메뉴
End-to-end ML system design for production. Use when designing ML pipelines, feature stores, model training infrastructure, or serving systems. Covers the complete lifecycle from data ingestion to model deployment and monitoring.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/majiayu000/claude-skill-registry --skill ml-system-design명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
LLM token logprobs and calibration. Per-decision confidence, ECE, Brier, reliability diagrams, low-confidence triage.
Analyze LLM token logprobs and calibration. Use for per-decision confidence, ECE, Brier scores, reliability diagrams, and low-confidence triage.
回顾最近 N 天的 Claude Code 使用记录——扫描原始会话数据，按主题分组汇总"我都做了什么"，并从个人操作系统视角输出模式、风险与增删建议。当用户说 /recap、"看看我这几天做了什么"、"回顾一下我最近的会话"、"这两天我用 claude 干了啥"、"活动回顾" 时使用。
SOC 직업 분류 기준
SKILL.md 표시 중
| name | ml-system-design |
| description | End-to-end ML system design for production. Use when designing ML pipelines, feature stores, model training infrastructure, or serving systems. Covers the complete lifecycle from data ingestion to model deployment and monitoring. |
| allowed-tools | Read, Glob, Grep |
This skill provides frameworks for designing production machine learning systems, from data pipelines to model serving.
Keywords: ML pipeline, machine learning system, feature store, model training, model serving, ML infrastructure, MLOps, A/B testing ML, feature engineering, model deployment
Use this skill when:
┌─────────────────────────────────────────────────────────────────────────┐
│ ML SYSTEM LIFECYCLE │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│ │ Data │──▶│ Feature │──▶│ Model │──▶│ Model │──▶│ Monitor│ │
│ │ Ingestion│ │ Pipeline │ │ Training │ │ Serving │ │ & Eval │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └────────┘ │
│ │ │ │ │ │ │
│ ▼ ▼ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│ │ Data │ │ Feature │ │ Model │ │ Inference│ │ Metrics│ │
│ │ Lake │ │ Store │ │ Registry │ │ Cache │ │ Store │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────────┘
| Component | Purpose | Examples |
|---|---|---|
| Data Ingestion | Collect raw data from sources | Kafka, Kinesis, Pub/Sub |
| Feature Pipeline | Transform raw data to features | Spark, Flink, dbt |
| Feature Store | Store and serve features | Feast, Tecton, Vertex AI |
| Model Training | Train and validate models | SageMaker, Vertex AI, Kubeflow |
| Model Registry | Version and track models | MLflow, Weights & Biases |
| Model Serving | Serve predictions | TensorFlow Serving, Triton, vLLM |
| Monitoring | Track model performance | Evidently, WhyLabs, Arize |
Problems without a feature store:
┌─────────────────────────────────────────────────────────────────┐
│ FEATURE STORE │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────────┐ ┌─────────────────────┐ │
│ │ OFFLINE STORE │ │ ONLINE STORE │ │
│ │ │ │ │ │
│ │ - Historical data │ │ - Low-latency │ │
│ │ - Training queries │ ────▶ │ - Point lookups │ │
│ │ - Batch features │ sync │ - Real-time serving│ │
│ │ │ │ │ │
│ │ (Data Warehouse) │ │ (Redis, DynamoDB) │ │
│ └─────────────────────┘ └─────────────────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────────┐│
│ │ FEATURE REGISTRY ││
│ │ - Feature definitions - Version control ││
│ │ - Data lineage - Access control ││
│ └─────────────────────────────────────────────────────────────┘│
└─────────────────────────────────────────────────────────────────┘
| Type | Computation | Storage | Example |
|---|---|---|---|
| Batch | Scheduled (hourly/daily) | Offline → Online | User purchase count (30 days) |
| Streaming | Real-time event processing | Direct to online | Items in cart (current) |
| On-demand | Request-time computation | Not stored | Distance to nearest store |
TRAINING (Historical):
