Note. transl.: the author of the article - Erkan Erol, an engineer from SAP - shares his study of the mechanisms of the functioning of the team kubectl exec, so familiar to everyone who works with Kubernetes. He accompanies the entire algorithm with listings of the Kubernetes source code (and related projects), which allow you to understand the topic as deeply as necessary.
One Friday, a colleague approached me and asked how to execute a command in a pod using client-go. I could not answer him and suddenly realized that I knew nothing about the mechanism of work kubectl exec. Yes, I had certain ideas about his device, but I was not 100% sure of their correctness and therefore decided to address this issue. Having studied blogs, documentation and source code, I learned a lot, and in this article I want to share my discoveries and understanding. If something is wrong, please contact me at Twitter.
Prepare
To create a cluster on a MacBook, I cloned ecomm-integration-ballerina/kubernetes-cluster. Then I corrected the IP addresses of the nodes in the kubelet'a config, since the default settings did not allow kubectl exec. You can read more about the main reason for this here.
Any machine = my MacBook
master node IP = 192.168.205.10
IP worker node = 192.168.205.11
API server port = 6443
Components
kubectl exec process: when we do "kubectl exec..." the process is started. You can do this on any machine with access to the K8s API server. Note. transl.: Further in the console listings, the author uses the comment "any machine", implying that the following commands can be executed on any such machines with access to Kubernetes.
api server: A component on the master node that provides access to the Kubernetes API. This is the front end for the control plane in Kubernetes.
kublet: agent that runs on every node in the cluster. It provides the work of containers in the pod.
container runtime (container runtime): The software responsible for running containers. Examples: Docker, CRI-O, containerdβ¦
kernel: OS kernel on the worker node; responsible for process management.
target (target) Containers: a container that is part of a pod and runs on one of the worker nodes.
What did I discover
1. Activity on the client side
Create a pod in a namespace default:
// any machine
$ kubectl run exec-test-nginx --image=nginx
Then we execute the exec command and wait 5000 seconds for further observations:
// any machine
$ kubectl exec -it exec-test-nginx-6558988d5-fgxgg -- sh
# sleep 5000
The kubectl process appears (with pid=8507 in our case):
We can also observe the request on the api-server side:
handler.go:143] kube-apiserver: POST "/api/v1/namespaces/default/pods/exec-test-nginx-6558988d5-fgxgg/exec" satisfied by gorestful with webservice /api/v1
upgradeaware.go:261] Connecting to backend proxy (intercepting redirects) https://192.168.205.11:10250/exec/default/exec-test-nginx-6558988d5-fgxgg/exec-test-nginx?command=sh&input=1&output=1&tty=1
Headers: map[Connection:[Upgrade] Content-Length:[0] Upgrade:[SPDY/3.1] User-Agent:[kubectl/v1.12.10 (darwin/amd64) kubernetes/e3c1340] X-Forwarded-For:[192.168.205.1] X-Stream-Protocol-Version:[v4.channel.k8s.io v3.channel.k8s.io v2.channel.k8s.io channel.k8s.io]]
Note that the HTTP request includes a protocol change request. SPDY allows separate "streams" of stdin/stdout/stderr/spdy-error to be multiplexed over a single TCP connection.
The API server receives the request and converts it to PodExecOptions:
// PodExecOptions is the query options to a Pod's remote exec call
type PodExecOptions struct {
metav1.TypeMeta
// Stdin if true indicates that stdin is to be redirected for the exec call
Stdin bool
// Stdout if true indicates that stdout is to be redirected for the exec call
Stdout bool
// Stderr if true indicates that stderr is to be redirected for the exec call
Stderr bool
// TTY if true indicates that a tty will be allocated for the exec call
TTY bool
// Container in which to execute the command.
Container string
// Command is the remote command to execute; argv array; not executed within a shell.
Command []string
}
To perform the required actions, the api-server needs to know which pod it needs to contact:
// ExecLocation returns the exec URL for a pod container. If opts.Container is blank
// and only one container is present in the pod, that container is used.
func ExecLocation(
getter ResourceGetter,
connInfo client.ConnectionInfoGetter,
ctx context.Context,
name string,
opts *api.PodExecOptions,
) (*url.URL, http.RoundTripper, error) {
return streamLocation(getter, connInfo, ctx, name, opts, opts.Container, "exec")
}
Of course, the data about the endpoint is taken from the information about the node:
nodeName := types.NodeName(pod.Spec.NodeName)
if len(nodeName) == 0 {
// If pod has not been assigned a host, return an empty location
return nil, nil, errors.NewBadRequest(fmt.Sprintf("pod %s does not have a host assigned", name))
}
nodeInfo, err := connInfo.GetConnectionInfo(ctx, nodeName)
Hooray! Kubelet now has a port (node.Status.DaemonEndpoints.KubeletEndpoint.Port) to which the API server can connect:
// GetConnectionInfo retrieves connection info from the status of a Node API object.
