Hi Flink Community, I am doing some research work on Flink Datastream and Table API and I meet two major problems. I am using Flink 1.11.2, scala version 2.11, java 8. My use case looks like this. I plan to write a data processing pipeline with two stages. My goal is to construct a business object containing information from several Kafka streams with a primary key and emit the complete business object if such primary key doesn't appear in the pipeline for 10 seconds. In the first stage, I first consume three Kafka streams and transform it to Flink Datastream using a deserialization schema containing some type and date format transformation, and then I register these data streams as Table and do a full outer join one by one using Table API. I also add query configuration for this to avoid excessive state. The primary key is also the join key. In the second stage, I transform the joined table to a retracted stream and put it into KeyedProcessFunction to generate the business object if the business object's primary key is inactive for 10 second. Is this way of handling the data the suggested approach? (I understand I can directly consume kafka data in Table API. I haven't tried that yet, maybe that's better?) Any suggestion is welcomed. During implementing this, I meet two major problems and several smaller questions under each problem. 1. Some type cast behavior of retracted streams I can't explain. (1) In the initial stage, I registered some field as java.sql.Date or java.sql.timestamp following the examples at (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html#data-type-extraction) . After join and transform to retracted stream, it becomes java.time.LocalDate and java.time.LocalDateTime instead. For example, when first ingesting the Kafka streams, I registerd a attribute in java.sql.Timestamp type. @JsonAlias("ATTRIBUTE1") When I tried to cast the type information back after the retracted stream, the code gives me error information below. java.lang.ClassCastException: java.time.LocalDateTime cannot be cast to java.sql.Timestamp Maybe I should use toAppendStream instead since append stream could register type information, but toRetractedStream can't do that? (https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html#convert-a-table-into-a-datastream-or-dataset) My work around is to cast it to LocalDateTime first and extract the epoch time, this doesn't seem to be a final solution. (2) During timestamp conversion, the Flink to retracted stream seems to lost the AM/PM information in the stream and causing a 12 hour difference if it is PM. I use joda time to do some timestamp conversion in the first deserialization stage, my pattern looks like this. "a" means AM/PM information DateTimeFormatter format3 = DateTimeFormat.forPattern("dd-MMM-yy HH.mm.ss.SSSSSS a").withZone(DateTimeZone.getDefault()); After the retracted stream, the AM/PM information is not preserved. 2. My onTimer method in KeyedProcessFunction can not be triggered when I scheduled a event timer timer. I am using event time in my code. I am new to configure watermarks and I might miss something to configure it correctly. I also tried to register a processing time, it could enter and produce some results. I am trying to follow the example here: https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html#example My onTimer method looks like this and the scheduled event doesn't happen.. In processElement(): context.timerService().registerEventTimeTimer(current.getLastModifiedTime() + 10000); My onTimer function @Override (1) When I assign the timestamp and watermarks outside the process() method chain. The "context.timestamp()" will be null. If I put it inside the chain, it won't be null. Is this the expected behavior? In the null case, the strange thing is that, surprisingly, I can collect the business object immediately without a designed 10 second waiting time... This shouldn't happen, right...? The processing timer also seems to work. The code can enter the on timer method. retractStream.assignTimestampsAndWatermarks(new BoRetractStreamTimestampAssigner()); (This is a deprecated method) (2) For watermarks configuration. I use an field in the retracted stream as the event time. This time is usually 15-20 seconds before current time. In my environment, I have done some settings for streaming env based on information here( https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator). My event doesn't always come, so I think I need to set auto watermark interval to let the event timer on timer works correctly. I have added the code below. env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); 1> Which kind of watermark strategy should I use? General BoundOutofOrderness or Watermark generator? I tried to write a Watermark generator and I just don't how to apply it to the stream correctly. The documentation doesn't explain very clearly. My code looks like below and it doesn't work. assign part: .assignTimestampsAndWatermarks(WatermarkStrategy.forGenerator((WatermarkGeneratorSupplier<Tuple2<Boolean, Row>>) context -> new TableBoundOutofOrdernessGenerator())) watermark generater: I just assign the event time attribute following the example in the doc. (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator) 2> I also tried to use the static method in Water Strategy. The syntax is correct, but I meet the same problem in 2.(1). .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, Row>>forBoundedOutOfOrderness(Duration.ofSeconds(15)) (3) For the retracted datastream, do I need to explicitly attach it to the stream environment? I think it is done by default, right? Just want to confirm it. I do have the env.execute() at the end of the code. I understand this is a lot of questions, thanks a lot for your patience to look through my email! If there is anything unclear, please reach out to me. Thanks! Best regards, Fuyao Li |
Hi Flink Users and Community, For the first part of the question, the 12 hour time difference is caused by a time extraction bug myself. I can get the time translated correctly now. The type cast problem does have some workarounds to solve it.. My major blocker right now is the onTimer part is not properly triggered. I guess it is caused by failing to configure the correct watermarks & timestamp assigners. Please give me some insights. 1. If I don't chain the assignTimestampsAndWatermarks() method in together with keyedBy().. and process().. method. The context.timestamp() in my processElement() function will be null. Is this some expected behavior? The Flink examples didn't chain it together. (see example here: https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#using-watermark-strategies) 2. If I use registerEventTimeTimer() in processElement(). The onTimer method will not be triggered. However, I can trigger the onTimer method if I simply change it to registerProcessingTimeTimer(). I am using the settings below in the stream env. env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); env.getConfig().setAutoWatermarkInterval(1000L); My code for method the process chain: retractStream .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, Row>>forBoundedOutOfOrderness(Duration.ofSeconds(20)) .withTimestampAssigner((booleanRowTuple2, timestamp) -> { Row rowData = booleanRowTuple2.f1; LocalDateTime headerTime = (LocalDateTime)rowData.getField(3); LocalDateTime linesTime = (LocalDateTime)rowData.getField(7); LocalDateTime latestDBUpdateTime = null; if (headerTime != null && linesTime != null) { latestDBUpdateTime = headerTime.isAfter(linesTime) ? headerTime : linesTime; } else { latestDBUpdateTime = (headerTime != null) ? headerTime : linesTime; } if (latestDBUpdateTime != null) { return latestDBUpdateTime.atZone(ZoneId.of("America/Los_Angeles")).toInstant().toEpochMilli(); } // In the worst case, we use system time instead, which should never be reached. return System.currentTimeMillis(); })) // .assignTimestampsAndWatermarks(new MyWaterStrategy()) // second way to create watermark, doesn't work .keyBy(value -> { // There could be null fields for header invoice_id field String invoice_id_key = (String)value.f1.getField(0); if (invoice_id_key == null) { invoice_id_key = (String)value.f1.getField(4); } return invoice_id_key; }) .process(new TableOutputProcessFunction()) .name("ProcessTableOutput") .uid("ProcessTableOutput") .addSink(businessObjectSink) .name("businessObjectSink") .uid("businessObjectSink") .setParallelism(1); Best regards, Fuyao On Mon, Nov 2, 2020 at 4:53 PM Fuyao Li <[hidden email]> wrote:
|
Hi Fuyao, for your first question about the different behavior depending on whether you chain the methods or not: Keep in mind that you have to save the return value of the assignTimestampsAndWatermarks method call if you don't chain the methods together as it is also shown in [1]. At least the following example from your first message is indicating it: ``` retractStream.assignTimestampsAndWatermarks(new BoRetractStreamTimestampAssigner()); (This is a deprecated method) // instead of: retractStream = retractStream.assignTimestampsAndWatermarks(new BoRetractStreamTimestampAssigner()); retractStream .keyBy(<key selector>) .process(new TableOutputProcessFunction()) .name("ProcessTableOutput") .uid("ProcessTableOutput") .addSink(businessObjectSink) .name("businessObjectSink") .uid("businessObjectSink") .setParallelism(1); ``` For your second question about setting the EventTime I'm going to pull in Timo from the SDK team as I don't see an issue with your code right away. Best, Matthias On Wed, Nov 4, 2020 at 10:16 PM Fuyao Li <[hidden email]> wrote:
|
Hi Matthias, Thanks for your information. I have managed to figure out the first
issue you mentioned. Regarding the second issue. I have got some
progress on it. I have sent another email with the title 'BoundedOutOfOrderness Watermark Generator is NOT making the event time to advance' using another email of mine, [hidden email]. That email contains some more context on my issue. Please take a look. I have made some progress after sending that new email. Previously, I had managed to make timelag watermark strategy working in my code, but my bound out of orderness strategy or punctuated watermark strategy doesn't work well. It produces 8 watermarks each time. Two cycles are shown below. I managed to figure out the root cause is that Flink stream execution environment has a default parallelism as 8. I didn't notice in the doc, could the Community add this explicitly into the official doc to avoid some confusion? Thanks. From my understanding, the watermark advances based on the lowest watermark among the 8, so I can not advance the bound out of orderness watermark since I am only advancing 1 of the 8 parallelisms. If I set the entire stream execution environment to be of parallelism 1, it will reflect the watermark in the context correctly. One more thing is that this behavior is not reflected in the Flink Cluster web UI interface. I can see the watermark is advancing, but it is not in reality. That's causing the inconsistency problem I mentioned in the other email I mentioned above. Will this be considered as a bug in the UI? My current question is, since I have full outer join operation before the KeyedProcessFunction here. How can I let the bound of orderness watermark / punctuated watermark strategy work if the parallelism > 1? It can only update one of the 8 parallelisms for the watermark for this onTimer operator. Is this related to my Table full outer join operation before this step? According to the doc, https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html#table-exec-resource-default-parallelism Default parallelism should be the same like the stream environment. Why can't I update the watermarks for all 8 parallelisms? What should I do to enable this function with Parallelism larger than 1? Thanks. First round: (Note the first column of each log row is the timelag strategy, it is getting updated correctly for all 8 parallelism, but the other two strategies I mentioned above can't do that..) 14:28:01,199 INFO org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator - Emit Watermark: watermark based on system time:
