Automatic merge from submit-queue (batch tested with PRs 42692, 42169, 42173)
Add pprof trace support
Add support for `/debug/pprof/trace`
Can wait for master to reopen for 1.7.
cc @smarterclayton @wojtek-t @gmarek @timothysc @jeremyeder @kubernetes/sig-scalability-pr-reviews
* Properly return ImageNotFoundError
* Support inject "Images" or "ImageInspects" and keep both in sync.
* Remove the FakeDockerPuller and let FakeDockerClient subsumes its
functinality. This reduces the overhead to maintain both objects.
* Various small fixes and refactoring of the testing utils.
Automatic merge from submit-queue (batch tested with PRs 41890, 42593, 42633, 42626, 42609)
Pods pending due to insufficient OIR should get scheduled once sufficient OIR becomes available (e2e disabled).
#41870 was reverted because it introduced an e2e test flake. This is the same code with the e2e for OIR disabled again.
We can attempt to enable the e2e test cases one-by-one in follow-up PRs, but it would be preferable to get the main fix merged in time for 1.6 since OIR is broken on master (see #41861).
cc @timothysc
Automatic merge from submit-queue (batch tested with PRs 42506, 42585, 42596, 42584)
provide active pods to cgroup cleanup
**What this PR does / why we need it**:
This PR provides more information for when a pod cgroup is considered orphaned. The running pods cache is based on the runtime's view of the world. we create pod cgroups before containers so we should just be looking at activePods.
**Which issue this PR fixes**
Fixes https://github.com/kubernetes/kubernetes/issues/42431
Automatic merge from submit-queue
cgroup names created by kubelet should be lowercased
**What this PR does / why we need it**:
This PR modifies the kubelet to create cgroupfs names that are lowercased. This better aligns us with the naming convention for cgroups v2 and other cgroup managers in ecosystem (docker, systemd, etc.)
See: https://www.kernel.org/doc/Documentation/cgroup-v2.txt
"2-6-2. Avoid Name Collisions"
**Special notes for your reviewer**:
none
**Release note**:
```release-note
kubelet created cgroups follow lowercase naming conventions
```
Automatic merge from submit-queue (batch tested with PRs 31783, 41988, 42535, 42572, 41870)
Pods pending due to insufficient OIR should get scheduled once sufficient OIR becomes available.
This appears to be a regression since v1.5.0 in scheduler behavior for opaque integer resources, reported in https://github.com/kubernetes/kubernetes/issues/41861.
- [X] Add failing e2e test to trigger the regression
- [x] Restore previous behavior (pods pending due to insufficient OIR get scheduled once sufficient OIR becomes available.)
Automatic merge from submit-queue
[Bug] Fix gpu initialization in Kubelet
Kubelet incorrectly fails if `AllAlpha=true` feature gate is enabled with container runtimes that are not `docker`.
Replaces #42407
- Added schedulercache.Resource.SetOpaque helper.
- Amend kubelet allocatable sync so that when OIRs are removed from capacity
they are also removed from allocatable.
- Fixes#41861.
Automatic merge from submit-queue (batch tested with PRs 42369, 42375, 42397, 42435, 42455)
[Bug Fix]: Avoid evicting more pods than necessary by adding Timestamps for fsstats and ignoring stale stats
Continuation of #33121. Credit for most of this goes to @sjenning. I added volume fs timestamps.
**why is this a bug**
This PR attempts to fix part of https://github.com/kubernetes/kubernetes/issues/31362 which results in multiple pods getting evicted unnecessarily whenever the node runs into resource pressure. This PR reduces the chances of such disruptions by avoiding reacting to old/stale metrics.
Without this PR, kubernetes nodes under resource pressure will cause unnecessary disruptions to user workloads.
This PR will also help deflake a node e2e test suite.
The eviction manager currently avoids evicting pods if metrics are old. However, timestamp data is not available for filesystem data, and this causes lots of extra evictions.
See the [inode eviction test flakes](https://k8s-testgrid.appspot.com/google-node#kubelet-flaky-gce-e2e) for examples.
