Function Dependency

Function Dependency Policies

When using ModelArts to develop algorithms or manage training jobs, you are required to use other Cloud services. For example, before submitting a training job, select an OBS path for storing the dataset and logs, respectively. Therefore, when configuring fine-grained authorization policies for a user, the administrator must configure dependent permissions so that the user can use required functions.

If you use ModelArts as the root user (default IAM user with the same name as the account), the root user has all permissions by default.

**Table 1** Basic configuration
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Global configuration	IAM	iam:users:listUsers	Obtain a user list. This action is required by the administrator only.
Basic function	IAM	iam:tokens:assume	(Mandatory) Use an agency to obtain temporary authentication credentials.

**Table 2** Managing workspaces
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Workspace	IAM	iam:users:listUsers	Authorize workspaces by user.
Workspace	ModelArts	modelarts::delete	Clear resources in a workspace when the workspace is deleted.

**Table 3** Managing development environment notebook
Application Scenario	Dependent Service	Dependent Policy	Supported Function
DevEnviron	SWR	swr:repository:getNamespace (obtaining the details about an organization) swr:repository:listNamespace (querying the list of organizations) swr:repository:deleteTag (deleting an artifact version) swr:repository:getRepository (obtaining the details about artifact repositories) swr:repository:listTags (obtaining the version list of an artifact) swr:instance:createTempCredential (creating a temporary access credential)	Create a notebook instance using a custom image.
DevEnviron	MRS	mrs:cluster:get (obtaining details about a cluster)	Interconnect notebook with an MRS cluster.
DevEnviron	ECS	ecs:serverKeypairs:list (querying the list of SSH key pairs) ecs:serverKeypairs:get (obtaining ECS key pairs) ecs:serverKeypairs:delete (deleting an SSH key pair) ecs:serverKeypairs:create (creating and importing an SSH key pair)	Configure a login key for a notebook instance.
DevEnviron	IAM	iam:users:listUsers (querying users)	View the creators of notebook instances on the ModelArts console.
VS Code plug-in (on-premises)/PyCharm Toolkit (on-premises)	ModelArts	modelarts:notebook:update (updating the notebook development environment) modelarts:notebook:list (obtaining the list of notebook development environments) modelarts:notebook:start (starting a development environment instance) modelarts:notebook:stop (stopping a development environment instance) modelarts:notebook:get (obtaining the details about a notebook development environment)	Access a notebook instance on the cloud through on-premises VS Code.
VS Code plug-in (on-premises)/PyCharm Toolkit (on-premises)	OBS	obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:object:GetObject (obtaining object content and metadata) obs:object:GetObjectVersion (obtaining object content and metadata) obs:bucket:HeadBucket (obtaining bucket metadata)	Manage OBS data through local PyCharm.
PyCharm Toolkit (on-premises)	IAM	iam:projects:listProjects (querying tenant projects)	Obtain an IAM project list through local PyCharm for access configurations.
PyCharm Toolkit (on-premises)	ModelArts	modelarts:pool:list (viewing dedicated resource pools) modelarts:trainJob:list (viewing training job details) modelarts:trainJob:update (modifying a training job) modelarts:trainJobVersion:delete (deleting a training job version)	Use ModelArts through local PyCharm.
DevEnviron	AOM	aom:alarm:put (reporting alarms) aom:metric:get aom:metric:list aom:alarm:list (querying alarms)	Call the AOM API to obtain monitoring data and events of notebook instances and display them in ModelArts notebook.

**Table 4** Managing training jobs
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Training management	IAM	iam:credentials:listCredentials (querying permanent access keys) iam:agencies:listAgencies (querying agencies based on specified conditions)	Use the configured agency authorization.
	SFS	sfsturbo:shares:getShare (obtaining details about a file system) sfsturbo:shares:getAllShares (obtaining details about all file systems)	Use SFS in a training job.
	SWR	swr:repository:listTags (obtaining the version list of an artifact) swr:repository:getRepository (obtaining the details about artifact repositories) swr:repository:listRepositories (obtaining the list of artifact repositories)	Use a custom image to create a training job.
	SMN	smn:topic:publish (publishing a message) smn:topic:list (obtaining a topic)	Notify training job status changes through SMN.
	OBS	obs:bucket:ListAllMybuckets (obtaining a bucket list) obs:bucket:HeadBucket (obtaining bucket metadata) obs:bucket:ListBucket (listing objects in a bucket) obs:bucket:GetBucketLocation (obtaining the bucket location) obs:object:GetObject (obtaining object content and metadata) obs:object:GetObjectVersion (obtaining object content and metadata) obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:object:DeleteObject (deleting an object or batch deleting objects) obs:object:DeleteObjectVersion (deleting an object or batch deleting objects) obs:object:ListMultipartUploadParts (listing uploaded parts) obs:object:AbortMultipartUpload (aborting multipart uploads) obs:object:GetObjectAcl (obtaining an object ACL) obs:object:GetObjectVersionAcl (obtaining an object ACL) obs:bucket:PutBucketAcl (configuring a bucket ACL) obs:object:PutObjectAcl (configuring an object ACL) obs:object:ModifyObjectMetaData (modifying object metadata)	Run a training job using a dataset in an OBS bucket.
Federated training	IEF	ief:node:get (obtaining edge node information)	Run a federated learning-powered training job.

