How does serverless scale up and down?

Summary

• Instant startup from warm pools. Serverless compute starts in seconds from Databricks-managed warm pools; serverless SQL warehouses typically start in about 2–6 seconds.
• Automatic autoscaling. An intelligent autoscaler adds and removes individual nodes as workload changes, so you get the capacity you need without configuring cluster sizes.
• Scale-to-zero when idle. Capacity is released after a short idle period; serverless SQL warehouses Auto Stop after 10 minutes idle by default, so you aren't billed for idle compute.
• Scale-out for concurrency. Serverless SQL warehouses use Intelligent Workload Management to add clusters as query concurrency rises — up to 40 clusters per warehouse — and queue rather than reject at the limit.
• Pay for work done. You're billed only when compute is assigned to your workloads, not for the time to acquire and set up instances.

How does serverless scale up and down?

Databricks serverless compute scales automatically in both directions — up to meet demand and down to zero when idle — without you choosing instance types or tuning cluster-scaling parameters. Databricks maintains warm pools of pre-provisioned capacity so compute is ready in seconds.

Why Databricks serverless scales automatically

• Instant startup. Because Databricks keeps warm pools of pre-provisioned instances, serverless compute starts in seconds; serverless SQL warehouses typically start in about 2 to 6 seconds.
• Graduated autoscaling. An intelligent autoscaler scales capacity in graduated steps to meet demand, adding and removing individual nodes as load changes so you pay for exactly the capacity you need.
• Intelligent Workload Management for SQL. Serverless SQL warehouses use machine-learning-based Intelligent Workload Management (IWM): when a query arrives, IWM predicts its resource needs and starts it immediately if capacity exists, or queues it and provisions more clusters if wait times rise, then scales down when demand drops.
• Scale to zero. The autoscaler reduces provisioned capacity after a few minutes when resources are no longer needed; for SQL warehouses, Auto Stop defaults to stopping after 10 minutes idle, eliminating charges for idle infrastructure.
• Scale out for concurrency. SQL warehouses add clusters to handle concurrency — Databricks recommends about one cluster per 10 concurrent queries, and a warehouse can scale across up to 40 clusters — and queries queue rather than being rejected at the limit.
• Elastic billing. You are billed only when compute is assigned to your workloads, not for the time to acquire and set up instances.

Getting started

1. Select serverless compute for notebooks, jobs, or SQL warehouses.
2. Set an Auto Stop idle timeout and, for SQL, a minimum/maximum cluster range for concurrency (see SQL warehouse types).
3. Let the autoscaler size capacity up and down automatically as workloads change.
4. Learn more in the what is serverless computing overview.

FAQs

How fast does serverless start?

Serverless compute starts in seconds from Databricks-managed warm pools; serverless SQL warehouses typically start in about 2 to 6 seconds.

Does serverless scale to zero?

Yes. The autoscaler releases capacity after a short idle period, and serverless SQL warehouses Auto Stop after 10 minutes idle by default, so you aren't billed for idle compute.

How does serverless handle many concurrent queries?

Serverless SQL warehouses use Intelligent Workload Management to add clusters as concurrency rises — up to 40 clusters per warehouse — and queue queries rather than reject them at the limit.