Merge pull request kubernetes#2031 from robscott/endpointslice-subsetting

KEP 2030: EndpointSlice Subsetting and Selection

Author: k8s-ci-robot

keps/sig-network/2030-endpointslice-subsetting/README.md

@@ -0,0 +1,268 @@
+# KEP-2030: EndpointSlice Subsetting and Selection
+
+<!-- toc -->
+- [Release Signoff Checklist](#release-signoff-checklist)
+- [Summary](#summary)
+- [Motivation](#motivation)
+  - [Goals](#goals)
+  - [Non-Goals](#non-goals)
+- [Proposal](#proposal)
+  - [Risks and Mitigations](#risks-and-mitigations)
+- [Design Details](#design-details)
+  - [API Consumption](#api-consumption)
+    - [Select EndpointSlices Without Labels](#select-endpointslices-without-labels)
+    - [Select EndpointSlices With Matching Zone](#select-endpointslices-with-matching-zone)
+    - [Select EndpointSlices With Matching Region](#select-endpointslices-with-matching-region)
+    - [Kube-Proxy](#kube-proxy)
+  - [Controller Implementation](#controller-implementation)
+  - [Backwards Compatibility](#backwards-compatibility)
+- [Test Plan](#test-plan)
+  - [Unit Tests](#unit-tests)
+- [Graduation Criteria](#graduation-criteria)
+  - [Alpha Release](#alpha-release)
+  - [Alpha -&gt; Beta Graduation](#alpha---beta-graduation)
+- [Upgrade / Downgrade Strategy](#upgrade--downgrade-strategy)
+- [Version Skew Strategy](#version-skew-strategy)
+- [Implementation History](#implementation-history)
+- [Drawbacks](#drawbacks)
+- [Alternatives](#alternatives)
+<!-- /toc -->
+
+## Release Signoff Checklist
+
+Items marked with (R) are required *prior to targeting to a milestone / release*.
+
+- [ ] (R) Enhancement issue in release milestone, which links to KEP dir in [kubernetes/enhancements] (not the initial KEP PR)
+- [ ] (R) KEP approvers have approved the KEP status as `implementable`
+- [ ] (R) Design details are appropriately documented
+- [ ] (R) Test plan is in place, giving consideration to SIG Architecture and SIG Testing input
+- [ ] (R) Graduation criteria is in place
+- [ ] (R) Production readiness review completed
+- [ ] Production readiness review approved
+- [ ] "Implementation History" section is up-to-date for milestone
+- [ ] User-facing documentation has been created in [kubernetes/website], for publication to [kubernetes.io]
+- [ ] Supporting documentation—e.g., additional design documents, links to mailing list discussions/SIG meetings, relevant PRs/issues, release notes
+
+[kubernetes.io]: https://kubernetes.io/
+[kubernetes/enhancements]: https://git.k8s.io/enhancements
+[kubernetes/kubernetes]: https://git.k8s.io/kubernetes
+[kubernetes/website]: https://git.k8s.io/website
+
+## Summary
+
+Segment EndpointSlices into logical chunks enabling API consumers like
+kube-proxy to watch only a subset of all EndpointSlices. This provides a natural
+starting point for topology aware routing and dramatically increases
+scalability.
+
+## Motivation
+
+There has been some discussion that the Alpha Service Topology API requires the
+user to do too much. Users generally want the same behavior for all services
+where topology is concerned: avoid cross-zone traffic when in-zone endpoints are
+available and have enough capacity. This problem is exacerbated by the
+introduction of Multi-Cluster services where cross-region services become a
+reality and some regions may be more desirable than others as failover
+locations.
+
+As clusters and services grow, we’re also seeing that the current proxy
+implementations are stretched to their limits as all endpoints for the entire
+cluster are tracked independently by each node.
+
+The best user experience seems like it would be to have the platform - with a
+first class controller and/or a provider specific implementation - intelligently
+prioritize endpoints based on topology, capacity, and any other useful cost
+metrics to aid in traffic shaping, or just to reduce the amount of global
+resource tracking required by every node. However, we are currently missing a
+way to allow endpoints to be targeted to specific subsets of nodes.
+
+### Goals
+
+- Provide the building blocks to allow EndpointSlices to target specific subsets
+  of nodes.
+- EndpointSlices subsetting will be fully backwards compatible for older
+  consumers of the EndpointSlice API.
+- Design is flexible enough for multiple implementations and experimentation.
+- Minimal duplication of data.
+- Room for future enhancements, for example weighted endpoints or Slices.
+
+### Non-Goals
+
+- Define how subsetting should be used.
+- Design the controller responsible for subsetting endpoints.
+- An API for telling controllers how a service should be subsetted.
+
+Many of these are being tackled by the follow up [KEP
+#2004](https://github.com/kubernetes/enhancements/issues/2004).
+
+## Proposal
+
+Two new topology based labels will be introduced for EndpointSlices to support
+subsetting:
+
+```
+endpointslice.kubernetes.io/for-zone
+endpointslice.kubernetes.io/for-region
+```
+
+In the future this pattern may be expanded to include other concepts or
+topologies. A simple pattern like this will allow EndpointSlices to be delivered
+to consumers in a specific zone or region.
+
+
+### Risks and Mitigations
+
+This approach does not allow a single EndpointSlice to target multiple zones or
+regions. Any approach that enabled that would be significantly more complicated.
+The initial proposal in [KEP
+#2004](https://github.com/kubernetes/enhancements/issues/2004) suggests that
+this won't be an issue. 
+
+## Design Details
+
+### API Consumption
+
+EndpointSlice API consumers can use the following selectors to consume
+EndpointSlices:
+
+#### Select EndpointSlices Without Labels
+This is required to support producers of EndpointSlices that don't set these
+labels, including older versions of the EndpointSlice controller.
+
+```
+matchExpressions:
+  - {key: endpointslice.kubernetes.io/for-zone, operator: DoesNotExist}
