AI 클라우드 아키텍처

The design must meet the task’s reliability, data, latency, and cost constraints. A powerful accelerator is only one component of that design.

주요 시사점

• Separate workloads by their operating needs.
• Enforce data and permission boundaries.
• Design capacity, retries, and rollback together.

Separate interactive and background workloads where their requirements differ. A user waiting for an answer needs bounded response time, while batch processing can use queues and longer-running jobs. Make queue status and retry behavior observable. Define data boundaries and access roles. Documents, embeddings, model artifacts, and logs may have different retention and permission requirements. Keep credentials in appropriate secret management and avoid assuming that network location alone establishes authorization. Plan for capacity changes and dependency failures. Autoscaling can take time, model loading can be expensive, and a provider can impose rate limits. Use admission controls, backpressure, bounded retries, and clear unavailable states to prevent one overloaded dependency from overwhelming the whole service. Version the deployment and test recovery. Check compatible model and preprocessing versions, data migrations, and rollback procedures. Measure cost per useful completed task, including storage, transfer, failed attempts, and idle resources. A low price for one API call may hide a more expensive overall workflow.

1. Imagine 100 application workers calling one rate-limited model endpoint. Each failed request is retried immediately five times.
2. The extra attempts increase load without adding endpoint capacity.
3. Apply a bounded retry policy that respects provider backoff, limit concurrent requests, and show the queue or unavailable state to users.

This constructed example explains how architecture can prevent an overload from spreading.