federation-phase-1

Ubernetes Design Spec (phase one)

Huawei PaaS Team

INTRODUCTION

In this document we propose a design for the “Control Plane” of Kubernetes (K8S) federation (a.k.a. “Ubernetes”). For background of this work please refer to this proposal. The document is arranged as following. First we briefly list scenarios and use cases that motivate K8S federation work. These use cases drive the design and they also verify the design. We summarize the functionality requirements from these use cases, and define the “in scope” functionalities that will be covered by this design (phase one). After that we give an overview of the proposed architecture, API and building blocks. And also we go through several activity flows to see how these building blocks work together to support use cases.

REQUIREMENTS

There are many reasons why customers may want to build a K8S federation:

  • High Availability: Customers want to be immune to the outage of a single availability zone, region or even a cloud provider.
  • Sensitive workloads: Some workloads can only run on a particular cluster. They cannot be scheduled to or migrated to other clusters.
  • Capacity overflow: Customers prefer to run workloads on a primary cluster. But if the capacity of the cluster is not sufficient, workloads should be automatically distributed to other clusters.
  • Vendor lock-in avoidance: Customers want to spread their workloads on different cloud providers, and can easily increase or decrease the workload proportion of a specific provider.
  • Cluster Size Enhancement: Currently K8S cluster can only support a limited size. While the community is actively improving it, it can be expected that cluster size will be a problem if K8S is used for large workloads or public PaaS infrastructure. While we can separate different tenants to different clusters, it would be good to have a unified view.

Here are the functionality requirements derived from above use cases:

  • Clients of the federation control plane API server can register and deregister clusters.
  • Workloads should be spread to different clusters according to the workload distribution policy.
  • Pods are able to discover and connect to services hosted in other clusters (in cases where inter-cluster networking is necessary, desirable and implemented).
  • Traffic to these pods should be spread across clusters (in a manner similar to load balancing, although it might not be strictly speaking balanced).
  • The control plane needs to know when a cluster is down, and migrate the workloads to other clusters.
  • Clients have a unified view and a central control point for above activities.

SCOPE

It’s difficult to have a perfect design with one click that implements all the above requirements. Therefore we will go with an iterative approach to design and build the system. This document describes the phase one of the whole work. In phase one we will cover only the following objectives:

  • Define the basic building blocks and API objects of control plane
  • Implement a basic end-to-end workflow
    • Clients register federated clusters
    • Clients submit a workload
    • The workload is distributed to different clusters
    • Service discovery
    • Load balancing

The following parts are NOT covered in phase one:

  • Authentication and authorization (other than basic client authentication against the ubernetes API, and from ubernetes control plane to the underlying kubernetes clusters).
  • Deployment units other than replication controller and service
  • Complex distribution policy of workloads
  • Service affinity and migration

ARCHITECTURE

The overall architecture of a control plane is shown as following:

Ubernetes Architecture

Some design principles we are following in this architecture:

  1. Keep the underlying K8S clusters independent. They should have no knowledge of control plane or of each other.
  2. Keep the Ubernetes API interface compatible with K8S API as much as possible.
  3. Re-use concepts from K8S as much as possible. This reduces customers’ learning curve and is good for adoption. Below is a brief description of each module contained in above diagram.

Ubernetes API Server

The API Server in the Ubernetes control plane works just like the API Server in K8S. It talks to a distributed key-value store to persist, retrieve and watch API objects. This store is completely distinct from the kubernetes key-value stores (etcd) in the underlying kubernetes clusters. We still use etcd as the distributed storage so customers don’t need to learn and manage a different storage system, although it is envisaged that other storage systems (consol, zookeeper) will probably be developedand supported over time.

Ubernetes Scheduler

The Ubernetes Scheduler schedules resources onto the underlying Kubernetes clusters. For example it watches for unscheduled Ubernetes replication controllers (those that have not yet been scheduled onto underlying Kubernetes clusters) and performs the global scheduling work. For each unscheduled replication controller, it calls policy engine to decide how to spit workloads among clusters. It creates a Kubernetes Replication Controller on one ore more underlying cluster, and post them back to etcd storage.

One subtlety worth noting here is that the scheduling decision is arrived at by combining the application-specific request from the user (which might include, for example, placement constraints), and the global policy specified by the federation administrator (for example, “prefer on-premise clusters over AWS clusters” or “spread load equally across clusters”).

Ubernetes Cluster Controller

The cluster controller performs the following two kinds of work:

  1. It watches all the sub-resources that are created by Ubernetes components, like a sub-RC or a sub-service. And then it creates the corresponding API objects on the underlying K8S clusters.
  2. It periodically retrieves the available resources metrics from the underlying K8S cluster, and updates them as object status of the cluster API object. An alternative design might be to run a pod in each underlying cluster that reports metrics for that cluster to the Ubernetes control plane. Which approach is better remains an open topic of discussion.

Ubernetes Service Controller

The Ubernetes service controller is a federation-level implementation of K8S service controller. It watches service resources created on control plane, creates corresponding K8S services on each involved K8S clusters. Besides interacting with services resources on each individual K8S clusters, the Ubernetes service controller also performs some global DNS registration work.

