Broadcast variables in Apache Spark - Step-by-Step Execution

Concept Flow - Broadcast variables

Create large read-only variable

↓

Broadcast variable to all worker nodes

↓

Workers receive broadcast variable

↓

Workers use broadcast variable in tasks

↓

Tasks execute efficiently without sending full data each time

↓

Job completes

Broadcast variables are created once and sent to all worker nodes to efficiently share large read-only data during distributed tasks.

Execution Sample

Apache Spark

from pyspark.sql import SparkSession
spark = SparkSession.builder.getOrCreate()
data = [1, 2, 3, 4]
bc_var = spark.sparkContext.broadcast(data)
result = spark.sparkContext.parallelize([10, 20]).map(lambda x: x + sum(bc_var.value)).collect()

This code broadcasts a list to all workers and uses it in a map operation to add the sum of the list to each element.

Execution Table

Step	Action	Broadcast Variable State	Task Input	Task Output
1	Create list data	[1, 2, 3, 4]	N/A	N/A
2	Broadcast data to workers	[1, 2, 3, 4] broadcasted	N/A	N/A
3	Parallelize input [10, 20]	[1, 2, 3, 4] broadcasted	[10, 20]	N/A
4	Map: For 10, add sum(bc_var.value)=10	[1, 2, 3, 4] broadcasted	10	20
5	Map: For 20, add sum(bc_var.value)=10	[1, 2, 3, 4] broadcasted	20	30
6	Collect results	[1, 2, 3, 4] broadcasted	[10, 20]	[20, 30]

💡 All tasks completed using broadcast variable without re-sending data.

Variable Tracker

Variable	Start	After Step 2	After Step 6
data	N/A	[1, 2, 3, 4]	[1, 2, 3, 4]
bc_var.value	N/A	[1, 2, 3, 4]	[1, 2, 3, 4]
result	N/A	N/A	[20, 30]

Key Moments - 2 Insights

Why do we broadcast the variable instead of sending it with each task?

Broadcasting sends the variable once to all workers, so tasks reuse it without repeated data transfer, as shown in steps 2 and 4-5 of the execution table.

Can the broadcast variable be changed after broadcasting?

No, broadcast variables are read-only after creation. The execution table shows the broadcast variable state stays the same from step 2 to 6.

Visual Quiz - 3 Questions

Look at the execution table, what is the output of the map operation for input 10 at step 4?

A 30

B 10

C 20

D 40

💡 Hint

Check the 'Task Output' column at step 4 in the execution table.

At which step is the broadcast variable first sent to the workers?

A Step 2

B Step 1

C Step 3

D Step 4

💡 Hint

Look at the 'Broadcast Variable State' column to see when it changes to 'broadcasted'.

If the broadcast variable was not used, how would the tasks behave?

A Tasks would run faster because no broadcast is needed.

B Each task would receive the full data separately, increasing network traffic.

C Tasks would fail because data is missing.

D Tasks would ignore the data and produce the same output.

💡 Hint

Recall the purpose of broadcast variables shown in the concept flow and key moments.

Concept Snapshot

Broadcast variables in Spark:
- Created once and sent to all workers
- Used to share large read-only data efficiently
- Avoid repeated data transfer in tasks
- Accessed via bc_var.value
- Immutable after broadcasting

Full Transcript

Broadcast variables are a way to share large read-only data efficiently in Apache Spark. First, you create the variable on the driver. Then Spark sends it once to all worker nodes. Workers use this broadcasted data in their tasks without needing to receive it repeatedly. This saves network traffic and speeds up jobs. The broadcast variable cannot be changed after sending. In the example, a list is broadcasted and each task adds the sum of the list to its input. The execution table shows each step from creating the data, broadcasting, running tasks, to collecting results. This helps beginners see how broadcast variables improve distributed computing.

Practice

1. What is the main purpose of broadcast variables in Apache Spark?

easy

A. To efficiently share small read-only data across all worker nodes

B. To store large datasets in memory on the driver node

C. To cache RDDs for faster repeated access

D. To shuffle data between partitions during a join

Solution

1. Step 1: Understand broadcast variable role

   Broadcast variables send a copy of small read-only data to all worker nodes once.

2. Step 2: Compare with other options

   Caching RDDs and shuffling data are different Spark features, not broadcast variables.