1605047266198, periodicEmitWatermarkTime: 0, currentMaxTimestamp:
15000 Best regards, Fuyao On Fri, Nov 13, 2020 at 9:03 AM Matthias Pohl <[hidden email]> wrote:
|
Hi Matthias, One more question regarding Flink table parallelism, is it possible to configure the parallelism for Table operation at operator level, it seems we don't have such API available, right? Thanks! Best, Fuyao On Fri, Nov 13, 2020 at 11:48 AM Fuyao Li <[hidden email]> wrote:
|
Hi Matthias, Just to provide more context on this problem. I only have 1 partition per each Kafka Topic at the beginning before the join operation. After reading the doc: https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html#kafka-consumers-and-timestamp-extractionwatermark-emission Maybe that is the root cause of my problem here, with less than 8 partitions (only 1 partition in my case), using the default parallelism of 8 will cause this wrong behavior. This is my guess, it takes a while to test it out... What's your opinion on this? Thanks! Best, Fuyao On Fri, Nov 13, 2020 at 11:57 AM Fuyao Li <[hidden email]> wrote:
|
Hi Fuyao,
sorry for not replying earlier. You posted a lot of questions. I scanned the thread quickly, let me try to answer some of them and feel free to ask further questions afterwards. "is it possible to configure the parallelism for Table operation at operator level" No this is not possible at the moment. The reason is 1) we don't know how to expose such a functionality in a nice way. Maybe we will use SQL hints in the future [1]. 2) Sometime the planner sets the paralellism of operators explicitly to 1. All other operators will use the globally defined parallelism for the pipeline (also to not mess up retraction messages internally). You will be able to set the parallelism of the sink operation in Flink 1.12. "BoundedOutOfOrderness Watermark Generator is NOT making the event time to advance" Have you checked if you can use an interval join instead of a full join with state retention? Table/SQL pipelines that don't preserve a time attribute in the end might also erase the underlying watermarks. Thus, event time triggers will not work after your join. "Why can't I update the watermarks for all 8 parallelisms?" You could play around with idleness for your source [2]. Or you set the source parallelism to 1 (while keeping the rest of the pipeline globally set to 8), would that be an option? "Some type cast behavior of retracted streams I can't explain." toAppendStream/toRetractStream still need an update to the new type system. This is explained in FLIP-136 which will be part of Flink 1.13 [3]. I hope I could help a bit. Regards, Timo [1] https://cwiki.apache.org/confluence/display/FLINK/FLIP-113%3A+Supports+Dynamic+Table+Options+for+Flink+SQL [2] https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html#dealing-with-idle-sources [3] https://cwiki.apache.org/confluence/display/FLINK/FLIP-136%3A++Improve+interoperability+between+DataStream+and+Table+API On 13.11.20 21:39, Fuyao Li wrote: > Hi Matthias, > > Just to provide more context on this problem. I only have 1 partition > per each Kafka Topic at the beginning before the join operation. After > reading the doc: > https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html#kafka-consumers-and-timestamp-extractionwatermark-emission > <https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html#kafka-consumers-and-timestamp-extractionwatermark-emission> > > Maybe that is the root cause of my problem here, with less than 8 > partitions (only 1 partition in my case), using the default > parallelism of 8 will cause this wrong behavior. This is my guess, it > takes a while to test it out... What's your opinion on this? Thanks! > > Best, > > Fuyao > > > On Fri, Nov 13, 2020 at 11:57 AM Fuyao Li <[hidden email] > <mailto:[hidden email]>> wrote: > > Hi Matthias, > > One more question regarding Flink table parallelism, is it possible > to configure the parallelism for Table operation at operator level, > it seems we don't have such API available, right? Thanks! > > Best, > Fuyao > > On Fri, Nov 13, 2020 at 11:48 AM Fuyao Li <[hidden email] > <mailto:[hidden email]>> wrote: > > Hi Matthias, > > Thanks for your information. I have managed to figure out the > first issue you mentioned. Regarding the second issue. I have > got some progress on it. > > I have sent another email with the title 'BoundedOutOfOrderness > Watermark Generator is NOT making the event time to advance' > using another email of mine, [hidden email] > <mailto:[hidden email]>. That email contains some more > context on my issue. Please take a look. I have made some > progress after sending that new email. > > Previously, I had managed to make timelag watermark strategy > working in my code, but my bound out of orderness strategy or > punctuated watermark strategy doesn't work well. It produces 8 > watermarks each time. Two cycles are shown below. > > I managed to figure out the root cause is that Flink stream > execution environment has a default parallelism as 8.*I didn't > notice in the doc, could the Community add this explicitly into > the official doc to avoid some confusion? Thanks.* > > From my understanding, the watermark advances based on the > lowest watermark among the 8, so I can not advance the bound out > of orderness watermark since I am only advancing 1 of the 8 > parallelisms. If I set the entire stream execution environment > to be of parallelism 1, it will reflect the watermark in the > context correctly. One more thing is that this behavior is not > reflected in the Flink Cluster web UI interface. I can see the > watermark is advancing, but it is not in reality. *That's > causing the inconsistency problem I mentioned in the other email > I mentioned above. Will this be considered as a bug in the UI?* > > My current question is, since I have full outer join operation > before the KeyedProcessFunction here. How can I let the bound of > orderness watermark / punctuated watermark strategy work if the > parallelism > 1? It can only update one of the 8 parallelisms > for the watermark for this onTimer operator. Is this related to > my Table full outer join operation before this step? According > to the doc, > https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html#table-exec-resource-default-parallelism > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html#table-exec-resource-default-parallelism> > > Default parallelism should be the same like the stream > environment. Why can't I update the watermarks for all 8 > parallelisms? What should I do to enable this function with > Parallelism larger than 1? Thanks. > > First round: (Note the first column of each log row is the > timelag strategy, it is getting updated correctly for all 8 > parallelism, but the other two strategies I mentioned above > can't do that..) > > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266199, > periodicEmitWatermarkTime: 1605047172881, currentMaxTimestamp: > 1605047187881 (only one of the 8 parallelism for bound out of > orderness is getting my new watermark) > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266199, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:01,199 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047266198, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > > Second round: (I set the autoWatermark interval to be 5 seconds) > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 1605047172881, currentMaxTimestamp: > 1605047187881 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > 14:28:06,200 INFO > org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > - Emit Watermark: watermark based on system time: 1605047271200, > periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > > > Best regards, > > Fuyao > > > On Fri, Nov 13, 2020 at 9:03 AM Matthias Pohl > <[hidden email] <mailto:[hidden email]>> wrote: > > Hi Fuyao, > for your first question about the different behavior > depending on whether you chain the methods or not: Keep in > mind that you have to save the return value of the > assignTimestampsAndWatermarks method call if you don't chain > the methods together as it is also shown in [1]. > At least the following example from your first message is > indicating it: > ``` > retractStream.assignTimestampsAndWatermarks(new > BoRetractStreamTimestampAssigner()); (This is a deprecated > method) > // instead of: retractStream = > retractStream.assignTimestampsAndWatermarks(new > BoRetractStreamTimestampAssigner()); > retractStream > .keyBy(<key selector>) > .process(new TableOutputProcessFunction()) > .name("ProcessTableOutput") > .uid("ProcessTableOutput") > .addSink(businessObjectSink) > .name("businessObjectSink") > .uid("businessObjectSink") > .setParallelism(1); > ``` > > For your second question about setting the EventTime I'm > going to pull in Timo from the SDK team as I don't see an > issue with your code right away. > > Best, > Matthias > > [1] > https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#using-watermark-strategies > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#using-watermark-strategies> > > On Wed, Nov 4, 2020 at 10:16 PM Fuyao Li > <[hidden email] <mailto:[hidden email]>> wrote: > > Hi Flink Users and Community, > > For the first part of the question, the 12 hour time > difference is caused by a time extraction bug myself. I > can get the time translated correctly now. The type cast > problem does have some workarounds to solve it.. > > My major blocker right now is the onTimer part is not > properly triggered. I guess it is caused by failing to > configure the correct watermarks & timestamp assigners. > Please give me some insights. > > 1. If I don't chain the assignTimestampsAndWatermarks() > method in together with keyedBy().. and process().. > method. The context.timestamp() in my processElement() > function will be null. Is this some expected behavior? > The Flink examples didn't chain it together. (see > example here: > https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#using-watermark-strategies > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#using-watermark-strategies>) > 2. If I use registerEventTimeTimer() in > processElement(). The onTimer method will not be > triggered. However, I can trigger the onTimer method if > I simply change it to registerProcessingTimeTimer(). I > am using the settings below in the stream env. > > env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); > env.getConfig().setAutoWatermarkInterval(1000L); > > My code for method the process chain: > retractStream > > .