AI-200 cheat sheet
Last-minute reference for AI-200. The data domain is the largest; the other three are roughly a quarter each.
Exam facts
| Duration | 120 minutes |
| Passing score | 700 / 1000 |
| Price | Set per country or region; shown at booking |
| Replaces | AZ-204 |
| Language focus | Python |
Skill areas
| Area | Weight |
|---|---|
| Develop AI solutions by using Azure data management services | 25–30% |
| Develop containerized solutions on Azure | 20–25% |
| Connect to and consume Azure services | 20–25% |
| Secure, monitor, and troubleshoot Azure solutions | 20–25% |
Requirement to service
| Requirement | Answer |
|---|---|
| Rebuild the image when the base image is patched | ACR task with base image trigger |
| Run exactly the tested image | Reference by digest |
| Secret in a container on App Service | Key Vault reference + managed identity |
| Canary release, instant rollback | Container Apps multiple revision mode |
| Scale a worker on queue length, to zero | KEDA Service Bus rule, min replicas 0 |
| Existing Kubernetes manifests | AKS |
| Read your own writes, cheapest | Cosmos DB session consistency |
| React to new documents in Cosmos DB | Change feed processor |
| Vector search with SQL filters and joins | PostgreSQL + pgvector |
| Semantic cache, lowest latency | Azure Managed Redis vector index |
| One consumer per message, retries | Service Bus queue |
| Every subscriber gets a copy, with filters | Service Bus topic + subscriptions |
| Notify many handlers of an event | Event Grid |
| Ordered processing per entity | Service Bus sessions |
| No credential stored anywhere | Managed identity + DefaultAzureCredential |
| Same key, different value per environment | App Configuration labels |
| Follow one request across services | OpenTelemetry context propagation |
Vector indexes
| Service | Options |
|---|---|
| Cosmos DB | flat (exact, small dimensions), quantizedFlat, diskANN (large scale) |
| PostgreSQL | IVFFlat (needs data first), HNSW (works on empty tables, higher recall) |
| Redis | Search index with a vector field |
pgvector operators: <-> Euclidean, <=> cosine, <#> negative inner product. The operator must match the index’s operator class.
KQL in one line
requests | where timestamp > ago(1d) | summarize percentile(duration, 95) by bin(timestamp, 5m) | render timechart
Join requests to dependencies on operation_Id.
Traps
- Redis is a cache, not a system of record. Durable data goes in Cosmos DB or PostgreSQL.
- CPU scale rules cannot scale to zero. Use an event-driven KEDA rule.
- Default indexing indexes everything. Exclude unused paths, including the embedding path from the range index.
- Messages that “disappear” are in the dead-letter queue. Something has to read it.
- Lost Event Grid events mean dead-lettering was not configured.
- Reading a secret by version ignores rotation. Read by name.
- AZ-204 material covers topics that are gone and misses the vector content.
Night-before checklist
- Name the four areas and which is largest
- One sentence each on when to pick Cosmos DB, PostgreSQL or Redis
- Service Bus versus Event Grid, cold
- Check ID and proctoring rules; see exam day
Take the 20-question practice test and check your weakest area.