Skip to content

Workbench Technical Overview

Objectives

  • Gain an overall understanding of the Workbench architecture
  • Explain the included components and integration methods
  • Examples and guide documentation

devops

System Components

The amamba-system mainly includes the following three components:

  • apiserver: The entry point for all services, stateless, providing access through gRPC and REST. It calls Providers services via REST or gRPC or creates resources through kube-client or kpanda to interact with Controllers.
  • devops-server: Mainly communicates with Jenkins and provides relevant interfaces to apiserver to support lazy loading, event-based synchronization, and other functions.
  • syncer: Primarily responsible for maintaining data consistency between Providers, Controllers, and the system, ensuring that the system configuration is correctly distributed to each component.

Architecture Diagram

CI

Jenkins

Amamba Jenkins is enhanced based on the Kubernetes plugin. Compared to native Jenkins, Workbench offers the following advantages:

  1. Complete Cloud-Native Solution

    Compared to traditional VM deployments, cloud-native offers elasticity, scalability, high availability, and flexibility. We provide a full set of Jenkins cloud-native solutions, including running in Kubernetes, building images and artifacts, publishing, integrating with other toolchains, and DR.

  2. Optimized Read Speed

    We provide a relational database-based cache where pipeline and run-related data are stored in the DB, ensuring timely updates through an event mechanism.

  3. Deploy Once, Run Anywhere

    Thanks to the DB, the deployment location and persistence of Jenkins are no longer important. Workbench can always connect to another Jenkins instance deployed in a different cluster and run previously created pipelines on it.

  4. Lower Barrier to Entry

    Jenkinsfile is powerful but has a relatively high understanding and usage barrier. Workbench provide a graphical interface to help users create and edit pipelines through drag-and-drop, allowing users to get started quickly.

  5. Multi-Tenancy

    Based on the Workspace mechanism of DCE5, we have isolated pipelines and credentials.

Additionally, the Workbench currently has the following issues:

  1. Slow Startup

    Compared to using registered node agents traditionally, running pipelines on Kubernetes-based Jenkins requires creating Pods, waiting for scheduling, and pulling images, resulting in slower startup speeds.

  2. Cache

    Due to the sandbox environment of Pods, the cache cannot be effectively utilized. Workbench can mount persistent storage for it, but this almost always means additional planning and management costs.

  3. Plugins

    Installing plugins still requires operations and configuration on the Jenkins management interface, and they will be lost when migrating to other clusters. We recommend forking amamba-io/jenkins-agent and modifying formula.yaml to rebuild and persist plugins into the image.

  4. Noisy Neighbour

    We have not implemented resource quotas and limits for pipelines because we believe CI and deployment resources should not be considered together. Instead, it relies on the administrator's pre-planning of the CI system, which may lead to some pipelines monopolizing all resources.

  5. Slow Image Build

    Besides the cache issue mentioned earlier, when the Kubernetes runtime is not Dockerd, you can only use methods like dind or pind to build images. Running containers inside containers can lead to some IO performance loss.

Based on this, Workbench is also planning to support Jenkins deployments using traditional methods.

Overall Architecture

Architecture

  • By utilizing the event mechanism provided by the Generic Event plugin, you can update the pipeline's running status timely. To avoid event loss caused by different network environments, we deploy event-proxy as a sidecar, which provides stronger network traffic control and a reliable event system. For Amamba, it is completely transparent and can be omitted if the network environment is controllable.
  • To ensure apiserver is completely stateless and scalable, we use devops-server to handle state-related services, including receiving Jenkins events, synchronizing pipeline status, lazy loading of runtime data, exposing Jenkins event metrics, and providing an SDK for apiserver to call.
  • Since Jenkins instances may run in any cluster, syncer helps us synchronize part of the system to different instances, including:
    • Our credentials are managed uniformly using secret, and syncer will synchronize the pipeline part to Jenkins.
    • Jenkins configurations are effective based on CasC. We retain the changed parts in the system and use syncer to synchronize them to the sub-clusters.
  • Pipelines are actually executed in Pods, communicating with Jenkins through the jnlp container. This architecture allows most tasks to continue running even if Jenkins goes down.

More Resources

CD

Argo CD

Amamba uses Argo CD as the engine to implement GitOps capabilities. Compared to native Argo CD, we have mainly enhanced integration with DCE 5.0:

  1. Applications are isolated by tenant and can only be deployed to the corresponding clusters and namespaces.
  2. Permissions are controlled according to the global RBAC strategy, and only users with proper rules can perform corresponding operations.

argo-cd

Issues:

  1. Code repositories are not isolated by tenant: This is a design flaw of Argo CD, and we are actively pushing the community to optimize it: argoproj/argo-cd #18290
  2. ApplicationSet is not yet supported.

Overall Architecture

  • We access argo-server through the SDK provided by Argo, using AppProject - Workspace one-to-one mapping to create and update corresponding resources for users.
  • syncer constantly watches the changes of tenants in the system and applies the changes to Argo's resources.

More Resources

Argo Rollouts

Amamba provides progressive release capabilities based on Argo Rollouts. Compared to native Argo Rollouts, we have mainly enhanced the following aspects:

  1. Easier to use, providing a Step By Step interface to convert workloads in the current cluster into Rollout and start progressive release.
  2. Support for cross-cluster creation and management.
  3. Better integration with Mspider (when using istio as a traffic control tool) and skoala (when using contour).

Issues:

  1. In Mspider's managed mesh mode, the VirtualService and DestinationRule deployed in sub-clusters will not take effect, making it impossible to use Istio-based Rollout in this scenario. Mspider is optimizing it and is expected to support it in version v0.26.
  2. Creating Rollout based on Deployment will cause temporary access failures. The community is working to fix this issue argoproj/argo-rollouts #3111, which is expected to be resolved in version v1.7.

More Resources

Applications

We support native applications, Helm apps, OAM applications, and OLM Apps.

Native Applications

Native applications refer to resources of Kubernetes native workload types, such as Deployment, StatefulSet, and DaemonSet. We have not added any cognitive load, but to associate related resources, we use kubernetes-sigs/application to indicate the resource types and labels included in the current application.

The controller corresponding to this CRD is optional and needs to be manually installed in each sub-cluster. Not deploying it will not cause resource creation failures, but deploying it allows continuous synchronization of resource status, returning the readiness of resources.

OAM Applications

OAM applications are implemented based on kubevela. We have mainly enhanced multi-tenancy:

  1. When creating an OAM application for the first time under the current tenant, it will prompt to create or specify a namespace as the location for all applications.core.oam.dev under that Workspace, achieving application isolation under different tenants.
  2. syncer constantly watches the changes of tenants in the system and applies the changes to KubeVela's resources.

kubevela

Helm and OLM Apps

OLM and Helm apps are mainly encapsulated based on proper capabilities of Kpanda.

Versions

All Addons of Amamba can be deployed in the app store using Helm from v0.21 (for installer v0.12) onwards, with no version restrictions. However, the default version has undergone complete testing and can be better integrated into the current system. The following table shows the version relationship between Addons and Amamba:

Addon Start Version End Version
amamba-io/Jenkins 0.3.1 (2.413) v0.21 v0.24
argocd 5.34.6 (2.7.3) v0.21 -
argo-rollouts 2.32.0 (1.6.0) v0.21 -
vela-core 1.7.7 (1.7.7) v0.21 v0.23
vela-core 1.9.7 (1.9.7) v0.23 -
amamba-io/Jenkins 0.3.2 (2.413) v0.24 -
amamba/kube-app-manager 0.1.4 (0.8.3) ? -

Comments