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, > Row>>forBoundedOutOfOrderness(Duration.ofSeconds(20)) > > .withTimestampAssigner((booleanRowTuple2, timestamp) -> { > Row rowData = > booleanRowTuple2.f1; > LocalDateTime headerTime = > (LocalDateTime)rowData.getField(3); > LocalDateTime linesTime = > (LocalDateTime)rowData.getField(7); > > LocalDateTime > latestDBUpdateTime = null; > if (headerTime != null && > linesTime != null) { > latestDBUpdateTime = > headerTime.isAfter(linesTime) ? headerTime : linesTime; > } > else { > latestDBUpdateTime = > (headerTime != null) ? headerTime : linesTime; > } > if (latestDBUpdateTime != > null) { > return > latestDBUpdateTime.atZone(ZoneId.of("America/Los_Angeles")).toInstant().toEpochMilli(); > } > // In the worst case, we > use system time instead, which should never be reached. > return > System.currentTimeMillis(); > })) > // .assignTimestampsAndWatermarks(new > MyWaterStrategy()) // second way to create watermark, > doesn't work > .keyBy(value -> { > // There could be null fields for > header invoice_id field > String invoice_id_key = > (String)value.f1.getField(0); > if (invoice_id_key == null) { > invoice_id_key = > (String)value.f1.getField(4); > } > return invoice_id_key; > }) > .process(new TableOutputProcessFunction()) > .name("ProcessTableOutput") > .uid("ProcessTableOutput") > .addSink(businessObjectSink) > .name("businessObjectSink") > .uid("businessObjectSink") > .setParallelism(1); > > Best regards, > Fuyao > > On Mon, Nov 2, 2020 at 4:53 PM Fuyao Li > <[hidden email] <mailto:[hidden email]>> > wrote: > > Hi Flink Community, > > I am doing some research work on Flink Datastream > and Table API and I meet two major problems. I am > using Flink 1.11.2, scala version 2.11, java 8. My > use case looks like this. I plan to write a data > processing pipeline with two stages. My goal is to > construct a business object containing information > from several Kafka streams with a primary key and > emit the complete business object if such primary > key doesn't appear in the pipeline for 10 seconds. > > In the first stage, I first consume three Kafka > streams and transform it to Flink Datastream using a > deserialization schema containing some type and date > format transformation, and then I register these > data streams as Table and do a full outer join one > by one using Table API. I also add query > configuration for this to avoid excessive state. The > primary key is also the join key. > > In the second stage, I transform the joined table to > a retracted stream and put it into > KeyedProcessFunction to generate the business object > if the business object's primary key is inactive for > 10 second. > > Is this way of handling the data the suggested > approach? (I understand I can directly consume kafka > data in Table API. I haven't tried that yet, maybe > that's better?) Any suggestion is welcomed. During > implementing this, I meet two major problems and > several smaller questions under each problem. > > > 1. Some type cast behavior of retracted streams I > can't explain. > > (1) In the initial stage, I registered some field as > *java.sql.Date* or *java.sql.timestamp* following > the examples at > (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html#data-type-extraction > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html#data-type-extraction>) > . After join and transform to retracted stream, it > becomes *java.time.LocalDate* and > *java.time.LocalDateTime* instead. > > For example, when first ingesting the Kafka streams, > I registerd a attribute in java.sql.Timestamp type. > > @JsonAlias("ATTRIBUTE1") > private @DataTypeHint(value = "TIMESTAMP(6)", > bridgedTo = java.sql.Timestamp.class) Timestamp > ATTRIBUTE1; > > When I tried to cast the type information back after > the retracted stream, the code gives me error > information below. > > java.lang.ClassCastException: > java.time.LocalDateTime cannot be cast to > java.sql.Timestamp > > Maybe I should use toAppendStream instead since > append stream could register type information, but > toRetractedStream can't do that? > (https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html#convert-a-table-into-a-datastream-or-dataset > <https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html#convert-a-table-into-a-datastream-or-dataset>) > > My work around is to cast it to LocalDateTime first > and extract the epoch time, this doesn't seem to be > a final solution. > > (2) During timestamp conversion, the Flink to > retracted stream seems to lost the AM/PM information > in the stream and causing a 12 hour difference if it > is PM. > > I use joda time to do some timestamp conversion in > the first deserialization stage, my pattern looks > like this. "a" means AM/PM information > > DateTimeFormatter format3 = > DateTimeFormat.forPattern("dd-MMM-yy HH.mm.ss.SSSSSS > a").withZone(DateTimeZone.getDefault()); > > After the retracted stream, the AM/PM information is > not preserved. > > > 2. My onTimer method in KeyedProcessFunction can not > be triggered when I scheduled a event timer timer. > > I am using event time in my code. I am new to > configure watermarks and I might miss something to > configure it correctly. I also tried to register a > processing time, it could enter and produce some > results. > > I am trying to follow the example here: > https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html#example > <https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html#example> > > My onTimer method looks like this and the scheduled > event doesn't happen.. > > In processElement(): > > context.timerService().registerEventTimeTimer(current.getLastModifiedTime() > + 10000); > > My onTimer function > > @Override > public void onTimer(long timestamp, > OnTimerContext ctx, Collector<BusinessObject> > collector) throws Exception { > TestBusinessObjectState result = > testBusinessObjectState.value(); > log.info <http://log.info/>("Inside onTimer Method, > current key: {}, timestamp: {}, last modified time: > {}", ctx.getCurrentKey(), timestamp, > result.getLastModifiedTime()); > > // check if this is an outdated timer or > the latest timer > if (timestamp >= > result.getLastModifiedTime() + 10000) { > // emit the state on timeout > log.info <http://log.info/>("Collecting a business > object, {}", result.getBusinessObject().toString()); > > collector.collect(result.getBusinessObject()); > > cleanUp(ctx); > } > } > > private void cleanUp(Context ctx) throws > Exception { > Long timer = > testBusinessObjectState.value().getLastModifiedTime(); > ctx.timerService().deleteEventTimeTimer(timer); > testBusinessObjectState.clear(); > } > > > (1) When I assign the timestamp and watermarks > outside the process() method chain. The > "context.timestamp()" will be null. If I put it > inside the chain, it won't be null. Is this the > expected behavior? In the null case, the strange > thing is that, surprisingly, I can collect the > business object immediately without a designed 10 > second waiting time... This shouldn't happen, > right...? The processing timer also seems to work. > The code can enter the on timer method. > > retractStream.assignTimestampsAndWatermarks(new > BoRetractStreamTimestampAssigner()); (This is a > deprecated method) > > retractStream > .keyBy(<key selector>) > .process(new TableOutputProcessFunction()) > .name("ProcessTableOutput") > .uid("ProcessTableOutput") > .addSink(businessObjectSink) > .name("businessObjectSink") > .uid("businessObjectSink") > .setParallelism(1); > > (2) For watermarks configuration. I use an field in > the retracted stream as the event time. This time is > usually 15-20 seconds before current time. > > In my environment, I have done some settings for > streaming env based on information here( > https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator>). > My event doesn't always come, so I think I need to > set auto watermark interval to let the event timer > on timer works correctly. I have added the code below. > > env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); > env.getConfig().setAutoWatermarkInterval(1000L); > > 1> Which kind of watermark strategy should I use? > General BoundOutofOrderness or Watermark generator? > > I tried to write a Watermark generator and I just > don't how to apply it to the stream correctly. The > documentation doesn't explain very clearly. My code > looks like below and it doesn't work. > > assign part: > > .assignTimestampsAndWatermarks(WatermarkStrategy.forGenerator((WatermarkGeneratorSupplier<Tuple2<Boolean, > Row>>) context -> new > TableBoundOutofOrdernessGenerator())) > > watermark generater: > > I just assign the event time attribute following the > example in the doc. > (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator>) > > 2> I also tried to use the static method in Water > Strategy. The syntax is correct, but I meet the same > problem in 2.(1). > > > .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, > Row>>forBoundedOutOfOrderness(Duration.ofSeconds(15)) > > .withTimestampAssigner((booleanRowTuple2, timestamp) > -> { > <Select a event time > attribute in the booleanRowTuple2> > })) > > > (3) For the retracted datastream, do I need to > explicitly attach it to the stream environment? I > think it is done by default, right? Just want to > confirm it. I do have the env.execute() at the end > of the code. > > I understand this is a lot of questions, thanks a > lot for your patience to look through my email! If > there is anything unclear, please reach out to me. > Thanks! > > > Best regards, > > Fuyao Li > |
Hi Timo, Thanks for your reply! I think your suggestions is really
helpful! The good news is that I had managed to figure out it
something by myself few days ago. 1. Thanks for the update about the table parallelism issue! 2. After trying out the idleness setting. It prevents some idle subtasks from blocking the pipeline's overall watermark and it works for me. Based on my observation and reading the source code, I have summarized some notes. Please correct me if I am wrong.
3. In order to route the correct information to the subtask in the join step, I have added the keyed() logic in the source based on the join key in the join step. It seems to work correctly and could route the message to a current place. 4. For the interval join, I think I can't use it directly since I need to use full outer join to not lose any information from any upstream datastream. I think interval join is a inner join it can't do this task. I guess my only option is to do full outer join with query configuration. 5. One more question about the data replay issue. I read the
ververica blog
(https://www.ververica.com/blog/replayable-process-functions-time-ordering-and-timers)
and I think with replay use case, we will face some similar
issues. I think the suggested approach mentioned (1). Puts each incoming track record
in a map keyed by its timestamp (2). creates an event timer to
process that record once the watermark hits that point. I kind of understand the idea here. Buffer all the data(maybe delete some of the old track if processed) in a track ordered by timestamp and trigger the event timer sequentially with this buffered track. Based on my understanding, this
buffered design is only suitable for offline data
processing, right? (It is a waste of resource to buffer this in
real time. ) Also, from the article, I think they
are using periodic watermark strategy[1]. how can they process the
last piece of data records with periodic watermark strategy since
there is no more incoming data to advance the watermark? So the
last piece of data will never be processed here? Is there a way to
gracefully handle this? My use case doesn't allow me to lose any
information.
