Previously, the `helm ls --$state` operation would display outdated
releases under certain conditions.
Given the following set of releases:
```
NAME REVISION UPDATED STATUS CHART APP VERSION NAMESPACE
bar 1 Wed Apr 8 16:54:39 2020 DEPLOYED bar-4.0.0 1.0 default
foo 1 Fri Feb 7 06:16:56 2020 DEPLOYED foo-0.1.0 1.0 default
foo 2 Mon May 4 07:16:56 2020 FAILED foo-0.1.0 1.0 default
foo 3 Mon May 4 07:20:00 2020 FAILED foo-0.1.0 1.0 default
foo 4 Tue May 5 08:16:56 2020 DEPLOYED foo-0.2.0 1.0 default
qux 1 Tue Jun 9 10:32:00 2020 DEPLOYED qux-4.0.3 1.0 default
qux 2 Tue Jun 9 10:57:00 2020 FAILED qux-4.0.3 1.0 default
```
`helm ls --failed` produced the following output:
```
NAME REVISION UPDATED STATUS CHART APP VERSION NAMESPACE
foo 3 Mon May 4 07:20:00 2020 FAILED foo-0.1.0 1.0 default
qux 2 Tue Jun 9 10:57:00 2020 FAILED qux-4.0.0 1.0 default
```
Including the `qux` release in that `helm ls --failed` output is not
controversial; the most recent revision of `qux` was not successful
and an operator should investigate.
Including the `foo` release in the output, however, is
questionable. Revision 3 of `foo` is _not_ the most recent release of
`foo`, and that FAILED release was fixed in a susubsequent upgrade. A
user may see that FAILED deploy and start taking inappropriate
action. Further, that issue was fixed months ago in this example --
troubleshooting an old deploy may not be safe if significant changes
have occurred. Concern over this behavior was raised in
https://github.com/helm/helm/issues/7495.
This behavior applied to all the state filter flags (--deployed,
--failed, --pending, etc.), and a user could pass multiple state
filter flags to a single command. The previous behavior can be
summarized as follows:
For each release name, all release revisions having any of the
supplied state flags were retrieved, and the most recent revision
among these was returned (regardless of whether a newer revision of an
unspecified state exists).
This change request alters the helm list action to match user
expectations such that only "current" releases are shown when
filtering on release state. After this change, the following output
would be produced by `helm ls --failed`:
```
NAME REVISION UPDATED STATUS CHART APP VERSION NAMESPACE
qux 2 Tue Jun 9 10:57:00 2020 FAILED qux-4.0.0 1.0 default
```
The command now returns only `qux` because it is the only "current" FAILED release.
This behavior change applies to all the state filters _except_
`superseded`, which now becomes a special case. By definition, at
least one newer release exists ahead of each superseded release. A
conditional is included in this change request to maintain the
preexisting behavior (return "most recent" superseded revison for
each release name) if the superseded state filter is requested.
---
Note that there is an alternate perspective that a state filter flag
should return all releases of a given state rather than only the
"current" releases. In the above example, `helm ls --failed` with this
approach would return the following:
```
NAME REVISION UPDATED STATUS CHART APP VERSION NAMESPACE
foo 2 Mon May 4 07:16:56 2020 FAILED foo-0.1.0 1.0 default
foo 3 Mon May 4 07:20:00 2020 FAILED foo-0.1.0 1.0 default
qux 2 Tue Jun 9 10:57:00 2020 FAILED qux-4.0.0 1.0 default
```
Multiple FAILED `foo` revisions are included in the output, unlike the current behavior.
This approach is logical and achievable. It allows a user to find
exactly what is requested: all historical releases of a given
state. In order to achieve continuity with helm behavior, however, a
new filter (something like "current") would probably need to be
implemented and become the new default.
Given current helm behavior as well as the comments in the #7495, I
did not pursue this approach.
---
Technical details:
- Moved list action state mask filter after latest release filter
Previously, the list operation in helm/pkg/action/list.go skipped
releases that were not covered by the state mask on _retrieval_ from
the Releases store:
```
results, err := l.cfg.Releases.List(func(rel *release.Release) bool {
// Skip anything that the mask doesn't cover
currentStatus := l.StateMask.FromName(rel.Info.Status.String())
if l.StateMask¤tStatus == 0 {
return false
}
...
```
8ea6b970ec/pkg/action/list.go (L154-L159)
While filtering on retrieval in this manner avoided an extra iteration
through the entire list to check on the supplied condition later, it
introduced the possibility of returning an outdated release to the
user because newer releases (that would have otherwise squashed
outdated releases in the `filterList` function) are simply not
included in the set of working records.
This change moves the state mask filtering process to _after_ the set
of current releases is built. Outdated, potentially misleading
releases are scrubbed out prior to the application of the state mask
filter.
As written, this state mask filtration (in the new `filterStateMask`
method on `*List`) incurs an additional, potentially expensive
iteration over the set of releases to return to the user. An
alternative approach could avoid that extra iteration and fit this
logic into the existing `filterList` function at the cost of making
`filterList` function a little harder to understand.
- Rename filterList to filterLatestReleases for clarity
Another function that filters the list is added, so update
to the more descriptive name here.
- List superseded releases without filtering for latest
This change makes superseded releases a special case, as they would
_never_ be displayed otherwise (by definition, as superseded releases have been
replaced by a newer release), so a conditional maintains current
behavior ("return newest superseded revision for each release name")
Fixes#7495.
Signed-off-by: Andrew Melis <andrewmelis@gmail.com>
The existing unit test doesn't cover the most important functionality
of filterList(). Adding that check. Separately test the specific
namespace-related behavior.
Remove the dead variable anotherOldOne.
Signed-off-by: Jakub Bielecki <jakub.bielecki@codilime.com>
* feat: add pkg/action to encapsulate action logic
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
* feat: replace client/server internals with action package
While we removed Tiller, we left the internal client/server architecture mostly intact. This replaces that architecture with the `pkg/action` package.
This implements the action package for list, but nothing else.
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
* feat: Add install and refactor some tests
This adds install to the action package, and then fixes up a lot of testing.
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
* fix: Move a bunch of sorters to the releaseutils package
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
* fix: updated APIs and fixed a failed test
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
* Use var for timestamper, instead of adding as a struct field
Signed-off-by: Matt Butcher <matt.butcher@microsoft.com>
Dmitry Chepurovskiy
Andrew Melis
jabielecki
Karuppiah Natarajan
Matt Farina
Abhilash Gnan
Adam Reese
Matthew Fisher
Matt Butcher