Re: [DISCUSS] FLIP-328: Allow source operators to determine isProcessingBacklog based on watermark lag
Hi, Xuannan. Thanks for preparing the FLIP. After this FLIP, we will have two ways to report isProcessingBacklog: 1. >From the source; 2. Judged by the watermark lag. What is the priority between them? For example, what is the status isProcessingBacklog when the source report `isProcessingBacklog=false` and the watermark lag exceeds the threshold? Best, Hang Xuannan Su 于2023年8月30日周三 10:06写道： > Hi Jing, > > Thank you for the suggestion. > > The definition of watermark lag is the same as the watermarkLag metric in > FLIP-33[1]. More specifically, the watermark lag calculation is computed at > the time when a watermark is emitted downstream in the following way: > watermarkLag = CurrentTime - Watermark. I have added this description to > the FLIP. > > I hope this addresses your concern. > > Best, > Xuannan > > [1] > https://cwiki.apache.org/confluence/display/FLINK/FLIP-33%3A+Standardize+Connector+Metrics > > > > On Aug 28, 2023, at 01:04, Jing Ge wrote: > > > > Hi Xuannan, > > > > Thanks for the proposal. +1 for me. > > > > There is one tiny thing that I am not sure if I understand it correctly. > > Since there will be many different WatermarkStrategies and different > > WatermarkGenerators. Could you please update the FLIP and add the > > description of how the watermark lag is calculated exactly? E.g. > Watermark > > lag = A - B with A is the timestamp of the watermark emitted to the > > downstream and B is(this is the part I am not really sure after > reading > > the FLIP). > > > > Best regards, > > Jing > > > > > > On Mon, Aug 21, 2023 at 9:03 AM Xuannan Su > wrote: > > > >> Hi Jark, > >> > >> Thanks for the comments. > >> > >> I agree that the current solution cannot support jobs that cannot define > >> watermarks. However, after considering the pending-record-based > solution, I > >> believe the current solution is superior for the target use case as it > is > >> more intuitive for users. The backlog status gives users the ability to > >> balance between throughput and latency. Making this trade-off decision > >> based on the watermark lag is more intuitive from the user's > perspective. > >> For instance, a user can decide that if the job lags behind the current > >> time by more than 1 hour, the result is not usable. In that case, we can > >> optimize for throughput when the data lags behind by more than an hour. > >> With the pending-record-based solution, it's challenging for users to > >> determine when to optimize for throughput and when to prioritize > latency. > >> > >> Regarding the limitations of the watermark-based solution: > >> > >> 1. The current solution can support jobs with sources that have event > >> time. Users can always define a watermark at the source operator, even > if > >> it's not used by downstream operators, such as streaming join and > unbounded > >> aggregate. > >> > >> 2.I don't believe it's accurate to say that the watermark lag will keep > >> increasing if no data is generated in Kafka. The watermark lag and > backlog > >> status are determined at the moment when the watermark is emitted to the > >> downstream operator. If no data is emitted from the source, the > watermark > >> lag and backlog status will not be updated. If the WatermarkStrategy > with > >> idleness is used, the source becomes non-backlog when it becomes idle. > >> > >> 3. I think watermark lag is more intuitive to determine if a job is > >> processing backlog data. Even when using pending records, it faces a > >> similar issue. For example, if the source has 1K pending records, those > >> records can span from 1 day to 1 hour to 1 second. If the records span > 1 > >> day, it's probably best to optimize for throughput. If they span 1 > hour, it > >> depends on the business logic. If they span 1 second, optimizing for > >> latency is likely the better choice. > >> > >> In summary, I believe the watermark-based solution is a superior choice > >> for the target use case where watermark/event time can be defined. > >> Additionally, I haven't come across a scenario that requires low-latency > >> processing and reads from a source that cannot define watermarks. If we > >> encounter such a use case, we can create another FLIP to address those > >> needs in the future. What do you think? > >> > >> > >> Best, > >> Xuannan > >> > >> > >> > >>> On Aug 20, 2023, at 23:27, Jark Wu >> imj...@gmail.com>> wrote: > >>> > >>> Hi Xuannan, > >>> > >>> Thanks for opening this discussion. > >>> > >>> This current proposal may work in the mentioned watermark cases. > >>> However, it seems this is not a general solution for sources to > determine > >>> "isProcessingBacklog". > >>> From my point of view, there are 3 limitations of the current proposal: > >>> 1. It doesn't cover jobs that don't have watermark/event-time defined, > >>> for example streaming join and unbounded aggregate. We may still need > to > >>> figure out solutions for them. > >>> 2. Watermark lag can not be trusted, because it increases unlim
[jira] [Created] (FLINK-32995) TPC-DS end-to-end test fails with chmod: cannot access '../target/generator/dsdgen_linux':
Sergey Nuyanzin created FLINK-32995:
---
Summary: TPC-DS end-to-end test fails with chmod: cannot access
'../target/generator/dsdgen_linux':
Key: FLINK-32995
URL: https://issues.apache.org/jira/browse/FLINK-32995
Project: Flink
Issue Type: Bug
Affects Versions: 1.19.0
Reporter: Sergey Nuyanzin
This build
https://dev.azure.com/apache-flink/apache-flink/_build/results?buildId=52773&view=logs&j=af184cdd-c6d8-5084-0b69-7e9c67b35f7a&t=0f3adb59-eefa-51c6-2858-3654d9e0749d&l=5504
fails as
{noformat}
Aug 29 10:03:20 [INFO] 10:03:20 Generating TPC-DS qualification data, this need
several minutes, please wait...
