- Identity Check: What This Cheat Sheet Covers
- The Three-Domain Map at a Glance
- Domain 1 Sheet: Design and Develop Database Solutions
- Domain 2 Sheet: Secure, Optimize, and Deploy
- Domain 3 Sheet: AI and Analytics Capabilities
- Confusable Pairs That Cost Points
- Logistics Facts: Score, Renewal, and What Is Not Fixed
- Sequencing the Review by Domain
- Frequently Asked Questions
- DP-420 is now Microsoft Certified: Azure Cosmos DB AI Developer Associate; the exam code and exam title "Building AI Applications with Azure Cosmos DB" carry...
- The exam has three domains: design and develop (40-45%), secure/optimize/deploy (30-35%), and AI and analytics (20-25%).
- The passing score is 700, and renewal is annual through Microsoft's free online assessment.
- Partition key choice drives most Domain 1 decisions, from hierarchical keys to synthetic keys to multitenant anti-patterns.
Identity Check: What This Cheat Sheet Covers
This page reviews one credential: Microsoft Certified: Azure Cosmos DB AI Developer Associate (DP-420). Microsoft renamed it from Azure Cosmos DB Developer Specialty effective October 6, 2026. The exam is still DP-420, now titled Building AI Applications with Azure Cosmos DB. This is the same exam identity with the same domain structure, not a new certificate and not a replacement. If you studied from older material built around five domains, set that outline aside; the current skills measured are organized into three.
If you are still orienting yourself, the explainers on what the DP-420 certification is and what DP-420 stands for cover the naming history. For a deeper treatment of every objective, pair this sheet with the complete guide to all three DP-420 content areas.
The Three-Domain Map at a Glance
The weights below are the published ranges from the exam's domain outline. Use them to allocate review time, not to guess how many questions each domain will contribute.
| Domain | Weight | Core Theme |
|---|---|---|
| 1. Design and develop database solutions | 40-45% | Data modeling, partition keys, SDK operations, change feed, AI-assisted development |
| 2. Secure, optimize, and deploy database solutions | 30-35% | Security controls, resilience, indexing and query performance, monitoring, Fleet |
| 3. Implement AI and analytics capabilities | 20-25% | Vector/full-text/hybrid search, agent memory, Cosmos DB in Fabric, Spark |
Domain 1 plus Domain 2 account for the large majority of the exam, which means classic Cosmos DB engineering skill still dominates. The AI domain is smaller, but it is where candidates with older Cosmos DB experience tend to have the least muscle memory.
Domain 1 Sheet: Design and Develop Database Solutions
This is the heaviest domain at 40-45%, and it rewards candidates who have actually written code against the SDKs rather than only read about them.
Resource model, SDKs, and account settings
Foundations You Must Recall Cold
Know the resource hierarchy (account, database, container, item) and which settings live at which level.
- Consistency levels and what each trades away in latency, availability, or read guarantees
- Request units (RUs) as the currency of cost, and the difference between provisioned throughput and autoscale
- Throughput scope: database-level shared versus container-level dedicated
- Account and container settings that affect behavior, including TTL and indexing policy
Operations: reads, queries, writes
- Point reads (ID plus partition key) are the cheapest and fastest way to fetch one item; a query that filters on ID is not equivalent.
- Queries should include the partition key where possible to avoid cross-partition fan-out.
- Paging and continuation tokens: understand how to resume a query and why you should not assume result counts per page.
- CRUD and patch: partial document updates avoid replacing the whole item.
- Bulk, batch, and transactions: transactional batch is scoped to a single logical partition, a recurring exam theme.
- TTL and optimistic concurrency: know how expiration is configured and how ETags protect against lost updates.
Data modeling and partition keys
Partition key design is the single most testable skill in this domain. Be ready to evaluate a scenario and say which key distributes storage and request load evenly, supports the dominant query pattern, and stays within logical partition limits.
Partition Key Patterns
The objectives name several specific patterns and anti-patterns.
- Hierarchical partition keys: multi-level keys that help targeted queries and avoid hot logical partitions in large tenants
- Synthetic keys: combining properties when no single property distributes well
- Multitenant patterns: how tenant isolation interacts with key choice and throughput sharing
- Anti-patterns: low-cardinality keys, monotonically increasing keys that concentrate writes, and keys that force most queries to be cross-partition
Also be fluent in the embed-versus-reference decision: embed data that is read together and changes together; reference data that grows without bound or is shared widely.
Change feed and AI-assisted development
- Change feed modes: know the difference between modes in terms of what changes they surface (for example, whether deletes and prior versions are visible).
- Leases: the lease container coordinates distributed consumers; understand what happens when you add or remove processors.
- Copy jobs: moving data between containers is in scope.
- AI-assisted development and Agent Kit: the objectives include using AI tooling to build Cosmos DB solutions, so expect questions about when such assistance is appropriate and how to validate its output.
Domain 2 Sheet: Secure, Optimize, and Deploy
At 30-35%, this domain is where operations-minded candidates pick up points that pure developers sometimes lose.
Security
Access and Network Controls
Be able to choose the right control for a stated requirement.
- Authentication and authorization: Microsoft Entra ID with role-based access versus key-based access, and why identity-based access is generally preferred
- Network controls: private endpoints, firewall rules, and virtual network integration
- Dynamic data masking: restricting what sensitive values certain principals can see
Resilience and recovery
- Backups and restores: know the available backup modes and what restoration looks like.
- Multiregion configuration: how read and write regions are configured and how this interacts with consistency.
