System Architecture & Design Flashcards

An organization deploys AI models across cloud and on-premises environments. Which architectural pattern best manages this complexity?

Answer: Hybrid ML platform with abstraction layers enabling portable model packaging (e.g., ONNX, containers)

Abstraction layers using open standards like ONNX and containerized serving runtimes make models portable across environments without re-engineering for each target infrastructure.

In an AI system processing sensitive PII for predictions, which architectural control ensures data minimization at the system boundary?

Answer: Data masking and tokenization applied before data enters the AI pipeline

Masking and tokenizing PII before it enters the pipeline ensures the AI components never process raw sensitive data, enforcing data minimization at the architectural boundary.

Which design pattern enables an AI system to continue serving predictions during a model backend outage?

Answer: Circuit breaker with a deterministic fallback model or cached predictions

A circuit breaker detects backend failures and redirects traffic to a simpler fallback model or stale cached results, maintaining degraded-but-functional service during outages.

What is concept drift, and which architectural component is responsible for detecting it in production?

Answer: A change in the statistical relationship between input features and target labels, detected by model monitoring with statistical drift tests

Concept drift occurs when the real-world relationship between inputs and outputs changes post-deployment; a model monitoring component tracks prediction distributions and applies statistical tests (PSI, KS test) to detect it.

In designing a multi-tenant AI SaaS platform, which pattern best ensures model output isolation between tenants?

Answer: Tenant-specific model fine-tuning with isolated serving endpoints per tenant

Isolated fine-tuned models per tenant ensures that one tenant's data never influences another tenant's predictions, providing both privacy and personalization guarantees.

A streaming AI pipeline using Apache Kafka experiences prediction lag during traffic spikes. Which architectural change most directly resolves this?

Answer: Implementing consumer group auto-scaling that dynamically adds inference workers based on consumer lag metrics

Auto-scaling consumers based on consumer lag allows the inference layer to match throughput to demand elastically, directly addressing the backpressure during traffic spikes.

Which architectural consideration is unique to designing AI systems compared to traditional software systems?

Answer: Managing non-deterministic behavior, model versioning, and data dependencies as first-class architectural concerns

AI systems introduce non-determinism, data-model coupling, and the need to version both code and data artifacts together, requiring architectural patterns not found in traditional software design.