Flux: Server-side reconciliation is coming

Posted on October 7, 2021 by Daniel Holbach

CNCF projects highlighted in this post

Guest post originally published on Flux’ blog by Daniel Holbach

tl;dr: Server-side reconciliation will make Flux more performant, improve overall observability and going forward will allow us to add new capabilities, like being able to preview local changes to manifests without pushing to upstream.

⚠ Changes required: Due to a Kubernetes issue, we require a certain set of Kubernetes releases (starting 1.16.11 – more on this below) as a minimum. The logs, events and alerts that report Kubernetes namespaced object changes are now using the Kind/Namespace/Name format instead of Kind/Name.

We rarely do this, but this time we want to give you some advance notice of a big upcoming feature you will be pleased about. Since Kubernetes moved server-side apply to GA, we are offering you a new reconciler based on it, and graduating the API to v1beta2.

What’s happening

• When does this happen?
  With the release of Flux 0.18, we will move to the new reconciler. It will be released in the coming weeks. Refer to this PR for more information.
• Do I have to use the new thing?
  Yes. Flux will be more performant, less error-prone and from a maintenance perspective will be a lot easier for us. We understand that this new feature will require changes on your end, but we are certain you are going to like the new experience!
• Will my clusters stop working?
  No, but you will need to do a little preparation to make sure Flux can still apply your configurations. See below.
  Note: The pre-flight checks should be able to catch issues like meeting the minimum required Kubernetes version.

Here is what you get

• The new reconciler improves performance (CPU, memory, network, FD usage) and reduces the number of calls to Kubernetes API by replacing kubectl exec calls with a specialized applier written in Go.
• We are able to validate and reconcile sources that contain both CRDs and CRs.
• Detects and reports drift between the desired state (git, s3, etc) and cluster state reliably.
• In the future: Preview of local changes to manifests without pushing to upstream (flux diff -k command TBA).
• Being able to wait for all applied resources to become ready without requiring users to fill in the health checks list.
• Improves the overall observability of the reconciliation process by reporting in real-time the garbage collection and health assessment actions.

This is what you need to do to prepare

Check the Kubernetes version you are running in your cluster. All the versions below fix a regression in the managed fields and field type.

Namespaced objects must contain metadata.namespace, defaulting to the default namespace is no longer supported. This means you will need to chase down any namespaced resources in your configuration files that are left to default, and give them a namespace. Keep in mind that kustomizations are often used to assign a namespace, so even if a particular file doesn’t have a namespace in it, it may not represent a problem.

The logs, events and alerts that report Kubernetes namespaced object changes are now using the Kind/Namespace/Name format instead of Kind/Name e.g.:

Service/flux-demo/podinfo unchanged
Deployment/flux-demo/podinfo configured
HorizontalPodAutoscaler/flux-demo/podinfo deleted

Any automation or monitoring that relies on a particular format in the logs will need to be adapted. Ideally, you should try to handle both the old and new formats.

In terms of API changes, the kustomize.toolkit.fluxcd.io/v1beta2 API is backwards compatible with v1beta1. This is done automatically by the Kubernetes API server, and no preparation is required. You may wish to translate your Flux Kustomization resources, though, according to the following table.

Additions, deprecations and removals:

Why we are doing this

When we started Flux v2, we set a goal to stop relying on third party binaries for core features. While we have successfully replaced the Git CLI shell execs with Go libraries (go-git, git2go) and C libraries (libgit2, libssh2), the kustomize CLI with Go libraries (kustomize/api, kustomize/kyaml), we still depend on the kubectl CLI for the three-way-merge apply feature. With Kubernetes “server-side apply” being promoted to GA, we can finally get rid of kubectl and drive the reconciliation using exclusively the controller-runtime Go client.

Please take a look at the PR introducing this change, as it talks at length about the issues which are solved by this.

Sneak-preview and leaving feedback

If you would like a sneak-preview of the feature before it gets released and try out the v1beta2 API on your own test cluster, please follow the following steps:

1. Install the latest Flux controllersflux install
2. Apply the CRDs from this branchkubectl apply -k https://github.com/fluxcd/kustomize-controller/config/crd?ref=v1beta2
3. Deploy the kustomize-controller build of this branchkubectl -n flux-system set image deployment/kustomize-controller \ manager=ghcr.io/fluxcd/kustomize-controller:v1beta2-8426b396

Please comment on this PR and let us know your thoughts about this.

What’s next?

The biggest parts of the work have been done, here is what is still on our TODO list until the release:

• Move the SSA resource manager to fluxcd/pkg/ssa
• Use the SSA manager in Flux CLI to replace kubectl shell execs for flux bootstrap and flux install
• Use the SSA manager in Flux CLI to implement flux build and flux diff commands
• Update the minimum Kubernetes versions in flux check --pre

This is great – I want to participate in this

Please join us in the #flux channel on CNCF Slack ( get an invite here) to discuss this.

Or find out other ways of connecting (including our weekly meetings) on our Community page.

We are looking forward to having you in our community!