chmod: cannot access '../target/generator/dsdgen_linux': No such file or
directory
Aug 29 10:03:20 [FAIL] Test script contains errors.
Aug 29 10:03:20 Checking for errors...
{noformat}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
[jira] [Created] (FLINK-32996) CheckpointAfterAllTasksFinishedITCase.testFailoverAfterSomeTasksFinished fails on AZP
Sergey Nuyanzin created FLINK-32996:
---
Summary:
CheckpointAfterAllTasksFinishedITCase.testFailoverAfterSomeTasksFinished fails
on AZP
Key: FLINK-32996
URL: https://issues.apache.org/jira/browse/FLINK-32996
Project: Flink
Issue Type: Bug
Components: Runtime / Checkpointing
Affects Versions: 1.18.0
Reporter: Sergey Nuyanzin
This build
https://dev.azure.com/apache-flink/apache-flink/_build/results?buildId=52810&view=logs&j=8fd9202e-fd17-5b26-353c-ac1ff76c8f28&t=ea7cf968-e585-52cb-e0fc-f48de023a7ca&l=8366
fails as
{noformat}
Aug 30 02:20:32 02:20:32.726 [ERROR] Failures:
Aug 30 02:20:32 02:20:32.726 [ERROR]
CheckpointAfterAllTasksFinishedITCase.testFailoverAfterSomeTasksFinished:162
Aug 30 02:20:32 expected: 20
Aug 30 02:20:32 but was: 40
Aug 30 02:20:32 02:20:32.726 [INFO]
{noformat}
it is very likely it is related to FLINK-32907
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
[jira] [Created] (FLINK-32997) [JUnit5 Migration] Module: flink-table-planner (StreamingTestBase)
Jiabao Sun created FLINK-32997: -- Summary: [JUnit5 Migration] Module: flink-table-planner (StreamingTestBase) Key: FLINK-32997 URL: https://issues.apache.org/jira/browse/FLINK-32997 Project: Flink Issue Type: Sub-task Components: Tests Affects Versions: 1.18.0 Reporter: Jiabao Sun Fix For: 1.19.0 JUnit5 Migration Module: flink-table-planner (StreamingTestBase) -- This message was sent by Atlassian Jira (v8.20.10#820010)
[jira] [Created] (FLINK-32998) if function result not correct
zhou created FLINK-32998: Summary: if function result not correct Key: FLINK-32998 URL: https://issues.apache.org/jira/browse/FLINK-32998 Project: Flink Issue Type: Bug Components: API / Core Affects Versions: 1.15.4 Reporter: zhou Attachments: image-2023-08-30-18-29-16-277.png, image-2023-08-30-18-30-05-568.png !image-2023-08-30-18-29-16-277.png! !image-2023-08-30-18-30-05-568.png! if function result not correct,not result in origin field value, cut off the filed(word) value -- This message was sent by Atlassian Jira (v8.20.10#820010)
[jira] [Created] (FLINK-32999) Remove HBase connector from master branch
Sergey Nuyanzin created FLINK-32999: --- Summary: Remove HBase connector from master branch Key: FLINK-32999 URL: https://issues.apache.org/jira/browse/FLINK-32999 Project: Flink Issue Type: Sub-task Components: Connectors / HBase Reporter: Sergey Nuyanzin The connector was externalized at FLINK-30061 Once it is released it would make sense to remove it from master branch -- This message was sent by Atlassian Jira (v8.20.10#820010)
[jira] [Created] (FLINK-33000) SqlGatewayServiceITCase should utilize TestExecutorExtension instead of using a ThreadFactory
Matthias Pohl created FLINK-33000:
-
Summary: SqlGatewayServiceITCase should utilize
TestExecutorExtension instead of using a ThreadFactory
Key: FLINK-33000
URL: https://issues.apache.org/jira/browse/FLINK-33000
Project: Flink
Issue Type: Bug
Components: Table SQL / Gateway, Tests
Affects Versions: 1.17.1, 1.16.2, 1.18.0, 1.19.0
Reporter: Matthias Pohl
{{SqlGatewayServiceITCase}} uses a {{ExecutorThreadFactory}} for its
asynchronous operations. Instead, one should use {{TestExecutorExtension}} to
ensure proper cleanup of threads.
