kubernetes: make the terminating-pod policy configurable in pods mode

未关闭
#224 0 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
53/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
elixir, kubernetes

调研方向

Start by reading the Kubernetes strategy code for pods mode and the change associated with issue #114, which introduced the current terminating-pod behavior. Trace how pod deletion timestamps and running state affect discovery, then check the strategy tests and configuration documentation. Done means an opt-in :include setting keeps running terminating pods visible while the default :exclude behavior remains unchanged.

由索引模型根据 Issue 内容生成。

描述

Steps to reproduce
  • Configuration Used: Cluster.Strategy.Kubernetes with kubernetes_ip_lookup_mode: :pods.
  • Strategy Used: Cluster.Strategy.Kubernetes (pods mode).
  • Errors/Incorrect Behaviour Encountered: the terminating-pod policy is fixed. 3.6.0 excludes pods with a metadata.deletionTimestamp, which is the right default for deployments that want an unreachable target removed quickly.
Description of issue

This is an improvement request, not a bug: make the terminating-pod policy configurable in pods mode, keeping 3.6.0 as the default. 3.6.0 introduced the current behavior to address #114, but it also removed the ability to express the opposite case.

The use case is a distributed application where a node has to stay reachable while it shuts down. Process groups are one example: if a node leaves discovery as soon as its pod is marked for deletion, it can disappear from :pg before it has finished draining, and peers may start replacement processes while the old ones are still unwinding. Other examples are a node handing off a lease, or a supervised job it owns, before it goes away.

Today there is no way to keep that node visible: pods mode removes it, and endpoints and EndpointSlice modes drop terminating endpoints too. A field selector can't express it either, because a terminating pod keeps phase: Running and metadata.deletionTimestamp is not a selectable field.

Would an opt-in option be acceptable?

kubernetes_pod_terminating: :exclude | :include
  • :exclude (default): current behavior, remove a pod as soon as it has a deletionTimestamp.
  • :include: keep it until it is no longer Running, which is what a draining node needs.

The documentation is correct for the default; this would add an option instead of changing behavior.

主要语言
Elixir
星标
2.2k
派生
202
平均合并
2 分钟
30 天内合并 PR
1

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

bitwalker/libcluster 的其他 Issue

查看 bitwalker/libcluster 的全部 Issue

相似的 Issue

更多 Elixir Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。