This patch adds a global timeout flag (viewable with `kubectl -h`) with
a default value of `0s` (meaning no timeout).
The timeout value is added to the default http client, so that zero
values and default behavior are enforced by the client.
**Example**
```
$ kubectl get pods # no timeout flag set - default to 0s (which means no
timeout)
NAME READY STATUS RESTARTS AGE
docker-registry-1-h7etw 1/1 Running 1 2h
router-1-uv0f9 1/1 Running 1 2h
$ kubectl get pods --timeout=0 # zero means no timeout no timeout flag set
NAME READY STATUS RESTARTS AGE
docker-registry-1-h7etw 1/1 Running 1 2h
router-1-uv0f9 1/1 Running 1 2h
$kubectl get pods --timeout=1ms
Unable to connect to the server: net/http: request canceled while
waiting for connection (Client.Timeout exceeded while awaiting headers)
```
This patch provides a more relevant error message when a client
configuration option is passed with an invalid or non-existent value.
`$ kubectl get pods --cluster="non-existent"`
```
error: No configuration file found, please login or point to an existing
file
```
`$ kubectl get pods --cluster="non-existent"`
```
error: cluster "non-existent" does not exist
```
This commit moves away from using a global variable for default
configuration checking, and instead exposes a method on LoadingRules to
determine whether a particular restclient.Config should be considered
"default". This allows kubectl to provide its own defaults (the same
as before, KUBERNETES_MASTER and the static localhost:8080 values) while
allowing other clients to avoid defining them.
In-cluster config defaulting is now easier to read.
Automatic merge from submit-queue
Add a check in ConfirmUsable() to validate the contextName
**What this PR does / why we need it**:
When a context name is provided, but can't be found (miss spelling), it currently
uses the defaults. This PR will cause the command to fail, to prevent unexpected side effects
of using the wrong configuration.
**Which issue this PR fixes**
fixes#21062
**Special notes for your reviewer**:
None
**Release note**:
```release-note
Error if a contextName is provided but not found in the kubeconfig.
```
By removing the default "localhost:8080" behavior several paths in
client config began returning err == ErrEmptyConfig rather than err ==
nil. The code checking for in cluster config was wrong - the logic
should be:
1. If loading the underlying config returns a non-empty error, fail
2. If the underlying config is not equal to the default config,
return that config (it's got user input)
3. If it is possible to use in cluster config, do so
4. Otherwise return the default config (and or default EmptyConfig
error).
This changes clientcmd to skip the default cluster, but preserves the
behavior in kubectl. This prevents the possibility of an administrator
misconfiguration in kubelet or other server component from allowing a
third party who can bind to 8080 on that host from potentially
impersonating an API server and gaining root access.
In the getting started example of AWS, the master uses an IP that is changed on
stop/start. If you are playing with a cluster and stop and start the master,
the IP is changed and you can't connect again, even using the
"--insecure-skip-tls-verify" option.
This patch fixes it and makes the option work on those cases too by making
sure no CA/CAData is added when it shouldn't.
Current BuildConfigFromFlags() seems to print a log like below when
config path isn't specified:
```
W0120 15:30:06.196820 13323 client_config.go:359] error creating
inClusterConfig, falling back to default config: %vunable to load
in-cluster configuration, KUBERNETES_SERVICE_HOST and
KUBERNETES_SERVICE_PORT must be defined
```
This commit removes the needless "%v"
For UX, it would be better if we presented an error for validation that
is "this config is empty" rather than the inaccurate "no server name" or
"no context name" errors. Returns a typed error.
