Make JoinAsync and JoinSeedNodesAsync more robust by checking cluster UP status #6033
Conversation
by design, although the member up actors could be programmed to shut themselves down after they've executed. |
Unlikely with closures, but theoretically possible. I wouldn't worry about it - end-user's responsibility. |
|
Changed how the state is handled, no longer using null as a marker. @Aaronontheweb @to11mtm, Would be great if you can re-review this code again. |
src/core/Akka.Cluster/Cluster.cs
Outdated
| { | ||
| _isUp.GetAndSet(true); | ||
| // If there is an async join operation in progress, complete it. | ||
| _asyncJoinTaskSource?.Complete(); |
There was a problem hiding this comment.
This looks a -little- closer...
IMO Ideally, we should only lock long enough to see if there is an existing continuation set up, and if so re-use that.
Does _asyncJoinTaskSource have a TryComplete to use instead?
There was a problem hiding this comment.
It uses TryComplete internally, the lock is to guard against possible null value
There was a problem hiding this comment.
_asyncJoinTaskSource is actually a custom wrapper class around TaskCompletionSource to provide timeout with custom exception support.
….com:Arkatufus/akka.net into cluster/fix_JoinAsync_and_JoinSeedNodesAsync
|
Here is how I would simplify this:
|
|
I've checked the cluster status change listener actor, its appropriately designed, no problem there. |
….com:Arkatufus/akka.net into cluster/fix_JoinAsync_and_JoinSeedNodesAsync
Arkatufus
changed the title
| _clusterDaemons.Tell(new InternalClusterAction.AddOnMemberRemovedListener(() => tcs.TrySetResult(null))); | ||
| _clusterDaemons.Tell(new InternalClusterAction.AddOnMemberRemovedListener(() => | ||
| { | ||
| tcs.TrySetResult(null); |
We used the `cluster.seed-node-timeout` property incorrectly in akkadotnet#6033 - that is a "how long did it take from me to hear back from a seed" setting, but we're using it like a "how much time do I have to join the cluster?" setting. Added a function designed to give us at least 20s of leeway.
* fix timing regressions in `Cluster.JoinAsync` methods We used the `cluster.seed-node-timeout` property incorrectly in #6033 - that is a "how long did it take from me to hear back from a seed" setting, but we're using it like a "how much time do I have to join the cluster?" setting. Added a function designed to give us at least 20s of leeway. * Update Cluster.cs * fixed addr * Fix timeout calculation * Fix CancellationTokenSource --------- Co-authored-by: Gregorius Soedharmo <[email protected]>
Changes
JoinAsync and JoinSeedNodesAsync are now idempotent, always returning the same Task instance when called multiple times.More robust memory clean-upKISS and code reuse, both JoinAsync and JoinSeedNodesAsync shares the same async instance.