Home › Learn › Databricks permissions
Databricks · governance and security
Databricks permissions: who can do what, from account admin to a single table
A plain guide to Databricks permissions: the five layers (identity, entitlements, workspace object ACLs, Unity Catalog privileges, storage), admin roles, the USE CATALOG + USE SCHEMA + SELECT rule, ownership vs MANAGE, row filters, column masks and ABAC, service principals, workspace binding and audit, with ready-made grants for common roles, mistakes to avoid and a verdict for each choice.
Read from the Databricks documentation on 21 November 2026 (AWS pages). Several pages were only partly readable, and some features are Beta. The "Not confirmed" list at the bottom says what I could not check. SQL marked "sketch" has not been run.
The big picture
Five layers, and all of them must say yes
People say "permissions" as if it were one thing. In Databricks it is five separate checks, and a failed request usually means one layer said no while you were looking at another.
Five checks, one after another. All must pass.
| Layer | Question | Where you set it |
|---|---|---|
| 1. Identity | Who is asking: a person, a group, or a service principal (a non-human identity for jobs and apps)? | Account console, identity provider sync |
| 2. Entitlements | May this identity use this kind of feature in this workspace (notebooks, SQL, create compute)? | Workspace settings, identity and access |
| 3. Workspace object ACLs | May it open, run, edit or manage this notebook, job, cluster or warehouse? | The Permissions button on each object |
| 4. Unity Catalog | May it read or change this table, volume or function? Which rows and columns? | GRANT, row filters, column masks, ABAC policies |
| 5. Storage | May anything read the cloud files directly? | Storage credentials and external locations |
Layer 1: identity
Users, groups, service principals and the three admins
| Identity | What it is | Use it for |
|---|---|---|
| User | A person, identified by email | Interactive work |
| Group | A set of users, service principals and other groups | All grants. Account groups can get Unity Catalog data access and workspace permissions. |
| Service principal | An identity for code, with no password and no person behind it | Production jobs, pipelines and apps |
- Identities live in the account and are assigned to workspaces. Identity federation is on by default for new workspaces.
- Sync from your identity provider. The docs recommend automatic identity management (Microsoft Entra ID and Okta), where the provider stays the source of truth. SCIM provisioning is the fallback, and it does not support nested groups or service principals.
- Old workspace-local groups cannot be given Unity Catalog data access or used across workspaces. Convert them to account groups.
- Limits: accounts created on or after 17 November 2025 can hold up to 50,000 users and service principals combined and 10,000 groups. Older accounts: 10,000 and 5,000.
The three admin roles
| Role | Scope | Can do | Cannot do |
|---|---|---|---|
| Account admin | Whole account | Create metastores and workspaces, assign admin roles and workspace access, grant privileges on a metastore, configure storage credentials, enable system tables | Day-to-day object grants are not its job |
| Metastore admin (optional) | One metastore | Change ownership and grant privileges on any object, read metadata of all objects, set tags, delete the metastore | Metastores set up by automatic enablement have none by default |
| Workspace admin | One workspace | Manage membership and jobs' run-as, manage workspace objects, create catalogs and top-level securables (workspaces created after 8 November 2023) | Cannot grant on, or change the owner of, existing objects they do not own |
The docs recommend giving the metastore admin role to a group, not a person. Workspace admins also own the default workspace catalog, so they can manage everything in it.
Layer 4: Unity Catalog
The hierarchy and the rule for reading a table
Grants flow down the left column. The right column is the rule for reading a table.
| To do this | You need |
|---|---|
| Read a table or view | USE CATALOG, USE SCHEMA, SELECT |
| Write to a table | USE CATALOG, USE SCHEMA, MODIFY |
| Create a table | USE CATALOG, USE SCHEMA, CREATE TABLE on the schema |
| Create a schema | USE CATALOG, CREATE SCHEMA on the catalog |
| Run a function | USE CATALOG, USE SCHEMA, EXECUTE |
| Read files in a volume | USE CATALOG, USE SCHEMA, READ VOLUME |
Inheritance
- A privilege on a catalog applies to every current and future schema, table, view, volume and function in it. On a schema, to everything inside.
- Metastore-level privileges do not flow down to data. They cover metastore actions like
CREATE CATALOG. The exception isREAD METADATA. - Ownership does not flow down. A catalog owner does not own its schemas, though they can manage them.
- Views: a user reading a view does not need privileges on its underlying tables. The view owner's privileges are used.
- Workspace binding wins. A catalog not bound to a workspace cannot be reached from it, even with an explicit
SELECT(see s10).
