Pipes and signals: Step Functions, EventBridge, SQS/SNS, CloudWatch
Lesson 4 of 4 in The Supporting Cast: Data, Compute, and Pipes.
ML systems are mostly plumbing, and four pipes carry nearly all of it. Step Functions is the orchestrator: visual state machines with retries, branching, and error handling, with direct service integrations into SageMaker and Bedrock — the exam's default for "multi-step ML workflow". EventBridge is the trigger: rules match events ("drift alarm fired") and schedules ("every Sunday at 02:00"), then start targets like a state machine. SQS vs SNS is a classic pair: a queue decouples one producer from one consumer pool that pulls at its own pace; a topic broadcasts one message to many subscribers at once. And MWAA, in one line: managed Apache Airflow, for teams whose pipelines are already DAGs.
| Service | The question it answers | Classic exam phrase |
|---|---|---|
CloudWatch | What is happening? (metrics, logs, alarms, dashboards) | "alert when endpoint latency exceeds a threshold" |
CloudTrail | Who did what, when? (an audit history of API calls) | "audit who accessed the model artifact bucket" |
X-Ray | Where did the time go? (distributed traces across services) | "trace a request across API Gateway → Lambda → endpoint" |
IAM — one line
Who (and what) may perform which actions on which resources: roles for services, least privilege for everything. Every SageMaker job and Bedrock call runs under a role.
KMS — one line
The encryption keys. S3 objects, training volumes, and model artifacts encrypt under KMS keys; "encrypt with keys we control" means KMS.
Secrets Manager — one line
Stores and rotates credentials — database passwords, API keys — so they never live in code, notebooks, or environment files.
Macie — one line
The ML-powered scanner that finds sensitive data (PII) sitting in S3. Point it at the training lake before the regulators do.
Inspector — one line
Automated vulnerability scanning for EC2 instances, ECR container images, and Lambda functions — your training and serving images included.
X-Ray — one line
Distributed tracing: follow one request across every hop and see which one ate the latency.
Key terms: AWS Step Functions, Amazon EventBridge
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.