Kafka SQL Connector: dropping events if more partitions then source tasks
Locked
Kafka SQL Connector: dropping events if more partitions then source tasks
 
|
Hi, i have a question regarding FlinkSQL connector for Kafka. I have 3 Kafka partitions and 1 Kafka SQL source connector (Parallelism 1). The data within the Kafka parttitons are sorted based on a event-time field, which
is also my event-time in Flink. My Watermark is generated with a delay of 12 hours WATERMARK
FOR
eventtime
as
eventtime -
INTERVAL
'12'
HOUR But the problem is that I see dropping events due arriving late in Prometheus. But with parallelism of 3 there are no drops.
Do I always have to have as much source-tasks as I have Kafka partitions? Best, Jan |
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
By using the DataStream API with the same business logic I’m getting no dropped events. Von: Jan Oelschlegel <[hidden email]>
Hi, i have a question regarding FlinkSQL connector for Kafka. I have 3 Kafka partitions and 1 Kafka SQL source connector (Parallelism 1). The data within the Kafka parttitons are sorted based on a event-time field, which
is also my event-time in Flink. My Watermark is generated with a delay of 12 hours WATERMARK
FOR
eventtime
as
eventtime -
INTERVAL
'12'
HOUR But the problem is that I see dropping events due arriving late in Prometheus. But with parallelism of 3 there are no drops.
Do I always have to have as much source-tasks as I have Kafka partitions? Best, Jan HINWEIS: Dies ist eine vertrauliche Nachricht und nur für den Adressaten bestimmt. Es ist nicht erlaubt, diese Nachricht zu kopieren oder Dritten zugänglich zu machen. Sollten Sie diese Nachricht irrtümlich
erhalten haben, bitte ich um Ihre Mitteilung per E-Mail oder unter der oben angegebenen Telefonnummer.
|
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
If i increase the watermark, the dropped events getting lower. But why is the DataStream API Job still running with 12 hours watermark delay? Best, Jan Von: Jan Oelschlegel <[hidden email]>
By using the DataStream API with the same business logic I’m getting no dropped events. Von: Jan Oelschlegel <[hidden email]>
Hi, i have a question regarding FlinkSQL connector for Kafka. I have 3 Kafka partitions and 1 Kafka SQL source connector (Parallelism 1). The data within the Kafka parttitons are sorted based on a event-time field, which
is also my event-time in Flink. My Watermark is generated with a delay of 12 hours WATERMARK
FOR
eventtime
as
eventtime -
INTERVAL
'12'
HOUR But the problem is that I see dropping events due arriving late in Prometheus. But with parallelism of 3 there are no drops.
Do I always have to have as much source-tasks as I have Kafka partitions? Best, Jan HINWEIS: Dies ist eine vertrauliche Nachricht und nur für den Adressaten bestimmt. Es ist nicht erlaubt, diese Nachricht zu kopieren oder Dritten zugänglich zu machen. Sollten Sie diese Nachricht irrtümlich
erhalten haben, bitte ich um Ihre Mitteilung per E-Mail oder unter der oben angegebenen Telefonnummer.
|
Re: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Jan, Are you running on historic data? Then your partitions might drift apart quickly. However, I still suspect that this is a bug (Watermark should only be from the slowest partition). I'm pulling in Timo who should know more. On Fri, Feb 19, 2021 at 10:50 AM Jan Oelschlegel <[hidden email]> wrote:
|
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Arvid, thanks for bringing back this topic.
Yes, I’m running on historic data, but as you mentioned that should not be the problem, even there is a event-time skew between partitions. But maybe this issue with the missing watermark pushdown per partition is the important fact: https://issues.apache.org/jira/browse/FLINK-20041 Best, Jan Von: Arvid Heise <[hidden email]> Hi Jan, Are you running on historic data? Then your partitions might drift apart quickly.
