k8s_container monitored resource type requires cloud.platform: gcp_kubernetes_engine even when running outside of Google Cloud

[jsirianni]

In the past, we were able to send logs to Cloud Logging by setting the following resource attributes

• k8s.container.name
• k8s.pod.name
• k8s.namespace.name
• k8s.cluster.name

Recently, I noticed that my logs are no longer shown under the k8s_container monitored resource type, instead they appear as generic_node.

This is because the internal/resourcemapping/resourcemapping.go is checking specifically for cloud.platform: gcp_kubernetes_engine.

We work with many Google customers who operate Kubernetes clusters outside of Google Cloud, therefore, the requirement for cloud.platform: gcp_kubernetes_engine is surprising, especially when using OpenTelemetry.

I can reproduce this with Minikube running on my workstation.

With cloud.platform set

Using the resource attribute processor, I can set the cloud.platform value to "trick" the exporter.

without cloud.platform

If I do not add the attribute, my logs come through but as generic_node.

[damemi]

Hi @jsirianni, it looks like the line you're referring to has been part of the resource mapping for at least ~2 years based on the blame: https://github.com/GoogleCloudPlatform/opentelemetry-operations-go/blame/3a2c57209765f4752bd140a007612a81f58489c1/exporter/collector/monitoredresource.go#L131 (moved from that file to resourcemapping.go in f4bd197 ~1 year ago), so maybe it is a different change that has caused this behavior to show up recently for you. Do you know what version of these modules you were using when this worked, or if you've recently upgraded them?

I do agree that the implied dependence on the GCP platform attribute is weird, so I'll check if there's a specific reason we have that or if this could be opened up to other k8s platforms.

[damemi]

Talked to the team and it sounds like this may have possibly been due to us originally writing to gke_container. We can look more into this in an upcoming sprint but unless a hard reason comes up for why we are strictly relying on GKE then we will probably open this up.

[dashpole]

Seems like if it is mapped to any hosted kubernetes cloud.platform (e.g. EKS, AKS, etc), then we should map to k8s_container.

[damemi]

Opened #683 to fix this

[damemi]

This should be fixed in the next release with #683, and anything that sets k8s.cluster.name should be, at minimum, mapped to k8s_cluster. Also opened issues in js, python, and java repos to make sure the behavior matches.