This should probably be treated as a bugfix, as it should help mitigate extra evictions.
cc: @kubernetes/sig-storage-pr-reviews @kubernetes/sig-node-pr-reviews @vishh @derekwaynecarr @sjenning
Automatic merge from submit-queue (batch tested with PRs 42369, 42375, 42397, 42435, 42455)
Kubelet: return container runtime's version instead of CRI's one
**What this PR does / why we need it**:
With CRI enabled by default, kubelet reports the version of CRI instead of container runtime version. This PR fixes this problem.
**Which issue this PR fixes**
Fixes#42396.
**Special notes for your reviewer**:
Should also cherry-pick to 1.6 branch.
**Release note**:
```release-note
NONE
```
cc @yujuhong @kubernetes/sig-node-bugs
Automatic merge from submit-queue
Eviction Manager Enforces Allocatable Thresholds
This PR modifies the eviction manager to enforce node allocatable thresholds for memory as described in kubernetes/community#348.
This PR should be merged after #41234.
cc @kubernetes/sig-node-pr-reviews @kubernetes/sig-node-feature-requests @vishh
** Why is this a bug/regression**
Kubelet uses `oom_score_adj` to enforce QoS policies. But the `oom_score_adj` is based on overall memory requested, which means that a Burstable pod that requested a lot of memory can lead to OOM kills for Guaranteed pods, which violates QoS. Even worse, we have observed system daemons like kubelet or kube-proxy being killed by the OOM killer.
Without this PR, v1.6 will have node stability issues and regressions in an existing GA feature `out of Resource` handling.
Automatic merge from submit-queue (batch tested with PRs 41919, 41149, 42350, 42351, 42285)
kubelet: enable qos-level memory limits
```release-note
Experimental support to reserve a pod's memory request from being utilized by pods in lower QoS tiers.
```
Enables the QoS-level memory cgroup limits described in https://github.com/kubernetes/community/pull/314
**Note: QoS level cgroups have to be enabled for any of this to take effect.**
Adds a new `--experimental-qos-reserved` flag that can be used to set the percentage of a resource to be reserved at the QoS level for pod resource requests.
For example, `--experimental-qos-reserved="memory=50%`, means that if a Guaranteed pod sets a memory request of 2Gi, the Burstable and BestEffort QoS memory cgroups will have their `memory.limit_in_bytes` set to `NodeAllocatable - (2Gi*50%)` to reserve 50% of the guaranteed pod's request from being used by the lower QoS tiers.
If a Burstable pod sets a request, its reserve will be deducted from the BestEffort memory limit.
The result is that:
- Guaranteed limit matches root cgroup at is not set by this code
- Burstable limit is `NodeAllocatable - Guaranteed reserve`
- BestEffort limit is `NodeAllocatable - Guaranteed reserve - Burstable reserve`
The only resource currently supported is `memory`; however, the code is generic enough that other resources can be added in the future.
@derekwaynecarr @vishh
Automatic merge from submit-queue (batch tested with PRs 41672, 42084, 42233, 42165, 42273)
ExecProbes should be able to do simple env var substitution
For containers that don't have bash, we should support env substitution
like we do on command and args. However, without major refactoring
valueFrom is not supportable from inside the prober. For now, implement
substitution based on hardcoded env and leave TODOs for future work.
Improves the state of #40846, will spawn a follow up issue for future refactoring after CRI settles down
Automatic merge from submit-queue (batch tested with PRs 41980, 42192, 42223, 41822, 42048)
CRI: Make dockershim better implements CRI.
When thinking about CRI Validation test, I found that `PodSandboxStatus.Linux.Namespaces.Options.HostPid` and `PodSandboxStatus.Linux.Namespaces.Options.HostIpc` are not populated. Although they are not used by kuberuntime now, we should populate them to conform to CRI.
/cc @yujuhong @feiskyer
Automatic merge from submit-queue (batch tested with PRs 41931, 39821, 41841, 42197, 42195)
Use `docker logs` directly if the docker logging driver is not `json-file`
Fixes https://github.com/kubernetes/kubernetes/issues/41996.
Post the PR first, I still need to manually test this, because we don't have test coverage for journald logging pluggin.
@yujuhong @dchen1107
/cc @kubernetes/sig-node-pr-reviews
Automatic merge from submit-queue (batch tested with PRs 41644, 42020, 41753, 42206, 42212)
Burstable QoS cgroup has cpu shares assigned
**What this PR does / why we need it**:
This PR sets the Burstable QoS cgroup cpu shares value to the sum of the pods cpu requests in that tier. We need it for proper evaluation of CPU shares in the new QoS hierarchy.
