Elegantly sharing state in a streaming environment

[leon_mclare]

Hello Flink team,

How can i partition and share static state among instances of a streaming operator?

I have a huge list of keys and values, which are used to filter tuples in a stream. The list does not change. Currently i am sharing the list with each operator instance via the constructor, although only a subset of the list is required per operator (the assignment of subset to operator instance is known). I cannot use DataSet based functions in a streaming execution environment to assign sub lists. I also cannot use DataStream based partitioning functions as the list is static, i.e. not a DataStream. The dilemma exists as i am mixing static (DataSet type) content with streaming content. Is there any other approach aside from using an additional tool (e.g. distributed cache)?

Thanks in advance.

Regards
Leon

[Philippe CAPARROY]

Just transform the list in a DataStream. A datastream can be finite.

One solution, in the context of a Streaming environment is to use Kafka, or any other distributed broker, although Flink ships with a KafkaSource.

 

1)Create a Kafka Topic dedicated to your list of key/values. Inject your values into this topic, partitionned by the keys. So that you recover the keys in Flink.

 

2) Create a source for the stream of tuple your analysing -> output1 (Tuples).

 

3) Create a KafkaSource, and parse/recover your key value pairs from this source (e.g a first map operator) : map1 -> output 2 (K,V), then :

 

 

 

                 a)  If you need all key/Value pairs at each operator :  broadcast all partitions from the output 1 to the analysis operator

 

                  b) if you dont need all key/values pairs, just chain output1 to the analysis operator. Partitioning of K,V pairs will depend on Kafka partitioning strategy, and can be controlled in Flink      anyway.

 

4) The analysis operator :  will perform a RichCoFlatMapFunction, and can be Checkpointed.

When receiving K,V pairs from output2, store them in a local state.

When receiving tuple, should be able to to filter with the help of the local state, and propagate downstream or not.

 

 

 

 

 

 

 

 

 

 

 

> Message du 30/05/16 13:41
> De : [hidden email]
> A : "User" <[hidden email]>
> Copie à :
> Objet : Elegantly sharing state in a streaming environment
>
>Hello Flink team,

How can i partition and share static state among instances of a streaming operator?

I have a huge list of keys and values, which are used to filter tuples in a stream. The list does not change. Currently i am sharing the list with each operator instance via the constructor, although only a subset of the list is required per operator (the assignment of subset to operator instance is known). I cannot use DataSet based functions in a streaming execution environment to assign sub lists. I also cannot use DataStream based partitioning functions as the list is static, i.e. not a DataStream. The dilemma exists as i am mixing static (DataSet type) content with streaming content. Is there any other approach aside from using an additional tool (e.g. distributed cache)?

Thanks in advance.

Regards
Leon

[leon_mclare]

Dear Philippe,

that is exactly what i need. Thank you for the concise explanation.

This approach is excellent, as it also permits the values to be easily updated externally.

Kind regards
Leon

30. May 2016 14:31 by [hidden email]:

Just transform the list in a DataStream. A datastream can be finite.

One solution, in the context of a Streaming environment is to use Kafka, or any other distributed broker, although Flink ships with a KafkaSource.

 

1)Create a Kafka Topic dedicated to your list of key/values. Inject your values into this topic, partitionned by the keys. So that you recover the keys in Flink.

 

2) Create a source for the stream of tuple your analysing -> output1 (Tuples).

 

3) Create a KafkaSource, and parse/recover your key value pairs from this source (e.g a first map operator) : map1 -> output 2 (K,V), then :

 

 

 

                 a)  If you need all key/Value pairs at each operator :  broadcast all partitions from the output 1 to the analysis operator

 

                  b) if you dont need all key/values pairs, just chain output1 to the analysis operator. Partitioning of K,V pairs will depend on Kafka partitioning strategy, and can be controlled in Flink      anyway.

 

4) The analysis operator :  will perform a RichCoFlatMapFunction, and can be Checkpointed.

When receiving K,V pairs from output2, store them in a local state.

When receiving tuple, should be able to to filter with the help of the local state, and propagate downstream or not.

 

 

 

 

 

 

 

 

 

 

 

> Message du 30/05/16 13:41
> De : [hidden email]
> A : "User" <[hidden email]>
> Copie à :
> Objet : Elegantly sharing state in a streaming environment
>
>Hello Flink team,

How can i partition and share static state among instances of a streaming operator?

I have a huge list of keys and values, which are used to filter tuples in a stream. The list does not change. Currently i am sharing the list with each operator instance via the constructor, although only a subset of the list is required per operator (the assignment of subset to operator instance is known). I cannot use DataSet based functions in a streaming execution environment to assign sub lists. I also cannot use DataStream based partitioning functions as the list is static, i.e. not a DataStream. The dilemma exists as i am mixing static (DataSet type) content with streaming content. Is there any other approach aside from using an additional tool (e.g. distributed cache)?

Thanks in advance.

Regards
Leon

[Ufuk Celebi]

Aljoscha is working to properly expose this in Flink. The design
document is here:
https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-MKQYN3m4/edit#heading=h.pqg5z6g0mjm7

On Mon, May 30, 2016 at 2:31 PM, Philippe CAPARROY
<[hidden email]> wrote:

>
> Just transform the list in a DataStream. A datastream can be finite.
>
>
> One solution, in the context of a Streaming environment is to use Kafka, or
> any other distributed broker, although Flink ships with a KafkaSource.
>
>
>
> 1)Create a Kafka Topic dedicated to your list of key/values. Inject your
> values into this topic, partitionned by the keys. So that you recover the
> keys in Flink.
>
>
>
> 2) Create a source for the stream of tuple your analysing -> output1
> (Tuples).
>
>
>
> 3) Create a KafkaSource, and parse/recover your key value pairs from this
> source (e.g a first map operator) : map1 -> output 2 (K,V), then :
>
>
>
>
>
>
>
>                  a)  If you need all key/Value pairs at each operator :
> broadcast all partitions from the output 1 to the analysis operator
>
>
>
>                   b) if you dont need all key/values pairs, just chain
> output1 to the analysis operator. Partitioning of K,V pairs will depend on
> Kafka partitioning strategy, and can be controlled in Flink      anyway.
>
>
>
> 4) The analysis operator :  will perform a RichCoFlatMapFunction, and can be
> Checkpointed.
>
> When receiving K,V pairs from output2, store them in a local state.
>
> When receiving tuple, should be able to to filter with the help of the local
> state, and propagate downstream or not.
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>> Message du 30/05/16 13:41
>> De : [hidden email]
>> A : "User" <[hidden email]>
>> Copie à :
>> Objet : Elegantly sharing state in a streaming environment
>
>>
>>Hello Flink team,
>
> How can i partition and share static state among instances of a streaming
> operator?
>
> I have a huge list of keys and values, which are used to filter tuples in a
> stream. The list does not change. Currently i am sharing the list with each
> operator instance via the constructor, although only a subset of the list is
> required per operator (the assignment of subset to operator instance is
> known). I cannot use DataSet based functions in a streaming execution
> environment to assign sub lists. I also cannot use DataStream based
> partitioning functions as the list is static, i.e. not a DataStream. The
> dilemma exists as i am mixing static (DataSet type) content with streaming
> content. Is there any other approach aside from using an additional tool
> (e.g. distributed cache)?
>
> Thanks in advance.
>
> Regards
> Leon
>
>
>