Access modes & compute policies
Databricks · Compute
Theory
Access mode decides who can use a cluster and how Unity Catalog security is enforced. Standard access mode (formerly Shared) lets many users share one cluster with full user isolation; it supports Python, SQL and Scala and is the recommended mode for most work. Dedicated access mode (formerly Single user) assigns the cluster to one user or one group, which you need for things like R, RDD APIs, GPUs or some ML libraries. The default Auto setting picks a mode for you; Legacy modes such as No isolation shared are not recommended and do not support Unity Catalog.
Compute policies are admin-defined rule sets that limit what users can create: allowed runtimes and node types, maximum workers, required tags, forced auto-termination. Users pick a policy in the Policy drop-down and the form only shows allowed values. For example, the sales team's policy forces team=sales, Max 4 workers and 60-minute auto-termination.
Why we use it
Without access modes, users on a shared cluster could see each other's data; without policies, anyone could start a huge cluster and forget it. Together they give secure sharing and cost control.
When to use it
Use Standard access mode for shared team clusters and most notebooks. Use Dedicated when a single user or group needs features Standard does not allow. Use policies whenever non-admins create compute; serverless needs neither setting.
How it works
- User creates compute
- Picks a compute policy in the Policy drop-down
- Form shows only values the policy allows
- Does one user or group need R, RDDs, GPUs or ML-only features? Yes: Access mode: Dedicated. No: Access mode: Standard.
- Cluster enforces Unity Catalog permissions per user
Syntax
Create compute → Advanced → Access mode: Auto | Manual → Standard | Dedicated
Compute → Policies → Create policy
Policy JSON: {"autotermination_minutes": {"type": "fixed", "value": 60}}Example
{
"autotermination_minutes": {"type": "fixed", "value": 60, "hidden": true},
"autoscale.max_workers": {"type": "range", "maxValue": 4},
"custom_tags.team": {"type": "fixed", "value": "sales"},
"data_security_mode": {"type": "fixed", "value": "USER_ISOLATION"},
"spark_version": {"type": "unlimited", "defaultValue": "auto:latest-lts"}
}Key points
- Standard = formerly Shared; Dedicated = formerly Single user (can be a group)
- Both Standard and Dedicated work with Unity Catalog; legacy modes do not
- In APIs, Standard is data_security_mode USER_ISOLATION, Dedicated is SINGLE_USER
- Policies limit settings and can hide them; users need permission to use a policy
- Policies (and tags) are the main tool to control classic compute cost