Architecture Series

Modern Data Warehouse: Schema-on-Write at Cloud Scale

Explore the Modern Data Warehouse (MDW) — how it combines schema-on-write governance with cloud scale, MPP query execution, and structured BI performance.

The Modern Data Warehouse (MDW) is the evolution of the traditional enterprise data warehouse — rebuilt for the cloud era. It retains the governance, schema discipline, and BI performance that made data warehouses the backbone of enterprise analytics, while shedding the on-premises hardware constraints and adding massively parallel processing (MPP) at cloud scale.


Overview

A Modern Data Warehouse combines data lake ingestion capabilities with relational store governance and schema-on-write enforcement. It delivers optimised performance for BI and reporting workloads, with structured access control and consistent data quality.

Core philosophy: Govern data at ingestion. Deliver consistent, high-performance analytics at scale.


Data Organisation

Staging Layer

Raw data lands in a temporary staging area where it is extracted and lightly validated before transformation:

  • ETL tools: SSIS, Informatica, Talend, Azure Data Factory, AWS Glue

Raw / Archive Zone

Original source data is preserved in a data lake (ADLS, S3) for audit, compliance, and reprocessing needs. The warehouse itself stores only the transformed representation.

Curated Zone

Data is transformed, cleaned, and modelled into dimensional structures. Two dominant modelling approaches:

Star Schema: A central fact table surrounded by dimension tables. Optimised for BI queries with predictable join patterns.

Fact_Sales ←→ Dim_Customer
            ←→ Dim_Product
            ←→ Dim_Date
            ←→ Dim_Geography

Snowflake Schema: Normalised dimensions (dimensions of dimensions). Reduces redundancy but adds join complexity.

Semantic / Presentation Layer

Pre-aggregated marts and views tailored for specific business domains (Finance, Sales, HR). Query tools and BI platforms connect here.


Strengths

StrengthDetail
Query PerformanceMPP engines execute queries across hundreds of nodes in parallel
Schema EnforcementData quality is guaranteed at ingestion — consumers trust the data
ACID ComplianceFull transactional consistency with rollback support
BI Tool OptimisationNative integration with Power BI, Tableau, Looker, Qlik
Mature GovernanceFine-grained row/column-level security, masking, auditing
Cost TransparencyCompute and storage are metered separately — pause when not querying

Limitations

LimitationImpact
Rigid SchemaSchema changes (adding columns, renaming fields) require careful migration
Limited Unstructured SupportSemi-structured and unstructured data are second-class citizens
Higher Cost at ScaleStoring all data in the warehouse is more expensive than object storage
ETL ComplexityTransformation pipelines must be built and maintained before data is available
Slower IterationAnalysts must wait for ETL pipelines to deliver new data shapes
ML FrictionNot natively designed for model training; data must be exported

Common Technology Stack

Cloud MPP Warehouses

PlatformProviderKey Characteristic
SnowflakeMulti-cloudFully separated storage/compute; virtual warehouses; near-zero management
Google BigQueryGCPServerless; automatic scaling; built-in ML with BQML
Amazon RedshiftAWSTight S3 integration via Redshift Spectrum; RA3 nodes
Azure Synapse AnalyticsAzureUnified workspace; Spark + SQL pool; ADLS native
Databricks SQLMulti-cloudSQL warehouse on Lakehouse foundation; Delta Lake native

Transformation Layer

  • dbt (data build tool): SQL-first transformation framework with version control, testing, and lineage
  • Informatica / Talend: Enterprise ETL with visual pipelines and data quality rules
  • Azure Data Factory / AWS Glue: Cloud-native managed ETL

Dimensional Modelling

  • Kimball Methodology: Star schema, conformed dimensions, bus architecture
  • Inmon Methodology: Normalised 3NF enterprise data warehouse with downstream data marts

Dimensional Modelling Deep Dive

Fact Tables

Contain measurable, quantitative events (sales amounts, click counts, transaction values). Characteristics:

  • High row count (billions of rows common)
  • Foreign keys to dimension tables
  • Additive, semi-additive, or non-additive measures

Dimension Tables

Contain descriptive attributes that provide context to facts. Characteristics:

  • Lower row count (thousands to millions)
  • Slowly Changing Dimensions (SCD) for historical tracking

SCD Types:

TypeBehaviourUse Case
SCD Type 1Overwrite old valueCorrections, non-historical attributes
SCD Type 2Insert new row with date rangeTrack history (e.g. customer address changes)
SCD Type 3Add previous-value columnTrack one version of history

When to Choose a Modern Data Warehouse

Good fit when:

  • Your primary use case is structured BI reporting with predictable query patterns
  • Your data is predominantly structured/relational (transactional systems, CRM, ERP)
  • Data governance and compliance (SOX, GDPR, HIPAA) are non-negotiable requirements
  • Business users need self-service analytics through BI tools with consistent metrics
  • You need fast query response for dashboards serving hundreds of concurrent users

Poor fit when:

  • You ingest unstructured data (logs, images, video, IoT raw telemetry)
  • You need schema flexibility for exploratory and data science workloads
  • ML/AI is a primary use case and data scientists need raw data access
  • Cost optimisation requires storing petabytes of historical data cheaply
  • You need near-real-time data availability (sub-minute latency)

MDW vs Data Lake

DimensionModern Data WarehouseData Lake
SchemaSchema-on-writeSchema-on-read
Data typesStructured primarilyAll formats
Storage costHigherLower
Query performanceExcellent (MPP)Varies (depends on format/engine)
ACID transactionsNativeRequires table format layer
ML/AI readinessLimitedExcellent
GovernanceMature, built-inRequires tooling

The line between the Modern Data Warehouse and the Lakehouse is increasingly blurred. Platforms like Snowflake, BigQuery, and Synapse now offer direct querying of object storage, while Lakehouses add warehouse-grade governance. The Lakehouse is the convergence point.