func (k *NodeConnectionInfoGetter) GetConnectionInfo(ctx context.Context, nodeName types.NodeName) (*ConnectionInfo, error) {
node, err := k.nodes.Get(ctx, string(nodeName), metav1.GetOptions{})
if err != nil {
return nil, err
}
// Find a kubelet-reported address, using preferred address type
host, err := nodeutil.GetPreferredNodeAddress(node, k.preferredAddressTypes)
if err != nil {
return nil, err
}
// Use the kubelet-reported port, if present
port := int(node.Status.DaemonEndpoints.KubeletEndpoint.Port)
if port <= 0 {
port = k.defaultPort
}
return &ConnectionInfo{
Scheme: k.scheme,
Hostname: host,
Port: strconv.Itoa(port),
Transport: k.transport,
}, nil
}
These connections are terminated at the kubelet's HTTPS endpoint. By default, apiserver does not verify the kubelet's certificate, which makes the connection vulnerable to "man-in-the-middle attacks" (MITM) and unsafe to work in untrusted and/or public networks.
Now the API server knows the endpoint and establishes a connection:
// Connect returns a handler for the pod exec proxy
func (r *ExecREST) Connect(ctx context.Context, name string, opts runtime.Object, responder rest.Responder) (http.Handler, error) {
execOpts, ok := opts.(*api.PodExecOptions)
if !ok {
return nil, fmt.Errorf("invalid options object: %#v", opts)
}
location, transport, err := pod.ExecLocation(r.Store, r.KubeletConn, ctx, name, execOpts)
if err != nil {
return nil, err
}
return newThrottledUpgradeAwareProxyHandler(location, transport, false, true, true, responder), nil
}
First, find out the IP of the working node. In our case, this is 192.168.205.11:
// any machine
$ kubectl get nodes k8s-node-1 -o wide
NAME STATUS ROLES AGE VERSION INTERNAL-IP EXTERNAL-IP OS-IMAGE KERNEL-VERSION CONTAINER-RUNTIME
k8s-node-1 Ready <none> 9h v1.15.3 192.168.205.11 <none> Ubuntu 16.04.6 LTS 4.4.0-159-generic docker://17.3.3
Then set the kubelet port (10250 in our case):
// any machine
$ kubectl get nodes k8s-node-1 -o jsonpath='{.status.daemonEndpoints.kubeletEndpoint}'
map[Port:10250]
Now it's time to test the network. Is there a connection to the worker node (192.168.205.11)? It is! If you "kill" the process exec, it will disappear, so I know that the connection was established by the api-server as a result of the executed exec command.
But wait, how did kubelet pull this off? The kubelet has a daemon that opens access to the API through the port for api-server requests:
// Server is the library interface to serve the stream requests.
type Server interface {
http.Handler
// Get the serving URL for the requests.
// Requests must not be nil. Responses may be nil iff an error is returned.
GetExec(*runtimeapi.ExecRequest) (*runtimeapi.ExecResponse, error)
GetAttach(req *runtimeapi.AttachRequest) (*runtimeapi.AttachResponse, error)
GetPortForward(*runtimeapi.PortForwardRequest) (*runtimeapi.PortForwardResponse, error)
// Start the server.
// addr is the address to serve on (address:port) stayUp indicates whether the server should
// listen until Stop() is called, or automatically stop after all expected connections are
// closed. Calling Get{Exec,Attach,PortForward} increments the expected connection count.
// Function does not return until the server is stopped.
Start(stayUp bool) error
// Stop the server, and terminate any open connections.
Stop() error
}
Kubelet implements an interface RuntimeServiceClient, which is part of the Container Runtime Interface (we wrote more about it, for example, here - approx. transl.):
Long listing from cri-api in kubernetes/kubernetes
// For semantics around ctx use and closing/ending streaming RPCs, please refer to https://godoc.org/google.golang.org/grpc#ClientConn.NewStream.
type RuntimeServiceClient interface {
// Version returns the runtime name, runtime version, and runtime API version.
Version(ctx context.Context, in *VersionRequest, opts ...grpc.CallOption) (*VersionResponse, error)
// RunPodSandbox creates and starts a pod-level sandbox. Runtimes must ensure
// the sandbox is in the ready state on success.
RunPodSandbox(ctx context.Context, in *RunPodSandboxRequest, opts ...grpc.CallOption) (*RunPodSandboxResponse, error)
// StopPodSandbox stops any running process that is part of the sandbox and
// reclaims network resources (e.g., IP addresses) allocated to the sandbox.
// If there are any running containers in the sandbox, they must be forcibly
// terminated.
