AWS Lambda
Serverless functions — the glue of AI pipelines.
❓ What is it?
Event-driven compute that runs code without servers, billed per millisecond, triggered by S3 uploads, API Gateway requests, queues, and schedules.
💡 Why does it exist?
AI architectures are mostly plumbing between managed services: new file → extract → analyse → store. Lambda is that plumbing with zero idle cost and automatic scaling.
⏱️ When should you use it?
Use it to orchestrate AI service calls (invoke Bedrock, chain Textract→Comprehend), fulfil Lex intents, and process events; avoid it for long training jobs (15-minute limit) — that is SageMaker's job.
🗺️ Where does it fit?
Between every pair of services in a serverless AI pipeline: triggers in, SDK calls out, results to storage or queues; Bedrock Agents call Lambda as their action-group executor.
🔌 How do you integrate it?
Write a handler, grant its role least-privilege access to the AI services it calls, attach a trigger, and set memory/timeout to fit the call pattern.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Bedrock Agents execute business actions THROUGH Lambda action groups — a fixture in agentic-architecture questions.
- 15-minute maximum runtime is why Lambda is orchestration, never model training.