Best, Fuyao
On 11/20/20 08:55, Timo Walther wrote:
Hi Fuyao, |
Hi Timo, One more question, the blog also mentioned a jira task to solve this issue. https://issues.apache.org/jira/browse/FLINK-10886. Will this feature be available in 1.12? Thanks! Best, Fuyao On 11/20/20 11:37, [hidden email]
wrote:
|
AFAIK, FLINK-10886 is not implemented yet. cc @Becket may know more plans about this feature. Best, Jark On Sat, 21 Nov 2020 at 03:46, <[hidden email]> wrote:
|
In reply to this post by Fuyao Li-2
Hi Fuyao,
great that you could make progress. 2. Btw nice summary of the idleness concept. We should have that in the docs actually. 4. By looking at tests like `IntervalJoinITCase` [1] it seems that we also support FULL OUTER JOINs as interval joins. Maybe you can make use of them. 5. "buffered design is only suitable for *offline* data processing, right" I guess this depends on the size of data and the SLAs for checkpointing time. In general, when using the RocksDB state backend you can add a lot of data into state. One backfill option could be to warmup a pipeline and only perform a savepoint at the end. Then one could start the online pipeline (maybe with a different flag set) from this savepoint again. Or you even configure the state handling of your backfill operator via a control stream. We are also working on a better offline data processing story for DataStream API programs in Flink 1.12. This might be interesting for you as well [2]. "how can they process the last piece of data" Maybe they are emitting a max watermark at some point. Bounded sources will also do that. The watermarks are not checkpointed and start from scratch when restarting the Flink job. Regards, Timo [1] https://github.com/apache/flink/blob/master/flink-table/flink-table-planner-blink/src/test/scala/org/apache/flink/table/planner/runtime/stream/sql/IntervalJoinITCase.scala [2] https://cwiki.apache.org/confluence/display/FLINK/FLIP-134%3A+Batch+execution+for+the+DataStream+API On 20.11.20 20:37, [hidden email] wrote: > Hi Timo, > > Thanks for your reply! I think your suggestions is really helpful! The > good news is that I had managed to figure out it something by myself few > days ago. > > 1. Thanks for the update about the table parallelism issue! > > 2. After trying out the idleness setting. It prevents some idle subtasks > from blocking the pipeline's overall watermark and it works for me. > Based on my observation and reading the source code, I have summarized > some notes. Please correct me if I am wrong. > > 1. (1)Watermark is independent within each subtask for an Flink operator. > 2. (2)The watermark of the multi-parallelism table operator is always > dominated by least watermark of the current*ACTIVE*subtasks. > 3. (3)With withIdleness() configured. A subtask will be mark as idle if > it hasn’t receive message for configured period of time. It will NOT > execute onPeriodEmit() and emit watermark after reaching the idle > state. Between [the start of the application/receive a new message] > and [reaching into the idle state], the onPeriodEmit() will still > emit watermark and dominate the overall context watermark if it > holds the smallest watermark among the subtasks. > 4. (4)Once an idle subtask receive a new message, it will switch its > status from idle to active and start to influence the overall > context watermark. > > 3. In order to route the correct information to the subtask in the join > step, I have added the keyed() logic in the source based on the join key > in the join step. It seems to work correctly and could route the message > to a current place. > > 4. For the interval join, I think I can't use it directly since I need > to use full outer join to not lose any information from any upstream > datastream. I think interval join is a inner join it can't do this task. > I guess my only option is to do full outer join with query configuration. > > 5. One more question about the data replay issue. I read the ververica > blog > (https://www.ververica.com/blog/replayable-process-functions-time-ordering-and-timers) > and I think with replay use case, we will face some similar issues. I > think the suggested approach mentioned > > (1). Puts each incoming track record in a map keyed by its timestamp > > (2). creates an event timer to process that record once the watermark > hits that point. > > I kind of understand the idea here. Buffer all the data(maybe delete > some of the old track if processed) in a track ordered by timestamp and > trigger the event timer sequentially with this buffered track. > > Based on my understanding, this buffered design is only suitable for > *offline* data processing, right? (It is a waste of resource to buffer > this in real time. ) > > Also, from the article, I think they are using periodic watermark > strategy[1]. how can they process the last piece of data records with > periodic watermark strategy since there is no more incoming data to > advance the watermark? So the last piece of data will never be processed > here? Is there a way to gracefully handle this? My use case doesn't > allow me to lose any information. > > > [1]https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator > > Best, > > Fuyao > > > On 11/20/20 08:55, Timo Walther wrote: >> Hi Fuyao, >> >> sorry for not replying earlier. >> >> You posted a lot of questions. I scanned the thread quickly, let me >> try to answer some of them and feel free to ask further questions >> afterwards. >> >> "is it possible to configure the parallelism for Table operation at >> operator level" >> >> No this is not possible at the moment. The reason is 1) we don't know >> how to expose such a functionality in a nice way. Maybe we will use >> SQL hints in the future [1]. 2) Sometime the planner sets the >> paralellism of operators explicitly to 1. All other operators will use >> the globally defined parallelism for the pipeline (also to not mess up >> retraction messages internally). You will be able to set the >> parallelism of the sink operation in Flink 1.12. >> >> "BoundedOutOfOrderness Watermark Generator is NOT making the event >> time to advance" >> >> Have you checked if you can use an interval join instead of a full >> join with state retention? Table/SQL pipelines that don't preserve a >> time attribute in the end might also erase the underlying watermarks. >> Thus, event time triggers will not work after your join. >> >> "Why can't I update the watermarks for all 8 parallelisms?" >> >> You could play around with idleness for your source [2]. Or you set >> the source parallelism to 1 (while keeping the rest of the pipeline >> globally set to 8), would that be an option? >> >> "Some type cast behavior of retracted streams I can't explain." >> >> toAppendStream/toRetractStream still need an update to the new type >> system. This is explained in FLIP-136 which will be part of Flink 1.13 >> [3]. >> >> I hope I could help a bit. >> >> Regards, >> Timo >> >> >> [1] >> https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-113*3A*Supports*Dynamic*Table*Options*for*Flink*SQL__;JSsrKysrKys!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J6qWrWNk$ >> [2] >> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html*dealing-with-idle-sources__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JMW06Who$ >> [3] >> https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-136*3A**AImprove*interoperability*between*DataStream*and*Table*API__;JSsrKysrKysr!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JfAjyGyQ$ >> >> On 13.11.20 21:39, Fuyao Li wrote: >>> Hi Matthias, >>> >>> Just to provide more context on this problem. I only have 1 partition >>> per each Kafka Topic at the beginning before the join operation. >>> After reading the doc: >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$ >>> > >>> >>> Maybe that is the root cause of my problem here, with less than 8 >>> partitions (only 1 partition in my case), using the default >>> parallelism of 8 will cause this wrong behavior. This is my guess, it >>> takes a while to test it out... What's your opinion on this? Thanks! >>> >>> Best, >>> >>> Fuyao >>> >>> >>> On Fri, Nov 13, 2020 at 11:57 AM Fuyao Li <[hidden email] >>> <mailto:[hidden email]>> wrote: >>> >>> Hi Matthias, >>> >>> One more question regarding Flink table parallelism, is it possible >>> to configure the parallelism for Table operation at operator level, >>> it seems we don't have such API available, right? Thanks! >>> >>> Best, >>> Fuyao >>> >>> On Fri, Nov 13, 2020 at 11:48 AM Fuyao Li <[hidden email] >>> <mailto:[hidden email]>> wrote: >>> >>> Hi Matthias, >>> >>> Thanks for your information. I have managed to figure out the >>> first issue you mentioned. Regarding the second issue. I have >>> got some progress on it. >>> >>> I have sent another email with the title 'BoundedOutOfOrderness >>> Watermark Generator is NOT making the event time to advance' >>> using another email of mine, [hidden email] >>> <mailto:[hidden email]>. That email contains some more >>> context on my issue. Please take a look. I have made some >>> progress after sending that new email. >>> >>> Previously, I had managed to make timelag watermark strategy >>> working in my code, but my bound out of orderness strategy or >>> punctuated watermark strategy doesn't work well. It produces 8 >>> watermarks each time. Two cycles are shown below. >>> >>> I managed to figure out the root cause is that Flink stream >>> execution environment has a default parallelism as 8.*I didn't >>> notice in the doc, could the Community add this explicitly into >>> the official doc to avoid some confusion? Thanks.* >>> >>> From my understanding, the watermark advances based on the >>> lowest watermark among the 8, so I can not advance the bound out >>> of orderness watermark since I am only advancing 1 of the 8 >>> parallelisms. If I set the entire stream execution environment >>> to be of parallelism 1, it will reflect the watermark in the >>> context correctly. One more thing is that this behavior is not >>> reflected in the Flink Cluster web UI interface. I can see the >>> watermark is advancing, but it is not in reality. *That's >>> causing the inconsistency problem I mentioned in the other email >>> I mentioned above. Will this be considered as a bug in the UI?