KEP-4743: Kubernetes-etcd interface

keps/sig-etcd/4743-kuberernetes-etcd-interface/README.md

@@ -0,0 +1,353 @@
+
+<!-- toc -->
+- [KEP-4743 The Kubernetes-etcd interface](#kep-4743-the-kubernetes-etcd-interface)
+  - [Summary](#summary)
+    - [Goals](#goals)
+    - [Non-Goals](#non-goals)
+  - [Proposal](#proposal)
+  - [User Stories](#user-stories)
+    - [Kubernetes Backporting an etcd client patch version](#kubernetes-backporting-an-etcd-client-patch-version)
+    - [Making Changes/Patching Bugs in the Interface](#making-changespatching-bugs-in-the-interface)
+    - [Kubernetes Leveraging New etcd Functionality](#kubernetes-leveraging-new-etcd-functionality)
+  - [Code location](#code-location)
+  - [The Interface](#the-interface)
+    - [KV interface](#kv-interface)
+    - [Design considerations](#design-considerations)
+    - [Watch interface](#watch-interface)
+    - [Design considerations](#design-considerations-1)
+  - [Alternatives](#alternatives)
+    - [Code location](#code-location-1)
+      - [Part of the etcd Client Struct](#part-of-the-etcd-client-struct)
+      - [New Package in etcd Repository](#new-package-in-etcd-repository)
+      - [New Repository under etcd-io](#new-repository-under-etcd-io)
+<!-- /toc -->
+# KEP-4743 The Kubernetes-etcd interface
+
+## Summary 
+
+This design proposal introduces an etcd-Kubernetes interface to be added to the 
+etcd client and adopted by Kubernetes. This interface aims to create a clear and
+standardized contract between the two projects, codifying the interactions
+outlined in the [Implicit Kubernetes-ETCD Contract]. By formalizing this contract,
+we will improve the testability of both Kubernetes and etcd, prevent common
+errors in their interaction, and establish a framework for the future evolution 
+of this critical contract.
+
+[Implicit Kubernetes-ETCD Contract]: https://docs.google.com/document/d/1NUZDiJeiIH5vo_FMaTWf0JtrQKCx0kpEaIIuPoj9P6A/edit#heading=h.tlkin1a8b8bl
+
+### Goals
+
+* **Improved Testability:** Enable thorough testing of etcd and Kubernetes 
+  interactions through a well-defined interface, as envisioned in [#15820].
+* **Error Prevention:** Reduce incorrect contract usage, addressing issues like Kubernetes [#110210].
+* **Reviewable Changes:** Make contract modifications easily reviewable and
+  trackable, ensuring a transparent and collaborative evolution.
+
+[#15820]: https://github.com/etcd-io/etcd/issues/15820
+[#110210]: https://github.com/kubernetes/kubernetes/issues/110210
+
+In scope
+* [etcd3 store]: The primary Kubernetes object storage interface.
+* [Master leases]: Lease management for Kubernetes control plane components (utilizing the [etcd3 store])
+
+[etcd3 store]: https://github.com/kubernetes/kubernetes/blob/master/staging/src/k8s.io/apiserver/pkg/storage/etcd3/store.go
+[Master leases]: https://github.com/kubernetes/kubernetes/blob/dae1859c896d742de1ee60a349475f8e28b61995/pkg/controlplane/reconcilers/lease.go#L47-L66
+
+### Non-Goals
+* **Alternative Interface Implementations:** This KEP focuses solely on defining
+  the interface for the existing etcd backend and ensuring its compatibility
+  with Kubernetes. It does not encompass the development or support of
+  alternative storage backends or implementations for the interface.
+* **Non-Storage Usage of etcd Client:**
+  * [Kubeadm] - Primarily used for etcd cluster administration, not Kubernetes object storage.
+  * [Compaction] - Kubernetes aims to encourage native etcd compaction. See [#80513].
+  * [Monitor] - Don’t see benefits of standardizing etcd metrics used for Kubernetes, at least for now.
+  * [Prober], [Feature checker] - These have planned migrations to native etcd features. See [Design Doc: etcd livez and readyz probes] and [KEP-4647].
