Pinniped v0.5.0: Now With Even More Pinnipeds
Feb 4, 2021
Pinniped is a “batteries included” authentication system for Kubernetes clusters that tightly integrates with Kubernetes using native API patterns. Pinniped is built using custom resource definitions (CRDs) and API aggregation, both of which are core to the configuration and runtime operation of the app.
We encountered a problem that’s familiar to many Kubernetes controller developers: we need to support multiple instances of our controller on one cluster.
You may have a similar need for several reasons, such as:
- Soft Multi-Tenancy: several teams share a cluster and they each want to manage their own instance of a controller.
- Scaling: you have outgrown the vertical scaling limit for your controller and would like to shard it along some dimension that’s easy to operate and reason about.
- Backwards Compatibility: you want to deploy two versions of your controller and provide a window of time for consumers to smoothly upgrade to the new version.
- Controller Development: you want to run, for example, the stable and alpha versions of your controller on the same cluster. Most cluster users will only rely on the stable version, but some test workloads will use the alpha version.
With Pinniped v0.5.0, we wanted to be able to bundle an opinionated configuration of Pinniped into our downstream commercial products while also allowing our customers to install their own Pinniped instance and configure it however they like.
This post describes how we approached the need for multiple Pinnipeds in v0.5.0.
For many Kubernetes controllers, there are existing best practices that will work well:
Add a “controller class” field: the most well-known example of this pattern is the
spec.ingressClassNamefield in the
networking.k8s.io/v1Ingress resource (formerly the
This field tags a particular object so that only the designated controller instance will watch it. This means that you must configure all the participating controllers to do the proper filtering and ignore any resources that they’re not intended to manage.
Use API versioning machinery: the other key technique is to strictly adhere to Kubernetes API contracts and take advantage of Kubernetes versioning machinery. Your CRD can have multiple versions and you can write a webhook to handle gracefully converting between versions so that several versions of your controller can co-exist peacefully.
These two techniques are sufficient for many situations but have some limitations. If your app uses Kubernetes API aggregation then a controller class annotation may not be sufficient, since each version of your API group must be registered with a single APIService resource.
Even in a purely CRD-based app, the CRD definition and associated webhook conversion service can only be defined once for each API type. At a minimum, this requires that you carefully manage the deployment of these resources. For example, in the soft multi-tenancy use case several teams must coordinate to deploy these singleton resources.
Building and maintaining webhook conversion functionality also carries a cost, especially if you need to handle many versions worth of version skew.
Our solution is to have a single controller codebase where the names of all the API groups can be adjusted via configuration.
This is controlled via a new
--api-group-suffix flag on the Pinniped server commands.
When unset, Pinniped defaults to the
pinniped.dev API group, which is the “true” name we use in our API definitions and generated code.
When a user deploys Pinniped with a custom API group suffix such as
--api-group-suffix=pinniped1.example.com, several new behaviors are triggered:
Outgoing Controller Requests: throughout our controller code, we use a consistent set of Kubernetes clients via the
go.pinniped.dev/internal/kubeclientpackage. These clients use
k8s.io/client-go/rest#Config.Wrapto inject a custom
http.RoundTripperthat can act as a client middleware layer.
For each outbound request from our controller API clients, the RoundTripper applies a set of transformations:
- It decodes the request from JSON/Protobuf.
- It rewrites the request’s
apiVersionto match the configured API group suffix.
- It renames other API group references in well-known object fields such as
- It re-encodes the request for wire transport and passes it along to the server.
- It decodes the response from JSON/Protobuf.
- It apply the inverse renaming operation to reverse step three and restore the default API group suffix (
- Finally, it re-encodes the response and passes it back to the client.
Steps 5-7 must also handle the case of streaming response to a
Incoming Aggregated API Requests: our aggregated API server is built using the
k8s.io/apiserver/pkg/serverpackage. We have only a single aggregated API called TokenCredentialRequest, and we were able to get the functionality we needed by creating a custom
k8s.io/apimachinery/pkg/runtime#Schemethat registers our API kinds under the custom group (in
With this configuration, all the builtin functionality of the generic API server works correctly.
Requests and responses are unmarshaled and marshalled correctly, and the OpenAPI discovery API even serves the custom API group names.
App-Specific Code: the Pinniped concierge server dynamically updates the TokenCredentialRequest APIService to rotate its TLS certificate authority bundle. This code had to become aware of the dynamic API group, but it was as easy as wiring through a new parameter from the CLI flag (see
With this system in place, we’ve achieved our goal. A user can deploy several instances of Pinniped, each interacting only with its own distinct set of API objects.
The default behavior of Pinniped remains unchanged, and we made sure to implement the changes such that they cause little to no overhead when no custom API group has been configured.
Advantages and Disadvantages
With v0.5.0, each instance of Pinniped can be upgraded and operated 100% independently, with no coordination or shared state needed. One remaining constraint is that each instance should be deployed into its own namespace. This ensures that any other standard Kubernetes objects such as Secrets and ConfigMaps referenced by the configuration do not overlap.
Our middleware solution carries some ongoing costs:
- It took a non-trivial amount of code to implement all the required transformations. We now have the maintenance burden of ensuring this code continues to work in future versions of the Kubernetes API machinery.
- Other API consumers (including
kubectlusers) need to know which API group to use. This might be as simple as knowing to run
kubectl get jwtauthenticators.authentication.concierge.team1.example.cominstead of simply
kubectl get jwtauthenticators. There is no builtin upgrade path between these versions, as there would be with a versioned CRD.
- The extra encoding/decoding steps cause some performance impact when this feature is in use. None of the Pinniped APIs are used in high throughput use cases, so this was not much a problem for us.
We’re happy to have shipped this for Pinniped v0.5.0, but we have more ideas about how to extend the concept. One idea is to extract the renaming middleware we’ve written for Pinniped into a standalone Go library that other Kubernetes apps can adopt.
We could also take this a step further and extract the behavior of our middleware into an out-of-process API proxy that can apply these transformations to an unmodified Kubernetes app. This would require major changes and it would be challenging to support some features seamlessly, such as Protobuf encoding.
As a team, we have no immediate plans for either of these ideas, but if you are interested please reach out in GitHub.
With the release of v0.10.0, Pinniped now supports Kubernetes clusters behind firewalls or in restricted environments
With the release of v0.7.0, Pinniped now supports a much wider range of real-world Kubernetes clusters, including managed Kubernetes environments on all major cloud providers.
With the release of v0.11.0, Pinniped offers CRDs for easy Active Directory configuration, OIDC password grant flow for CLI workflows, and Distroless images for security and performance