**Table 5** Using workflows
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Using a dataset	ModelArts	modelarts:dataset:getDataset modelarts:dataset:createDataset modelarts:dataset:createDatasetVersion modelarts:dataset:createImportTask modelarts:dataset:updateDataset modelarts:processTask:createProcessTask modelarts:processTask:getProcessTask modelarts:dataset:listDatasets	Use ModelArts datasets in a workflow.
Managing AI applications	ModelArts	modelarts:model:list modelarts:model:get modelarts:model:create modelarts:model:delete modelarts:model:update	Manage ModelArts AI applications in a workflow.
Deploying a service	ModelArts	modelarts:service:get modelarts:service:create modelarts:service:update modelarts:service:delete modelarts:service:getLogs	Manage ModelArts real-time services in a workflow.
Training jobs	ModelArts	modelarts:trainJob:get modelarts:trainJob:create modelarts:trainJob:list modelarts:trainJobVersion:list modelarts:trainJobVersion:create modelarts:trainJob:delete modelarts:trainJobVersion:delete modelarts:trainJobVersion:stop	Manage ModelArts training jobs in a workflow.
Workspace	ModelArts	modelarts:workspace:get modelarts:workspace:getQuotas	Use ModelArts workspaces in a workflow.
Managing data	OBS	obs:bucket:ListAllMybuckets (obtaining a bucket list) obs:bucket:HeadBucket (obtaining bucket metadata) obs:bucket:ListBucket (listing objects in a bucket) obs:bucket:GetBucketLocation (obtaining the bucket location) obs:object:GetObject (obtaining object content and metadata) obs:object:GetObjectVersion (obtaining object content and metadata) obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:object:DeleteObject (deleting an object or batch deleting objects) obs:object:DeleteObjectVersion (deleting an object or batch deleting objects) obs:object:ListMultipartUploadParts (listing uploaded parts) obs:object:AbortMultipartUpload (aborting multipart uploads) obs:object:GetObjectAcl (obtaining an object ACL) obs:object:GetObjectVersionAcl (obtaining an object ACL) obs:bucket:PutBucketAcl (configuring a bucket ACL) obs:object:PutObjectAcl (configuring an object ACL)	Use OBS data in a workflow.
Executing a workflow	IAM	iam:users:listUsers (querying users) iam:agencies:getAgency (obtaining details about a specified agency) iam:tokens:assume (obtaining an agency token)	Call other ModelArts services when the workflow is running.
Integrating DLI	DLI	dli:jobs:get (obtaining job details) dli:jobs:list_all (viewing a job list) dli:jobs:create (creating a job)	Integrate DLI into a workflow.
Integrating MRS	MRS	mrs:job:get (obtaining job details) mrs:job:submit (creating and executing a job) mrs:job:list (viewing a job list) mrs:job:stop (stopping a job) mrs:job:batchDelete (batch deleting jobs) mrs:file:list (viewing a file list)	Integrate MRS into a workflow.