Privileges
The privileges you will use, owners, MANAGE and GRANT
| Privilege | Where | What it allows |
|---|---|---|
USE CATALOG, USE SCHEMA | Catalog, schema | Required to reach anything inside. Gives no data access by itself. |
SELECT | Table, view, materialized view, or the schema or catalog above | Query it |
MODIFY | Table, schema, catalog | Insert, update and delete rows. INSERT, UPDATE, DELETE are finer Beta versions. |
CREATE TABLE, CREATE SCHEMA, CREATE VOLUME, CREATE FUNCTION, CREATE MATERIALIZED VIEW | Schema or catalog (CREATE SCHEMA: catalog) | Create that kind of object |
EXECUTE | Function, model, schema, catalog | Call the function or load the model |
READ VOLUME, WRITE VOLUME | Volume, schema, catalog | Read, or add and change, files |
REFRESH | Materialized view | Refresh it by hand |
APPLY TAG | Most objects | Add and edit tags |
BROWSE | Catalog, external location | See that objects exist and request access, without USE CATALOG |
READ METADATA | Most objects | See definitions, grants, filters and masks, without reading data. Not in ALL PRIVILEGES. |
MANAGE | Most objects | Grant and revoke, transfer ownership, rename, drop. Not in ALL PRIVILEGES. |
EXTERNAL USE SCHEMA | Schema, catalog | Temporary credentials so outside engines can read through open APIs. Not in ALL PRIVILEGES. |
ALL PRIVILEGES | Most objects | The ordinary privileges for that object. It leaves out MANAGE, READ METADATA and the two external-use privileges. |
Metastore-level creation rights (CREATE CATALOG, CREATE EXTERNAL LOCATION, CREATE STORAGE CREDENTIAL, CREATE SHARE and similar) are also privileges. The full reference lists many more, such as secrets, features, flows and connections.
Owner, MANAGE and ALL PRIVILEGES are three different things
| Owner | MANAGE | ALL PRIVILEGES | |
|---|---|---|---|
| How many | Exactly one principal (can be a group) | Any number | Any number |
| Can grant and revoke | Yes | Yes | No |
| Reads data automatically | Has every capability on the object | No, but can grant themselves SELECT | Yes, for the privileges it bundles |
Shown in SHOW GRANTS | No (implicit) | Yes | Yes |
Who may grant on an object: its owner, the owner of the containing catalog or schema, a MANAGE holder, or a metastore admin. The creator of an object becomes its owner, so give production objects to a group or service principal so they survive someone leaving. Transfer with ALTER TABLE main.sales.orders OWNER TO `data-eng`. For views, functions and models, only a metastore admin can hand them to anyone else; other owners can hand them only to themselves or a group they belong to.
GRANT, REVOKE, SHOW GRANTS
GRANT USE CATALOG ON CATALOG main TO `finance-team`; GRANT USE SCHEMA, SELECT ON SCHEMA main.gold TO `finance-team`; -- covers every current and future table in gold REVOKE SELECT ON SCHEMA main.gold FROM `finance-team`; SHOW GRANTS `finance-team` ON SCHEMA main.gold; SHOW GRANTS ON TABLE main.gold.orders; -- needs READ METADATA, MANAGE or ownership to see everyone's -- a service principal is named by its application ID, in backticks GRANT SELECT ON TABLE main.gold.orders TO `fab9e00e-ca35-11ec-9d64-0242ac120002`;
No SQL DENY. I found no DENY statement in the docs. Denials exist only as ABAC DENY policies (Beta, s7). Access is therefore built by granting the right amount, not by granting broadly and denying parts.
Starting state
Users start with no access to data. Workspace users get USE CATALOG on the workspace catalog, and USE SCHEMA plus the main CREATE privileges on its default schema. The docs also say all users have USE CATALOG on main. The sample catalog samples is read-only.
Layer 3: workspace objects
ACLs on notebooks, jobs, compute and warehouses
Workspace admins have CAN MANAGE on every workspace object, and people get CAN MANAGE (or IS OWNER) on what they create. Objects in a folder inherit the folder's permissions. In the UI the lowest read level is called CAN VIEW; the Permissions API calls it CAN READ.
