Flux from End-to-End

A narrative of the life of a commit as it relates to Flux components.

Below we describe the flow of data through Flux, from End to End.

We assume a standard Flux installation, with all optional features enabled, then explain how Flux users can expect their change to flow through the stages of its life as a commit, and describe how the commit data passes through the Flux system and the cluster, in rough chronological order. We tried to cover every supported opportunity that users have to inspect and interact with their changes through Flux, with a goal of showing the cohesive behavior of each component of the GitOps toolkit together in a single document.

Users can start their journey here: Getting Started for the beginner’s introduction to Flux. Then follow each successive guide in order to arrive at a fully configured Flux installation with all features enabled. Most users will not need or use all of the features, but we present an overview of every required or optional part of Flux here.

This document is not a comprehensive guide in and of itself, as that would be far too long to consume in one sitting. It should be clear from reading this document though, when and how Flux components interact with the cluster resources or APIs, or any commit or registry data or other external resources outside of Flux and the cluster. This document narrates through those interactions so that there can be an end-to-end analysis of how Flux works that includes mention of any authentication or security hardening procedures that are in play for Flux at runtime. One specific goal of this document is to provide an anchor and a starting point for security auditors.

Security and hardening procedures that the Flux development team may be taking to guarantee Flux release-engineering standards for testing, runtime safety, and release quality are considered outside of the scope of this document. The operational details of each feature should already have their own documentation supporting them, and we can link out to those documents as references wherever possible to make this document evergreen and easy to maintain; so we should not try to reflect every detail of every feature here.

See one of: Security, Contributing: Acceptance Policy for more information about the standards and practices in general around those topics. An exhaustive description of the precautions with regard to sensitive and/or secret data and flow of information related to sensitive access, is out of the scope of this document.


Flux uses the following terms throughout the codebase and documentation:

  • Cluster - Any number of Kubernetes nodes, joined into a group to run containerized applications.
  • Commit - A snapshot of a Git repository’s state (or any Version Control System) at any given time.
  • Client - Any application or resource manager which implements the “customer” side of any API or Service.
  • Resource - In Kubernetes, a YAML data structure represents cluster objects that drive workloads like: Deployment, Pod, Service, Ingress, StatefulSet, Job, and many others.
  • Custom Resource - Kubernetes provides a Custom Resource Definition (CRD) type for defining Custom Resources to be implemented by a controller. In Flux, examples include: GitRepository, Bucket, HelmRepository, Kustomization, HelmRelease, Alert, and others.
  • Field - YAML resources are collections of data fields, which can be nested to create complex structures like with Array and Map.
  • Event - A YAML resource emits events while undergoing state transitions, which themselves (Event) are also resources.
  • API - In Kubernetes, an API consists of (usually) a CRD, a control loop, and optionally one or more admission or mutation hooks. Flux’s APIs are also known, collectively, as the GitOps Toolkit.
  • Agent - A process which runs in the cluster and does some work on behalf of users. Flux’s controllers are “software agents” that implement a control loop.
  • Service - When Kubernetes Deployments spawn workloads, they are placed in ephemeral Pods which usually are not directly addressable. Services are used to connect these Endpoints to a stable address that can usually be discovered through a DNS lookup in the cluster.

Microservice architecture

Flux is composed of four separable core components or controllers: Source Controller, Kustomize Controller, Helm Controller, and Notification Controller, with two extra components: Image Automation Controller and Image Reflector Controller. These controllers or Agents run on the Cluster, and they define APIs which are based on Custom Resources that altogether implement the GitOps Toolkit.

Source Controller is the Agent responsible for pulling Commit data into the Cluster. Commits are made available as a read-only Service to Clients, which can connect with the Source Controller and fetch Artifacts, .tar.gz files containing Kubernetes Resource manifest data.

Besides artifact acquisition from external sources, the responsibilities of the Source Controller also include: verification of source authenticity through cryptographic signatures, detection of source changes based on semantic version policies, outbound notification of Cluster subscribers when updates are available, and also reacting to inbound notifications that represent Git push and Helm chart upload events.

Kustomize Controller is the Agent responsible for reconciling the cluster state with the desired state as defined by Commit manifests retrieved through Source controller. Kustomize controller delivers, or applies, resources into a cluster. The Kustomization is the Custom Resource or API through which a Flux user defines how Kustomize controller delivers workloads from sources.

