Adds the design proposal for self-hosting authz webhook.

This proposal implements feature request:
  kubernetes/enhancements#516

contributors/design-proposals/auth/self-hosting-authorization-webhook.md

@@ -0,0 +1,337 @@
+# Self-hosting authorization webhooks
+
+Status: Pending
+
+Version: **Alpha**
+
+Implementation Owner: @filmil
+
+## Motivation
+
+See: kubernetes/features#516 kubernetes/kubernetes#54733
+kubernetes/kubernetes#54163
+
+Make it possible to have the authorization (hereafter authz for short) webhook
+run as a pod in the cluster being authorized. This addresses part of the
+concerns raised in kubernetes/kubernetes#52511.
+
+This requires a change in the way the `kube-apiserver` webhook configuration is
+specified (a new configuration format), and a special initialization of the
+webhook module in `kube-apiserver` (in contrast to any other apiservers out
+there) as [outlined in an earlier discussion on this topic][9].
+
+As of this writing, the authorization webhook in Kubernetes is configured
+through the `kube-apiserver` flag `--authorization-webhook-config-file` which is
+a [kubeconfig][3]-formatted [configuration file][2].  An example taken from the
+documentation:
+
+```yaml
+# clusters refers to the remote service.
+clusters:
+  - name: name-of-remote-authz-service
+    cluster:
+      # CA for verifying the remote service.
+      certificate-authority: /path/to/ca.pem
+      # URL of remote service to query. Must use 'https'. May not include parameters.
+      server: https://authz.example.com/authorize
+
+# users refers to the API Server's webhook configuration.
+users:
+  - name: name-of-api-server
+    user:
+      client-certificate: /path/to/cert.pem # cert for the webhook plugin to use
+      client-key: /path/to/key.pem          # key matching the cert
+
+# kubeconfig files require a context. Provide one for the API Server.
+current-context: webhook
+contexts:
+- context:
+    cluster: name-of-remote-authz-service
+    user: name-of-api-server
+  name: webhook
+```
+
+The kubeconfig file format does not easily admit reference to an endpoint that
+is hosted within the cluster.  Earlier proposals to extend the syntax of the
+`clusters[*].cluster.server` field to admit a custom dialer that calls a service
+endpoint [were rejected][4] (also see [another rationale][5]).  The [suggested
+viable alternative][6] is to provide a new webhook file format that references
+a service explicitly, and build on top of that.
+
+One further concern to address is that this proposed change is only [relevant
+for the `kube-apiserver`][7], and not to any of the generic apiservers.  This
+means that the proposed changes will be set up such that only the
+`kube-apiserver` can make use of it.  This approach has [garnered support][8]
+from the community in the earlier proposal.
+
+## Proposal
+
+* Define a new configuration format, supported only by `kube-apiserver` that
+  admits a webhook endpoint within the cluster to be specified.
+
+  This new configuration file is still specified using the same flag as for the
+  original webhook configuration, `--auhtorization-webhook-config-file`.
+
+* Instantiate a custom subject access review client based on this configuration
+  that has special behavior as follows:
+
+  - If the new configuration has been specified, a custom dialer is supplied
+    which resolves a service endpoint and uses that service endpoint to dial
+    into when making a call.
+
+## User Experience
+
+### Use Cases
+
+1. As a cluster administrator, I would like to install Kubernetes-based products
+   that have self-contained deployments in a cluster and include authorization.
+
+   Kubernetes products typically come with an install script, say based on Helm
+   or `kubeadm`.  These allow custom deployments, including reconfiguring
+   `kube-apiserver`.  When such a product requires an authorizer, the authorizer
+   is added as a system component with a pod-based deployment in a cluster and
+   the apiserver is instructed to consult the endpoint based on that deployment
+   for authorization decisions.
+
+   The cluster administrator experience changes in that they are able to install
+   a wider range of Kubernetes-based products that include this feature,
+   compared to when this is not the case.
+
+## Implementation
+
+
+### Configuration format
+
+Let us define a new configuration file format.  This is based on the [admission
+configuration file format][12] as well as the proposal from @frankfarzan for an
+analogous [configuration format for authz][13].
