Driver, executors & cluster managers
Apache Spark · Spark Architecture
Theory
Every Spark application has one driver and many executors. The driver is the process that runs your main program (your PySpark script or notebook): it holds the SparkSession, turns your code into a plan, splits it into tasks and tracks their progress. Executors are worker processes on the cluster machines; each one has a number of cores (task slots) and memory, runs the tasks the driver sends, and can cache data for reuse.
The cluster manager is the service that hands out machines: Spark's own Standalone manager, Hadoop YARN or Kubernetes, while local mode runs everything in one process. For example, with 10 executors of 4 cores each, Spark can run 40 tasks at the same time; if a 400-partition dataset is processed, the driver schedules 40 tasks at a time across the executors until all 400 are done. Calling collect() brings results back to the driver, so large results can crash it.
Why we use it
Knowing which code runs on the driver and which runs on executors explains most performance and memory problems: out-of-memory on the driver after collect(), idle executors from too few partitions, or slow jobs from too few cores. It is also a standard interview question.
When to use it
Think about this whenever you size a cluster (executor count, cores, memory), choose a deploy mode, or debug failures. In client mode the driver runs where you submit (handy for notebooks); in cluster mode it runs inside the cluster (better for production jobs).
How it works
- spark-submit or notebook starts the driver
- Driver asks the cluster manager for executors
- Cluster manager launches executor processes
- Driver turns actions into stages and tasks
- Executors run tasks on their partitions
- Did an executor die? Yes: Driver re-schedules its tasks elsewhere. No: Executors return results.
- Driver finishes the job and releases resources
Syntax
spark-submit --master yarn --deploy-mode cluster \
--num-executors 10 --executor-cores 4 --executor-memory 8g \
--driver-memory 4g job.py
--master local[*] | spark://host:7077 | yarn | k8s://https://host:443Example
from pyspark.sql import SparkSession, functions as F
spark = (SparkSession.builder.master("local[1]")
.config("spark.ui.enabled", "false")
.getOrCreate())
orders = spark.createDataFrame([
(101, 1, "North", "Laptop", 1, 55000.0, "2026-01-05"),
(102, 2, "West", "Phone", 2, 30000.0, "2026-01-06"),
(103, 1, "North", "Mouse", 3, 1500.0, "2026-01-06"),
(104, 3, "South", "Laptop", 1, 52000.0, "2026-01-07"),
(105, 4, "West", "Tablet", 1, 25000.0, "2026-01-08"),
(106, 2, "East", "Phone", 1, 15000.0, "2026-01-08"),
(107, 5, "North", "Tablet", 2, 48000.0, "2026-01-09"),
(108, 3, "South", "Mouse", 4, 2000.0, "2026-01-10"),
], "order_id INT, customer_id INT, region STRING, product STRING, "
"qty INT, amount DOUBLE, order_date STRING"
).withColumn("order_date", F.to_date("order_date"))
sc = spark.sparkContext
print("master:", sc.master)
print("default parallelism (task slots):", sc.defaultParallelism)
# Executors work on partitions; spark_partition_id() shows which one holds each row
parts = orders.repartition(3)
print("partitions:", parts.rdd.getNumPartitions())
(parts.groupBy(F.spark_partition_id().alias("partition"))
.count().orderBy("partition").show())
# collect() brings rows back to the DRIVER: fine for small results only
top = orders.orderBy(F.desc("amount")).limit(2).collect()
print([r.order_id for r in top])
# Output:
# master: local[1]
# default parallelism (task slots): 1
# partitions: 3
# +---------+-----+
# |partition|count|
# +---------+-----+
# | 0| 3|
# | 1| 3|
# | 2| 2|
# +---------+-----+
#
# [101, 104]
Key points
- Driver: runs your program, builds the plan, schedules tasks, collects results
- Executors: run tasks on partitions, hold cached data, report back to the driver
- One task per partition per stage; parallelism = total executor cores
- Cluster managers: Standalone, YARN, Kubernetes; local mode for development
- collect(), toPandas() and large broadcasts use driver memory