Loop through large dataframe in Pyspark - alternative - pyspark

df_hrrchy
|lefId |Lineage |
|-------|--------------------------------------|
|36326 |["36326","36465","36976","36091","82"]|
|36121 |["36121","36908","36976","36091","82"]|
|36380 |["36380","36465","36976","36091","82"]|
|36448 |["36448","36465","36976","36091","82"]|
|36683 |["36683","36465","36976","36091","82"]|
|36949 |["36949","36908","36976","36091","82"]|
|37349 |["37349","36908","36976","36091","82"]|
|37026 |["37026","36908","36976","36091","82"]|
|36879 |["36879","36465","36976","36091","82"]|
df_trans
|tranID | T_Id |
|-----------|-------------------------------------------------------------------------|
|1000540 |["36121","36326","37349","36949","36380","37026","36448","36683","36879"]|
df_creds
|T_Id |T_val |T_Goal |Parent_T_Id |Parent_Val |parent_Goal|
|-------|-------|-------|---------------|----------------|-----------|
|36448 |100 |1 |36465 |200 |1 |
|36465 |200 |1 |36976 |300 |2 |
|36326 |90 |1 |36465 |200 |1 |
|36091 |500 |19 |82 |600 |4 |
|36121 |90 |1 |36908 |200 |1 |
|36683 |90 |1 |36465 |200 |1 |
|36908 |200 |1 |36976 |300 |2 |
|36949 |90 |1 |36908 |200 |1 |
|36976 |300 |2 |36091 |500 |19 |
|37026 |90 |1 |36908 |200 |1 |
|37349 |100 |1 |36908 |200 |1 |
|36879 |90 |1 |36465 |200 |1 |
|36380 |90 |1 |36465 |200 |1 |
Desired Result
T_id
children
T_Val
T_Goal
parent_T_id
parent_Goal
trans_id
36091
["36976"]
500
19
82
4
1000540
36465
["36448","36326","36683","36879","36380"]
200
1
36976
2
1000540
36908
["36121","36949","37026","37349"]
200
1
36976
2
1000540
36976
["36465","36908"]
300
2
36091
19
1000540
36683
null
90
1
36465
1
1000540
37026
null
90
1
36908
1
1000540
36448
null
100
1
36465
1
1000540
36949
null
90
1
36908
1
1000540
36326
null
90
1
36465
1
1000540
36380
null
90
1
36465
1
1000540
36879
null
90
1
36465
1
1000540
36121
null
90
1
36908
1
1000540
37349
null
100
1
36908
1
1000540
Code Tried
from pyspark.sql import functions as F
from pyspark.sql import DataFrame
from pyspark.sql.functions import explode, collect_set, expr, col, collect_list,array_contains, lit
from functools import reduce
for row in df_transactions.rdd.toLocalIterator():
# def find_nodemap(row):
dfs = []
df_hy_set = (df_hrrchy.filter(df_hrrchy. lefId.isin(row["T_ds"]))
.select(explode("Lineage").alias("Terrs"))
.agg(collect_set(col("Terrs")).alias("hierarchy_list"))
.select(F.lit(row["trans_id"]).alias("trans_id "),"hierarchy_list")
)
df_childrens = (df_creds.join(df_ hy _set, expr("array_contains(hierarchy_list, T_id)"))
.select("T_id", "T_Val","T_Goal","parent_T_id", "parent_Goal", "trans _id" )
.groupBy("parent_T_id").agg(collect_list("T_id").alias("children"))
)
df_filter_creds = (df_creds.join(df_ hy _set, expr("array_contains(hierarchy_list, T_id)"))
.select ("T_id", "T_val","T_Goal","parent_T_id", "parent_Goal”, "trans_id")
)
df_nodemap = (df_filter_ creds.alias("A").join(df_childrens.alias("B"), col("A.T_id") == col("B.parent_T_id"), "left")
.select("A.T_id","B.children", "A.T_val","A.terr_Goal","A.parent_T_id", "A.parent_Goal", "A.trans_ id")
)
display(df_nodemap)
# dfs.append(df_nodemap)
# df = reduce(DataFrame.union, dfs)
# display(df)
# # display(df)
My problem - Its a bad design. df_trans is having millions of data and looping through dataframe , its taking forever. Without looping can I do it. I tried couple of other methods, not able to get the desired result.


