Steve is a lightweight API proxy for Kubernetes whose aim is to create an interface layer suitable for dashboards to efficiently interact with Kubernetes.
Requests made to /api, /api/*, /apis/*, /openapi/* and /version will
be proxied directly to Kubernetes.
Steve registers all Kubernetes resources as schemas in the /v1 API. Any endpoint can support methods GET, POST, PATCH, PUT, or DELETE, depending on what the underlying Kubernetes endpoint supports and the user's permissions.
/v1/{type}- all cluster-scoped resources OR all resources in all namespaces of type{type}that the user has access to/v1/{type}/{name}- cluster-scoped resource of type{type}and unique name{name}/v1/{type}/{namespace}- all resources of type{type}under namespace{namespace}/v1/{type}/{namespace}/{name}- resource of type{type}under namespace{namespace}with name{name}unique within the namespace
Steve supports query parameters to perform actions or process data on top of what Kubernetes supports. In-depth, auto-generated API examples can be found in rancher.
Trigger a link handler, which is registered with the schema. Examples are calling the shell for a cluster, or following logs during cluster or catalog operations:
GET /v1/management.cattle.io.clusters/local?link=log
Trigger an action handler, which is registered with the schema. Examples are generating a kubeconfig for a cluster, or installing an app from a catalog:
POST /v1/catalog.cattle.io.clusterrepos/rancher-partner-charts?action=install
Only applicable to list requests (/v1/{type} and /v1/{type}/{namespace}).
Set the maximum number of results to retrieve from Kubernetes. The limit is passed on as a parameter to the Kubernetes request. The purpose of setting this limit is to prevent a huge response from overwhelming Steve and Rancher. For more information about setting limits, review the Kubernetes documentation on retrieving results in chunks.
The limit controls the size of the set coming from Kubernetes, and then filtering, sorting, and pagination are applied on that set. Because of this, if the result set is partial, there is no guarantee that the result returned to the client is fully sorted across the entire list, only across the returned chunk.
The returned response will include a continue token, which indicates that the
result is partial and must be used in the subsequent request to retrieve the
next chunk.
The default limit is 100000. To override the default, set limit=-1.
Only applicable to list requests (/v1/{type} and /v1/{type}/{namespace}).
Continue retrieving the next chunk of a partial list. The continue token is included in the response of a limited list and indicates that the result is partial. This token can then be used as a query parameter to retrieve the next chunk. All chunks have been retrieved when the continue field in the response is empty.
Only applicable to list requests (/v1/{type} and /v1/{type}/{namespace}).
Filter results by a designated field. Filter keys use dot notation to denote the subfield of an object to filter on. The filter value is matched as a substring.
Example, filtering by object name:
/v1/{type}?filter=metadata.name=foo
One filter can list multiple possible fields to match, these are ORed together:
/v1/{type}?filter=metadata.name=foo,metadata.namespace=foo
Stacked filters are ANDed together, so an object must match all filters to be included in the list.
/v1/{type}?filter=metadata.name=foo&filter=metadata.namespace=bar
Filters can be negated to exclude results:
/v1/{type}?filter=metadata.name!=foo
Arrays are searched for matching items. If any item in the array matches, the item is included in the list.
/v1/{type}?filter=spec.containers.image=alpine
Resources can also be filtered by the Rancher projects their namespaces belong
to. Since a project isn't an intrinsic part of the resource itself, the filter
parameter for filtering by projects is separate from the main filter
parameter. This query parameter is only applicable when steve is running in
concert with Rancher.
The list can be filtered by either projects or namespaces or both.
Filtering by a single project or a single namespace:
/v1/{type}?projectsornamespaces=p1
Filtering by multiple projects or namespaces is done with a comma separated list. A resource matching any project or namespace in the list is included in the result:
/v1/{type}?projectsornamespaces=p1,n1,n2
The list can be negated to exclude results:
/v1/{type}?projectsornamespaces!=p1,n1,n2
Only applicable to list requests (/v1/{type} and /v1/{type}/{namespace}).
Results can be sorted lexicographically by primary and secondary columns.
Sorting by only a primary column, for example name:
/v1/{type}?sort=metadata.name
Reverse sorting by name:
/v1/{type}?sort=-metadata.name
The secondary sort criteria is comma separated.
Example, sorting by name and creation time in ascending order:
/v1/{type}?sort=metadata.name,metadata.creationTimestamp
Reverse sort by name, normal sort by creation time:
/v1/{type}?sort=-metadata.name,metadata.creationTimestamp
Normal sort by name, reverse sort by creation time:
/v1/{type}?sort=metadata.name,-metadata.creationTimestamp
Only applicable to list requests (/v1/{type} and /v1/{type}/{namespace}).
