karmada/docs/proposals/mcs/remove-derived-mcs.md

title	authors	reviewers	creation-date
Remove "derived-" prefix of MCS	@zishen	@kevin-wangzefeng @RainbowMango @XiShanYongYe-Chang @Garrybest @GitHubxsy	2023-06-30

Remove "derived-" prefix of MCS

Summary

As mentioned in issue-2384, currently Karmada uses ServiceExport, ServiceImport and other CR to form mcs-controller and manage multi-cluster services. But in the cluster where the service is not located, we need to use a different IP or a service name with a "derived-" prefix to access. We hope to reduce this difference so that different clusters can use the same domain name to access service.

Motivation

Currently, in different cluster members of Karmada, if you need to access the service across clusters, you need to use different ip or domain names with a "derived-" prefix to access. This will increase the complexity of the user's use. We hope that users can use the same domain name to achieve cross-cluster access.

Goals

• Keep the original access method. Avoid the need to make modifications to the customer's system after the upgrade.
• Added the use of a unified domain name for cross-cluster access. The domain name is the access domain name of the cluster where the service resides.

Non-Goals

none

Proposal

User Stories

According to the current Karmada's mcs configuration , if a service in member1, the user needs to access the service in member1, needs to use the domain name: "serve.default.svc.cluster.local", and the domain name in member2 will be: "derived-serve.default.svc.cluster.local". As a result, when customers use the service, they need to add the prefix "derived-" to access the domain name of the service. This brings additional inconvenience. So I hope to be able to:

1. Provide a unified access method across clusters.

2. The original access method remains unchanged. So that existing business can also work normally.

Notes/Constraints/Caveats (Optional)

1. The coreDNS version must be v1.9.3 or above, and the matching k8s version is at v1.21.0 or above, otherwise the multicluster feature is not supported.

Risks and Mitigations

Design Details

API change

none

Components change

Here is give the implementation plan:ServiceImport + karmada.

1) Use ServiceImport to generate rr records

This solution is mainly realized by combining the multicluster plug-in of corends. It is divided into two parts: coreDNS transformation and karmada. For the coreDNS, here is two implementation plan.

(1) coreDNS+multicluster

This solution is mainly realized by the multicluster plug-in of corends.

Principle Description

In order to adapt to mcs, coreDNS launched the multicluster plug-in. After the plug-in completes the configuration and deployment, it will parse ServiceImport. The main content of the analysis is the name, namespace and ips, and an rr record is generated accordingly. Then client pod can cross-cluster access services.

The workflow of the multicluster plugin is as follows:

The diagram is explained below:

• Here give a service named foo in member1. We should use the full domain name: foo.default.svc.cluster.local to access this service. But we cannot use the same domain name in member2.

• Karmada exports the service through ServiceExport and imports it into member2 through ServiceImport. At this time, the shadow service derived-foo will appear in member2. User in member2 can access to the foo service in member1 by using derived-foo.default.svc.cluster.local.

• After the coreDNS installed with multicluster found the ServiceImport had been created, it will analyze name, namespace, and ips fields of the ServiceImport and generate the rr records. In this example, the ips in ServiceImport can be the clusterIP of derived-foo.

• When the client accesses the service across clusters (the suffix may be different from the local access), coreDNS can access it according to the corresponding rr record.

Compilation-instructions

According to the documentation on the official website multicluster plugin. Download coreDNS, the corresponding version of k8s 1.26 is v1.9.3 (the latest v1.10.1 is also available):

git clone https://github.com/coredns/coredns
cd coredns
git checkout v1.9.3

• Add multicluster plugin

Open the plugin.cfg file and add multicluster in it as follow:

...
kubernetes:kubernetes
multicluster: github.com/coredns/multicluster
...

• Execute compilation

make

After the compilation is successful, the corresponding binary will be generated in the root directory.

Mirror making and installation or upgrading please refer to coreDNS github website.

Configuration-instructions

The suffix clusterset.local is used as an example to illustrate

• Configure coredns permissions in member2

kubectl edit clusterrole system:coredns

Add the following to the end

...
- apiGroups:
   - multicluster.x-k8s.io
   resources:
   - serviceimports
   verbs:
   - list
   - watch

• Configure multicluster zone rules

Add multicluster processing rules to corefile in coredns

kubectl edit configmap coredns -n kube-system

  ....
    Corefile: |
     .:53 {
         errors
         health {
            lameduck 5s
         }
         ready
         multicluster clusterset.local # Add this line.
         kubernetes cluster.local in-addr.arpa ip6.arpa {
            pods insecure
            fallthrough in-addr.arpa ip6.arpa
            ttl 30
         }
         prometheus:9153
         forward ./etc/resolv.conf {
            max_concurrent 1000
         }
         cache 30
         the loop
         reload
         load balance
     }     
...

• Configure dns search item "clusterset.local"

Configure node /etc/resolv.conf file

Edit the /etc/resolv.conf file and add the "clusterset.local" entry to give priority to local services when using the domain name foo.default.svc. It is worth reminding that the remote service is only accessed when the local service does not exist, not unavailable.

root@btg:/# cat /etc/resolv.conf
search default.svc.cluster.local svc.cluster.local cluster.local clusterset.local
...

