(DEPRECATED) Apache Flink User Mailing List archive.

[Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

classic Classic list List threaded Threaded
♦
♦
 10 messages Options
Jungtaek Lim
Reply | Threaded
Open this post in threaded view
♦
♦
|

[Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Jungtaek Lim
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Jungtaek Lim
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Jungtaek Lim
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Chesnay Schepler
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Chesnay Schepler
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.

On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Fabian Hueske-2
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Fabian Hueske-2
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Jungtaek Lim
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Jungtaek Lim
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Fabian Hueske-2
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Fabian Hueske-2
Hi Jungtaek,

If it is "only" about the missing support to parse a string as timestamp, you could also implement a custom TimestampExtractor that works similar to the ExistingField extractor [1].
You would need to adjust a few things and use the expression "Cast(Cast('tsString, SqlTimeTypeInfo.TIMESTAMP), Types.LONG)" to convert the String to a Long.
So far this works only if the date is formatted like "2018-05-28 12:34:56.000"

Regarding the side outputs, these would not be handled as results but just redirect late records into separate data streams. We would offer a configuration to write them to a sink like HDFS or Kafka.

Best, Fabian


2018-07-04 11:54 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Jungtaek Lim
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Jungtaek Lim
Thanks again Fabian for providing nice suggestion!

Finally I got it working with applying your suggestion. Couple of tricks was needed:

1. I had to apply a hack (create new TimestampExtractor class to package org.apache.flink.blabla...) since Expression.resultType is defined as "package private" for flink. I feel adjusting scope of Explain's methods (at least resultType) to "public" would help on implementing custom TimestampExtractor in users' side: please let me know your thought about this. If you think it makes sense, I will file an issue and submit a PR, or initiate a new thread in dev mailing list to discuss it if the step is recommend.

2. To ensure KafkaTableSource's verification of rowtime field type, the type of field (here in "eventTime") should be defined as SQL_TIMESTAMP whereas the type of field in JSON should be defined as STRING.
Kafka010JsonTableSource.builder()
  .forTopic(topic)
  .withSchema(TableSchema.builder()
    .field("eventTime", Types.SQL_TIMESTAMP)
    .build())
  .forJsonSchema(TableSchema.builder()
    .field("eventTime", Types.STRING)
    .build())
  .withKafkaProperties(prop)
  .withRowtimeAttribute(
    "eventTime",
    new IsoDateStringAwareExistingField("eventTime"),
    new BoundedOutOfOrderTimestamps(Time.minutes(1).toMilliseconds)
  )
  .build()
Thanks again!
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 8:18, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

If it is "only" about the missing support to parse a string as timestamp, you could also implement a custom TimestampExtractor that works similar to the ExistingField extractor [1].
You would need to adjust a few things and use the expression "Cast(Cast('tsString, SqlTimeTypeInfo.TIMESTAMP), Types.LONG)" to convert the String to a Long.
So far this works only if the date is formatted like "2018-05-28 12:34:56.000"

Regarding the side outputs, these would not be handled as results but just redirect late records into separate data streams. We would offer a configuration to write them to a sink like HDFS or Kafka.

Best, Fabian


2018-07-04 11:54 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Fabian Hueske-2
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Fabian Hueske-2
Hi,

Glad you could get it to work! That's great :-)

Regarding you comments:

1) Yes, I think we should make resultType() public. Please open a Jira issue and describe your use case.
Btw. would you like to contribute your TimestampExtractor to Flink (or even a more generic one that allows to configure the format of the timestamp string)? There is FLINK-8094 [1].
2) This is "expected" because you define two different schemas, the JSON schema which defines how to read the data and the Table schema that defines how it is exposed to the Table API / SQL.

Thanks, Fabian


2018-07-04 14:52 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks again Fabian for providing nice suggestion!

Finally I got it working with applying your suggestion. Couple of tricks was needed:

1. I had to apply a hack (create new TimestampExtractor class to package org.apache.flink.blabla...) since Expression.resultType is defined as "package private" for flink. I feel adjusting scope of Explain's methods (at least resultType) to "public" would help on implementing custom TimestampExtractor in users' side: please let me know your thought about this. If you think it makes sense, I will file an issue and submit a PR, or initiate a new thread in dev mailing list to discuss it if the step is recommend.

