DP-800 Cheat Sheet: Key Facts to Review

How to use this sheet

This is a review aid, not a full course, and it does not cover everything on the exam. Skim it after you have studied, and mark every line you cannot explain from memory. Then go back to the domains explained page or Microsoft’s own documentation for those lines. The facts come from the skills list dated October 19, 2026.

Exam facts at a glance

Item	Fact
Exam	DP-800, Developing AI-Enabled Database Solutions
Credential	Microsoft Certified: SQL AI Developer Associate
Time	120 minutes (check your booking for the exact time)
Passing score	700 on a scale of 1 to 1,000
Delivery	Proctored through Pearson VUE
Question count	Not published
Domain 1	Design and develop database solutions, 35-40%
Domain 2	Secure, optimize, and deploy database solutions, 35-40%
Domain 3	Implement AI capabilities in database solutions, 25-30%
Renewal	Yearly, free, open-book assessment

Domain 1: Design and develop database solutions

Key concepts

Table types: in-memory, temporal, external, ledger, graph. Remember one purpose each.
Constraints: PRIMARY KEY, FOREIGN KEY, UNIQUE, CHECK, DEFAULT.
Programmability: views, scalar and table-valued functions, stored procedures, triggers.
T-SQL: CTEs, window functions, JSON functions, regular-expression functions, fuzzy matching, graph MATCH, error handling.
AI-assisted tools: GitHub Copilot, instruction files, MCP server connections, and the security impact of letting a tool see or run code.

Easily confused pairs

Temporal is for row history. Ledger is for tamper evidence. Graph is for relationships.
JSON_VALUE returns a scalar. JSON_QUERY returns an object or an array.
CROSS APPLY drops left rows with no match. OUTER APPLY keeps them.
Inline table-valued function is one query. Multi-statement table-valued function builds a table variable in steps.
EDIT_DISTANCE counts changes. JARO_WINKLER_DISTANCE scores similarity, with a bonus for matching starts.

Decision rules

If the scenario needs history of changes to rows, think temporal.
If it says “prove no one altered the data”, think ledger.
If it says “who is connected to whom, how many hops”, think graph and MATCH.
If you need a value that changes unpredictably in the results, think about NULL handling first (JSON_ARRAY and JSON_OBJECT differ here).

Traps

Assuming a trigger fires once per row. It fires once per statement.
Forgetting that an AI assistant that runs model-written T-SQL carries more risk than one that calls fixed operations.

Domain 2: Secure, optimize, and deploy

Key concepts

Security: Always Encrypted, column-level encryption, Dynamic Data Masking, Row-Level Security, permissions, auditing, passwordless access, Managed Identity, securing GraphQL, REST, and MCP endpoints.
Performance: isolation levels, execution plans, DMVs, Query Store, Query Performance Insight, blocking, deadlocks.
CI/CD: SQL Database Projects (including SDK-style), reference data in source control, branching and pull requests, secrets management, schema drift, pipeline controls.
Integration: Data API builder, Azure Monitor with Application Insights and Log Analytics, change event streaming, CDC, Change Tracking, Azure Functions, Logic Apps.

Easily confused pairs

Masking hides values in results. Always Encrypted keeps plaintext from administrators.
Query Store keeps query history. DMVs show the current state.
Estimated plan does not run the query. Actual plan does.
CDC records changed data. Change Tracking records that a row changed.
Project means desired state. Deployed database means actual state. Drift is the gap.

Decision rules

“Even administrators must not read it” points to Always Encrypted, not masking.
“Different users see different rows” points to Row-Level Security.
“Which query got slow last Tuesday” points to Query Store.
“A secret in a repository” is always wrong; use secrets management.
In Data API builder, check whether a runtime-level setting overrides an entity-level one.

Traps

Raising the isolation level when an index is the real fix.
Treating masking as protection against someone who can read the files or run unmasked queries.

Domain 3: Implement AI capabilities

Key concepts

External models: evaluate (multimodal, multilingual, size, structured output), create, manage, secure.
Embeddings: which columns, how to chunk, how to keep them current (triggers, Change Tracking, Azure Functions, Logic Apps, CDC, Microsoft Foundry).
Search: full-text, vector, hybrid. VECTOR_NORMALIZE, VECTOR_DISTANCE, VECTORPROPERTY, VECTOR_SEARCH. Vector index types and metrics. Reciprocal rank fusion.
RAG: when it fits, building a prompt with sp_invoke_external_rest_endpoint, shaping data as JSON, sending to a model, reading the response.

Easily confused pairs

Full-text matches words. Vector search matches meaning. Hybrid combines both.
ANN is faster and may miss a true neighbor. KNN is exact and slower.
RAG adds fresh data at question time. Fine-tuning changes the model.

Decision rules

“Results must be exact” points to KNN.
“Data changes often” points to RAG, not fine-tuning.
“Users type exact product codes and also vague descriptions” points to hybrid search.
“Source rows change” means you need an embedding maintenance method.
Match the distance metric to how the embedding model was built.

Traps

Embedding IDs and dates, which adds noise.
Chunks that are too large to be specific.
Stale embeddings after a source update.

Last-day checklist

Re-read the pairs above out loud and explain each difference in one sentence.
Write one line per table type and one per search type.
Skim your own mistakes from practice sets, not new material.
Check your booking: time, language, ID, and room or online setup.
Remember that you can mark questions, and that a break stops you from returning to earlier ones.
Sleep. A rested brain reads the one deciding word in a scenario.

More guides for DP-800

Next step

Time to test recall? Try the sample questions or revisit the exam format. Back to the DP-800 overview. For longer practice, our DP-800 practice tests include explanations for every answer.

Try the DP-800 practice exams