This is the first of a few related changes that re-organize our API after the big recent changes that introduced the supervisor component.
Signed-off-by: Matt Moyer <moyerm@vmware.com>
- TLS certificates can be configured on the OIDCProviderConfig using
the `secretName` field.
- When listening for incoming TLS connections, choose the TLS cert
based on the SNI hostname of the incoming request.
- Because SNI hostname information on incoming requests does not include
the port number of the request, we add a validation that
OIDCProviderConfigs where the issuer hostnames (not including port
number) are the same must use the same `secretName`.
- Note that this approach does not yet support requests made to an
IP address instead of a hostname. Also note that `localhost` is
considered a hostname by SNI.
- Add port 443 as a container port to the pod spec.
- A new controller watches for TLS secrets and caches them in memory.
That same in-memory cache is used while servicing incoming connections
on the TLS port.
- Make it easy to configure both port 443 and/or port 80 for various
Service types using our ytt templates for the supervisor.
- When deploying to kind, add another nodeport and forward it to the
host on another port to expose our new HTTPS supervisor port to the
host.
- When two different Issuers have the same host (i.e. they differ
only by path) then they must have the same secretName. This is because
it wouldn't make sense for there to be two different TLS certificates
for one host. Find any that do not have the same secret name to
put an error status on them and to avoid serving OIDC endpoints for
them. The host comparison is case-insensitive.
- Issuer hostnames should be treated as case-insensitive, because
DNS hostnames are case-insensitive. So https://me.com and
https://mE.cOm are duplicate issuers. However, paths are
case-sensitive, so https://me.com/A and https://me.com/a are
different issuers. Fixed this in the issuer validations and in the
OIDC Manager's request router logic.
When using kind we forward the node's port to the host, so we only
really care about the `nodePort` value. For acceptance clusters,
we put an Ingress in front of a NodePort Service, so we only really
care about the `port` value.
Based on our experiences today with GKE, it will be easier for our users
to configure Ingress health checks if the healthz endpoint is available
on the same public port as the OIDC endpoints.
Also add an integration test for the healthz endpoint now that it is
public.
Also add the optional `containers[].ports.containerPort` to the
supervisor Deployment because the GKE docs say that GKE will look
at that field while inferring how to invoke the health endpoint. See
https://cloud.google.com/kubernetes-engine/docs/concepts/ingress#def_inf_hc
- New optional ytt value called `into_namespace` means install into that
preexisting namespace rather than creating a new namespace for each app
- Also ensure that every resource that is created statically by our yaml
at install-time by either app is labeled consistently
- Also support adding custom labels to all of those resources from a
new ytt value called `custom_labels`
Add install-pinniped-supervisor.yaml and rename install-pinniped.yaml
to install-pinniped-concierge.yaml in the release process and
installation/demo documentation.
- Tiltfile and prepare-for-integration-tests.sh both specify the
NodePort Service using `--data-value-yaml 'service_nodeport_port=31234'`
- Also rename the namespaces used by the Concierge and Supervisor apps
during integration tests running locally
- Also continue renaming things related to the concierge app
- Enhance the uninstall test to also test uninstalling the supervisor
and local-user-authenticator apps
This will hopefully come in handy later if we ever decide to add
support for multiple OIDC providers as a part of one supervisor.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
This needs to be overridden for Tilt usage, since the main image referenced by Tilt isn't actually pullable.
Signed-off-by: Matt Moyer <moyerm@vmware.com>
Right now in the YTT templates we assume that the agent pods are gonna use
the same image as the main Pinniped deployment, so we can use the same logic
for the image pull secrets.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
New resource naming conventions:
- Do not repeat the Kind in the name,
e.g. do not call it foo-cluster-role-binding, just call it foo
- Names will generally start with a prefix to identify our component,
so when a user lists all objects of that kind, they can tell to which
component it is related,
e.g. `kubectl get configmaps` would list one named "pinniped-config"
- It should be possible for an operator to make the word "pinniped"
mostly disappear if they choose, by specifying the app_name in
values.yaml, to the extent that is practical (but not from APIService
names because those are hardcoded in golang)
- Each role/clusterrole and its corresponding binding have the same name
- Pinniped resource names that must be known by the server golang code
are passed to the code at run time via ConfigMap, rather than
hardcoded in the golang code. This also allows them to be prepended
with the app_name from values.yaml while creating the ConfigMap.
- Since the CLI `get-kubeconfig` command cannot guess the name of the
CredentialIssuerConfig resource in advance anymore, it lists all
CredentialIssuerConfig in the app's namespace and returns an error
if there is not exactly one found, and then uses that one regardless
of its name
- For now, build the test-webhook binary in the same container image as
the pinniped-server binary, to make it easier to distribute
- Also fix lots of bugs from the first draft of the test-webhook's
`/authenticate` implementation from the previous commit
- Add a detailed README for the new deploy-test-webhook directory
- We are not setting an upper limit because Kubernetes might randomly
decide to unschedule our pod in ways that we can't anticipate in
advance, causing very hard to reproduce production bugs.