+  - {key: endpointslice.kubernetes.io/for-region, operator: DoesNotExist}
+```
+
+#### Select EndpointSlices With Matching Zone
+```
+matchLabels:
+  endpointslice.kubernetes.io/for-zone: example-zone
+```
+
+#### Select EndpointSlices With Matching Region
+```
+matchLabels:
+  endpointslice.kubernetes.io/for-region: example-region
+```
+
+#### Kube-Proxy
+When the `EndpointSliceSubsetting` feature gate is set to true on `Kube-Proxy`,
+it will use these selectors to filter EndpointSlices.
+
+### Controller Implementation
+Although a controller implementation is out of scope for this KEP, it is worth
+discussing what that might look like. For reference, [KEP
+#2004](https://github.com/kubernetes/enhancements/issues/2004) discusses how this
+could be implemented for the EndpointSlice controller. That proposal involves 3
+potential approaches - Original, PreferZone, and RequireZone.
+
+None of the proposed approaches would involve data duplication. Each
+Pod/endpoint would continue to live in a single EndpointSlice. The reason they
+might end up with more EndpointSlices would be less efficient packing. Here's
+the number of EndpointSlices that would result based on the number of endpoints
+a Service has in a 3 zone cluster. In each case, the numbers in parentheses
+represent how many endpoints would exist in each slice.
+
+| # endpoints | Original # slices | PreferZone # slices | RequireZone # slices |
+|-|-|-|-|
+| 6 | 1 (6) | 1 (6) | 3 (2) |
+| 90 | 1 (90) | 3 (30) | 3 (30) |
+| 270 | 3 (90) | 3 (90) | 3 (90) |
+
+The RequireZone approach requires at least one EndpointSlice per zone per
+Service. The PreferZone also has the same requirement unless the minimum
+threshold has been met. Before that threshold, a single shared EndpointSlice (no
+additional labels) is used. There's some padding involved here to make sure
+we're not flapping back and forth between these states.
+
+With this approach, EndpointSlices can be delivered everywhere (no additional
+labels), or to a zone (for-zone), or to a region (for-region). None of the
+proposed approaches involve a single Service having separate sets of
+EndpointSlices for each use case. As defined by [KEP
+1659](https://github.com/kubernetes/enhancements/tree/master/keps/sig-architecture/1659-standard-topology-labels),
+"region" and "zone" are strictly hierarchical ("zones" are subsets of "regions")
+and zone names are unique across regions.
+
+### Backwards Compatibility
+We don't need create EndpointSlices without labels for backwards compatibility,
+we just need to ensure that consumers always support consuming EndpointSlices
+without these labels. Even if we updated the EndpointSlice controller to
+consistently label these EndpointSlices with `for-zone` or `for-region`, we
+couldn't guarantee that other producers would.
+
+There's nothing in any current consumer implementation that would break if
+additional labels like `for-zone` or `for-region` were added to EndpointSlices.
+All consumers will need to care about for this or the original approach is the
+`kubernetes.io/service-name` label. If they want to support subsetting, they can
+update their selectors as described in this KEP, but subsetting won't actually
+break any existing functionality.
+
+## Test Plan
+This KEP is quite small in scope. The only new functionality being added will be
+an adjustment to the EndpointSlices kube-proxy consumes when a feature gate is
+enabled. We will need to add more test coverage for when this feature is enabled
+or disabled.
+
+### Unit Tests
+* Ensure kube-proxy will continue to consume all EndpointSlices when this
+  feature is disabled.
+* Ensure EndpointSlices delivered to a specific zone will be consumed by
+  kube-proxy running in the same zone when this feature is enabled.
+* Ensure EndpointSlices delivered to a specific zone will not be consumed by
+  kube-proxy running in a different zone when this feature is enabled.
+* Ensure EndpointSlices delivered to a specific region will be consumed by
+  kube-proxy running in the same region when this feature is enabled.
+* Ensure EndpointSlices delivered to a specific region will not be consumed by
+  kube-proxy running in a different region when this feature is enabled.
+
+## Graduation Criteria
+
+### Alpha Release
+
+- Proposed labels are added as well known labels in Discovery API types.
+- Implement new selectors in kube-proxy.
+- Implement test plan.
+
+### Alpha -> Beta Graduation
+
+- EndpointSlice controller supports publishing EndpointSlices in subsets. (See
+  [KEP 2004](https://github.com/kubernetes/enhancements/issues/2004) for more
+  info).
+
+## Upgrade / Downgrade Strategy
+
+This functionality will be guarded by the `EndpointSliceSubsetting` feature gate
+on kube-proxy. This will be fully backwards compatible and will only make a
+difference in a cluster if EndpointSlices are being published with the labels
+described in this KEP.
+
+## Version Skew Strategy
+
+This is designed with backwards compatibility in mind. Enabling this feature is
+not reliant on any other feature being enabled in any other release. [KEP
+#2004](https://github.com/kubernetes/enhancements/issues/2004) will be dependent
+on this KEP though.
+
+## Implementation History
+
+September 2020: Initial Proposal Submitted
+
+## Drawbacks
+
+Although an optional feature, this adds more complexity to the consumption of
+EndpointSlices for anyone that wants to support the feature.
+
+## Alternatives
+
+An alternative would be to use an approach that would allow delivery to multiple
+zones. With labels, this would require including the zone name in the label key:
+
+```
+endpointslice.kubernetes.io/for-zone-a
+endpointslice.kubernetes.io/for-zone-b
+```
+
+Unfortunately it would be much more difficult to build backwards compatible
+selectors to consume these labels.