| Object | Permission levels, lowest to highest |
|---|---|
| Notebook, folder, file, Git folder | CAN VIEW (CAN READ) · CAN RUN · CAN EDIT · CAN MANAGE |
| Job | CAN VIEW · CAN MANAGE RUN · CAN EDIT (Beta, folders only) · IS OWNER · CAN MANAGE |
| Lakeflow pipeline | CAN VIEW · CAN RUN · CAN EDIT (Beta) · CAN MANAGE · IS OWNER |
| Classic compute | CAN VIEW (preview) · CAN ATTACH TO · CAN RESTART · CAN MANAGE |
| SQL warehouse | CAN VIEW · CAN MONITOR · CAN USE · IS OWNER · CAN MANAGE |
| Dashboard, query | CAN VIEW (CAN READ) · CAN RUN · CAN EDIT · CAN MANAGE |
| Alert | CAN RUN · CAN MANAGE |
| Model serving endpoint | CAN VIEW · CAN QUERY · CAN MANAGE |
| Instance pool | CAN ATTACH TO · CAN MANAGE |
| Secret scope | READ · WRITE · MANAGE |
Jobs, in detail
| Ability | CAN VIEW | CAN MANAGE RUN | CAN EDIT | IS OWNER / CAN MANAGE |
|---|---|---|---|---|
| See job and results | Yes | Yes | Yes | Yes |
| See Spark UI and run logs | Yes | Yes | Yes | |
| Run now, with other parameters, cancel | Yes | Yes | Yes | |
| Edit settings | Yes (not the run-as identity) | Yes | ||
| Delete, change permissions | Yes |
- A job has one owner, and a group cannot be the owner.
- Run now runs with the permissions of the job owner, not the person who pressed the button. That is why production jobs should be owned by, or run as, a service principal.
Compute and warehouses, in detail
| Ability | Compute | Needs |
|---|---|---|
| Attach a notebook | Classic compute | CAN ATTACH TO |
| Start, restart, terminate | Classic compute | CAN RESTART |
| Edit, resize, add libraries, view driver logs | Classic compute | CAN MANAGE |
| Run queries | SQL warehouse | CAN USE (CAN MONITOR also runs queries) |
| Stop, edit, delete | SQL warehouse | IS OWNER or CAN MANAGE |
By default driver logs are visible only to CAN MANAGE. On the legacy No isolation shared compute, people with lower levels can see service account keys in the log4j file, which is one reason to avoid it.
Folder visibility. Someone with CAN VIEW on an object inside a folder can see the parent folder's name without any permission on the folder. The docs also warn that membership of account groups will soon affect workspace object permissions, so review existing grants.
Layer 2: entitlements, policies and access modes
Which features and which compute a person may use
| Entitlement | Lets the identity |
|---|---|
| Workspace access | Use notebooks, jobs, models and pipelines |
| Databricks SQL access | Use dashboards, queries and SQL warehouses |
| Consumer access | View shared dashboards, Genie and apps, and query warehouses from BI tools. Not create or edit. |
| Allow unrestricted cluster creation | Create compute with any settings. Without it, people create compute only through cluster policies granted to them. |
| Allow pool creation | Create instance pools |
- A user needs at least one access entitlement to enter a workspace. Consumer access is assigned individually, not through the
usersgroup. - Cluster policies limit the settings and cost a person can pick. Everyone has the Personal Compute policy by default. Editing a policy does not change existing compute until you enforce it, and job compute is enforced immediately.
- The
usersgroup changed in 2026. The docs say that from 15 June 2026 theusersgroup has no entitlements andadminshas them all, with enforcement for every workspace on 14 September 2026. The entitlements page still shows the older defaults, so check Settings, then Identity and access, in your own workspace.
Access modes decide what compute does with Unity Catalog
| Mode | Who uses it | Note |
|---|---|---|
| Standard | Many users, with data isolation between them | Python, SQL, Scala. The docs recommend it unless a feature you need is missing. |
| Dedicated | One user or one group | Also R. Filters, masks and dynamic views on it need Runtime 15.4 LTS or later and serverless enabled for the workspace. Unity Catalog Python UDFs are not supported. |
| No isolation shared (legacy) | Not described on the pages I read | Avoid |
Row and column security
Row filters, column masks, ABAC and dynamic views
Grants decide which tables. These decide which rows and columns inside a table the person sees.
| Table-level filter or mask | ABAC policy | Dynamic view | |
|---|---|---|---|
| What it is | A SQL function attached to one table or column | A rule on a catalog, schema or table that matches governed tags | A view whose SQL checks who is asking |
| Covers new tables automatically | No | Yes, through tags | No |
| Who is exempt | Not stated in the pages I read | Principals listed under EXCEPT | Anyone who can read the base table |
| Status | Standard | Core feature not labelled GA on the pages I read; DENY, views and metastore-level policies are Beta | Standard |
-- table-level row filter (from the docs)
CREATE FUNCTION salary_filter(dept STRING) RETURNS BOOLEAN RETURN dept IS NULL;
ALTER TABLE employees SET ROW FILTER salary_filter ON (department);
-- ABAC column mask on every column tagged pii=ssn in a schema, for one group (from the docs)
CREATE POLICY mask_ssn ON SCHEMA prod.customers
COLUMN MASK ssn_to_last_nr TO us_analysts EXCEPT admins
FOR TABLES MATCH COLUMNS has_tag_value('pii', 'ssn') AS ssn
ON COLUMN ssn USING COLUMNS (4);
-- dynamic view: use account-level group membership
CREATE VIEW sales_redacted AS
SELECT user_id, CASE WHEN is_account_group_member('auditors') THEN email ELSE 'REDACTED' END AS email
FROM sales_raw;
- A policy never grants access. The person still needs
SELECT. It only hides rows and columns they could otherwise read. - One distinct filter and one mask can resolve per person per table or column, or access is blocked with an error.