**Table 6** Managing AI applications
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Managing AI applications	SWR	swr:repository:deleteRepository swr:repository:deleteTag swr:repository:getRepository swr:repository:listTags	Import a model from a custom image. Use a custom engine when importing a model from OBS.
Managing AI applications	OBS	obs:bucket:ListAllMybuckets (obtaining a bucket list) obs:bucket:HeadBucket (obtaining bucket metadata) obs:bucket:ListBucket (listing objects in a bucket) obs:bucket:GetBucketLocation (obtaining the bucket location) obs:object:GetObject (obtaining object content and metadata) obs:object:GetObjectVersion (obtaining object content and metadata) obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:object:DeleteObject (deleting an object or batch deleting objects) obs:object:DeleteObjectVersion (deleting an object or batch deleting objects) obs:object:ListMultipartUploadParts (listing uploaded parts) obs:object:AbortMultipartUpload (aborting multipart uploads) obs:object:GetObjectAcl (obtaining an object ACL) obs:object:GetObjectVersionAcl (obtaining an object ACL) obs:bucket:PutBucketAcl (configuring a bucket ACL) obs:object:PutObjectAcl (configuring an object ACL)	Import a model from a template. Specify an OBS path for model conversion.

**Table 7** Managing service deployment
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Deploying a service	LTS	lts:logs:list (querying the log list)	Show LTS logs.
Batch services	OBS	obs:object:GetObject (obtaining object content and metadata) obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:bucket:CreateBucket (creating a bucket) obs:bucket:ListBucket (listing objects in a bucket) obs:bucket:ListAllMyBuckets (obtaining a bucket list)	Create a batch service.
Edge services	CES	ces:metricData:list: (obtaining metric data)	View monitoring metrics.
Edge services	IEF	ief:deployment:delete (deleting a deployment)	Manage edge services.

**Table 8** Managing datasets
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Managing datasets and labels	OBS	obs:bucket:ListBucket (listing objects in a bucket) obs:object:GetObject (obtaining object content and metadata) obs:object:PutObject (uploading objects using PUT method, uploading objects using POST method, copying objects, appending an object, initializing a multipart task, uploading parts, and merging parts) obs:object:DeleteObject (deleting an object or batch deleting objects) obs:bucket:HeadBucket (obtaining bucket metadata) obs:bucket:GetBucketAcl (obtaining a bucket ACL) obs:bucket:PutBucketAcl (configuring a bucket ACL) obs:bucket:GetBucketPolicy (obtaining a bucket policy) obs:bucket:PutBucketPolicy (configuring a bucket policy) obs:bucket:DeleteBucketPolicy (deleting a bucket policy) obs:bucket:PutBucketCORS (configuring or deleting CORS rules of a bucket) obs:bucket:GetBucketCORS (obtaining the CORS rules of a bucket) obs:object:PutObjectAcl (configuring an object ACL)	Manage datasets in OBS. Label OBS data. Create a data management job.
Managing table datasets	DLI	dli:database:displayAllDatabases dli:database:displayAllTables dli:table:describe_table	Manage DLI data in a dataset.
Managing table datasets	DWS	dws:openAPICluster:list dws:openAPICluster:getDetail	Manage DWS data in a dataset.
Managing table datasets	MRS	mrs:job:submit mrs:job:list mrs:cluster:list mrs:cluster:get	Manage MRS data in a dataset.
Auto labeling	ModelArts	modelarts:service:list modelarts:model:list modelarts:model:get modelarts:model:create modelarts:trainJobInnerModel:list modelarts:workspace:get modelarts:workspace:list	Enable auto labeling.
Team labeling	IAM	iam:projects:listProjects (querying tenant projects) iam:users:listUsers (querying users) iam:agencies:createAgency (creating an agency) iam:quotas:listQuotasForProject (querying the quotas of a project)	Manage labeling teams.

**Table 9** Managing resources
Application Scenario	Dependent Service	Dependent Policy	Supported Function
Managing resource pools	BSS	bss:coupon:view bss:order:view bss:balance:view bss:discount:view bss:renewal:view bss:bill:view bss:contract:update bss:order:pay bss:unsubscribe:update bss:renewal:update bss:order:update	Create, renew, and unsubscribe from a resource pool. Dependent permissions must be configured in the IAM project view.
Managing resource pools	ECS	ecs:availabilityZones:list	Show AZs. Dependent permissions must be configured in the IAM project view.
Network management	VPC	vpc:routes:create vpc:routes:list vpc:routes:get vpc:routes:delete vpc:peerings:create vpc:peerings:accept vpc:peerings:get vpc:peerings:delete vpc:routeTables:update vpc:routeTables:get vpc:routeTables:list vpc:vpcs:create vpc:vpcs:list vpc:vpcs:get vpc:vpcs:delete vpc:subnets:create vpc:subnets:get vpc:subnets:delete vpcep:endpoints:list vpcep:endpoints:create vpcep:endpoints:delete vpcep:endpoints:get vpc:ports:create vpc:ports:get vpc:ports:update vpc:ports:delete vpc:networks:create vpc:networks:get vpc:networks:update vpc:networks:delete	Create and delete ModelArts networks, and interconnect VPCs. Dependent permissions must be configured in the IAM project view.
Network management	SFS Turbo	sfsturbo:shares:addShareNic sfsturbo:shares:deleteShareNic sfsturbo:shares:showShareNic sfsturbo:shares:listShareNics	Interconnect your network with SFS Turbo. Dependent permissions must be configured in the IAM project view.
Edge resource pool	IEF	ief:node:list ief:group:get ief:application:list ief:application:get ief:node:listNodeCert ief:node:get ief:IEFInstance:get ief:deployment:list ief:group:listGroupInstanceState ief:IEFInstance:list ief:deployment:get ief:group:list	Add, delete, modify, and search for edge pools

