We thought about this pattern for a while and ultimately decided we didn't want app compute to directly provision supporting infrastructure.
You end up in this weird place where your compute nodes can, at any time, assume super powerful roles that can change nacls, security groups, completely destroy database clusters or what have you.
It feels like running your app as root and database admin, but more. I guess if you completely trust your kube control plane?
We're trying it out in a dev-cluster/dev-account setup to let teams provision their own buckets/SQS and some supporting infrastructure easily without AWS access.
With properly configured RBAC this shouldn’t be an issue. The problem is making sure those policies are properly enforced. The tools are there, though.
Yeah I think this could probably work well in gcp.
In AWS you end up with a manually built/managed root cluster and if you've automated deploying / upgrading (which you probably should) that then you might as well just use those pipelines and code for all your other clusters too instead of having 1 special thing.
If the infrastructure gets complex there will start to be dependencies between stuff, there is definitely a tipping point where you need a real iac pipeline (tf/cf) instead in any case.
I think captn3mo proposed one of the more reasonable use cases which is provisioning stuff like queues and dbs for test envs or ci, we ended up stubbing out aws for ci and using the same pipeline code we use to deploy to prod instead for "full fat" testing but I can see where it might make sense to use kube manifests. Probably some creative uses for "deployment per customer" SaaSs too.
The app wouldn't build the infrastructure though, a controller on the cluster would. If you're using e.g. kube2iam or kiam, then only the controller would receive permissions. If you're in AWS and the controller uses CloudFormation to provision the resources, you can define a service role that CloudFormation will assume. Then your controller mainly just needs PassRole permissions for the service role, which means you don't need anything close to "root" privileges on your cluster.
Spin up a k8s cluster to manage your other resources that will be used by your... other k8s clusters. Is kubernetes just one big snake eating its own tail now?
Definitely cool... although AWS is really entrenched and Azure's getting there too for a lot of orgs. Especially historically MS oriented shops and those with some apps tethered to windows where others can be containerized.
K8s is in one sense a convergence engine, bringing container resources across a cluster into line with a given desired state.
So why not cloud resources too?
Having to run a cluster to do so doesn't feel quite right, however. But then if you're using Terraform - which is the most popular "cloud convergence engine" - you've probably provisioned compute resources with which to run Terraform. So not that strange.
I am really interested in using this GCP Operator and the corresponding ones from AWS and Azure to replace functionality from Open Service Brokers with a more Kubernetes-native experience.
Even better, this model uses the RBAC, namespaces, quotas and tools that developers are already using.
Provisioning cloud stuff ends up being about the relationships between all the things. Plus a bunch of dirty domain knowledge about retries and resource schemas.
It's very difficult to delete infra in AWS without a resource graph once you have enough things that refer to each other.
You can only get so far with a bunch of isolated yaml manifests for single resources.
So the MVP of anything that can solve the general case basically ends up doing the exact same things as terraform... and might as well actually be terraform.
25 comments
[ 3.2 ms ] story [ 70.6 ms ] threadIt is too bad because declarative specifications and control loops to have state converge to what it should be is a nice pattern.
Keeping up with all of AWS’s surface area though is hard if their heart isn’t in it.
You end up in this weird place where your compute nodes can, at any time, assume super powerful roles that can change nacls, security groups, completely destroy database clusters or what have you.
It feels like running your app as root and database admin, but more. I guess if you completely trust your kube control plane?
Like everything I'm sure it has a use somewhere.
In AWS you end up with a manually built/managed root cluster and if you've automated deploying / upgrading (which you probably should) that then you might as well just use those pipelines and code for all your other clusters too instead of having 1 special thing.
If the infrastructure gets complex there will start to be dependencies between stuff, there is definitely a tipping point where you need a real iac pipeline (tf/cf) instead in any case.
I think captn3mo proposed one of the more reasonable use cases which is provisioning stuff like queues and dbs for test envs or ci, we ended up stubbing out aws for ci and using the same pipeline code we use to deploy to prod instead for "full fat" testing but I can see where it might make sense to use kube manifests. Probably some creative uses for "deployment per customer" SaaSs too.
So much better to get an immediate feel for a service.
https://gardener.cloud/
So why not cloud resources too?
Having to run a cluster to do so doesn't feel quite right, however. But then if you're using Terraform - which is the most popular "cloud convergence engine" - you've probably provisioned compute resources with which to run Terraform. So not that strange.
Even better, this model uses the RBAC, namespaces, quotas and tools that developers are already using.
It's very difficult to delete infra in AWS without a resource graph once you have enough things that refer to each other.
You can only get so far with a bunch of isolated yaml manifests for single resources.
So the MVP of anything that can solve the general case basically ends up doing the exact same things as terraform... and might as well actually be terraform.