Independent guide

Data Warehouse Architecture: Understanding Each Layer

Data warehouse architecture determines how data flows from source systems into the analytical layer your team queries every day. The architecture is not a single component — it is a stack of layers that each serve a distinct purpose: ingestion, staging, storage, and presentation. Getting these layers right affects query speed, cost, and how easily the system adapts as your data sources grow. The cost worksheet on this site helps you model the financial impact of architectural choices on your compute and storage spend.

Work it out for your own case

Change the inputs and the figures update as you type. Nothing you enter leaves your browser.

Illustrative defaults — replace the unit prices with the ones on your own contract or price sheet.

Two line items only: what sits on disk, and what runs. Transfer, tooling and seat licences are separate bills and are not counted here.

Estimates for general guidance only. Real figures depend on the details you enter and on the provider you deal with.

The Ingestion Layer

Ingestion is how data enters the warehouse. Source systems — databases, APIs, file feeds, event streams — push or pull data into the pipeline through connectors that normalize the format and handle scheduling. The two primary patterns are batch ingestion, where data loads at scheduled intervals, and streaming ingestion, where data flows continuously.

Batch is simpler to build, easier to monitor, and sufficient for most analytical workloads. A nightly load that captures the previous day's transactions serves the majority of reporting needs without the complexity and cost of real-time streaming. Streaming is necessary when the business requires near-real-time analytics — fraud detection, live dashboards, operational alerts — and it costs more in both infrastructure and engineering attention.

Choose the pattern that matches your analytical latency requirement, not the one that sounds more modern. If your reports run daily and your stakeholders review them in the morning, a nightly batch load delivers everything they need. Overengineering the ingestion layer with streaming adds cost and maintenance burden without improving outcomes.

Source connector reliability also matters. A connector that fails silently and stops loading data without raising an alert creates a gap in the warehouse that analysts discover only when a report shows missing rows. Build alerting into the ingestion layer so failures are detected and addressed before they affect downstream queries and reports.

The Staging and Transformation Layer

Raw data that enters through the ingestion layer is not ready for analytical queries. It contains duplicates, inconsistent formats, null values, and records that violate business rules. The staging layer is where this raw data lands before transformation logic cleans and shapes it.

Transformation handles deduplication, data type standardization, business-rule application, and referential integrity checks. A customer record that appears in both the CRM and the billing system needs to be resolved into a single identity. A revenue figure denominated in multiple currencies needs to be converted to a common base. These operations happen in the staging layer so the storage layer receives only clean, consistent data.

Data quality checks belong here as well. Automated tests that verify row counts, check for unexpected nulls, and validate value ranges catch pipeline failures before bad data reaches the analysts. A pipeline without quality checks at the staging layer will eventually deliver incorrect results that someone uses in a decision before the error is discovered.

The Storage Layer

The storage layer holds the transformed, query-ready data in a structure optimized for analytical access. Columnar storage is the standard for warehouses because it reads only the columns a query references rather than scanning entire rows, which reduces compute time and cost for the aggregation-heavy workloads warehouses serve.

Partitioning divides tables into segments based on a key — typically date — so queries that filter by time period scan only the relevant partition instead of the full table. Compression reduces storage costs by encoding columnar data efficiently, which is particularly effective for columns with low cardinality where the same values repeat frequently.

Hot, warm, and cold tiering controls cost by keeping frequently accessed data on faster, more expensive storage and moving older or rarely accessed data to cheaper storage. The cost-drivers guide on this site covers how to tune these parameters to balance performance against budget as your data volume grows.

The Presentation Layer

The presentation layer is what analysts and business users interact with. It includes semantic models that translate technical table structures into business-friendly names, BI tool connections that serve dashboards and reports, access controls that limit who can see which data, and materialized views that pre-compute expensive queries for faster retrieval.

A well-built presentation layer hides the complexity of the underlying architecture. An analyst should be able to ask a question using business terms — revenue by region, customer churn by quarter — without needing to understand how the data was ingested, staged, or stored. If users regularly need to know the technical details to write their queries, the semantic layer needs more work.

Access controls at this layer enforce data governance. Not every user should see every table. Financial data, personally identifiable information, and competitive intelligence each require permission boundaries that the presentation layer enforces. Role-based access, row-level security, and column-level masking are tools that keep sensitive data visible only to authorized users. The star-schema guide on this site covers the modeling patterns that make the presentation layer intuitive and efficient for common analytical workloads.

Architecture decisions should match your current workload and team size — overengineering creates maintenance burden that outweighs the theoretical benefits.

Questions

Common questions

What is the most common data warehouse architecture?

A three-tier architecture with an ingestion/staging layer, a core storage layer, and a presentation layer is the most widely adopted pattern. Variations exist depending on whether transformation happens before or after loading, and whether the storage layer uses a traditional warehouse engine or a lakehouse hybrid.

How do I choose between batch and streaming ingestion?

Match the choice to your analytical latency requirement. If reports run daily and decisions are made on yesterday's data, batch ingestion is simpler, cheaper, and sufficient. If the business needs minute-level freshness for operational dashboards or alerting, streaming is necessary but carries higher cost and complexity.

Can I add layers to an existing warehouse architecture?

Yes. Most warehouse architectures are designed to be extended. Adding a new data source requires a new ingestion connector and transformation logic, but the storage and presentation layers can accommodate the additional data without a redesign. Plan for extensibility from the start by keeping schemas modular and pipelines independent.

What role does the semantic layer play?

The semantic layer translates technical table and column names into business terms so analysts and BI tools can query the warehouse using language that maps to how the business thinks. It also standardizes metric definitions — ensuring that revenue, churn, and conversion mean the same thing across every report — which prevents conflicting numbers from different queries.

Written & maintained by

Mustafa Bilgic — sole publisher, DataWarehousing.us

Mustafa Bilgic publishes independent, source-cited guides and free tools. This site takes no vendor sponsorship and sells no leads. Where a figure comes from a published source, that source is named on the page so you can check it yourself.

  • Sources: listed in full at the end of each guide.
  • Last reviewed: see the date shown on this page.

Compare on the things that actually differ

Read the comparison guides before you shortlist. Most of the difference between options sits in the detail, not the headline.

Back to the tool