DP-420 Study Guide: Passing the Azure Cosmos DB AI Developer Associate Exam

What Changed on October 6, 2026

Microsoft announced a new name and level for the certification. The certification page says the change goes into effect on October 6, 2026 and that no action is required from existing holders. The study guide's change log marks the audience profile as a major change and lists the following:

• Names: the exam title changes from "Designing and Implementing Cloud-Native Applications Using Microsoft Azure Cosmos DB" to "Building AI Applications with Azure Cosmos DB", and the certification moves from Specialty to Associate.
• Structure: five domains become three. Data distribution, integration and maintenance no longer exist as separate domains; their surviving content is folded into the new ones.
• Removed: server-side programming with JavaScript (stored procedures, triggers, UDFs), data movement (Azure Data Factory, Kafka, Stream Analytics), the DevOps process (ARM templates), Synapse Link and analytical store workloads, and "implement solutions across services".
• New: vector search, full-text and hybrid search, sharded DiskANN, Retrieval-Augmented Generation (RAG), agent memory stores, Fabric mirroring and querying, the Azure Cosmos DB Agent Kit, fleets, dynamic data masking, copy container jobs, and the all versions and deletes change feed mode.
• Audience: you should have experience developing applications by using C# or Python and be familiar with generative AI concepts and retrieval-based systems. The previous profile asked for reading C# or Java code and creating server-side objects with JavaScript.

Microsoft updates the English version of an exam first; localized versions are updated about eight weeks later. If you take DP-420 in another language soon after October 6, check which version you will get. Our DP-420 practice bank has already been updated to the October 6 objectives.

DP-420 vs DP-900: Why "I Already Know Cosmos DB" Isn't Enough

Many DP-420 candidates already hold DP-900 (Azure Data Fundamentals), and Cosmos DB appears in both exams. The overlap is small. DP-900's non-relational data domain gives Cosmos DB two bullets, identifying use cases and describing its APIs, and moves on. It never asks you to reason about consistency levels, partition key design under a real workload, change feed, or RU cost.

DP-900 does have an analytics domain that mentions Microsoft Fabric, which gives you some context for DP-420's Fabric objectives. But DP-420 asks you to configure and monitor mirroring and query mirrored JSON with T-SQL and Spark, and it adds vector search, RAG and agent memory, none of which DP-900 covers. If your only Cosmos DB exposure is DP-900, treat this guide as a first pass, not a refresher. The DP-900 study guide and this one share little beyond the product name.

DP-420 vs AZ-204: Coming From General Azure Development

The other common starting point is AZ-204 (Developing Solutions for Microsoft Azure). AZ-204 was retired on July 31, 2026, but many developers still hold it or have the equivalent hands-on experience. Those candidates already understand SDK patterns, async operations, retry logic, managed identities and Microsoft Entra ID authentication.

AZ-204 covered Cosmos DB with three bullets: perform operations on containers and items with the SDK, set the consistency level, and implement change feed notifications. DP-420 assumes that base and goes much further on one data platform: partition key strategy (including hierarchical and synthetic keys), RU cost per operation, multi-region writes and partition failover, dynamic data masking, query and index metrics, and an entire domain on vector, full-text and hybrid search, RAG, agent memory and Fabric.

In practice, an AZ-204-level developer can skip the SDK ramp-up and go straight to Cosmos DB-specific mechanics and the AI domain. A DP-900 holder needs to build application-development context in C# or Python first.

The Three Domains and Where the Points Are

This table compares a Fundamentals-level view of each topic with what DP-420 asks. Weightings are from Microsoft's official DP-420 study guide (skills measured as of October 6, 2026, checked October 2, 2026).

Design and develop database solutions carries the most weight at 40–45%. It holds the partition key and data modeling decisions that the other two domains depend on: a poor partition key makes failover, query tuning and vector search scoping harder later.

