Free DP-420 practice test: 30 questions with answers

Bounded staleness lets the team configure a maximum lag in time or versions, which gives a predictable freshness guarantee with lower latency than strong. Eventual has no bound, and consistent prefix guarantees order but not a time limit.

Session is the default and guarantees read-your-own-writes for a client session. Strong and bounded staleness are stricter options, and eventual is weaker than the default.

A point read addresses one item directly, which costs far fewer request units than queries that fan out or scan. Cross-partition queries, bulk deletes, and full scans all consume substantially more throughput.

High cardinality and alignment with common filters both help distribute load and keep queries inside one logical partition. Low-cardinality keys create hot partitions, and cross-partition queries cost more than targeted ones.

Hierarchical partition keys let a container partition on up to three levels, such as tenant then user, which supports targeted queries at each level. Azure Cosmos DB does not use secondary indexes as a substitute for partitioning.

A synthetic key combines existing fields into a higher-cardinality value that spreads writes while keeping the source fields intact. A single fixed value or a per-country duplication does not improve distribution in the same way.

Embedding pays off when data is read together and the bounded size keeps the item within limits. Frequent independent updates and shared children favor separate items or references.

The pull model lets the application request changes in batches using continuation tokens that it manages itself. The processor model handles distributed leases automatically, and triggers and nightly exports serve different purposes.

The lease container records which instance owns each partition range and where it stopped, which lets work be shared and resumed after failures. It does not hold application data, backups, or caches.

Transactional batch executes multiple operations against one logical partition atomically. Separate writes can leave partial results, and bulk ingestion tolerates partial success by design.

Managed identities with Entra ID remove long-lived secrets from application code and allow access to be governed by role assignments. Embedded or shared keys create exposure that identity-based access avoids.

Data-plane role-based access control assigns read-only data roles that can be scoped, which enforces least privilege. Full keys and subscription ownership grant far more than the service needs.

A private endpoint gives the account a private IP address inside the virtual network, and restricting public access enforces the policy. Public endpoints and open ports do not meet this requirement.

Continuous backup with point-in-time restore lets the team recover to a chosen moment within the retention window, which covers logical corruption. Replicas copy bad writes too, so they do not protect against that failure.

Multi-region writes allow each region to accept writes, and automatic failover keeps service available when a region fails. A single write region without failover leaves the service exposed to regional outages.

Filtering on the partition key limits the query to one logical partition, which avoids fan-out and reduces cost. More threads add load without reducing work, and disabling indexing breaks efficient filtering.

The service returns a retry-after hint, and honoring it with backoff avoids hammering the account. Throughput can be raised or autoscaled to meet sustained demand. Immediate tight retries and silent drops both make the problem worse or hide it.

Azure Monitor metric alerts evaluate platform metrics such as throttled request counts against thresholds and notify the team. Manual reviews, commit hooks, and DNS checks do not evaluate account metrics.

The IP firewall restricts which public addresses may connect to the account. Throughput, consistency, and region count do not control network access.

Azure Cosmos DB replicates data to multiple regions, and the SDK can prefer the closest read region to lower latency. Client-side duplication and deleting data do not provide that, and a VPN to the primary region adds latency.

The vector embedding policy and matching vector index are part of the container definition, which the vector search depends on. The other options are unrelated to the container configuration.

DiskANN is a graph-based approximate nearest-neighbor index that is designed for large vector collections. Hash and B-tree indexes do not support vector similarity search, and client-side scans do not scale.

FullTextScore returns a BM25-based relevance score for full-text search, which can be used for ranking. VectorDistance measures similarity between vectors, ST_Distance computes geospatial distance, and COUNT aggregates rows.

Reciprocal rank fusion merges rankings from different retrieval methods without requiring their scores to be comparable, which is why it is common in hybrid search. Averaging unnormalized scores mixes incompatible scales.

RAG retrieves relevant content and supplies it to a language model so that the answer is grounded in that content. Copying training data into sessions and firewall rules are not part of the RAG pattern.

Partitioning by session or user keeps related items together, and time-to-live removes stale conversations automatically. Browser memory is not durable, a shared item creates contention, and local log files are hard to query and expire.

The distance function, such as cosine, euclidean, or dot product, is set in the vector embedding policy when the container is defined. The other settings are unrelated to similarity ranking.

Vector dimensions must match the policy the container was created with, so mismatched vectors are rejected. Silent truncation or automatic policy changes are not how the service behaves, and dimension mismatch is a configuration error to fix.

The full-text policy's default language determines how terms are tokenized and stemmed for search. TTL, connection locales, and partition paths do not affect language processing.

A database gives durable, queryable storage that survives individual sessions and scales beyond a prompt window. Prompt context is temporary, and databases do not train models automatically.