[jira] [Commented] (GEARPUMP-32) Minimum clock of source Tasks maybe inaccurate

["ASF GitHub Bot (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/GEARPUMP-32?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15425584#comment-15425584 ] 

ASF GitHub Bot commented on GEARPUMP-32:
----------------------------------------

Github user kkasravi commented on the issue:

    https://github.com/apache/incubator-gearpump/pull/67
  
    Is the basic idea to enable event time processing vs operator time processing by using KafkaMessages?


> Minimum clock of source Tasks maybe inaccurate
> ----------------------------------------------
>
>                 Key: GEARPUMP-32
>                 URL: https://issues.apache.org/jira/browse/GEARPUMP-32
>             Project: Apache Gearpump
>          Issue Type: Bug
>          Components: streaming
>    Affects Versions: 0.8.0
>            Reporter: Manu Zhang
>            Assignee: Manu Zhang
>             Fix For: 0.8.1
>
>
> Moved from [https://github.com/gearpump/gearpump/issues/1835] and reported by [Zhu Yueqian|https://github.com/yueqianzhu]
> {quote}
> Source tasks have not any upstreamClocks. So, startClock is the minimum of pending clocks when recover happen.
> eg below:
> source task1: timeStamp:15,not ACK, minClockValue maybe is 15(<= 15).
> source task2: timeStamp:10,ACKed, minClockValue maybe is Long.MaxValue
> when recover happen,startClock maybe is 15. where is the data between 10 to 15 at task2?
> {quote}
> More context on this issue:
> In Gearpump, we maintain a global minimum clock tracked from a message's timestamp across all tasks. It means messages with timestamp before this clock have all been processed. An application will restart from this value on failure, and thus at-least-once message delivery could be guaranteed. 
> The global minimum clock is the lower bound of all the Tasks' minimum clocks. 
> For a task, the minimum clock is the lower of 
> # upstream minimum clock
> # a. the minimum timestamp of unacked messages
>    b. Long.MaxValue if all messages have been acked.
>  
> Note that 2.b allows the global minimum clock to progress and it is almost safe since the clock is also bounded by the upstream minimum clock. I said "almost safe" because a source task has no upstream but we assume the upstream minimum clock is Long.MaxValue. Thus, the scenario described by Zhu Yueqian could happen and breaks at-least-once guarantee. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)