• Common Spark Concepts which may be asked in an interview
  • Types of join strateries
  • Types of Joins
  • Common Spark optimization techniques
  • Different phases of Spark-SQL engine
  • Common reasons for analysis exception in Spark
  • Flow of how Spark works internally
  • Explain DAG in Spark
  • Explain spark.sql.shuffle.partitions Variable

Common Spark Concepts which may be asked in an interview

In this article I will try to put some common Spark concepts in a tabular format(So that its compact). These are good concepts to remember. Also, they may be asked in Interview questions.

Types of join strateries

Join Strategy	Description	Use Case	Example
Shuffle Hash Join	Both DataFrames are shuffled based on the join keys, and then a hash join is performed.	Useful when both DataFrames are large and have a good distribution of keys.	spark.conf.set("spark.sql.join.preferSortMergeJoin", "false") df1.join(df2, "key")
Broadcast Hash Join	One of the DataFrames is small enough to fit in memory and is broadcasted to all worker nodes. A hash join is then performed.	Efficient when one DataFrame is much smaller than the other.	val broadcastDF = broadcast(df2) df1.join(broadcastDF, "key")
Sort-Merge Join	Both DataFrames are sorted on the join keys and then merged. This requires a shuffle if the data is not already sorted.	Suitable for large DataFrames when the join keys are sorted or can be sorted efficiently.	spark.conf.set("spark.sql.join.preferSortMergeJoin", "true") df1.join(df2, "key")
Cartesian Join (Cross Join)	Every row of one DataFrame is paired with every row of the other DataFrame.	Generally not recommended due to its computational expense, but can be used for generating combinations of all rows.	df1.crossJoin(df2)
Broadcast Nested Loop Join	The smaller DataFrame is broadcasted, and a nested loop join is performed.	Used when there are no join keys or the join condition is complex and cannot be optimized with hash or sort-merge joins.	val broadcastDF = broadcast(df2) df1.join(broadcastDF)
Shuffle-and-Replicate Nested Loop Join	Both DataFrames are shuffled and replicated to perform a nested loop join.	Used for complex join conditions that cannot be handled by other join strategies.	df1.join(df2, expr("complex_condition"))

Types of Joins

Join Type	Description	Example
Inner Join	Returns rows that have matching values in both DataFrames.	df1.join(df2, "key")
Outer Join	Returns all rows when there is a match in either DataFrame. Missing values are filled with nulls.	df1.join(df2, Seq("key"), "outer")
Left Outer Join	Returns all rows from the left DataFrame, and matched rows from the right DataFrame.	df1.join(df2, Seq("key"), "left_outer")
Right Outer Join	Returns all rows from the right DataFrame, and matched rows from the left DataFrame.	df1.join(df2, Seq("key"), "right_outer")
Left Semi Join	Returns only the rows from the left DataFrame that have a match in the right DataFrame.	df1.join(df2, Seq("key"), "left_semi")
Left Anti Join	Returns only the rows from the left DataFrame that do not have a match in the right DataFrame.	df1.join(df2, Seq("key"), "left_anti")
Cross Join	Returns the Cartesian product of both DataFrames. Every row in the left DataFrame will be combined with every row in the right DataFrame.	df1.crossJoin(df2)
Self Join	A join in which a DataFrame is joined with itself. This can be an inner, outer, left, or right join.	df.join(df, df("key1") === df("key2"))