keps/sig-network/2030-endpointslice-subsetting/kep.yaml

@@ -0,0 +1,41 @@
+title: EndpointSlice Subsetting and Selection
+kep-number: 2030
+authors:
+  - "@robscott"
+  - "@jeremyot"
+owning-sig: sig-network
+participating-sigs:
+  - sig-multicluster
+  - sig-scalability
+status: implementable
+creation-date: 2020-09-29
+reviewers:
+  - "@andrewsykim"
+  - "@bowei"
+  - "@danwinship"
+  - "@thockin"
+  - "@wojtek-t"
+approvers:
+  - "@thockin"
+
+# The target maturity stage in the current dev cycle for this KEP.
+stage: alpha
+
+# The most recent milestone for which work toward delivery of this KEP has been
+# done. This can be the current (upcoming) milestone, if it is being actively
+# worked on.
+latest-milestone: "v1.20"
+
+# The milestone at which this feature was, or is targeted to be, at each stage.
+milestone:
+  alpha: "v1.20"
+  beta: "v1.21"
+  stable: "v1.23"
+
+# The following PRR answers are required at alpha release
+# List the feature gate name and the components for which it must be enabled
+feature-gates:
+  - name: EndpointSliceSubsetting
+    components:
+      - kube-proxy
+disable-supported: true