// This call is idempotent, and must not return an error if all relevant
// resources have already been reclaimed. kubelet will call StopPodSandbox
// at least once before calling RemovePodSandbox. It will also attempt to
// reclaim resources eagerly, as soon as a sandbox is not needed. Hence,
// multiple StopPodSandbox calls are expected.
StopPodSandbox(ctx context.Context, in *StopPodSandboxRequest, opts ...grpc.CallOption) (*StopPodSandboxResponse, error)
// RemovePodSandbox removes the sandbox. If there are any running containers
// in the sandbox, they must be forcibly terminated and removed.
// This call is idempotent, and must not return an error if the sandbox has
// already been removed.
RemovePodSandbox(ctx context.Context, in *RemovePodSandboxRequest, opts ...grpc.CallOption) (*RemovePodSandboxResponse, error)
// PodSandboxStatus returns the status of the PodSandbox. If the PodSandbox is not
// present, returns an error.
PodSandboxStatus(ctx context.Context, in *PodSandboxStatusRequest, opts ...grpc.CallOption) (*PodSandboxStatusResponse, error)
// ListPodSandbox returns a list of PodSandboxes.
ListPodSandbox(ctx context.Context, in *ListPodSandboxRequest, opts ...grpc.CallOption) (*ListPodSandboxResponse, error)
// CreateContainer creates a new container in specified PodSandbox
CreateContainer(ctx context.Context, in *CreateContainerRequest, opts ...grpc.CallOption) (*CreateContainerResponse, error)
// StartContainer starts the container.
StartContainer(ctx context.Context, in *StartContainerRequest, opts ...grpc.CallOption) (*StartContainerResponse, error)
// StopContainer stops a running container with a grace period (i.e., timeout).
// This call is idempotent, and must not return an error if the container has
// already been stopped.
// TODO: what must the runtime do after the grace period is reached?
StopContainer(ctx context.Context, in *StopContainerRequest, opts ...grpc.CallOption) (*StopContainerResponse, error)
// RemoveContainer removes the container. If the container is running, the
// container must be forcibly removed.
// This call is idempotent, and must not return an error if the container has
// already been removed.
RemoveContainer(ctx context.Context, in *RemoveContainerRequest, opts ...grpc.CallOption) (*RemoveContainerResponse, error)
// ListContainers lists all containers by filters.
ListContainers(ctx context.Context, in *ListContainersRequest, opts ...grpc.CallOption) (*ListContainersResponse, error)
// ContainerStatus returns status of the container. If the container is not
// present, returns an error.
ContainerStatus(ctx context.Context, in *ContainerStatusRequest, opts ...grpc.CallOption) (*ContainerStatusResponse, error)
// UpdateContainerResources updates ContainerConfig of the container.
UpdateContainerResources(ctx context.Context, in *UpdateContainerResourcesRequest, opts ...grpc.CallOption) (*UpdateContainerResourcesResponse, error)
// ReopenContainerLog asks runtime to reopen the stdout/stderr log file
// for the container. This is often called after the log file has been
// rotated. If the container is not running, container runtime can choose
// to either create a new log file and return nil, or return an error.
// Once it returns error, new container log file MUST NOT be created.
ReopenContainerLog(ctx context.Context, in *ReopenContainerLogRequest, opts ...grpc.CallOption) (*ReopenContainerLogResponse, error)
// ExecSync runs a command in a container synchronously.
ExecSync(ctx context.Context, in *ExecSyncRequest, opts ...grpc.CallOption) (*ExecSyncResponse, error)
// Exec prepares a streaming endpoint to execute a command in the container.
Exec(ctx context.Context, in *ExecRequest, opts ...grpc.CallOption) (*ExecResponse, error)
// Attach prepares a streaming endpoint to attach to a running container.
Attach(ctx context.Context, in *AttachRequest, opts ...grpc.CallOption) (*AttachResponse, error)
// PortForward prepares a streaming endpoint to forward ports from a PodSandbox.
PortForward(ctx context.Context, in *PortForwardRequest, opts ...grpc.CallOption) (*PortForwardResponse, error)
// ContainerStats returns stats of the container. If the container does not
// exist, the call returns an error.
ContainerStats(ctx context.Context, in *ContainerStatsRequest, opts ...grpc.CallOption) (*ContainerStatsResponse, error)
// ListContainerStats returns stats of all running containers.
ListContainerStats(ctx context.Context, in *ListContainerStatsRequest, opts ...grpc.CallOption) (*ListContainerStatsResponse, error)
// UpdateRuntimeConfig updates the runtime configuration based on the given request.
UpdateRuntimeConfig(ctx context.Context, in *UpdateRuntimeConfigRequest, opts ...grpc.CallOption) (*UpdateRuntimeConfigResponse, error)
// Status returns the status of the runtime.