You certainly need to process entire DataFrame in batch, not iterate row by row.
Key points are to "reverse" df_hrrchy, ie. from parent lineage obtain list of children for every T_Id:
val df_children = df_hrrchy.withColumn("children", slice($"Lineage", lit(1), size($"Lineage") - 1))
.withColumn("parents", slice($"Lineage", 2, 999999))
.select(explode(arrays_zip($"children", $"parents")).as("rels"))
.distinct
.groupBy($"rels.parents".as("T_Id"))
.agg(collect_set($"rels.children").as("children"))
df_children.show(false)
+-----+-----------------------------------+
|T_Id |children |
+-----+-----------------------------------+
|36091|[36976] |
|36465|[36448, 36380, 36326, 36879, 36683]|
|36976|[36465, 36908] |
|82 |[36091] |
|36908|[36949, 37349, 36121, 37026] |
+-----+-----------------------------------+
then expand list of T_Ids in df_trans and also include all T_Ids from the hierarchy:
val df_trans_map = df_trans.withColumn("T_Id", explode($"T_Id"))
.join(df_hrrchy, array_contains($"Lineage", $"T_Id"))
.select($"tranID", explode($"Lineage").as("T_Id"))
.distinct
df_trans_map.show(false)
+-------+-----+
|tranID |T_Id |
+-------+-----+
|1000540|36976|
|1000540|82 |
|1000540|36091|
|1000540|36465|
|1000540|36326|
|1000540|36121|
|1000540|36908|
|1000540|36380|
|1000540|36448|
|1000540|36683|
|1000540|36949|
|1000540|37349|
|1000540|37026|
|1000540|36879|
+-------+-----+
With this it is just a simple join to obtain final result:
df_trans_map.join(df_creds, Seq("T_Id"))
.join(df_children, Seq("T_Id"), "left_outer")
.show(false)
+-----+-------+-----+------+-----------+----------+-----------+-----------------------------------+
|T_Id |tranID |T_val|T_Goal|Parent_T_Id|Parent_Val|parent_Goal|children |
+-----+-------+-----+------+-----------+----------+-----------+-----------------------------------+
|36976|1000540|300 |2 |36091 |500 |19 |[36465, 36908] |
|36091|1000540|500 |19 |82 |600 |4 |[36976] |
|36465|1000540|200 |1 |36976 |300 |2 |[36448, 36380, 36326, 36879, 36683]|
|36326|1000540|90 |1 |36465 |200 |1 |null |
|36121|1000540|90 |1 |36908 |200 |1 |null |
|36908|1000540|200 |1 |36976 |300 |2 |[36949, 37349, 36121, 37026] |
|36380|1000540|90 |1 |36465 |200 |1 |null |
|36448|1000540|100 |1 |36465 |200 |1 |null |
|36683|1000540|90 |1 |36465 |200 |1 |null |
|36949|1000540|90 |1 |36908 |200 |1 |null |
|37349|1000540|100 |1 |36908 |200 |1 |null |
|37026|1000540|90 |1 |36908 |200 |1 |null |
|36879|1000540|90 |1 |36465 |200 |1 |null |
+-----+-------+-----+------+-----------+----------+-----------+-----------------------------------+


You need to re-write this to use the full cluster, using a localIterator means that you aren't fully utilizing the cluster for shared work.
Below code was not run as you didn't provide a workable data set to test. If you do I'll run the code to make sure it's sound.
from pyspark.sql import functions as F
from pyspark.sql import DataFrame
from pyspark.sql.functions import explode, collect_set, expr, col, collect_list,array_contains, lit
from functools import reduce
#uses explode I know this will create a lot of short lived records but the flip side is it will use the entire cluster to complete the work instead of the driver.
df_trans_expld = df_trans.select( df_trans.tranID, explode(df_trans.T_Id).alias("T_Id") )
#uses explode
df_hrrchy_expld = df_hrrchy.select( df_hrrchy.leftId, explode( df_hrrchy.Lineage ).alias("Lineage") )
#uses exploded data to join which is the same as a filter.
df_hy_set = df_trans_expld.join( df_hrrchy_expld, df_hrrchy_expld.lefId === df_trans_expld.T_id, "left").select( "trans_id" ).agg(collect_set(col("Lineage")).alias("hierarchy_list"))
.select(F.lit(col("trans_id")).alias("trans_id "),"hierarchy_list")
#logic unchanged from here down
df_childrens = (df_creds.join(df_hy_set, expr("array_contains(hierarchy_list, T_id)"))
.select("T_id", "T_Val","T_Goal","parent_T_id", "parent_Goal", "trans _id" )
.groupBy("parent_T_id").agg(collect_list("T_id").alias("children"))
)
df_filter_creds = (df_creds.join(ddf_hy_set, expr("array_contains(hierarchy_list, T_id)"))
.select ("T_id", "T_val","T_Goal","parent_T_id", "parent_Goal”, "trans_id")
)
df_nodemap = (df_filter_creds.alias("A").join(df_childrens.alias("B"), col("A.T_id") == col("B.parent_T_id"), "left")
.select("A.T_id","B.children", "A.T_val","A.terr_Goal","A.parent_T_id", "A.parent_Goal", "A.trans_ id")
)
# no need to append/union data as it's now just one dataframe df_nodemap
I'd have to look into this more but I'm pretty sure you are pulling all the data through the driver(with your existing code), which will really slow things down, this will make use of all executors to complete the work.
There may be another optimization to get rid of the array_contains (and use a join instead). I'd have to look at the explain to see if you could get even more performance out of it. Don't remember off the top of my head, you are avoiding a shuffle so it may be better as is.

