In #91342 attempting to evict a Pod with a DeletionTimestamp caused
checking of PDBs to be ignored due to the fact that a Pod scheduled for
deletion should not be factored into a disruption budget. However, PDB
eviction tests currently will sometimes select a Pod already scheduled
for deletion, expecting that attempting to evict it will conflict with
the PDB. This updates those tests to make sure a Pod with deletion
timestamp is not selected for eviction when it is intended to violate a
PDB.
Signed-off-by: hasheddan <georgedanielmangum@gmail.com>
The multi-arch container images used in tests live in quay.io which
doesn't support nesting. By making the /volume/ images repo configurable,
we are able to override them despite our current limitation.
Signed-off-by: Rafael Fonseca <r4f4rfs@gmail.com>
This extends the existing "ephemeral volume" tests to also cover
generic ephemeral inline volumes. They get instantiated for all
drivers (CSI and others) which support persistent volume provisioning,
for several different filesystems.
Configuring the number of inline volumes via a flag with a computed
name had been identified as problematic before and now gets removed
because re-using the tests as a stress test with a higher number of
volumes should be added and configured separately.
Windows test for subPath is failing due to an issue related to
removeUnusedContainers calls. After image is changed to Agnhost, it
automatically has a args by default. However, there are places to use
container commands instead of args and causing issues.
This is the first step to fix this issue. Next plan to replace
busybox used in Linux with Agnhost which can work for both linux and
windows.
I also mark two subPath tests as LinuxOnly. I think they are not ready
for windows yet. Before they were passing due to wrong reason. The tests
checks failed container status but the contain fails due to other
reasons than what we expected.
This is useful in case that the pod owns some resources, because then
waiting for the pod ensures that those resources also were removed.
This should not matter at the moment because pods typically are not
owners of any other object, but that will change with the introduction
of generic ephemeral inline
volumes (https://github.com/kubernetes/enhancements/tree/master/keps/sig-storage/1698-generic-ephemeral-volumes).
- Allow client-side to server-side apply upgrade.
Ensure that a user can change management of an object from client-side apply to
server-side apply without conflicts.
- Allow server-side apply to client-side downgrade.
For an object managed with client-side apply, a user may upgrade to
managing the object with server-side apply, then decide to downgrade.
We can support this downgrade by keeping the last-applied-configuration
annotation for client-side apply updated with server-side apply.
The test steps are as follows:
1. Write some data
2. Take a snapshot
3. Write more data
4. Create a new volume from snapshot
5. Validate data is the old data
1. Use ginkgo before each to do common setup
2. Use volume resource to create SC, PV, PVC and handle cleanup
3. Add SnapshotResource to handle creating and cleanup of VS, VSC, VSClass
4. Add test pattern for deletion policy: Delete vs Retain
5. Use test pattern to determine test behaviour
6. Add test pattern for preprovisioned snapshot (not implemented)
These changes are made to consolidate common setup steps and stop resource
leaks by waiting for objects to be deleted.
By creating CSIStorageCapacity objects in advance, we get the
FailedScheduling pod event if (and only if!) the test is expected to
fail because of insufficient or missing capacity. We can use that as
indicator that waiting for pod start can be stopped early. However,
because we might not get to see the event under load, we still need
the timeout.
Setting testParameters.scName had no effect because
StorageClassTest.StorageClassName isn't used anywhere. Instead, the
storage class name is generated dynamically.
DeprecatedMightBeMasterNode() has been marked as deprecated and we need to
find alternative way for callers of the function.
In NewResourceUsageGatherer(), the function was called for distinguishing
the specified pods are running on master nodes, and the gatherer gathers
those pods' resource usage.
This adds nodeHasControlPlanePods() to gistinguish the specified pods
are running on nodes which are operating control plane pods (kube-scheduler
and kube-controller-manager) and replace callers of DeprecatedMightBeMasterNode()
with this new function as better way.
If a bearer token is present in a request, the exec credential plugin should accept that as the chosen method of authentication. Judging by an [earlier comment in exec.go](c18bc7e9f7/staging/src/k8s.io/client-go/plugin/pkg/client/auth/exec/exec.go (L217)), this was already intended. This would however not work since UpdateTransportConfig would set the GetCert callback which would then get called by the transport, triggering the exec plugin action even with a token present in the request. See linked issue for further details.
See #87369 for further details.
Signed-off-by: Anders Eknert <anders.eknert@bisnode.com>
When using the entire test name as file name, the name became too
long (> 256 characters, which wasn't supported by all file systems)
and the artifact directory got cluttered.
The original reason (a limitation in Gubernator) no longer applies
because Spyglass is used now for log viewing.
Using NodeWrapper in the integration tests gives more flexibility when
creating nodes. For instance, tests can create nodes with labels or
with a specific sets of resources.
Also, NodeWrapper initialises a node with a capacity of 32 pods, which
can be overridden by the caller. This makes sure that a node is usable
as soon as it is created.
Now the test covers 6 different api calls
- verify create with a get
- verify patch with a list (all namespaces)
- verify delete with a list (single namespace)
The thing is, for this test at least, I'm pretty sure there's nothing
we need to wait on. Instead of waiting for a deleted event, we will
relist configmaps and expect 0, to confirm the deletion took effect
This drops testfiles.ReadOrDie and updated testfiles.Exists to return an
error, forcing the caller to decide whether to call framework.Fail or do
something else.
It makes for a slightly less friendly API, but also means the package is
decoupled from framework again, as per the comments at the top of the
file
Currently when checking for unscheduled pods an exception will be raised
if a pod is not scheduled and the status is unknown. This update modifies
the logic to include any pod without a NodeName in the not scheduled
pods returned.
Signed-off-by: hasheddan <georgedanielmangum@gmail.com>
If a reserve plugin's Reserve method returns an error, there could be
previously allocated resources from successfully completed reserve
plugins that must be unallocated by the corresponding Unreserve
operation. Since Unreserve operations are idempotent, this patch runs
the Unreserve operation of ALL reserve plugins when a Reserve operation
fails.
This runs much faster than before. This change removes all of the
async status output because all of the compute time is spent inside
go/packages, with no opportunity to update the status.
Adds testdata code to prove it fails when expected.
When node scheduling tests were updated to use worker instead of master
nodes the GetPodsScheduled function, which is tasked with getting all
scheduled and not scheduled pods inadvertently was changed to ignore all
pods that have an empty NodeName before checking whether pods had been
scheduled or not. This updates the function to include pods without a
NodeName in the check for unscheduled pods.
Signed-off-by: hasheddan <georgedanielmangum@gmail.com>
As its name, DeprecatedMightBeMasterNode is deprecated.
In e2e metrics, the function was used for knowing master node name to
get metrics from kube-scheduler and kube-controller-manager pods.
This make e2e metrics get these metrics directly by getting those pod
names without calling DeprecatedMightBeMasterNode().