The Kustomize controller is responsible for validating manifests against the Kubernetes API, and managing access to permissions in a way that is safe for multi-tenant clusters through Kubernetes Service Account impersonation. The controller supports health checking of deployed resources and dependency ordering, optionally enabled garbage collection or “pruning” of deleted resources from the cluster when they are removed from the source, and also notification of when cluster state changes – Kustomizations can also target and deliver resources onto a remote cluster, (which can, but does not necessarily also run its own local independent set of Flux controllers.)

The Kustomization API is designed from the beginning with support for multi-tenancy as a primary concern.

Helm Controller is the Agent responsible for managing Helm artifacts (with some parts of the work shared in the Source Controller). The Source Controller acquires Helm charts from Helm repositories or other sources. The desired state of a Helm release is described through a Custom Resource named HelmRelease. Based on the creation, mutation or removal of a HelmRelease resource in the cluster, Helm actions are performed by the controller.

Helm Controller (and its predecessor, Helm Operator) stand alone in the GitOps world as Go client implementations of the Helm package library. While there are many projects in the GitOps space that can perform “Helm Chart Inflation” which can also be explained through the activities of merging values from different sources, rendering templates, applying the changes to the cluster, and then waiting to become healthy, other projects usually cannot claim strict compatibility with all Helm features. Helm Controller boasts full compatibility and reliably identical behavior in Flux with all released Helm features.

Examples of some Helm Controller features that directly leverage upstream features of Helm today include Helm Chart Hooks, Helm Release Lifecycle events and the optional health checking that’s performed by helm --wait to determine if a release is successful, Helm tests, rollbacks and uninstalls, and an implementation of Helm’s Post Rendering feature that allows safety and security while using the Kustomize post-renderer in Flux Continuous Delivery pipelines (that is, without requiring untrusted execution of any external scripts or binaries).

Notification Controller

The Notification Controller is a Kubernetes operator, specialized in handling inbound and outbound events. The controller handles:

  • events coming from external systems (GitHub, GitLab, Bitbucket, Harbor, Jenkins, etc) and notifies the GitOps toolkit controllers about source changes.
  • events emitted by the GitOps toolkit controllers (source, kustomize, helm) and dispatches them to external systems (Slack, Microsoft Teams, Discord, RocketChat) based on event severity and involved objects.

Image Reflector Controller

The Image Reflector Controller scans image repositories and reflects the image metadata in Kubernetes resources. The reflector reconciles ImageRepository resources, which can be used with ImagePolicy rules to select a “latest” image. This can be used to drive automation, as with Image Automation Controller; or more generally, for processes which need to know about the state of tags in image repositories.

With ImageRepository, the users can specify how to scan an OCI image repository, such as repository authentication, as well as client and server TLS certificates to connect to the repository host.

Image Automation Controller

The Image Automation Controller automates updates to YAML files based on the latest images scanned by the Image Reflector Controller, and commits the changes to a given Git repository. The behavior of the automation process is defined by a custom resource named ImageUpdateAutomation.

That resource defines the way that automated commits are created and pushed. The ImagePolicy is another custom resource that determines what image tags go where. ImagePolicy can be defined to select the latest image from images within a SemVer range, or more flexible RegEx filters with alphabetical or numerical sorting to select the “latest” image. Image tags can also be filtered with FilterTags before they are considered as candidate images by the policy rule.

The updates are governed by marking fields to be updated in each YAML file. For each field marked, the automation process checks the image policy named, and updates the field value if there is a new image selected by the policy. The marker format is shown in the image automation guide. Within the spec of ImageUpdateAutomation, the Git commit branch and message can be customized.

Commit flow


A brief outline of the life cycle of a change as it’s processed through Flux, centered around a Git commit:

  1. Bootstrapping Flux is the first step to get started using Flux.

  2. Generating a Flux resource with flux create ....

  3. Previewing changes prior to making or pushing a commit with flux diff kustomization.

  4. Automating upgrades through Image Update Automation resources uses Flux to generate commits when there are updated images available.

