Version 2.17

Release Content - April 14, 2024¶

• Deprecation notifications
• Breaking changes

Researcher¶

Scheduler¶

• Added functionality to configure over provisioning ratios for node pools running any kind of workload. Over provisioning assumes that workloads are either under utilizing or intermittently using GPUs. This indicates that the real utilization is lower than the actual GPU allocation requested. Over provisioning allows the administrator to condense more workloads on a single GPU than what the workload required. For more information, see Optimize performance with Node Level Scheduler.

• Added the GPU Resource Optimization feature to the UI. Now you can enable and configure GPU Portion (Fraction) limit and GPU Memory Limit from the UI. For more information, see Compute resources UI with Dynamic Fractions.

• Added the ability to set Run:ai as the default scheduler for any project or namespace. This provides the administrator the ability to ensure that all workloads in a project or namespace are scheduled using the Run:ai scheduler. For more information, see Setting Run:ai as default scheduler.

Jobs, Workloads, and Workspaces¶

• Added to the workload details view, the ability to filter by pod. You can now filter metrics and logs per pod or all the pods. Also, the Workloads table now has additional columns including connections and preemtability adding more at a glance information about the workload. In addition, using the Copy & edit button, you can submit a new workload via CLI based on the selected workload. For more information, see Workloads.

• Added Inference to workload types. Inference workloads can now be created and managed from the unified Workloads table. The Deployments workload type has been deprecated, and replaced with Inference workloads which are submitted using the workload form. For more information, see Inference and for submitting an Inference workload, see Submitting workloads.

• Added functionality that supports a single workloads submission selection. Now you can submit workloads by pressing + New workloads in the Workloads table. You can submit the following workloads from this table:

  • Workspace
  • Training
  • Inference

  This improvement phases out the previous version's Workspace and Jobs tables. The Jobs table and submission forms have been deprecated and can be reactivated. To reenable the Jobs table and forms, press Tools & settings, then General, then Workloads, and then Toggle the Jobs view and the Jobs submission buttons. For more information, see Submitting workloads.

• Added the ability to configure a Kubernetes readiness probe. The readiness probe detects resources and workloads that are ready to receive traffic.

Assets¶

• Added the capability to use a ConfigMap as a data source. The ability to use a ConfigMap as a data source can be configured in the Data sources UI, the CLI, and as part of a policy. For more information, see Setup a ConfigMap as a data source, Setup a ConfigMap as a volume using the CLI, or Setup a ConfigMap Resource description fields in training policies.

• Added a Status column to the Credentials table, and the Data sources table. The Status column displays the state of the resource and provides troubleshooting information about that asset. For more information, see the Credentials table and the Data sources table.

• Added functionality for asset creation that validates the asset based on version compatibility of the cluster or the control plane within a specific scope. At time of asset creation, invalid scopes will appear greyed out and will show a pop-up with the reason for the invalidation. This improvement is designed to increase the confidence that an asset is created properly and successfully.

Run:ai Administrator¶

Configuration and Administration¶

• Introducing a new Tools & Settings menu. The new Tools & Settings menu provides a streamlined UI for administrators to configure the Run:ai environment. The new UI is divided into categories that easily identify the areas where the administrator can change settings. The new categories include:

  • Analytics—features related to analytics and metrics.
  • Resources—features related to resource configuration and allocation.
  • Workloads—features related to configuration and submission of workloads.
  • Security—features related to configuration of SSO (Single Sign On).
  • Notifications—used for system notifications.
  • Cluster authentication—snippets related to Researcher authentication.

  Some features are now labeled either Experimental or Legacy. Experimental features are new features in the environment, that may have certain instabilities and may not perform as expected. Legacy features are features that are in the process of being deprecated, and may be removed in future versions.

Clusters¶

• Added new columns to the Clusters table to show Kubernetes distribution and version. This helps administrators view potential compatibility issues that may arise.

• Improved the location of the cluster filter. The cluster filter has been relocated to filter bar and the drop down cluster filter in the header of the page has been removed. This improvement creates the following:

  • Filter assets by cluster in the following tables:

    • Data sources
    • Environments
    • Computer resources
    • Templates
    • Credentials

  • Creating a new asset, will automatically display only the scope of the selected cluster.

  • Prevention of account (top most level in the Scope) from being selected when creating assets.
  • Enforcement a cluster specific scope. This increases the confidence that an asset is created properly and successfully.

  Note

  This feature is only applicable if the all the clusters are version 2.17 and above.

Monitoring and Analytics¶

• Improved GPU Overview dashboard. This improvement provides rich and extensive GPU allocation and performance data and now has interactive tiles that provide direct links to the Nodes, Workloads, and Departments tables. Hover over tiles with graphs to show rich data in the selected time frame filter. Tiles with graphs can be downloaded as CSV files. The new dashboard is enabled by default. Use the Go back to legacy view to return to the previous dashboard style. For more information, see Dashboard analysis.

• Updated the knative and autoscaler metrics. Run:ai currently supports the following metrics:

  • Throughput
  • Concurrency

  For more information, see Autoscaling metrics.

• Improved availability of metrics by using Run:ai APIs. Using the API endpoints is now the preferred method to retrieve metrics for use in any application. For more information, see Metrics.

Authentication and Authorization¶

• Added new functionality to SAML 2.0 identity provider configuration in the Security category of the General settings. The added functionality assists with troubleshooting SSO configuration and authentication issues that may arise. Now administrators now have the ability to:

  • View and edit the identity provider settings for SAML 2.0
  • Upload or download the SAML 2.0 identity provider metadata XML file.

For more information, see SSO UI configuration.

Deprecation Notifications¶

Deprecation notifications allow you to plan for future changes in the Run:ai Platform.

Feature deprecations¶

Deprecated features will be available for two versions ahead of the notification. For questions, see your Run:ai representative. The following features have been marked for deprecation:

• Jobs—the Jobs feature (submission form and view) has been moved to the category of Legacy. To enable them, go to Tools & Settings, General, open the Workloads pane, and then toggle the Jobs view and Job submission switch to the enabled position.
• Deployments—the Deployments feature has been removed. It has been replaced by Inference workloads. For more information, see Jobs, Workloads, and Workspaces above.
• Workspaces view—the Workspaces menu has been removed. You can now submit a Workspace workload using the + New workload form from the Workloads table.

API support and endpoint deprecations¶

The endpoints and parameters specified in the API reference are the ones that are officially supported by Run:ai. For more information about Run:ai's API support policy and deprecation process, see Developer overview.

Deprecated APIs and API fields¶

The following list of API endpoints and fields that have been marked for deprecation:

Jobs and Pods API¶

Clusters API¶

Authorization and Authentication API¶

Projects API¶

Departments API¶

Policy API¶

Removed APIs and API fields (completed deprecation)¶

The following list of API endpoints and fields that have completed their deprecation process and therefore will be changed as follows:

Assets API¶

Metrics deprecations¶

The following metrics are deprecated and replaced by API endpoints. For details about the replacement APIs, see Changed Metrics:

Breaking changes¶

Breaking changes notifications allow you to plan around potential changes that may interfere your current workflow when interfacing with the Run:ai Platform.

Metrics¶

Be aware that some names of metrics have been changed. For more information, see Changed Metrics.