* >>> >>> My current question is, since I have full outer join operation >>> before the KeyedProcessFunction here. How can I let the bound of >>> orderness watermark / punctuated watermark strategy work if the >>> parallelism > 1? It can only update one of the 8 parallelisms >>> for the watermark for this onTimer operator. Is this related to >>> my Table full outer join operation before this step? According >>> to the doc, >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$ >>> > >>> >>> Default parallelism should be the same like the stream >>> environment. Why can't I update the watermarks for all 8 >>> parallelisms? What should I do to enable this function with >>> Parallelism larger than 1? Thanks. >>> >>> First round: (Note the first column of each log row is the >>> timelag strategy, it is getting updated correctly for all 8 >>> parallelism, but the other two strategies I mentioned above >>> can't do that..) >>> >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266199, >>> periodicEmitWatermarkTime: 1605047172881, currentMaxTimestamp: >>> 1605047187881 (only one of the 8 parallelism for bound out of >>> orderness is getting my new watermark) >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266199, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:01,199 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047266198, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> >>> Second round: (I set the autoWatermark interval to be 5 seconds) >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 1605047172881, currentMaxTimestamp: >>> 1605047187881 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> 14:28:06,200 INFO >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator >>> - Emit Watermark: watermark based on system time: 1605047271200, >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 >>> >>> >>> Best regards, >>> >>> Fuyao >>> >>> >>> On Fri, Nov 13, 2020 at 9:03 AM Matthias Pohl >>> <[hidden email] <mailto:[hidden email]>> wrote: >>> >>> Hi Fuyao, >>> for your first question about the different behavior >>> depending on whether you chain the methods or not: Keep in >>> mind that you have to save the return value of the >>> assignTimestampsAndWatermarks method call if you don't chain >>> the methods together as it is also shown in [1]. >>> At least the following example from your first message is >>> indicating it: >>> ``` >>> retractStream.assignTimestampsAndWatermarks(new >>> BoRetractStreamTimestampAssigner()); (This is a deprecated >>> method) >>> // instead of: retractStream = >>> retractStream.assignTimestampsAndWatermarks(new >>> BoRetractStreamTimestampAssigner()); >>> retractStream >>> .keyBy(<key selector>) >>> .process(new TableOutputProcessFunction()) >>> .name("ProcessTableOutput") >>> .uid("ProcessTableOutput") >>> .addSink(businessObjectSink) >>> .name("businessObjectSink") >>> .uid("businessObjectSink") >>> .setParallelism(1); >>> ``` >>> >>> For your second question about setting the EventTime I'm >>> going to pull in Timo from the SDK team as I don't see an >>> issue with your code right away. >>> >>> Best, >>> Matthias >>> >>> [1] >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ >>> > >>> >>> On Wed, Nov 4, 2020 at 10:16 PM Fuyao Li >>> <[hidden email] <mailto:[hidden email]>> >>> wrote: >>> >>> Hi Flink Users and Community, >>> >>> For the first part of the question, the 12 hour time >>> difference is caused by a time extraction bug myself. I >>> can get the time translated correctly now. The type cast >>> problem does have some workarounds to solve it.. >>> >>> My major blocker right now is the onTimer part is not >>> properly triggered. I guess it is caused by failing to >>> configure the correct watermarks & timestamp assigners. >>> Please give me some insights. >>> >>> 1. If I don't chain the assignTimestampsAndWatermarks() >>> method in together with keyedBy().. and process().. >>> method. The context.timestamp() in my processElement() >>> function will be null. Is this some expected behavior? >>> The Flink examples didn't chain it together. (see >>> example here: >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ >>> >) >>> 2. If I use registerEventTimeTimer() in >>> processElement(). The onTimer method will not be >>> triggered. However, I can trigger the onTimer method if >>> I simply change it to registerProcessingTimeTimer(). I >>> am using the settings below in the stream env. >>> >>> env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); >>> env.getConfig().setAutoWatermarkInterval(1000L); >>> >>> My code for method the process chain: >>> retractStream >>> .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, >>> Row>>forBoundedOutOfOrderness(Duration.ofSeconds(20)) >>> .withTimestampAssigner((booleanRowTuple2, timestamp) -> { >>> Row rowData = >>> booleanRowTuple2.f1; >>> LocalDateTime headerTime = >>> (LocalDateTime)rowData.getField(3); >>> LocalDateTime linesTime = >>> (LocalDateTime)rowData.getField(7); >>> >>> LocalDateTime >>> latestDBUpdateTime = null; >>> if (headerTime != null && >>> linesTime != null) { >>> latestDBUpdateTime = >>> headerTime.isAfter(linesTime) ? headerTime : linesTime; >>> } >>> else { >>> latestDBUpdateTime = >>> (headerTime != null) ? headerTime : linesTime; >>> } >>> if (latestDBUpdateTime != >>> null) { >>> return >>> latestDBUpdateTime.atZone(ZoneId.of("America/Los_Angeles")).toInstant().toEpochMilli(); >>> } >>> // In the worst case, we >>> use system time instead, which should never be reached. >>> return >>> System.currentTimeMillis(); >>> })) >>> // .assignTimestampsAndWatermarks(new >>> MyWaterStrategy()) // second way to create watermark, >>> doesn't work >>> .keyBy(value -> { >>> // There could be null fields for >>> header invoice_id field >>> String invoice_id_key = >>> (String)value.f1.getField(0); >>> if (invoice_id_key == null) { >>> invoice_id_key = >>> (String)value.f1.getField(4); >>> } >>> return invoice_id_key; >>> }) >>> .process(new >>> TableOutputProcessFunction()) >>> .name("ProcessTableOutput") >>> .uid("ProcessTableOutput") >>> .addSink(businessObjectSink) >>> .name("businessObjectSink") >>> .uid("businessObjectSink") >>> .setParallelism(1); >>> >>> Best regards, >>> Fuyao >>> >>> On Mon, Nov 2, 2020 at 4:53 PM Fuyao Li >>> <[hidden email] <mailto:[hidden email]>> >>> wrote: >>> >>> Hi Flink Community, >>> >>> I am doing some research work on Flink Datastream >>> and Table API and I meet two major problems. I am >>> using Flink 1.11.2, scala version 2.11, java 8. My >>> use case looks like this. I plan to write a data >>> processing pipeline with two stages. My goal is to >>> construct a business object containing information >>> from several Kafka streams with a primary key and >>> emit the complete business object if such primary >>> key doesn't appear in the pipeline for 10 seconds. >>> >>> In the first stage, I first consume three Kafka >>> streams and transform it to Flink Datastream using a >>> deserialization schema containing some type and date >>> format transformation, and then I register these >>> data streams as Table and do a full outer join one >>> by one using Table API. I also add query >>> configuration for this to avoid excessive state. The >>> primary key is also the join key. >>> >>> In the second stage, I transform the joined table to >>> a retracted stream and put it into >>> KeyedProcessFunction to generate the business object >>> if the business object's primary key is inactive for >>> 10 second. >>> >>> Is this way of handling the data the suggested >>> approach? (I understand I can directly consume kafka >>> data in Table API. I haven't tried that yet, maybe >>> that's better?) Any suggestion is welcomed. During >>> implementing this, I meet two major problems and >>> several smaller questions under each problem. >>> >>> >>> 1. Some type cast behavior of retracted streams I >>> can't explain. >>> >>> (1) In the initial stage, I registered some field as >>> *java.sql.Date* or *java.sql.timestamp* following >>> the examples at >>> (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$ >>> >) >>> . After join and transform to retracted stream, it >>> becomes *java.time.LocalDate* and >>> *java.time.LocalDateTime* instead. >>> >>> For example, when first ingesting the Kafka streams, >>> I registerd a attribute in java.sql.Timestamp type. >>> >>> @JsonAlias("ATTRIBUTE1") >>> private @DataTypeHint(value = "TIMESTAMP(6)", >>> bridgedTo = java.sql.Timestamp.class) Timestamp >>> ATTRIBUTE1; >>> >>> When I tried to cast the type information back after >>> the retracted stream, the code gives me error >>> information below. >>> >>> java.lang.ClassCastException: >>> java.time.LocalDateTime cannot be cast to >>> java.sql.Timestamp >>> >>> Maybe I should use toAppendStream instead since >>> append stream could register type information, but >>> toRetractedStream can't do that? >>> (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$ >>> >) >>> >>> My work around is to cast it to LocalDateTime first >>> and extract the epoch time, this doesn't seem to be >>> a final solution. >>> >>> (2) During timestamp conversion, the Flink to >>> retracted stream seems to lost the AM/PM information >>> in the stream and causing a 12 hour difference if it >>> is PM. >>> >>> I use joda time to do some timestamp conversion in >>> the first deserialization stage, my pattern looks >>> like this. "a" means AM/PM information >>> >>> DateTimeFormatter format3 = >>> DateTimeFormat.forPattern("dd-MMM-yy HH.mm.ss.SSSSSS >>> a").withZone(DateTimeZone.getDefault()); >>> >>> After the retracted stream, the AM/PM information is >>> not preserved. >>> >>> >>> 2. My onTimer method in KeyedProcessFunction can not >>> be triggered when I scheduled a event timer timer. >>> >>> I am using event time in my code. I am new to >>> configure watermarks and I might miss something to >>> configure it correctly. I also tried to register a >>> processing time, it could enter and produce some >>> results. >>> >>> I am trying to follow the example here: >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$ >>> > >>> >>> My onTimer method looks like this and the scheduled >>> event doesn't happen.. >>> >>> In processElement(): >>> >>> context.timerService().registerEventTimeTimer(current.getLastModifiedTime() >>> + 10000); >>> >>> My onTimer function >>> >>> @Override >>> public void onTimer(long timestamp, >>> OnTimerContext ctx, Collector<BusinessObject> >>> collector) throws Exception { >>> TestBusinessObjectState result = >>> testBusinessObjectState.value(); >>> log.info >>> <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$ >>> >("Inside onTimer Method, >>> current key: {}, timestamp: {}, last modified time: >>> {}", ctx.getCurrentKey(), timestamp, >>> result.getLastModifiedTime()); >>> >>> // check if this is an outdated timer or >>> the latest timer >>> if (timestamp >= >>> result.getLastModifiedTime() + 10000) { >>> // emit the state on timeout >>> log.info >>> <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$ >>> >("Collecting a business >>> object, {}", result.getBusinessObject().toString()); >>> collector.collect(result.getBusinessObject()); >>> >>> cleanUp(ctx); >>> } >>> } >>> >>> private void cleanUp(Context ctx) throws >>> Exception { >>> Long timer = >>> testBusinessObjectState.value().getLastModifiedTime(); >>> ctx.timerService().deleteEventTimeTimer(timer); >>> testBusinessObjectState.clear(); >>> } >>> >>> >>> (1) When I assign the timestamp and watermarks >>> outside the process() method chain. The >>> "context.timestamp()" will be null. If I put it >>> inside the chain, it won't be null. Is this the >>> expected behavior? In the null case, the strange >>> thing is