Common Spark optimization techniques

Technique	What it is	When to use	Example
Caching and Persistence	Storing data in memory for quick access	When you need to reuse the same data multiple times	scala val cachedData = df.cache()
Broadcast Variables	Sending a small dataset to all worker nodes	When one dataset is much smaller than the other	scala val broadcastData = spark.broadcast(smallDF) largeDF.join(broadcastData.value, "key")
Partitioning	Dividing data into smaller, manageable chunks	When dealing with large datasets to improve parallel processing	scala val partitionedData = df.repartition(10, $"key")
Avoiding Shuffles	Reducing the movement of data between nodes	To improve performance by minimizing network overhead	Use mapPartitions instead of groupBy when possible
Coalesce	Reducing the number of partitions	When the data has become sparse after a transformation	scala val coalescedData = df.coalesce(1)
Predicate Pushdown	Filtering data as early as possible in the processing	To reduce the amount of data read and processed	scala val filteredData = df.filter($"column" > 10)
Using the Right Join Strategy	Choosing the most efficient way to join two datasets	Based on the size and distribution of data	Prefer broadcast joins for small datasets
Tuning Spark Configurations	Adjusting settings to optimize resource usage	To match the workload and cluster resources	scala spark.conf.set("spark.executor.memory", "4g")
Using DataFrames/Datasets API	Leveraging the high-level APIs for optimizations	To benefit from Catalyst optimizer and Tungsten execution engine	scala val df = spark.read.csv("data.csv") df.groupBy("column").count()
Vectorized Query Execution	Processing multiple rows of data at a time	For high-performance operations on large datasets	Use built-in SQL functions and DataFrame methods

Different phases of Spark-SQL engine

Phase	Description	Details	Example
Parsing	Converting SQL queries into a logical plan.	The SQL query is parsed into an abstract syntax tree (AST).	Converting SELECT * FROM table WHERE id = 1 into an internal format.
Analysis	Resolving references and verifying the logical plan.	Resolves column names, table names, and function names; checks for errors.	Ensuring that the table table and the column id exist in the database.
Optimization	Improving the logical plan for better performance.	Transforms the logical plan using various optimization techniques; applies rules via Catalyst optimizer.	Reordering filters to reduce the amount of data processed early on.
Physical Planning	Converting the logical plan into a physical plan.	Converts the optimized logical plan into one or more physical plans; selects the most efficient plan.	Deciding whether to use a hash join or a sort-merge join.
Code Generation	Generating executable code from the physical plan.	Generates Java bytecode to execute the physical plan; this code runs on Spark executors.	Creating code to perform join operations, filter data, and compute results.
Execution	Running the generated code on the Spark cluster.	Distributes generated code across the Spark cluster; executed by Spark executors; results collected and returned.	Running join and filter operations on different nodes in the cluster and aggregating results.

Common reasons for analysis exception in Spark

Here are some common reasons why you might encounter an AnalysisException in Spark:

Reason	Description	Example
Non-Existent Column or Table	Column or table specified does not exist.	Referring to a non-existent column id.
Ambiguous Column Reference	Same column name exists in multiple tables without qualification.	Joining two DataFrames with the same column name id.
Invalid SQL Syntax	SQL query has syntax errors.	Using incorrect SQL syntax like SELCT.
Unsupported Operations	Using an operation that Spark SQL does not support.	Using an unsupported function.
Schema Mismatch	Schema of the DataFrame does not match the expected schema.	Inserting data with different column types.
Missing File or Directory	Specified file or directory does not exist.	Referring to a non-existent CSV file.
Incorrect Data Type	Operations expecting a specific data type are given the wrong type.	Performing a math operation on a string column.

