WIP: Expose a handle to cleanly cancel a reflector #6602
Conversation
|
This is pretty complicated, it's not clear to me what problem you're solving. Please see my pending #6546. I think once that is in you should be able to easily wait on a condition with something like the below (can package this up into a few lines of code for easy usage multiple places). outputChan := make(chan interface{})
stopCh := make(chan struct{})
checkFunc := func(obj interface{}) {
if /* obj meets condition */ {
close(stopCh)
outputChan <- obj
}
}
_, controller := framework.NewInformer(
c.createAssignedPodLW(),
&api.Pod{},
time.Second*30,
framework.ResourceEventHandlerFuncs{
AddFunc: checkFunc,
UpdateFunc: checkFunc,
},
)
go controller.Run(stopCh)
obj := <- outputChan |
|
|
||
| // Stopping the watcher should be idempotent and if we return from this function there's no way | ||
| // we're coming back in with the same watch interface. | ||
| defer w.Stop() |
There was a problem hiding this comment.
This line's change is good :)
But otherwise, I think I pretty strongly prefer just plumbing two channels down to watchHandler. 'exitWatch' can be renamed to resyncChan and we can add a 'stopCh'. Resyncing and stopping are just two different things. Also--go design style--if you're using both mutexes and channels, something is wrong. :)
There was a problem hiding this comment.
Agree about the refactor, don't completely agree about the difference between resync and stop. In general I feel like they're different, but the way listAndWatch behaves today, they feel essentially the same. A resync is a cancel unless called from withing a while loop.
|
feedback received, will iterate in another pr that addresses both RunUntil stopping and the leaked channel |
@lavalamp need opinions before I move on.
This is the first step in teaching
kubectl stopto watch status.replicas of an rc, the unittest has a fairly complete example of how I plan to use reflector cancellation in creating an edge trigger watcher (watches a resource for a single value, in this case status.replicas=0).This PR exposes a CancellationController capable of dictating when a reflector should cleanup and unwind the listAndWatch stack. The cancellation controller uses a single channel as a conduit for cancellation signals. The signal might take the form of either a timeout (resync period of the reflector) or a stop event sent from some client of the reflector that no longer cares about what it's watching (all the resizers in kubectl).
For this to WAI we also need to send a stop signal to reflector.RunUnti, or the reflector will just retry. This is how we'd get the stop signal to stop watching, while the resync timeout will not stop watching.
Also fixes #6488