2. To ensure KafkaTableSource's verification of rowtime field type, the type of field (here in "eventTime") should be defined as SQL_TIMESTAMP whereas the type of field in JSON should be defined as STRING.
Kafka010JsonTableSource.builder()
  .forTopic(topic)
  .withSchema(TableSchema.builder()
    .field("eventTime", Types.SQL_TIMESTAMP)
    .build())
  .forJsonSchema(TableSchema.builder()
    .field("eventTime", Types.STRING)
    .build())
  .withKafkaProperties(prop)
  .withRowtimeAttribute(
    "eventTime",
    new IsoDateStringAwareExistingField("eventTime"),
    new BoundedOutOfOrderTimestamps(Time.minutes(1).toMilliseconds)
  )
  .build()
Thanks again!
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 8:18, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

If it is "only" about the missing support to parse a string as timestamp, you could also implement a custom TimestampExtractor that works similar to the ExistingField extractor [1].
You would need to adjust a few things and use the expression "Cast(Cast('tsString, SqlTimeTypeInfo.TIMESTAMP), Types.LONG)" to convert the String to a Long.
So far this works only if the date is formatted like "2018-05-28 12:34:56.000"

Regarding the side outputs, these would not be handled as results but just redirect late records into separate data streams. We would offer a configuration to write them to a sink like HDFS or Kafka.

Best, Fabian


2018-07-04 11:54 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Jungtaek Lim
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Jungtaek Lim
Thanks Fabian, filed FLINK-9742 [1].

I'll submit a PR for FLINK-8094 via providing my TimestampExtractor. The implementation is also described as FLINK-9742. I'll start with current implementation which just leverages automatic cast from STRING to SQL_TIMESTAMP, but we could improve it from PR. Feedbacks are welcome!

Btw, maybe need to initiate from another thread, but I also had to struggle to find a solution to convert table to retract stream. Looks like "implicit conversion" comes into play prior to toRetractStream and raise error. outTable is the result of "distinct" which looks like requiring retract mode. (Not even easy for me to know I should provide implicit  TypeInformation for Row, but I'm fairly new to Scala so it's just me.)
// below doesn't work as below line implicitly converts table as 'append stream'
// via org.apache.flink.table.api.scala.package$.table2RowDataStream
// though we are calling toRetractStream
//outTable.toRetractStream[Row](outTable.dataType).print()

implicit val typeInfo = Types.ROW(outTable.getSchema.getColumnNames,
  outTable.getSchema.getTypes)
tableEnv.toRetractStream[Row](outTable).print()

Thanks again,

2018년 7월 4일 (수) 오후 10:03, Fabian Hueske <[hidden email]>님이 작성:
Hi,

Glad you could get it to work! That's great :-)

Regarding you comments:

1) Yes, I think we should make resultType() public. Please open a Jira issue and describe your use case.
Btw. would you like to contribute your TimestampExtractor to Flink (or even a more generic one that allows to configure the format of the timestamp string)? There is FLINK-8094 [1].
2) This is "expected" because you define two different schemas, the JSON schema which defines how to read the data and the Table schema that defines how it is exposed to the Table API / SQL.

Thanks, Fabian


2018-07-04 14:52 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks again Fabian for providing nice suggestion!

Finally I got it working with applying your suggestion. Couple of tricks was needed:

1. I had to apply a hack (create new TimestampExtractor class to package org.apache.flink.blabla...) since Expression.resultType is defined as "package private" for flink. I feel adjusting scope of Explain's methods (at least resultType) to "public" would help on implementing custom TimestampExtractor in users' side: please let me know your thought about this. If you think it makes sense, I will file an issue and submit a PR, or initiate a new thread in dev mailing list to discuss it if the step is recommend.

