A user has one role per tenant. The same person can be a DEVELOPER in tenant A and a VIEWER in tenant B.

Roles

RoleForIn one line
VIEWERPeople who only look at resultsEvery screen, read-only
DEVELOPERML engineers, data staffDatasets, training, experiments, model registration, up to Staging
OPERATOROperations staffWhat DEVELOPER does + deployment, serving, alert handling, pipelines
APPROVERQuality or operations ownersView + approve or reject, promote to Production
ADMINTenant administratorsEvery permission + members, tokens, tenant settings, rollback

Permission table

PermissionWhat it allowsVIEWERDEVELOPEROPERATORAPPROVERADMIN
ViewView every screen read-only✅✅✅✅✅
DevelopDEVELOPER · OPERATOR · ADMINCreate training runs, log and stop runs, issue MLflow tokens✅✅✅
Register modelDEVELOPER · OPERATOR · ADMINRegister a run's result as a new model version✅✅✅
Promote to StagingDEVELOPER · OPERATOR · ADMINMove a model version between stages such as Staging and Archived✅✅✅
Promote to ProductionAPPROVER · ADMINRequest promotion to Production (it takes effect after approval). Demoting a live Production version also needs this permission✅✅
ApproveAPPROVER · ADMINApprove or reject promotion requests✅✅
DeployOPERATOR · ADMINDeploy and serve models, build serving images✅✅
OperateOPERATOR · ADMINOperational actions: acknowledge and resolve alerts, run and stop pipelines, delete training runs✅✅
Roll backADMINRecord a deployment as rolled back (devices must pull the previous version again; see Roll back a deployment)✅
Write datasetsDEVELOPER · OPERATOR · ADMINCreate datasets, upload and delete files✅✅✅
SettingsADMINManage members, tokens and tenant settings in the admin console✅

Global administrator

Global administrator status is granted separately from tenant roles. It creates and deletes tenants, manages every user and uses the system console (cluster, GPUs, build cache, weights mirror, Prometheus and so on). A global administrator can access every tenant. In tenants where they are not a member they act as ADMIN; where they hold a direct membership, that membership's role applies. Grant it only to those who need it.

Example flow — promoting to Production

Promotion to Production: ① v3, registered by the AI developer, is in Staging. ② An approver requests promotion with a reason, rollback conditions and scope, which creates an approval request. ③ When an approver approves, v3 becomes Production and the former Production v2 moves to Archived automatically. Safety-critical models need two approvers, and a rejection notifies the requester with the reason

  1. A DEVELOPER registers the training result as a model version (automatically from the wizard, or from Experiments (실험)). New versions go to Staging. They summarise the validation results for the approvers.
  2. An APPROVER (or ADMIN) presses Request production promotion on the model version and enters a reason. The version does not become Production yet — an approval request is created and appears under Approvals (승인).
  3. Someone with the approve permission reviews it and presses Approve: the version becomes Production and the previous Production version moves to Archived. Reject notifies the requester with the reason.

Written for the platform as of 2026-10-02.

© MLOps Central Platform