This configuration method needs to configure the /etc/resolv.conf file on the client node. If you don't want to change the node's resolv.conf file, you can also use the following method.

Configure pod's dnsConfig in yaml

That is, add the configuration item searches of dnsConfig in yaml, and set the value to clusterset.local, as shown below:

...
       dnsPolicy: ClusterFirst
       dnsConfig:
         searches:
           - clusterset.local
...

Access method

Since the search function of dns is used, the foo service access method is:

1. foo.default.svc access. This method gives priority to local access, and if the local service with the same name is inexistence, it will access the remote cluster service.

2. foo.default.svc.clusterset.local full domain name access. The request will go directly to the remote server for processing.

Advantages and disadvantages

shortcoming:

1. Using foo.default.svc access cannot achieve load balancing. The remote service will only be accessed when the local service is inexistence. Therefore, if you need to achieve local and remote access load balancing, you need to cooperate with other solutions.

2. To use foo.default.svc, you need to modify the dns configuration. If you use the resolv.conf method, you need to modify the resolv.conf file of each possible nodes. If you use the dnsConfig method, you need to modify the yaml or api of the delivered pod.

3. The multicluster plug-in cannot be loaded dynamically. Customers need to recompile coredns and modify the coredns configuration of member clusters.

4. The original access method will be changed. This solution uses different suffixes to distinguish single-cluster Service and multi-cluster Service. Therefore, the access domain name foo.default.svc.cluster.local will become foo.default.svc.clusterset.local. Customers need to modify the code to adapt.

advantage:

1. The impact is controllable. No new components will be added. The way coredns adds multicluster is to comply with the official proposal of coredns, and the code is open.

2. The amount of modification is small. It can be done by using the existing code and scheme of multicluster.

(2) Additional dns server

Similar to submariner's SERVICE DISCOVERY. It is necessary to develop an extended dns resolver, and install the corresponding plug-in in coreDNS to transfer the domain name with the specified suffix (such as clusterset.local) to this additional dns resolver. Therefore, the domain name can be resolved by dns.

The workflow of the additional dns server as follows:

The advantages and disadvantages of this method are as follows:

advantage:

1. Load balancing can be achieved. Since the extended dns resolver is customized, customers can achieve load balancing here.

2. Access in the original way can be realized. Rules for local services and remote services can be determined in custom plug-ins. Therefore, the original access method can be maintained, in this case is foo.default.svc.cluster.local.

3. No other configuration of dns is required. Since there is a custom dns resolver, there is no need to configure dns search. Therefore, there is no configuration of dns search in solution one.

shortcoming:

1. The coreDNS component needs to be modified. It is also necessary to recompile coreDNS in order to install the corresponding plug-ins, and also need to modify the corresponding coreDNS configuration.

2. New components need to be deployed in member clusters. This solution requires a dns extension resolver to be installed.

2) Karmada adaptation

What needs to be done in the karmada part is to fill in the ips field of the ServiceImport which imported to the member cluster on the existing basis and correspond to the clusterIP of derived-foo. Its flow diagram is as follows:

As shown in the figure above, we still take the foo service as an example. After configuring coreDNS, we add the logic to obtain the derived-foo service in the cluster members in the ServiceImport-controller of the control plane.

1. Wait for cluster member config the derived-foo service's clusterIP .

2. The controller of the control plane obtains the value, and fills the value into the work corresponding to the ServiceImport of the control plane.

3. According to the synchronization principle of Karmada, the ServiceImport object in the member cluster will be filled with the clusterIP value of derived-foo.

That is, the effect of ServiceImport after completion needs to be:

apiVersion: multicluster.x-k8s.io/v1alpha1
kind: ServiceImport
metadata:
   name: foo
   namespace: default
spec:
   type: ClusterSetIP
   ports:
   - name: "http"
     port: 80
     protocol: TCP
   ips:
   - "10.10.0.5" # clusterIP for cross-cluster access

3) Suggested operation

The recommended configuration instructions for the coreDNS+multicluster plug-in solution are given below.

Environment preparation:

A normal Karmada cluster, and can perform existing mcs operations.

The following uses the cluster member2 where the client resides as an example to access the fooservice using the suffix clusterset.local. This solution needs to be configured as follows:

coredns compile

Refer to subsection Compilation-Instructions.

coredns configuration

Refer to subsection Configuration-instructions.

Client pod configuration

Taking yaml as an example, you need to add the configuration item searches of dnsConfig in yaml, and set the value to clusterset.local. A simple and complete example is as follows:

apiVersion: apps/v1
kind: Deployment
metadata:
   name: dns-test
spec:
   replicas: 2
   selector:
     matchLabels:
       app: dns-test
   template:
     metadata:
       labels:
         app: dns-test
     spec:
       containers:
       - name: dns-test
         image: ubuntu:18.04
         command:
         - "/bin/bash"
         - "-c"
         args: [ "while true; do sleep 3; cat /etc/resolv.conf; done" ]
       dnsPolicy: ClusterFirst
       dnsConfig:
         searches:
           -clusterset.local
         options:
           - name: timeout
             value: "2

Test Plan

• All current testing should be passed, no break change would be involved by this feature.

Alternatives

Here are another two ways I know:

First, using the same service name, not service withderived- prefix on other cluster. More details see bivas 's pr proposal for native service discovery

Second, install and using submariner .