that, surprisingly, I can collect the >>> business object immediately without a designed 10 >>> second waiting time... This shouldn't happen, >>> right...? The processing timer also seems to work. >>> The code can enter the on timer method. >>> >>> retractStream.assignTimestampsAndWatermarks(new >>> BoRetractStreamTimestampAssigner()); (This is a >>> deprecated method) >>> >>> retractStream >>> .keyBy(<key selector>) >>> .process(new TableOutputProcessFunction()) >>> .name("ProcessTableOutput") >>> .uid("ProcessTableOutput") >>> .addSink(businessObjectSink) >>> .name("businessObjectSink") >>> .uid("businessObjectSink") >>> .setParallelism(1); >>> >>> (2) For watermarks configuration. I use an field in >>> the retracted stream as the event time. This time is >>> usually 15-20 seconds before current time. >>> >>> In my environment, I have done some settings for >>> streaming env based on information here( >>> https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ >>> >). >>> My event doesn't always come, so I think I need to >>> set auto watermark interval to let the event timer >>> on timer works correctly. I have added the code >>> below. >>> >>> env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); >>> env.getConfig().setAutoWatermarkInterval(1000L); >>> >>> 1> Which kind of watermark strategy should I use? >>> General BoundOutofOrderness or Watermark generator? >>> >>> I tried to write a Watermark generator and I just >>> don't how to apply it to the stream correctly. The >>> documentation doesn't explain very clearly. My code >>> looks like below and it doesn't work. >>> >>> assign part: >>> >>> .assignTimestampsAndWatermarks(WatermarkStrategy.forGenerator((WatermarkGeneratorSupplier<Tuple2<Boolean, >>> Row>>) context -> new >>> TableBoundOutofOrdernessGenerator())) >>> >>> watermark generater: >>> >>> I just assign the event time attribute following the >>> example in the doc. >>> (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ >>> <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ >>> >) >>> >>> 2> I also tried to use the static method in Water >>> Strategy. The syntax is correct, but I meet the same >>> problem in 2.(1). >>> >>> .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, >>> Row>>forBoundedOutOfOrderness(Duration.ofSeconds(15)) >>> .withTimestampAssigner((booleanRowTuple2, timestamp) >>> -> { >>> <Select a event time >>> attribute in the booleanRowTuple2> >>> })) >>> >>> >>> (3) For the retracted datastream, do I need to >>> explicitly attach it to the stream environment? I >>> think it is done by default, right? Just want to >>> confirm it. I do have the env.execute() at the end >>> of the code. >>> >>> I understand this is a lot of questions, thanks a >>> lot for your patience to look through my email! If >>> there is anything unclear, please reach out to me. >>> Thanks! >>> >>> >>> Best regards, >>> >>> Fuyao Li >>> >> |
Hi Timo, Thanks for your information. I saw the Flink SQL can actually do the full outer join in the test code with interval join semantic. However, this is not explicitly shown in the Flink SQL documentation. That makes me thinking this might not be available for me to use. Maybe the doc could be updated to explicitly show what kind of join it can do with interval join. (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/sql/queries.html#joins) Thanks! Fuyao On Tue, Nov 24, 2020 at 9:06 AM Timo Walther <[hidden email]> wrote: Hi Fuyao, |
Hi Fuyao,
yes I agree. The code evolved quicker than the docs. I will create an issue for this. Regards, Timo On 25.11.20 19:27, Fuyao Li wrote: > Hi Timo, > Thanks for your information. I saw the Flink SQL can actually do the > full outer join in the test code with interval join semantic. However, > this is not explicitly shown in the Flink SQL documentation. That makes > me thinking this might not be available for me to use. Maybe the doc > could be updated to explicitly show what kind of join it can do with > interval join. > (https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/sql/queries.html#joins > <https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/sql/queries.html#joins>) > > Thanks! > Fuyao > > > > On Tue, Nov 24, 2020 at 9:06 AM Timo Walther <[hidden email] > <mailto:[hidden email]>> wrote: > > Hi Fuyao, > > great that you could make progress. > > 2. Btw nice summary of the idleness concept. We should have that in the > docs actually. > > 4. By looking at tests like `IntervalJoinITCase` [1] it seems that we > also support FULL OUTER JOINs as interval joins. Maybe you can make use > of them. > > 5. "buffered design is only suitable for *offline* data processing, > right" > I guess this depends on the size of data and the SLAs for checkpointing > time. > > In general, when using the RocksDB state backend you can add a lot of > data into state. One backfill option could be to warmup a pipeline and > only perform a savepoint at the end. Then one could start the online > pipeline (maybe with a different flag set) from this savepoint again. > > Or you even configure the state handling of your backfill operator > via a > control stream. > > We are also working on a better offline data processing story for > DataStream API programs in Flink 1.12. This might be interesting for > you > as well [2]. > > "how can they process the last piece of data" > > Maybe they are emitting a max watermark at some point. Bounded sources > will also do that. The watermarks are not checkpointed and start from > scratch when restarting the Flink job. > > Regards, > Timo > > [1] > https://github.com/apache/flink/blob/master/flink-table/flink-table-planner-blink/src/test/scala/org/apache/flink/table/planner/runtime/stream/sql/IntervalJoinITCase.scala > <https://github.com/apache/flink/blob/master/flink-table/flink-table-planner-blink/src/test/scala/org/apache/flink/table/planner/runtime/stream/sql/IntervalJoinITCase.scala> > [2] > https://cwiki.apache.org/confluence/display/FLINK/FLIP-134%3A+Batch+execution+for+the+DataStream+API > <https://cwiki.apache.org/confluence/display/FLINK/FLIP-134%3A+Batch+execution+for+the+DataStream+API> > > > > > On 20.11.20 20:37, [hidden email] <mailto:[hidden email]> > wrote: > > Hi Timo, > > > > Thanks for your reply! I think your suggestions is really > helpful! The > > good news is that I had managed to figure out it something by > myself few > > days ago. > > > > 1. Thanks for the update about the table parallelism issue! > > > > 2. After trying out the idleness setting. It prevents some idle > subtasks > > from blocking the pipeline's overall watermark and it works for me. > > Based on my observation and reading the source code, I have > summarized > > some notes. Please correct me if I am wrong. > > > > 1. (1)Watermark is independent within each subtask for an Flink > operator. > > 2. (2)The watermark of the multi-parallelism table operator is > always > > dominated by least watermark of the current*ACTIVE*subtasks. > > 3. (3)With withIdleness() configured. A subtask will be mark as > idle if > > it hasn’t receive message for configured period of time. It > will NOT > > execute onPeriodEmit() and emit watermark after reaching the idle > > state. Between [the start of the application/receive a new > message] > > and [reaching into the idle state], the onPeriodEmit() will still > > emit watermark and dominate the overall context watermark if it > > holds the smallest watermark among the subtasks. > > 4. (4)Once an idle subtask receive a new message, it will switch its > > status from idle to active and start to influence the overall > > context watermark. > > > > 3. In order to route the correct information to the subtask in > the join > > step, I have added the keyed() logic in the source based on the > join key > > in the join step. It seems to work correctly and could route the > message > > to a current place. > > > > 4. For the interval join, I think I can't use it directly since I > need > > to use full outer join to not lose any information from any upstream > > datastream. I think interval join is a inner join it can't do > this task. > > I guess my only option is to do full outer join with query > configuration. > > > > 5. One more question about the data replay issue. I read the > ververica > > blog > > > (https://www.ververica.com/blog/replayable-process-functions-time-ordering-and-timers > <https://www.ververica.com/blog/replayable-process-functions-time-ordering-and-timers>) > > > and I think with replay use case, we will face some similar > issues. I > > think the suggested approach mentioned > > > > (1). Puts each incoming track record in a map keyed by its > timestamp > > > > (2). creates an event timer to process that record once the > watermark > > hits that point. > > > > I kind of understand the idea here. Buffer all the data(maybe delete > > some of the old track if processed) in a track ordered by > timestamp and > > trigger the event timer sequentially with this buffered track. > > > > Based on my understanding, this buffered design is only suitable for > > *offline* data processing, right? (It is a waste of resource to > buffer > > this in real time. ) > > > > Also, from the article, I think they are using periodic watermark > > strategy[1]. how can they process the last piece of data records > with > > periodic watermark strategy since there is no more incoming data to > > advance the watermark? So the last piece of data will never be > processed > > here? Is there a way to gracefully handle this? My use case doesn't > > allow me to lose any information. > > > > > > > [1]https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator > <https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html#writing-a-periodic-watermarkgenerator> > > > > Best, > > > > Fuyao > > > > > > On 11/20/20 08:55, Timo Walther wrote: > >> Hi Fuyao, > >> > >> sorry for not replying earlier. > >> > >> You posted a lot of questions. I scanned the thread quickly, let me > >> try to answer some of them and feel free to ask further questions > >> afterwards. > >> > >> "is it possible to configure the parallelism for Table operation at > >> operator level" > >> > >> No this is not possible at the moment. The reason is 1) we don't > know > >> how to expose such a functionality in a nice way. Maybe we will use > >> SQL hints in the future [1]. 