Results can be batched by pages for easier display.
Example initial request returning a page with 10 results:
/v1/{type}?pagesize=10
Pages are one-indexed, so this is equivalent to
/v1/{type}?pagesize=10&page=1
To retrieve subsequent pages, the page number and the list revision number must be included in the request. This ensures the page will be retrieved from the cache, rather than making a new request to Kubernetes. If the revision number is omitted, a new fetch is performed in order to get the latest revision. The revision is included in the list response.
/v1/{type}?pagezie=10&page=2&revision=107440
The total number of pages and individual items are included in the list
response as pages and count respectively.
If a page number is out of bounds, an empty list is returned.
page and pagesize can be used alongside the limit and continue
parameters supported by Kubernetes. limit and continue are typically used
for server-side chunking and do not guarantee results in any order.
Steve is typically imported as a library. The calling code starts the server:
import (
"fmt"
"context"
"github.com/rancher/steve/pkg/server"
"github.com/rancher/wrangler/v2/pkg/kubeconfig"
)
func steve() error {
restConfig, err := kubeconfig.GetNonInteractiveClientConfigWithContext("", "").ClientConfig()
if err != nil {
return err
}
ctx := context.Background()
s, err := server.New(ctx, restConfig, nil)
if err != nil {
return err
}
fmt.Println(s.ListenAndServe(ctx, 9443, 9080, nil))
return nil
}steve can be run directly as a binary for testing. By default it runs on ports 9080 and 9443:
export KUBECONFIG=your.cluster
go run main.goThe API can be accessed by navigating to https://localhost:9443/v1.
Steve's main use is as an opinionated consumer of rancher/apiserver, which it uses to dynamically register every Kubernetes API as its own. It implements apiserver Stores to use Kubernetes as its data store.
Steve uses apiserver Stores to transform and store data, mainly in Kubernetes. The main mechanism it uses is the proxy store, which is actually a series of four nested stores and a "partitioner". It can be instantiated by calling NewProxyStore. This gives you:
proxy.errorStore- translates any returned errors into HTTP errorsproxy.WatchRefresh- wraps the nested store's Watch method, canceling the watch if access to the watched resource changespartition.Store- wraps the nested store's List method and parallelizes the request according to the given partitioner, and additionally implements filtering, sorting, and pagination on the unstructured data from the nested storeproxy.rbacPartitioner- the partitioner fed to thepartition.Storewhich allows it to parallelize requests based on the user's access to certain namespaces or resourcesproxy.Store- the Kubernetes proxy store which performs the actual connection to Kubernetes for all operations
The default schema additionally wraps this proxy store in
metrics.Store,
which records request metrics to Prometheus, by calling
metrics.NewMetricsStore
on it.
Steve provides two additional exported stores that are mainly used by Rancher's catalogv2 package:
selector.Store- wraps the list and watch commands with a label selector
switchschema.Store- transforms the object's schema
Steve watches all Kubernetes API resources, including built-ins, CRDs, and APIServices, and registers them under its own /v1 endpoint. The component responsible for watching and registering these schemas is the schema controller. Schemas can be queried from the /v1/schemas endpoint. Steve also registers a few of its own schemas not from Kubernetes to facilitate certain use cases.
Steve creates a fake local cluster to use in standalone scenarios when there is not a real clusters.management.cattle.io resource available. Rancher overrides this and sets its own customizations on the cluster resource.
User preferences in steve provides a way to configure dashboard preferences
through a configuration file named prefs.json. Rancher overrides this and
uses the
preferences.management.cattle.io
resource for preference storage instead.
Counts keeps track of the number of resources and updates the count in a buffered stream that the dashboard can subscribe to.
Steve exposes a websocket endpoint on /v1/subscribe for sending streams of events. Connect to the endpoint using a websocket client like websocat:
websocat -k wss://127.0.0.1:9443/v1/subscribeReview the apiserver guide for details.
In addition to regular Kubernetes resources, steve allows you to subscribe to special steve resources. For example, to subscribe to counts, send a websocket message like this:
{"resourceType":"count"}
Existing schemas can be customized using schema templates. You can customize individual schemas or apply customizations to all schemas.