+  * [Lease manager] - Planned removal in favor of one lease per key to address [#110210].
+
+[Kubeadm]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/cmd/kubeadm/app/util/etcd/etcd.go#L66-L93
+[Compaction]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/staging/src/k8s.io/apiserver/pkg/storage/etcd3/compact.go#L133-L162
+[#80513]: https://github.com/kubernetes/kubernetes/issues/80513
+[Monitor]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/staging/src/k8s.io/apiserver/pkg/storage/storagebackend/factory/etcd3.go#L270-L283
+[Prober]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/staging/src/k8s.io/apiserver/pkg/storage/storagebackend/factory/etcd3.go#L256-L268
+[Feature checker]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/staging/src/k8s.io/apiserver/pkg/storage/feature/feature_support_checker.go#L143-L151
+[Design Doc: etcd livez and readyz probes]: https://docs.google.com/document/d/1SkzmO4RT_GI9YhT0dw4a6nEwKVCciwrwbCDxK0D7ASM/edit?usp=sharing
+[KEP-4647]: https://github.com/kubernetes/enhancements/pull/4662
+[Lease manager]: https://github.com/kubernetes/kubernetes/blob/6ba9fa89fb5889550649bfde847c742a55d3d29c/staging/src/k8s.io/apiserver/pkg/storage/etcd3/lease_manager.go#L90-L120
+[#110210]: https://github.com/kubernetes/kubernetes/issues/110210
+
+## Proposal
+
+This KEP proposes creating an etcd-Kubernetes code interface owned and
+maintained by SIG-etcd. The interface will serve as a formalization of the
+existing etcd-Kubernetes contract, ensuring the correct usage of etcd within
+Kubernetes and enabling improved testing and validation.
+
+The interface will prioritize etcd's existing capabilities and behaviors,
+focusing on compatibility with the current etcd API. It will not introduce
+features or behaviors not supported by etcd, adhering to the existing
+SIG API Machinery policy outlined in [Storage for Extension API Servers].
+This policy designates etcd as the sole supported storage backend for Kubernetes
+for the foreseeable future.
+
+[Storage for Extension API Servers]: https://docs.google.com/document/d/1i0xzRFB-uGLmLYueLMBTpHrOot9ScFxpkkcVcZHVbyA/edit?usp=sharing]
+
+## User Stories
+
+To better understand the importance of code location let’s visit the following use cases:
+
+### Kubernetes Backporting an etcd client patch version
+
+**The Journey:** Kubernetes regularly updates to newer etcd versions to leverage
+bug fixes, or security patches. However, ensuring compatibility between the 
+codified etcd-Kubernetes interface and etcd client is essential.
+
+**Considerations:**
+
+*   Even minor etcd client updates might inadvertently introduce changes that 
+    break the interface's assumptions or functionality.
+*   Kubernetes needs to choose a compatible interface version before upgrading 
+    the etcd client.
+*   Tight coupling to the etcd client could necessitate backporting the 
+    interface to older etcd branches, a complex and time-consuming process.
+
+
+### Making Changes/Patching Bugs in the Interface
+
+**The Journey:** Despite careful design, the complex etcd-Kubernetes contract 
+might reveal bugs or require adjustments.
+
+**Considerations:**
+
+*   Changes and bug fixes need to be implemented and released with minimal
+    disruption to both Kubernetes and etcd users.
+*   Tightly coupled interfaces (Options 1 and 2) might require a full etcd
+    client release for bug fixes, slowing down the process.
+
+### Kubernetes Leveraging New etcd Functionality
+
+**The Journey:** Kubernetes wants to expand its use of the etcd API beyond the
+current interface's scope.
+
+**Considerations:** The interface codifies a minimal subset of the etcd API
+currently used by Kubernetes, and new features will initially be outside its
+scope. Balancing new feature adoption with interface stability is crucial.
+
+**Mitigation:**
+
+*   Allow Kubernetes to directly use new etcd client features during alpha/beta
+    stages, bypassing the interface temporarily.
+*   Extending etcd robustness test to cover the new functionality before
+    formalizing them in interface.
