What you will be able to do
- Grant the minimum privileges for Feature Store producer and consumer roles, including grants that cover feature views created later
- Fix a feature view registration that fails because change tracking is off on a source table owned by another role
- Choose between OWNERSHIP, USAGE and READ on a registered model for each kind of inference
- Decide which compute pool privilege a role needs to run, operate, inspect or own ML compute
- Identify which object each ML component's privileges are granted on, and which role grants them
Key concept
ML assets are governed as ordinary Snowflake objects — Feature stores, models and compute pools are schema- or account-level objects. You control access to them with the same role-based grants used for tables and warehouses, and you usually arrange those grants in a role hierarchy rather than granting them user by user.
1.Producer and consumer roles on a Feature Store schema
A Feature Store has two kinds of user, and each gets its own role. Producers create and operate feature views. Consumers only read feature views and entities. A Feature Store is a schema (see Feature Store schema and entities (join keys)), and its feature views are dynamic tables or views. That makes every Feature Store permission a grant on that schema or on the objects inside it. There is no separate ML permission model to learn.
The consumer minimum is short. It covers USAGE on the database and schema, SELECT and MONITOR on the schema's dynamic tables, SELECT and REFERENCE on its views, and USAGE on the warehouse passed to the feature store initializer. None of these grants touches the source tables behind the feature views. A consumer reads what the store has already materialized. Consumers who also build training datasets need more: CREATE TABLE and CREATE DATASET on the destination schema, SELECT and REFERENCE on the tables there, and USAGE on the DATASETs.
Producers hold everything a consumer holds and add the privileges that create and run feature views. They need CREATE DYNAMIC TABLE, CREATE VIEW and CREATE TAG on the Feature Store schema, CREATE TABLE and CREATE DATASET for training data, OPERATE on the schema's dynamic tables and tasks to manage refresh settings, and USAGE on the warehouse. Producers inherit the consumer grants through the role hierarchy, not through duplicate grants. The setup script grants the consumer role to the producer role and the producer role to SYSADMIN.
| Privilege | Granted on | Consumer | Producer |
|---|---|---|---|
| USAGE | Feature store database and schema | Required | Required (inherited) |
| SELECT, MONITOR | Dynamic tables in the schema | Required | Required (inherited) |
| SELECT, REFERENCE | Views in the schema | Required | Required (inherited) |
| USAGE | Warehouse passed to the initializer | Required | Required |
| CREATE DYNAMIC TABLE, CREATE VIEW, CREATE TAG | Feature store schema | Not granted | Required |
| CREATE TABLE, CREATE DATASET | Feature store or destination schema | Optional, for training datasets | Required for training datasets |
| OPERATE | Dynamic tables and tasks in the schema | Not granted | Required to manage refresh |
Producers add new feature views all the time, so consumer grants must reach objects that do not exist yet. The setup script grants each object privilege twice. The ON ALL grant covers dynamic tables, views and datasets that already exist, and the ON FUTURE grant covers the ones producers create later. Without the FUTURE grant, every new feature view stays invisible to consumers until an administrator runs another GRANT.
-- Grant CONSUMER role privileges
GRANT USAGE ON DATABASE IDENTIFIER($FS_DATABASE) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT USAGE ON SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT SELECT, MONITOR ON FUTURE DYNAMIC TABLES IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT SELECT, MONITOR ON ALL DYNAMIC TABLES IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT SELECT, REFERENCES ON FUTURE VIEWS IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT SELECT, REFERENCES ON ALL VIEWS IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);Only a role with MANAGE GRANTS, CREATE ROLE and CREATE SCHEMA on the database can set all of this up. That can be ACCOUNTADMIN or a narrower custom role. From snowflake-ml-python 1.6.3, the setup_feature_store utility creates the producer and consumer roles and their grants in one call. With several feature stores, give each store, or each logical group of stores, its own producer and consumer pair.
