Init rfc for piped plugin arch

docs/rfcs/0015-pipecd-plugin-arch-meta.md

@@ -0,0 +1,143 @@
+- Start Date: 2024-10-01
+- Target Version: 1.0
+
+# Summary
+
+This can be consider as the biggest step forward to make PipeCD fit its vision: The one CD for all.
+
+# Motivation
+
+Up to this time, PipeCD archived serveral goals as its defined design in the initial stage, which included "bing a consistent interface which can be used to make progressive delivery process easy for many application platforms". As its goal, the PipeCD currently supports 5 kind of application platforms, which are Kubernetes, ECS, Terraform, Lambda, and Cloud Run. But we want it not just like that, we want to make PipeCD support progressive delivery for whatever application platforms the market use today.
+
+To archive that goal, there is only one way, which is enable the PipeCD accept new platforms as plugin of the piped agent, which can free the limit of what platform PipeCD could support as a CD system.
+
+Also, with the power of plugin architecture, PipeCD can allow multi-versions of the same platform plugin, say the same plugin for CD on Kubernetes, but one is built-in and the other implemented by other developers. That is the future of PipeCD, and the maintainer team wants to bring it to.
+
+# Detailed design
+
+At the time this RFC is writen, there was serveral issues created on PipeCD main repo for tracking the plugin architecture tasks. This RFC summarizes the decisions the maintainer team has agreed to regarding the design of the PipeCD plugin architecture.
+
+After this line in the documentation, pipedv1 is a mention to plugin-arch piped, while pipedv0 is a mention to up-to-now piped.
+
+### The approach
+
+We agreed that pipedv0 will be supported as least until the end of 2025, which mean we have to find a way to ensure our single PipeCD control plane can work with both pipedv0 and pipedv1 at the same time. That leads to this issue at [pipecd/issues/5252](https://github.com/pipe-cd/pipecd/issues/5252).
+
+The key point of the control plane supports both pipedv0 and v1 approach is: platform related concepts like platform provider and kind are remained on the data model (for pipedv0), but we don't adding logic based on those concepts anymore. Pipedv1 logic will be built only around the plugins.
+
+As at this point, we have migration plan for platform related concepts in configuration as below
+
+For platform provider
+
+Instead of Platform Provider, we plan to introduce the config for the plugin and define deployTargets.
+
+```yaml
+apiVersion: pipecd.dev/v1beta1
+kind: Piped
+spec:
+...
+  plugin:
+    - name: k8s_plugin
+      port: 8081
+      deployTargets:
+        - name: dev
+          labels: 
+            env: dev
+          config: # depends on plugins
+            masterURL: http://cluster-dev
+            kubeConfigPath: ./kubeconfig-dev
+```
+
+For kind
+
+Instead of Kind, we plan to introduce the label to represent the application kind.
+
+```yaml
+apiVersion: pipecd.dev/v1beta1
+kind: Application
+metadata:
+  labels:
+    kind: KUBERNETES # <- like this
+spec:
+  name: myApp
+```
+
+The control plane will be updated so that it can accept platform provider from pipedv0 and deployTargets from pipedv1.
+
+### The protocol
+
+The protocol means to protocol used to connect between piped and its plugins.
+As at the starting point, we have 2 options
+- WASM
+- gRPC
+
+For less study maintaining cost, we choose gRPC. Also, other components connections in PipeCD system used gRPC as well.
+
+![](assets/0015-piped-protocol.png)
+
+Boostrap flow:
+- Start piped (as entrypoint of piped pod container for example)
+- Piped loads plugins' configuration
+- Piped pull plugins from registry (if it's remote URI specified)
+- Piped starts plugins with plugins configuration passed at this point
+- Plugins start its grpc servers
+- Plugins ping and send status ready to piped plugin management grpc server
+-> ready to be used at this point
+
+In case of failure:
+- Piped failured: Restart containers / taskset as piped is the entrypoint
+- Plugins failured:
+  - Ping failed -> piped will be notified that plugins are not working
+  - Based on piped logic, will restart the failured plugins (plugin management grpc logic)
+
+### The configuration
+
+The configuration which up-to-now placed in piped and application (kind specific) manifests will be rearrangement as below
+
+![](assets/0015-config-context.png)
+
+We need futher discussion on this but currently, we agreed to separate configuration as piped context and plugin context as above.
+
+### The interfaces
+
+The Piped agent has serveral main features, which are
+
+- Plan and execute deployment
+- Calculate the plan preview result
+- Calculate the drift between live and "git source" manifests
+- Fetch and build the state of the live manifest
+- Insight feature
+
+Across these features, we have to support 5 kinds of platforms (up to now), and each deployment we have serveral kind of stages, which are
+
+- K8s stages
+- Cloud run stages
+- Terraform stages
+- ECS stages
+- Lambda stages
+- Wait stage
+- Wait approval stage
+- Analysis stage
+- Script run stage
+- Custom sync stage (deprecated)
+
+Based on the requirement, we would create plugins based on 3 big interfaces
+
+- Deployment interface: response to how deployment should be done (include: plan, execute deployment tasks)
+- LiveState interface: response to fetch and build state of live manifest
+- Drift interface: response to calculate the drift between 2 given manifests (live - git or running - target)
+
+Plugin can choose how and which interface it implement such as
+
+- K8s plugin can implement all 3 interfaces
+- Wait approval plugin can implement only Deployment interface (since it has nothing to deal with livestate and drift features)
+
+Also, we may need support interface which implemented in piped side (optional) for tasks such as: Decrypt piped secret.
+
+### The running flow
+
+# Unresolved questions
+
+What parts of the design are still TBD?
+
+There are still many areas for improvement, and we welcome your ideas and suggestions.