+
+An example configuration file for authz looks like this:
+
+```yaml
+apiVersion: authorizationconfig.k8s.io/v1alpha1
+kind: ExternalAuthorizationHookConfiguration
+metadata:
+  name: example-config
+externalAuthorizationHooks:
+- name: webhook-name
+  clientConfig:
+    serverCaFile: ... # the path to the CA file for the webhook server
+    clientCertificateFile: ... # path to the cert file for the webhook plugin to use
+    clientKeyFile: ... # path to the private key matching the cert
+    # Optional.  When specified, the result of the service resolution for this
+    # endpoint is substituted in place of the Host portion of the url field
+    # below.
+    service:
+      name: some-service # name of the front-end service
+      namespace: some-namespace #<namespace of the front-end service>
+    url: <URL of the remote webhook>
+```
+
+The go data definition matching the file above is as follows.  This captures
+both the service reference and a URL.
+
+```go
+// Required well-known import statements elided for brevity.
+
+// Top-level
+type ExternalAuthorizationHookConfiguration struct {
+  meta.TypeMeta
+  meta.ObjectMeta
+
+  ExternalAuthorizationHooks map[string]ExternalAuthorizationHook
+}
+
+type ExternalAuthorizationHook struct {
+
+  Name string
+  ClientConfig ClientConfig
+}
+
+type ClientConfig struct {
+  ServerCaFile string
+  ClientCertificateFile string
+  ClientKeyFile string
+  // +optional
+  Service ServiceReference
+  // https only
+  Url string
+}
+
+type ServiceReference struct {
+  Name string
+  Namespace string
+}
+```
+
+When `service` is omitted, then the `url` alone is used.  If service is defined,
+the endpoint obtained as result of service resolution is substituted for the
+`Host` in the url.
+
+Example:
+
+```yaml
+# Elided irrelevant fields.
+externalAuthorizationHooks:
+- name: webhook-name
+  clientConfig:
+    service:
+      name: some-service
+      namespace: some-namespace
+    url: https:/some/endpoint/path
+```
+
+This would make a request to `https://(endpointOf name=some-service,
+namespace=some-namespace)/some/endpoint/path`.
+
+
+### Parsing the configuration file
+
+This will edit kube-apiserver `config.go` parsing to first try parsing the
+authz configuration file as new format.  If the new format is found, new
+behavior is invoked.  Otherwise, the old code paths are reused.
+
+While strictly speaking parsing the file twice, looking for different formats is
+not as efficient as parsing it once looking for both formats, it has other
+benefits:
+
+- It does not require changing the old code path which uses a deep integration
+  with the generic webhook client code.  By inspection it seems that unifying
+  the two would be a sizable undertaking, and does not seem to make sense from a
+  cost/benefit perspective.
+
+- The amortized cost of this extra read is zero, since the read only happens at
+  `kube-apiserver` start-up.
+
+### Wiring through extra information
+
+To construct a service-aware dialer for the new code path, we need to wire
+through a proxy-aware transport and the service resolver from the
+`kube-apiserver`.  This will be done by packaging the two components into an
+interface provided by the webhook library as follows, to decouple the
+two implementations:
+
+```go
+// In server.go, kube-apiserver:
+type resolver struct {
+  proxyTransport *http.Transport
+  resolver ServiceResolver
+}
+func (r *resolver) NewDialer(namespace, name string) func (net.Conn, error) {
+  // ...
+}
+
+// In webhook.go:
+type DialerFactory interface {
+  func NewDialer(namespace, name string) func (net.Conn, error)
+}
+```
+
+This dialer will be subsituted when creating a REST client.
+
+
+### Wiring things together and testing
+
+The wire-through and testing will follow the approach already outlined in
+the PR kubernetes/kubernetes#54733 ([link][14]).
+
+### Server Backwards/Forwards compatibility
+
+The `kube-apiserver` will continue to support the kubeconfig file format for
+configuring authorization for as long as it is required, and will retain the
+respective behavior.