- We noticed that our app currently uses ~30 MB of memory when idle,
and ~35 MB of memory under some load. So a memory request of 128
MB should be reasonable.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
- Indicate the success or failure of the cluster signing key strategy
- Also introduce the concept of "capabilities" of an integration test
cluster to allow the integration tests to be run against clusters
that do or don't allow the borrowing of the cluster signing key
- Tests that are not expected to pass on clusters that lack the
borrowing of the signing key capability are now ignored by
calling the new library.SkipUnlessClusterHasCapability test helper
- Rename library.Getenv to library.GetEnv
- Add copyrights where they were missing
The rotation is forced by a new controller that deletes the serving cert
secret, as other controllers will see this deletion and ensure that a new
serving cert is created.
Note that the integration tests now have an addition worst case runtime of
60 seconds. This is because of the way that the aggregated API server code
reloads certificates. We will fix this in a future story. Then, the
integration tests should hopefully get much faster.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
This switches us back to an approach where we use the Pod "exec" API to grab the keys we need, rather than forcing our code to run on the control plane node. It will help us fail gracefully (or dynamically switch to alternate implementations) when the cluster is not self-hosted.
Signed-off-by: Matt Moyer <moyerm@vmware.com>
Co-authored-by: Ryan Richard <richardry@vmware.com>
- Call the auto-generated /healthz endpoint of our aggregated API server
- Use http for liveness even though tcp seems like it might be
more appropriate, because tcp probes cause TLS handshake errors
to appear in our logs every few seconds
- Use conservative timeouts and retries on the liveness probe to avoid
having our container get restarted when it is temporarily slow due
to running in an environment under resource pressure
- Use less conservative timeouts and retries for the readiness probe
to remove an unhealthy pod from the service less conservatively than
restarting the container
- Tuning the settings for retries and timeouts seem to be a mysterious
art, so these are just a first draft
- We want to follow the <noun>Request convention.
- The actual operation does not login a user, but it does retrieve a
credential with which they can login.
- This commit includes changes to all LoginRequest-related symbols and
constants to try to update their names to follow the new
CredentialRequest type.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
- For high availability reasons, we would like our app to scale linearly
with the size of the control plane. Using a DaemonSet allows us to run
one pod on each node-role.kubernetes.io/master node.
- The hope is that the Service that we create should load balance
between these pods appropriately.
- Refactors the existing cert generation code into controllers
which read and write a Secret containing the certs
- Does not add any new functionality yet, e.g. no new handling
for cert expiration, and no leader election to allow for
multiple servers running simultaneously
- This commit also doesn't add new tests for the cert generation
code, but it should be more unit testable now as controllers
- Previously the golang code would create a Service and an APIService.
The APIService would be given an owner reference which pointed to
the namespace in which the app was installed.
- This prevented the app from being uninstalled. The namespace would
refuse to delete, so `kapp delete` or `kubectl delete` would fail.
- The new approach is to statically define the Service and an APIService
in the deployment.yaml, except for the caBundle of the APIService.
Then the golang code will perform an update to add the caBundle at
runtime.
- When the user uses `kapp deploy` or `kubectl apply` either tool will
notice that the caBundle is not declared in the yaml and will
therefore avoid editing that field.
- When the user uses `kapp delete` or `kubectl delete` either tool will
destroy the objects because they are statically declared with names
in the yaml, just like all of the other objects. There are no
ownerReferences used, so nothing should prevent the namespace from
being deleted.
- This approach also allows us to have less golang code to maintain.
- In the future, if our golang controllers want to dynamically add
an Ingress or other objects, they can still do that. An Ingress
would point to our statically defined Service as its backend.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
- Why? Because the discovery URL is already there in the kubeconfig; let's
not make our lives more complicated by passing it in via an env var.
- Also allow for ytt callers to not specify data.values.discovery_url - there
are going to be a non-trivial number of installers of placeholder-name
that want to use the server URL found in the cluster-info ConfigMap.
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
- Seems like the next step is to allow override of the CA bundle; I didn't
do that here for simplicity of the commit, but seems like it is the right
thing to do in the future.
- Also includes bumping the api and client-go dependencies to the newer
version which also moved LoginDiscoveryConfig to the
crds.placeholder.suzerain-io.github.io group in the generated code
This is a somewhat more basic way to get access to the certificate and private key we need to issue short lived certificates.
The host path, tolerations, and node selector here should work on any kubeadm-derived cluster including TKG-S and Kind.
Signed-off-by: Matt Moyer <moyerm@vmware.com>
Add initial aggregated API server (squashed from a bunch of commits).
Signed-off-by: Andrew Keesler <akeesler@vmware.com>
Signed-off-by: Aram Price <pricear@vmware.com>
Signed-off-by: Ryan Richard <richardry@vmware.com>
- Also fix mistakes in the deployment.yaml
- Also hardcode the ownerRef kind and version because otherwise we get an error
Signed-off-by: Monis Khan <mok@vmware.com>
