|
| 1 | +# KEP-1959: Service Type=LoadBalancer Class Field |
| 2 | + |
| 3 | +<!-- toc --> |
| 4 | +- [Release Signoff Checklist](#release-signoff-checklist) |
| 5 | +- [Summary](#summary) |
| 6 | +- [Motivation](#motivation) |
| 7 | + - [Goals](#goals) |
| 8 | + - [Non-Goals](#non-goals) |
| 9 | +- [Proposal](#proposal) |
| 10 | + - [User Stories (Optional)](#user-stories-optional) |
| 11 | + - [Story 1](#story-1) |
| 12 | + - [Story 2](#story-2) |
| 13 | + - [Risks and Mitigations](#risks-and-mitigations) |
| 14 | +- [Design Details](#design-details) |
| 15 | + - [Test Plan](#test-plan) |
| 16 | + - [Graduation Criteria](#graduation-criteria) |
| 17 | + - [Upgrade / Downgrade Strategy](#upgrade--downgrade-strategy) |
| 18 | + - [Version Skew Strategy](#version-skew-strategy) |
| 19 | +- [Implementation History](#implementation-history) |
| 20 | +- [Drawbacks](#drawbacks) |
| 21 | +- [Alternatives](#alternatives) |
| 22 | + - [ServiceClass Resource](#serviceclass-resource) |
| 23 | + - [Generic Annotation](#generic-annotation) |
| 24 | + - [Provider-Specific Annotations](#provider-specific-annotations) |
| 25 | +- [Infrastructure Needed (Optional)](#infrastructure-needed-optional) |
| 26 | +<!-- /toc --> |
| 27 | + |
| 28 | +## Release Signoff Checklist |
| 29 | + |
| 30 | +Items marked with (R) are required *prior to targeting to a milestone / release*. |
| 31 | + |
| 32 | +- [ ] (R) Enhancement issue in release milestone, which links to KEP dir in [kubernetes/enhancements] (not the initial KEP PR) |
| 33 | +- [ ] (R) KEP approvers have approved the KEP status as `implementable` |
| 34 | +- [ ] (R) Design details are appropriately documented |
| 35 | +- [ ] (R) Test plan is in place, giving consideration to SIG Architecture and SIG Testing input |
| 36 | +- [ ] (R) Graduation criteria is in place |
| 37 | +- [ ] (R) Production readiness review completed |
| 38 | +- [ ] Production readiness review approved |
| 39 | +- [ ] "Implementation History" section is up-to-date for milestone |
| 40 | +- [ ] User-facing documentation has been created in [kubernetes/website], for publication to [kubernetes.io] |
| 41 | +- [ ] Supporting documentation—e.g., additional design documents, links to mailing list discussions/SIG meetings, relevant PRs/issues, release notes |
| 42 | + |
| 43 | +[kubernetes.io]: https://kubernetes.io/ |
| 44 | +[kubernetes/enhancements]: https://git.k8s.io/enhancements |
| 45 | +[kubernetes/kubernetes]: https://git.k8s.io/kubernetes |
| 46 | +[kubernetes/website]: https://git.k8s.io/website |
| 47 | + |
| 48 | +## Summary |
| 49 | + |
| 50 | +When Service Type=LoadBalancer is enabled by a Kubernetes cloud provider, it is a global |
| 51 | +configuration that applies for all Service Type=LoadBalancer resources in a given cluster. |
| 52 | +This becomes problematic if users want to leverage multiple Service Type=LoadBalancer |
| 53 | +implementations in a cluster. |
| 54 | + |
| 55 | +The new [Services APIs](https://github.com/kubernetes-sigs/service-apis) addresses this already |
| 56 | +with the GatewayClass resource. However, until Gateway/GatewayClass APIs become mature, we should |
| 57 | +support similar functionality for Services of Type=LoadBalancer. Introducing a new resource like |
| 58 | +`ServiceClass` is probably not worthwhile given that there are new APIs already in development. |
| 59 | +This KEP proposes a light-weight approach for Service Type=LoadBalancer by introducing a Service |
| 60 | +field `service.spec.loadBalancerClass`. |
| 61 | + |
| 62 | +## Motivation |
| 63 | + |
| 64 | +The main use-case for this feature is being able to support multiple Service Type=LoadBalancer |
| 65 | +implementations in a cluster, as different workloads may want to leverage different loadbalancer |
| 66 | +providers based on efficiency, availability, cost and other factors. |
| 67 | + |
| 68 | +For example, a cluster admin may want to use a public load balancer from a cloud provider |
| 69 | +for workloads that must be assigned a publically routable address, but they may want to |