Agency authorization

To simplify operations when you use ModelArts to run jobs, certain operations are automatically performed on the ModelArts backend, for example, downloading the datasets in an OBS bucket to a workspace before a training job is started and dumping training job logs to the OBS bucket.

ModelArts does not save your token authentication credentials. Before performing operations on your resources (such as OBS buckets) in a backend asynchronous job, you are required to explicitly authorize ModelArts through an IAM agency. ModelArts will use the agency to obtain a temporary authentication credential for performing operations on your resources. For details, see Adding Authorization.

Figure 1 Agency authorization

As shown in Figure 1, after authorization is configured on ModelArts, ModelArts uses the temporary credential to access and operate your resources, relieving you from some complex and time-consuming operations. The agency credential will also be synchronized to your jobs (including notebook instances and training jobs). You can use the agency credential to access your resources in the jobs.

You can use either of the following methods to authorize ModelArts using an agency:

One-click authorization

ModelArts provides one-click automatic authorization. You can quickly configure agency authorization on the Global Configuration page of ModelArts. Then, ModelArts will automatically create an agency for you and configure it in ModelArts.

In this mode, the authorization scope is specified based on the preset system policies of dependent services to ensure sufficient permissions for using services. The created agency has almost all permissions of dependent services. If you want to precisely control the scope of permissions granted to an agency, use the second method.

Custom authorization

The administrator creates different agency authorization policies for different users in IAM, and configures the created agency for ModelArts users. When creating an agency for an IAM user, the administrator specifies the minimum permissions for the agency based on the user's permissions to control the resources that the user can access when they use ModelArts.

Risks in Unauthorized Operations

The agency authorization of a user is independent. Theoretically, the agency authorization scope of a user can be beyond the authorization scope of the authorization policy configured for the user group. Any improper configuration will result in unauthorized operations.

To prevent unauthorized operations, only a tenant administrator is allowed to configure agencies for users in the ModelArts global configuration to ensure the security of agency authorization.

Minimal Agency Authorization

When configuring agency authorization, an administrator must strictly control the authorization scope.

ModelArts asynchronously and automatically performs operations such as job preparation and clearing. The required agency authorization is within the basic authorization scope. If you use only some functions of ModelArts, the administrator can filter out the basic permissions that are not used according to the agency authorization configuration. Conversely, if you need to obtain resource permissions beyond the basic authorization scope in a job, the administrator can add new permissions to the agency authorization configuration. In a word, the agency authorization scope must be minimized and customized based on service requirements.

Basic Agency Authorization Scope

To customize the permissions for an agency, select permissions based on your service requirements.