However, I still suspect that this is a bug (Watermark should only be from the slowest partition). I'm pulling in Timo who should know more. On Fri, Feb 19, 2021 at 10:50 AM Jan Oelschlegel <[hidden email]> wrote:
|
Re: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Jan, What you are observing is correct for the current implementation. Current watermark generation is based on subtask instead of partition. Hence if there are more than on partition in the same subtask, it's very easy to see more data dropped. AFAIK, FLIP-27 could solve this problem, however the Kafka Connector has not been migrated to FLIP-27 for now. Jan Oelschlegel <[hidden email]> 于2021年2月24日周三 下午10:07写道：
Best, Benchao Li |
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Benchao, i’m observing this behaviour only for the SQL API. With the Datastream API i can take more or less source-tasks then kafka partition count. And FLIP-27 seems to belong to the Datastream
API. The problem is only on the SQL site. Best, Jan Von: Benchao Li <[hidden email]> Hi Jan, What you are observing is correct for the current implementation. Current watermark generation is based on subtask instead of partition. Hence if there are more than on partition in the same subtask, it's very easy to see more data dropped. AFAIK, FLIP-27 could solve this problem, however the Kafka Connector has not been migrated to FLIP-27 for now. Jan Oelschlegel <[hidden email]>
于2021年2月24日周三
下午10:07写道：
--
|
Re: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi, Jan. Could you tell us which Flink version you use? As far as I know, the kafka sql connector has implemented `SupportWatermarkPushDown` in Flink-1.12. The `SupportWatermarkPushDown` pushes the watermark generator into the source and emits the minimum watermark among all the partitions. For more details, you can refer to the doc for more details[1]. If you use the version before FLINK-1.12, I think the best approach to solve this problem is to increase source tasks. Best, Shengkai Jan Oelschlegel <[hidden email]> 于2021年2月25日周四 下午4:24写道：
|
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Shengkai, i’m using Flink 1.11.2. The problem is if I use a parallelism higher than my kafka partition count, the watermarks are not increasing and so the windows are never ggot fired.
I suspect that then a source task is not marked as idle and thus the watermark is not increased. In any case I have observed how with a larger number of source tasks no results are produced. Best, Jan Von: Shengkai Fang <[hidden email]> Hi, Jan. Could you tell us which Flink version you use? As far as I know, the kafka sql connector has implemented `SupportWatermarkPushDown` in Flink-1.12. The `SupportWatermarkPushDown` pushes the watermark generator into the source and emits the
minimum watermark among all the partitions. For more details, you can refer to the doc for more details[1]. If you use the version before FLINK-1.12, I think the best approach to solve this problem is to increase source tasks. Best, Shengkai Jan Oelschlegel <[hidden email]>
于2021年2月25日周四
下午4:24写道：
|
Re: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Jan.
Thanks for your reply. Do you set the option `table.exec.source.idle-timeout` and `pipeline.auto-watermark-interval` ? If the `pipeline.auto-watermark-interval ` is zero, it will not trigger the detection of the idle source. Best, Shengkai Jan Oelschlegel <[hidden email]> 于2021年2月26日周五 下午11:09写道：
|
AW: Kafka SQL Connector: dropping events if more partitions then source tasks
|
|
Hi Shengkai, thanks for this hint. I solves the issue having more consumer tasks than kafka partitions. But the case with dropping events while having less consumer tasks than kafka partitions is still there. I think it will be okay in version 1.12 [1] [1]
https://issues.apache.org/jira/browse/FLINK-20041 Best, Jan Von: Shengkai Fang <[hidden email]> Hi Jan. Thanks for your reply. Do you set the option `table.exec.source.idle-timeout` and `pipeline.auto-watermark-interval` ? If the `pipeline.auto-watermark-interval
` is zero, it will not trigger the detection of the idle source. Best, Shengkai Jan Oelschlegel <[hidden email]>
于2021年2月26日周五
下午11:09写道：
|
| Free forum by Nabble | Edit this page |