beam-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Stephan Ewen (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (BEAM-1612) Support real Bundle in Flink runner
Date Thu, 04 May 2017 08:49:04 GMT

    [ https://issues.apache.org/jira/browse/BEAM-1612?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15996375#comment-15996375
] 

Stephan Ewen commented on BEAM-1612:
------------------------------------

[~kenn] Ah, I think there are two notions floating around about what a bundle should be:

  - The bundle of records between persistence boundaries (flush, commit, checkpoint, whatever)
- [~kenn]'s notion
  - Some arbitrary form of mini batching [~aljoscha]'s notion

I think that notifying the user code about persistence boundaries is actually a good idea,
if they want to coordinate their flushes / rollover / etc at these boundaries. For the just
arbitrary mini batching - what speaks against small/fast processing time windows for that?

> Support real Bundle in Flink runner
> -----------------------------------
>
>                 Key: BEAM-1612
>                 URL: https://issues.apache.org/jira/browse/BEAM-1612
>             Project: Beam
>          Issue Type: Improvement
>          Components: runner-flink
>            Reporter: Jingsong Lee
>
> The Bundle is very important in the beam model. Users can use the bundle to flush buffer,
can reuse many heavyweight resources in a bundle. Most IO plugins use the bundle to flush.

> Moreover, FlinkRunner can also use Bundle to reduce access to the FlinkState, such as
first placed in JavaHeap, flush into RocksDbState when invoke finishBundle , this can reduce
the number of serialization.
> But now FlinkRunner calls the finishBundle every processElement. We need support real
Bundle.
> I think we can have the following implementations:
> 1.Invoke finishBundle and next startBundle in {{snapshot}} of Flink. But sometimes this
"Bundle" maybe too big. This depends on the user's checkpoint configuration.
> 2.Manually control the size of the bundle. The half-bundle will be flushed to a full-bundle
by count or eventTime or processTime or {{snapshot}}. We do not need to wait, just call the
startBundle and finishBundle at the right time.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Mime
View raw message