  5. Git as a Single Source of Truth means that Flux takes its instructions from a Git commit.

  6. Flux’s default configuration for NetworkPolicy protects Flux services from arbitrary access across the cluster.

  7. Trigger Reconciling on Git push with Webhook Receivers.

  8. GitRepository and other Sources - Artifacts and Revisions.

  9. Secret Decryption via SOPS is performed if an optional decryption configuration is provided.

  10. Kustomize Controller builds and validates resources on the cluster through a server-side apply dry-run operation.

  11. Kustomize Controller applies changes with Server-Side Apply in stages, as an atomic transaction.

  12. Helm Controller reconciles HelmRelease resources through the Helm client library.

  13. Sources and HelmReleases generate HelmChart resources from a HelmRepository before HelmReleases can be installed.

  14. Using a GitRepository-backed or S3-backed HelmRelease is an alternative to use Helm without a HelmRepository.

  15. Channel-based Providers for Notifications re-publish Events from Flux resources at-large to a channel where users can see them.

  16. Git Commit Status Provider Notifications re-publish Events from the Kustomize Controller as commit checks.

  17. Waiting and Health Checking for Flux Kustomization.

Bootstrapping Flux

Bootstrapping is the process of installing Flux so that Flux manages itself. This is the primary official supported way to install and use Flux.

We recommend all new Flux users start reading the documentation from Core Concepts, then follow the introductory guide, Get Started with Flux which covers bootstrapping Flux in-depth.

The bootstrap process installs all of Flux’s core components to a cluster, as well as creating a GitRepository and Kustomization resource to keep all of Flux’s deployed resources updated when any new changes are detected in Git.

Bootstrap also optionally connects with a Git host provider API if needed, to create a Deploy Key in the cluster and apply it to the repository so that private repositories can be used. Flux can be bootstrapped into an existing repository, or Flux can create the repository from scratch. The default bootstrap creates a private repository and generates a deploy key with read-only access for Flux, but there are many other configurations possible.

Once the repository is created and Flux adds its components there as a commit, bootstrap applies the Flux components and Custom Resource Definitions to the cluster, waits for the components to become ready, then applies the Flux sync resources (GitRepository and Kustomization) and finally waits for a successful reconciliation before reporting back to the user that this was successful.

Diagram: Bootstrapping over SSH

sequenceDiagram actor me as admin participant cli as Flux

CLI participant kube as Kubernetes

API server participant flux as Flux

controllers participant git as Git

repository me->>cli: 1. flux bootstrap cli-->>git: 2. push install config cli->>kube: 3. install controllers cli-->>git: 4. set deploy key cli->>kube: 5. set private key cli-->>git: 6. push sync config cli->>kube: 7. apply sync config git-->>flux: 8. pull config flux->>kube: 9. reconcile kube->>cli: 10. report status cli->>me: 11. return status

Generating a Flux resource

After bootstrapping, the Flux CLI provides create generators to help users build more of Flux’s Custom Resources that drive the operation of Flux.

flux create source git --help

flux create kustomization --help

These generators can be used imperatively to create Flux resources in the cluster, or as preferred: when called with the --export option, flux create ... can emit YAML on stdout. That output can be captured in a file, then committed and pushed to create the resource.

Following this process (rather than applying the resources directly to the cluster) maintains the manifests in the repository as the Single Source of Truth, according to GitOps principles.

Some resource options are not available through generators and can only be accessed through fields in YAML; users are generally expected to write resources in YAML and commit them, and should do so when they require access to those features.

For more information, see: flux create.

Flux’s OpenAPI specification can also be integrated with editors to assist Flux users in producing valid YAML for Flux APIs; for a popular example see the Kubernetes Tools for VS Code. This method retrieves the OpenAPI spec from the cluster CRDs and doesn’t require any special configuration for Flux.

Previewing changes

Users have an opportunity to inspect the result of Flux building manifests from a repository from the Flux CLI. This can be done ahead of where Flux actually applies it to the cluster, with flux diff kustomization / flux build kustomization.

Run flux diff kustomization --path=./clusters/my-cluster flux-system from the bootstrap repo, or point it at any other Flux Kustomization and the matching path in your configuration repository to observe what changes Flux will apply, even before they are committed and pushed. This takes account of the cluster state and so it can also be used at any time to check for drift on the cluster that Flux would revert back to the state in Git as soon as the Kustomization is reconciled, or at its next interval.

Any diff containing secret data is obscured so that no secret data is accidentally divulged from the diff output.

