Assembling and Deploying Applications
Free Databricks Certified GenAI Engineer Associate practice — 6 questions on Assembling and Deploying Applications, with explanations. No sign-up. Full 12-question mixed test →
A GenAI application deployed using Mosaic AI Model Serving is experiencing intermittent timeout errors during peak usage hours. The model inference endpoint has autoscaling enabled with min_workers=1 and max_workers=10, but CloudWatch metrics show queue depth spiking to 500+ requests. The application requires sub-2-second response times. Which configuration change would BEST resolve this issue?
Request batching with proper configuration (max_batch_size=32, max_wait_time=100ms) processes multiple requests simultaneously, dramatically reducing queue depth while maintaining sub-2s latency. Setting min_workers=3 ensures baseline capacity during traffic spikes. This approach optimizes throughput without over-provisioning.
A team needs to deploy a RAG application that queries a 50GB vector index stored in Databricks Vector Search. The application must support A/B testing of two different embedding models while maintaining separate vector indices. Deployment requires rollback capability within 30 seconds if error rates exceed 2%. Which deployment architecture BEST meets these requirements?
Deploying both models to a shared GPU endpoint minimizes infrastructure costs while maintaining isolation via separate vector indices. Feature flags enable instant A/B test control and sub-30-second rollback by simply toggling the flag based on error rate monitoring. This architecture separates deployment (models) from traffic routing (feature flags), providing maximum flexibility.
An organization is deploying a multi-agent LLM application using LangGraph that orchestrates three specialized agents (SQL generation, document retrieval, and response synthesis). The application must maintain conversation state across user sessions, support concurrent users, and comply with data residency requirements that mandate all conversation history remains within the Databricks workspace. Which state management approach is MOST appropriate?
LangGraph's MemorySaver with Delta table checkpointing natively integrates with the LangGraph execution model, storing agent state and conversation history directly in Unity Catalog. This ensures data residency compliance, provides ACID guarantees, supports concurrent access through Delta's versioning, and enables automatic lifecycle management through Unity Catalog policies.
A production GenAI application uses a fine-tuned Llama 3.1 70B model served via Mosaic AI Model Serving. After deployment, the application experiences P95 latency of 8 seconds, exceeding the 3-second SLA. Token generation metrics show 12 tokens/second throughput per request. The endpoint uses g5.12xlarge instances (4x A10 GPUs) with tensor parallelism enabled. Which optimization would provide the GREATEST latency reduction?
Speculative decoding with a smaller draft model (Llama 3.1 8B) can achieve 2-3x speedup for large models by generating multiple tokens per forward pass when the draft model's predictions are accepted. This directly addresses the 12 tokens/sec bottleneck without requiring hardware changes. KV-cache optimization further reduces memory bandwidth constraints during generation.
A financial services company needs to deploy a GenAI application that generates investment advice based on customer data. The application must log all model inputs/outputs for regulatory audit, implement real-time PII detection with automatic redaction before logging, and support model version rollback within 5 minutes if compliance violations are detected. Which deployment pattern BEST addresses these requirements?
Mosaic AI Gateway provides native middleware support for custom PII detection that processes requests/responses before logging, ensuring redacted data in audit trails. Unity Catalog tables with Delta sharing enable secure compliance team access. MLflow Model Registry stage transitions (Production→Archived, Staging→Production) enable sub-5-minute rollback without redeployment, meeting all requirements in an integrated platform.
A GenAI application uses a compound AI system with three components: (1) embedding model for vector search, (2) reranker model for result refinement, and (3) LLM for response generation. The system must minimize total cost while maintaining P99 latency under 5 seconds for 1000 requests/hour with bursty traffic patterns (80% of requests in 20% of time). Which resource allocation strategy is MOST cost-effective?
Serverless endpoints for embedding/reranker handle bursty traffic efficiently with instant autoscaling and zero cost during idle periods (80% of the hour). Provisioned GPU with min_workers=2 for the LLM ensures the latency-critical component has baseline capacity during bursts while avoiding cold-start delays that would violate P99<5s. This right-sizes infrastructure for actual usage patterns.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$199.99 $66.99 with code FREETEST33 — valid through August 23.
Get my $66.99 deal →