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Historical │───▶│ Point-in-Time│───▶│ Training │
│ Events │ │ Join │ │ Dataset │
└──────────────┘ └──────────────┘ └──────────────┘
│
Uses feature
definitions
│
SERVING (Real-time): ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Online │───▶│ Same Feature │───▶│ Prediction │
│ Store │ │ Definitions │ │ Request │
└──────────────┘ └──────────────┘ └──────────────┘
┌───────────────────────────────────────────────────────────────────────┐
│ TRAINING PIPELINE │
├───────────────────────────────────────────────────────────────────────┤
│ │
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ Data │──▶│ Feature │──▶│ Model │──▶│ Model │ │
│ │ Loader │ │ Transform│ │ Train │ │ Validate │ │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ Experiment │ │ Hyperparameter│ │ Checkpoint │ │ Model │ │
│ │ Tracking │ │ Tuning │ │ Storage │ │ Registry │ │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘ │
│ │
└───────────────────────────────────────────────────────────────────────┘
| Pattern | Use Case | Tools |
|---|---|---|
| Single-node | Small datasets, quick experiments | Jupyter, local GPU |
| Distributed data-parallel | Large datasets, same model | Horovod, PyTorch DDP |
| Model-parallel | Large models that don't fit in memory | DeepSpeed, FSDP, Megatron |
| Hyperparameter tuning | Automated model optimization | Optuna, Ray Tune |
Track for reproducibility:
| What to Track | Why |
|---|---|
| Hyperparameters | Reproduce training runs |
| Metrics | Compare model performance |
| Artifacts | Model files, datasets |
| Code version | Git commit hash |
| Environment | Docker image, dependencies |
| Data version | Dataset hash or snapshot |
| Pattern | Latency | Throughput | Use Case |
|---|---|---|---|
| Online (REST/gRPC) | Low (<100ms) | Medium | Real-time predictions |
| Batch | High (hours) | Very high | Bulk scoring |
| Streaming | Medium | High | Event-driven predictions |
| Embedded | Very low | Varies | Edge/mobile inference |
┌─────────────────────────────────────────────────────────────────────┐
│ MODEL SERVING SYSTEM │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ │
│ │ Clients │ │
│ └──────┬───────┘ │
│ │ │
│ ▼ │
│ ┌──────────────┐ │
│ │ Load Balancer│ │
│ └──────┬───────┘ │
│ │ │
│ ▼ │
│ ┌──────────────────────────────────────────────────────────────┐ │
│ │ API Gateway │ │
│ │ - Authentication - Rate limiting - Request validation │ │
│ └──────────────────────────────┬───────────────────────────────┘ │
│ │ │
│ ┌───────────────────────┼───────────────────────┐ │
│ ▼ ▼ ▼ │
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ Model A │ │ Model B │ │ Model C │ │
│ │ (v1.2) │ │ (v2.0) │ │ (v1.0) │ │
│ └────────────┘ └────────────┘ └────────────┘ │
│ │ │ │ │
│ └───────────────────────┼───────────────────────┘ │
│ ▼ │
│ ┌────────────────┐ │
│ │ Feature Store │ │
│ │ (Online) │ │
│ └────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────┘
| Technique | Latency Impact | Trade-off |
|---|---|---|
| Batching | Reduces per-request | Increases latency for first request |
| Caching | 10-100x faster | May serve stale predictions |
| Quantization | 2-4x faster | Slight accuracy loss |
| Distillation | Variable | Training overhead |
| GPU inference | 10-100x faster | Cost increase |
┌─────────────────────────────────────────────────────────────────────┐
│ A/B TESTING ARCHITECTURE │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ │
│ │ Traffic │ │
│ └──────┬───────┘ │
│ │ │
│ ▼ │
│ ┌──────────────────────┐ │
│ │ Experiment Assignment │ ◀─────── Experiment Config │
│ │ - User bucketing │ - Allocation % │
│ │ - Feature flags │ - Target segments │
│ └──────────┬───────────┘ - Guardrails │
│ │ │
│ ┌────────┴────────┐ │
│ ▼ ▼ │