Automating upgrades through Image Update Automation resources

Flux can create Git commits to apply updates to the cluster, that are applied in the standard GitOps way to the cluster (as a Git commit), written by a Flux agent called Image Automation Controller. The Image Automation Controller with the help of the Image Reflector Controller works to determine when updates are available and apply them to the cluster.

ImageRepository and ImageUpdateAutomation resources, along with ImagePolicy, are documented in the image automation guide, and in the image automation API docs.

Diagram: Image update to Git

sequenceDiagram actor me participant oci as Image

repository participant irc as Flux

image-reflector-controller participant iac as Flux

image-automation-controller participant kube as Kubernetes

api-server participant nc as Flux

notification-controller participant git as Git

repository me->>oci: 1. docker push irc->>oci: 2. list tags irc->>irc: 3. match tags to policies irc->>kube: 4. update status irc-->>nc: 5. emit events kube->>iac: 6. notify about new tags iac->>git: 7. git clone iac->>iac: 8. patch manifests with new tags iac->>git: 9. git push iac->>kube: 10. update status iac-->>nc: 11. emit events nc-->>me: 12. send alerts

Git as a Single Source of Truth

Flux takes instructions from Git (or another source) which is meant to be the Single Source of Truth. Users create Git commits and push them to the repository that Flux is watching. Except when it happens as a result of Image Automation, the commit itself happens outside of Flux’s purview. The source-controller pulls “commit data” into the cluster.

When the cluster reconciles a Source resource (like GitRepository or Bucket) the content in the new revision is captured on the cluster through a set of filters (.sourceignore, spec.ignore, …) and collected in a tar file to be stored; this file is known as an Artifact in Flux, and it can be accessed by any agent in the flux-system namespace.

When pushed, the receipt of a new commit activates the Git host to fire a webhook to notify subscribers about a push event, which Flux can consume via its Receiver API.

Flux’s default configuration for NetworkPolicy

Arbitrary clients cannot connect to any service in the flux-system namespace, as a precaution to limit the potential for new features to create and expose attack surfaces within the cluster. A set of default network policies restricts communication in and out of the flux-system namespace according to three rules:

  1. allow-scraping permits clients from any namespace to reach port 8080 on any pods in the cluster, for the purpose of collecting metrics. (This can be further restricted when the metrics collectors are known to be deployed in a specific namespace.)

  2. allow-webhooks permits clients from any namespace to reach the Flux notification controller on any port, for the purpose of sending notifications via webhook when events are emitted from sources that the Notification Controller can be subscribed to with a Receiver.

  3. allow-egress permits agents in the flux-system namespace to send traffic outside of the namespace, (for the purpose of reaching any remote GitRepository, HelmRepository, ImageRepository, or Provider), and denies ingress for any traffic from pods or other clients outside of flux-system to prevent any traffic directed into the namespace.

Trigger Reconciling on Git push with Webhook Receivers

When activated by an event from a Receiver, Flux’s Notification controller activates GitRepository or other Flux “sources” ahead of schedule, without first waiting for a spec.interval to elapse.

If Receivers are not configured, the GitRepository will activate on an interval, or can be reconciled on-demand ahead of the interval through the flux reconcile source git. This is a difference between Flux controllers and the Kubernetes Controller Runtime at-large, which Flux’s code is directly based upon, where reconciling is usually done immediately upon detecting a change, rather than at intervals. That behavior is able to be accomplished roughly instantaneously through a publisher-subscriber model.

Resources like GitRepository and Bucket (and other Source API kinds) can also be activated by webhook receivers to provide a similar experience. Webhook receivers are used to make Flux’s pull-based model as fast and responsive as push-based pipelines, but importantly they do not make Flux “push-based” as the event contains no instructions, and only serves as an “early wake-up call” to notify Flux controllers. (It is not intended to be possible for Receivers to change anything else about the behavior of Flux, except for reconciling ahead of schedule.)

Inside the cluster, such subscriptions are auto-negotiated and implied by manifest references like sourceRef; for resources that come from the outside internet, or other non-native resources to Kubernetes, the Flux Receiver and a corresponding webhook configuration at the outside service is manually configured for this purpose instead.

Any Flux resource that subscribes to any outside service (those that are external to the Kubernetes cluster) can be instrumented via webhooks that publish the events. These events can come from container image registries, Helm chart repositories, Git Repositories, or other arbitrary systems that are capable of sending a generic event through HTTP POST to notify when a new artifact is made available.