**Special notes for your reviewer**:
It builds against the framework proposed for https://github.com/kubernetes/kubernetes/pull/41833
Automatic merge from submit-queue (batch tested with PRs 41644, 42020, 41753, 42206, 42212)
Ensure pod cgroup is deleted prior to deletion of pod
**What this PR does / why we need it**:
This PR ensures that the kubelet removes the pod cgroup sandbox prior to deletion of a pod from the apiserver. We need this to ensure that the default behavior in the kubelet is to not leak resources.
Automatic merge from submit-queue
Re-writing of the resolv.conf file generated by docker
Fixes#17406
Docker 1.12 will contain feature "The option --dns and --net=host should not be mutually exclusive" (docker/docker#22408)
This patch adds optional support for this ability in kubelet (for now in case of "hostNetwork: true" set all dns settings are ignored if any).
To enable feature use newly added kubelet flag: --allow-dns-for-hostnet=true
Introduced chages:
1. Re-writing of the resolv.conf file generated by docker.
Cluster dns settings aren't passed anymore to docker api in all cases, not only for pods with host network:
the resolver conf will be overwritten after infra-container creation to override docker's behaviour.
2. Added new one dnsPolicy - 'ClusterFirstWithHostNet', so now there are:
- ClusterFirstWithHostNet - use dns settings in all cases, i.e. with hostNet=true as well
- ClusterFirst - use dns settings unless hostNetwork is true
- Default
Fixes#17406
Automatic merge from submit-queue
Extend experimental support to multiple Nvidia GPUs
Extended from #28216
```release-note
`--experimental-nvidia-gpus` flag is **replaced** by `Accelerators` alpha feature gate along with support for multiple Nvidia GPUs.
To use GPUs, pass `Accelerators=true` as part of `--feature-gates` flag.
Works only with Docker runtime.
```
1. Automated testing for this PR is not possible since creation of clusters with GPUs isn't supported yet in GCP.
1. To test this PR locally, use the node e2e.
```shell
TEST_ARGS='--feature-gates=DynamicKubeletConfig=true' FOCUS=GPU SKIP="" make test-e2e-node
```
TODO:
- [x] Run manual tests
- [x] Add node e2e
- [x] Add unit tests for GPU manager (< 100% coverage)
- [ ] Add unit tests in kubelet package
Automatic merge from submit-queue (batch tested with PRs 41921, 41695, 42139, 42090, 41949)
Unify fake runtime helper in kuberuntime, rkt and dockertools.
Addresses https://github.com/kubernetes/kubernetes/pull/42081#issuecomment-282429775.
Add `pkg/kubelet/container/testing/fake_runtime_helper.go`, and change `kuberuntime`, `rkt` and `dockertools` to use it.
@yujuhong This is a small unit test refactoring PR. Could you help me review it?
Automatic merge from submit-queue (batch tested with PRs 38676, 41765, 42103, 41833, 41702)
kubelet: cm: refactor QoS logic into seperate interface
This commit has no functional change. It refactors the QoS cgroup logic into a new `QOSContainerManager` interface to allow for better isolation for QoS cgroup features coming down the pike.
This is a breakout of the refactoring component of my QoS memory limits PR https://github.com/kubernetes/kubernetes/pull/41149 which will need to be rebased on top of this.
@vishh @derekwaynecarr
Automatic merge from submit-queue (batch tested with PRs 42162, 41973, 42015, 42115, 41923)
Increase Min Timeout for kill pod
Should mitigate #41347, which describes flakes in the inode eviction test due to "GracePeriodExceeded" errors.
When we use gracePeriod == 0, as we do in eviction, the pod worker currently sets a timeout of 2 seconds to kill a pod.
We are hitting this timeout fairly often during eviction tests, causing extra pods to be evicted (since the eviction manager "fails" to evict that pod, and kills the next one).
This PR increases the timeout from 2 seconds to 4, although we could increase it even more if we think that would be appropriate.
cc @yujuhong @vishh @derekwaynecarr
For containers that don't have bash, we should support env substitution
like we do on command and args. However, without major refactoring
valueFrom is not supportable from inside the prober. For now, implement
substitution based on hardcoded env and leave TODOs for future work.
