[lochan2014]

list of top 20 scenario-based PySpark questions that are frequently asked in interviews:
🎯 #1: You have a large dataset. How do you optimize reading only specific partitions using spark.read.parquet()?
🎯 #2: Explain how you would handle schema mismatch when reading multiple JSON/Parquet files with different structures.
🎯 #3: You want to write a DataFrame back to Parquet but keep the original file size consistent. What options do you use?
🎯 #4: Why might a join operation cause executor OOM errors , and how can you avoid it?
🎯 #5: You’re reading a CSV with missing values. How would you replace nulls dynamically across all columns?
🎯 #6: How do you read a nested JSON and flatten it into a tabular format using PySpark?
🎯 #7: How would you implement slowly changing dimension Type 2 (SCD2) logic using PySpark?
🎯 #8: You need to read from a Kafka topic where each message is a JSON. How would you parse and process it?
🎯 #9: How would you perform incremental data loading from a source like MySQL to Delta Lake?
🎯 #10: What’s the difference between cache() and persist()? When would you use one over the other?
🎯 #11: You have skewed data during a groupBy. How would you optimize this?
🎯 #12: How would you write unit tests for your PySpark code?
🎯 #13: You want to add a row number to a DataFrame partitioned by a key and ordered by timestamp. How do you do it?
🎯 #14: You’re writing to a Delta table and want to merge only new records. How do you use MERGE INTO?
🎯 #15: Your Spark job is taking too long. How would you go about debugging and optimizing performance?
🎯 #16: How would you read a huge CSV file and write it as partitioned Parquet based on a date column?
🎯 #17: You want to broadcast a small DataFrame in a join. How do you do it and what are the caveats?
🎯 #18: You’re processing streaming data from Kafka. How would you ensure exactly-once semantics?
🎯 #19: You have a list of dates per user and want to generate a daily activity flag for each day in a month. How do you do it?
🎯 #20: You have a PySpark script that runs fine locally but fails on the cluster. What could be the possible reasons?
Thanks to https://www.linkedin.com/feed/update/urn:li:activity:7337048186416439296/

[lochan2014]

To **optimize reading only specific partitions** using spark.read.parquet(), you can leverage **partition pruning**, which allows Spark to skip reading unnecessary files and only load relevant partitions.

### ✅ Here’s how:

—

### 💡 **Assume your dataset is partitioned like this:**

/data/events/year=2023/month=01/
/month=02/
/year=2024/month=01/
`

—

### ✅ Option 1: **Partition Pruning via Filter**

If the Parquet data is partitioned by year and month, Spark will **automatically prune partitions** when you filter on those columns **after reading**:

df = spark.read.parquet(“/data/events”)
df_filtered = df.filter((df.year == 2023) & (df.month == 1))
`

➡️ Spark reads **only the matching folders** using directory structure without scanning full data.

—

### ✅ Option 2: **Partition Directory Path Filtering (Manual)**

You can also directly specify the partition path to **only read that portion** of the dataset:

df = spark.read.parquet(“/data/events/year=2023/month=01”)
`

➡️ This skips the full directory scan and reads only data from specified partition paths.

—

### ✅ Option 3: **Passing multiple paths**

If you want to read a few selected partitions:

df = spark.read.parquet(
“/data/events/year=2023/month=01”,
“/data/events/year=2024/month=01”
)
`

➡️ Only the specified paths are read—efficient when you know exact partitions to load.

—

### 🛠️ Best Practices

* **Use filter on partition columns early** (for lazy pruning).
* Always **partition your data** based on query patterns (e.g., time-based like year/month/day).
* Avoid filtering on non-partitioned columns if performance is a concern during reads.
* **Don’t cast partition columns**, as it can disable pruning (e.g., df.filter(df.year.cast("string") == "2023") may break pruning).

—

### 🔍 Check if pruning is applied:

You can verify partition pruning by enabling physical plan explain:

df_filtered.explain(True)
`

Look for PushedFilters: and ensure only necessary partitions are being accessed.

