Window functions in Apache Spark

Introduction

Window functions let you do calculations across rows related to the current row without grouping data. This helps keep all rows while adding useful info.

Calculate running totals or moving averages in sales data.

Rank items within categories, like top products per region.

Compare each row to previous or next rows, like daily temperature changes.

Add row numbers or dense ranks without losing detail rows.

Syntax

Apache Spark

from pyspark.sql.window import Window
from pyspark.sql.functions import row_number, rank, sum

windowSpec = Window.partitionBy('category').orderBy('value')

# Example: Add row number within each category
result = df.withColumn('row_num', row_number().over(windowSpec))

partitionBy splits data into groups (like categories).

orderBy defines the order inside each group for calculations.

Examples

Adds a row number to each employee within their department, ordered by salary.

Apache Spark

from pyspark.sql.window import Window
from pyspark.sql.functions import row_number

windowSpec = Window.partitionBy('department').orderBy('salary')
df.withColumn('row_num', row_number().over(windowSpec))

Calculates running total of sales per store ordered by date.

Apache Spark

from pyspark.sql.window import Window
from pyspark.sql.functions import sum

windowSpec = Window.partitionBy('store').orderBy('date').rowsBetween(Window.unboundedPreceding, 0)
df.withColumn('running_total', sum('sales').over(windowSpec))

Ranks items by profit within each region.

Apache Spark

from pyspark.sql.window import Window
from pyspark.sql.functions import rank

windowSpec = Window.partitionBy('region').orderBy('profit')
df.withColumn('rank', rank().over(windowSpec))

Sample Program

This program creates sales data for two categories over dates. It adds a running total of sales per category ordered by date and a row number for each row within the category.

Apache Spark

from pyspark.sql import SparkSession
from pyspark.sql.window import Window
from pyspark.sql.functions import row_number, sum

spark = SparkSession.builder.appName('WindowFunctionsExample').getOrCreate()

data = [
    ('A', '2024-01-01', 100),
    ('A', '2024-01-02', 150),
    ('A', '2024-01-03', 200),
    ('B', '2024-01-01', 50),
    ('B', '2024-01-02', 80)
]

columns = ['category', 'date', 'sales']
df = spark.createDataFrame(data, columns)

windowSpec = Window.partitionBy('category').orderBy('date').rowsBetween(Window.unboundedPreceding, 0)

result = df.withColumn('running_total', sum('sales').over(windowSpec))
result = result.withColumn('row_num', row_number().over(Window.partitionBy('category').orderBy('date')))

result.show()

Important Notes

Window functions do not reduce the number of rows like groupBy does.

Use partitionBy to define groups, and orderBy to define order inside groups.

Common window functions include row_number(), rank(), sum(), avg(), and lag().

Summary

Window functions let you add calculations across related rows without grouping.

They keep all rows and add new columns with useful info like running totals or ranks.

Use partitionBy and orderBy to control how data is grouped and ordered for these calculations.

Practice

1. What is the main purpose of window functions in Apache Spark?

easy

A. To perform calculations across rows related to the current row without collapsing the data

B. To group data and reduce the number of rows

C. To filter rows based on a condition

D. To join two datasets together

Solution

1. Step 1: Understand window function behavior

   Window functions calculate values across related rows but keep all original rows intact.

2. Step 2: Compare with grouping

   Grouping reduces rows by aggregation, but window functions add columns without reducing rows.

3. Final Answer:

   To perform calculations across rows related to the current row without collapsing the data -> Option A

4. Quick Check:

   Window functions add calculations without grouping ≠ A [OK]

Hint: Window functions add columns, grouping reduces rows [OK]

Common Mistakes:

• Confusing window functions with groupBy aggregations
• Thinking window functions filter rows
• Assuming window functions join datasets

2. Which of the following is the correct way to define a window specification in PySpark to partition by 'department' and order by 'salary'?

easy

A. Window.groupBy('department').sortBy('salary')

B. Window.sortBy('salary').partitionBy('department')

C. Window.partitionBy('department').orderBy('salary')

D. Window.partition('department').order('salary')

Solution

1. Step 1: Recall PySpark window spec syntax

   The correct methods are partitionBy() and orderBy() chained on Window.

2. Step 2: Check method order and names

   partitionBy comes first, then orderBy; groupBy and sortBy are incorrect here.