+
+The new file format will be versioned to allow a smooth transition between
+apiserver versions.
+
+One will, of course, not be able to specify a new configuration file format
+to a `kube-apiserver` that was compiled without the new file format support.
+
+None of the above changes seem to be disruptive to the daily use of the
+`kube-apiserver` in deployments.  Only the clients that want to use the new
+feature need ever know that it exists.
+
+## Alternatives considered
+
+This is a list of rejected alternatives for specific components of the proposal.
+Each one is contrasted to its currently accepted solution from the proposal
+above.
+
+### Extending kubeconfig format
+
+Versus: a new configuration format.
+
+Feedback from the community has indicated that extending the kubeconfig file
+format is not acceptable.  Concerns raised were:
+
+- Service resolution based on the hostname in a URL is not the direction that
+  Kubernetes evolves towards.
+
+- Kubeconfig files already have a widely accepted use, and extending them to
+  make services first-class citizens is a non-goal.
+
+### Using a new configuration flag
+
+Versus: overloading the configuration flag with a new meaning.
+
+More flags mean a larger configuration space.  It seems beneficial to avoid
+introducing new flags if possible.  Reusing a flag also makes it obvious that
+there may only be one authz configuration active at a time.
+
+### Using dynamic authorizator configuration
+
+Versus: static authorizator configuration in a file.
+
+[Concerns were raised][11] about the ability of the cluster admin to
+misconfigure a cluster, or undo a previously established hosted configuration
+if dynamic configuration is used.
+
+This proposal side-steps that concern as it still relies on the use of a
+configuration file, which is out of reach of cluster admins in hosted solutions.
+At the same time, it does not prevent a future change that would, if required,
+admit dynamic configuration.
+
+### Using PEM-encoded configuration in the configuration file
+
+Versus: specifying filesystem paths to the certificates.
+
+Specifying the filesystem paths allows the authz configuration file to remain
+the same across deployments, while only the mounted contents of the secret
+varies where it applies.
+
+Contrast that to self-registration, where certificates and keys would have to be
+passed along as inline content in the API request to register a webhook.
+
+Since we are currently not considering dynamic registration (see previous
+subsection), we can accept only having fields that point at filesystem paths.
+This decision does not prevent changing the direction in a followup version, say
+by allowing filesystem references and inline content to coexist.
+
+### Extending generic apiserver
+
+Versus: extending kube-apiserver only.
+
+Only the `kube-apiserver` has the ability to resolve service endpoints without
+special configuration.  Also, `kube-apiserver` is the only apiserver that has
+the use case described in this document.
+
+Therefore, it seems reasonable to confine the behavior changes to
+`kube-apiserver` alone.
+
+### Unifying code paths for webhooks
+
+
+[1]: https://github.com/kubernetes/kubernetes/issues/52511
+[2]: https://kubernetes.io/docs/admin/authorization/webhook/
+[3]: https://kubernetes.io/docs/tasks/access-application-cluster/configure-access-multiple-clusters/
+[4]: https://github.com/kubernetes/kubernetes/pull/54889#issuecomment-343045279
+[5]: https://github.com/kubernetes/kubernetes/pull/54733#issuecomment-343160937
+[6]: https://github.com/kubernetes/kubernetes/pull/54733#issuecomment-343292540
+[7]: https://github.com/kubernetes/kubernetes/issues/52511#issuecomment-331489326
+[8]: https://github.com/kubernetes/kubernetes/issues/52511#issuecomment-333224542
+[9]: https://github.com/kubernetes/kubernetes/issues/52511#issuecomment-333541769
+[10]: https://github.com/kubernetes/kubernetes/issues/52511#issuecomment-329803092
+[11]: https://github.com/kubernetes/kubernetes/issues/52511#issuecomment-329803092
+[12]: https://kubernetes.io/docs/admin/extensible-admission-controllers/#configure-webhook-admission-controller-on-the-fly
+[13]: https://github.com/kubernetes/kubernetes/pull/54733#issuecomment-343287758
+[14]: https://github.com/kubernetes/kubernetes/pull/54733