Automatic merge from submit-queue
make client-go more authoritative
Builds on https://github.com/kubernetes/kubernetes/pull/40103
This moves a few more support package to client-go for origination.
1. restclient/watch - nodep
1. util/flowcontrol - used interface
1. util/integer, util/clock - used in controllers and in support of util/flowcontrol
Automatic merge from submit-queue
controller: decouple cleanup policy from deployment strategies
Deployments get cleaned up only when they are paused, they get scaled up/down,
or when the strategy that drives rollouts completes. This means that stuck
deployments that fall into none of the above categories will not get cleaned
up. Since cleanup is already safe by itself (we only delete old replica sets
that are synced by the replica set controller and have no replicas) we can
execute it for every deployment when there is no intention to rollback.
Fixes https://github.com/kubernetes/kubernetes/issues/40068
Deployments get cleaned up only when they are paused, they get scaled up/down,
or when the strategy that drives rollouts completes. This means that stuck
deployments that fall into none of the above categories will not get cleaned
up. Since cleanup is already safe by itself (we only delete old replica sets
that are synced by the replica set controller and have no replicas) we can
execute it for every deployment when there is no intention to rollback.
Automatic merge from submit-queue (batch tested with PRs 34763, 38706, 39939, 40020)
Use Statefulset instead in e2e and controller
Quick fix ref: #35534
We should finish the issue to meet v1.6 milestone.
Automatic merge from submit-queue
genericapiserver: cut off pkg/serviceaccount dependency
**Blocked** by pkg/api/validation/genericvalidation to be split up and moved into apimachinery.
Automatic merge from submit-queue
Do not list CronJob unmet starting times beyond deadline
**What this PR does / why we need it**:
See #36311. `getRecentUnmetScheduleTimes` gives up after 100 unmet times to avoid wasting too much CPU or memory generating all the times, as it generates them sequentially.
When concurrency is forbidden, this is conceptually un-necessary: we only need the last unmet start time. This suggests that when concurrency is forbidden, we could generate times by going backward in time from now. This is not very practical as CronJob currently relies on a package that only provides `Next` and no `Prev`. Hand-cooking a `Prev` does not seem like a good idea. I could submit a PR to the cron library to add a `Prev` method, and use that when concurrency is forbidden through something like `getLastUnmetScheduleTime`. This would be `O(1)` and there would be no limit involved.
(edit: actually, even for the other concurrency settings, we only start the last unmet start times -- there is a `TODO` in the controller to actually start all of them, but that is not implemented at the moment. This means the solution would apply, at least temporarily, to all concurrency settings).
cc @soltysh what do you think?
In the meantime, I would suggest to do something simple. Currently, the user has no way to configure anything to ensure that his CronJob will not get stuck if one job takes more that 100 unmet times.
`getRecentUnmetScheduleTimes` starts with an initial time corresponding to the last start (or to the creation of the CronJob, if nothing has started yet). However, when `StartingDeadlineSeconds` is set, the controller will not start anything that is older than the deadline, so if the last start is way beyond the deadline, we are generating potentially lots of unmet start times that will not be considered by the scheduler for scheduling anyway.
Consider a job running every minute, where the last instance has taken 120 minutes. This means there are more than 100 unmet times when we start counting from the last start time.
**The PR makes `getRecentUnmetScheduleTimes` only consider times that do not fall beyond the deadline.** Here, the CronJob can be configured with a `StartingDeadlineSeconds` of, say, 10 minutes. After the 120min job has run, `getRecentUnmetScheduleTimes` will only consider the times in the last 10 minutes from now, and will not get stuck.
As a side note on the max. number of unmet times to use as limits in terms of CPU used by the controller: I have run a quick benchmark on my i7 mac. Schedules corresponding to "once a week" tend to be more expensive to generate unmet times for. Just FYI.
