Skip to content

Lesson 4 — Architecture: ETL/ELT, Lakes, and Service Layers

ETL vs ELT: when does integration happen?

ETL: Extract → Transform/Integrate → Load. Traditional warehouse logic often imposes the integrated structure before or during loading.

ELT: Extract → Load/Ingest → Transform/Integrate for use. Big Data environments may land varied source data first because future analytical uses are not fully known.

ELT does not mean “never transform.” It moves transformation/integration later, making Metadata, inventory, quality, and governance even more important.

Data lake vs data swamp

A lake can legitimately hold raw or minimally transformed data for exploration, history, archive, or streams. Rawness is not the failure.

A lake becomes a swamp when users lose: - inventory/discoverability; - source/provenance; - definitions/structure; - lineage; - quality context; - relationships/keys; - ownership/intended use.

Memory rule:

LAKE + METADATA + INVENTORY + QUALITY + LINKS = usable environment
LAKE − CONTEXT = SWAMP

Batch, speed, serving

  • Batch layer: complete processed history; more complete, higher latency.
  • Speed layer: current/real-time information; lower latency, less history/completeness.
  • Serving layer: exposes/merges the views for consumers.

The design is a business trade-off among completeness, latency, complexity, and cost. Serving is not a third independent data source; it is the consumption/interface layer over the other views.

Source: pp. 471–480, 495–500.

← Lesson 3 · Next: ML & Mining →