[lochan2014]

from pyspark.sql import SparkSession
from pyspark.sql.utils import AnalysisException

spark = SparkSession.builder \
.appName(“SchemaMismatchHandler”) \
.getOrCreate()

from pyspark.sql.functions import lit

# Example file paths
file_paths = [“s3://bucket/data1.json”, “s3://bucket/data2.json”, “s3://bucket/data3.json”]

# Step 1: Read all files separately and collect schemas
dfs = []
all_fields = set()

for path in file_paths:
try:
df = spark.read.json(path)
dfs.append(df)
all_fields.update(df.schema.fieldNames())
except AnalysisException as e:
print(f”Failed to read {path}: {e}”)

# Step 2: Align all DataFrames to a common schema
def align_schema(df, all_columns):
for col in all_columns:
if col not in df.columns:
df = df.withColumn(col, lit(None))
return df.select(sorted(all_columns)) # Sort for consistent column order

# Step 3: Apply schema alignment and union all
aligned_dfs = [align_schema(df, all_fields) for df in dfs]
final_df = aligned_dfs[0]
for df in aligned_dfs[1:]:
final_df = final_df.unionByName(df)

# final_df is now a unified DataFrame with aligned schema
final_df.show()
`

This handles schema mismatches by:

* Reading each file independently.
* Building a union of all column names.
* Padding missing columns with null.
* Ensuring consistent column order before unionByName.

[lochan2014]

`python
# Write DataFrame to Parquet while controlling file size

# Step 1: Estimate desired file size per file (in bytes)
# For example, target ~128MB per file
target_file_size = 128 * 1024 * 1024 # 128 MB

# Step 2: Use df.rdd.getNumPartitions() and data size to tune partitions
# Alternatively, use coalesce/repartition if size is known or can be estimated
# Suppose you want 10 files around 128MB each
df = df.repartition(10)

# Step 3: Write with compression to reduce actual storage size
df.write \
.mode(“overwrite”) \
.option(“compression”, “snappy”) \
.parquet(“s3://your-bucket/target-path/”)

# Optional: Use maxRecordsPerFile to cap records per file if you know average row size
df.write \
.mode(“overwrite”) \
.option(“compression”, “snappy”) \
.option(“maxRecordsPerFile”, 1_000_000) \
.parquet(“s3://your-bucket/target-path/”)
`

**Key options used:**

* .repartition(n) to control the number of output files.
* .option("maxRecordsPerFile", n) to cap the number of rows per file.
* Compression (snappy) to ensure consistent file size and performance.

[lochan2014]

# REASON: Join operations can cause executor Out Of Memory (OOM) errors when:
# – One side of the join (especially in a shuffle join) is too large to fit in executor memory.
# – Skewed keys cause data to pile up in specific partitions.
# – Spark performs a wide transformation and shuffles too much data.

# HOW TO AVOID OOM IN JOINS:

# 1. Use Broadcast Join (if one side is small enough)
from pyspark.sql.functions import broadcast

result = large_df.join(broadcast(small_df), “join_key”)

# 2. Repartition by join key to evenly distribute data
df1 = df1.repartition(“join_key”)
df2 = df2.repartition(“join_key”)

# 3. Filter unnecessary columns before join to reduce memory footprint
df1_trimmed = df1.select(“join_key”, “needed_col1”)
df2_trimmed = df2.select(“join_key”, “needed_col2”)

# 4. Handle skew by salting
from pyspark.sql.functions import rand, concat_ws

# Salt one side of the skewed join
df1 = df1.withColumn(“salted_key”, concat_ws(“_”, df1[“join_key”], (rand() * 10).cast(“int”)))
df2 = df2.withColumn(“salted_key”, concat_ws(“_”, df2[“join_key”], (rand() * 10).cast(“int”)))

# Then join on salted_key

# 5. Use spark.sql.autoBroadcastJoinThreshold wisely
spark.conf.set(“spark.sql.autoBroadcastJoinThreshold”, 50 * 1024 * 1024) # 50MB

# 6. Monitor partition sizes and optimize memory settings
# Avoid shuffling massive partitions; tune spark.sql.shuffle.partitions

# 7. Cache intermediate DataFrames if reused
df1.cache()
`

**Summary:**
OOM during joins typically occurs due to shuffling large or skewed data. Mitigate it using broadcast joins, repartitioning, filtering columns, salting for skew, tuning configs, and caching when needed.

[lochan2014]

from pyspark.sql import SparkSession
from pyspark.sql.functions import col, lit, when

spark = SparkSession.builder.appName(“NullHandling”).getOrCreate()

# Step 1: Read CSV with header and infer schema
df = spark.read.option(“header”, True).option(“inferSchema”, True).csv(“path/to/file.csv”)

# Step 2: Dynamically replace nulls based on column data types
for column, dtype in df.dtypes:
if dtype in [“int”, “bigint”, “double”, “float”]:
df = df.withColumn(column, when(col(column).isNull(), lit(0)).otherwise(col(column)))
elif dtype == “string”:
df = df.withColumn(column, when(col(column).isNull(), lit(“missing”)).otherwise(col(column)))
elif dtype == “boolean”:
df = df.withColumn(column, when(col(column).isNull(), lit(False)).otherwise(col(column)))
# Add other type-specific defaults as needed

# Final cleaned DataFrame
df.show()
`

### ✅ Summary:

* Automatically detects column types.
* Fills nulls with type-appropriate defaults:
0 for numbers, "missing" for strings, False for booleans.
* Avoids hardcoding column names.