| 70 | +enable a lower-cost solution for workloads that are only internally accessible. |
| 71 | + |
| 72 | +### Goals |
| 73 | + |
| 74 | +* allow users to opt-out of the Service Type=LoadBalancer implementation by the cloud provider. |
| 75 | +* allow multiple implementations of Service Type=LoadBalancer in a cluster. |
| 76 | +* prevent every cloud provider from implementing a custom "opt-out" annotation for their load balancer. |
| 77 | + |
| 78 | +### Non-Goals |
| 79 | + |
| 80 | +* performance improvements for Service Type=LoadBalancer. |
| 81 | +* changing any other existing behaviors for Service Type=LoadBalancer aside from being able |
| 82 | +to disabling it from the cloud provider. |
| 83 | + |
| 84 | +## Proposal |
| 85 | + |
| 86 | +This KEP proposes to add a new field `spec.loadBalancerClass` in Service which allows for |
| 87 | +multiple implementations of Service Type=LoadBalancer in a cluster. |
| 88 | + |
| 89 | +### User Stories (Optional) |
| 90 | + |
| 91 | +#### Story 1 |
| 92 | + |
| 93 | +As a cluster admin: |
| 94 | +* I want to use my cloud provider's public load balancer service for applications that require |
| 95 | +public ingress. |
| 96 | +* I want to use my own load balancing solution for any applications that only talk internally |
| 97 | +within my own network because I want to save costs. |
| 98 | + |
| 99 | +#### Story 2 |
| 100 | + |
| 101 | +As an application developer: |
| 102 | +* I MUST use a hardware-based loadbalancer for certain applications due to specific protocols |
| 103 | +only available there. |
| 104 | +* I want to use the cloud provider's default load balancer for any applications that do not |
| 105 | +rely on protocols from hardware load balancers. |
| 106 | + |
| 107 | +### Risks and Mitigations |
| 108 | + |
| 109 | +Many cloud providers today support an "opt-out" annotation for this behavior. The annotation is specific |
| 110 | +to the cloud provider. Introduction of the `loadBalancerClass` field at this point would mean that |
| 111 | +cloud providers need to start accounting for both existing annotations and the new field. |
| 112 | + |
| 113 | +## Design Details |
| 114 | + |
| 115 | +Introduce a new field to Service `spec.loadBalancerClass`. |
| 116 | + |
| 117 | +If the field `spec.loadBalancerClass` is not set, the existing cloud provider will assume |
| 118 | +ownership of the Service Type=LoadBalancer resource. This is required to not break existing clusters |
| 119 | +that assume Service Type=LoadBalancer is always managed by the cloud provider. |
| 120 | + |
| 121 | +Required API changes: |
| 122 | +```go |
| 123 | +// ServiceSpec describes the attributes that a user creates on a service. |
| 124 | +type ServiceSpec struct { |
| 125 | + ... |
| 126 | + ... |
| 127 | + |
| 128 | + // loadBalancerClass is the name of the load balancer implementation this Service belongs to. |
| 129 | + // This field can only be set when the Service type is 'LoadBalancer'. If not set, the default load |
| 130 | + // balancer implementation is used, today this is typically done through the cloud provider integration, |
| 131 | + // but should apply for any default implementation. If set, it is assumed that a load balancer |
| 132 | + // implementation is watching for Services with a matching class name. Any default load balancer |
| 133 | + // implementation (e.g. cloud providers) should ignore Services that set this field. |
| 134 | + // +optional |
| 135 | + LoadBalancerClassName string `json:"loadBalancerClass,omitempty"` |
| 136 | +} |
| 137 | +``` |
| 138 | + |
| 139 | +* `loadBalancerClass` will be immutable only when the Service type is `LoadBalancer`, this way existing and future implementations |
| 140 | +do not have to worry about handling Services that change the class name. The class name is mutable and can be cleared when the |
| 141 | +type changes. |
| 142 | +* `loadBalancerClass` will be validated against label-style format. |
| 143 | +* the `loadBalancerClass` field will be feature gated. The field will be dropped during API strategy unless |