3. Final Answer:

   To efficiently share small read-only data across all worker nodes -> Option A

4. Quick Check:

   Broadcast variables = share small data efficiently [OK]

Hint: Broadcast = share small read-only data once [OK]

Common Mistakes:

• Confusing broadcast variables with caching
• Thinking broadcast variables handle large datasets
• Mixing broadcast variables with shuffle operations

2. Which of the following is the correct way to create a broadcast variable in PySpark?

easy

A. broadcastVar = sc.broadcast(data)

B. broadcastVar = SparkContext.broadcast(data)

C. broadcastVar = sc.broadcastVariable(data)

D. broadcastVar = sc.share(data)

Solution

1. Step 1: Recall PySpark broadcast syntax

   The correct method is calling broadcast() on SparkContext instance sc.

2. Step 2: Check other options

   Other options use incorrect method names or class calls.

3. Final Answer:

   broadcastVar = sc.broadcast(data) -> Option A

4. Quick Check:

   Use sc.broadcast() to create broadcast variable [OK]

Hint: Use sc.broadcast(data) to create broadcast variable [OK]

Common Mistakes:

• Using wrong method names like broadcastVariable
• Calling broadcast on SparkContext class instead of instance
• Trying to use sc.share() which doesn't exist

3. Given the following PySpark code, what will be the output?

data = {'a': 1, 'b': 2}
broadcastVar = sc.broadcast(data)
rdd = sc.parallelize(['a', 'b', 'c'])
result = rdd.map(lambda x: broadcastVar.value.get(x, 0)).collect()
print(result)

medium

A. [1, 2, None]

B. [0, 0, 0]

C. [1, 2]

D. [1, 2, 0]

Solution

1. Step 1: Understand broadcast variable usage in map

   The broadcast variable holds a dictionary. The map looks up each key in the dictionary, returning 0 if key not found.

2. Step 2: Evaluate each element in RDD

   'a' maps to 1, 'b' maps to 2, 'c' is missing so returns 0.

3. Final Answer:

   [1, 2, 0] -> Option D

4. Quick Check:

   Missing keys return 0, so output is [1, 2, 0] [OK]

Hint: Missing keys return default 0 in get() [OK]

Common Mistakes:

• Expecting None instead of 0 for missing keys
• Forgetting to use broadcastVar.value to access data
• Assuming output length matches dictionary size

4. Identify the error in this PySpark code snippet using broadcast variables:

data = [1, 2, 3]
broadcastVar = sc.broadcast(data)
rdd = sc.parallelize([1, 2, 3])
result = rdd.map(lambda x: broadcastVar.get(x)).collect()
print(result)

medium

A. The lambda function should not use broadcastVar

B. The broadcast variable should be created with sc.share()

C. broadcastVar.get() is not a valid method to access broadcast data

D. RDD should be created with sc.broadcast() instead of sc.parallelize()

Solution

1. Step 1: Check broadcast variable access method

   Broadcast variables are accessed via broadcastVar.value, not broadcastVar.get().

2. Step 2: Verify other parts of code

   sc.share() does not exist, and sc.parallelize() is correct for RDD creation.

3. Final Answer:

   broadcastVar.get() is not a valid method to access broadcast data -> Option C

4. Quick Check:

   Use broadcastVar.value to access data [OK]

Hint: Access broadcast data with .value, not .get() [OK]

Common Mistakes:

• Using broadcastVar.get() instead of broadcastVar.value
• Confusing RDD creation methods
• Trying to use non-existent sc.share() method

5. You have a large dataset of user logs and a small dictionary mapping user IDs to user names. You want to replace user IDs in the logs with names efficiently using broadcast variables. Which approach is best?

hard

A. Join the logs RDD with the user dictionary RDD using a shuffle join

B. Broadcast the user ID to name dictionary and map logs using broadcastVar.value.get()

C. Collect the logs RDD to driver and replace IDs locally

D. Use sc.broadcast() on the logs RDD to share it across workers

Solution

1. Step 1: Identify efficient use of broadcast variables

   Broadcast small user dictionary to all workers to avoid shuffling large logs data.

2. Step 2: Apply mapping using broadcast variable

   Use broadcastVar.value.get() in map to replace user IDs with names efficiently.

3. Step 3: Evaluate other options

   Shuffle join is expensive, collecting large logs to driver is inefficient, broadcasting large logs is incorrect.

4. Final Answer:

   Broadcast the user ID to name dictionary and map logs using broadcastVar.value.get() -> Option B

5. Quick Check:

   Broadcast small data, map large data [OK]

Hint: Broadcast small lookup data, map large dataset [OK]

Common Mistakes:

• Trying to broadcast large datasets
• Using shuffle join instead of broadcast
• Collecting large RDDs to driver causing memory issues