This capability allows Flux to manage resources outside of the cluster with like-native performance, and helps users avoid creating a performance bottleneck by polling on excessively short intervals.

The period of waiting or spec.interval can be increased or reduced for each Flux resource that reconciles an external data source. The lower bound on that interval is probably best kept above 30s; if tighter responsiveness is needed from Flux, Receivers are expected to be used to close the gap.

This is mentioned here because Flux is specifically used by many organizations seeking to increase velocity and aiming to improve the DORA metrics (from the DevOps Research and Assessment team at Google Cloud).

One of the measures generally considered important is how long it takes for developers to get feedback from CI/CD systems. It’s commonly put forth that “the CI feedback loop should not take longer than 10 minutes.” It should be clear from those relevant materials that for tasks we do many times every day, seconds add up to minutes quickly. For this reason it is recommended to use Receivers wherever possible, at least whenever shortening the feedback loop is to be considered as an important goal.

Sources - (Artifacts and Revisions)

A GitRepository is a Custom Resource which fulfills a more general interface, Source, that saves a read-only view of the latest revision of a repository (usually from outside of the cluster) and hosts its data as a service in the cluster.

The GitRepository Custom Resource describes a (usually remote) Git repository, including the URL of the repository host, and the Git reference (such as a branch name or tag) to monitor for changes. Additionally you may find a secret reference with SSH or TLS keys to verify that host, a secret reference containing authentication credentials, a secret reference containing keys for verifying commit signatures, a configurable polling interval, and other meta information related to the source repository.

The Source Controller treats its connection to each Git repository as read-only, even if the authentication credentials supplied would enable writes to the repository. (Other components, such as the Image Update Automation Controller, may use the Git repository for its access and can write to the repository.)

The Source Controller connects to the repository host, pulls the latest commit for the specified reference, and stores the contents in a bundled and compressed format (currently a gzipped tarball file).

The Source Controller also supports:

Note that it does not make any difference to the Source Controller whether a source is hosted within the cluster or on an external service or server. The Source Controller will still attempt to verify the host using the SSH or TLS keys supplied in the GitRepository Custom Resource and will store its contents as a read-only tarball.

Features include:

  • Validate source definitions
  • Authenticate to sources (SSH, user/password, API token)
  • Validate source authenticity (PGP)
  • Detect source changes based on update policies (semver)
  • Fetch resources both on-demand (via webhooks) and on-a-schedule (at a configured polling interval)
  • Package the fetched resources into a well-known format (tar.gz, yaml)
  • Make the artifacts addressable by their source identifier (sha, version, ts)
  • Make the artifacts available in-cluster to interested 3rd parties (such as the Kustomize Controller and Helm Controller)
  • Notify interested 3rd parties of source changes and availability (status conditions, events, hooks)

Diagram: Cluster sync from Git

sequenceDiagram actor me participant git as Git

repository participant sc as Flux

source-controller participant kc as Flux

kustomize-controller participant kube as Kubernetes

api-server participant nc as Flux

notification-controller me->>git: 1. git push sc->>git: 2. git pull sc->>sc: 3. build artifact for revision sc->>kube: 4. update status for revision sc-->>nc: 5. emit events kube->>kc: 6. notify about new revision kc->>sc: 7. fetch artifact for revision kc->>kc: 8. build manifests to objects kc-->>kc: 9. decrypt secrets kc->>kube: 10. validate objects kc->>kube: 11. apply objects kc-->>kube: 12. delete objects kc-->>kube: 13. wait for readiness kc->>kube: 14. update status for revision kc-->>nc: 15. emit events nc-->>me: 16. send alerts for revision

(This diagram spans some coverage of sections above and below.)

Secret Decryption via SOPS

The Kustomize Controller will decrypt Secret values encrypted using Mozilla’s SOPS CLI with OpenPGP, AWS KMS, GCP KMS or Azure Key Vault and stored in the source Git repository as Kubernetes Secret manifests. The encrypted secret can be safely stored in a public or private Git repository.

The Kustomize Controller will pull the Kubernetes Secret manifest with encrypted values from the Source, (the metadata is stored in plain-text), then Kustomize Controller decrypts its encrypted values using the supplied key.

