How reliable is serverless?

Summary

• Databricks serverless compute is built for reliability: Databricks operates a secure, fault-tolerant fleet on your behalf, so workloads keep running through infrastructure failures.
• Automatic multi-availability-zone failover is built in with no configuration — Databricks selects zones, replaces VMs lost to a zonal outage, and redistributes workloads automatically.
• Automatic instance-type failover and VM fallback protect workloads from cloud-provider capacity shortages by switching to compatible instances when a preferred type is unavailable.
• Versionless and automatically maintained: Databricks applies runtime upgrades and security patches for you, and can automatically roll a workload back to the last stable version if an issue is detected.
• Serverless inherits the platform's high availability — the control plane recovers automatically from a zone failure, typically within a 15-minute RTO with an RPO of 0 — and Model Serving runs as a highly available, low-latency service.

How reliable is serverless?

Serverless compute on Databricks is designed to be highly reliable because Databricks operates the underlying fleet for you. Rather than provisioning and maintaining clusters yourself, you run workloads on a managed pool of compute that Databricks keeps patched, secured, and resilient to infrastructure failures. Reliability is built in at several layers: availability-zone failover, capacity failover across instance types, automatic maintenance, and platform-level high availability. See Serverless compute.

Why Databricks serverless compute is reliable

• Automatic multi-zone failover. Serverless compute provides automatic multi-availability-zone failover with no additional configuration. Databricks manages zone selection, automatically replaces VMs lost to a zonal outage, and redistributes workloads so recovery is transparent to users. See Design a high availability strategy.
• Warm pools for fast, dependable startup. Databricks maintains warm pools of hot-standby instances with the runtime and critical services pre-loaded, so compute is ready in seconds rather than waiting on infrastructure to provision.
• Instance-type failover and VM fallback. When a preferred VM type is unavailable, serverless compute automatically falls back to a compatible instance type, shielding workloads from cloud-provider capacity shortages and stockout errors.
• Versionless and automatically maintained. Serverless uses a versionless architecture with no manual runtime upgrades. Databricks applies runtime upgrades and security patches automatically, with a 99.998% success rate across upgrades, and if an issue is detected during an upgrade, workloads are automatically rolled back to the previous stable version while the problem is resolved.
• Hardened and continuously patched. Serverless VM and container base images are hardened to at least CIS Level 1, new images are released regularly, and serverless compute restarts automatically before a maximum lifetime to pick up the latest image.
• Platform high availability underneath. The Databricks control plane is multi-zonal and fails over automatically to healthy zones, typically within a 15-minute RTO and with an RPO of 0 during a zonal outage. Model Serving runs as a highly available, low-latency service with automatic scaling.

Getting started

• Read Serverless compute to see how the managed fleet works.
• Review Design a high availability and disaster recovery strategy for the platform's HA guarantees.
• Follow the reliability best practices, including designing for failure and testing recovery procedures.
• Read What is serverless computing? for the concept and benefits.

FAQs

Does serverless survive an availability-zone failure?

Yes. Serverless compute provides automatic multi-availability-zone failover with no additional configuration: Databricks selects zones, replaces VMs lost to a zonal outage, and redistributes workloads automatically.

What happens if my preferred VM type runs out of capacity?

Serverless automatically falls back to a compatible instance type, so workloads are protected from cloud-provider capacity shortages and stockout errors without any action from you.

Do I have to manage runtime upgrades or patches on serverless?

No. Serverless is versionless: Databricks applies runtime upgrades and security patches automatically and can roll a workload back to the last stable version if an issue is detected during an upgrade.

How reliable is the platform serverless runs on?

The Databricks control plane is multi-zonal and fails over automatically to healthy zones, typically within a 15-minute RTO and with an RPO of 0 during a zonal outage, and Model Serving runs as a highly available service.