│ ┌────────┐ ┌────────┐ │
│ │Control │ │Treatment│ │
│ │Model A │ │Model B │ │
│ └────┬───┘ └────┬───┘ │
│ │ │ │
│ └────────┬───────┘ │
│ ▼ │
│ ┌────────────────┐ │
│ │ Metrics Logger │ │
│ └────────┬───────┘ │
│ ▼ │
│ ┌────────────────┐ │
│ │ Statistical │ ─────▶ Decision: Ship / Iterate / Kill │
│ │ Analysis │ │
│ └────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────┘
| Metric Type | Examples | Purpose |
|---|---|---|
| Model metrics | AUC, RMSE, precision/recall | Model quality |
| Business metrics | CTR, conversion, revenue | Business impact |
| Guardrail metrics | Latency, error rate, engagement | Prevent regressions |
| Segment metrics | Metrics by user segment | Detect heterogeneous effects |
| Category | Metrics | Alert Threshold |
|---|---|---|
| Data quality | Missing values, schema drift | >1% change |
| Feature drift | Distribution shift (PSI, KL) | PSI >0.2 |
| Prediction drift | Output distribution shift | Depends on use case |
| Model performance | Accuracy, AUC (when labels available) | >5% degradation |
| Operational | Latency, throughput, errors | SLO violations |
┌─────────────────────────────────────────────────────────────────────┐
│ DRIFT DETECTION PIPELINE │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ Training Data Production Data │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Reference │ │ Current │ │
│ │ Distribution │ │ Distribution │ │
│ └──────┬───────┘ └──────┬───────┘ │
│ │ │ │
│ └──────────────┬──────────────┘ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ Statistical Test │ │
│ │ - PSI (Population Stability Index) │
│ │ - KS Test │
│ │ - Chi-squared │
│ └────────┬─────────┘ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ Drift Score │ │
│ └────────┬─────────┘ │
│ │ │
│ ┌───────────┼───────────┐ │
│ ▼ ▼ ▼ │
│ No Drift Warning Critical │
│ (< 0.1) (0.1-0.2) (> 0.2) │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ Continue Investigate Retrain │
│ │
└─────────────────────────────────────────────────────────────────────┘
Components needed:
- Candidate Generation (retrieve 100s-1000s)
- Ranking Model (score and sort)
- Feature Store (user features, item features)
- Real-time personalization (recent behavior)
- A/B testing infrastructure
Components needed:
- Real-time feature computation
- Low-latency model serving (<50ms)
- High recall focus (can't miss fraud)
- Explainability for compliance
- Human-in-the-loop review
- Feedback loop for labels
Components needed:
- Two-stage ranking (retrieval + ranking)
- Feature store for query/document features
- Low latency (<200ms end-to-end)
- Learning to rank models
- Click-through rate prediction
- A/B testing with interleaving
Training time estimation:
- Dataset size: 100M examples
- Model: Transformer (100M params)
- GPU: A100 (80GB, 312 TFLOPS)
- Batch size: 32
- Training steps: Dataset / batch = 3.1M steps
- Time per step: ~100ms
- Total time: ~86 hours single GPU
- With 8 GPUs (data parallel): ~11 hours
Inference estimation:
- QPS: 10,000
- Model latency: 20ms
- Batch size: 1 (real-time)
- GPU utilization: 50% (latency constraint)
- Requests per GPU/sec: 25
- GPUs needed: 10,000 / 25 = 400 GPUs
- With batching (batch 8): 100 GPUs (4x reduction)
llm-serving-patterns - LLM-specific serving and optimizationrag-architecture - Retrieval-Augmented Generation patternsvector-databases - Vector search and embeddingsml-inference-optimization - Latency and cost optimizationestimation-techniques - Back-of-envelope calculationsquality-attributes-taxonomy - NFR definitions/sd:ml-pipeline <problem> - Design ML system interactively/sd:estimate <scenario> - Capacity calculationsml-systems-designer - Design ML architecturesml-interviewer - Mock ML system design interviewsDate: 2025-12-26 Model: claude-opus-4-5-20251101