Upgrade Advisory

This documentation is for Flux (v1) and Helm Operator (v1). Both projects are in maintenance mode and will soon reach end-of-life. We strongly recommend you familiarise yourself with the newest Flux and start looking at your migration path.

For documentation regarding the latest Flux, please refer to this section.

Frequently asked questions

A HelmRelease is stuck in a failed release state. How do I force it through?

You need to manually perform a helm rollback or helm upgrade to get it in a healthy deployed state again before the Helm Operator will continue. Read more about this in the HelmRelease guide.

A HelmRelease reports a deployed release status but also an error. What is going on?

The release status recorded in the status sub-resource of the HelmRelease is the status as given by Helm for the release. It is polled from Helm on a interval (see --status-update-interval).

The message is the error reported during the latest release failure. When this failure was for example due to a validation error, the upgrade has not been completed, and the status of the release will thus still be deployed.

I want to automate chart releases from Helm chart repositories. Is this possible?

This is possible by wrapping the chart from the Helm chart repository in an umbrella chart sourced from a Git source, read more about this here.

It takes a long time before the operator processes an update to a resource. How can I speed up the processing of releases?

The operator watches for changes of Custom Resources of kind HelmRelease. It receives Kubernetes Events and queues these to be processed. All resources will also be re-queued every --charts-sync-interval (default 3m) for a dry-run to detect and revert manual changes made to a release.

Depending on how many resources the operator is watching and the complexity of the charts (umbrella charts for example generally take a longer time to process), the default number of workers may not be sufficient to instantly process a release the moment it enters the queue as the queue works on a FIFO basis, and other releases may have to be processed first.

The solution is to increase the number of workers processing the releases using the --workers flag (default 2), i.e. by steps of 2 until the releases are processed within a time frame that feels right to you.

If this does not give the desired effect, or if the number of workers required is unacceptably high, there are two other tweaks possible:

  1. increasing the --charts-sync-interval; this causes the queue to be less heavy occupied at the cost of detecting mutations less rapidly
  2. using multiple Helm Operator instances, i.e. by having one operator per namespace; namespace scoping is possible by configuring the --allow-namespace flag

The Helm Operator is taking up too many resources. How can I limit the resource usage?

When you are running the Helm Operator on an edge, home lab, or other resource constrained cluster it may take up too many resources for your liking.

  1. In case of a high number of HelmRelease resources, think about a namespace restricted multi-tenant setup as an option to spread the load and have finer grain control over configuration.

  2. Increase the various polling intervals based on what your environment really needs:

    - `--status-update-interval`: The interval at which Helm is polled for
      the current state of the release, generally safe to increase.
    - `--charts-sync-interval`: The interval at which `HelmRelease` resources
      are re-queued for reconciliation. Increasing this interval will result
      in mutations to a chart release being detected slower.
    - `--git-poll-interval`: The interval at which Git repositories are
      polled for changes for charts from Git sources.

I have manually deleted a Helm release using helm. Why is the operator not able to restore it?

When you delete a Helm release with helm delete <name> using Helm 2, the release name can not be re-used as the history of the old release is still stored in the Helm storage under the same name. You need to use the helm delete --purge option, only then the release history will be removed from the Helm storage and the Helm Operator will then be able to reinstall the release.

I am using SSL between Helm and Tiller. How can I configure the operator to use the certificate?

When installing the Helm Operator, you can supply the CA and client-side certificate using the tls options on the chart. More details about this on:

I am using Flux and have deleted a HelmRelease file from Git. Why is the Helm release still running on my cluster?

Flux does not delete resources by default; you can enable Flux garbage collection feature by passing the command-line flag --sync-garbage-collection to fluxd.

With Flux garbage collection enabled, the Helm Operator will receive the delete event and will purge the Helm release.

Are there prerelease builds I can run?

There are builds from CI for each merge to master branch. See fluxcd/helm-operator-prerelease.