Same rule as SAA-C03: connect service behavior to pipeline design, not to a memorized definition.

Streams gives you custom consumers, configurable retention, and application-controlled processing. Firehose trades that control for simple, managed delivery to a supported destination. Trap: managed delivery is not the same capability as arbitrary stream processing — don’t reach for Firehose when you need to process records yourself.

Glue is managed, serverless data integration and ETL. EMR gives you broader cluster-based control over big-data frameworks. Trap: pick based on transformation complexity and how much operational control you actually need, not on which tool you’re more familiar with.

Athena queries data where it already sits using SQL — no loading required. Redshift is a full data-warehouse platform built for modeled, repeated analytical workloads. Trap: a query engine and a warehouse have very different performance and cost shapes; “SQL” alone doesn’t tell you which one fits.

Object storage is a durable landing and sharing layer. Organize it into raw, validated, and curated zones with real metadata and access controls. Trap: storage alone gives you none of catalog, quality, or governance — those are separate jobs you still have to do.

Batch processes bounded data on a schedule. Streaming handles records as they arrive. Trap: low latency is a requirement you test for — it’s not automatic justification to stream everything just because streaming exists.

ETL transforms data before it loads into the target. ELT loads first and lets the target or query layer do the transforming. Trap: the right choice depends on data quality needs, governance, cost, and what the target system can actually process — it’s not a style preference.

Provides technical metadata so consumers and services can discover schemas and locations. Trap: a catalog entry doesn’t guarantee quality, correct permissions, or business meaning — it just means the data is findable.

Adds governed access patterns around data-lake resources. Trap: governance is broader than a single access policy — identity, encryption, catalog, lineage, and sharing all have to work together.

Choose migration tooling based on source, target, change-capture needs, downtime tolerance, and validation requirements. Trap: copying bytes across successfully is not the same as validating that schemas, permissions, and data quality all survived the trip.

Lambda handles a suitable event-driven unit of work. Step Functions coordinates stateful workflow steps and error paths across multiple units. Trap: burying orchestration logic inside one giant function makes it much harder to observe and retry safely — that’s exactly what Step Functions exists to fix.

Glue Data Catalog is primarily about organizing technical metadata and schemas. Lake Formation adds governance-oriented access controls on top of that. Trap: discoverability and governed authorization are two different problems — solving one doesn’t solve the other.

Follow the Data: A Conceptual End-to-End Pipeline

DEA-C01 Decision Signals

• Near real-time — Evaluate streaming ingestion and processing, but confirm latency, ordering, replay, and consumer requirements before committing.
• Data quality — Look for validation, completeness, consistency, schema handling, and observable quality checks — not just a passing job.
• Governed sharing — Consider identity, authorization, catalog, lineage, privacy, sovereignty, and genuinely controlled data access.
• Cost and performance — Inspect partitioning, file format, compression, query pattern, compute duration, storage lifecycle, and data movement together.