Access to any Snowflake ML component comes from privileges granted to roles, and what changes from one component to the next is the object the grant lands on. For a Feature Store, the grants sit on its schema and the dynamic tables and views inside it. For a registered model, the owner grants USAGE or READ on the model itself to the role that will call it. For a compute pool, the grants sit on the pool: USAGE to run work there, while creating a new pool takes CREATE COMPUTE POOL on the account. Grant to roles rather than to individual users, and give each role only the privileges its job needs.
Checkpoint 1 of 6· Fill the gap
Producers keep adding feature views to this Feature Store schema. The consumer role must be able to read the dynamic tables that exist now and also the ones producers create later, without anyone running another GRANT. Which keyword completes the first statement?
GRANT SELECT, MONITOR ON ? DYNAMIC TABLES IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);
GRANT SELECT, MONITOR ON ALL DYNAMIC TABLES IN SCHEMA IDENTIFIER($SCHEMA_FQN) TO ROLE IDENTIFIER($FS_ROLE_CONSUMER);The ON ALL grant only covers dynamic tables that already exist. The ON FUTURE grant covers dynamic tables that producers create later. Without the FUTURE grant, each new feature view stays hidden from consumers until an administrator grants access again.
Source: docs.snowflake.comCheckpoint 2 of 6· Exam question
A data science team needs a consumer role that can read feature views for batch inference in a Feature Store schema, but must not create or refresh feature views. Which set of grants is the minimal one?
Correct answer: A — USAGE on the database and schema, SELECT and MONITOR on dynamic tables, SELECT and REFERENCE on views in the schema, and USAGE on a warehouse.
- A. Correct. Consumers need USAGE on the container objects, SELECT and MONITOR on the dynamic tables backing feature views, SELECT and REFERENCE on the views, and a warehouse to run queries.
- B. Incorrect. CREATE DYNAMIC TABLE and CREATE TAG are producer privileges for creating feature views, which exceeds the read-only requirement.
- C. Incorrect. Owning the schema gives broad administrative control, violating least privilege, and it does not cover objects owned by other roles.
- D. Incorrect. Feature views are backed by dynamic tables and views in the Feature Store schema, so privileges on those objects are required, not on source tables.
2.Change tracking on source tables you do not own
An incremental feature view needs change tracking on its source tables, and turning it on is an owner's action. When the producer role owns the source tables, Snowflake enables change tracking automatically as the feature view is created. When another team's role owns them, registration fails with a 'change tracking is not enabled' error.
You can fix this without taking the table away from its owner in two ways. The owning role can enable change tracking itself with one ALTER TABLE. The alternative is to transfer OWNERSHIP to the creating role, create the dynamic table so change tracking turns on automatically, and then transfer OWNERSHIP back. The ALTER TABLE route is the cleaner one, because ownership never changes hands.
ALTER TABLE mydb.myschema.raw_orders SET CHANGE_TRACKING = TRUE;OPERATE does not solve this. OPERATE applies to the Feature Store's own dynamic tables and tasks, and it controls their refresh settings. It grants nothing on the upstream source table. A view over untracked tables raises the same error, so enable change tracking on the base tables that the view references.
Checkpoint 3 of 6· Check yourself
A producer role tries to register an incremental feature view on raw_orders, a table owned by another team's role. Registration fails with a 'change tracking is not enabled' error. Which fix resolves it while the other team keeps ownership of the table throughout?
Only the source table's owner can turn on change tracking, and running ALTER TABLE as the owner means ownership never changes hands. OPERATE only manages refresh on the Feature Store's own dynamic tables, SELECT does not turn on change tracking, and a view over untracked tables raises the same error.
“the source tables must have change tracking enabled, or the user must have OWNERSHIP of these tables to automatically enable change tracking”Source: docs.snowflake.com
Checkpoint 4 of 6· Exam question
A producer role registers a feature view with incremental refresh over a source table owned by another team's role. Registration fails because change tracking is not enabled on the source table. What resolves this while keeping ownership with the other team?
Correct answer: B — Have the source table's owner role run ALTER TABLE ... SET CHANGE_TRACKING = TRUE, or let the producer role hold OWNERSHIP to enable it automatically.