For example, if you wanted to customize the store for secrets so that secret data is always redacted, you could implement a store like this:
import (
"github.com/rancher/apiserver/pkg/store/empty"
"github.com/rancher/apiserver/pkg/types"
)
type redactStore struct {
empty.Store // must override the other interface methods as well
// or use a different nested store
}
func (r *redactStore) ByID(_ *types.APIRequest, _ *types.APISchema, id string) (types.APIObject, error) {
return types.APIObject{
ID: id,
Object: map[string]string{
"value": "[redacted]",
},
}, nil
}
func (r *redactStore) List(_ *types.APIRequest, _ *types.APISchema) (types.APIObjectList, error) {
return types.APIObjectList{
Objects: []types.APIObject{
{
Object: map[string]string{
"value": "[redacted]",
},
},
},
}, nil
}and then create a schema template for the schema with ID "secrets" that uses that store:
import (
"github.com/rancher/steve/pkg/schema"
)
template := schema.Template{
ID: "secret",
Store: &redactStore{},
}You could specify the same by providing the group and kind:
template := schema.Template{
Group: "", // core resources have an empty group
Kind: "secret",
Store: &redactStore{},
}then add the template to the schema factory:
schemaFactory.AddTemplate(template)As another example, if you wanted to add custom field to all objects in a collection response, you can add a schema template with a collection formatter to omit the ID or the group and kind:
template := schema.Template{
Customize: func(schema *types.APISchema) {
schema.CollectionFormatter = func(apiOp *types.APIRequest, collection *types.GenericCollection) {
schema.CollectionFormatter = func(apiOp *types.APIRequest, collection *types.GenericCollection) {
for _, d := range collection.Data {
obj := d.APIObject.Object.(*unstructured.Unstructured)
obj.Object["tag"] = "custom"
}
}
}
}
}Steve implements access control on schemas based on the user's RBAC in Kubernetes.
The apiserver
Server
object exposes an AccessControl field which is used to customize how access
control is performed on server requests.
An
accesscontrol.AccessStore
is stored on the schema factory. When a user makes any request, the request
handler first finds all the schemas that are available to the user. To do this,
it first retrieves an
accesscontrol.AccessSet
by calling
AccessFor
on the user. The AccessSet contains a map of resources and the verbs that can
be used on them. The AccessSet is calculated by looking up all of the user's
role bindings and cluster role bindings for the user's name and group. The
result is cached, and the cached result is used until the user's role
assignments change. Once the AccessSet is retrieved, each registered schema is
checked for existence in the AccessSet, and filtered out if it is not
available.
This final set of schemas is inserted into the
types.APIRequest
object and passed to the apiserver handler.
Steve authenticates incoming requests using a customizable authentication middleware. The default authenticator in standalone steve is the AlwaysAdmin middleware, which accepts all incoming requests and sets admin attributes on the user. The authenticator can be overridden by passing a custom middleware to the steve server:
import (
"context"
"github.com/rancher/steve/pkg/server"
"github.com/rancher/steve/pkg/auth"
"k8s.io/apiserver/pkg/authentication/user"
)
func run() {
restConfig := getRestConfig()
authenticator := func (req *http.Request) (user.Info, bool, error) {
username, password, ok := req.BasicAuth()
if !ok {
return nil, false, nil
}
if username == "hello" && password == "world" {
return &user.DefaultInfo{
Name: username,
UID: username,
Groups: []string{
"system:authenticated",
},
}, true, nil
}
return nil, false, nil
}
server := server.New(context.TODO(), restConfig, &server.Options{
AuthMiddleware: auth.ToMiddlware(auth.AuthenticatorFunc(authenticator)),
}
server.ListenAndServe(context.TODO(), 9443, 9080, nil)
}Once the user is authenticated, if the request is for a Kubernetes resource, then steve must proxy the request to Kubernetes, so it needs to transform the request. Steve passes the user Info object from the authenticator to a proxy handler, either a generic handler or an impersonating handler. The generic Handler mainly sets transport options and cleans up the headers on the request in preparation for forwarding it to Kubernetes. The ImpersonatingHandler uses the user Info object to set Impersonate-* headers on the request, which Kubernetes uses to decide access.
Steve is designed to be consumed by a graphical user interface and therefore serves one by default, even in the test server. The default UI is the Rancher Vue UI hosted on releases.rancher.com. It can be viewed by visiting the running steve instance on port 9443 in a browser.
The UI can be enabled and customized by passing options to
NewUIHandler.
For example, if you have an alternative index.html file, add the file to
a directory called ./ui, then create a route that serves a custom UI handler:
import (
"net/http"
"github.com/rancher/steve/pkg/ui"
"github.com/gorilla/mux"
)
func routes() http.Handler {
custom := ui.NewUIHandler(&ui.Options{
Index: func() string {
return "./ui/index.html"
},
}
router := mux.NewRouter()
router.Handle("/hello", custom.IndexFile())
return routerIf no options are set, the UI handler will serve the latest index.html file from the Rancher Vue UI.