2. To ensure KafkaTableSource's verification of rowtime field type, the type of field (here in "eventTime") should be defined as SQL_TIMESTAMP whereas the type of field in JSON should be defined as STRING.
Kafka010JsonTableSource.builder()
  .forTopic(topic)
  .withSchema(TableSchema.builder()
    .field("eventTime", Types.SQL_TIMESTAMP)
    .build())
  .forJsonSchema(TableSchema.builder()
    .field("eventTime", Types.STRING)
    .build())
  .withKafkaProperties(prop)
  .withRowtimeAttribute(
    "eventTime",
    new IsoDateStringAwareExistingField("eventTime"),
    new BoundedOutOfOrderTimestamps(Time.minutes(1).toMilliseconds)
  )
  .build()
Thanks again!
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 8:18, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

If it is "only" about the missing support to parse a string as timestamp, you could also implement a custom TimestampExtractor that works similar to the ExistingField extractor [1].
You would need to adjust a few things and use the expression "Cast(Cast('tsString, SqlTimeTypeInfo.TIMESTAMP), Types.LONG)" to convert the String to a Long.
So far this works only if the date is formatted like "2018-05-28 12:34:56.000"

Regarding the side outputs, these would not be handled as results but just redirect late records into separate data streams. We would offer a configuration to write them to a sink like HDFS or Kafka.

Best, Fabian


2018-07-04 11:54 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Fabian Hueske-2
Reply | Threaded
Open this post in threaded view
♦
♦
|

Re: [Table API/SQL] Finding missing spots to resolve why 'no watermark' is presented in Flink UI

Fabian Hueske-2
Hi,

Thanks for the PR! I'll have a look at it later today.

The problem of the retraction stream conversion is probably that the return type is a Tuple2[Boolean, Row].
The boolean flag indicates whether the row is added or retracted.

Best, Fabian

2018-07-04 15:38 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Fabian, filed FLINK-9742 [1].

I'll submit a PR for FLINK-8094 via providing my TimestampExtractor. The implementation is also described as FLINK-9742. I'll start with current implementation which just leverages automatic cast from STRING to SQL_TIMESTAMP, but we could improve it from PR. Feedbacks are welcome!

Btw, maybe need to initiate from another thread, but I also had to struggle to find a solution to convert table to retract stream. Looks like "implicit conversion" comes into play prior to toRetractStream and raise error. outTable is the result of "distinct" which looks like requiring retract mode. (Not even easy for me to know I should provide implicit  TypeInformation for Row, but I'm fairly new to Scala so it's just me.)
// below doesn't work as below line implicitly converts table as 'append stream'
// via org.apache.flink.table.api.scala.package$.table2RowDataStream
// though we are calling toRetractStream
//outTable.toRetractStream[Row](outTable.dataType).print()

implicit val typeInfo = Types.ROW(outTable.getSchema.getColumnNames,
  outTable.getSchema.getTypes)
tableEnv.toRetractStream[Row](outTable).print()

Thanks again,

2018년 7월 4일 (수) 오후 10:03, Fabian Hueske <[hidden email]>님이 작성:
Hi,

Glad you could get it to work! That's great :-)

Regarding you comments:

1) Yes, I think we should make resultType() public. Please open a Jira issue and describe your use case.
Btw. would you like to contribute your TimestampExtractor to Flink (or even a more generic one that allows to configure the format of the timestamp string)? There is FLINK-8094 [1].
2) This is "expected" because you define two different schemas, the JSON schema which defines how to read the data and the Table schema that defines how it is exposed to the Table API / SQL.

Thanks, Fabian


2018-07-04 14:52 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks again Fabian for providing nice suggestion!

Finally I got it working with applying your suggestion. Couple of tricks was needed:

1. I had to apply a hack (create new TimestampExtractor class to package org.apache.flink.blabla...) since Expression.resultType is defined as "package private" for flink. I feel adjusting scope of Explain's methods (at least resultType) to "public" would help on implementing custom TimestampExtractor in users' side: please let me know your thought about this. If you think it makes sense, I will file an issue and submit a PR, or initiate a new thread in dev mailing list to discuss it if the step is recommend.

