Resilience Beyond the Code

In the previous lessons, we learned how to build and monitor. But as an AWS AI Practitioner, you must also be aware of the "Infrastructure Limits" that can kill your project before it even starts.

If you don't plan for Quotas and Scalability, your AI will stop working exactly when it's most popular.

1. Service Quotas (The Speed Limit)

By default, AWS puts "Limits" on every account to prevent accidental spending.

• Example: By default, you might only be allowed to process 50,000 tokens per minute in Amazon Bedrock.
• If you suddenly have a viral marketing campaign that tries to process 1,000,000 tokens per minute, your users will get a "Throttling Error" (Error 429).

The Solution: You must use the AWS Service Quotas console to "Ask" AWS for a higher limit weeks before your big launch.

2. Regional Availability

Not every AI service is available in every AWS region.

• You might have your data in London, but the specific Claude 3.5 Opus model is only available in US-East (N. Virginia).
• Moving data between regions (Cross-region) increases Latency and Cost.

The Strategy: Build your AI pipeline in the "Primary" regions (US-East-1, US-West-2, EU-West-1) if you need access to the newest models first.

3. The "Managed Service" Risk

When you use a managed service like Bedrock, you are at the mercy of the model provider.

• If Anthropic updates their model from v1 to v2, your prompt might stop working as expected.
• The Defense: Always use Pinned Versions (e.g., set your code to use claude-3-opus-20240229 specifically) so that your app doesn't break when a new version is released.

4. Visualizing the Risks

graph TD
    A[Public User] --> B[Your Application]
    B -->|Request| C{AWS Infrastructure}
    
    subgraph Potential_Failure_Points
    C -->|Limit Exceeded| D[Throttling Error]
    C -->|Region Down| E[Service Outage]
    C -->|Deprecated Model| F[Model Execution Error]
    end

5. Summary: High Availability is a Choice

A "Hobbyist" builds for the "Happy Path" (when everything works). A Practitioner builds for the "Failure Path."

• You should have an "Emergency" plan for what your app does if Bedrock is unavailable (e.g., Fail over to a simple rule-based response).

Exercise: Identify the Risk Mitigation

A company is planning a "Black Friday" sale. They expect 10x the normal traffic to their AI recommendation engine. Which AWS action should they take one week before the sale?

• A. Re-train the model.
• B. Check and increase their AWS Service Quotas.
• C. Change the IAM password.
• D. Enable AWS Artifact.

The Answer is B! You must ensure your account has a high enough "Speed Limit" (Quota) to handle the spikes in traffic during a promotion.

Knowledge Check

What's Next?

We know the risks; now let's scale. In our final lesson of Module 14, we look at Scaling AI workloads responsibly.