Logs, alerts, and metrics

What is the log level number for every severity ?

1: Fatal

2: Error

3: Warn

4: Info

5: Debug

Is there a metric in CAST AI I can use to check the number of empty node in the cluster?

CAST AI deletes empty nodes, and all events are logged in the audit log. You can pull them with an API if needed.

ListAuditEntries returns audit entries for a given cluster; learn more about it here.

I added a panel with compute hourly costs to our Grafana board but I'm not sure how it should be interpreted?

These prices are the hourly expenses. To get a typical monthly cost, you need to multiply this number by 730.

I see the following log in the castai-cluster-controller, can you explain what it means

Log:

{ message:I0524 06:00:28.976409 1 request.go:682] Waited for 1.037042311s due to client-side throttling, not priority and fairness, request: GET:https://172.20.0.1:443/apis/vpcresources.k8s.aws/v1beta1 }

request.go is part of the Kubernetes REST API package. Kubernetes uses a discovery cache to populate the names of all API resources in the cluster, and this generates potentially 100s of GET requests behind the scenes. This log is most often seen concerning this behavior. The throttling is used to prevent kubelet/kube-apiserver from being overloaded with requests and degrading performance.

This pull request explains the issue in detail.

When I run Terraform, I get an error saying that role_arn is an unsupported attribute?

Log:

Error: Unsupported attribute on castai.tf line 18, in module "castai-eks-cluster-use1-1":   18:   aws_assume_role_arn      = module.castai-aws-iam-use1-1.role_arn  This object does not have an attribute named "role_arn".

role_arn should work with the module "castai-eks-iam-role". Similar to the usage in our examples:

aws_assume_role_arn = module.castai-eks-role-iam.role_arn

Learn more: Unsupported node affinity rules

The pod wasn't scheduled due to nodeSelector, why does the log say it failed on node affinity?

Log:

[autoscaler.cast.ai  Unsupported node affinity rules]

Node affinity and node selectors are essentially the same concept, just that affinity allows for conditional matching while selectors are a strict match.

If I apply Pod Disruption Budgets in a deployment, does CAST AI generate a log if the policy is not met and it cannot drain the node?

In this scenario, the drain would fail, and CAST AI would revert the node to active.

To prevent this issue, we recommend not using percentages for unavailable, but rather static numbers based on the min/max size of the deployment. For instance, with 10 replicas minimum having 3 unavailable would be acceptable. But in the case of having 2 or 3, 1 unavailable might be all it can tolerate.

I’m trying to add a Grafana panel that will calculate a number of CAST AI-managed spot instances over a period of time. Do those spot instances get terminated or does CAST AI drain and gracefully shut them down in advance?

Typically, CAST AI will drain and shut them down before the hard termination from AWS so long as the pods can move relatively quickly.

You can scrape our metrics which will show the CAST AI spot nodes and other types of nodes.

I am seeing a status of "WARNING" for my cluster. How can I determine the reason behind that warning?

Usually, a warning means, "CAST AI-managed cluster has encountered a transient error and is currently attempting to recover from it automatically. Autoscaling is not working."

Our system attempts an automatic recovery and only sends notifications if the recovery fails, resulting in a "Failed" status.

Learn more: Cluster status

Is it possible to send alert notification to email with CAST AI?

Currently, we only support notifications via the CAST AI console as a webhook for PagerDuty and Slack.

Does CAST AI rely in any shape or form on Datadog for scaling etc?

CAST AI doesn't rely on Datadog for scaling.

Sometimes, I have unscheduled pods. Is there a way to see those pods? What should I change in my Unscheduled pods policy to prevent that?

We recommend using kubectl to find the unscheduled pod and running kubectl describe to check the events for insight into why it's unscheduled. To see the unscheduled pods in your cluster, you can use the following command:

kubectl get pods --field-selector=status.phase=Pending

Does CAST AI take out pod spec and if so, which of its parts?

CAST AI takes all pod configuration YAML content, but we remove env variables to avoid sending passwords/keys, etc. The CAST AI agent running inside your cluster is in charge of doing this.

NodeSelectors, Affinity, Labels, CPU/Memory requests/limits, and PVC mounts - we use all that information either for reporting or making decisions, such as which AZ to create a node so that the pod gets scheduled.