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.

Scala Spark dataframe filter using multiple column based on available value

I need to filter a dataframe with the below criteria.
I have 2 columns 4Wheel(Subaru, Toyota, GM, null/empty) and 2Wheel(Yamaha, Harley, Indian, null/empty).
I have to filter on 4Wheel with values (Subaru, Toyota), if 4Wheel contain empty/null then filter on 2Wheel with values (Yamaha, Harley)
I couldn't find this type of filtering in different examples. I am new to spark/scala, so could not get enough idea to implement this.
Thanks,
Barun.

You can use spark SQL built-in function when to check if a column is null or empty, and filter accordingly:
import org.apache.spark.sql.functions.{col, when}
dataframe.filter(when(col("4Wheel").isNull || col("4Wheel").equalTo(""),
col("2Wheel").isin("Yamaha", "Harley")
).otherwise(
col("4Wheel").isin("Subaru", "Toyota")
))
So if you have the following input:
+---+------+------+
|id |4Wheel|2Wheel|
+---+------+------+
|1 |Toyota|null |
|2 |Subaru|null |
|3 |GM |null |
|4 |null |Yamaha|
|5 | |Yamaha|
|6 |null |Harley|
|7 | |Harley|
|8 |null |Indian|
|9 | |Indian|
|10 |null |null |
+---+------+------+
You get the following filtered ouput:
+---+------+------+
|id |4Wheel|2Wheel|
+---+------+------+
|1 |Toyota|null |
|2 |Subaru|null |
|4 |null |Yamaha|
|5 | |Yamaha|
|6 |null |Harley|
|7 | |Harley|
+---+------+------+

Delete values lower than cummax on multiple spark dataframe columns in scala

I am having a data frame as shown below. The number of signals are more than 100, so there will be more than 100 columns in the data frame.
+---+------------+--------+--------+--------+
|id | date|signal01|signal02|signal03|......
+---+------------+--------+--------+--------+
|050|2021-01-14 |1 |3 |1 |
|050|2021-01-15 |null |4 |2 |
|050|2021-02-02 |2 |3 |3 |
|051|2021-01-14 |1 |3 |0 |
|051|2021-01-15 |2 |null |null |
|051|2021-02-02 |3 |3 |2 |
|051|2021-02-03 |1 |3 |1 |
|052|2021-03-03 |1 |3 |0 |
|052|2021-03-05 |3 |3 |null |
|052|2021-03-06 |2 |null |2 |
|052|2021-03-16 |3 |5 |5 |.......
+-------------------------------------------+
I have to find out cummax of each signal and then compare with respective signal columns and delete the signal records which are having value lower than cummax and null values.
step1. find cumulative max for each signal with respect to id column.
step2. delete the records which are having lower value than cummax for each signal.
step3. Take count of records which are having cummax less than signal value(excluded of null) for each signals with respect to id.
After the count the final output should be as shown below.
+---+------------+--------+--------+--------+
|id | date|signal01|signal02|signal03|.....
+---+------------+--------+--------+--------+
|050|2021-01-14 |1 | 3 | 1 |
|050|2021-01-15 |null | null | 2 |
|050|2021-02-02 |2 | 3 | 3 |
|
|051|2021-01-14 |1 | 3 | 0 |
|051|2021-01-15 |2 | null | null |
|051|2021-02-02 |3 | 3 | 2 |
|051|2021-02-03 |null | 3 | null |
|
|052|2021-03-03 |1 | 3 | 0 |
|052|2021-03-05 |3 | 3 | null |
|052|2021-03-06 |null | null | 2 |
|052|2021-03-16 |3 | 5 | 5 | ......
+----------------+--------+--------+--------+
I have tried by using window function as below and it worked for almost all records.
val w = Window.partitionBy("id").orderBy("date").rowsBetween(Window.unboundedPreceding, Window.currentRow)
val signalList01 = ListBuffer[Column]()
signalList01.append(col("id"), col("date"))
for (column <- signalColumns) {
// Applying the max non null aggregate function on each signal column
signalList01 += (col(column), max(column).over(w).alias(column+"_cummax")) }
val cumMaxDf = df.select(signalList01: _*)
But I am getting error values as shown below for few records.
Is there any idea about how this error records in the cummax column? Any leads appreciated!