2. To ensure KafkaTableSource's verification of rowtime field type, the type of field (here in "eventTime") should be defined as SQL_TIMESTAMP whereas the type of field in JSON should be defined as STRING.
Kafka010JsonTableSource.builder()
  .forTopic(topic)
  .withSchema(TableSchema.builder()
    .field("eventTime", Types.SQL_TIMESTAMP)
    .build())
  .forJsonSchema(TableSchema.builder()
    .field("eventTime", Types.STRING)
    .build())
  .withKafkaProperties(prop)
  .withRowtimeAttribute(
    "eventTime",
    new IsoDateStringAwareExistingField("eventTime"),
    new BoundedOutOfOrderTimestamps(Time.minutes(1).toMilliseconds)
  )
  .build()
Thanks again!
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 8:18, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

If it is "only" about the missing support to parse a string as timestamp, you could also implement a custom TimestampExtractor that works similar to the ExistingField extractor [1].
You would need to adjust a few things and use the expression "Cast(Cast('tsString, SqlTimeTypeInfo.TIMESTAMP), Types.LONG)" to convert the String to a Long.
So far this works only if the date is formatted like "2018-05-28 12:34:56.000"

Regarding the side outputs, these would not be handled as results but just redirect late records into separate data streams. We would offer a configuration to write them to a sink like HDFS or Kafka.

Best, Fabian


2018-07-04 11:54 GMT+02:00 Jungtaek Lim <[hidden email]>:
Thanks Chesnay! Great news to hear. I'll try out with latest master branch.

Thanks Fabian for providing the docs! 

I guess I already tried out with KafkaJsonTableSource and failed back to custom TableSource since the type of rowtime field is string unfortunately, and I needed to parse and map to new SQL timestamp field in order to use it to rowtime attribute.

I guess JSON -> table fields mapping is provided only for renaming, and "withRowtimeAttribute" doesn't help defining new field to use it as rowtime.

Are there better approaches on this scenario? Or would we be better to assume the type of rowtime field is always timestamp?

Btw, providing late-data side output in Table API might be just a matter of how to define it correctly (not a technical or syntactic issue), though providing in SQL might be tricky (as the semantic of SQL query is not for multiple outputs).

Thanks,
Jungtaek Lim (HeartSaVioR)

2018년 7월 4일 (수) 오후 5:49, Fabian Hueske <[hidden email]>님이 작성:
Hi Jungtaek,

Flink 1.5.0 features a TableSource for Kafka and JSON [1], incl. timestamp & watemark generation [2].
It would be great if you could let us know, if that addresses your use case and if not what's missing or not working.

So far Table API / SQL does not have support for late-data side outputs. However, that's on the road map. The idea is to filter streams during ingestion for late events and passing them to a side output.
Currently, operators just drop late events.

Best, Fabian


2018-07-04 10:39 GMT+02:00 Chesnay Schepler <[hidden email]>:
The watermark display in the UI is bugged in 1.5.0.

It is fixed on master and the release-1.5 branch, and will be included in 1.5.1 that is slated to be released next week.


On 04.07.2018 10:22, Jungtaek Lim wrote:
Sorry I forgot to mention the version: Flink 1.5.0, and I ran the app in IntelliJ, not tried from cluster.

2018년 7월 4일 (수) 오후 5:15, Jungtaek Lim <[hidden email]>님이 작성:
Hi Flink users,

I'm new to Flink and trying to evaluate couple of streaming frameworks via implementing same apps.

While implementing apps with both Table API and SQL, I found there's 'no watermark' presented in Flink UI, whereas I had been struggling to apply row time attribute.

For example, below is one of TableSource implementation which wraps DataStream reading from Kafka.


(Actually I ended up implementing TableSource to address adding rowtime attribute as well as reading and parsing JSON. I'd be really happy if someone can guide a way to get rid of needed of custom implementation of TableSource.)

and below is one of app I implemented:


Btw, I'm about to experiment side-output with late events, but is it possible to leverage side-output with Table API / SQL? Looks like DataStream exposes late events only when it's converted to AllWindowedStream.

Thanks in advance!

Best Regards,
Jungtaek Lim (HeartSaVioR)
Free forum by Nabble Edit this page