System Architecture & Design Flashcards
7 cards from real CAIC practice questions. Tap to flip, then mark Knew It or Still Learning — missed cards come back until you master them.
Read the first 7 System Architecture & Design flashcards as text
A company wants to serve multiple AI use cases (recommendation, fraud detection, NLP) from a single platform. Which architectural approach best enables this?
Answer: An AI platform with shared infrastructure, common SDKs, and modular ML pipelines
A unified AI platform with shared infrastructure and tooling reduces duplication, standardizes best practices, and enables teams to focus on their specific models rather than platform concerns.
What distinguishes an online learning system from a batch retraining system in AI architecture?
Answer: Online learning updates the model incrementally with each new data point or mini-batch in production
Online learning continuously updates model parameters using a stream of incoming examples, allowing the model to adapt to distribution shifts without full retraining cycles.
Which pattern addresses the challenge of serving recommendations to millions of users with sub-100ms latency?
Answer: Two-stage retrieval: fast approximate nearest-neighbor retrieval followed by a lightweight re-ranker
Two-stage systems use efficient ANN search to narrow candidates quickly, then apply a more accurate but lightweight model to rank only those candidates, achieving both speed and quality.
In a microservices-based AI system, what is the primary risk of tight coupling between the prediction service and the data preprocessing service?
Answer: Changes to preprocessing logic can break the prediction service without clear version boundaries
Tight coupling means that changes in one service's interface or behavior directly impact dependent services, violating the independent deployability principle of microservices.
Which database type is most architecturally suited as a vector store for similarity search in a RAG (Retrieval-Augmented Generation) system?
Answer: Purpose-built vector database with ANN index support like Pinecone or Weaviate
Purpose-built vector databases use specialized ANN indexes (HNSW, IVF) that enable efficient high-dimensional similarity search, which is the core operation in RAG retrieval.
When an AI system must maintain a 99.99% SLA, which architectural pattern is essential for the model serving layer?
Answer: Active-active multi-region deployment with circuit breakers and fallback logic
A 99.99% SLA (~52 minutes downtime/year) requires active-active redundancy across failure domains plus circuit breakers to degrade gracefully rather than fail completely.
What is the primary purpose of a model registry in an MLOps architecture?
Answer: Centralizing model artifact versioning, metadata, and stage promotion across the model lifecycle
A model registry tracks model versions, associated metadata (metrics, lineage, parameters), and manages promotion stages (staging → production) to enable governed model lifecycle management.