Leader remove runlocal

[jmrodri]

No description provided.

[coveralls]

Pull Request Test Coverage Report for Build 181321825

• 23 of 36 (63.89%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-25.9%) to 74.074%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
leader/leader.go	23	36	63.89%

Totals
Change from base Build 180379840:	-25.9%
Covered Lines:	220
Relevant Lines:	297

---

💛 - Coveralls

[joelanford]

One more review round (sorry!)

And another thing that just popped into my head. It occurs to me that controller-runtime has some pre- and post- leader-election runnables as part of the manager. I wonder if it would be possible to shim this leader election implementation into controller-runtime. This point is not a PR a blocker, but it would be nice to handle in a follow-up if it is possible. I'll investigate and report back.

[joelanford]

I think this Describe layer is unnecessary. Every test in this suite (i.e. package) is a leader test.

[jmrodri]

If I remove it, it doesn't seem to work. I renamed it to Leader election to better describe it. I would say, it works so I'd rather not change this.

[joelanford]

Ah crap. I totally messed this up. Now that I'm seeing this, I feel like it's a little user-hostile. If we don't export Config, users can't provide their own Option implementations.

Given that we only have one field to set, it might not seem important now, but if we add more fields in the future, we want to give users the flexibility to write their own Options.

Opening this up later would be a breaking change to the Option definition.

So given all of this, I think we want:

type Option func(*Config) error

type Config struct { ... }

// now that Config is exported, setDefaults would need 
// to be unexported since this only needs to be called
// by us after all options have run.
func (c *Config) setDefaults() error { ... }

Sorry for the back and forth!

[jmrodri]

Reverted the commit that unexported Config. I then added a new commit to unexport just setDefaults

[joelanford]

Following up about controller-runtime manager integration.

It doesn't seem immediately do-able given controller-runtime's direct use of resourcelock.Interface, which seems to have a hardcoded expectation of using leases.

Here's where the manager starts the pre-leader-election runnables vs the post-leader-election runnables: https://github.com/kubernetes-sigs/controller-runtime/blob/ab02953d0bcb6467333fcc5809eb1df83a8d89c4/pkg/manager/internal.go#L465-L478.

The only way I can see this working would be for controller-runtime to give callers a hook into startLeaderElection to override the default. Perhaps it could abstract the leader election implementation to something that allows a LeaderCallbacks to be injected and implements Run(context.Context). So perhaps:

type LeaderElector interface {
    Run(context.Context)
}

And then a new manager Option field:

type Option struct {
	// NewLeaderElector overrides the default lease-based leader elector
	// and allows callers to use their own leader election implementation.
	NewLeaderElector func(leaderelection.LeaderCallbacks) LeaderElector
}

This would definitely require a controller-runtime proposal and thorough discussion, so not anything that could happen anytime really soon.

[camilamacedo86]

We replaced the sdk impl for the controller-runtime one. So, I do not think that we need to keep it at all.

@joelanford why we will keep our implementation since all opers now are using the controller-runtime one? If yes, i need to revert the change made in operator-framework/operator-sdk#3514.

c/c @jmrodri

[jmrodri]

@camilamacedo86 my understanding was that this is a leader for life and we kept it for someone that wanted to have leader for life. It will not be used in the SDK but could be used by an operator developer.

[mhrivnak]

A lot of existing operators are using the leader-for-life implementation. They'll either need a way to continue using that, or will need an easy path to migrating to the lease-based election.

These days the lease-based election is easy to use and robust, which hasn't always been the case. But even still, each pattern has advantages and disadvantages.

[joelanford]

Can we add a test case for returning ErrNoNamespace from Become and using errors.Is(err, ErrNoNamespace)?

[joelanford]

/lgtm

[openshift-ci-robot]

New changes are detected. LGTM label has been removed.