- A. Incorrect. OPERATE controls suspending and resuming refresh, but incremental refresh still needs change tracking on the source table.
- B. Correct. Incremental refresh depends on change tracking; the table owner can enable it, or the producer needs OWNERSHIP so the Feature Store can enable it.
- C. Incorrect. A grant option lets a role pass on SELECT, but it has no effect on whether change tracking is enabled for the table.
- D. Incorrect. DOWNSTREAM only sets the refresh timing relative to dependents and never derives change data from query history.
3.OWNERSHIP, USAGE and READ on registered models
A registered model has exactly three privileges. The difference between USAGE and READ follows where inference runs. USAGE allows warehouse inference and SHOW commands but blocks the code, weights and artifacts. READ allows inference on Snowpark Container Services (SPCS) and access to the model files. OWNERSHIP covers everything else: managing versions, updating metadata and granting the other two privileges.
| Privilege | Allows | Does not allow |
|---|---|---|
| OWNERSHIP | Managing model versions, accessing artifacts, updating metadata | Being held by more than one role at a time |
| USAGE | Warehouse inference, SHOW MODELS, SHOW VERSIONS IN MODEL | Model code, weights or other artifacts |
| READ | SPCS inference, model files and metadata, SHOW MODELS, SHOW VERSIONS IN MODEL | Managing versions or updating metadata |
GRANT USAGE ON MODEL my_model TO ROLE prod_role;
-- OR
GRANT READ ON MODEL my_model TO ROLE prod_role;Only one role can own a model. Shared control comes from granting that owning role to several users or roles. This matters once promotion to production is someone else's job. When production lives in a separate schema, the role that copies a version into it needs OWNERSHIP or READ on the source model. USAGE is not enough, because USAGE cannot reach the artifacts.
Checkpoint 5 of 6· Check yourself
Data scientists own a model in a development schema. A separate platform role will copy approved versions into a production schema and serve them on Snowpark Container Services, but it must not take ownership of the source model. Which privilege on the source model should the model owner grant to the platform role?
Copying a version into another schema requires the model's artifacts, so the promoting role needs OWNERSHIP or READ on the source model. Ownership is ruled out here, which leaves READ. READ also covers SPCS inference, while USAGE only allows warehouse inference and blocks the code, weights and other artifacts.
“Note that the role that promotes models to production should have OWNERSHIP or READ privilege on the source model.”Source: docs.snowflake.com
Sources2
4.Privileges for ML compute pools
Compute pools supply the container compute for ML Jobs and SPCS services, and each pool privilege matches one job function. USAGE lets a role run services and jobs on the pool. OPERATE lets it suspend and resume the pool. MODIFY lets it change the pool's properties. MONITOR lets it see the pool and what is running there. OWNERSHIP gives full control and, as with models, belongs to one role at a time.
| Privilege | Enables | Typical holder |
|---|---|---|
| USAGE | Running a service or a job; communicating with the service | Data scientists submitting ML Jobs, inference services |
| OPERATE | Suspending or resuming the compute pool | Platform operators |
| MODIFY | Altering the compute pool and setting properties | Platform administrators |
| MONITOR | Viewing pool usage, properties, and listing pools | Observability and cost reviewers |
| OWNERSHIP | Full control; one role at a time | The platform team that created the pool |
USAGE on the pool alone does not let a role run an ML Job. The role also needs USAGE on the database and schema where the job runs, CREATE SERVICE on that schema, and USAGE on a stage for the job payload. If the stage does not exist yet, the role needs CREATE STAGE so the job can create it. The role also needs privileges on whatever tables and warehouses the workload reads. Creating a new pool instead of reusing one takes CREATE COMPUTE POOL on the account.
Checkpoint 6 of 6· Check yourself
An ML Job will run on an existing compute pool. The data scientist's role already holds USAGE on that pool, on the job's database and schema, and on the payload stage, plus SELECT on the tables the job reads. Which grant is still missing before the role can run the job?
