You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@flink.apache.org by Aljoscha Krettek <al...@apache.org> on 2016/03/15 11:02:55 UTC
[DISCUSS] Removing Non-Keyed WindowOperator
Hi folks,
I opened a Jira Issue: https://issues.apache.org/jira/browse/FLINK-3614
This is the text of the issue:
I propose to remove the special Non-Keyed Window Operator and implement non-parallel windows by using the standard WindowOperator with a dummy KeySelector.
Maintaining everything for two WindowOperators is a huge burden. The implementation is completely separate by now. For example, the Non-Keyed window operator does not use the StateBackend for state, i.e. cannot use RocksDB. Also, with upcoming changes (Merging/Session windows, aligned windows) this will only increase the maintenance burden.
Also, the fast AlignedProcessingTimeWindows operators also only support the Parallel/Keyed case.
Would anyone be opposed to me removing the non-keyed window operator?
Cheers,
Aljoscha
Re: [DISCUSS] Removing Non-Keyed WindowOperator
Posted by Ufuk Celebi <uc...@apache.org>.
On Tue, Mar 15, 2016 at 11:06 AM, Ufuk Celebi <uc...@apache.org> wrote:
> Are there any performance differences to be expected?
To clarify: I don't think that it would be a blocker expect for if it
had a yuge impact.
– Ufuk
Re: [DISCUSS] Removing Non-Keyed WindowOperator
Posted by Ufuk Celebi <uc...@apache.org>.
I'm not too familiar with this part of Flink, but your arguments
certainly make sense.
Are there any performance differences to be expected?
– Ufuk
On Tue, Mar 15, 2016 at 11:02 AM, Aljoscha Krettek <al...@apache.org> wrote:
> Hi folks,
> I opened a Jira Issue: https://issues.apache.org/jira/browse/FLINK-3614
>
> This is the text of the issue:
> I propose to remove the special Non-Keyed Window Operator and implement non-parallel windows by using the standard WindowOperator with a dummy KeySelector.
>
> Maintaining everything for two WindowOperators is a huge burden. The implementation is completely separate by now. For example, the Non-Keyed window operator does not use the StateBackend for state, i.e. cannot use RocksDB. Also, with upcoming changes (Merging/Session windows, aligned windows) this will only increase the maintenance burden.
>
> Also, the fast AlignedProcessingTimeWindows operators also only support the Parallel/Keyed case.
>
> Would anyone be opposed to me removing the non-keyed window operator?
>
> Cheers,
> Aljoscha