Introducing Kubernetes for Snowflake

Since the dawn of civilization, humanity has strived towards one goal: running data warehouse workloads as cheaply and efficiently as possible. 

Today, we’re excited to announce AI-powered workload scheduling for Snowflake to do just that: scheduling can instantly save 40%+ on your Snowflake bill with no overhead and no downside.

What Is Compute Scheduling?

Compute scheduling is the process of deciding where your workloads execute. In Snowflake, this means deciding which warehouses to run your queries on, and by default, Snowflake’s scheduling is static. Each workload is assigned to a specific warehouse and will always run on that warehouse.

This is simple to set up, but the downside is that it’s not efficient. Our users’ warehouse utilization is typically in the range of 40%-60%; in other words, half of their compute bill is paying for idle or underutilized hardware.

Enter Kubernetes. The gold standard for efficient hardware utilization in physical data centers is using a scheduler, like Kubernetes or Borg, to dynamically allocate each incoming workload based on the workload’s resource requirements and available resources in your fleet. This brings utilization rates of 80%-90% - figures which are usually impossible to hit with a static configuration.

Scheduling in Action

Here’s an example of a user running three workloads on three underutilized warehouses: 

‍

This is a reasonable, static setup for ETL, BI, and ad-hoc analysis. However, if we know all three warehouses are underutilized, we can do better. Here’s what it looks like with scheduling:

‍

Real-world examples are more complex and, of course, depend on actual utilization and runtimes. Here’s what things look like in practice:

This diagram shows which warehouses a query was sent to (on the left) and which it ran on (on the right).

How do you use AI?

Our scheduling agent makes real-time routing decisions to optimize cost and performance based on the global state of your warehouses each time a query comes in. This relies on ML in two crucial ways.

First, Kubernetes scheduling depends on two inputs: resource requirements of incoming workloads and available hardware resources. Snowflake queries don’t come with CPU and memory requirements, and Snowflake warehouses don’t expose available capacity - so we use LLMs to bridge the gap. 

Our models are trained on trillions of SQL tokens to accurately predict warehouse capacity and query runtime.

Second, our scheduling agent is trained end-to-end using Reinforcement Learning (RL) to make optimal decisions. Specifically, the agent is trained to minimize your Snowflake costs, subject to latency constraints. 

This is a classic RL setup: actions (job placement), rewards (cost), and constraints (latency guarantees). Over time, the model learns how to:

• Queue jobs for smarter packing
• Anticipate spikes in resource needs
• Upsize queries opportunistically when there’s spare capacity - so jobs finish faster
• Balance job latency (how fast a query runs) with queue latency (how long it waits to start)

This combination of models enables us to orchestrate your entire Snowflake workload intelligently, efficiently and automatically.

How do I try it?

Kubernetes for Snowflake is now generally available. Book a demo to learn more, or follow the instructions here to get a savings estimate. Your Snowflake bill could be 40% lower by the end of the week.

‍