[lochan2014]

Absolutely! Here’s a crisp, sharp snippet-style answer for each of the commonly asked questions:

—

### 🔹 **How do you handle corrupt records while reading JSON/CSV?**

# For JSON
df = spark.read.option(“badRecordsPath”, “s3://path/to/badRecords”).json(“s3://path/to/json”)

# For CSV
df = spark.read.option(“mode”, “PERMISSIVE”) \
.option(“columnNameOfCorruptRecord”, “_corrupt_record”) \
.csv(“s3://path/to/csv”)

# Options: PERMISSIVE (default), DROPMALFORMED, FAILFAST

### 🔹 **How do you handle duplicate rows in a DataFrame?**

# Remove exact duplicates
df = df.dropDuplicates()

# Remove duplicates based on specific columns
df = df.dropDuplicates([“col1”, “col2”])
`

—

### 🔹 **How to handle nulls before aggregation?**

# Replace nulls before groupBy
df = df.fillna(0) # or df.fillna({“col1”: 0, “col2”: “unknown”})

# Then safely aggregate
df.groupBy(“group_col”).agg({“metric_col”: “sum”})
`

—

### 🔹 **How do you read only specific columns from a Parquet file?**

df = spark.read.parquet(“s3://path/to/data”).select(“col1”, “col2”)
`

—

### 🔹 **How do you optimize wide transformations like joins and groupBy?**

# Use broadcast join if one side is small
from pyspark.sql.functions import broadcast
df = large_df.join(broadcast(small_df), “key”)

# Repartition on group key to reduce shuffle pressure
df = df.repartition(“group_key”)

# Cache intermediate results if reused
df.cache()
`

—

### 🔹 **How do you write partitioned Parquet files with overwrite behavior?**

df.write.mode(“overwrite”) \
.partitionBy(“year”, “month”) \
.parquet(“s3://path/output/”)
`

—

### 🔹 **How do you check for skew in Spark join keys?**

df.groupBy(“join_key”).count().orderBy(“count”, ascending=False).show(10)
`

—

Would you like a few more on file formats, cache vs persist, or checkpointing?

[lochan2014]

🎯 #6: Flatten Nested JSON to Tabular Format

from pyspark.sql.functions import col

df = spark.read.json(“path/to/nested.json”)

# Flatten all nested structs
flattened_df = df.selectExpr(“id”, “details.name”, “details.address.city”, “details.address.zip”)

🎯 #7: SCD Type 2 in PySpark

from delta.tables import DeltaTable
from pyspark.sql.functions import current_date, lit

# Existing Delta table
target = DeltaTable.forPath(spark, “s3://path/delta_table”)

# New data
incoming_df = spark.read.parquet(“s3://path/incoming”)

# Merge logic
target.alias(“t”).merge(
incoming_df.alias(“s”),
“t.id = s.id AND t.current_flag = true AND (t.attr1 != s.attr1 OR t.attr2 != s.attr2)”
).whenMatchedUpdate(set={
“current_flag”: lit(False),
“end_date”: current_date()
}).whenNotMatchedInsert(values={
“id”: “s.id”,
“attr1”: “s.attr1”,
“attr2”: “s.attr2”,
“start_date”: current_date(),
“end_date”: lit(None),
“current_flag”: lit(True)
}).execute()

🎯 #8: Parse Kafka JSON Messages

from pyspark.sql.functions import from_json, col
from pyspark.sql.types import StructType, StringType

schema = StructType().add(“user_id”, StringType()).add(“action”, StringType())

df = spark.readStream \
.format(“kafka”) \
.option(“subscribe”, “topic-name”) \
.load()

json_df = df.select(from_json(col(“value”).cast(“string”), schema).alias(“data”)).select(“data.*”)

🎯 #9: Incremental Load from MySQL to Delta

# Step 1: Load latest watermark
last_ts = spark.read.format(“delta”).table(“delta_table”).agg({“updated_at”: “max”}).collect()[0][0]

# Step 2: Pull only new rows
jdbc_df = spark.read \
.format(“jdbc”) \
.option(“url”, “jdbc:mysql://…”) \
.option(“query”, f”SELECT * FROM table WHERE updated_at > ‘{last_ts}'”) \
.load()