Note, it’s a good idea to also back up your secret values in a password manager or other form of secure external storage.

The decrypted manifests are kept in memory and passed on to the next stage.

Kustomize Controller builds and validates resources

The Kustomize Controller runs the go library equivalent of a kustomize build against the Kustomization.spec.path to recursively generate and render (or inflate) any Kustomize overlays. (All manifests are passed through Kustomize, even those that don’t include a kustomization.yaml.)

Before it applies YAML or JSON resource declarations to the Kubernetes API for its cluster, the Kustomize Controller reads the artifact files from its source path and builds them using the Kustomize Go library’s build call. This call returns any custom resource definitions (CRDs), namespaces, or other cluster-wide resources it renders before subordinate custom resources or namespace-scoped resources so that they will be available in the API for the resources that refer to or use them.

Kustomize Controller applies changes with Server-Side Apply

Needs a link to Flux documentation for Server-Side Apply process! The only current on-point reference is the Server-side reconciliation is coming blog post.

The Kustomize Controller communicates directly with the Kubernetes API using server-side apply and update API operations instead of running the kubectl apply command as a separate forked process and passing it manifest data through a system pipe. Applying resource manifests directly to the Kubernetes API is both more efficient and provides more control over the process, enabling the Kustomize Controller to give real-time feedback on validation errors, garbage-collection and resource “health assessment” or health checking. It also allows the Kubernetes API to track field management, so different management tools or controllers can set field values within the same resource without interfering with each other.

The server-side apply operation is synchronous rather than asynchronous. If any resources fail to become ready before a specified timeout, the controller can abort the entire transaction. The timeout value is used in two separate contexts, such that either or both of them can take up to spec.timeout seconds before being cancelled or timing out. So the theoretical maximum time for Kustomize to reconcile is 2x spec.timeout but this will only be the case when each of Apply, and Health Checking both take fully up to the maximum allowed time.

The Kustomize Controller applies resource manifests to match the order in which they were rendered by the kustomize build call. It therefore applies any custom resource definition (CRD), namespace, or cluster-scoped resources before their subordinate custom resource or namespace-scoped resources so that they will be available in the API for the resources that refer to or use them.

Helm Controller reconciles HelmRelease resources

A HelmRelease is a composition of a chart, the chart values (parameters to the chart), and any inputs like secrets or config maps that are used to compose the values.

Declaring a HelmRelease will cause the Helm Controller to perform an install using the Helm client libraries. Any changes made to the HelmRelease will trigger the Helm Controller to perform an upgrade to actuate those changes.

You can find more information about HelmReleases here and more general info about Helm for Flux users here.

Sources and HelmReleases generate HelmChart resources from a HelmRepository

When a HelmRelease is first reconciled, the Source Controller polls the source and creates a HelmChart artifact from the data that it retrieves.

A Helm Repository is the native and preferred source for Helm. The Helm Controller works in tandem with the Source Controller to provide a HelmRepository API that collects the correct release version from the helm repo and republishes its data as a HelmChart artifact (another .tar.gz).

The helm repo itself is represented internally in the Source Controller as a YAML index of all releases, including any charts in the repository.

Using a GitRepository-backed or S3-backed HelmRelease

The GitRepository and Bucket sources are also valid for use with Helm Controller.

A GitRepository can be used as a source for Helm Release. The Git repo is not a native storage format for helm and there are some idiosyncrasies when you’re using Helm Controller with a Git repository source. While one can use a GitRepository as a source for HelmRelease, a best practice is to not package many HelmReleases through the same GitRepository, but instead limit it to one chart per GitRepository.

It is a bad idea to create a GitRepository with 400 helm charts. Why exactly? The problem is that Git repo sources are simple .tgz files under the hood, and since it’s not possible to partially fetch such an artifact, this configuration will end up with lots of oversized artifacts that are all pulled each time any of them changes.

Using an analogy, if Flux is a juicemaker and Helm Controller is the business end where you put raw fruit in order to make juice, this is like trying to put the entire bag of oranges into the machine without even opening it up first. While this might work, a clearly better option is to remove the fruits from the bag, feed them in one at a time, and (to stretch the analogy) peel and cut the fruit?

In the analogy, the fruits are the charts and the GitRepository is the bag – clearly it doesn’t make logical sense to do things like this.

