DP-600 (Implementing Analytics Solutions Using Microsoft Fabric) is often the next step after PL-300, and that's exactly why it catches people. Power BI habits get you most of the way, and then the exam asks about lakehouses, Delta tables and OneLake.
First, where the points are. Prepare data is the biggest domain at 45–50%, with Maintain a data analytics solution and Implement and manage semantic models at 25–30% each. (Microsoft's pages don't fully agree on the outline right now, so check the official certification page before you book.)
Here are seven answers that look right and aren't.
1. Choosing a lakehouse for multi-table transactions
Both stores hold Delta tables, so it's tempting to treat them as interchangeable. They aren't:
- Lakehouse: files plus Delta tables, Spark-first, great for a medallion architecture. T-SQL users get a read-only SQL analytics endpoint.
- Warehouse: full T-SQL DML, stored procedures and multi-table transactions.
If the scenario says a stored procedure must update several tables in one transaction, the answer is a warehouse. And if it's streaming telemetry that needs analysis within seconds in KQL, it's neither: that's an eventhouse.
2. Copying data a shortcut could reference
Fabric has five ways to bring data in: pipeline Copy activity, Copy job, Dataflow Gen2, notebooks and shortcuts. The first four copy data. A shortcut doesn't: it points at data where it already lives (ADLS, S3, another Fabric item).
When the scenario says the data only needs to be read from another location, "copy it nightly with a pipeline" is the distractor. Deleting a shortcut also doesn't delete the target data, which is another detail questions like to test.
3. Scheduling dependent steps at guessed times
A dataflow loads staging at 02:00, and a stored procedure that builds the gold tables is scheduled for 02:30 "because the dataflow usually takes 20 minutes".
That works until the night it takes 40. The exam's answer is one pipeline that runs the dataflow, then the stored procedure on success, then the semantic model refresh, with the pipeline as the only thing on a schedule.
4. Assuming Direct Lake never falls back to DirectQuery
Direct Lake reads Delta tables from OneLake straight into memory, so it's fast without copying data. But there are two flavors, and they behave differently:
- Direct Lake on SQL (through the SQL analytics endpoint) can fall back to DirectQuery, for example when the table is a SQL view, or when the warehouse uses RLS.
- Direct Lake on OneLake never falls back.
To find out why a model is falling back, set the model's DirectLakeBehavior to DirectLakeOnly in development: queries then error instead of silently getting slower.
One more thing worth knowing: "refreshing" a Direct Lake model is framing, which just points the model at the latest version of each Delta table. It's metadata only and takes seconds.
5. Partitioning by the column with the most values
(sales_df.write.format("delta")
.partitionBy("Year", "Month")
.mode("overwrite")
.saveAsTable("sales"))
Partitioning writes one folder per value, so queries that filter on those columns read less. That only helps when the table is large and the column has few distinct values that queries actually filter on, such as date parts or region.
Partitioning a few-million-row table by CustomerID creates thousands of tiny files and makes Spark, SQL and Direct Lake slower. If an option partitions by a high-cardinality key, it's the trap.
6. Treating data masking as security
Dynamic data masking in a warehouse hides values in query results, like showing only part of an email address. It isn't a security boundary, and workspace Admins, Members and Contributors always see unmasked data.
To actually restrict access, use:
-
GRANT/DENYon columns, tables or schemas (column- and object-level security) - a security policy with a predicate function for row-level security
In a lakehouse, the equivalent is OneLake security roles, which apply to Viewers and people the item was shared with. Remember to narrow the DefaultReader role, or everyone with ReadAll keeps seeing everything.
7. Incremental refresh that reloads everything anyway
Incremental refresh needs a few details exactly right:
- two parameters named exactly
RangeStartandRangeEnd, of type Date/Time (not Date) - one inclusive and one exclusive boundary:
[OrderDate] >= RangeStart and [OrderDate] < RangeEnd. Using>=on both duplicates rows at the edges - a filter that folds to the source. Without folding, every partition reads the whole table
Also note that Direct Lake models don't need incremental refresh at all: their freshness comes from the Delta tables and framing.
How to study for this
Prepare data is half the exam, so spend half your time there. Build one small end-to-end project:
- Land data in a lakehouse with a pipeline, and add a shortcut to a second source.
- Load a warehouse star schema with a stored procedure, including a Type 2 dimension.
- Build a Direct Lake model on top, then try a SQL view and watch it fall back.
We run Skillmarked, an independent exam-prep site (not affiliated with Microsoft). The full DP-600 study guide is free: 24 lessons across all three domains, with a practice scenario after every lesson.
Top comments (0)