We might also want to remove the {{AbstractTestBase}} parent class because that
uses JUnit4 whereas {{SqlGatewayServiceITCase}} is already based on JUnit5
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
[jira] [Created] (FLINK-33001) KafkaSource in batch mode failing with exception if topic partition is empty
Abdul created FLINK-33001:
-
Summary: KafkaSource in batch mode failing with exception if topic
partition is empty
Key: FLINK-33001
URL: https://issues.apache.org/jira/browse/FLINK-33001
Project: Flink
Issue Type: Bug
Components: Connectors / Kafka
Affects Versions: 1.17.1, 1.14.6, 1.12.7
Environment: The only workaround that works fine right now is to
change the DEBUG level to INFO for logging.
{code:java}
logger.KafkaPartitionSplitReader.name =
org.apache.flink.connector.kafka.source.reader.KafkaPartitionSplitReader
logger.KafkaPartitionSplitReader.level = INFO{code}
It is strange that changing this doesn't cause the above exception.
Reporter: Abdul
If the Kafka topic is empty in Batch mode, there is an exception while
processing it. This bug was supposedly fixed but unfortunately, the exception
still occurs. The original bug was reported as this
https://issues.apache.org/jira/browse/FLINK-27041
We tried to backport it but it still doesn't work.
* The problem will occur in case of DEBUG level of logger for class
org.apache.flink.connector.kafka.source.reader.KafkaPartitionSplitReader
* The same problems will occur in other versions of Flink, at least in the
1.15 release branch and tag release-1.15.4
* Same problem also occur in Flink 1.7.1 and 1.14
The minimal code to produce this is
final StreamExecutionEnvironment env =
StreamExecutionEnvironment.getExecutionEnvironment();
env.setRuntimeMode(RuntimeExecutionMode.BATCH);
KafkaSource kafkaSource = KafkaSource
.builder()
.setBootstrapServers("localhost:9092")
.setTopics("test_topic")
.setValueOnlyDeserializer(new
SimpleStringSchema())
.setBounded(OffsetsInitializer.latest())
.build();
DataStream stream = env.fromSource(
kafkaSource,
WatermarkStrategy.noWatermarks(),
"Kafka Source" );
stream.print();
env.execute("Flink KafkaSource test job");
This produces exception:
{code:java}
Caused by: java.lang.RuntimeException: One or more fetchers have encountered
exception at
org.apache.flink.connector.base.source.reader.fetcher.SplitFetcherManager.checkErrors(SplitFetcherManager.java:199)
at
org.apache.flink.connector.base.source.reader.SourceReaderBase.getNextFetch(SourceReaderBase.java:154)
at
org.apache.flink.connector.base.source.reader.SourceReaderBase.pollNext(SourceReaderBase.java:116)
at
org.apache.flink.streaming.api.operators.SourceOperator.emitNext(SourceOperator.java:275)
at
org.apache.flink.streaming.runtime.io.StreamTaskSourceInput.emitNext(StreamTaskSourceInput.java:67)
at
org.apache.flink.streaming.runtime.io.StreamOneInputProcessor.processInput(StreamOneInputProcessor.java:65)
at
org.apache.flink.streaming.runtime.tasks.StreamTask.processInput(StreamTask.java:398)
at
org.apache.flink.streaming.runtime.tasks.mailbox.MailboxProcessor.runMailboxLoop(MailboxProcessor.java:191)
at
org.apache.flink.streaming.runtime.tasks.StreamTask.runMailboxLoop(StreamTask.java:619)
at
org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:583)
at org.apache.flink.runtime.taskmanager.Task.doRun(Task.java:758)
at org.apache.flink.runtime.taskmanager.Task.run(Task.java:573) at
java.lang.Thread.run(Thread.java:748)Caused by: java.lang.RuntimeException:
SplitFetcher thread 0 received unexpected exception while polling the records
at
org.apache.flink.connector.base.source.reader.fetcher.SplitFetcher.runOnce(SplitFetcher.java:146)
at
org.apache.flink.connector.base.source.reader.fetcher.SplitFetcher.run(SplitFetcher.java:101)
at
java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
at java.util.concurrent.FutureTask.run(FutureTask.java:266) at
java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
at
java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
... 1 moreCaused by: java.lang.IllegalStateException: You can only check
the position for partitions assigned to this consumer. at
org.apache.kafka.clients.consumer.KafkaConsumer.position(KafkaConsumer.java:1737)
at
org.apache.kafka.clients.consumer.KafkaConsumer.position(KafkaConsumer.java:1704)
at