Flux provides tools that you have at your disposal for making sources narrowly scoped, here’s one example:

apiVersion: source.toolkit.fluxcd.io/v1
kind: GitRepository
  name: my-chart-git-repo
  namespace: flux-system
  interval: 1h
  url: https://github.com/example/chartsrepo
    branch: main
  ignore: |
    # exclude all
    # include singular charts directory

In the example above, we have a monorepo with many charts in the deploy directory. The artifact (.tgz) that source controller generates will only carry the weight of one single Helm chart inside of it. Flux users with a monorepo configuration such as this should take care that artifacts are not over-inflated in size.

Avoid this potential issue stemming from Helm Controller accidentally pulling in too many resources and causing source controller to repackage them again, when you did not intend to include copies of many charts together in one artifact. There is no partial or sparse checkout for Flux source artifacts; once they are packed, they can only be downloaded in full.

Take care also that with the approach you use, many releases are not being triggered all at once. Setting HelmChartTemplate.spec.reconcileStrategy to Revision can reduce the amount of work required to publish changes via Helm, however this strategy can be broadly ill-advised, and is not recommended to use without caution as all charts will be released again for every commit, even when there are no changes.

The default ChartVersion strategy behavior is not as potentially harmful as Revision, but may have other drawbacks as its function is not obvious. With ChartVersion, the source controller will only reconcile a new version of the HelmChart when the version is changed in Chart.yaml.

A GitRepository for every chart is a lot of boilerplate, but this is currently the most optimal way to host charts directly in a GitRepository.

It is recommended that users who hit scaling issues publish the chart in a HelmRepository for a better overall experience.

Diagram: Helm release upgrade from Git

sequenceDiagram actor me participant git as Git

repository participant sc as Flux

source-controller participant hc as Flux

helm-controller participant kube as Kubernetes

api-server participant nc as Flux

notification-controller me->>git: 1. git push sc->>git: 2. git pull sc->>sc: 3. build chart for revision sc->>kube: 4. update chart status sc-->>nc: 5. emit events kube->>hc: 6. notify about new revision hc->>sc: 7. fetch chart hc->>kube: 8. get values hc->>hc: 9. render and customize manifests hc-->>kube: 10. apply CRDs hc->>kube: 11. upgrade release hc-->>kube: 12. run tests hc-->>kube: 13. wait for readiness hc->>kube: 14. update status hc-->>nc: 15. emit events nc-->>me: 16. send alerts

Channel-based Providers for Notifications

Notification Providers are used by Flux for outbound notifications to platforms like Slack, Microsoft Teams, Discord and others. The Notification Provider manifest must contain an identifier to connect to the receiver platforms, usually spec.address, and an authorization token which should be stored in a secret that we will reference in the notification Provider as spec.secretRef.name.

They are driven by Alerts, another CRD in the Flux Notification Controller’s API. Alerts create notifications from events, and all of the flux reconcilers generate events while they are undergoing status transitions. The Alerts are used to filter the events generated by flux reconcilers using spec.eventSources and the spec.eventSeverity and then they are forwarded to a Provider specified on spec.providerRef.

To avoid duplicated alerts Events are rate limited based on the InvolvedObject.Name, InvolvedObject.Namespace, InvolvedObject.Kind, Message, and Metadata.revision. The interval of the rate limit is set by default to 5m, but it’s configurable.

Git Commit Status Provider Notifications

Git Commit Status Providers work similarly to other notification providers however they target a specific commit with their event. If you [set up Git commit status notifications][Setup Git Commit Status Notifications] through an integration for GitHub, GitLab, Bitbucket (or any supported Git providers) Flux will display success or failure reported on each commit from any alerts targeting the provider. This feature is restricted to Kustomization as an event source since the Git Commit Status Providers require a commit hash to be present in the metadata.

The provider will continuously receive events as they happen, and multiple events may be received for the same commit hash. The Git providers are configured to update the status only if it has changed. This avoids repeatedly spamming the commit status history.

Waiting and Health Checking for Flux Kustomization

Kustomize Controller can be configured with or without spec.wait which decides whether the Kustomization will be considered ready as soon as the resources are applied, or if the Kustomization will not be considered ready until the resources it created are all marked as ready.

The health checking feature is called Health Checks in the Flux Kustomization API.

Last modified 2024-05-13: upgrade apis for helm GA (b3dba1e)