Status(ctx context.Context, in *StatusRequest, opts ...grpc.CallOption) (*StatusResponse, error)
}
Container Runtime is responsible for the implementation RuntimeServiceServer:
Long listing from cri-api in kubernetes/kubernetes
// RuntimeServiceServer is the server API for RuntimeService service.
type RuntimeServiceServer interface {
// Version returns the runtime name, runtime version, and runtime API version.
Version(context.Context, *VersionRequest) (*VersionResponse, error)
// RunPodSandbox creates and starts a pod-level sandbox. Runtimes must ensure
// the sandbox is in the ready state on success.
RunPodSandbox(context.Context, *RunPodSandboxRequest) (*RunPodSandboxResponse, error)
// StopPodSandbox stops any running process that is part of the sandbox and
// reclaims network resources (e.g., IP addresses) allocated to the sandbox.
// If there are any running containers in the sandbox, they must be forcibly
// terminated.
// This call is idempotent, and must not return an error if all relevant
// resources have already been reclaimed. kubelet will call StopPodSandbox
// at least once before calling RemovePodSandbox. It will also attempt to
// reclaim resources eagerly, as soon as a sandbox is not needed. Hence,
// multiple StopPodSandbox calls are expected.
StopPodSandbox(context.Context, *StopPodSandboxRequest) (*StopPodSandboxResponse, error)
// RemovePodSandbox removes the sandbox. If there are any running containers
// in the sandbox, they must be forcibly terminated and removed.
// This call is idempotent, and must not return an error if the sandbox has
// already been removed.
RemovePodSandbox(context.Context, *RemovePodSandboxRequest) (*RemovePodSandboxResponse, error)
// PodSandboxStatus returns the status of the PodSandbox. If the PodSandbox is not
// present, returns an error.
PodSandboxStatus(context.Context, *PodSandboxStatusRequest) (*PodSandboxStatusResponse, error)
// ListPodSandbox returns a list of PodSandboxes.
ListPodSandbox(context.Context, *ListPodSandboxRequest) (*ListPodSandboxResponse, error)
// CreateContainer creates a new container in specified PodSandbox
CreateContainer(context.Context, *CreateContainerRequest) (*CreateContainerResponse, error)
// StartContainer starts the container.
StartContainer(context.Context, *StartContainerRequest) (*StartContainerResponse, error)
// StopContainer stops a running container with a grace period (i.e., timeout).
// This call is idempotent, and must not return an error if the container has
// already been stopped.
// TODO: what must the runtime do after the grace period is reached?
StopContainer(context.Context, *StopContainerRequest) (*StopContainerResponse, error)
// RemoveContainer removes the container. If the container is running, the
// container must be forcibly removed.
// This call is idempotent, and must not return an error if the container has
// already been removed.
RemoveContainer(context.Context, *RemoveContainerRequest) (*RemoveContainerResponse, error)
// ListContainers lists all containers by filters.
ListContainers(context.Context, *ListContainersRequest) (*ListContainersResponse, error)
// ContainerStatus returns status of the container. If the container is not
// present, returns an error.
ContainerStatus(context.Context, *ContainerStatusRequest) (*ContainerStatusResponse, error)
// UpdateContainerResources updates ContainerConfig of the container.
UpdateContainerResources(context.Context, *UpdateContainerResourcesRequest) (*UpdateContainerResourcesResponse, error)
// ReopenContainerLog asks runtime to reopen the stdout/stderr log file
// for the container. This is often called after the log file has been
// rotated. If the container is not running, container runtime can choose
// to either create a new log file and return nil, or return an error.
// Once it returns error, new container log file MUST NOT be created.
ReopenContainerLog(context.Context, *ReopenContainerLogRequest) (*ReopenContainerLogResponse, error)
// ExecSync runs a command in a container synchronously.
ExecSync(context.Context, *ExecSyncRequest) (*ExecSyncResponse, error)
// Exec prepares a streaming endpoint to execute a command in the container.
Exec(context.Context, *ExecRequest) (*ExecResponse, error)
// Attach prepares a streaming endpoint to attach to a running container.
Attach(context.Context, *AttachRequest) (*AttachResponse, error)
// PortForward prepares a streaming endpoint to forward ports from a PodSandbox.
PortForward(context.Context, *PortForwardRequest) (*PortForwardResponse, error)
// ContainerStats returns stats of the container. If the container does not
// exist, the call returns an error.
ContainerStats(context.Context, *ContainerStatsRequest) (*ContainerStatsResponse, error)
// ListContainerStats returns stats of all running containers.
ListContainerStats(context.Context, *ListContainerStatsRequest) (*ListContainerStatsResponse, error)
// UpdateRuntimeConfig updates the runtime configuration based on the given request.
UpdateRuntimeConfig(context.Context, *UpdateRuntimeConfigRequest) (*UpdateRuntimeConfigResponse, error)
// Status returns the status of the runtime.
Status(context.Context, *StatusRequest) (*StatusResponse, error)
}