Domain-by-Domain Study Notes

1. Design and develop database solutions (40–45%)

• Resource models and SDK: the five consistency levels and their tradeoffs, RU consumption, provisioned vs autoscale throughput, account- and container-level settings, and initializing SDK clients (one client instance per application, not per request).
• Database operations: parameterized queries, paging with continuation tokens, point reads, create/replace/upsert/patch/delete, bulk and transactional batch operations, TTL, and optimistic concurrency with ETags.
• Data modeling and partitioning: embed vs reference, choosing partition keys by cardinality and write distribution, hierarchical partition keys for multi-tenant designs, synthetic keys, spotting hot partitions and other anti-patterns, and estimating workload cost.
• Change feed: change feed processor vs pull model, lease containers, and the two modes. Latest version mode doesn't capture deletes. All versions and deletes mode captures creates, updates and deletes (including TTL expirations) but requires continuous backup and only returns changes within the backup retention period.
• Copy container jobs: used for scenarios such as changing a container's partition key. Offline jobs need writes to the source stopped; online jobs use the all versions and deletes change feed to replicate incremental changes. Microsoft's documentation labels copy jobs as preview, and the study guide notes the exam may include preview features that are commonly used.
• Agent Kit: an open-source set of skills that gives AI coding assistants (GitHub Copilot, Claude Code, Gemini CLI and others) Azure Cosmos DB best-practice rules for data modeling, partition keys, query optimization, SDK usage, vector and full-text search. It provides guidance only and doesn't run operations on your database.

2. Secure, optimize, and deploy database solutions (30–35%)

• Security: Microsoft Entra ID and data plane RBAC vs account keys, network security, and dynamic data masking. Masking is applied at query time for users without the unmask data action; the stored data doesn't change. It requires Microsoft Entra ID (account keys aren't supported) and can't be turned off once enabled on an account.
• Availability and recovery: periodic vs continuous backup and restore, multi-region read and write configurations, and partition failover. Per-partition automatic failover moves writes only for the affected partitions of a single-write-region account to the next region instead of failing over the whole account.
• Performance: query and index metrics, end-to-end latency against service objectives, Azure Monitor alerts, and diagnostic logs in Log Analytics.
• Fleets: a fleet groups many accounts (typically one per tenant) into fleetspaces, even across subscriptions and resource groups. Pools let those accounts draw shared RU/s on top of their dedicated throughput, and fleet analytics exports usage data to Microsoft Fabric or an Azure Storage account.

3. Implement AI and analytics capabilities in database solutions (20–25%)

• Vector data and indexes: container vector policies and the three vector index types. flat is exact search up to 505 dimensions; quantizedFlat and diskANN support up to 4,096 dimensions. Microsoft recommends quantizedFlat when a search is scoped to roughly 50,000 vectors or fewer and DiskANN above that. Both need at least 1,000 vectors indexed before the index is used. Vector policies and indexes can't be changed after creation.
• Sharded DiskANN: setting vectorIndexShardKey (for example /tenantId) creates one DiskANN index per value, so filtered searches hit a smaller index.
• Full-text and hybrid search: a full-text policy and index, FullTextScore (BM25), and hybrid queries that combine VectorDistance and FullTextScore with ORDER BY RANK RRF(...), optionally weighted.
• RAG and agent memory: retrieval patterns that ground an LLM in your data, durable memory stores, conversation state (short-term turns, often expired with TTL) and semantic memory (summaries and facts retrieved by vector or hybrid search).
• Operational analytics: choosing between Azure Cosmos DB and Cosmos DB in Microsoft Fabric, configuring and monitoring mirroring into Fabric OneLake (continuous backup is a prerequisite), querying mirrored JSON with T-SQL and Spark, and reading and writing Cosmos DB data from Fabric with the Cosmos DB Spark connector.

Studying DP-420 From DP-900 vs From AZ-204

There is no single generic study plan here. Where you start determines which weeks you can compress.

Pitfalls on the New DP-420

Common DP-420 Questions

Related Resources