API OBJECTS

Cluster

Cluster is a new first-class API object introduced in this design. For each registered K8S cluster there will be such an API resource in control plane. The way clients register or deregister a cluster is to send corresponding REST requests to following URL: /api/{$version}/clusters. Because control plane is behaving like a regular K8S client to the underlying clusters, the spec of a cluster object contains necessary properties like K8S cluster address and credentials. The status of a cluster API object will contain following information:

  1. Which phase of its lifecycle
  2. Cluster resource metrics for scheduling decisions.
  3. Other metadata like the version of cluster

$version.clusterSpec

Name
Description
Required
Schema
Default
Address
address of the cluster
yes
address

Credential
the type (e.g. bearer token, client certificate etc) and data of the credential used to access cluster. It's used for system routines (not behalf of users)
yes
string

$version.clusterStatus

Name
Description
Required
Schema
Default
Phase
the recently observed lifecycle phase of the cluster
yes
enum

Capacity
represents the available resources of a cluster
yes
any

ClusterMeta
Other cluster metadata like the version
yes
ClusterMeta

For simplicity we didn’t introduce a separate “cluster metrics” API object here. The cluster resource metrics are stored in cluster status section, just like what we did to nodes in K8S. In phase one it only contains available CPU resources and memory resources. The cluster controller will periodically poll the underlying cluster API Server to get cluster capability. In phase one it gets the metrics by simply aggregating metrics from all nodes. In future we will improve this with more efficient ways like leveraging heapster, and also more metrics will be supported. Similar to node phases in K8S, the “phase” field includes following values:

  • pending: newly registered clusters or clusters suspended by admin for various reasons. They are not eligible for accepting workloads
  • running: clusters in normal status that can accept workloads
  • offline: clusters temporarily down or not reachable
  • terminated: clusters removed from federation

Below is the state transition diagram.

Cluster State Transition Diagram

Replication Controller

A global workload submitted to control plane is represented as a replication controller in the Cluster Federation control plane. When a replication controller is submitted to control plane, clients need a way to express its requirements or preferences on clusters. Depending on different use cases it may be complex. For example:

  • This workload can only be scheduled to cluster Foo. It cannot be scheduled to any other clusters. (use case: sensitive workloads).
  • This workload prefers cluster Foo. But if there is no available capacity on cluster Foo, it’s OK to be scheduled to cluster Bar (use case: workload )
  • Seventy percent of this workload should be scheduled to cluster Foo, and thirty percent should be scheduled to cluster Bar (use case: vendor lock-in avoidance). In phase one, we only introduce a clusterSelector field to filter acceptable clusters. In default case there is no such selector and it means any cluster is acceptable.

Below is a sample of the YAML to create such a replication controller.

apiVersion: v1
kind: ReplicationController
metadata:
  name: nginx-controller
spec:
  replicas: 5
  selector:
    app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: nginx
        ports:
        - containerPort: 80
      clusterSelector:
      name in (Foo, Bar)

Currently clusterSelector (implemented as a LabelSelector) only supports a simple list of acceptable clusters. Workloads will be evenly distributed on these acceptable clusters in phase one. After phase one we will define syntax to represent more advanced constraints, like cluster preference ordering, desired number of split workloads, desired ratio of workloads spread on different clusters, etc.

Besides this explicit “clusterSelector” filter, a workload may have some implicit scheduling restrictions. For example it defines “nodeSelector” which can only be satisfied on some particular clusters. How to handle this will be addressed after phase one.

Federated Services

The Service API object exposed by the Cluster Federation is similar to service objects on Kubernetes. It defines the access to a group of pods. The federation service controller will create corresponding Kubernetes service objects on underlying clusters. These are detailed in a separate design document: Federated Services.

Pod

In phase one we only support scheduling replication controllers. Pod scheduling will be supported in later phase. This is primarily in order to keep the Cluster Federation API compatible with the Kubernetes API.

ACTIVITY FLOWS

Scheduling

The below diagram shows how workloads are scheduled on the Cluster Federation control
plane:

  1. A replication controller is created by the client.
  2. APIServer persists it into the storage.
  3. Cluster controller periodically polls the latest available resource metrics from the underlying clusters.
  4. Scheduler is watching all pending RCs. It picks up the RC, make policy-driven decisions and split it into different sub RCs.
  5. Each cluster control is watching the sub RCs bound to its corresponding cluster. It picks up the newly created sub RC.
  6. The cluster controller issues requests to the underlying cluster API Server to create the RC. In phase one we don’t support complex distribution policies. The scheduling rule is basically:
    1. If a RC does not specify any nodeSelector, it will be scheduled to the least loaded K8S cluster(s) that has enough available resources.
    2. If a RC specifies N acceptable clusters in the clusterSelector, all replica will be evenly distributed among these clusters.

There is a potential race condition here. Say at time T1 the control plane learns there are m available resources in a K8S cluster. As the cluster is working independently it still accepts workload requests from other K8S clients or even another Cluster Federation control plane. The Cluster Federation scheduling decision is based on this data of available resources. However when the actual RC creation happens to the cluster at time T2, the cluster may don’t have enough resources at that time. We will address this problem in later phases with some proposed solutions like resource reservation mechanisms.

Federated Scheduling

Service Discovery

This part has been included in the section “Federated Service” of document “Federated Cross-cluster Load Balancing and Service Discovery Requirements and System Design)”. Please refer to that document for details.