1 files changed, 148 insertions, 0 deletions
diff --git a/doc/development/agent/gitops.md b/doc/development/agent/gitops.md
new file mode 100644
index 00000000000..8c8586326fa
--- /dev/null
+++ b/doc/development/agent/gitops.md
@@ -0,0 +1,148 @@
+---
+stage: Configure
+group: Configure
+info: To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#designated-technical-writers
+---
+
+# GitOps with the Kubernetes Agent **(PREMIUM ONLY)**
+
+The [GitLab Kubernetes Agent](../../user/clusters/agent/index.md) supports the
+[pull-based version](https://www.gitops.tech/#pull-based-deployments) of
+[GitOps](https://www.gitops.tech/). To be useful, the feature must be able to perform these tasks:
+
+- Connect one or more Kubernetes clusters to a GitLab project or group.
+- Synchronize cluster-wide state from a Git repository.
+- Synchronize namespace-scoped state from a Git repository.
+- Control the following settings:
+
+  - The kinds of objects an agent can manage.
+  - Enabling the namespaced mode of operation for managing objects only in a specific namespace.
+  - Enabling the non-namespaced mode of operation for managing objects in any namespace, and
+    managing non-namespaced objects.
+
+- Synchronize state from one or more Git repositories into a cluster.
+- Configure multiple agents running in different clusters to synchronize state
+  from the same repository.
+
+## GitOps architecture
+
+In this architecture, the Kubernetes cluster (`agentk`) periodically fetches
+configuration from (`kas`), spawning a goroutine for each configured GitOps
+repository. Each goroutine makes a streaming `GetObjectsToSynchronize()` gRPC call.
+`kas` accepts these requests, then checks if this agent is authorized to access
+this GitLab repository. If authorized, `kas` polls Gitaly for repository updates
+and sends the latest manifests to the agent.
+
+Before each poll, `kas` verifies with GitLab that the agent's token is still valid.
+When `agentk` receives an updated manifest, it performs a synchronization using
+[`gitops-engine`](https://github.com/argoproj/gitops-engine).
+
+If a repository is removed from the list, `agentk` stops the `GetObjectsToSynchronize()`
+calls to that repository.
+
+```mermaid
+graph TB
+  agentk -- fetch configuration --> kas
+  agentk -- fetch GitOps manifests --> kas
+
+  subgraph "GitLab"
+  kas[kas]
+  GitLabRoR[GitLab RoR]
+  Gitaly[Gitaly]
+  kas -- poll GitOps repositories --> Gitaly
+  kas -- authZ for agentk --> GitLabRoR
+  kas -- fetch configuration --> Gitaly
+  end
+
+  subgraph "Kubernetes cluster"
+  agentk[agentk]
+  end
+```
+
+## Architecture considered but not implemented
+
+As part of the implementation process, this architecture was considered, but ultimately
+not implemented.
+
+In this architecture, `agentk` periodically fetches configuration from `kas`. For each
+configured GitOps repository, it spawns a goroutine. Each goroutine then spawns a
+copy of [`git-sync`](https://github.com/kubernetes/git-sync). It polls a particular
+repository and invokes a corresponding webhook on `agentk` when it changes. When that
+happens, `agentk` performs a synchronization using
+[`gitops-engine`](https://github.com/argoproj/gitops-engine).
+
+For repositories no longer in the list, `agentk` stops corresponding goroutines
+and `git-sync` copies, also deleting their cloned repositories from disk:
+
+```mermaid
+graph TB
+  agentk -- fetch configuration --> kas
+  git-sync -- poll GitOps repositories --> GitLabRoR
+
+  subgraph "GitLab"
+  kas[kas]
+  GitLabRoR[GitLab RoR]
+  kas -- authZ for agentk --> GitLabRoR
+  kas -- fetch configuration --> Gitaly[Gitaly]
+  end
+
+  subgraph "Kubernetes cluster"
+  agentk[agentk]
+  git-sync[git-sync]
+  agentk -- control --> git-sync
+  git-sync -- notify about changes --> agentk
+  end
+```
+
+## Comparing implemented and non-implemented architectures
+
+Both architectures attempt to answer the same question: how to grant an agent
+access to a non-public repository?
+
+In the **implemented** architecture:
+
+- Favorable: Fewer moving parts, as `git-sync` and `git` are not used, making this
+  design more reliable.
+- Favorable: Uses existing connectivity and authentication mechanisms are used (gRPC + `agentk` token).
+- Favorable: No polling through external infrastructure. Saves traffic and avoids
+  noise in access logs.
+
+In the **unimplemented** architecture:
+
+- Favorable: `agentk` uses `git-sync` to access repositories with standard protocols
+  (either HTTPS, or SSH and Git) with accepted authentication and authorization methods.
+
+  - Unfavorable: The user must put credentials into a `secret`. GitLab doesn't have
+    a mechanism for per-repository tokens for robots.
+  - Unfavorable: Rotating all credentials is more work than rotating a single `agentk` token.
+
+- Unfavorable: A dependency on an external component (`git-sync`) that can be avoided.
+- Unfavorable: More network traffic and connections than the implemented design
+
+### Ideas considered for the unimplemented design
+
+As part of the design process, these ideas were considered, and discarded:
+
+- Running `git-sync` and `gitops-engine` as part of `kas`.
+
+  - Favorable: More code and infrastructure under our control for GitLab.com
+  - Unfavorable: Running an arbitrary number of `git-sync` processes would require
+    an unbounded amount of RAM and disk space.
+  - Unfavorable: Unclear which `kas` replica is responsible for which agent and
+    repository synchronization. If done as part of `agentk`, leader election can be
+    done using [client-go](https://pkg.go.dev/k8s.io/client-go/tools/leaderelection?tab=doc).
+
+- Running `git-sync` and a "`gitops-engine` driver" helper program as a separate
+  Kubernetes `Deployment`.
+
+  - Favorable: Better isolation and higher resiliency. For example, if the node
+    with `agentk` dies, not all synchronization stops.
+  - Favorable: Each deployment has its own memory and disk limits.
+  - Favorable: Per-repository synchronization identity (distinct `ServiceAccount`)
+    can be implemented.
+  - Unfavorable: Time consuming to implement properly:
+
+    - Each `Deployment` needs CRUD (create, update, and delete) permissions.
+    - Users may want to customize a `Deployment`, or add and remove satellite objects
+      like `PodDisruptionBudget`, `HorizontalPodAutoscaler`, and `PodSecurityPolicy`.
+    - Metrics, monitoring, logs for the `Deployment`.