# Step 3: Upsert into Delta
from delta.tables import DeltaTable
delta_table = DeltaTable.forPath(spark, “s3://delta_table”)

delta_table.alias(“t”).merge(
jdbc_df.alias(“s”),
“t.id = s.id”
).whenMatchedUpdateAll().whenNotMatchedInsertAll().execute()

🎯 #10: cache() vs persist()

df.cache() # Uses MEMORY_AND_DISK by default
df.persist() # Customize storage level, e.g., MEMORY_ONLY, DISK_ONLY

# Use cache() when memory suffices and the data is reused multiple times.
# Use persist() when memory may not be enough or if you want control.

🎯 #11: Optimize Skewed groupBy

# Add random salt to reduce skew
from pyspark.sql.functions import expr

df = df.withColumn(“salted_key”, expr(“concat(group_key, ‘_’, floor(rand()*10))”))

df.groupBy(“salted_key”).agg(…) # Later remove salt if needed

🎯 #12: Unit Tests in PySpark

import unittest
from pyspark.sql import SparkSession

class MyTest(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.spark = SparkSession.builder.master(“local[1]”).appName(“test”).getOrCreate()

def test_transformation(self):
df = self.spark.createDataFrame([(1, “a”), (2, “b”)], [“id”, “val”])
result = df.filter(“id = 1”).collect()
self.assertEqual(result[0][“val”], “a”)

🎯 #13: Row Number by Partition

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

window_spec = Window.partitionBy(“user_id”).orderBy(“event_time”)

df = df.withColumn(“row_num”, row_number().over(window_spec))

🎯 #14: MERGE INTO Delta for New Records

from delta.tables import DeltaTable

delta = DeltaTable.forPath(spark, “s3://delta_table”)
new_data = spark.read.parquet(“s3://incoming”)

delta.alias(“t”).merge(
new_data.alias(“s”),
“t.id = s.id”
).whenNotMatchedInsertAll().execute()

🎯 #15: Debug & Optimize Slow Spark Jobs

# 🔍 Steps:
# 1. Use Spark UI → check stages, skew, shuffle size.
# 2. Cache/persist reused DataFrames.
# 3. Avoid UDFs unless necessary.
# 4. Repartition wisely.
# 5. Tune configs:
spark.conf.set(“spark.sql.shuffle.partitions”, 200)
spark.conf.set(“spark.sql.autoBroadcastJoinThreshold”, 50*1024*1024)

🎯 #16: CSV → Partitioned Parquet by Date

df = spark.read.option(“header”, True).csv(“s3://path/data.csv”)

df.write.partitionBy(“date_col”) \
.mode(“overwrite”) \
.parquet(“s3://path/output/”)

🎯 #17: Broadcast Join + Caveats

from pyspark.sql.functions import broadcast

df = big_df.join(broadcast(small_df), “id”)

# ⚠️ Caveat: Ensure small_df fits in executor memory (~<10MB-50MB)

🎯 #18: Kafka Streaming with Exactly-Once

stream_df = spark.readStream.format("kafka")... # your stream config

# Ensure checkpointing + idempotent sink (like Delta)
stream_df.writeStream \
.format("delta") \
.option("checkpointLocation", "s3://chk/") \
.outputMode("append") \
.start("s3://delta/output")

🎯 #19: Daily Activity Flag

from pyspark.sql.functions import explode, sequence, lit, col, to_date

# List of active dates per user
df = df.withColumn("date", to_date("date"))

# Generate full month range
full_dates = spark.range(1).select(explode(sequence(lit("2024-01-01"), lit("2024-01-31"))).alias("date"))

# Cross join users with date range
users = df.select("user_id").distinct()
calendar = users.crossJoin(full_dates)

# Left join and flag activity
result = calendar.join(df, ["user_id", "date"], "left") \
.withColumn("active_flag", (col("df.date").isNotNull()).cast("int"))

🎯 #20: Script Fails on Cluster but Not Locally – Possible Causes

1. Missing files/resources on cluster (e.g., JARs, dependencies).
2. Schema inference issues (e.g., inferSchema=True works locally but fails on cluster scale).
3. Cluster memory limits (OOM).
4. Code depends on local paths not accessible in distributed mode.
5. Local mode uses single JVM – cluster uses distributed shuffle, causing real execution paths.
6. Version mismatch (e.g., Spark/Python/PyArrow).
7. Missing permissions on HDFS/S3.

[Author]

Posts