Github user EronWright commented on a diff in the pull request:
https://github.com/apache/flink/pull/4628#discussion_r140051191
--- Diff:
flink-mesos/src/main/java/org/apache/flink/mesos/runtime/clusterframework/MesosFlinkResourceManager.java
---
@@ -663,6 +666,7 @@ private void taskTerminated(Protos.TaskID taskID,
Protos.TaskStatus status) {
//
------------------------------------------------------------------------
private LaunchableMesosWorker createLaunchableMesosWorker(Protos.TaskID
taskID) {
+ setCoTaskGetter();
--- End diff --
I believe that this approach is violating an Akka rule by passing a direct
reference to the RM object to the launch coordinator. I understand the
challenge to be that Fenzo's `BalancedHostAttrConstraint` expects a
`coTasksGetter` function
([ref](http://netflix.github.io/Fenzo/fenzo-core/com/netflix/fenzo/plugins/BalancedHostAttrConstraint.html#BalancedHostAttrConstraint-com.netflix.fenzo.functions.Func1-java.lang.String-int-)).
But that function cannot keep a reference to the RM object.
An imperfect solution would be to pass the `taskIds` to the
`LaunchableMesosWorker` constructor; imperfect because the list wouldn't be
dynamic.
In truth we want all tasks to be balanced and so the getter is actually
overkill. If you look at the implementation of `BalancedHostAttrConstraint`
([ref](https://github.com/Netflix/Fenzo/blob/master/fenzo-core/src/main/java/com/netflix/fenzo/plugins/BalancedHostAttrConstraint.java))
it uses the function output to lookup tasks in the `TaskTrackerState`. A
custom constraint could simply use all tasks in `TaskTrackerState` without
need for a `coTasksGetter`.
Once we get this sorted out, some of the code should be moved to a utility
class.
---
