Skip to content

Azure Cosmos DB Developer Specialty (DP-420) - Practice Questions

15 questions for DP-420 prep, weighted toward data modeling and partitioning, which is where most of the exam's difficulty and most real-world Cosmos DB failure lives.

Cert page: exams/azure/dp-420/

Question 1

Scenario: A container stores telemetry for 10,000 devices. Queries are almost always for one device over a time range.

A. Partition key /date B. Partition key /deviceId C. Partition key /id D. No partition key

Answer **Correct: B** **Why:** A good partition key has high cardinality and matches the dominant query filter, which keeps queries single-partition and cheap. `/date` creates a hot partition because all of today's writes land in one place. `/id` gives perfect distribution but makes every device query fan out across all partitions.

Question 2

Scenario: A query returns the right data but consumes 500 RUs instead of the expected 3.

A. The container is under-provisioned B. The query is cross-partition; add the partition key to the filter or reshape the model C. The consistency level is too weak D. The SDK version is old

Answer **Correct: B** **Why:** Without the partition key in the WHERE clause the query fans out to every physical partition and the cost scales with partition count. Check the `x-ms-request-charge` header and the query metrics. Provisioning more RUs makes the same inefficiency affordable rather than fixing it.

Question 3

Scenario: Which consistency level guarantees monotonic reads within a session and is the default?

A. Strong B. Bounded staleness C. Session D. Eventual

Answer **Correct: C** **Why:** Session consistency guarantees read-your-writes and monotonic reads for a client holding the session token, at much lower cost and latency than Strong. Strong requires a single write region or region-bounded configuration and costs roughly double the RUs for reads. Eventual is cheapest with no ordering guarantee.

Question 4

Scenario: An account has multiple write regions and two users update the same document in different regions simultaneously.

A. One write fails B. Conflict resolution applies: last-writer-wins by a defined property, or a custom stored procedure resolves it C. The document is deleted D. The account goes read-only

Answer **Correct: B** **Why:** Multi-region writes accept both and then resolve. The default policy is last-writer-wins on `_ts`, which you can point at your own numeric property, or you can register a merge procedure and inspect the conflicts feed. If silent overwrite is unacceptable, custom resolution is mandatory, not optional.

Question 5

Scenario: Documents contain a large attachments array that is rarely read but always retrieved with the item.

A. Keep it embedded for atomicity B. Reference the attachments as separate items in the same logical partition C. Move to a different database D. Increase the RU budget

Answer **Correct: B** **Why:** Embed when data is read together, is bounded, and changes together. An unbounded, rarely read array inflates item size, which raises the RU cost of every read and write of the parent. Keeping the referenced items in the same logical partition preserves transactional batch capability.

Question 6

Scenario: Only three properties are ever filtered on, but writes cost far more than expected.

A. Lower the consistency level B. Customize the indexing policy to exclude paths that are never queried C. Increase throughput D. Add a composite index for every path

Answer **Correct: B** **Why:** Cosmos DB indexes every path by default, and every indexed path adds RU cost to writes. Excluding unused paths (or switching to an include-list) is one of the highest-leverage cost optimizations available. Adding more indexes would increase write cost further.

Question 7

Scenario: A query does ORDER BY category ASC, price DESC and fails or is slow.

A. Add a composite index on those two paths in that order and direction B. Add a range index on price C. Use a spatial index D. Increase MaxItemCount

Answer **Correct: A** **Why:** Multi-property ORDER BY requires a composite index matching the property order and sort directions (and their exact reverse). Single-path range indexes cover single-property ordering only. This is a frequent DP-420 question because the requirement is not obvious until the query fails.

Question 8

Scenario: Downstream systems must react to every item change in a container, in order, per partition.

A. Poll the container on a timer B. Change feed, consumed via the change feed processor or an Azure Functions trigger C. A stored procedure D. Enable analytical store

Answer **Correct: B** **Why:** The change feed gives a durable, ordered-per-partition-key log of creates and updates, with lease-based checkpointing so consumers resume where they left off. Note the standard mode does not surface deletes, so a soft-delete flag with TTL is the usual pattern. Polling is expensive and misses intermediate states.

Question 9

Scenario: Analytics teams want to run large aggregations without affecting the transactional workload's RUs.

A. Provision more RUs B. Enable the analytical store and query it with Azure Synapse Link or Fabric C. Run the queries at night D. Add a read region

Answer **Correct: B** **Why:** The analytical store is a column-oriented copy maintained automatically with no RU cost on the transactional store, which is the point of Synapse Link. Adding a read region still consumes RUs and replicates the row store. Shifting the time of day does not remove the contention.

Question 10

Scenario: Old telemetry should be removed automatically after 90 days.

A. A scheduled delete job B. Set container TTL to 7,776,000 seconds, optionally overriding per item C. Drop and recreate the container D. Use the change feed

Answer **Correct: B** **Why:** TTL expiry is performed by the engine using leftover throughput and does not consume your provisioned RUs the way a delete job would. Container-level TTL sets the default and an item's own `ttl` property overrides it, including `-1` to never expire.

Question 11

Scenario: A transactional batch must update three documents atomically.

A. They must share the same logical partition key B. They can be in any partition C. They must be in different containers D. Use a distributed transaction coordinator

Answer **Correct: A** **Why:** Transactional batch and stored procedure transactions are scoped to a single logical partition. This constraint should drive your model: if two entities must change atomically, give them the same partition key. There is no cross-partition transaction.

Question 12

Scenario: A workload is spiky, with 10x bursts a few times a day and near-zero baseline.

A. Provisioned throughput sized for the peak B. Autoscale throughput, which scales between 10% and 100% of the maximum C. Serverless for a high-volume production workload D. Manual scaling scripts

Answer **Correct: B** **Why:** Autoscale bills for the highest RU/s used each hour within a 10x band, which fits exactly this shape. Provisioning for the peak pays for it around the clock. Serverless suits low-volume and development workloads and has lower ceilings than a bursty production system usually needs.

Question 13

Scenario: The SDK reports HTTP 429 responses under load.

A. The request was unauthorized B. Rate limiting: requests exceeded provisioned RU/s, so honor the retry-after header, tune the query, or raise throughput C. The item was not found D. Consistency conflict

Answer **Correct: B** **Why:** 429 means throttled. The SDKs retry automatically using `x-ms-retry-after-ms`, so occasional 429s are normal and not an outage. Persistent throttling means either a hot partition, an expensive query, or genuinely insufficient throughput, and the fix differs for each.

Question 14

Scenario: A globally distributed app must read with the lowest latency from each region.

A. Add read regions and configure the client's preferred regions list B. Use Strong consistency everywhere C. Increase RU/s D. Use a single region with a CDN

Answer **Correct: A** **Why:** Replicas serve reads locally, and the SDK's preferred regions list determines which replica a client uses and the failover order. Strong consistency across many regions increases latency because writes must be acknowledged more broadly. A CDN cannot cache authenticated document reads.

Question 15

Scenario: Access must be granted to an application without using the account primary key.

A. Microsoft Entra ID with the Cosmos DB data plane RBAC roles, assigned to a managed identity B. A resource token broker only C. A read-only key D. IP firewall rules

Answer **Correct: A** **Why:** Data plane RBAC lets you grant read or write on specific databases and containers to an Entra identity, with no shared secret to rotate. Read-only keys are still account-wide shared secrets. Firewall rules restrict where a key can be used, which is a useful layer but not authentication.

Where to go deeper