- Tags are a security boundary. Whoever can apply a tag can change which policies apply, so give
APPLY TAGand tag assignment sparingly. - Pipelines run as their owner. Streaming tables and materialized views evaluate policies as the pipeline owner or run-as identity. Add that identity to
EXCEPT, or the stored data may be permanently masked. - ABAC needs serverless, or standard or dedicated compute on Runtime 16.4 or later. Older runtimes cannot read ABAC-secured tables.
- Limits to know for table-level filters: not on views, not with time travel, no deep or shallow clones, not shareable through OpenSharing, no path-based file access. Dynamic views: use
is_account_group_member, notis_member, and do not grant people the base tables. - Quotas for ABAC: 10,000 policies per metastore, 100 per catalog or schema, 50 per table, 20 principals per policy.
For the same idea inside metric views, see row-level security with metric views.
Layer 5: storage
Storage credentials and external locations
| Object | What it is | Key privileges |
|---|---|---|
| Storage credential | A long-lived cloud credential, such as an IAM role | CREATE EXTERNAL LOCATION on it and on the metastore; keep the rights narrow |
| External location | A storage path plus a credential | READ FILES, WRITE FILES (needs READ FILES too), CREATE EXTERNAL TABLE, MANAGE |
| External table or volume | A pointer to data at an external location | The usual table and volume privileges |
GRANT READ FILES, WRITE FILES ON EXTERNAL LOCATION raw_landing TO `ingest-team`; GRANT CREATE EXTERNAL TABLE ON EXTERNAL LOCATION raw_landing TO `data-eng`;
The rule that matters most: do not give people or service principals direct access to the cloud buckets behind Unity Catalog managed tables or volumes. It bypasses the controls, leaves no audit or lineage, and writes from outside can corrupt data. Let Unity Catalog be the only identity that touches those files, and share file access through volumes (/Volumes paths).
An external location can be marked read-only, which can take up to 60 minutes to apply.
Jobs and apps
Service principals and run-as
- Why: a job that runs as a service principal uses that identity, not a person's. The docs say this stops jobs failing when someone leaves or a group changes.
- Create and assign: account admins add service principals, and workspace admins can add them to a workspace. Grant them Unity Catalog privileges and group membership like users, and give them access to the compute or warehouse they use.
- Managers: creating a service principal makes you its manager, but you still must give yourself its user role to run jobs as it.
- OAuth M2M: generate a secret for the service principal (up to 5 secrets, lifetime up to 730 days, unused ones deleted after 90 days). Request a token at the workspace
/oidc/v1/tokenendpoint. Tokens last 3,600 seconds, and the SDKs and CLI refresh them. Set a scope when you create the secret, since it cannot be changed later. - Secret scopes grant READ, WRITE or MANAGE on the whole scope, not single secrets, so anyone who can read one secret can read them all.
Isolation
Workspace binding
By default every catalog is open to every workspace on the metastore. Switch a catalog to ISOLATED and bind it to chosen workspaces, so production data cannot be reached from the dev workspace even by someone who holds a grant.
databricks catalogs update prod --isolation-mode ISOLATED
databricks workspace-bindings update-bindings catalog prod \
--json '{"add": [{"workspace_id": 123, "binding_type": "BINDING_TYPE_READ_WRITE"}], "remove": []}'
- Access level is read-write (default) or read-only. External locations can be bound too, but not read-only.
- Setting bindings needs a metastore admin, the catalog owner, or
MANAGE. - A binding is checked when a privilege is used, so unbinding later does not break existing external tables.
Audit
Who has what, and who changed it
- Current grants:
SHOW GRANTSon an object, and the Permissions tab in Catalog Explorer. You see everyone's grants only withREAD METADATA,MANAGEor ownership; others see their own. - Which policies apply:
SHOW EFFECTIVE POLICIES ON SCHEMA prod.customers. - History:
system.access.audit(Public Preview) records events. Filter onevent_datefor speed.
