Databricks Spark Developer Associate exam: what to study

To pass the Databricks Certified Associate Developer for Apache Spark exam, you need to read a few lines of PySpark and know what they do without running them. Half the 45 questions sit in 2 sections: the DataFrame API (30%) and Spark SQL (20%). The rest test how Spark runs that code, plus 3 small sections that people tend to skip.

The guide was updated on 8 October 2026. The exam runs 90 minutes, with up to 10 unscored questions on top of the 45, and costs USD 200. You can take it online or at a test centre, in English. The guide asks for DataFrame work in Python and names no Spark version, so study from the current Apache Spark documentation. Databricks does not publish the pass mark.

One oddity: the published PDF still carries its authors' drafting notes in square brackets, such as "[Author: list current, non-deprecated training only by name...]". They are notes to the people who wrote it, not things you need to know.

What "reading code" looks like

Most questions show a short block and ask what it returns, what it costs, or which line is wrong. Take this one:

orders = spark.read.table("sales.orders")
big = orders.filter("amount > 100").select("id", "amount")
big.cache()
print(big.count())
big.write.mode("overwrite").saveAsTable("sales.big_orders")

Filter, select and cache do no work. Spark only builds a plan until an action asks for a result, and here that is count(), which runs a job and fills the cache. The write is the second action and reads from the cache. If you can say that without thinking, you have objective 1.6 (actions, transformations and lazy evaluation). Along the way you will also have touched caching (1.3) and save modes (2.2). The how Spark runs a job lessons go through jobs, stages and tasks with the Spark UI.

The pairs to tell apart

A lot of wrong answers look right because they use the twin of the correct method. These are the pairs worth drilling:

Join questions add one more: a broadcast join copies the small side to every executor and skips the shuffle. Adaptive Query Execution, on by default since Spark 3.2, can switch a sort-merge join to a broadcast join at runtime when one side turns out to be small (4.2). See the joins guide for when each wins.

The 20% people skip

Structured Streaming (10%), Spark Connect (5%) and the pandas API on Spark (5%) add up to 9 questions. That is too many to give away, and each needs only a few ideas.

• Streaming. Know the micro-batch model, the 3 output modes (append, update, complete), which sinks allow which mode, and deduplication with and without a watermark. The event time and exactly-once lessons cover the core.
• Spark Connect. The client sends unresolved plans to the server over gRPC. Because the client is not in the driver's process, there is no SparkContext and no RDD API. That one fact explains most of its limits.
• Pandas API on Spark. Know why a Pandas UDF beats a plain Python UDF (Arrow moves the data and pandas works on it in vectorized form), and why toPandas() on a big DataFrame runs the driver out of memory.

How to practise

Databricks Free Edition is enough for most of this, and the guide points to it for hands-on work. Keep the PySpark cheat sheet open while you write. For jobs and stages, a local PySpark install shows the Spark UI at port 4040, where you can see where each job shuffled. The troubleshooting section (10%) asks you to read driver and executor logs for out-of-memory errors and idle clusters, which only sticks once you have caused a few yourself.

When you can predict the output of code blocks like the one above, try the free Spark Developer practice questions. If you are deciding between this exam and the Data Engineer track, read which Databricks certification to take first.