Files
website/content/zh/docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods.md
Zach Corleissen abcee2dccd Update localization guidelines (#10485)
* Update localization guidelines for language labels

Continuing work

Continuing work

Continuing work

More work in progress

Add local OWNERS folders

Add an OWNERS file to Chinese

Remove shortcode for repos

Add Japanese

Alphabetize languages, change weights accordingly

More updates

Add Korean in Korean

Add English to languageName

Feedback from gochist

Move Chinese content from cn/ to zh/

Move OWNERS from cn/ to zh/

Resolve merge conflicts by updating from master

Add files back in to prep for resolution

After rebase on upstream/master, remove files

Review and update localization guidelines

Feedback from gochist, tnir, cstoku

Add a trailing newline to content/ja/OWNERS

Add a trailing newline to content/zh/OWNERS

Drop requirement for GH repo project

Clarify language about forks/branches

Edits and typos

Remove a shortcode specific to a multi-repo language setup

Update aliases and owners

Add explicit OWNERS for content/en

Migrate content from Chinese repo, update regex in config.toml

Remove untranslated strings

Add trailing newline to content/en/OWNERS

Add trailing newlines to OWNERS files

add Jaguar project description (#10433)

* add Jaguar project description

[Jaguar](https://gitlab.com/sdnlab/jaguar) is an open source solution for Kubernetes's network based on OpenDaylight.
Jaguar provides overlay network using vxlan and Jaguar CNIPlugin provides one IP address per pod.

* Minor newline tweak

blog post for azure vmss (#10538)

Add microk8s to pick-right-solution.md (#10542)

* Add microk8s to pick-right-solution.md

Microk8s is a single-command installation of upstream Kubernetes on any Linux and should be included in the list of local-machine solutions.

* capitalized Istio

Add microk8s to foundational.md (#10543)

* Add microk8s to foundational.md

Adding microk8s as credible and stable alternative to get started with Kubernetes on a local machine. This is especially attractive for those not wanting to incur the overhead of running a VM for a local cluster.

* Update foundational.md

Thank you for your suggestions! LMK if this works now?

* Rewrote first paragraph

And included a bullet list of features of microk8s

* Copyedit

fix typo (#10545)

Fix the kubectl subcommands links. (#10550)

Signed-off-by: William Zhang <warmchang@outlook.com>

Fix command issue (#10515)

Signed-off-by: mooncake <xcoder@tenxcloud.com>

remove imported community files per issue 10184 (#10501)

networking.md: Markdown fix (#10498)

Fix front matter, federation command-line tools (#10500)

Clean up glossary entry (#10399)

update slack link (#10536)

typo in StatefulSet docs (#10558)

fix discription about horizontal pod autoscale (#10557)

Remove redundant symbols (#10556)

Fix issue #10520 (#10554)

Signed-off-by: William Zhang <warmchang@outlook.com>

Update api-concepts.md (#10534)

Revert "Fix command issue (#10515)"

This reverts commit c02a7fb9f9.

Update memory-constraint-namespace.md (#10530)

update memory request to 100MiB corresponding the yaml content

Blog: Introducing Volume Snapshot Alpha for Kubernetes (#10562)

* blog post for azure vmss

* snapshot blog post

Resolve merge conflicts in OWNERS*

Minor typo fix (#10567)

Not sure what's supposed to be here, proposing removing it.

* Feedback from gochist

Tweaks to feedback

* Feedback from ClaudiaJKang
2018-10-12 14:25:01 -07:00

2.9 KiB
Raw Permalink Blame History

approvers, title
approvers title
davidopp
filipg
piosz
关键插件 Pod 的调度保证

{{< toc >}}

概览

除了 Kubernetes 核心组件,像运行在 master 机器上的 api-server、scheduler、controller-manager,由于各种原因,还有很多插件必须运行在一个普通的集群节点上(而不是 Kubernetes master)。 这些插件中的一些对于一个功能完备的集群来说是非常关键的,例如 Heapster、DNS 以及 UI。 如果一个关键的插件被移除(或者手动,或是类似升级这样具有副作用的其它操作),或者变成挂起状态(例如,当集群利用率过高,以及或者其它被调度到该空间中的挂起 Pod 被清理关键插件 Pod 给移除,或者由于其它原因导致节点上可用资源的总量发生变化),集群可能会停止正常工作。

二次调度器:关键插件的调度保证

二次调度器确保关键的插件总是能被调度(假定普通的 Pod 不存在时,集群具有足够的资源去运行关键插件 Pod)。 如果调度器确定没有节点有足够资源去运行关键插件 Pod,假使有 Pod 已经运行在集群中(通过将关键插件 Pod 的条件 PodScheduled 设置为 false,原因设置为 Unschedulable 来指示),这时二次调度器通过清理一些 Pod 尽量去释放空间,然后调度器将调度该插件 Pod。

为了避免这种情况,当另一个 Pod 被调度到该空间,为了该关键插件,被选择的节点导致临时变成 "CriticalAddonsOnly" 的 taint(查看 更多详情)。 每个关键插件不得不容忍它,然而其它一些 Pod 不可能容忍该 taint。一旦该插件被成功调度,该 taint 会被移除。

警告: 当前对选中哪个节点,以及为了调度该关键 Pod 杀掉哪个 Pod 并没有任何保证,所以如果启用了二次调度器,任何 Pod 都可能被偶然的杀掉以达到目的。

配置

二次调度器应该被 作为 static Pod 默认启用。 它没有任何面向用户的配置(组件配置),或者 API,可以通过如下方式来禁用:

  • 在集群安装过程中通过设置 ENABLE_RESCHEDULER 标志为 false
  • 在运行中的集群中从 master 节点删除它的 manifest(默认路径 /etc/kubernetes/manifests/rescheduler.manifest

标记关键插件

想变成关键插件,必须运行在 kube-system Namespace 中(是基于标志可配置的),并且:

  • scheduler.alpha.kubernetes.io/critical-pod annotation 设置为空字符串,并且
  • 将 PodSpec 的 tolerations 字段设置为 [{"key":"CriticalAddonsOnly", "operator":"Exists"}]

第一个表示是一个关键 Pod。第二个是二次调度器算法必需的。