This commit adds support for optional values references, as discussions
have brought to light that there are some valid use cases in which
having this option is desirable. For example to allow a user to extend
behaviour at a later date with their own repository without modifying
the HelmRelease object.
When a values reference is marked as optional, not found errors for the
value reference are ignored, but any ValuesKey, TargetPath or transient
error will still result in a reconciliation failure.
This also removes the sorting from the `HelmChartWatcher`, as with
the current `HelmChartTemplateSpec` a chart is only used by a single
`HelmRelease`. Rendering the action obsolete.
This removes:
- Installed, Upgraded, RolledBack, and Uninstalled status conditions
since they did not represent current state, but rather actions
taken, which are already recorded by events.
- status.observedStateReconciled since it solved the problem of
remembering past release (install/upgrade/test) success, but not
past release failures, after other subsequent failures such as
dependency failures, k8s API failures, etc.
This adds:
- Remediated status condition which records whether the release is
currently in a remediated state. It is used to prevent release retries
after remediation failures. We were previously not doing this for
rollback failures.
- Released status condition which records whether the current state
has been successfully released (install/upgrade/test). This is used to
remember the last release attempt status, regardless of any subsequent
other failures such as dependency failures, k8s API failures, etc.
This renames:
- Tested > TestsSuccess status condition, for forward compatibility
with interval based helm tests.
During high custom resource count / low interval tests, I was greated
with a `cannot patch resource "events"` message. This happened due to
event compaction, where it will perform a patch instead of a create.
By giving the role the permission to do so this should no longer pose
a problem.
This adds a .status.lastObservedTime field which records when the
HelmRelease was last observed by the controller. This allows the user
to observe whether the spec.interval and reconcileAt annotations are
triggering reconciliation attempts as desired.
- Ensure upgrade actually occurs if known state was not reached
for any reason (other than install failure).
- After transient failures not tied to new state application, ensure
spurious upgrades do not occur and ready state is again reached,
by remembering that the known state was already successfully applied.
- Reset failure counts after success so they're not stale.
- Only lookup post-deployment release revision on remediation,
since otherwise we already have it.
- Push status update after finding new state so user can observe.
This commit adds support for conditional remediation, enabling the user
to:
* configure if test failures should be ignored
* configure what action should taken when a Helm install or upgrade
action fails (e.g. rollback, uninstall)
* configure if a failed Helm action should be retried
* configure if a failed release should be kept for debugging purposes
The previous behaviour where failed Helm tests did not mark the
`HelmRelease` as not `Ready` has changed, it now marks them as failed
by default.
Co-authored-by: Hidde Beydals <hello@hidde.co>
After taking a closer look at the Helm code, this is the behaviour
we are actually after, given we want to mimic the merging behaviour
of the -f flag available to Helm install and/or upgrade commands.
To allow the Helm Operator and helm-controller HelmReleases to
co-exist in the cluster, while being handled by separate controllers
during e.g. the migration period.
This is not possible without using another domain due to how Custom
Resource Definitions work, as newer API versions are seen as a
replacement of older versions, and are expected to be handled by a
single controller.
As "condition type names should describe the current observed state of
the resource, rather than describing the current state transitions".
Described by the draft convention for update conditions in
kubernetes/community#4521.
* Ensure Status.LastReleaseRevision is 0 after uninstall.
* Ensure uninstall leads to ready condition being false. This has
no effect now since uninstall is only triggered on install failure,
but will once other triggers (such as test failure) are supported.
Hidde Beydals
Sean Eagan
Michael Bridgen
Sean Eagan
stefanprodan