```
+--------------+---------------+--------------+
| SCHEDULE | MISSED STARTS | TIMING |
+--------------+---------------+--------------+
| */1 * * * ? | 100 | 383.645µs |
| */30 * * * ? | 100 | 354.765µs |
| 30 1 * * ? | 100 | 1.065124ms |
| 30 1 * * 0 | 100 | 1.80034ms |
| */1 * * * ? | 500 | 1.341365ms |
| */30 * * * ? | 500 | 1.814441ms |
| 30 1 * * ? | 500 | 8.475012ms |
| 30 1 * * 0 | 500 | 10.020613ms |
| */1 * * * ? | 1000 | 2.551697ms |
| */30 * * * ? | 1000 | 4.075813ms |
| 30 1 * * ? | 1000 | 17.674945ms |
| 30 1 * * 0 | 1000 | 19.149324ms |
| */1 * * * ? | 10000 | 25.725531ms |
| */30 * * * ? | 10000 | 87.520022ms |
| 30 1 * * ? | 10000 | 174.29216ms |
| 30 1 * * 0 | 10000 | 196.565748ms |
+--------------+---------------+--------------+
```
using
```.go
package main
import (
"fmt"
"time"
"os"
"strconv"
"github.com/robfig/cron"
"github.com/olekukonko/tablewriter"
)
func timeSchedule(schedule string, iterations int) (time.Duration) {
sched, err := cron.ParseStandard(schedule)
if err != nil {
panic(fmt.Sprintf("Unparseable schedule: %s", err))
}
start := time.Now()
t := time.Now()
for i := 1; i <= iterations; i++ {
t = sched.Next(t)
}
return time.Since(start)
}
func main() {
table := tablewriter.NewWriter(os.Stdout)
table.SetHeader([]string{"Schedule", "Missed starts", "Timing"})
schedules := []string{"*/1 * * * ?", "*/30 * * * ?", "30 1 * * ?", "30 1 * * 0"}
iteration_nums := []int{100, 500, 1000, 10000}
for _, iterations := range iteration_nums {
for _, schedule := range schedules {
table.Append([]string{schedule,
strconv.Itoa(iterations),
timeSchedule(schedule, iterations).String()})
}
}
table.Render()
}
```
**Which issue this PR fixes**: fixes#36311
**Special notes for your reviewer**:
**Release note**:
```release-note
```
Automatic merge from submit-queue (batch tested with PRs 38592, 39949, 39946, 39882)
move api/errors to apimachinery
`pkg/api/errors` is a set of helpers around `meta/v1.Status` that help to create and interpret various apiserver errors. Things like `.NewNotFound` and `IsNotFound` pairings. This pull moves it into apimachinery for use by the clients and servers.
@smarterclayton @lavalamp First commit is the move plus minor fitting. Second commit is straight replace and generation.
Automatic merge from submit-queue
Updated unit tests
@janetkuo updated the flaky unit test to have the same structure with regard to uncasting as the rest of the tests. ptal
Automatic merge from submit-queue (batch tested with PRs 39807, 37505, 39844, 39525, 39109)
Update deployment equality helper
@mfojtik @janetkuo this is split out of https://github.com/kubernetes/kubernetes/pull/38714 to reduce the size of that PR, ptal
Automatic merge from submit-queue (batch tested with PRs 39807, 37505, 39844, 39525, 39109)
Made cache.Controller to be interface.
**What this PR does / why we need it**:
#37504
Automatic merge from submit-queue
replace global registry in apimachinery with global registry in k8s.io/kubernetes
We'd like to remove all globals, but our immediate problem is that a shared registry between k8s.io/kubernetes and k8s.io/client-go doesn't work. Since client-go makes a copy, we can actually keep a global registry with other globals in pkg/api for now.
@kubernetes/sig-api-machinery-misc @lavalamp @smarterclayton @sttts
Automatic merge from submit-queue (batch tested with PRs 39661, 39740, 39801, 39468, 39743)
fix nodeStatusUpdateRetry count exceeding condition judgement
When tryUpdateNodeStatus() return err,err!=nil, but nc.kubeClient.Core().Nodes().Get() return no err, err==nil,
And we run nodeStatusUpdateRetry times, when for loop ends, err == nil, we can not print error info and run continue, so maybe the condition judgement is not right
Maybe caused #38671
When tryUpdateNodeStatus() return err,err!=nil, but nc.kubeClient.Core().Nodes().Get() return no err, err==nil,
And we run nodeStatusUpdateRetry times, when for loop ends, err == nil, we can not print error info and run continue, so the condition judgement is wrong.
