Kubernetes (K8s) [1]

Feature of: Service Discovery

Quick Links: K8s Website | K8s Documentation | K8s GitHub

Ocelot will call the K8s endpoints API in a given namespace to get all of the endpoints for a pod and then load balance across them. Ocelot used to use the services API to send requests to the K8s service but this was changed in pull request 1134 because the service did not load balance as expected.

Our NuGet Ocelot.Provider.Kubernetes extension package is based on the KubeClient package. For a comprehensive understanding, it is essential refer to the KubeClient documentation.

Install

The first thing you need to do is install the package that provides support in Ocelot:

Install-Package Ocelot.Provider.Kubernetes

AddKubernetes(bool) method

public static class OcelotBuilderExtensions
{
    public static IOcelotBuilder AddKubernetes(this IOcelotBuilder builder, bool usePodServiceAccount = true);
}

This extension-method adds K8s services with or without using a pod service account. Then add the following to your Program:

builder.Services
    .AddOcelot(builder.Configuration)
    .AddKubernetes(); // usePodServiceAccount is true

If you have services deployed in Kubernetes, you will normally use the naming service to access them.

1. By default the useServiceAccount argument is true, which means that Service Account using Pod to access the service of the K8s cluster needs to be Service Account based on RBAC authorization:

   You can replicate a Permissive using RBAC role bindings (see Permissive RBAC Permissions), K8s API server and token will read from pod.

   kubectl create clusterrolebinding permissive-binding --clusterrole=cluster-admin --user=admin --user=kubelet --group=system:serviceaccounts
   

   Finally, it creates the KubeClient from pod service account.

2. When the useServiceAccount argument is false, you need to provide KubeClientOptions to create KubeClient using them. You have to bind the options configuration section for the DI IOptions<KubeClientOptions> interface or register a custom action to initialize the options:

    Action<KubeClientOptions> configureKubeClient = opts =>
    {
        opts.ApiEndPoint = new UriBuilder("https", "my-host", 443).Uri;
        opts.AccessToken = "my-token";
        opts.AuthStrategy = KubeAuthStrategy.BearerToken;
        opts.AllowInsecure = true;
    };
    builder.Services
        .AddOptions<KubeClientOptions>()
        .Configure(configureKubeClient); // manual binding options via IOptions<KubeClientOptions>
    builder.Services
        .AddOcelot(builder.Configuration)
        .AddKubernetes(false); // don't use pod service account, and IOptions<KubeClientOptions> is reused
   

   Note, this could also be written like this (shortened version):

   builder.Services
       .AddKubeClientOptions(opts =>
       {
           opts.ApiEndPoint = new UriBuilder("https", "my-host", 443).Uri;
           opts.AuthStrategy = KubeAuthStrategy.BearerToken;
           opts.AccessToken = "my-token";
           opts.AllowInsecure = true;
       })
       .AddOcelot(builder.Configuration)
       .AddKubernetes(false); // don't use pod service account, and client options provided via AddKubeClientOptions
   

   Finally, it creates the KubeClient from your options.

   Note 1: For understanding the IOptions<TOptions> interface, please refer to the Microsoft Learn documentation: Options pattern in .NET.

   Note 2: Please consider this Case 2 as an example of manual setup when you do not use a pod service account. We recommend using our official extension method, which receives an Action<KubeClientOptions> argument with your options: refer to the AddKubernetes(Action<KubeClientOptions>) method 2 below.

AddKubernetes(Action<KubeClientOptions>) method [2]

public static class OcelotBuilderExtensions
{
    public static IOcelotBuilder AddKubernetes(this IOcelotBuilder builder, Action<KubeClientOptions> configureOptions, /*optional params*/);
}

This extension method adds K8s services without using a pod service account, explicitly calling an action to initialize configuration options for KubeClient. It operates in two modes:

1. If configureOptions is provided (action is not null), it calls the action, ignoring all optional arguments.

   Action<KubeClientOptions> configureKubeClient = opts =>
   {
       opts.ApiEndPoint = new UriBuilder("https", "my-host", 443).Uri;
       // ...
   };
   builder.Services
       .AddOcelot(builder.Configuration)
       .AddKubernetes(configureKubeClient); // without optional arguments
   

Note: Optional arguments do not make sense; all settings are defined inside the configureKubeClient action.

2. If configureOptions is not provided (action is null), it reads the global ServiceDiscoveryProvider Configuration options and reuses them to initialize the following properties: ApiEndPoint, AccessToken, and KubeNamespace, finally initializing the rest of the properties with optional arguments.

   builder.Services
       .AddOcelot(builder.Configuration)
       .AddKubernetes(null, allowInsecure: true, /*optional args*/) // shortened version
       // or
       .AddKubernetes(configureOptions: null, allowInsecure: true, /*optional args*/); // long version
   

Note: Optional arguments must be used here in addition to the options coming from the global ServiceDiscoveryProvider Configuration. Find the comprehensive documentation in the C# code of the AddKubernetes methods.

Configuration

The following examples show how to set up a route that will work in Kubernetes. The most important thing is the ServiceName which is made up of the Kubernetes service name. We also need to set up the ServiceDiscoveryProvider in GlobalConfiguration.

Regarding global and route configurations, if your downstream service resides in a different namespace, you can override the global setting at the route level by specifying a ServiceNamespace.