2) Sometime the planner sets the > >> paralellism of operators explicitly to 1. All other operators > will use > >> the globally defined parallelism for the pipeline (also to not > mess up > >> retraction messages internally). You will be able to set the > >> parallelism of the sink operation in Flink 1.12. > >> > >> "BoundedOutOfOrderness Watermark Generator is NOT making the event > >> time to advance" > >> > >> Have you checked if you can use an interval join instead of a full > >> join with state retention? Table/SQL pipelines that don't > preserve a > >> time attribute in the end might also erase the underlying > watermarks. > >> Thus, event time triggers will not work after your join. > >> > >> "Why can't I update the watermarks for all 8 parallelisms?" > >> > >> You could play around with idleness for your source [2]. Or you set > >> the source parallelism to 1 (while keeping the rest of the pipeline > >> globally set to 8), would that be an option? > >> > >> "Some type cast behavior of retracted streams I can't explain." > >> > >> toAppendStream/toRetractStream still need an update to the new type > >> system. This is explained in FLIP-136 which will be part of > Flink 1.13 > >> [3]. > >> > >> I hope I could help a bit. > >> > >> Regards, > >> Timo > >> > >> > >> [1] > >> > https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-113*3A*Supports*Dynamic*Table*Options*for*Flink*SQL__;JSsrKysrKys!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J6qWrWNk$ > <https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-113*3A*Supports*Dynamic*Table*Options*for*Flink*SQL__;JSsrKysrKys!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J6qWrWNk$> > > >> [2] > >> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html*dealing-with-idle-sources__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JMW06Who$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/event_timestamps_watermarks.html*dealing-with-idle-sources__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JMW06Who$> > > >> [3] > >> > https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-136*3A**AImprove*interoperability*between*DataStream*and*Table*API__;JSsrKysrKysr!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JfAjyGyQ$ > <https://urldefense.com/v3/__https://cwiki.apache.org/confluence/display/FLINK/FLIP-136*3A**AImprove*interoperability*between*DataStream*and*Table*API__;JSsrKysrKysr!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JfAjyGyQ$> > > >> > >> On 13.11.20 21:39, Fuyao Li wrote: > >>> Hi Matthias, > >>> > >>> Just to provide more context on this problem. I only have 1 > partition > >>> per each Kafka Topic at the beginning before the join operation. > >>> After reading the doc: > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/connectors/kafka.html*kafka-consumers-and-timestamp-extractionwatermark-emission__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JnAwo_lc$> > > >>> > > >>> > >>> Maybe that is the root cause of my problem here, with less than 8 > >>> partitions (only 1 partition in my case), using the default > >>> parallelism of 8 will cause this wrong behavior. This is my > guess, it > >>> takes a while to test it out... What's your opinion on this? > Thanks! > >>> > >>> Best, > >>> > >>> Fuyao > >>> > >>> > >>> On Fri, Nov 13, 2020 at 11:57 AM Fuyao Li > <[hidden email] <mailto:[hidden email]> > >>> <mailto:[hidden email] <mailto:[hidden email]>>> > wrote: > >>> > >>> Hi Matthias, > >>> > >>> One more question regarding Flink table parallelism, is it > possible > >>> to configure the parallelism for Table operation at > operator level, > >>> it seems we don't have such API available, right? Thanks! > >>> > >>> Best, > >>> Fuyao > >>> > >>> On Fri, Nov 13, 2020 at 11:48 AM Fuyao Li > <[hidden email] <mailto:[hidden email]> > >>> <mailto:[hidden email] <mailto:[hidden email]>>> > wrote: > >>> > >>> Hi Matthias, > >>> > >>> Thanks for your information. I have managed to figure > out the > >>> first issue you mentioned. Regarding the second issue. > I have > >>> got some progress on it. > >>> > >>> I have sent another email with the title > 'BoundedOutOfOrderness > >>> Watermark Generator is NOT making the event time to > advance' > >>> using another email of mine, [hidden email] > <mailto:[hidden email]> > >>> <mailto:[hidden email] <mailto:[hidden email]>>. That > email contains some more > >>> context on my issue. Please take a look. I have made some > >>> progress after sending that new email. > >>> > >>> Previously, I had managed to make timelag watermark > strategy > >>> working in my code, but my bound out of orderness > strategy or > >>> punctuated watermark strategy doesn't work well. It > produces 8 > >>> watermarks each time. Two cycles are shown below. > >>> > >>> I managed to figure out the root cause is that Flink stream > >>> execution environment has a default parallelism as 8.*I > didn't > >>> notice in the doc, could the Community add this > explicitly into > >>> the official doc to avoid some confusion? Thanks.* > >>> > >>> From my understanding, the watermark advances based on the > >>> lowest watermark among the 8, so I can not advance the > bound out > >>> of orderness watermark since I am only advancing 1 of the 8 > >>> parallelisms. If I set the entire stream execution > environment > >>> to be of parallelism 1, it will reflect the watermark > in the > >>> context correctly. One more thing is that this behavior > is not > >>> reflected in the Flink Cluster web UI interface. I can > see the > >>> watermark is advancing, but it is not in reality. *That's > >>> causing the inconsistency problem I mentioned in the > other email > >>> I mentioned above. Will this be considered as a bug in > the UI?* > >>> > >>> My current question is, since I have full outer join > operation > >>> before the KeyedProcessFunction here. How can I let the > bound of > >>> orderness watermark / punctuated watermark strategy > work if the > >>> parallelism > 1? It can only update one of the 8 > parallelisms > >>> for the watermark for this onTimer operator. Is this > related to > >>> my Table full outer join operation before this step? > According > >>> to the doc, > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/config.html*table-exec-resource-default-parallelism__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7J4wxLjc0$> > > >>> > > >>> > >>> Default parallelism should be the same like the stream > >>> environment. Why can't I update the watermarks for all 8 > >>> parallelisms? What should I do to enable this function with > >>> Parallelism larger than 1? Thanks. > >>> > >>> First round: (Note the first column of each log row is the > >>> timelag strategy, it is getting updated correctly for all 8 > >>> parallelism, but the other two strategies I mentioned above > >>> can't do that..) > >>> > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266199, > >>> periodicEmitWatermarkTime: 1605047172881, > currentMaxTimestamp: > >>> 1605047187881 (only one of the 8 parallelism for bound > out of > >>> orderness is getting my new watermark) > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266199, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:01,199 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047266198, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> > >>> Second round: (I set the autoWatermark interval to be 5 > seconds) > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 1605047172881, > currentMaxTimestamp: > >>> 1605047187881 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> 14:28:06,200 INFO > >>> org.myorg.quickstart.operator.PeriodicTableOutputWatermarkGenerator > >>> - Emit Watermark: watermark based on system time: > 1605047271200, > >>> periodicEmitWatermarkTime: 0, currentMaxTimestamp: 15000 > >>> > >>> > >>> Best regards, > >>> > >>> Fuyao > >>> > >>> > >>> On Fri, Nov 13, 2020 at 9:03 AM Matthias Pohl > >>> <[hidden email] <mailto:[hidden email]> > <mailto:[hidden email] <mailto:[hidden email]>>> wrote: > >>> > >>> Hi Fuyao, > >>> for your first question about the different behavior > >>> depending on whether you chain the methods or not: > Keep in > >>> mind that you have to save the return value of the > >>> assignTimestampsAndWatermarks method call if you > don't chain > >>> the methods together as it is also shown in [1]. > >>> At least the following example from your first > message is > >>> indicating it: > >>> ``` > >>> retractStream.assignTimestampsAndWatermarks(new > >>> BoRetractStreamTimestampAssigner()); (This is a > deprecated > >>> method) > >>> // instead of: retractStream = > >>> retractStream.assignTimestampsAndWatermarks(new > >>> BoRetractStreamTimestampAssigner()); > >>> retractStream > >>> .keyBy(<key selector>) > >>> .process(new TableOutputProcessFunction()) > >>> .name("ProcessTableOutput") > >>> .uid("ProcessTableOutput") > >>> .addSink(businessObjectSink) > >>> .name("businessObjectSink") > >>> .uid("businessObjectSink") > >>> .setParallelism(1); > >>> ``` > >>> > >>> For your second question about setting the > EventTime I'm > >>> going to pull in Timo from the SDK team as I don't > see an > >>> issue with your code right away. > >>> > >>> Best, > >>> Matthias > >>> > >>> [1] > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$> > > >>> > > >>> > >>> On Wed, Nov 4, 2020 at 10:16 PM Fuyao Li > >>> <[hidden email] > <mailto:[hidden email]> <mailto:[hidden email] > <mailto:[hidden email]>>> > >>> wrote: > >>> > >>> Hi Flink Users and Community, > >>> > >>> For the first part of the question, the 12 hour > time > >>> difference is caused by a time extraction bug > myself. I > >>> can get the time translated correctly now. The > type cast > >>> problem does have some workarounds to solve it.. > >>> > >>> My major blocker right now is the onTimer part > is not > >>> properly triggered. I guess it is caused by > failing to > >>> configure the correct watermarks & timestamp > assigners. > >>> Please give me some insights. > >>> > >>> 1. If I don't chain the > assignTimestampsAndWatermarks() > >>> method in together with keyedBy().. and process().. > >>> method. The context.timestamp() in my > processElement() > >>> function will be null. Is this some expected > behavior? > >>> The Flink examples didn't chain it together. (see > >>> example here: > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*using-watermark-strategies__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JYxd6Sb4$> > > >>> >) > >>> 2. If I use registerEventTimeTimer() in > >>> processElement(). The onTimer method will not be > >>> triggered. However, I can trigger the onTimer > method if > >>> I simply change it to > registerProcessingTimeTimer(). I > >>> am using the settings below in the stream env. > >>> > >>> env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); > >>> env.getConfig().setAutoWatermarkInterval(1000L); > >>> > >>> My code for method the process chain: > >>> retractStream > >>> .