DP-900 Study Guide: Three Questions to Ask About Any Data Before You Pick an Azure Service
DP-900 (Microsoft Azure Data Fundamentals) looks like a list of services, but almost every question is really asking you to describe the data first. What shape is it? What is it used for? Who is working on it? Answer those three questions and the right Azure service usually follows, often before you read the options. This guide teaches the three questions, then shows where each answer lands on Azure.
Study guide · Microsoft Azure Data Fundamentals (DP-900)

The exam
45 minutes, passing score 700 on a scale of 1 to 1,000. Microsoft does not publish a fixed number of questions (this site's mock exam assumes 45), so plan your pace by time. Skill areas: Describe core data concepts (25–30%), Identify considerations for relational data on Azure (20–25%), Describe considerations for working with non-relational data on Azure (15–20%), Describe an analytics workload (25–30%). Every skill is worded as "describe" or "identify": you recognise and choose rather than build, although the exam page notes that some items may be interactive.
Among the 225 DP-900 questions we wrote for this question bank, one in four is a plain definition, and the services that come up most are Power BI, Azure Cosmos DB and Microsoft Fabric (about one question in ten each), then Azure SQL Database. Notice that two of those three are analytics tools: the analytics domain is as heavy as core concepts.
Question 1: what shape is the data?
- Structured: a fixed schema of rows and columns, every record with the same fields. Customers, orders, payments.
- Semi-structured: each record carries its own field names and may differ from the next. JSON documents, where one customer has three email addresses and another none.
- Unstructured: no fields a program can read directly. Images, video, audio, scanned contracts.
File formats follow the same logic. CSV, JSON and XML are human-readable; Parquet is columnar and compressed for analytics; Avro is row-based; Delta Lake adds a transaction log on top of Parquet files. Expect questions such as "which format suits large analytical reads of a few columns" (Parquet).
Question 2: what is it being used for?
| Clue in the stem | Workload |
|---|---|
| Many small reads and writes, each must be correct | Transactional (OLTP), protected by ACID: atomicity, consistency, isolation, durability. |
| Large volumes of history, read for reports and trends | Analytical (OLAP): a warehouse, a lake or a lakehouse. |
| Results needed tomorrow morning | Batch processing. |
| Each event handled as it arrives, in seconds | Stream processing. |
Question 3: who is doing the work?
Follow the data. The database administrator keeps operational databases available, secure, backed up and fast. The data engineer moves, cleans and loads data and keeps pipelines healthy. The data analyst models the prepared data and builds reports. Look for the verb: "restore a deleted table" is the DBA, "fix a nightly load" is the engineer, "build a sales dashboard" is the analyst.
Beginner tip: if SQL statement types blur together, spend twenty minutes with SQLite on your laptop or any free online SQL playground. Create a Customer table and an Order table with a foreign key (that is DDL), insert a few rows and update one (DML), then join the two tables in a SELECT. Write "DDL" or "DML" in a comment above each statement. Having typed each kind once, you will not mix up CREATE and INSERT under exam pressure.
Where the answers land on Azure
Structured and transactional: the relational services
Azure SQL Database for new cloud apps, Azure SQL Managed Instance for SQL Server instance features with near-complete compatibility, SQL Server on Azure Virtual Machines for full control of the operating system, and Azure Database for PostgreSQL or MySQL for those open-source engines. Only the VM option is IaaS. Our separate comparison of the three SQL options goes through the clues that decide between them. Watch for retired or retiring distractors: Azure Database for MariaDB is on the retirement path (Microsoft recommends migrating to Azure Database for MySQL), and Azure Cosmos DB for PostgreSQL is on a retirement path and no longer recommended for new projects.
Unstructured files: Azure Storage
Blob Storage for objects (block blobs for ordinary files, page blobs for VM disks, append blobs for logs), Azure Data Lake Storage when the hierarchical namespace is enabled for analytics, Azure Files for shares that clients map as a drive, Table storage for simple key-attribute data in one region.
Going deeper: why only the archive tier needs rehydration
Hot, cool and cold are all online tiers: each step down lowers the price of storing a blob and raises the price of reading it, but the blob still opens at once. Archive is offline. Its storage cost is the lowest, but the data is not kept in a form that can be read directly, so a blob must first be rehydrated to an online tier, which can take hours. That is why "must be readable immediately when an auditor asks" rules archive out even for data that is rarely touched. Each cooler tier also has a minimum period (30 days for cool, 90 for cold, 180 for archive), and moving or deleting a blob earlier costs an early deletion charge.
Semi-structured at global scale: Azure Cosmos DB
Choose Cosmos DB when the stem says low latency, users around the world, or a schema that changes often. One API per account: NoSQL (native, JSON, the default for new apps), MongoDB, Table, Apache Cassandra (column-family) and Apache Gremlin (graph). Know the five consistency levels from strongest to weakest: Strong, Bounded staleness, Session, Consistent prefix, Eventual. A bank balance needs Strong; a shopping cart where each user must see their own additions fits Session.
Analytics: Fabric, Databricks, streaming, Power BI
ETL transforms before loading, ELT loads first and transforms in the target. Microsoft Fabric is one SaaS platform where every workload shares OneLake and Delta tables. A Fabric lakehouse suits Spark and mixed data, and its SQL analytics endpoint is read-only; a Fabric warehouse supports full T-SQL writes. Shortcuts reference data where it is; mirroring replicates a database into OneLake. Azure Databricks is a managed analytics platform built on Apache Spark and Delta Lake. For streams: Event Hubs ingests, Stream Analytics queries over time windows (tumbling windows never overlap, hopping windows do). In Power BI, a report is one or more interactive pages over one model, a dashboard is a single page of pinned tiles, and a measure is calculated at query time while a calculated column is stored for each row.
Going deeper: how Delta Lake turns a folder of files into a table
Parquet files on their own are just files: a job that fails halfway leaves half a dataset behind, and nothing stops a file with the wrong columns from landing next to the others. A Delta table adds a transaction log beside the Parquet files. A write becomes visible only when its entry is committed to the log, which gives ACID transactions; the log records the schema, so bad rows can be rejected; and because old entries are kept, you can query an earlier version of the table. That is what lets a lakehouse offer warehouse-like guarantees on cheap lake storage, and why both Fabric and Databricks use Delta as their default table format.
A two-week order
- Days 1–4: core concepts (shapes, formats, workloads, roles). They are a quarter of the exam and make everything else easier.
- Days 5–7: relational concepts and SQL, then the Azure relational services.
- Days 8–9: Azure Storage and Cosmos DB.
- Days 10–13: analytics, Fabric, streaming and Power BI.
- Day 14: mixed practice sets, and a list of the pairs you confused.
Practise
Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.
- DP-900 exam page · All 285 DP-900 questions · Timed mock exam · 14-day study plan
- By domain: Describe core data concepts · Identify considerations for relational data on Azure · Describe considerations for working with non-relational data on Azure · Describe an analytics workload
Skills measured as listed in Microsoft's DP-900 study guide. Service names and features change, Fabric especially; check the current exam page and Microsoft Learn.