Data Platform · FinOps
Re-architecting a Snowflake Platform for Cost & Speed
A Snowflake bill growing far faster than the value it produced — and how disciplined warehouse design, query hygiene, and modelling cut spend by roughly 40% while queries got faster.
Problem
The Snowflake bill had nearly doubled year on year with no corresponding growth in value delivered. Nobody could say which teams or workloads were responsible, so every proposed fix was a guess.
Business Context
Finance had flagged data platform costs as a growing line item, and the easy-but-wrong answer — throttle everything and slow the business down — was on the table. The mandate was to cut cost without cutting capability.
Architecture
I separated workloads that had been sharing one oversized warehouse — ad-hoc analytics, scheduled transformations, and BI — onto right-sized warehouses with aggressive auto-suspend. On top of that I added clustering on the largest tables, leaned into result caching, and introduced curated marts so common dashboards stopped re-aggregating raw data on every load.
Technical Decisions
Cost had to become visible before it could be managed, so a usage dashboard built from Snowflake's own account views broke spend down by warehouse, team, and query pattern. Resource monitors capped runaway spend as a safety net rather than a bottleneck.
Trade-offs
Pre-aggregating into marts spends engineering effort and some build-time compute up front. We accepted that in exchange for eliminating repeated query-time compute — a clear net win at the platform's query volumes, verified against real usage before committing.
Implementation
We rolled changes out workload by workload, watching the cost dashboard after each move to confirm the saving was real and nothing regressed. Retiring 'zombie' jobs — pipelines feeding reports no one opened — turned out to be one of the largest single wins.
Challenges
The hardest part was cultural, not technical: getting teams to see the cost of their own queries. Once the dashboard made spend personal and visible, the worst offenders quietly optimised themselves without a single mandate.
Performance Improvements
Because most of the savings came from scanning less data and caching more, queries got faster as costs fell. The optimisation had almost no downside for end users — they noticed better dashboards, not restrictions.
Business Value
Compute spend dropped by roughly 40% while performance improved, turning a worrying cost trend into a well-governed, predictable line item — and rebuilding finance's confidence in the platform.
Future Enhancements
Automating the cost governance itself: an agent that watches usage patterns, flags newly-expensive queries and idle warehouses, and proposes optimisations before the bill ever moves.
Let's build something intelligent together.
Whether you're modernising a data platform or bringing agentic AI into production, I'd love to hear what you're working on.