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, > >>> Row>>forBoundedOutOfOrderness(Duration.ofSeconds(20)) > >>> .withTimestampAssigner((booleanRowTuple2, timestamp) -> { > >>> Row rowData = > >>> booleanRowTuple2.f1; > >>> LocalDateTime > headerTime = > >>> (LocalDateTime)rowData.getField(3); > >>> LocalDateTime > linesTime = > >>> (LocalDateTime)rowData.getField(7); > >>> > >>> LocalDateTime > >>> latestDBUpdateTime = null; > >>> if (headerTime != > null && > >>> linesTime != null) { > >>> latestDBUpdateTime = > >>> headerTime.isAfter(linesTime) ? headerTime : > linesTime; > >>> } > >>> else { > >>> latestDBUpdateTime = > >>> (headerTime != null) ? headerTime : linesTime; > >>> } > >>> if > (latestDBUpdateTime != > >>> null) { > >>> return > >>> > latestDBUpdateTime.atZone(ZoneId.of("America/Los_Angeles")).toInstant().toEpochMilli(); > >>> } > >>> // In the worst > case, we > >>> use system time instead, which should never be > reached. > >>> return > >>> System.currentTimeMillis(); > >>> })) > >>> // .assignTimestampsAndWatermarks(new > >>> MyWaterStrategy()) // second way to create > watermark, > >>> doesn't work > >>> .keyBy(value -> { > >>> // There could be null > fields for > >>> header invoice_id field > >>> String invoice_id_key = > >>> (String)value.f1.getField(0); > >>> if (invoice_id_key == null) { > >>> invoice_id_key = > >>> (String)value.f1.getField(4); > >>> } > >>> return invoice_id_key; > >>> }) > >>> .process(new > >>> TableOutputProcessFunction()) > >>> .name("ProcessTableOutput") > >>> .uid("ProcessTableOutput") > >>> .addSink(businessObjectSink) > >>> .name("businessObjectSink") > >>> .uid("businessObjectSink") > >>> .setParallelism(1); > >>> > >>> Best regards, > >>> Fuyao > >>> > >>> On Mon, Nov 2, 2020 at 4:53 PM Fuyao Li > >>> <[hidden email] > <mailto:[hidden email]> <mailto:[hidden email] > <mailto:[hidden email]>>> > >>> wrote: > >>> > >>> Hi Flink Community, > >>> > >>> I am doing some research work on Flink > Datastream > >>> and Table API and I meet two major > problems. I am > >>> using Flink 1.11.2, scala version 2.11, > java 8. My > >>> use case looks like this. I plan to write a > data > >>> processing pipeline with two stages. My > goal is to > >>> construct a business object containing > information > >>> from several Kafka streams with a primary > key and > >>> emit the complete business object if such > primary > >>> key doesn't appear in the pipeline for 10 > seconds. > >>> > >>> In the first stage, I first consume three Kafka > >>> streams and transform it to Flink > Datastream using a > >>> deserialization schema containing some type > and date > >>> format transformation, and then I register > these > >>> data streams as Table and do a full outer > join one > >>> by one using Table API. I also add query > >>> configuration for this to avoid excessive > state. The > >>> primary key is also the join key. > >>> > >>> In the second stage, I transform the joined > table to > >>> a retracted stream and put it into > >>> KeyedProcessFunction to generate the > business object > >>> if the business object's primary key is > inactive for > >>> 10 second. > >>> > >>> Is this way of handling the data the suggested > >>> approach? (I understand I can directly > consume kafka > >>> data in Table API. I haven't tried that > yet, maybe > >>> that's better?) Any suggestion is welcomed. > During > >>> implementing this, I meet two major > problems and > >>> several smaller questions under each problem. > >>> > >>> > >>> 1. Some type cast behavior of retracted > streams I > >>> can't explain. > >>> > >>> (1) In the initial stage, I registered some > field as > >>> *java.sql.Date* or > *java.sql.timestamp* following > >>> the examples at > >>> > (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/table/types.html*data-type-extraction__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JsB1tdos$> > > >>> >) > >>> . After join and transform to retracted > stream, it > >>> becomes *java.time.LocalDate* and > >>> *java.time.LocalDateTime* instead. > >>> > >>> For example, when first ingesting the Kafka > streams, > >>> I registerd a attribute in > java.sql.Timestamp type. > >>> > >>> @JsonAlias("ATTRIBUTE1") > >>> private @DataTypeHint(value = "TIMESTAMP(6)", > >>> bridgedTo = java.sql.Timestamp.class) Timestamp > >>> ATTRIBUTE1; > >>> > >>> When I tried to cast the type information > back after > >>> the retracted stream, the code gives me error > >>> information below. > >>> > >>> java.lang.ClassCastException: > >>> java.time.LocalDateTime cannot be cast to > >>> java.sql.Timestamp > >>> > >>> Maybe I should use toAppendStream instead since > >>> append stream could register type > information, but > >>> toRetractedStream can't do that? > >>> > (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/table/common.html*convert-a-table-into-a-datastream-or-dataset__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JQ99YqY0$> > > >>> >) > >>> > >>> My work around is to cast it to > LocalDateTime first > >>> and extract the epoch time, this doesn't > seem to be > >>> a final solution. > >>> > >>> (2) During timestamp conversion, the Flink to > >>> retracted stream seems to lost the AM/PM > information > >>> in the stream and causing a 12 hour > difference if it > >>> is PM. > >>> > >>> I use joda time to do some timestamp > conversion in > >>> the first deserialization stage, my pattern > looks > >>> like this. "a" means AM/PM information > >>> > >>> DateTimeFormatter format3 = > >>> DateTimeFormat.forPattern("dd-MMM-yy > HH.mm.ss.SSSSSS > >>> a").withZone(DateTimeZone.getDefault()); > >>> > >>> After the retracted stream, the AM/PM > information is > >>> not preserved. > >>> > >>> > >>> 2. My onTimer method in > KeyedProcessFunction can not > >>> be triggered when I scheduled a event timer > timer. > >>> > >>> I am using event time in my code. I am new to > >>> configure watermarks and I might miss > something to > >>> configure it correctly. I also tried to > register a > >>> processing time, it could enter and produce > some > >>> results. > >>> > >>> I am trying to follow the example here: > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/operators/process_function.html*example__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JNMi_YMc$> > > >>> > > >>> > >>> My onTimer method looks like this and the > scheduled > >>> event doesn't happen.. > >>> > >>> In processElement(): > >>> > >>> > context.timerService().registerEventTimeTimer(current.getLastModifiedTime() > >>> + 10000); > >>> > >>> My onTimer function > >>> > >>> @Override > >>> public void onTimer(long timestamp, > >>> OnTimerContext ctx, Collector<BusinessObject> > >>> collector) throws Exception { > >>> TestBusinessObjectState result = > >>> testBusinessObjectState.value(); > >>> log.info <http://log.info> > >>> > <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$ > <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$> > > >>> >("Inside onTimer Method, > >>> current key: {}, timestamp: {}, last > modified time: > >>> {}", ctx.getCurrentKey(), timestamp, > >>> result.getLastModifiedTime()); > >>> > >>> // check if this is an outdated > timer or > >>> the latest timer > >>> if (timestamp >= > >>> result.getLastModifiedTime() + 10000) { > >>> // emit the state on timeout > >>> log.info <http://log.info> > >>> > <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$ > <https://urldefense.com/v3/__http://log.info/__;!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JSt0BaYQ$> > > >>> >("Collecting a business > >>> object, {}", > result.getBusinessObject().toString()); > >>> collector.collect(result.getBusinessObject()); > >>> > >>> cleanUp(ctx); > >>> } > >>> } > >>> > >>> private void cleanUp(Context ctx) throws > >>> Exception { > >>> Long timer = > >>> testBusinessObjectState.value().getLastModifiedTime(); > >>> ctx.timerService().deleteEventTimeTimer(timer); > >>> testBusinessObjectState.clear(); > >>> } > >>> > >>> > >>> (1) When I assign the timestamp and watermarks > >>> outside the process() method chain. The > >>> "context.timestamp()" will be null. If I put it > >>> inside the chain, it won't be null. Is this the > >>> expected behavior? In the null case, the > strange > >>> thing is that, surprisingly, I can collect the > >>> business object immediately without a > designed 10 > >>> second waiting time... This shouldn't happen, > >>> right...? The processing timer also seems > to work. > >>> The code can enter the on timer method. > >>> > >>> retractStream.assignTimestampsAndWatermarks(new > >>> BoRetractStreamTimestampAssigner()); (This is a > >>> deprecated method) > >>> > >>> retractStream > >>> .keyBy(<key selector>) > >>> .process(new TableOutputProcessFunction()) > >>> .name("ProcessTableOutput") > >>> .uid("ProcessTableOutput") > >>> .addSink(businessObjectSink) > >>> .name("businessObjectSink") > >>> .uid("businessObjectSink") > >>> .setParallelism(1); > >>> > >>> (2) For watermarks configuration. I use an > field in > >>> the retracted stream as the event time. > This time is > >>> usually 15-20 seconds before current time. > >>> > >>> In my environment, I have done some > settings for > >>> streaming env based on information here( > >>> > https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$> > > >>> >). > >>> My event doesn't always come, so I think I > need to > >>> set auto watermark interval to let the > event timer > >>> on timer works correctly. I have added the > code > >>> below. > >>> > >>> env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime); > >>> env.getConfig().setAutoWatermarkInterval(1000L); > >>> > >>> 1> Which kind of watermark strategy should > I use? > >>> General BoundOutofOrderness or Watermark > generator? > >>> > >>> I tried to write a Watermark generator and > I just > >>> don't how to apply it to the stream > correctly. The > >>> documentation doesn't explain very clearly. > My code > >>> looks like below and it doesn't work. > >>> > >>> assign part: > >>> > >>> > .assignTimestampsAndWatermarks(WatermarkStrategy.forGenerator((WatermarkGeneratorSupplier<Tuple2<Boolean, > >>> Row>>) context -> new > >>> TableBoundOutofOrdernessGenerator())) > >>> > >>> watermark generater: > >>> > >>> I just assign the event time attribute > following the > >>> example in the doc. > >>> > (https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$> > > >>> > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$ > <https://urldefense.com/v3/__https://ci.apache.org/projects/flink/flink-docs-release-1.11/dev/event_timestamps_watermarks.html*writing-a-periodic-watermarkgenerator__;Iw!!GqivPVa7Brio!ItHlGfYT1dLQeAolQoFNfXPN876842lnF4hOE7cxmmTJY4tJkXUmkz7JqRHnniA$> > > >>> >) > >>> > >>> 2> I also tried to use the static method in > Water > >>> Strategy. The syntax is correct, but I meet > the same > >>> problem in 2.(1). > >>> > >>> .assignTimestampsAndWatermarks(WatermarkStrategy.<Tuple2<Boolean, > >>> Row>>forBoundedOutOfOrderness(Duration.ofSeconds(15)) > >>> .withTimestampAssigner((booleanRowTuple2, timestamp) > >>> -> { > >>> <Select a > event time > >>> attribute in the booleanRowTuple2> > >>> })) > >>> > >>> > >>> (3) For the retracted datastream, do I need to > >>> explicitly attach it to the stream > environment? I > >>> think it is done by default, right? Just > want to > >>> confirm it. I do have the env.execute() at > the end > >>> of the code. > >>> > >>> I understand this is a lot of questions, > thanks a > >>> lot for your patience to look through my > email! If > >>> there is anything unclear, please reach out > to me. > >>> Thanks! > >>> > >>> > >>> Best regards, > >>> > >>> Fuyao Li > >>> > >> > |
Free forum by Nabble | Edit this page |