[ https://issues.apache.org/jira/browse/FLINK-1675?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14711321#comment-14711321 ]
Ufuk Celebi commented on FLINK-1675: ------------------------------------ (3) has been addressed in 8261ed5438278331c9cd760463273ad94d4bd410. Except for (1), all of this is for 0.10. > Rework Accumulators > ------------------- > > Key: FLINK-1675 > URL: https://issues.apache.org/jira/browse/FLINK-1675 > Project: Flink > Issue Type: Bug > Components: JobManager, TaskManager > Affects Versions: 0.9 > Reporter: Stephan Ewen > Fix For: 0.10 > > > The accumulators need an overhaul to address various issues: > 1. User defined Accumulator classes crash the client, because it is not > using the user code classloader to decode the received message. > 2. They should be attached to the ExecutionGraph, not the dedicated > AccumulatorManager. That makes them accessible also for archived execution > graphs. > 3. Accumulators should be sent periodically, as part of the heart beat that > sends metrics. This allows them to be updated in real time > 4. Accumulators should be stored fine grained (per executionvertex, or per > execution) and the final value should be on computed by merging all involved > ones. This allows users to access the per-subtask accumulators, which is > often interesting. > 5. Accumulators should subsume the aggregators by allowing to be "versioned" > with a superstep. The versioned ones should be redistributed to the cluster > after each superstep. -- This message was sent by Atlassian JIRA (v6.3.4#6332)