Automatic merge from submit-queue (batch tested with PRs 39483, 39088, 38787)
daemonset: differentiate between cases in nodeShouldRun
specifically we need to differentiate between wanting to run,
should run and should continue running. This is required to
support all taint effects and will improve reporting and end
user debuggability.
fixes https://github.com/kubernetes/kubernetes/issues/28839 among other things
secifically we need to differentiate between wanting to run,
should run and should continue running. This is required to
support all taint effects and will improve reporting and end
user debuggability.
Automatic merge from submit-queue (batch tested with PRs 39684, 39577, 38989, 39534, 39702)
kubelet: request client auth certificates from certificate API.
This fixes kubeadm and --experiment-kubelet-bootstrap.
cc @liggitt
Automatic merge from submit-queue (batch tested with PRs 39694, 39383, 39651, 39691, 39497)
HPA Controller: Check for 0-sum request value
In certain conditions in which the set of metrics returned by Heapster
is completely disjoint from the set of pods returned by the API server,
we can have a request sum of zero, which can cause a panic (due to
division by zero). This checks for that condition.
Fixes#39680
**Release note**:
```release-note
Fixes an HPA-related panic due to division-by-zero.
```
Automatic merge from submit-queue
certificates: add a signing profile to the internal types
Here is a strawman of a CertificateSigningProfile type which would be used by the certificates controller when configuring cfssl. Side question: what magnitude of change warrants a design proposal?
@liggitt @gtank
In certain conditions in which the set of metrics returned by Heapster
is completely disjoint from the set of pods returned by the API server,
we can have a request sum of zero, which can cause a panic (due to
division by zero). This checks for that condition.
Fixes#39680
Automatic merge from submit-queue (batch tested with PRs 39628, 39551, 38746, 38352, 39607)
Increasing times on reconciling volumes fixing impact to AWS.
#**What this PR does / why we need it**:
We are currently blocked by API timeouts with PV volumes. See https://github.com/kubernetes/kubernetes/issues/39526. This is a workaround, not a fix.
**Special notes for your reviewer**:
A second PR will be dropped with CLI cobra options in it, but we are starting with increasing the reconciliation periods. I am dropping this without major testing and will test on our AWS account. Will be marked WIP until I run smoke tests.
**Release note**:
```release-note
Provide kubernetes-controller-manager flags to control volume attach/detach reconciler sync. The duration of the syncs can be controlled, and the syncs can be shut off as well.
```
Automatic merge from submit-queue (batch tested with PRs 37845, 39439, 39514, 39457, 38866)
Log a warning message when failed to find kind for resource in garbage collector controller
at this time, I do not think thirdparty api group version resources should be taken care by garbage collector controllers, and this line of call will fail actually: https://github.com/kubernetes/kubernetes/blob/master/pkg/controller/garbagecollector/garbagecollector.go#L565, and as a result, the garbagecollector controller failed to start.
Automatic merge from submit-queue
Remove jobs that do not exist from active list of CronJob
**What this PR does / why we need it**: This PR modifies the controller for CronJob to remove from the active job list any job that does not exist anymore, to avoid staying blocked in active state forever. See #37957.
**Which issue this PR fixes**: fixes#37957
**Special notes for your reviewer**:
**Release note**:
```
```
Automatic merge from submit-queue (batch tested with PRs 39284, 39367)
Remove HostRecord annotation (beta feature)
The annotation has made it to GA so this code should be deleted.
**Release note**:
```release-note
The 'endpoints.beta.kubernetes.io/hostnames-map' annotation is no longer supported. Users can use the 'Endpoints.subsets[].addresses[].hostname' field instead.
```
Automatic merge from submit-queue (batch tested with PRs 39092, 39126, 37380, 37093, 39237)
Endpoints with TolerateUnready annotation, should list Pods in state terminating
**What this PR does / why we need it**:
We are using preStop lifecycle hooks to gracefully remove a node from a cluster. This hook is potentially long running and after the preStop hook is fired, the DNS resolution of the soon to be stopped Pod is failing, which causes a failure there.