Just giving out hints here (as you suggested) to help you unblock the situation, but --WARNING-- haven't tested the code !
the code you provided in the comments looks good. It'll get you your max column
val nw_df = original_df.withColumn("singal01_cummax", sum(col("singal01")).over(windowCodedSO))
now, you need to be able to compare the two values in "singal01" and "singal01_cummax". A function like this, maybe:
def takeOutRecordsLessThanCummax (signal:Int, singal_cummax: Int) : Any =
{ if (signal == null || signal < singal_cummax) null
else singal_cummax }
since we'll be applying it to columns, we'll wrap it up in a UDF
val takeOutRecordsLessThanCummaxUDF : UserDefinedFunction = udf {
(i:Int, j:Int) => takeOutRecordsLessThanCummax(i,j)
}
and then, you can combine everything above so it can be applicable on your original dataframe. Something like this could work:
val signal_cummax_suffix = "_cummax"
val result = original_df.columns.foldLeft(original_df)(
(dfac, colname) => dfac
.withColumn(colname.concat(signal_cummax_suffix),
sum(col(colname)).over(windowCodedSO))
.withColumn(colname.concat("output"),
takeOutRecordsLessThanCummaxUDF(col(colname), col(colname.concat(signal_cummax_suffix))))
)

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 |
+---+----------+--------+------+

How to iterate grouped rows to produce multiple rows in spark structured streaming?

I have the input data set like:
id operation value
1 null 1
1 discard 0
2 null 1
2 null 2
2 max 0
3 null 1
3 null 1
3 list 0
I want to group the input and produce rows according to "operation" column.
for group 1, operation="discard", then the output is null,
for group 2, operation="max", the output is:
2 null 2
for group 3, operation="list", the output is:
3 null 1
3 null 1
So finally the output is like:
id operation value
2 null 2
3 null 1
3 null 1
Is there a solution for this?
I know there is a similar question how-to-iterate-grouped-data-in-spark
But the differences compared to that are:
I want to produce more than one row for each grouped data. Possible
and how?
I want my logic to be easily extended for more operation to be added in future. So User-defined aggregate functions (aka UDAF) is
the only possible solution?
Update 1:
Thank stack0114106, then more details according to his answer, e.g. for id=1, operation="max", I want to iterate all the item with id=2, and find the max value, rather than assign a hard-coded value, that's why I want to iterate the rows in each group. Below is a updated example:
The input:
scala> val df = Seq((0,null,1),(0,"discard",0),(1,null,1),(1,null,2),(1,"max",0),(2,null,1),(2,null,3),(2,"max",0),(3,null,1),(3,null,1),(3,"list",0)).toDF("id"
,"operation","value")
df: org.apache.spark.sql.DataFrame = [id: int, operation: string ... 1 more field]
scala> df.show(false)
+---+---------+-----+
|id |operation|value|
+---+---------+-----+
|0 |null |1 |
|0 |discard |0 |
|1 |null |1 |
|1 |null |2 |
|1 |max |0 |
|2 |null |1 |
|2 |null |3 |
|2 |max |0 |
|3 |null |1 |
|3 |null |1 |
|3 |list |0 |
+---+---------+-----+
The expected output:
+---+---------+-----+
|id |operation|value|
+---+---------+-----+
|1 |null |2 |
|2 |null |3 |
|3 |null |1 |
|3 |null |1 |
+---+---------+-----+