org.apache.flink.connector.kafka.source.reader.KafkaPartitionSplitReader.maybeLogSplitChangesHandlingResult(KafkaPartitionSplitReader.java:375)
at
org.apache.flink.connector.kafka.source.reader.K
Re: [DISCUSS] FLIP-328: Allow source operators to determine isProcessingBacklog based on watermark lag
Hi Xuannan, Thanks for the clarification. That is the part where I am trying to understand your thoughts. I have some follow-up questions: 1. It depends strongly on the watermarkStrategy and how customized watermark generation looks like. It mixes business logic with technical implementation and technical data processing mode. The value of the watermark lag threshold must be set very carefully. If the value is too small. any time, when the watermark generation logic is changed(business logic changes lead to the threshold getting exceeded), the same job might be running surprisingly in backlog processing mode, i.e. a butterfly effect. A comprehensive documentation is required to avoid any confusion for the users. 2. Like Jark already mentioned, use cases that do not have watermarks, like pure processing-time based stream processing[1] are not covered. It is more or less a trade-off solution that does not support such use cases and appropriate documentation is required. Forcing them to explicitly generate watermarks that are never needed just because of this does not sound like a proper solution. 3. If I am not mistaken, it only works for use cases where event times are very close to the processing times, because the wall clock is used to calculate the watermark lag and the watermark is generated based on the event time. Best regards, Jing [1] https://github.com/apache/flink/blob/2c50b4e956305426f478b726d4de4a640a16b810/flink-core/src/main/java/org/apache/flink/api/common/eventtime/WatermarkStrategy.java#L236 On Wed, Aug 30, 2023 at 4:06 AM Xuannan Su wrote: > Hi Jing, > > Thank you for the suggestion. > > The definition of watermark lag is the same as the watermarkLag metric in > FLIP-33[1]. More specifically, the watermark lag calculation is computed at > the time when a watermark is emitted downstream in the following way: > watermarkLag = CurrentTime - Watermark. I have added this description to > the FLIP. > > I hope this addresses your concern. > > Best, > Xuannan > > [1] > https://cwiki.apache.org/confluence/display/FLINK/FLIP-33%3A+Standardize+Connector+Metrics > > > > On Aug 28, 2023, at 01:04, Jing Ge wrote: > > > > Hi Xuannan, > > > > Thanks for the proposal. +1 for me. > > > > There is one tiny thing that I am not sure if I understand it correctly. > > Since there will be many different WatermarkStrategies and different > > WatermarkGenerators. Could you please update the FLIP and add the > > description of how the watermark lag is calculated exactly? E.g. > Watermark > > lag = A - B with A is the timestamp of the watermark emitted to the > > downstream and B is(this is the part I am not really sure after > reading > > the FLIP). > > > > Best regards, > > Jing > > > > > > On Mon, Aug 21, 2023 at 9:03 AM Xuannan Su > wrote: > > > >> Hi Jark, > >> > >> Thanks for the comments. > >> > >> I agree that the current solution cannot support jobs that cannot define > >> watermarks. However, after considering the pending-record-based > solution, I > >> believe the current solution is superior for the target use case as it > is > >> more intuitive for users. The backlog status gives users the ability to > >> balance between throughput and latency. Making this trade-off decision > >> based on the watermark lag is more intuitive from the user's > perspective. > >> For instance, a user can decide that if the job lags behind the current > >> time by more than 1 hour, the result is not usable. In that case, we can > >> optimize for throughput when the data lags behind by more than an hour. > >> With the pending-record-based solution, it's challenging for users to > >> determine when to optimize for throughput and when to prioritize > latency. > >> > >> Regarding the limitations of the watermark-based solution: > >> > >> 1. The current solution can support jobs with sources that have event > >> time. Users can always define a watermark at the source operator, even > if > >> it's not used by downstream operators, such as streaming join and > unbounded > >> aggregate. > >> > >> 2.I don't believe it's accurate to say that the