And the expected result would be:
+-------------------+------------------+-----------------+----------+-----------------+-----------------+
|id |zone |zone_father |father_id |country |zone_value |
+-------------------+------------------+-----------------+----------+-----------------+-----------------+
|2 |1 |123 |1 |0 |7 |
|2 |2 |123 |1 |0 |7 |
|3 |3 |1 |2 |0 |7 |
|2 |4 |123 |1 |0 |8 |
|3 |5 |2 |2 |0 |7 |
|3 |6 |4 |2 |0 |8 |
|3 |7 |19 |2 |0 |2 |
+-------------------+------------------+-----------------+----------+-----------------+-----------------+
I tried to join both dataframes, but due "or" operation, two results for each row is returned, because the last premise returns true regardless of the result of the other two.

Replace the values based on condition spark

I have dataset I want to replace the result column based on the least value of quantity by grouping id,date
id,date,quantity,result
1,2016-01-01,245,1
1,2016-01-01,345,3
1,2016-01-01,123,2
1,2016-01-02,120,5
2,2016-01-01,567,1
2,2016-01-01,568,1
2,2016-01-02,453,1
Here the output, replace the quantity which has least value in that groupby(id,date). Here ordering of rows doesn't matter, any order it can be.
id,date,quantity,result
1,2016-01-01,245,2
1,2016-01-01,345,2
1,2016-01-01,123,2
1,2016-01-02,120,5
2,2016-01-01,567,1
2,2016-01-01,568,1
2,2016-01-02,453,1
Use the Window and get the maximum by max.
import pyspark.sql.functions as f
from pyspark.sql import Window
w = Window.partitionBy('id', 'date')
df.withColumn('result', f.when(f.col('quantity') == f.min('quantity').over(w), f.col('result'))) \
.withColumn('result', f.max('result').over(w)).show(10, False)
+---+----------+--------+------+
|id |date |quantity|result|
+---+----------+--------+------+
|1 |2016-01-02|120 |5 |
|1 |2016-01-01|245 |2 |
|1 |2016-01-01|345 |2 |
|1 |2016-01-01|123 |2 |
|2 |2016-01-02|453 |1 |
|2 |2016-01-01|567 |1 |
|2 |2016-01-01|568 |1 |
+---+----------+--------+------+

PySpark: Creating a column with number of timesteps to an event

I have a dataframe that looks as follows:
|id |val1|val2|
+---+----+----+
|1 |1 |0 |
|1 |2 |0 |
|1 |3 |0 |
|1 |4 |0 |
|1 |5 |5 |
|1 |6 |0 |
|1 |7 |0 |
|1 |8 |0 |
|1 |9 |9 |
|1 |10 |0 |
|1 |11 |0 |
|2 |1 |0 |
|2 |2 |0 |
|2 |3 |0 |
|2 |4 |0 |
|2 |5 |0 |
|2 |6 |6 |
|2 |7 |0 |
|2 |8 |8 |
|2 |9 |0 |
+---+----+----+
only showing top 20 rows
I want to create a new column with the number of rows until a non-zero value appears in val2, this should be done groupby/partitionby 'id'... if the event never happens, I need to put a -1 in the steps field.
|id |val1|val2|steps|
+---+----+----+----+
|1 |1 |0 |4 |
|1 |2 |0 |3 |
|1 |3 |0 |2 |
|1 |4 |0 |1 |
|1 |5 |5 |0 | event
|1 |6 |0 |3 |
|1 |7 |0 |2 |
|1 |8 |0 |1 |
|1 |9 |9 |0 | event
|1 |10 |0 |-1 | no further events for this id
|1 |11 |0 |-1 | no further events for this id
|2 |1 |0 |5 |
|2 |2 |0 |4 |
|2 |3 |0 |3 |
|2 |4 |0 |2 |
|2 |5 |0 |1 |
|2 |6 |6 |0 | event
|2 |7 |0 |1 |
|2 |8 |8 |0 | event
|2 |9 |0 |-1 | no further events for this id
+---+----+----+----+
only showing top 20 rows
Your requirement seems easy but implementing in spark and preserving immutability is a difficult task. I am suggesting you would need a recursive function to generate the steps column. Below I have tried to suggest you a recursive way using a udf function.
import org.apache.spark.sql.functions._
//udf function to populate step column
def stepsUdf = udf((values: Seq[Row]) => {
//sorting the collected struct in reverse order according to val1 column in reverse order
val val12 = values.sortWith(_.getAs[Int]("val1") > _.getAs[Int]("val1"))
//selecting the first of sorted list
val val12Head = val12.head
//generating the first step column in the collected list
val prevStep = if(val12Head.getAs("val2") != 0) 0 else -1
//generating the first output struct
val listSteps = List(steps(val12Head.getAs("val1"), val12Head.getAs("val2"), prevStep))
//recursive function for generating the step column
def recursiveSteps(vals : List[Row], previousStep: Int, listStep : List[steps]): List[steps] = vals match {
case x :: y =>
//event changed so step column should be 0
if(x.getAs("val2") != 0) {
recursiveSteps(y, 0, listStep :+ steps(x.getAs("val1"), x.getAs("val2"), 0))
}
//event doesn't change after the last event change
else if(x.getAs("val2") == 0 && previousStep == -1) {
recursiveSteps(y, previousStep, listStep :+ steps(x.getAs("val1"), x.getAs("val2"), previousStep))
}
//val2 is 0 after the event change so increment the step column
else {
recursiveSteps(y, previousStep+1, listStep :+ steps(x.getAs("val1"), x.getAs("val2"), previousStep+1))
}
case Nil => listStep
}
//calling the recursive function
recursiveSteps(val12.tail.toList, prevStep, listSteps)
})
df
.groupBy("id") // grouping by id column
.agg(stepsUdf(collect_list(struct("val1", "val2"))).as("stepped")) //calling udf function after the collection of struct of val1 and val2
.withColumn("stepped", explode(col("stepped"))) // generating rows from the list returned from udf function
.select(col("id"), col("stepped.*")) // final desired output
.sort("id", "val1") //optional step just for viewing
.show(false)
where steps is a case class
case class steps(val1: Int, val2: Int, steps: Int)
which should give you
+---+----+----+-----+
|id |val1|val2|steps|
+---+----+----+-----+
|1 |1 |0 |4 |
|1 |2 |0 |3 |
|1 |3 |0 |2 |
|1 |4 |0 |1 |
|1 |5 |5 |0 |
|1 |6 |0 |3 |
|1 |7 |0 |2 |
|1 |8 |0 |1 |
|1 |9 |9 |0 |
|1 |10 |0 |-1 |
|1 |11 |0 |-1 |
|2 |1 |0 |5 |
|2 |2 |0 |4 |
|2 |3 |0 |3 |
|2 |4 |0 |2 |
|2 |5 |0 |1 |
|2 |6 |6 |0 |
|2 |7 |0 |1 |
|2 |8 |8 |0 |
|2 |9 |0 |-1 |
+---+----+----+-----+
I hope the answer is helpful

