[jira] [Commented] (KAFKA-4325) Improve processing of late records for window operations

2017-06-04 Thread Guozhang Wang (JIRA)

[ 
https://issues.apache.org/jira/browse/KAFKA-4325?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16036512#comment-16036512
 ] 

Guozhang Wang commented on KAFKA-4325:
--

I think the timestamp computation needs some further thoughts: 
https://issues.apache.org/jira/browse/KAFKA-3514. So I'd suggest let's not rush 
to fix one-toe-at-a-time, and I will upload a summary / proposal for the 
timestamp computation under KAFKA-3514 for people to discuss and then propose a 
KIP.

> Improve processing of late records for window operations
> 
>
> Key: KAFKA-4325
> URL: https://issues.apache.org/jira/browse/KAFKA-4325
> Project: Kafka
>  Issue Type: Improvement
>  Components: streams
>Reporter: Matthias J. Sax
>Priority: Minor
>
> Windows are kept until their retention time passed. If a late arriving record 
> is processed that is older than any window kept, a new window is created 
> containing this single late arriving record, the aggregation is computed and 
> the window is immediately discarded afterward (as it is older than retention 
> time).
> This behavior might case problems for downstream application as the original 
> window aggregate might we overwritten with the late single-record- aggregate 
> value. Thus, we should rather not process the late arriving record for this 
> case.
> However, data loss might not be acceptable for all use cases. In order to 
> enable the use to not lose any data, window operators should allow to 
> register a handler function that is called instead of just dropping the late 
> arriving record.



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


[jira] [Commented] (KAFKA-4325) Improve processing of late records for window operations

2017-06-03 Thread Matthias J. Sax (JIRA)

[ 
https://issues.apache.org/jira/browse/KAFKA-4325?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16036135#comment-16036135
 ] 

Matthias J. Sax commented on KAFKA-4325:


I assume yes. \cc [~guozhang]

> Improve processing of late records for window operations
> 
>
> Key: KAFKA-4325
> URL: https://issues.apache.org/jira/browse/KAFKA-4325
> Project: Kafka
>  Issue Type: Improvement
>  Components: streams
>Reporter: Matthias J. Sax
>Priority: Minor
>
> Windows are kept until their retention time passed. If a late arriving record 
> is processed that is older than any window kept, a new window is created 
> containing this single late arriving record, the aggregation is computed and 
> the window is immediately discarded afterward (as it is older than retention 
> time).
> This behavior might case problems for downstream application as the original 
> window aggregate might we overwritten with the late single-record- aggregate 
> value. Thus, we should rather not process the late arriving record for this 
> case.
> However, data loss might not be acceptable for all use cases. In order to 
> enable the use to not lose any data, window operators should allow to 
> register a handler function that is called instead of just dropping the late 
> arriving record.



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


[jira] [Commented] (KAFKA-4325) Improve processing of late records for window operations

2017-06-02 Thread Jeyhun Karimov (JIRA)

[ 
https://issues.apache.org/jira/browse/KAFKA-4325?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16035688#comment-16035688
 ] 

Jeyhun Karimov commented on KAFKA-4325:
---

[~mjsax], would this jira requre KIP? 

> Improve processing of late records for window operations
> 
>
> Key: KAFKA-4325
> URL: https://issues.apache.org/jira/browse/KAFKA-4325
> Project: Kafka
>  Issue Type: Improvement
>  Components: streams
>Reporter: Matthias J. Sax
>Priority: Minor
>
> Windows are kept until their retention time passed. If a late arriving record 
> is processed that is older than any window kept, a new window is created 
> containing this single late arriving record, the aggregation is computed and 
> the window is immediately discarded afterward (as it is older than retention 
> time).
> This behavior might case problems for downstream application as the original 
> window aggregate might we overwritten with the late single-record- aggregate 
> value. Thus, we should rather not process the late arriving record for this 
> case.
> However, data loss might not be acceptable for all use cases. In order to 
> enable the use to not lose any data, window operators should allow to 
> register a handler function that is called instead of just dropping the late 
> arriving record.



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