- Partition failover: understand failover behavior and how to plan for it.
Performance and monitoring
- Indexing policy: included and excluded paths, composite indexes for multi-property sorts and filters, and the write-cost versus query-cost trade-off.
- Query performance: reading RU charge and query metrics to diagnose expensive queries.
- Metrics and latency: interpreting throttling (HTTP 429) signals versus genuine latency problems.
- Azure Monitor and Log Analytics: where diagnostic logs go and how to query them.
- Fleet: pooled throughput and analytics across accounts, a newer topic worth reviewing from current documentation.
Key Takeaway
When a scenario describes slow or expensive queries, check three things in order: whether the partition key is in the filter, whether the indexing policy supports the filter and sort, and whether the RU charge shows throttling. That sequence resolves a large share of optimization questions.
Domain 3 Sheet: AI and Analytics Capabilities
This is the smallest domain at 20-25% but the one that most clearly justifies the renamed credential.
Search for AI workloads
Full-Text, Vector, and Hybrid Search
Know what each search type is for and how they combine.
- Full-text search: keyword and linguistic matching inside the database
- Vector search: similarity over embeddings; understand vector data types, vector indexes, embedding dimension and size considerations, and partitioning implications
- Sharded DiskANN: the index approach designed for large-scale vector workloads
- Hybrid search: combining vector and full-text signals to improve relevance
- RAG: retrieval-augmented generation, where retrieved items ground a model's response; know where Cosmos DB sits in that pipeline
Agent memory
The objectives distinguish several memory roles for AI agents. Be able to match a requirement to the right pattern:
- Durable memory: long-lived facts that persist across sessions
- Conversation-state memory: the running context of a single interaction
- Semantic memory: retrieval of relevant past information by meaning, typically via vector search
Analytics and Fabric
- Cosmos DB versus Cosmos DB in Fabric: understand which fits transactional application workloads and which fits analytics-oriented, Fabric-integrated scenarios.
- Mirroring: replicating operational data into Fabric for analysis without building a custom pipeline.
- T-SQL and Spark queries: analyzing mirrored data from the Fabric side.
- Spark connector: reading from and writing to Cosmos DB from Spark.
Confusable Pairs That Cost Points
Many wrong answers on this exam come from mixing up two plausible options. Review these pairs until the distinction is automatic.
| Pair | The Distinction to Remember |
|---|---|
| Point read vs. query by ID | Point read needs ID plus partition key and is the cheapest lookup; a query is more flexible but costs more. |
| Transactional batch vs. bulk | Batch is atomic within one logical partition; bulk optimizes throughput across many operations without that guarantee. |
| Hierarchical vs. synthetic key | Hierarchical uses multiple ordered key levels natively; synthetic concatenates properties into one value. |
| Vector search vs. hybrid search | Vector matches by embedding similarity alone; hybrid blends it with full-text signals. |
| Conversation-state vs. durable memory | Conversation state is scoped to an interaction; durable memory persists across sessions. |
| Cosmos DB vs. Cosmos DB in Fabric | The first serves application workloads; the second targets analytics inside Fabric, often via mirroring. |
Logistics Facts: Score, Renewal, and What Is Not Fixed
Two logistics facts are firm: the passing score is 700, and the credential is renewed annually through Microsoft's free online renewal assessment. For how scaled scoring works in practice, see the breakdown of the DP-420 passing score.
Several details are intentionally left out of this cheat sheet because they should be confirmed at the source rather than memorized from a secondary page: the exact item count, the exam duration, the regional fee, and any pass-rate figure. Fees vary by region, so check Microsoft's current exam page and the DP-420 certification cost breakdown before budgeting. Likewise, confirm scheduling windows using the DP-420 exam dates guide and the official study guide on Microsoft Learn.
Sequencing the Review by Domain
If you have a few weeks, order the work by dependency rather than by weight alone: Domain 1 modeling decisions underpin the optimization topics in Domain 2, and both feed the vector and analytics topics in Domain 3.
Partition keys and operations
- Build containers with good and bad keys; compare RU charges
- Practice point reads, patch, transactional batch, and TTL
Change feed and data modeling
- Run a change feed processor and observe leases
- Work embed-versus-reference scenarios
Security, resilience, and tuning
- Configure Entra ID access and a private endpoint
- Tune indexing policy and read metrics in Log Analytics
AI and analytics, then full review
- Build a vector index and a hybrid query
- Review confusable pairs and take timed practice sets
For a fuller plan, the DP-420 study guide expands this into a complete preparation strategy, and the difficulty guide helps you judge how much hands-on time you personally need. When you are ready to test recall under pressure, work through scenario-style questions on the DP-420 practice test site, and use a second pass on the weakest domain from your results.
Frequently Asked Questions
No. The exam code DP-420 is unchanged. Microsoft renamed the certification to Microsoft Certified: Azure Cosmos DB AI Developer Associate, and the exam is now titled Building AI Applications with Azure Cosmos DB. It is the same exam identity, not a replacement credential.
Three: Design and develop database solutions (40-45%), Secure, optimize, and deploy database solutions (30-35%), and Implement AI and analytics capabilities in database solutions (20-25%). Older five-domain outlines are outdated.
The passing score is 700. For context on how that figure is applied, read the passing score guide.
The certification is renewed annually through Microsoft's free online renewal assessment, so keeping it current does not require retaking the full exam.
Check Microsoft's official exam page, since fees vary by region and details can change. The DP-420 requirements guide and the cost breakdown linked above explain what to verify before you register.