How to use collect_set and collect_list functions in windowed aggregation in Spark 1.6?

In Spark 1.6.0 / Scala, is there an opportunity to get collect_list("colC") or collect_set("colC").over(Window.partitionBy("colA").orderBy("colB")?
Given that you have dataframe as
+----+----+----+
|colA|colB|colC|
+----+----+----+
|1 |1 |23 |
|1 |2 |63 |
|1 |3 |31 |
|2 |1 |32 |
|2 |2 |56 |
+----+----+----+
You can Window functions by doing the following
import org.apache.spark.sql.functions._
import org.apache.spark.sql.expressions._
df.withColumn("colD", collect_list("colC").over(Window.partitionBy("colA").orderBy("colB"))).show(false)
Result:
+----+----+----+------------+
|colA|colB|colC|colD |
+----+----+----+------------+
|1 |1 |23 |[23] |
|1 |2 |63 |[23, 63] |
|1 |3 |31 |[23, 63, 31]|
|2 |1 |32 |[32] |
|2 |2 |56 |[32, 56] |
+----+----+----+------------+
Similar is the result for collect_set as well. But the order of elements in the final set will not be in order as with collect_list
df.withColumn("colD", collect_set("colC").over(Window.partitionBy("colA").orderBy("colB"))).show(false)
+----+----+----+------------+
|colA|colB|colC|colD |
+----+----+----+------------+
|1 |1 |23 |[23] |
|1 |2 |63 |[63, 23] |
|1 |3 |31 |[63, 31, 23]|
|2 |1 |32 |[32] |
|2 |2 |56 |[56, 32] |
+----+----+----+------------+
If you remove orderBy as below
df.withColumn("colD", collect_list("colC").over(Window.partitionBy("colA"))).show(false)
result would be
+----+----+----+------------+
|colA|colB|colC|colD |
+----+----+----+------------+
|1 |1 |23 |[23, 63, 31]|
|1 |2 |63 |[23, 63, 31]|
|1 |3 |31 |[23, 63, 31]|
|2 |1 |32 |[32, 56] |
|2 |2 |56 |[32, 56] |
+----+----+----+------------+
I hope the answer is helpful
Existing answer is valid, just adding here a different style of writting window functions:
import org.apache.spark.sql.expressions.Window
val wind_user = Window.partitionBy("colA", "colA2").orderBy("colB", "colB2".desc)
df.withColumn("colD_distinct", collect_set($"colC") over wind_user)
.withColumn("colD_historical", collect_list($"colC") over wind_user).show(false)