watermark lag will keep > >> increasing if no data is generated in Kafka. The watermark lag and > backlog > >> status are determined at the moment when the watermark is emitted to the > >> downstream operator. If no data is emitted from the source, the > watermark > >> lag and backlog status will not be updated. If the WatermarkStrategy > with > >> idleness is used, the source becomes non-backlog when it becomes idle. > >> > >> 3. I think watermark lag is more intuitive to determine if a job is > >> processing backlog data. Even when using pending records, it faces a > >> similar issue. For example, if the source has 1K pending records, those > >> records can span from 1 day to 1 hour to 1 second. If the records span > 1 > >> day, it's probably best to optimize for throughput. If they span 1 > hour, it > >> depends on the business logic. If they span 1 second, optimizing for > >>
[jira] [Created] (FLINK-33002) Bump snappy-java from 1.1.4 to 1.1.10.1
Martijn Visser created FLINK-33002: -- Summary: Bump snappy-java from 1.1.4 to 1.1.10.1 Key: FLINK-33002 URL: https://issues.apache.org/jira/browse/FLINK-33002 Project: Flink Issue Type: Technical Debt Components: Stateful Functions Reporter: Martijn Visser Assignee: Martijn Visser -- This message was sent by Atlassian Jira (v8.20.10#820010)
[jira] [Created] (FLINK-33003) Flink ML add isolationForest algorithm
zhaozijun created FLINK-33003: - Summary: Flink ML add isolationForest algorithm Key: FLINK-33003 URL: https://issues.apache.org/jira/browse/FLINK-33003 Project: Flink Issue Type: New Feature Components: Library / Machine Learning Reporter: zhaozijun Attachments: IsolationForest.zip I want to use flink solve some problems related to anomaly detection, but currently flink ml lacks algorithms related to anomaly detection, so I want to add the isolation forest algorithm to library/flink ml. During the implementation process, when IterationBody is used, I try to understand the implementation of the Kmeans algorithm, and use iterative behavior to calculate the center point of the isolation forest algorithm, but in the test, I found that when the parallelism > 1, the number of iterations > 1, and there will be sometimes succeed sometimes fail (fail to find the broadcast variable). Please teachers help me to review and point out my problem. Thank you 🙏 -- This message was sent by Atlassian Jira (v8.20.10#820010)
[jira] [Created] (FLINK-33004) Decoupling topology and network memory to support complex job topologies
dalongliu created FLINK-33004: - Summary: Decoupling topology and network memory to support complex job topologies Key: FLINK-33004 URL: https://issues.apache.org/jira/browse/FLINK-33004 Project: Flink Issue Type: Improvement Components: Runtime / Network Affects Versions: 1.18.0, 1.19.0 Reporter: dalongliu Currently, the default value of taskmanager.memory.network.fraction option in Flink is 0.1, and after the topology of the job is complex enough, it will run with an insufficient network buffer. We currently encountered this issue when running TPC-DS test set q9, and bypassed it by adjusting taskmanager.memory.network.fraction to 0.2. Theoretically, we should have network memory decoupled from the job topology so that arbitrarily complex jobs can be supported. -- This message was sent by Atlassian Jira (v8.20.10#820010)
Re: [DISCUSS] FLIP-328: Allow source operators to determine isProcessingBacklog based on watermark lag
Hi Jing, Thanks for the reply. 1. You are absolutely right that the watermark lag threshold must be carefully set with a thorough understanding of watermark generation. It is crucial for users to take into account the WatermarkStrategy when setting the watermark lag threshold. 2. Regarding pure processing-time based stream processing jobs, alternative strategies will be implemented to determine whether the job is processing backlog data. I have outlined two possible strategies below: - Based on the source operator's state. For example, when MySQL CDC source is reading snapshot, it can claim isBacklog=true. - Based on metrics. For example, when busyTimeMsPerSecond (or backPressuredTimeMsPerSecond) > user_specified_threshold, then isBacklog=true. As of the strategies proposed in this FLIP, it rely on generated watermarks. Therefore, if a user intends for the job to detect backlog status based on watermark, it is necessary to generate the watermark. 