+*   Once a feature is mature and stable, extend the interface, ensuring backward
+    compatibility for existing Kubernetes versions.
+
+## Code location
+
+We propose locating the interface in a `kubernetes` subdirectory under
+https://github.com/etcd-io/etcd/tree/main/client/v3. 
+This approach allows for seamless integration with the etcd client while 
+maintaining a dedicated space for the interface code.
+
+Interfaced will be part of the etcd client package and it's release will be
+combined with etcd release. For immediate Kubernetes use etcd will backport the 
+client to `release-3.5` branch and introduce it in next etcd patch release for 
+Kubernetes to consume.
+
+Alternative code locations are discussed at the end of the document.
+
+## The Interface
+
+To ensure smoother transition we propose the adoption of the etcd-Kubernetes interface to be done in two stages:
+
+1. **KV Interface:** Covering the basic get, list, count, put, and delete operations.
+2. **Watch Interface: **Covering Watch operation and requesting progress notification for it.
+
+
+### KV interface
+
+For the reasoning please see the section below.
+
+```
+
+type Kubernetes interface {
+	Get(ctx context.Context, key string, opts GetOptions) (KubernetesGetResponse, error)
+	List(ctx context.Context, prefix string, opts ListOptions) (KubernetesListResponse, error)
+	Count(ctx context.Context, prefix string) (int64, error)
+	OptimisticPut(ctx context.Context, key string, value []byte, opts PutOptions) (KubernetesPutResponse, error)
+	OptimisticDelete(ctx context.Context, key string, opts DeleteOptions) (KubernetesDeleteResponse, error)
+}
+
+type GetOptions struct {
+	Revision int64
+}
+
+type ListOptions struct {
+	Revision int64
+	Limit    int64
+	Continue string
+}
+
+type WatchOptions struct {
+	StreamKey string
+	Revision  int64
+	Prefix    bool
+}
+
+type PutOptions struct {
+	ExpectedRevision int64
+	GetOnFailure bool
+       // Deprecated: Should be replaced with TTL when Kubernetes starts using one lease per object.
+	LeaseID          LeaseID
+}
+
+type DeleteOptions struct {
+	ExpectedRevision int64
+	GetOnFailure bool
+}
+
+type KubernetesGetResponse struct {
+	KV       *mvccpb.KeyValue
+	Revision int64
+}
+
+type KubernetesListResponse struct {
+	KVs      []*mvccpb.KeyValue
+	Count    int64
+	Revision int64
+}
+
+type KubernetesPutResponse struct {
+	KV        *mvccpb.KeyValue
+	Succeeded bool
+	Revision  int64
+}
+
+type KubernetesDeleteResponse struct {
+	KV        *mvccpb.KeyValue
+	Succeeded bool
+	Revision  int64
+}
+
+```
+
+### Design considerations
+
+**How should arguments be passed?** Proposed: Options struct. 
+
+*   It’s more extensible than a hardcoded list of arguments, allowing adding more fields in future.
+*   It’s more readable than the variadic options list when arguments are optional. 
+    Take a server code to manage [list limit options] as an example.
+*   Same arguments apply for response struct.
+
+[list limit options]: https://github.com/kubernetes/kubernetes/blob/97e87e2c40e5b83399a44738d38653fd59c58e99/staging/src/k8s.io/apiserver/pkg/storage/etcd3/store.go#L640-L645
+
+**Prefer Range vs List semantics?** Proposed: List
+
+*   List matches the intention of the Kubernetes behavior
+
+**Combine Create and Update?** Proposed: Combine them into Put
+
+*   They are the same from an argument standpoint. Create is a Update with ExpectedRevision set to 0.
+*   The difference in on failure can be solved by optional argument `GetOnFailure`
+
+
+### Watch interface
+
+For the reasoning please see the section below.