**Table 10** Basic agency authorization for a development environment
Application Scenario	Dependent Service	Agency Authorization	Description
ModelArts SDK	OBS	obs:object:DeleteObject obs:object:GetObject obs:object:GetObjectVersion obs:object:PutObject obs:bucket:CreateBucket obs:bucket:ListBucket obs:bucket:HeadBucket	Access OBS through ModelArts SDKs.
	ModelArts	modelarts:dataset:listDatasets modelarts:dataset:createDataset modelarts:dataset:updateDataset modelarts:dataset:deleteDataset modelarts:dataset:getDataset modelarts:dataset:createDatasetVersion modelarts:dataset:deleteDatasetVersion modelarts:sample:listSamples modelarts:sample:addSamples modelarts:sample:deleteSamples modelarts:sample:getSample modelarts:dataset:createImportTask modelarts:dataset:createExportTask modelarts:image:get modelarts:image:register modelarts:notebook:get modelarts:image:create modelarts:pool:list modelarts:dataset:list modelarts:trainJob:create modelarts:trainJob:update modelarts:trainJob:delete modelarts:model:create modelarts:model:list modelarts:model:get modelarts:model:delete modelarts:service:create modelarts:service:list modelarts:service:get modelarts:service:delete	Use ModelArts SDKs to operate ModelArts.
	SWR	swr:repository:createNamespace swr:repository:listNamespaces swr:repository:getNamespace swr:repository:deleteRepository swr:instance:createTempCredential	Access SWR through ModelArts SDKs.
Algorithm suite	ModelArts	modelarts:aiAlgorithm:create	Use algorithm suites through ModelArts notebook.
JupyterLab plug-ins	OBS	obs:object:DeleteObject obs:object:GetObject obs:object:GetObjectVersion obs:bucket:CreateBucket obs:bucket:ListBucket obs:object:PutObject obs:bucket:GetBucketAcl obs:bucket:PutBucketAcl obs:bucket:PutBucketCORS	Use OBS to upload and download data in JupyterLab through ModelArts notebook.
DevEnviron monitoring	AOM	aom:alarm:put aom:metric:get aom:metric:list aom:alarm:list	Call the AOM API to obtain monitoring data and events of notebook instances and display them in ModelArts notebook.

**Table 11** Basic agency authorization for managing training jobs
Application Scenario	Dependent Service	Agency Authorization	Description
Training jobs	OBS	obs:bucket:ListBucket obs:object:GetObject obs:object:PutObject	Download data, models, and code before starting a training job. Upload logs and models when a training job is running.

**Table 12** Basic agency authorization for deploying services
Application Scenario	Dependent Service	Agency Authorization	Description
Real-time services	LTS	lts:groups:create lts:groups:list lts:topics:create lts:topics:delete lts:topics:list	Configure LTS for reporting logs of real-time services.
Batch services	OBS	obs:bucket:ListBucket obs:object:GetObject obs:object:PutObject	Run a batch service.
Edge services	IEF	ief:deployment:list ief:deployment:create ief:deployment:update ief:deployment:delete ief:node:createNodeCert ief:iefInstance:list ief:node:list	Deploy an edge service using IEF.

**Table 13** Basic agency authorization for managing data
Application Scenario	Dependent Service	Agency Authorization	Description
Dataset and data labeling	OBS	obs:object:GetObject obs:object:PutObject obs:object:DeleteObject obs:object:PutObjectAcl obs:bucket:ListBucket obs:bucket:HeadBucket obs:bucket:GetBucketAcl obs:bucket:PutBucketAcl obs:bucket:GetBucketPolicy obs:bucket:PutBucketPolicy obs:bucket:DeleteBucketPolicy obs:bucket:PutBucketCORS obs:bucket:GetBucketCORS	Manage datasets in an OBS bucket.
Labeling data	ModelArts inference	modelarts:service:get modelarts:service:create modelarts:service:update	Perform auto labeling based on ModelArts inference.

**Table 14** Basic agency authorization for managing dedicated resource pools
Application Scenario	Dependent Service	Agency Authorization	Description
Network management (new version)	VPC	vpc:routes:create vpc:routes:list vpc:routes:get vpc:routes:delete vpc:peerings:create vpc:peerings:accept vpc:peerings:get vpc:peerings:delete vpc:routeTables:update vpc:routeTables:get vpc:routeTables:list vpc:vpcs:create vpc:vpcs:list vpc:vpcs:get vpc:vpcs:delete vpc:subnets:create vpc:subnets:get vpc:subnets:delete vpcep:endpoints:list vpcep:endpoints:create vpcep:endpoints:delete vpcep:endpoints:get vpc:ports:create vpc:ports:get vpc:ports:update vpc:ports:delete vpc:networks:create vpc:networks:get vpc:networks:update vpc:networks:delete	Create and delete ModelArts networks, and interconnect VPCs. Dependent permissions must be configured in the IAM project view.
Network management (new version)	SFS Turbo	sfsturbo:shares:addShareNic sfsturbo:shares:deleteShareNic sfsturbo:shares:showShareNic sfsturbo:shares:listShareNics	Interconnect your network with SFS Turbo. Dependent permissions must be configured in the IAM project view.
Managing resource pools	ECS	ecs:availabilityZones:list	Show AZs. Dependent permissions must be configured in the IAM project view.

Agencies and Dependencies

Function Dependency

Agency authorization