group everything collecting the values, then write logic for each operation :
import org.apache.spark.sql.functions._
val grouped=df.groupBy($"id").agg(max($"operation").as("op"),collect_list($"value").as("vals"))
val maxs=grouped.filter($"op"==="max").withColumn("val",explode($"vals")).groupBy($"id").agg(max("val").as("value"))
val lists=grouped.filter($"op"==="list").withColumn("value",explode($"vals")).filter($"value"!==0).select($"id",$"value")
//we don't collect the "discard"
//and we can add additional subsets for new "operations"
val result=maxs.union(lists)
//if you need the null in "operation" column add it with withColumn

You can use flatMap operation on the dataframe and generate required rows based on the conditions that you mentioned. Check this out
scala> val df = Seq((1,null,1),(1,"discard",0),(2,null,1),(2,null,2),(2,"max",0),(3,null,1),(3,null,1),(3,"list",0)).toDF("id","operation","value")
df: org.apache.spark.sql.DataFrame = [id: int, operation: string ... 1 more field]
scala> df.show(false)
+---+---------+-----+
|id |operation|value|
+---+---------+-----+
|1 |null |1 |
|1 |discard |0 |
|2 |null |1 |
|2 |null |2 |
|2 |max |0 |
|3 |null |1 |
|3 |null |1 |
|3 |list |0 |
+---+---------+-----+
scala> df.filter("operation is not null").flatMap( r=> { val x=r.getString(1); val s = x match { case "discard" => (0,0) case "max" => (1,2) case "list" => (2,1) } ; (0
until s._1).map( i => (r.getInt(0),null,s._2) ) }).show(false)
+---+----+---+
|_1 |_2 |_3 |
+---+----+---+
|2 |null|2 |
|3 |null|1 |
|3 |null|1 |
+---+----+---+
Spark assigns _1,_2 etc.. so you can map them to actual names by assigning them as below
scala> val df2 = df.filter("operation is not null").flatMap( r=> { val x=r.getString(1); val s = x match { case "discard" => (0,0) case "max" => (1,2) case "list" => (2,1) } ; (0 until s._1).map( i => (r.getInt(0),null,s._2) ) }).toDF("id","operation","value")
df2: org.apache.spark.sql.DataFrame = [id: int, operation: null ... 1 more field]
scala> df2.show(false)
+---+---------+-----+
|id |operation|value|
+---+---------+-----+
|2 |null |2 |
|3 |null |1 |
|3 |null |1 |
+---+---------+-----+
scala>
EDIT1:
Since you need the max(value) for each id, you can use window functions and get the max value in a new column, then use the same technique and get the results. Check this out
scala> val df = Seq((0,null,1),(0,"discard",0),(1,null,1),(1,null,2),(1,"max",0),(2,null,1),(2,null,3),(2,"max",0),(3,null,1),(3,null,1),(3,"list",0)).toDF("id","operation","value")
df: org.apache.spark.sql.DataFrame = [id: int, operation: string ... 1 more field]
scala> df.createOrReplaceTempView("michael")
scala> val df2 = spark.sql(""" select *, max(value) over(partition by id) mx from michael """)
df2: org.apache.spark.sql.DataFrame = [id: int, operation: string ... 2 more fields]
scala> df2.show(false)
+---+---------+-----+---+
|id |operation|value|mx |
+---+---------+-----+---+
|1 |null |1 |2 |
|1 |null |2 |2 |
|1 |max |0 |2 |
|3 |null |1 |1 |
|3 |null |1 |1 |
|3 |list |0 |1 |
|2 |null |1 |3 |
|2 |null |3 |3 |
|2 |max |0 |3 |
|0 |null |1 |1 |
|0 |discard |0 |1 |
+---+---------+-----+---+
scala> val df3 = df2.filter("operation is not null").flatMap( r=> { val x=r.getString(1); val s = x match { case "discard" => 0 case "max" => 1 case "list" => 2 } ; (0 until s).map( i => (r.getInt(0),null,r.getInt(3) )) }).toDF("id","operation","value")
df3: org.apache.spark.sql.DataFrame = [id: int, operation: null ... 1 more field]
scala> df3.show(false)
+---+---------+-----+
|id |operation|value|
+---+---------+-----+
|1 |null |2 |
|3 |null |1 |
|3 |null |1 |
|2 |null |3 |
+---+---------+-----+
scala>