**Special notes for your reviewer**:
Would be great to backport that to 1.4, 1.3
**Release note**:
```release-note
Endpoints, that tolerate unready Pods, are now listing Pods in state Terminating as well
```
@bprashanth
Automatic merge from submit-queue (batch tested with PRs 39075, 39350, 39353)
Move pkg/api.{Context,RequestContextMapper} into pkg/genericapiserver/api/request
**Based on #39350**
Automatic merge from submit-queue
DaemonSet ObservedGeneration
Extracting ObserverdGeneration part from #31693. It also implements #7328 for DaemonSets.
cc @kargakis
Automatic merge from submit-queue (batch tested with PRs 39150, 38615)
Add work queues to PV controller
PV controller should not use Controller.Requeue, as as it is not available in
shared informers. We need to implement our own work queues instead, where we
can enqueue volumes/claims as we want.
PV controller should not use Controller.Requeue, as as it is not available in
shared informers. We need to implement our own work queues instead where we
can enqueue volumes/claims as we want.
Automatic merge from submit-queue
Avoid unnecessary memory allocations
Low-hanging fruits in saving memory allocations. During our 5000-node kubemark runs I've see this:
ControllerManager:
- 40.17% k8s.io/kubernetes/pkg/util/system.IsMasterNode
- 19.04% k8s.io/kubernetes/pkg/controller.(*PodControllerRefManager).Classify
Scheduler:
- 42.74% k8s.io/kubernetes/plugin/pkg/scheduler/algrorithm/predicates.(*MaxPDVolumeCountChecker).filterVolumes
This PR is eliminating all of those.
Automatic merge from submit-queue (batch tested with PRs 39093, 34273)
start breaking up controller manager into two pieces
This PR addresses: https://github.com/kubernetes/features/issues/88
This commit starts breaking the controller manager into two pieces, namely,
1. cloudprovider dependent piece
2. coudprovider agnostic piece
the controller manager has the following control loops -
- nodeController
- volumeController
- routeController
- serviceController
- replicationController
- endpointController
- resourceQuotaController
- namespaceController
- deploymentController
etc..
among the above controller loops,
- nodeController
- volumeController
- routeController
- serviceController
are cloud provider dependent. As kubernetes has evolved tremendously, it has become difficult
for different cloudproviders (currently 8), to make changes and iterate quickly. Moreover, the
cloudproviders are constrained by the kubernetes build/release lifecycle. This commit is the first
step in moving towards a kubernetes code base where cloud providers specific code will move out of
the core repository, and will be maintained by the cloud providers themselves.
I have added a new cloud provider called "external", which signals the controller-manager that
cloud provider specific loops are being run by another controller. I have added these changes in such
a way that the existing cloud providers are not affected. This change is completely backwards compatible, and does not require any changes to the way kubernetes is run today.
Finally, along with the controller-manager, the kubelet also has cloud-provider specific code, and that will be addressed in a different commit/issue.
@alena1108 @ibuildthecloud @thockin @dchen1107
**Special notes for your reviewer**:
@thockin - Im making this **WIP** PR to ensure that I don't stray too far from everyone's view of how we should make this change. As you can see, only one controller, namely `nodecontroller` can be disabled with the `--cloudprovider=external` flag at the moment. I'm working on cleaning up the `rancher-controller-manger` that I wrote to test this.
Secondly, I'd like to use this PR to address cloudprovider specific code in kubelet and api-server.
**Kubelet**
Kubelet uses provider specific code for node registration and for checking node-status. I thought of two ways to divide the kubelet:
- We could start a cloud provider specific kubelet on each host as a part of kubernetes, and this cloud-specific-kubelet does node registration and node-status checks.
- Create a kubelet plugin for each provider, which will be started by kubelet as a long running service. This plugin can be packaged as a binary.
I'm leaning towards the first option. That way, kubelet does not have to manage another process, and we can offload the process management of the cloud-provider-specific-kubelet to something like systemd.
@dchen1107 @thockin what do you think?
**Kube-apiserver**
Kube-apiserver uses provider specific code for distributing ssh keys to all the nodes of a cluster. Do you have any suggestions about how to address this?
**Release note**:
``` release-note
```