| 144 | +the feature gate is enabled. |
| 145 | + |
| 146 | +Required updates to service controller: |
| 147 | +* if the class field is NOT set for a Service, allow the cloud provider to reconcile the load balancer. |
| 148 | +* if the class annotation IS set for a Service, skip reconciliation of the Service from the cloud provider. |
| 149 | + |
| 150 | +### Test Plan |
| 151 | + |
| 152 | +Unit tests: |
| 153 | +* test that service controller does not call the cloud provider if the class field is set. |
| 154 | +* test API strategy to ensure the `loadBalancerClass` field is dropped unless the feature gate is enabled |
| 155 | +or an existing Service has the field set. |
| 156 | +* test API validation for immutability. |
| 157 | + |
| 158 | +Integration tests: |
| 159 | +* test that the class field is propoerly cleared/validated when the Service type changes to and from `LoadBalancer`. |
| 160 | + |
| 161 | +E2E tests: |
| 162 | +* test that creating a Service with an unknown class name results in no load balancer being created for a Service. |
| 163 | + |
| 164 | +### Graduation Criteria |
| 165 | + |
| 166 | +Alpha: |
| 167 | +* the `loadBalancerClass` field is added to Service with an alpha feature gate. |
| 168 | +* when enabled, service controller will ignore Service LBs with a non-empty class name. |
| 169 | +* unit tests for service controller. |
| 170 | +* unit tests for API strategy (drop disabled fields). |
| 171 | + |
| 172 | +### Upgrade / Downgrade Strategy |
| 173 | + |
| 174 | +* Usage of `loadBalancerClass` will be off by default during the alpha stage but can handle existing Services that |
| 175 | +has the field set already. This ensures apiserver can handle the new field on downgrade. |
| 176 | +* On upgrade, if the feature gate is enabled, there should be no changes since the default behavior has not changed |
| 177 | +(service controller calls the cloud provider to reconcile load balancers). |
| 178 | + |
| 179 | +### Version Skew Strategy |
| 180 | + |
| 181 | +Since this feature will be alpha for at least 1 release, an n-1 kube-controller-manager or cloud-controller-manager should |
| 182 | +handle enablement of this feature if a new apiserver enabled it. |
| 183 | + |
| 184 | +## Implementation History |
| 185 | + |
| 186 | +- the `Summary`, `Motivation`, `Proposal` and `Design Details` sections was merged, signaling SIG acceptance |
| 187 | + |
| 188 | +## Drawbacks |
| 189 | + |
| 190 | +* Added complexity to Service. |
| 191 | +* In **most** clusters, a single Service Type=LoadBalancer implementation from the cloud provider is sufficient. |
| 192 | + |
| 193 | +## Alternatives |
| 194 | + |
| 195 | +### ServiceClass Resource |
| 196 | + |
| 197 | +Instead of a field specifying the name of the implemmentation, the class name can reference the name of a class resource |
| 198 | +similar to GatewayClass and IngressClass. This would enable more expressive configuration per load balancer implementation. |
| 199 | + |
| 200 | +### Generic Annotation |
| 201 | + |
| 202 | +A generic annotation can be used to store the class name. This is avoided since there would be no way to introduce |
| 203 | +the annotation in a safe way and we can't enforce immutability for annotations. |
| 204 | + |
| 205 | +### Provider-Specific Annotations |
| 206 | + |
| 207 | +Instead of a generic Kubernetes annotation read by service controller, each cloud provider could implement |
| 208 | +their own "skip this Service"-like logic with custom annotations. Given that many cloud providers have been |
| 209 | +asking for this feature, a generic field used across all providers may be more beneficial. |
| 210 | + |
| 211 | +## Infrastructure Needed (Optional) |
| 212 | + |
| 213 | +<!-- |
| 214 | +Use this section if you need things from the project/SIG. Examples include a |
| 215 | +new subproject, repos requested, or GitHub details. Listing these here allows a |
| 216 | +SIG to get the process for these resources started right away. |
| 217 | +--> |
0 commit comments