Release notes

Evictor – you can now install our algorithm that continuously bin packs pods into the smallest number of nodes via Autoscaler policy (in UI and Terraform). Previously, users had to follow a documented guide and install it via the command line. Please note: if you have Evictor already installed and configured it will continue to run, even though in the Autoscaler page it might indicate that it can’t be enabled. In order to correct this you would need to remove current Evictor installation manually and enable it from the Autoscaler page.

‘Cost per CPU’ reporting dimension was added to Available Savings report, as well as Node list and Rebalancing screens. This cost is calculated by dividing the compute cost by the number of provisioned CPUs. It is also exposed as a scrapeable metric for the whole cluster or per-instance life cycle: spot, on-demand, fallback. A full list of currently available metrics and the setup guide are available here.

We reacted to community feedback and improved the user interface and experience of our Cost report. The user interface of the Autoscaler policies page was also uplifted.

Added functionality that allows users to remove a team member from the organization.

Added the concept of a “Project” to the CAST AI console. Previously, if users had clusters with the same name across different GCP projects (or Azure Resource groups, AWS accounts) in the CAST AI console, there was no way to differentiate between these clusters. Now each cluster record also indicates the name of the GCP project / Azure Resource group / AWS account ID.

Node list view now displays the total node CPU and Memory capacity instead of allocatable values. We have also fixed an issue preventing the status of the cordoned node to be accurately presented in the node list.

We have updated the calculation formula for the root volume that is added to each CAST AI provisioned node. Before the change nodes could have had 100 GiB disks as a minimum or a larger disk based on CPU to Storage (GiB) ratio. This ratio couldn’t have been less than 1 CPU: 25 GiB. Now, a 100 GiB disk is a base and we add additional storage based on the CPU to Storage ratio, which can be as low as 1 CPU : 0 GiB.

For GKE clusters, the logic of the ‘Node constraints’ setting in the Autoscaler policy is now more flexible. We have removed CPU to RAM ratio validations so users can choose more flexible configurations.

For AWS EKS users who are using AWS CNI, we have improved the autoscaler logic, so when considering what node to add, autoscaler would react to an error if the target subnet is full and choose another available subnet (if any).

For AWS EKS users who are using CSI driver for Amazon EBS we have added the support for topology.ebs.csi.aws.com/zone label in the autoscaler so the new node is created in the correct zone, respecting the specification of the storage class.

We have optimized permissions required to onboard and run CAST AI in EKS clusters, more information can be found in the documentation.

If an EKS cluster is being onboarded using Terraform we now allow passing AWS credentials via environment variables to the agent helm chart.

Added Terraform examples for onboarding EKS cluster using custom IAM policies, creating EKS cluster with NAT Gateway and Application load balancer. All examples can be found here.

The latest version of the CAST AI agent is v0.27, the list of changes can be found here. To update the agent in your cluster, follow these steps.

The latest version of our Terraform provider is v0.23, the list of changes can be found here. The provider and the modules can be found in the registry.

Added support for ssh public keys. Using the updated cluster configuration API, users can set the public key (base64 encoded) or, in the case of AWS, also use the AWS key pair ID ("key-0123456789EXAMPLE") and connect to CAST AI provisioned nodes.

Spot nodes can now be interrupted directly from the node list. The interrupted node will change its status and eventually be removed from the cluster while the new spot node is provisioned instead. The whole process takes a few minutes to complete.

Added numerical values of requested and allocatable resources in CPU and GiB to the detailed node view in the Node list.

We have added an additional sheet that lists all workloads and their CPU & RAM requests in the Excel extract of the Available savings report.

If Evictor fails the leader election process, it will now restart automatically. Previously, users may have encountered a situation where the leader election process has failed, causing Evictor to fail silently.

“Unscheduled pods policy applied” audit log event JSON now has more context about what information was considered and led to the addition of a specific node, i.e. which nodes were skipped and why, which workload triggered the autoscaling event, what were the node constraints, etc. This feature greatly improves transparency into the decision-making process employed by CAST AI.

Introduced label selectors in the mutating webhook configuration, so customers can control the webhook in a much more flexible manner. Previously, users could set which pods should be scheduled on the on-demand nodes using regular expression values (namespaces). Now, they can use label selectors to force (or ignore) some pods to run on spot nodes (ability to force some pods to run on on-demand nodes based on a namespace remains).

Added another scrapable cluster metric – the hourly compute cost per pricing type (spot, on-demand, fallback). Check the documentation for more details.

Introduced a separate screen for the dashboard of the disconnected cluster.

Uplifted the design of console menu items.