"Routes": [
  {
    "ServiceName": "my-service",
    "ServiceNamespace": "my-namespace"
  }
]

Kube provider

The example here shows a typical configuration:

"Routes": [
  {
    "ServiceName": "my-service",
    // ...
  }
],
"GlobalConfiguration": {
  "ServiceDiscoveryProvider": {
    "Scheme": "https",
    "Host": "my-host",
    "Port": 443,
    "Token": "my-token",
    "Namespace": "Dev",
    "Type": "Kube"
  }
}

Service deployment in Dev namespace, and discovery provider type is Kube, you also can set PollKube or WatchKube provider type.

Note 1: Scheme, Host, Port, and Token are not used if usePodServiceAccount is true when KubeClient is created from a pod service account. Please refer to the Install section for technical details.

Note 2: The Kube provider searches for the service entry using ServiceName and then retrieves the first available port from the EndpointSubsetV1.Ports collection. Therefore, if the port name is not specified, the default downstream scheme will be http; Please refer to the “Downstream Scheme vs Port Names” section for technical details.

PollKube provider [3]

You use Ocelot to poll Kubernetes for latest service information rather than per request. If you want to poll Kubernetes for the latest services rather than per request (default behaviour of the Kube provider) then you need to set the following configuration:

"ServiceDiscoveryProvider": {
  "Namespace": "dev",
  "Type": "PollKube",
  "PollingInterval": 100 // ms
}

The polling interval is in milliseconds and tells Ocelot how often to call Kubernetes for changes in service configuration.

Note, there are tradeoffs here. If you poll Kubernetes, it is possible Ocelot will not know if a service is down depending on your polling interval and you might get more errors than if you get the latest services per request. This really depends on how volatile your services are. We doubt it will matter for most people and polling may give a tiny performance improvement over calling Kubernetes per request. There is no way for Ocelot to work these out for you, except perhaps through a discussion.

WatchKube provider [4]

With this configuration, Kubernetes API “watch requests” are used to fetch service configuration. Essentially, it establishes one streamed HTTP connection with the Kubernetes API per downstream service. Changes streamed through this connection will be used to update the list of available endpoints.

"ServiceDiscoveryProvider": {
  "Namespace": "dev",
  "Type": "WatchKube"
}

Note

The WatchKube provider is specifically designed for high-load Ocelot vs. Kubernetes environments with high RPS ratios. To better understand which type is suitable for your needs, we have added a table Comparing providers.

The provider has an implicit configuration for fine-tuned watching, which are available and can only be initialized in C# code.

• WatchKube.FirstResultsFetchingTimeoutSeconds: This is the default number of seconds to wait after Ocelot starts, following the provider’s creation, to fetch the first result from the Kubernetes endpoint. ¹

• WatchKube.FailedSubscriptionRetrySeconds: This is the default number of seconds to wait before scheduling the next retry for the subscription operation. ¹

¹ For both static int properties, the default value is 1 (one) second. The constraint ensures that the assigned value is greater than or equal to 1 (one). Therefore, the minimum value is 1 (one) second.

Comparing providers

This table explains the most important indicators that may influence Ocelot vs. Kubernetes deployment or DevOps strategy. The evolution path of all providers follows: Kube -> PollKube -> WatchKube, with WatchKube being the most advanced provider.

Indicators \ Providers	Kube	PollKube	WatchKube
Extra latency	One hop per route	-	-
Speed of response to endpoints changes	High	Low 1	High
Pressure on Kubernetes API	High	Low 1	Low
Ocelot load (estimated) 2	< 1000 RPS	> 1000 RPS	> 5000 RPS
Ocelot deployment 3	Single instance	Multiple instances	Cluster of instances

¹ Depends on the PollingInterval option.

² Please consider this a rough load estimation, as our team has not provided any tests or benchmarks.

³ The term “instance” refers to an Ocelot instance, not a Kubernetes one.

Downstream Scheme vs Port Names [5]

Kubernetes configuration permits the definition of multiple ports with names for each address of an endpoint subset. When binding multiple ports, you assign a name to each subset port. To allow the Kube provider to recognize the desired port by its name, you need to specify the DownstreamScheme with the port’s name; if not, the collection’s first port entry will be chosen by default.

For instance, consider a service on Kubernetes that exposes two ports: https for 443 and http for 80, as follows:

Name:         my-service
Namespace:    default
Subsets:
  Addresses:  10.1.161.59
  Ports:
    Name   Port  Protocol
    ----   ----  --------
    https  443   TCP
    http   80    TCP

When you need to use the http port while intentionally bypassing the default https port (first one), you must define DownstreamScheme to enable the provider to recognize the desired http port by comparing DownstreamScheme with the port name as follows:

"Routes": [
  {
    "ServiceName": "my-service",
    "DownstreamScheme": "http", // port name -> http -> port is 80
  }
]

Note

In the absence of a specified DownstreamScheme (the default behavior), the Kube provider—as well as other providers—will select the first available port from the EndpointSubsetV1.Ports collection. Consequently, if the port name is not designated, the default downstream scheme utilized will be http.

---

[1]

The “ Kubernetes (K8s) ” feature was requested as part of issue 345 to add support for Kubernetes Service Discovery provider, and released in version 13.5.0

[2]

The “AddKubernetes(Action{KubeClientOptions}) method” was requested as part of issue 2255 (pull request 2257), and released in version 24.0

[3]

The evolution of the “PollKube provider” began with pull request 772 (version 13.2.0). Since then, the provider’s design was reviewed due to reported bug 2304, and patch 2335 was applied and rolled out in version 24.1.

[4]

The “WatchKube provider” was first discussed in thread 2168, later implemented in pull request 2174, and released in version 24.1.

[5]

The “Downstream Scheme vs Port Names” feature was requested as part of issue 1967 and released in version 23.3