The cluster cache keeps watches of all resources with registered schemas. This is mainly used to update the summary cache and resource counts, but any module could add a handler to react to any resource change or get cached cluster data. For example, if we wanted a handler to log all "add" events for newly created secrets:
import (
"context"
"github.com/rancher/steve/pkg/server"
"k8s.io/apimachinery/pkg/runtime"
"github.com/sirupsen/logrus"
"k8s.io/apimachinery/pkg/runtime/schema"
)
func logSecretEvents(server *server.Server) {
server.ClusterCache.OnAdd(context.TODO(), func(gvk schema.GroupVersionKind, key string, obj runtime.Object) error {
if gvk.Kind == "Secret" {
logrus.Infof("[event] add: %s", key)
}
return nil
})
}Rancher uses a concept called "aggregation" to maintain connections to remote services. Steve implements an aggregation client in order to allow connections from Rancher and expose its API to Rancher.
Aggregation is enabled by defining a secret name and namespace in the steve server:
import (
"context"
"github.com/rancher/steve/pkg/server"
)
func run() {
restConfig := getRestConfig()
server := server.New(context.TODO(), restConfig, &server.Options{
AggregationSecretNamespace: "cattle-system",
AggregationSecretName: "stv-aggregation",
})
server.ListenAndServe(context.TODO(), 9443, 9080, nil)
}This prompts the steve server to start a controller that watches for this secret. The secret is expected to contain two pieces of data, a URL and a token:
$ kubectl -n cattle-system get secret stv-aggregation -o yaml
apiVersion: v1
data:
token: Zm9vYmFy
url: aHR0cHM6Ly8xNzIuMTcuMC4xOjg0NDMvdjMvY29ubmVjdA==
kind: Secret
metadata:
...Steve makes a websocket connection to the URL using the token to authenticate. When the secret changes, the steve aggregation server restarts with the up-to-date URL and token.
Through this websocket connection, the steve agent is exposed on the remote management server and the management server can route steve requests to it. The management server can also keep track of the availability of the agent by detecting whether the websocket session is still active. In Rancher, the connection endpoint runs on /v3/connect.
Rancher implements aggregation for other types of services as well. In Rancher, the user can define endpoints via a v3.APIService custom resource (which is distinct from the built-in Kubernetes v1.APIService resource). Then Rancher runs a middleware handler that routes incoming requests to defined endpoints. The external services follow the same process of using a defined secret containing a URL and token to connect and authenticate to Rancher. This aggregation is defined independently and does not use steve's aggregation client.
Steve supports query parameters filter, sort, page/pagesize/revision,
and projectsornamespaces for list requests as described
above. These formatting options exist to allow user
interfaces like dashboards to easily consume and display list data in a
friendly way.
This feature relies on the concept of stores and the RBAC partitioner. The proxy store provides raw access to Kubernetes and returns data as an unstructured.UnstructuredList. The partitioner calls the proxy store in parallel for each segment of resources the user has access to, such as for each namespace. The partitioner feeds the results of each parallelized request into a stream of unstructured.Unstructured. From here, the list is passed to the listprocessor to filter, sort, and paginate the list. The partition store formats the list as a types.APIObjectList and it is returned up the chain of nested stores.
Most stores in steve are implementations of the apiserver Store interface, which returns apiserver types. The partitioner implements its own store type called UnstructuredStore which returns unstructured.Unstructured objects. The reason for this is that the filtering and sorting functions in the listprocessor package need to operate on unstructured data because they work on arbitrary fields. However, it also needs to be run after the parallelized partitioner has accumulated all the results, because each concurrent fetcher will only contain partial results. Therefore, the data remains in an unstructured format until after the listprocessor has been run, then the data is converted to a structured type. The below diagram illustrates the conversion sequence.
The unit tests for these API features are located in two places:
pkg/stores/partition/listprocessor/processor_test.go contains tests for each individual query handler. All changes to listprocessor should include a unit test in this file.
pkg/stores/partition/store_test.go
contains tests for the List operation of the partition store. This is
especially important for testing the functionality for multiple partitions. It
also tests all supported query parameters, not limited to the
pagination-related ones, and tests them in combination with one another. Tests
should be added here when:
- the change is related to partitioning
- the change is related to parsing the query parameters
- the change is related to the
limitorcontinueparameters - the listprocessor change should be tested with other query parameters
It doesn't hurt to add a test here for any other listprocessor change.
Each table test runs several requests, so they are effectively each a bundle of
tests. Each table test has a list of apiOps which each specify the request
and the user running it, a list of access maps which declares the users
corresponding to each request and controls the
AccessSet
the user has, the partitions the users have access to, and the objects in
each partition. The requests in apiOps are run sequentially, and each item in
the lists want, wantCache, and wantListCalls correlate to the expected
results and side effects of each request. partitions and objects apply to
all requests in the table test.
Integration tests for the steve API are located among the rancher integration tests. See the documentation included there for running the tests and using them to generate API documentation.