The Right-Sized Data Stack for a Commerce Brand (No, You Don’t Need Databricks)

Commerce brands consistently buy data infrastructure two sizes wrong: spreadsheet-and-prayer long past the point of pain, then — stung — an enterprise platform sized for a bank. Between those failure modes sits a boring, proven, right-sized stack that runs a €2M–200M commerce business for hundreds, not tens of thousands, per month.

The shape

One managed warehouse (BigQuery, Snowflake, or increasingly DuckDB/MotherDuck at the smaller end — commerce data is small by data-industry standards; a million orders a year is nothing). Managed ELT connectors (Fivetran/Airbyte-class) pulling Shopify, Amazon, ad platforms, GA4, your 3PL and finance tools on schedule — the settlement reports included, because that is where the margin truth hides. dbt for transforms, version-controlled, tested, turning raw feeds into a modelled layer: orders, margin, inventory, marketing. One BI tool on top — and dashboards built like products, not inventories.

The entire stack is configuration and SQL. No Spark clusters, no Kafka, no data lakehouse — commerce analytics is joins and aggregates over modest volumes, and pretending otherwise is how eighty-thousand-euro platform bills happen to forty-person companies.

The part that is actually hard

Not the tools — the semantics. What is net revenue (after refunds? fees? vouchers)? Which timestamp is "the sale" (order, payment, ship)? How do returns hit last month's cohort? Which channel owns a marketplace order fulfilled from webshop stock? Every dashboard argument in your company's history is an undocumented definition. The valuable work is a metrics layer where each number is defined once, in writing, in code review — and every tool downstream inherits it. Two weeks of definition workshops save two years of "whose number is right" meetings.

Sequencing that pays back fast

1. Week 1–2: warehouse + connectors live; raw data landing on schedule. Immediate win: settlement reconciliation queries that find money.
2. Week 3–5: core dbt models — orders, margin (fully loaded), inventory positions. First honest per-SKU contribution numbers most brands have ever seen.
3. Week 6–8: the three dashboards that matter (plan tracking, SKU contribution, channel P&L) plus alerting on the anomalies that cost money (stock-outs on winners, fee drift, ad overspend).

After that, the stack earns extensions: MMM feeds, forecast models, feed-quality monitoring. Each addition rides infrastructure that already paid for itself.

What to defer indefinitely

Real-time streaming (you rarely need it); customer-data platforms until your channel count forces one; reverse-ETL until a workflow demands it; anything whose pitch involves "AI-ready lakehouse". The stack above is AI-ready — clean modelled tables are exactly what analytical agents want to query. Boring first; clever rides on boring.