[
https://issues.apache.org/jira/browse/FLINK-2590?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14724924#comment-14724924
]
ASF GitHub Bot commented on FLINK-2590:
---------------------------------------
Github user tillrohrmann commented on the pull request:
https://github.com/apache/flink/pull/1075#issuecomment-136619441
@s1ck, it's important to note that `1` will be subtracted from
`getRuntimeContext().getNumberOfParallelSubtasks()` and not `getBitSize()`. The
reason is that we have `0` based indices for the subtasks. Thus, we only have
to calculate the maximum needed bits for the highest index we can encounter.
And this is `getRuntimeContext().getNumberOfParallelSubtasks() - 1`. Thus if
`getNumberOfParallelSubtasks == 7`, then we would calculate `getBitSize(6) ==
3`.
> DataSetUtils.zipWithUniqueID creates duplicate IDs
> --------------------------------------------------
>
> Key: FLINK-2590
> URL: https://issues.apache.org/jira/browse/FLINK-2590
> Project: Flink
> Issue Type: Bug
> Components: Java API, Scala API
> Affects Versions: 0.10, master
> Reporter: Martin Junghanns
> Assignee: Martin Junghanns
> Priority: Minor
>
> The function creates IDs using the following code:
> {code:java}
> shifter = log2(numberOfParallelSubtasks)
> id = counter << shifter + taskId;
> {code}
> As the binary function + is executed before the bitshift <<, this results in
> cases where different tasks create the same ID. It essentially calculates
> {code}
> counter*2^(shifter+taskId)
> {code}
> which is 0 for counter = 0 and all values of shifter and taskID.
> Consider the following example.
> numberOfParallelSubtaks = 8
> shifter = log2(8) = 4 (maybe rename the function?)
> produces:
> {code}
> start: 1, shifter: 4 taskId: 4 label: 256
> start: 2, shifter: 4 taskId: 3 label: 256
> start: 4, shifter: 4 taskId: 2 label: 256
> {code}
> I would suggest the following:
> {code}
> counter*2^(shifter)+taskId
> {code}
> which in code is equivalent to
> {code}
> shifter = log2(numberOfParallelSubtasks);
> id = (counter << shifter) + taskId;
> {code}
> and for our example produces:
> {code}
> start: 1, shifter: 4 taskId: 4 label: 20
> start: 2, shifter: 4 taskId: 3 label: 35
> start: 4, shifter: 4 taskId: 2 label: 66
> {code}
> So we move the counter to the left and add the task id. As there is space for
> 2^shifter numbers, this prevents collisions.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
