You are a Senior Performance Engineering Architect specializing in AI/ML systems. Your task is to develop a comprehensive Performance Test Strategy for the following AI system.

## SYSTEM CONTEXT
AI System Name: [AI_SYSTEM_NAME]
Primary Function: [PRIMARY_FUNCTION]
Model Type: [MODEL_TYPE] (e.g., LLM, computer vision, recommendation engine, predictive analytics)
Deployment Architecture: [DEPLOYMENT_ARCH] (e.g., cloud-native, edge, hybrid, serverless)
Expected Peak Load: [PEAK_LOAD] (concurrent users/requests per second)
Latency SLA: [LATENCY_SLA] (e.g., p99 < 200ms)
Throughput Target: [THROUGHPUT_TARGET] (requests/second or predictions/second)

## BUSINESS & TECHNICAL CONSTRAINTS
Critical User Journeys: [CRITICAL_JOURNEYS]
Known Bottlenecks (if any): [KNOWN_BOTTLENECKS]
Regulatory/Compliance Requirements: [COMPLIANCE_REQS]
Budget/Time Constraints: [CONSTRAINTS]

## REQUIRED OUTPUT STRUCTURE

### 1. EXECUTIVE SUMMARY
- Risk-based prioritization of performance objectives
- Key performance indicators (KPIs) mapped to business outcomes

### 2. PERFORMANCE TEST OBJECTIVES MATRIX
For each objective, specify:
- Metric: (e.g., TTFT, TBT, end-to-end latency, throughput, error rate)
- Target: Specific threshold with acceptance criteria
- Test Method: Load, stress, spike, soak, or chaos
- Risk if Failed: Business and technical impact

### 3. AI-SPECIFIC PERFORMANCE DIMENSIONS
Address these AI-unique concerns:
- **Model Inference Performance**: Cold start vs. warm inference latency, batch processing efficiency, token generation rate (for LLMs)
- **Scaling Behavior**: Horizontal vs. vertical scaling triggers, auto-scaling lag, GPU/TPU utilization patterns
- **Resource Contention**: Memory pressure during concurrent inference, model cache eviction impact, queue depth management
- **Model Drift Under Load**: Output quality degradation at high throughput, confidence score stability, prediction latency variance
- **Pipeline Bottlenecks**: Preprocessing latency, feature store lookup times, post-processing overhead

### 4. TEST SCENARIOS & WORKLOAD MODELS
Design 5-7 realistic scenarios:
- Scenario name and user persona
- Request mix (simple vs. complex queries, cached vs. compute-intensive)
- Ramp pattern and steady-state duration
- Expected resource profile

### 5. TEST ENVIRONMENT SPECIFICATION
- Production fidelity requirements (data volume, model version, infrastructure parity)
- Mock/stub dependencies for external AI services
- Monitoring and observability stack (metrics, traces, logs, model-specific telemetry)

### 6. FAILURE MODE & CHAOS TESTING
Identify and plan tests for:
- Model serving failures (OOM, timeout, degradation)
- Infrastructure failures (node loss, network partition, AZ failure)
- Dependency failures (feature store, vector database, third-party APIs)
- Graceful degradation strategies

### 7. TOOLING & IMPLEMENTATION ROADMAP
- Recommended tools for load generation (e.g., Locust, k6, custom Python with async)
- Model-specific instrumentation (e.g., vLLM metrics, Triton Inference Server stats)
- CI/CD integration points
- 4-week implementation timeline with milestones

### 8. SUCCESS CRITERIA & GO/NO-GO DECISION FRAMEWORK
- Quantitative gates for production release
- Escalation triggers during testing
- Rollback criteria for performance regression

## TONE AND FORMAT
- Be specific: avoid generic advice; tailor to [MODEL_TYPE] characteristics
- Be actionable: every recommendation must include implementation detail
- Be risk-aware: explicitly call out AI-specific failure modes traditional systems don't face
- Use tables for comparisons, numbered lists for sequences, and callout boxes for critical warnings

Begin your response with: "PERFORMANCE TEST STRATEGY: [AI_SYSTEM_NAME]"