Scaling Snowflake Sustainably Through Informed Warehouse Sizing

Context

As analytics adoption increased, Snowflake costs and performance variability grew in parallel. Queries were technically correct, but execution times were unpredictable. Some workloads experienced queueing and latency, while others consumed resources inefficiently. Cost discussions were reactive and often disconnected from real usage patterns.

The problem was not a lack of optimization effort, but a lack of shared understanding. Warehouse sizing decisions were historically based on assumptions, defaults, or isolated complaints rather than observed behavior across workloads.

Initial Situation

Multiple Workloads
↓
Single / Poorly Sized Warehouses
↓
Queueing · Latency · Cost Spikes
↓
Reactive Tuning

Decisions & Trade-offs

I took ownership of warehouse sizing as a platform-level decision rather than a tuning exercise. Instead of scaling warehouses up to absorb the load, I started from query history to understand how the platform was actually being used.

Query history revealed clear patterns: short, frequent ad hoc queries; scheduled workloads with predictable windows; and a long tail of expensive queries driving a disproportionate share of costs. Average metrics were misleading—distributions and concurrency windows told the real story.

Several options were evaluated and deliberately rejected. Simply increasing warehouse size would have masked inefficiencies while accelerating cost growth. Applying a single warehouse size across all workloads would have optimized for convenience rather than behavior. Early enforcement of strict limits was postponed until usage patterns were understood and communicated.

The trade-off was intentional: slightly more configuration complexity in exchange for predictable performance, clearer cost attribution, and sustainable growth.

Observed via Query History

Many short queries  ──► concurrency pressure
Few long queries    ──► cost concentration
Peak windows        ──► predictable contention

Implementation Approach

Warehouses were resized and aligned to workload characteristics rather than organizational convenience. Sizing decisions were driven by observed concurrency, execution time distributions, and peak usage windows.

Oversizing was deliberate. Smaller, fit-for-purpose warehouses made inefficiencies visible and actionable, while workload separation reduced blast radius. Query history analysis became a recurring input into platform decisions, not a one-time optimization exercise.

Target State

Distinct Workloads
↓
Fit-for-purpose Warehouses
↓
Predictable Performance
Clear Cost Attribution

Outcome

Performance stabilized, and queueing was significantly reduced without linear cost growth. Cost drivers became visible and explainable, shifting conversations from “why is this expensive?” to “which workloads justify their cost?”

Optimization moved from reactive firefighting to informed decision-making. Teams experienced more consistent performance, and platform behavior became easier to reason about as usage scaled.

What Became Possible

With warehouse sizing grounded in real usage data, the analytics platform could grow sustainably. New teams and workloads were onboarded with clear expectations, cost governance became proactive, and Snowflake evolved from a black box into a predictable, governable part of the analytics infrastructure.

Back to case studies