Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
2.0.0
-
None
-
None
Description
Now for many concurrent access resource in Nimbus.java, we use AtomicReference to make them multi thread safe. The resources summarized below:
1. heartbeatsCache
2. schedulingStartTimeNs
3. idToSchedStatus
4. nodeIdToResources
5. idToWorkerResources
6. idToExecutors
The 1, 4, 5 and 6 may grows huge if we have hundreds of topologies on cluster, when we update AtomicReference, actually we passed in a Function and use compareAndSet to update the whole val to the new returned by the Function, in that case, we must do a reference copy and merge the changes, which seems not necessary.
I think the reason to use AtomicReference is a legacy from old Clojure code, we can replace them totally with ConcurrentHashMap which supported better performance.