3. I'm afraid I'm not fully grasping your question. From my understanding, it should work in both cases. When event times are close to the processing time, resulting in watermarks close to the processing time, the job is not processing backlog data. On the other hand, when event times are far from processing time, causing watermarks to also be distant, if the lag surpasses the defined threshold, the job is considered processing backlog data. Best, Xuannan > On Aug 31, 2023, at 02:56, Jing Ge wrote: > > Hi Xuannan, > > Thanks for the clarification. That is the part where I am trying to > understand your thoughts. I have some follow-up questions: > > 1. It depends strongly on the watermarkStrategy and how customized > watermark generation looks like. It mixes business logic with technical > implementation and technical data processing mode. The value of the > watermark lag threshold must be set very carefully. If the value is too > small. any time, when the watermark generation logic is changed(business > logic changes lead to the threshold getting exceeded), the same job might > be running surprisingly in backlog processing mode, i.e. a butterfly > effect. A comprehensive documentation is required to avoid any confusion > for the users. > 2. Like Jark already mentioned, use cases that do not have watermarks, > like pure processing-time based stream processing[1] are not covered. It is > more or less a trade-off solution that does not support such use cases and > appropriate documentation is required. Forcing them to explicitly generate > watermarks that are never needed just because of this does not sound like a > proper solution. > 3. If I am not mistaken, it only works for use cases where event times are > very close to the processing times, because the wall clock is used to > calculate the watermark lag and the watermark is generated based on the > event time. > > Best regards, > Jing > > [1] > https://github.com/apache/flink/blob/2c50b4e956305426f478b726d4de4a640a16b810/flink-core/src/main/java/org/apache/flink/api/common/eventtime/WatermarkStrategy.java#L236 > > On Wed, Aug 30, 2023 at 4:06 AM Xuannan Su wrote: > >> Hi Jing, >> >> Thank you for the suggestion. >> >> The definition of watermark lag is the same as the watermarkLag metric in >> FLIP-33[1]. More specifically, the watermark lag calculation is computed at >> the time when a watermark is emitted downstream in the following way: >> watermarkLag = CurrentTime - Watermark. I have added this description to >> the FLIP. >> >> I hope this addresses your concern. >> >> Best, >> Xuannan >> >> [1] >> https://cwiki.apache.org/confluence/display/FLINK/FLIP-33%3A+Standardize+Connector+Metrics >> >> >>> On Aug 28, 2023, at 01:04, Jing Ge wrote: >>> >>> Hi Xuannan, >>> >>> Thanks for the proposal. +1 for me. >>> >>> There is one tiny thing that I am not sure if I understand it correctly. >>> Since there will be many different WatermarkStrategies and different >>> WatermarkGenerators. Could you please update the FLIP and add the >>> description of how the watermark lag is calculated exactly? E.g. >> Watermark >>> lag = A - B with A is the timestamp of the watermark emitted to the >>> downstream and B is(this is the part I am not really sure after >> reading >>> the FLIP). >>> >>> Best regards, >>> Jing >>> >>> >>> On Mon, Aug 21, 2023 at 9:03 AM Xuannan Su >> wrote: >>> Hi Jark, Thanks for the comments. I agree that the current solution cannot support jobs that cannot define watermarks. However, after considering the pending-record-based >> solution, I believe the current solution is superior for the target use case as it >> is more intuitive for users. The backlog status gives users the ability to balance between throughput and latency. Making this trade-off decision based on the watermark lag is more intuitive from the user's >> perspective. For instance, a user can decide that if the job lags behind the c
Re: 退订
Please send email to user-unsubscr...@flink.apache.org if you want to unsubscribe the mail from u...@flink.apache.org, and you can refer [1][2] for more details. [1] https://flink.apache.org/zh/community/#%e9%82%ae%e4%bb%b6%e5%88%97%e8%a1%a8 [2] https://flink.apache.org/community.html#mailing-lists Best, Ron 喻凯 于2023年8月30日周三 14:17写道： > >
