Writing output with partitioning in Apache Spark - Time & Space Complexity

Time Complexity: Writing output with partitioning

O(n)

Understanding Time Complexity

When saving data in Apache Spark, partitioning splits data into parts. This affects how long writing takes.

We want to know how the time to write grows as data size and partitions change.

Scenario Under Consideration

Analyze the time complexity of the following code snippet.

df.repartition(10).write.mode("overwrite").parquet("output/path")

This code repartitions the data into 10 parts and writes it as parquet files.

Identify Repeating Operations

Identify the loops, recursion, array traversals that repeat.

• Primary operation: Writing each partition's data to disk.
• How many times: Once per partition (here, 10 times).

How Execution Grows With Input

As data size grows, each partition holds more data, so writing takes longer per partition.

Input Size (n)	Approx. Operations
10,000 rows	Writing 10 partitions, each with ~1,000 rows
100,000 rows	Writing 10 partitions, each with ~10,000 rows
1,000,000 rows	Writing 10 partitions, each with ~100,000 rows

Pattern observation: Total work grows roughly with data size, split across fixed partitions.

Final Time Complexity

Time Complexity: O(n)

This means the time to write grows linearly with the total data size, divided among partitions.

Common Mistake

[X] Wrong: "More partitions always make writing faster regardless of data size."

[OK] Correct: Too many partitions add overhead and small files, which can slow writing and later reading.

Interview Connect

Understanding how partitioning affects write time helps you explain data scaling in Spark jobs clearly and confidently.

Self-Check

"What if we change repartition(10) to repartition(100)? How would the time complexity change?"

Practice

1. What does the partitionBy method do when writing data in Apache Spark?

easy

A. It sorts the data before saving to disk.

B. It compresses the output data to save disk space.

C. It splits the output data into folders based on column values.

D. It encrypts the output data for security.

Solution

1. Step 1: Understand the purpose of partitionBy

   The partitionBy method organizes output files into folders named by unique values of specified columns.

2. Step 2: Recognize its effect on data storage

   This splitting helps in faster data reading and better organization by grouping similar data together.

3. Final Answer:

   It splits the output data into folders based on column values. -> Option C

4. Quick Check:

   partitionBy splits data folders by column values [OK]

Hint: partitionBy creates folders by column values when saving [OK]

Common Mistakes:

• Thinking partitionBy compresses data
• Confusing partitionBy with sorting
• Assuming partitionBy encrypts data

2. Which of the following is the correct syntax to write a DataFrame df partitioned by column country in Spark?

easy

A. df.write.save.partitionBy("country")("path/to/output")

B. df.partitionBy.write("country").save("path/to/output")

C. df.write.save("path/to/output").partitionBy("country")

D. df.write.partitionBy("country").save("path/to/output")

Solution

1. Step 1: Recall the method chaining order

   In Spark, partitionBy is called on the write object before save.

2. Step 2: Verify correct syntax

   The correct syntax is df.write.partitionBy("country").save("path"). Other options misuse method order or call.

3. Final Answer:

   df.write.partitionBy("country").save("path/to/output") -> Option D

4. Quick Check:

   partitionBy called after write, before save [OK]

Hint: Use df.write.partitionBy(...).save(...) syntax [OK]

Common Mistakes:

• Calling partitionBy after save
• Using partitionBy on df instead of df.write
• Incorrect method chaining order

3. Given the following code, what folder structure will be created when saving the DataFrame?

df.write.partitionBy("year", "month").mode("overwrite").parquet("/data/events")

Assuming year has values 2023 and 2024, and month has values 1 and 2.

medium

A. /data/events/year=2023/month=1/, /data/events/year=2023/month=2/, /data/events/year=2024/month=1/, /data/events/year=2024/month=2/

B. /data/events/2023/1/, /data/events/2023/2/, /data/events/2024/1/, /data/events/2024/2/

C. /data/events/year/month/2023/1/, /data/events/year/month/2023/2/, /data/events/year/month/2024/1/, /data/events/year/month/2024/2/

D. /data/events/year=month=2023_1/, /data/events/year=month=2023_2/, /data/events/year=month=2024_1/, /data/events/year=month=2024_2/

Solution

1. Step 1: Understand partition folder naming

   Spark creates folders named by each partition column and its value, nested in order of columns.

2. Step 2: Apply to given columns and values

   Folders will be nested as /year=2023/month=1/ etc., not just values or combined keys.

3. Final Answer:

   /data/events/year=2023/month=1/, /data/events/year=2023/month=2/, /data/events/year=2024/month=1/, /data/events/year=2024/month=2/ -> Option A

4. Quick Check:

   Partition folders named as column=value pairs [OK]

Hint: Partition folders use column=value format nested by columns [OK]

Common Mistakes:

• Using only values without column names in folders
• Combining columns into one folder name
• Adding extra folder levels incorrectly

4. Identify the error in this Spark code that tries to write a DataFrame partitioned by category:

df.write.save("/output/path").partitionBy("category")

medium

A. The path should be a directory, not a file.

B. partitionBy should be called before save, not after.

C. partitionBy requires a list of columns, not a single string.

D. The DataFrame must be cached before writing.

Solution

1. Step 1: Check method chaining order

   In Spark, partitionBy must be called on the write object before save.

2. Step 2: Identify the error in the code

   The code calls save first, then partitionBy, which is invalid and causes an error.

3. Final Answer:

   partitionBy should be called before save, not after. -> Option B

4. Quick Check:

   partitionBy before save is required [OK]

Hint: Call partitionBy before save to avoid errors [OK]

Common Mistakes:

• Calling partitionBy after save
• Passing a single string instead of list (both allowed but string is valid)
• Confusing caching with writing order

5. You have a large DataFrame df with columns country, year, and sales. You want to save it partitioned by country and year to speed up queries filtering by these columns. Which approach is best to write the data efficiently?

hard

A. Use df.write.partitionBy("country", "year").mode("overwrite").parquet("/data/sales") to save partitioned parquet files.

B. Save the DataFrame without partitioning and filter later during reading.

C. Use df.write.mode("overwrite").csv("/data/sales") without partitioning for simplicity.

D. Write the DataFrame multiple times, once per country and year combination.

Solution

1. Step 1: Understand partitioning benefits

   Partitioning by country and year organizes data into folders, speeding up queries filtering on these columns.

2. Step 2: Evaluate options for efficient writing

   Use df.write.partitionBy("country", "year").mode("overwrite").parquet("/data/sales") to save partitioned parquet files. This uses partitionBy with parquet format and overwrite mode, which is efficient and standard practice.

3. Step 3: Reject less efficient options

   Options B and C do not partition, causing slower queries. Writing the DataFrame multiple times, once per country and year combination, is inefficient and error-prone.

4. Final Answer:

   Use df.write.partitionBy("country", "year").mode("overwrite").parquet("/data/sales") to save partitioned parquet files. -> Option A

5. Quick Check:

   partitionBy with parquet speeds queries [OK]

Hint: Partition by query filter columns using parquet format [OK]

Common Mistakes:

• Not partitioning data before saving
• Writing multiple times instead of once
• Using CSV without partitioning for large data