Flow of how Spark works internally

Component/Step	Role/Process	Function/Example
Driver Program	Entry point for the Spark application	- Manages application lifecycle - Defines RDD transformations and actions
SparkContext	Acts as the master of the Spark application	- Connects to cluster manager - Coordinates tasks
Cluster Manager	Manages the cluster of machines	- Allocates resources to Spark applications - Examples: YARN, Mesos, standalone
Executors	Worker nodes that run tasks and store data	- Execute assigned code - Return results to the driver - Cache data in memory for quick access
Spark Application Submission	Submitting the driver program to the cluster manager	- Example: Submitting a job using spark-submit
SparkContext Initialization	Driver program initializes SparkContext	- Example: val sc = new SparkContext(conf)
Job Scheduling	Driver program defines transformations and actions on RDDs/DataFrames	- Example: val rdd = sc.textFile("data.txt").map(line => line.split(" "))
DAG (Directed Acyclic Graph) Creation	Constructing a DAG of stages for the job	- Stages are sets of tasks that can be executed in parallel - Example: A series of map and filter transformations create a DAG
Task Execution	Dividing the DAG into stages, creating tasks, and sending them to executors	- Tasks are distributed across executors - Each executor processes a partition of the data
Data Shuffling	Exchanging data between nodes during operations like reduceByKey	- Data is grouped by key across nodes - Example: Shuffling data for aggregations
Result Collection	Executors process the tasks and send the results back to the driver program	- Example: Final results of collect or count are returned to the driver
Job Completion	Driver program completes the execution	- Example: Driver terminates after executing sc.stop()

Explain DAG in Spark

Topic	Description	Details	Example
DAG in Spark	DAG stands for Directed Acyclic Graph.	- Series of steps representing the operations on data. - Directed: Operations flow in one direction. - Acyclic: No cycles or loops.	N/A
Why We Need DAG	Optimizes Execution, Fault Tolerance, and Parallel Processing.	- Optimizes Execution: Spark can optimize operations. - Fault Tolerance: Recomputes lost data if a node fails. - Parallel Processing: Divides tasks into stages for parallel execution.	N/A
Without DAG	No Optimization, No Fault Tolerance, and Less Parallelism.	- No Optimization: Operations would run as written, slower performance. - No Fault Tolerance: Inefficient data recomputation. - Less Parallelism: Harder to parallelize tasks.	N/A
Example	Example of a Spark job and DAG construction.	- Read Data: sc.textFile("file.txt") - Split Lines into Words: data.flatMap(...) - Map Words to Key-Value Pairs: words.map(...) - Reduce by Key: wordCounts.reduceByKey(...) - Collect Results: wordCounts.collect()	scala val data = sc.textFile("file.txt") val words = data.flatMap(line => line.split(" ")) val wordCounts = words.map(word => (word, 1)).reduceByKey(_ + _) wordCounts.collect()

Explain spark.sql.shuffle.partitions Variable

Topic	Description	Details	Example
spark.sql.shuffle.partitions	Configuration setting for shuffle partitions	- Default Value: 200 partitions - Defines the default number of partitions used when shuffling data for wide transformations	scala spark.conf.set("spark.sql.shuffle.partitions", "number_of_partitions")
Purpose	Optimize Performance and Control Data Distribution	- Optimize Performance: Balances workload across the cluster - Control Data Distribution: Manages how data is distributed and processed during shuffle operations	N/A
When It’s Used	Wide Transformations and SQL Queries	- Wide Transformations: reduceByKey, groupByKey, join, etc. - SQL Queries: Operations involving shuffling data like joins and aggregations	N/A
How to Set It	Setting via Configuration and spark-submit	- Configuration: spark.conf.set("spark.sql.shuffle.partitions", "number_of_partitions") - spark-submit: spark-submit --conf spark.sql.shuffle.partitions=number_of_partitions ...	N/A
Example	Default and Custom Settings	- Default Setting: scala val spark = SparkSession.builder.appName("Example").getOrCreate() println(spark.conf.get("spark.sql.shuffle.partitions")) // Output: 200 - Custom Setting: scala val spark = SparkSession.builder.appName("Example").getOrCreate() spark.conf.set("spark.sql.shuffle.partitions", "50") val df = spark.read.json("data.json") df.groupBy("column").count().show()	scala val spark = SparkSession.builder.appName("Example").getOrCreate() println(spark.conf.get("spark.sql.shuffle.partitions")) // Output: 200
Why Adjust This Setting?	Small and Large Datasets	- Small Datasets: Reduce the number of partitions to avoid too many small tasks, leading to overhead - Large Datasets: Increase the number of partitions to distribute data evenly and avoid large partitions that slow down processing	N/A