+```
+
+type Kubernetes interface {
+	Watch(ctx context.Context, key string, opts WatchOptions) KubernetesWatchChan
+	RequestProgress(ctx context.Context, opts RequestProgressOptions) error
+}
+
+type WatchOptions struct {
+	StreamKey string
+	Revision  int64
+	Prefix    bool
+}
+
+type RequestProgressOptions struct {
+	StreamKey string
+}
+
+type KubernetesWatchChan <-chan KubernetesWatchEvent
+
+type KubernetesEventType string
+
+const (
+	Added    KubernetesEventType = "ADDED"
+	Modified KubernetesEventType = "MODIFIED"
+	Deleted  KubernetesEventType = "DELETED"
+	Bookmark KubernetesEventType = "BOOKMARK"
+	Error    KubernetesEventType = "ERROR"
+)
+
+type KubernetesWatchEvent struct {
+	Type KubernetesEventType
+
+	Error error
+	Revision int64
+	Key string
+	Value []byte
+	PreviousValue []byte
+}
+```
+
+### Design considerations
+
+**What control does the user have over requesting progress?** Proposed: Allow user to set streamKey when create watch and requesting progress
+
+*   StreamKey is used to separate watch grpc streams. 
+    For Kubernetes we always use one stream as we don’t change grpc metadata between requests (e.g. WithRequireLeader). 
+    Currently etcd client doesn’t expose streamKey to the user, just calculates it based on grpc metadata taken from context.
+*   Having access to streamKey is useful as progress notifications cannot be requested on a per watch basis, only for the whole stream.
+    This isn’t a big problem in their current setup as Kubernetes opens only one watch per resource. However, this would become a scalability issue for CRDs.
+
+**Should Kubernetes explicitly pass WithRequireLeader or make it default?** Proposed: Make it default if Kubernetes interface is used.
+
+**Should we wrap the watch response?** Proposed: Yes, it allows us to codify the Kubernetes dependency on single revision per transaction and PrevKV dependency.
+
+## Alternatives
+
+### Code location
+
+#### Part of the etcd Client Struct
+
+**Pros:**
+
+*   **Seamless Integration:** The interface becomes inherently part of the client, fostering intuitive usage.
+*   **Code Reuse:** Leverage existing private client methods, reducing redundancy.
+
+**Cons:**
+
+*   **Tight Coupling:** Changes to the interface necessitate updates to the entire etcd client, impacting Kubernetes upgrades.
+*   **Limited Autonomy:** Release and bug-fix cycles are bound to the etcd project's schedule, which may not align with Kubernetes' needs.
+*   **Backporting Challenge:** Requires backporting to v3.5 for Kubernetes compatibility, going against the etcd project's goal of minimizing backports.
+
+#### New Package in etcd Repository
+
+**Pros:**
+
+*   **Versioning Flexibility:** Allows for independent versioning (e.g., `v3.5.13-interface.1`) to track interface changes separately from the etcd client.
+*   **Manageable Integration:** Separates the interface from the client but keeps it within the etcd project, simplifying coordination.
+
+**Cons:**
+
+*   **Backporting Challenge: **Still requires backporting to v3.5 for initial Kubernetes compatibility.
+*   **Maintenance Overhead:** Separate versioning introduces some additional maintenance effort to ensure compatibility between the interface and etcd versions.
+*   **Compatibility Risk:** Incompatibilities may arise between etcd and interface versions if not managed meticulously.
+
+
+#### New Repository under etcd-io
+
+**Pros:**
+
+*   **Maximum Autonomy:** Grants Kubernetes full control over development, releases, and bug fixes.
+
+**Cons:**
+
+*   **Increased Overhead:** Demands significant effort for maintenance, versioning, and compatibility across etcd client versions.
+*   **Dependency Management:** Introduces an additional dependency for Kubernetes, increasing the complexity of version management.
+*   **Potential for Code Duplication:** Implementing the interface might necessitate changes to internal client behavior, potentially requiring some code to be copied.

keps/sig-etcd/4743-kuberernetes-etcd-interface/kep.yaml

@@ -0,0 +1,15 @@
+title: Kubernetes-etcd interface
+kep-number: 4743
+authors:
+  - "@serathius"
+owning-sig: sig-etcd
+participating-sigs:
+  - sig-api-machinery
+status: implementable
+creation-date: 2024-06-28
+reviewers:
+  - "@dims"
+approvers:
+  - "@jpbetz"
+  - "@ahrtr"
+  - "@wojtek-t"