SELECT event_time, action_name, user_identity.email AS actor, response.status_code
FROM system.access.audit
WHERE service_name = 'unityCatalog'
AND action_name IN ('createPolicy', 'deletePolicy', 'createEntityTagAssignment', 'deleteEntityTagAssignment')
AND event_date >= current_date() - 30
ORDER BY event_time DESC;
I did not confirm the action names for grant and permission changes, so look at the distinct action_name values in your own table before building an alert.
Recipes
Ready-made grants and common mistakes
| Role | Grants (sketch, not run) | Verdict |
|---|---|---|
| Analysts read the gold layer | GRANT USE CATALOG ON CATALOG prod TO `analysts`;GRANT USE SCHEMA, SELECT ON SCHEMA prod.gold TO `analysts`; | Pursue: one grant covers future tables |
| ETL service principal writes silver | GRANT USE CATALOG ON CATALOG prod TO `<app-id>`;GRANT USE SCHEMA, SELECT, MODIFY, CREATE TABLE ON SCHEMA prod.silver TO `<app-id>`; | Pursue: least privilege for jobs |
| Data engineers build in dev | GRANT USE CATALOG, USE SCHEMA, CREATE TABLE ON CATALOG dev TO `data-eng`; | Pursue in dev, not prod |
| Auditors see who has what | GRANT READ METADATA ON CATALOG prod TO `auditors`; | Pursue: metadata only, no data |
| BI tool users | Consumer access entitlement, CAN USE on a warehouse, then the analyst grants above | Pursue |
Give everyone ALL PRIVILEGES on a catalog | — | Drop: far more than needed |
| Grant to individual users one table at a time | — | Drop: use groups and schema-level grants |
| Production jobs owned by a person | — | Drop: use a service principal |
| Direct bucket access for engineers | — | Drop: bypasses Unity Catalog |
| Filters only inside a metric view's YAML | — | Drop: that is scope, not security |
When something is denied, check in this order
- Identity and groups: who is the request really running as? A job runs as its owner or run-as identity.
- Entitlement: does it have Workspace access or Databricks SQL access in this workspace?
- Object ACL: CAN ATTACH TO on the cluster, CAN USE on the warehouse, CAN RUN on the notebook.
- Unity Catalog:
USE CATALOG,USE SCHEMA, thenSELECT. RunSHOW GRANTS. - Workspace binding: is the catalog bound to this workspace?
- Compute: is the access mode and runtime able to enforce filters, masks and ABAC?
- Storage: for external files, does the location grant
READ FILES?
Honest limits
Not confirmed
- Partly read pages. Several docs pages were truncated, and parts of the privilege reference, ownership transfer rules and external-location SQL came from a summary, not the full text. Check them before building an approval process.
- DENY. No SQL
DENYfound. The DENY policy syntax page was not read. - Who bypasses table-level filters and masks. The docs I read do not say whether owners or
MANAGEholders see unfiltered data. - Whether ABAC is GA. No GA label found on the pages I read; sub-features are labelled Beta.
- Job run-as. I confirmed job ownership rules, but not the run-as setting itself, or whether it can differ from the owner.
- The
usersgroup today. Which entitlement model your workspace uses depends on its state after the 2026 change. - Not read: cluster policy definition rules and the policy permission level name, the dedicated group access page, the standard-mode restricted settings list, Genie space permissions, Delta Sharing privileges, governed tag creation, SCIM setup, storage credential creation, grant-change
action_namevalues. - All grants shown as "sketch" are patterns from the docs, adapted to made-up names. None was run.
FAQ
Questions people ask
What do I need to read a table in Unity Catalog?
USE CATALOG on the catalog, USE SCHEMA on the schema, and SELECT on the table, or SELECT on the schema or catalog above it, which inherits downward. SELECT alone is not enough.
What is the difference between an owner and MANAGE?
An object has exactly one owner, who has every capability on it. MANAGE can be given to several principals and lets them manage privileges, transfer ownership, rename and drop, but it does not give them data access. A MANAGE holder can still grant themselves SELECT. ALL PRIVILEGES does not include MANAGE.
Is there a DENY statement?
I found no SQL DENY statement in the docs. Denial exists through ABAC DENY policies, which are labelled Beta.
Should I grant permissions to users or groups?
Groups. Account groups can be granted Unity Catalog data access and workspace permissions, and a person who changes team only has to move between groups. Service principals, not people, should own and run production jobs.
Next
Where to go next
Related on this site: Row-level security with metric views · Databricks compute and cost · Who maintains your tables · Databricks materialized views · Medallion architecture · AI governance