3. Final Answer:

   Window.partitionBy('department').orderBy('salary') -> Option C

4. Quick Check:

   Use partitionBy then orderBy methods [OK]

Hint: Use Window.partitionBy().orderBy() exactly [OK]

Common Mistakes:

• Using groupBy instead of partitionBy
• Using sortBy instead of orderBy
• Using incorrect method names like partition or order

3. Given the following PySpark code, what will be the output of the 'rank' column?

from pyspark.sql import SparkSession
from pyspark.sql.window import Window
from pyspark.sql.functions import rank

spark = SparkSession.builder.getOrCreate()
data = [("Alice", "Sales", 5000), ("Bob", "Sales", 4000), ("Charlie", "Sales", 5000)]
df = spark.createDataFrame(data, ["name", "department", "salary"])
windowSpec = Window.partitionBy("department").orderBy("salary")
df = df.withColumn("rank", rank().over(windowSpec))
df.orderBy("salary").show()

medium

A. [ ("Bob", 4000, 1), ("Alice", 5000, 2), ("Charlie", 5000, 2) ]

B. [ ("Bob", 4000, 1), ("Alice", 5000, 3), ("Charlie", 5000, 3) ]

C. [ ("Bob", 4000, 1), ("Alice", 5000, 2), ("Charlie", 5000, 3) ]

D. [ ("Bob", 4000, 1), ("Alice", 5000, 1), ("Charlie", 5000, 1) ]

Solution

1. Step 1: Understand rank() behavior

   Rank assigns the same rank to ties and skips ranks after ties.

2. Step 2: Apply rank over salary ascending

   Bob has salary 4000 ranked 1; Alice and Charlie both have 5000, so both get rank 2.

3. Final Answer:

   [ ("Bob", 4000, 1), ("Alice", 5000, 2), ("Charlie", 5000, 2) ] -> Option A

4. Quick Check:

   rank() ties share rank, next rank skips [OK]

Hint: rank() gives same rank to ties, skips next ranks [OK]

Common Mistakes:

• Confusing rank() with dense_rank()
• Assigning different ranks to tied values
• Not ordering data before applying rank

4. Identify the error in this PySpark code snippet using window functions:

from pyspark.sql.window import Window
from pyspark.sql.functions import row_number

windowSpec = Window.partitionBy('department')
df = df.withColumn('row_num', row_number().over(windowSpec.orderBy('salary')))

medium

A. Missing import for Window functions causes error

B. windowSpec.orderBy('salary') is valid; no error in code

C. row_number() cannot be used with partitionBy

D. windowSpec.orderBy('salary') is invalid; orderBy should be called on Window, not on windowSpec

Solution

1. Step 1: Check windowSpec usage

   windowSpec is a WindowSpec object; calling orderBy on it is valid and returns a new WindowSpec.

2. Step 2: Confirm chaining

   Chaining orderBy on windowSpec is allowed to refine the window specification.

3. Final Answer:

   windowSpec.orderBy('salary') is valid; no error in code -> Option B

4. Quick Check:

   WindowSpec supports chaining orderBy to refine window [OK]

Hint: WindowSpec supports chaining orderBy to refine window [OK]

Common Mistakes:

• Calling orderBy on windowSpec instead of Window
• Assuming row_number can't be used with partitionBy
• Forgetting to import Window or functions

5. You have a sales dataset with columns: 'store', 'date', and 'sales'. You want to add a column showing the 3-day moving average of sales per store ordered by date. Which PySpark window function setup and code will achieve this?

hard

A. Use Window.partitionBy('store').orderBy('date').rowsBetween(0, 2) and avg('sales').over(windowSpec)

B. Use Window.partitionBy('store').orderBy('date').rangeBetween(-2, 0) and avg('sales').over(windowSpec)

C. Use Window.orderBy('date').rowsBetween(-3, 0) and avg('sales').over(windowSpec)

D. Use Window.partitionBy('store').orderBy('date').rowsBetween(-2, 0) and avg('sales').over(windowSpec)

Solution

1. Step 1: Define window for 3-day moving average per store

   Partition by 'store' to separate stores, order by 'date' for time sequence, use rowsBetween(-2, 0) for current and previous 2 rows.

2. Step 2: Apply avg() over this window

   Use avg('sales').over(windowSpec) to compute moving average over the defined window.

3. Final Answer:

   Use Window.partitionBy('store').orderBy('date').rowsBetween(-2, 0) and avg('sales').over(windowSpec) -> Option D

4. Quick Check:

   rowsBetween(-2,0) for 3-day window, partition by store [OK]

Hint: Use rowsBetween(-2,0) for 3-row moving average [OK]

Common Mistakes:

• Using rangeBetween instead of rowsBetween for row count
• Not partitioning by store
• Using wrong rowsBetween range (like 0 to 2)