How do I approach ML System Design interview questions?

ML System Design questions require understanding of core concepts and practice. PracHub provides solutions with explanations to help you master ml system design interviews.

What difficulty level is this interview question?

This is a hard difficulty ML System Design question, commonly asked during Technical Screen rounds at Perplexity AI.

What role is this question designed for?

This question is commonly asked for Software Engineer candidates at Perplexity AI during technical interviews.

Implement resilient LLM provider pool | Perplexity AI Interview Question

Quick Overview

This English-language summary evaluates a candidate's ability to design a resilient, multi-provider LLM client library encompassing provider management, governance (rate limiting and quotas), resilience mechanisms (health checks, timeouts, circuit breakers), routing (cost- and latency-aware load balancing, retries and fallbacks), observability, security, and concurrency. Commonly asked in the ML System Design domain to assess architectural decision-making and trade-off reasoning for availability, cost, latency, and operational robustness, the question targets practical application and architectural-level thinking rather than purely conceptual or low-level implementation details.

System Design Task: Resilient Multi‑Provider LLM Client Library

Context

You are designing a client library used by backend services to call external Large Language Model (LLM) providers (e.g., OpenAI, Anthropic, etc.). The library must route requests across multiple providers to maximize availability, control cost, and meet latency SLOs.

Requirements

Provider management
- Provider registration/unregistration
- Capability mapping: available models, max context length, supported features (streaming, function calling, JSON mode, etc.)
Governance
- Per‑provider rate limiting (including per‑API key where applicable)
- Quotas (per provider, per model, per tenant)
Resilience
- Health checks (active and passive)
- Timeouts (connect, request, total)
- Circuit breakers (fail‑fast, half‑open probes)
Routing
- Cost‑ and latency‑aware load balancing
- Retry and fallback across providers when degraded or down
Observability
- Metrics, structured logs, and distributed tracing
Security
- Secure key management (storage, rotation, scoping, redaction)
Concurrency
- Thread‑safe, high‑throughput, supports parallel requests and streaming

Deliverables

Describe interfaces and data structures
Explain request routing, failure handling, and concurrency strategy
Include assumptions where needed

Quick Overview

Requirements

Provider management

Provider registration/unregistration
Capability mapping: available models, max context length, supported features (streaming, function calling, JSON mode, etc.)

Governance

Per‑provider rate limiting (including per‑API key where applicable)
Quotas (per provider, per model, per tenant)

Resilience

Health checks (active and passive)
Timeouts (connect, request, total)
Circuit breakers (fail‑fast, half‑open probes)

Routing

Cost‑ and latency‑aware load balancing
Retry and fallback across providers when degraded or down

Observability

Metrics, structured logs, and distributed tracing

Security

Secure key management (storage, rotation, scoping, redaction)

Concurrency

Thread‑safe, high‑throughput, supports parallel requests and streaming

Implement resilient LLM provider pool

Quick Overview

System Design Task: Resilient Multi‑Provider LLM Client Library

Context

Requirements

Deliverables

Solution

Submit Your Answer

Implement resilient LLM provider pool

Quick Overview

System Design Task: Resilient Multi‑Provider LLM Client Library

Context

Requirements

Deliverables

Solution

Submit Your Answer