Compute Instances
In addition to being able to create and curate "compute profiles" and "service profiles", admins with a PaaS Admin role can also launch and operate compute instances and services as a "managed service" for their users.
There are situations where administrators will need to deploy and operate compute and service instances on behalf of their end users. For example, an organization may want to provide their Data Scientists and ML Researchers with a "managed service" internally for MLOps. To deliver this, the administrator can deploy Rafay's Kubeflow based MLOps Platform (a service profile) and the "compute instance" needed to operate it.
Managed Compute Instance¶
To create a compute instance, access PaaS Studio and navigate to the home page. The page provides options to manage instances, including compute instances, which can be customized to meet specific requirements. Click on the Compute Instances menu on the left to access the compute instances page.
Click on New Compute Instance. A prompt will appear to specify the following in a wizard-like experience.
The workflow is identical to the end user workflows for a compute instance.
Info
A good example of a compute instance managed by an admin is a large, dedicated cluster with 100 GPUs meant for training of a LLM. Clusters like this need to have a high uptime and operational issues can result in significant delays to the training job.
Available Compute Instances¶
The list of compute Instances available for use with compute instances is displayed here. Choose the desired profile and fill in the necessary details
Unique Name¶
Provide a unique name for the service profile (e.g. gpu-instance-prod). Note that you cannot change the unique name once the profile has been created.
Display Name¶
This is typically a user friendly name such as "gpu-instance-prod" that will make sense to the end users. This name is what the end users will see an experience.
Description¶
Provide a multiple sentence description that will help the end users understand more details about the compute instance.
Compute Profile¶
When ready to proceed with the selected compute profile, click Save & Continue. Alternatively, click Change Compute Profile to select a different profile
The Overview page will display the Deploying Status as Waiting while the process is pending.
Labels¶
Labels are optional metadata in the form of key-value pairs that will be automatically attached to the resulting compute instances when they are launched. Labels help in categorizing and organizing compute instances based on specific criteria like environment, project, or team.
Annotations¶
Annotations are also optional metadata in key-value pair format, but unlike labels, they are typically used for descriptive purposes rather than for categorization or filtering. Annotations provide additional context or information about the compute instance.
- Users can either click Save Changes to save the configuration for later publishing, keeping the deployment status as Not Deployed
- Click Deploy to immediately proceed with publishing the changes
- On clicking Deploy, the instance deployment process begins, and the status updates to In Progress
- To add or edit the labels and annotations, click the Edit Configuration button. This will direct to the Overview page, where changes to labels and annotations can be made
Actions¶
The Actions in the instance view allow authorized users to view and execute actions configured within the environment template. Actions displayed here are derived from the associated environment template and reflect the configurations defined in the corresponding Compute Profile.
Each action appears with its variables as defined in the admin configuration. Variables marked as Allow Override can be modified during instance creation or execution, while variables with override disabled remain fixed and are not displayed during execution.
For example, if the override-allow action variable has the override option enabled, the authorized users can modify its value before executing the action. In contrast, variables such as not-override or run_strategy with the override option disabled will still appear in the form but remain non-editable during action execution.
In cases where override is allowed, the variable value can be updated (for example, modifying the run_strategy value in the Execute Action: start form) and then Apply can be clicked to execute the action with the updated input.
Viewing and Executing Actions
- The list of available actions is automatically displayed based on the selected profile.
- Only actions configured at the template level appear in this section.
- Each action includes its corresponding variables, displayed according to the alias and tooltip defined in the configuration.
- Variables marked as overridable can be edited before executing the action.
- Once configuration is verified, click Apply to trigger the action or Close to cancel.
- Click Deploy to start the instance deployment. The page displays the deployment details, with the status initially shown as
Pending, and then updated toIn Progress.
Once publishing is successful, the deployment status changes to Success
View Compute Instances¶
When PaaS Admins click on the compute instance menu on the left, the entire list of compute instances available are listed. They can search for the compute instances by providing input into the search box. Some basic information about a compute instance is shown to the admin in the main listing.
- Unique Name
- Creation Date
- Publish Status
- Sharing
- Actions Menu (ellipses)
To view additional details about a compute instance, simply click on it. This will display all the details the administrator had configured during initial creation.
Update Compute Instance¶
Follow the steps above for "View" compute instance. Alternatively, click on the actions (ellipses) on the far right of a compute instance and select "Edit Configuration". PaaS Admins can now make necessary changes to the compute instance and save it.
Info
Any changes will go into effect only for "new" services that are launched based on the profile. Already running instances will not be impacted.
View Instance Schedules¶
If the Compute Profile used to create a compute instance includes predefined schedules (for example, start and stop actions), these schedules are automatically displayed in the Compute Instance Deployment view.
In the example shown below, the instance computeinstance-scheduled includes two schedules:
- A start action scheduled to trigger every day at 9:00 AM.
- A stop action scheduled to trigger every day at 11:00 PM.
The next scheduled action is also displayed at the top, indicating the upcoming trigger time based on the cron expressions defined in the profile.
Delete Compute Instance¶
To delete a compute instance, click on the actions (ellipses) on the far right of a compute instance and select delete.
Info
This is a destruction action and cannot be reversed. Once a compute instance is deleted, it is no longer available to end users for use.
