Design an image detection system
Overview
This question evaluates a candidate's competency in ML system design, covering architectural reasoning for scalable image ingestion, model serving, data and model versioning, preprocessing and batching strategies, and operational concerns such as monitoring, cost control, privacy, and failure-mode handling.
Datadog
Sep 6, 2025, 12:00 AM
Software Engineer
Technical Screen
ML System Design
0
0
System Design: End-to-End Image Object-Detection Platform
Context
Design a production system that accepts user-submitted images, detects objects, and returns results via APIs. The system should be scalable, observable, cost-aware, and safe to operate.
Requirements to Specify
-
Functional
- Image ingestion (sync and async), object detection, result retrieval.
- API definitions and idempotency.
- Error handling and retries.
-
Non-Functional
- Accuracy (e.g., mAP targets), latency budgets (P50/P95), throughput, availability, durability, and SLOs.
- Cost targets and multi-tenancy considerations.
Architecture to Propose
- High-level components: API gateway, ingestion, storage (raw and derived), preprocessing, model serving, asynchronous workers/queues, result storage, and an admin console.
- Data and model versioning: training/eval datasets, model registry, schema evolution.
- Batching strategy, GPU utilization, autoscaling strategies, and caching layers.
ML Considerations
- Model choices: single-stage vs. two-stage detectors; when to use each.
- Training and labeling pipeline: data sources, labeling strategy, active learning, augmentation.
- Evaluation metrics: mAP, precision/recall, latency/throughput.
- Online/offline monitoring and A/B testing plan.
Operations and Safety
- Failure modes and backpressure handling.
- Privacy, security, and compliance controls.
- Cost controls and capacity planning.
- Deployment: blue/green, canary, and rollback strategy.
Solution
Login required
Show