USAGE on the pool is not enough by itself. Running ML Jobs in an existing environment also takes CREATE SERVICE on the schema where the jobs run. OPERATE and MODIFY are for platform operators who manage the pool, and CREATE COMPUTE POOL is needed only to create a new pool, not to reuse one.
“CREATE SERVICE privilege on the schema to create and manage ML Jobs”Source: docs.snowflake.com
Sources
(6)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Consumers can read information about feature views and entities in the feature store.”
↩︎ Producer and consumer roles on a Feature Store schema“SELECT and REFERENCE on views in the feature store schema.”
↩︎ Producer and consumer roles on a Feature Store schema“OPERATE on the dynamic tables and tasks in the feature store schema to manage Feature View refresh settings.”
↩︎ Producer and consumer roles on a Feature Store schema“GRANT ROLE IDENTIFIER($FS_ROLE_CONSUMER) TO ROLE IDENTIFIER($FS_ROLE_PRODUCER);”
↩︎ Producer and consumer roles on a Feature Store schema“A role with MANAGE GRANTS, CREATE ROLE, and CREATE SCHEMA ON DATABASE <DB> privileges is needed to configure the necessary Feature Store roles”
↩︎ Producer and consumer roles on a Feature Store schema“snowflake-ml-python package version 1.6.3 and later include a setup_feature_store utility API”
↩︎ Producer and consumer roles on a Feature Store schema“the source tables must have change tracking enabled, or the user must have OWNERSHIP of these tables”
↩︎ Change tracking on source tables you do not own“SELECT on and MONITOR on DYNAMIC TABLES in the feature store schema.”
↩︎ Prediction“the source tables must have change tracking enabled, or the user must have OWNERSHIP of these tables to automatically enable change tracking”
↩︎ Checkpoint - 2.https://docs.snowflake.com/en/developer-guide/snowflake-ml/model-registry/model-managementOfficial docs
“The owner of a model can grant access to any role as follows:”
↩︎ Producer and consumer roles on a Feature Store schema“Model objects have three privileges: OWNERSHIP, USAGE, and READ.”
↩︎ OWNERSHIP, USAGE and READ on registered models“Only one role can own the model, but you can grant that role to multiple users or to other roles.”
↩︎ OWNERSHIP, USAGE and READ on registered models“Note that the role that promotes models to production should have OWNERSHIP or READ privilege on the source model.”
↩︎ OWNERSHIP, USAGE and READ on registered models“Because machine learning models are first-class objects in Snowflake, you can use all standard Snowflake governance capabilities with them, including role-based access control”
↩︎ Key concept“Roles with only USAGE cannot access the model code, weights, or other artifacts.”
↩︎ Exam trap 2 - 3.https://docs.snowflake.com/en/developer-guide/snowpark-container-services/working-with-compute-poolOfficial docs
“To create a compute pool in an account, the current role needs the CREATE COMPUTE POOL privilege on the account.”
↩︎ Producer and consumer roles on a Feature Store schema - 4.
“Alternatively, grant OWNERSHIP on the base table to the creating role, create the dynamic table (which auto-enables change tracking), and then transfer OWNERSHIP back.”
↩︎ Change tracking on source tables you do not own“Views that reference tables without change tracking also cause this error.”
↩︎ Change tracking on source tables you do not own“If another role owns the base tables, ask the owner to enable change tracking:”
↩︎ Exam trap 1 - 5.
“Enables suspending or resuming a compute pool.”
↩︎ Privileges for ML compute pools“Enables viewing compute pool usage (number of services and jobs running), properties, and listing compute pools”
↩︎ Privileges for ML compute pools - 6.https://docs.snowflake.com/en/developer-guide/snowflake-ml/ml-jobs/access-control-requirementsOfficial docs
“USAGE privilege on the compute pool to allow it to be used for ML workloads”
↩︎ Privileges for ML compute pools“CREATE SERVICE privilege on the schema to create and manage ML Jobs”
↩︎ Privileges for ML compute pools