flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Robert Metzger (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-5719) Let LatencyMarkers completely bypass operators / chains
Date Wed, 08 Feb 2017 10:26:42 GMT

    [ https://issues.apache.org/jira/browse/FLINK-5719?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15857797#comment-15857797

Robert Metzger commented on FLINK-5719:

Yes, changing from PublicEvolving to Internal is good!

I don't care so much about changing this, if you think its necessary.
One thing to consider however is the following: If we ever want to add a functionality that
allows operators to "add" latency, we need to expose the latency markers to the operator.
One way to implement this feature would be adding an additional field to the marker "addedLatency",
where operators can add the milliseconds delay the operator is adding.

> Let LatencyMarkers completely bypass operators / chains
> -------------------------------------------------------
>                 Key: FLINK-5719
>                 URL: https://issues.apache.org/jira/browse/FLINK-5719
>             Project: Flink
>          Issue Type: Improvement
>          Components: DataStream API, Streaming
>            Reporter: Tzu-Li (Gordon) Tai
> Currently, {{LatencyMarker}} s are forwarded through operators via the operator interfaces
and methods, i.e. {{AbstractStreamOperator#processLatencyMarker()}},  {{Output#emitLatencyMarker()}},
{{OneInputStreamOperator#processLatencyMarker()}} etc.
> The main issue with this is that {{LatencyMarker}} s are essentially internal elements,
and the implementation on how to handle them should be final. Exposing them through operator
interfaces will allow the user to override the implementation, and also makes the user interface
for operators over-complicated.
> [~aljoscha] suggested to bypass such internal stream elements from the operator to keep
the operator interfaces minimal, in FLINK-5017.
> We propose a similar approach here for {{LatencyMarker}} as well. Since the chaining
output calls contribute very little to the measured latency and can be ignored, instead of
passing it through operator chains, latency markers can simply be passed downstream once tasks
receive them.

This message was sent by Atlassian JIRA

View raw message