Caching and persistence in Apache Spark - Time & Space Complexity

Time Complexity: Caching and persistence

O(n)

Understanding Time Complexity

When working with Apache Spark, caching and persistence help speed up repeated data use.

We want to see how caching changes the time it takes to run operations again.

Scenario Under Consideration

Analyze the time complexity of the following code snippet.

val data = spark.read.csv("data.csv")
val transformed = data.filter(_(2) === "USA").select("_c0", "_c1")
transformed.cache()
transformed.count()
transformed.collect()

This code loads data, filters rows, caches the result, then runs two actions that use the cached data.

Identify Repeating Operations

Identify the loops, recursion, array traversals that repeat.

• Primary operation: Filtering and selecting rows from the dataset.
• How many times: Without caching, these run every time an action is called; with caching, they run once.

How Execution Grows With Input

Without caching, each action repeats the full data scan, so time grows with data size every time.

Input Size (n)	Approx. Operations Without Cache	Approx. Operations With Cache
10	20 (2 actions x 10)	10 + 10 (cache once + reuse)
100	200 (2 x 100)	100 + 100
1000	2000 (2 x 1000)	1000 + 1000

Pattern observation: Without cache, work doubles with each action; with cache, work is done once and reused.

Final Time Complexity

Time Complexity: O(n)

This means the time grows linearly with data size, but caching avoids repeating the same work multiple times.

Common Mistake

[X] Wrong: "Caching makes the operation run instantly regardless of data size."

[OK] Correct: Caching only saves time on repeated actions by storing results; the first run still processes all data.

Interview Connect

Understanding caching shows you can make data work faster by avoiding repeated heavy tasks, a key skill in data projects.

Self-Check

"What if we used persistence with a different storage level instead of cache? How would the time complexity change?"

Practice

1. What is the main purpose of using cache() in Apache Spark?

easy

A. To delete data from memory

B. To permanently save data to disk

C. To store data in memory for faster repeated access

D. To convert data into a different format

Solution

1. Step 1: Understand what caching means in Spark

   Caching stores data in memory to speed up repeated operations on the same data.

2. Step 2: Identify the role of cache()

   The cache() method specifically stores the DataFrame or RDD in memory for quick access.

3. Final Answer:

   To store data in memory for faster repeated access -> Option C

4. Quick Check:

   Caching = store in memory [OK]

Hint: Cache means keep data in memory for speed [OK]

Common Mistakes:

• Thinking cache saves data permanently on disk
• Confusing cache with deleting data
• Assuming cache changes data format

2. Which of the following is the correct syntax to persist a DataFrame df with disk storage level in Spark?

easy

A. df.persist(StorageLevel.DISK_ONLY)

B. df.cache(StorageLevel.DISK_ONLY)

C. df.persist('disk_only')

D. df.cache('disk_only')

Solution

1. Step 1: Recall the correct method to specify storage level

   To persist with a specific storage level, use persist() with a StorageLevel object.

2. Step 2: Check the syntax for disk-only persistence

   The correct syntax is df.persist(StorageLevel.DISK_ONLY). The cache() method does not accept parameters.

3. Final Answer:

   df.persist(StorageLevel.DISK_ONLY) -> Option A

4. Quick Check:

   persist + StorageLevel object = correct syntax [OK]

Hint: Use persist() with StorageLevel, cache() has no params [OK]

Common Mistakes:

• Using cache() with parameters
• Passing storage level as string instead of StorageLevel object
• Using persist() without StorageLevel

3. Consider the following Spark code snippet:

val df = spark.read.csv("data.csv")
df.cache()
df.count()
df.count()

What is the effect of calling df.count() twice after caching?

medium

A. Both counts reload data from disk without caching

B. The first count triggers data loading and caching; the second count uses cached data for faster result

C. The second count triggers caching; the first count uses cached data

D. Both counts use cached data without loading from disk

Solution

1. Step 1: Understand caching behavior on first action

   The first action (count()) triggers loading data and caches it in memory.

2. Step 2: Effect on second action

   The second count() uses the cached data, so it runs faster without reloading from disk.

3. Final Answer:

   The first count triggers data loading and caching; the second count uses cached data for faster result -> Option B

4. Quick Check:

   First action caches, second uses cache [OK]

Hint: First action caches, second action reads cache [OK]

Common Mistakes:

• Assuming cache happens before first action
• Thinking both counts reload from disk
• Believing second count triggers caching

4. Given this Spark code:

val df = spark.read.json("data.json")
df.persist(StorageLevel.MEMORY_AND_DISK)
df.unpersist()

What is the main issue with this code?

medium

A. Calling unpersist() immediately removes cached data, so persistence is ineffective

B. The persist() method requires a string, not StorageLevel object

C. You must call cache() before persist()

D. The code will cause a syntax error due to missing parentheses

Solution

1. Step 1: Understand the effect of persist()

   The persist() call stores the DataFrame in memory and disk as specified.

2. Step 2: Effect of immediate unpersist()

   Calling unpersist() right after removes the cached data, negating the persistence benefit.

3. Final Answer:

   Calling unpersist() immediately removes cached data, so persistence is ineffective -> Option A

4. Quick Check:

   Immediate unpersist clears cache [OK]

Hint: Unpersist right after persist clears cache [OK]

Common Mistakes:

• Thinking persist needs string argument
• Believing cache must be called before persist
• Assuming syntax error due to parentheses

5. You have a large DataFrame df used multiple times in your Spark job. You want to cache it but ensure it uses memory first and spills to disk if memory is full. Which persistence level should you use and why?

hard

A. Use df.persist(StorageLevel.DISK_ONLY) to avoid memory usage

B. Use df.cache() because it automatically spills to disk

C. Use df.unpersist() to keep data only in memory

D. Use df.persist(StorageLevel.MEMORY_AND_DISK) to store in memory and spill to disk if needed

Solution

1. Step 1: Identify persistence levels for memory and disk

   StorageLevel.MEMORY_AND_DISK stores data in memory and spills to disk if memory is insufficient.

2. Step 2: Compare with other options

   cache() only stores in memory and may fail if memory is full; DISK_ONLY avoids memory; unpersist() removes cached data.

3. Final Answer:

   Use df.persist(StorageLevel.MEMORY_AND_DISK) to store in memory and spill to disk if needed -> Option D

4. Quick Check:

   MEMORY_AND_DISK = memory first, then disk spill [OK]

Hint: MEMORY_AND_DISK spills to disk if memory full [OK]

Common Mistakes:

• Assuming cache spills to disk automatically
• Choosing DISK_ONLY to save memory but slower
• Using unpersist to keep data cached