Streaming and batch jobs together

[Flavio Pompermaier]

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio

[Kostas Kloudas]

Hi Flavio,

If I understand correctly, you have a set of keys which evolves in two ways:

keys may be added/deleted

values associated with the keys can also be updated.

If this is the case, you can use a streaming job that:

1. has as a source the stream of events (ADD/DELETE…)

2. has after the source a flatmap that reads the initial set of keys in the open() method

as soon as it sees the first element from the stream, it flushes out all the keys it has read in the previous step, and then the element

(the previous step will guarantee that you will see the set of keys before the change events, if this is important. If not, then ignore that step)

3. then there is a keyBy that partitions the elements by key (e.g. countryId)

4. then a process function that takes care of deduplicating the initial keys and updating the states associated with each key.

Does this sound like a good starting point?

Thanks,

Kostas

On May 8, 2018, at 12:22 PM, Flavio Pompermaier <[hidden email]> wrote:

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio

[Michael Latta]

If you use a KeyedStream you can group records by key (city) and then use a RichFlatMap to aggregate state in a MapState or ListState per key.  You can then have that operator publish the updated results as a new aggregated record, or send it to a database or such as you see fit.

Michael

On May 8, 2018, at 4:22 AM, Flavio Pompermaier <[hidden email]> wrote:

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio

[Flavio Pompermaier]

Thanks! Both solutions are reasonable but ehat abiut max state size (per key)?is there any suggested database/nosql store to use?

On Tue, 8 May 2018, 18:09 TechnoMage, <[hidden email]> wrote:

If you use a KeyedStream you can group records by key (city) and then use a RichFlatMap to aggregate state in a MapState or ListState per key.  You can then have that operator publish the updated results as a new aggregated record, or send it to a database or such as you see fit.

Michael

On May 8, 2018, at 4:22 AM, Flavio Pompermaier <[hidden email]> wrote:

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio

[Kostas Kloudas]

Hi Flavio,

Flink has no inherent limitations as far as state size is concerned, apart from the fact that the state associated to a single key

(not the total state) should fit in memory. For production use, it is also advised to use the RocksDB state backend, as this will

allow you to spill on disk when the state grows too large.

Now for recommended DB/no-sql store, there is no recommendation from my part. It depends on what you and your team are 

more familiar with. I suppose you are talking about sink, right? In this case, it also depends on what will optimize your batch jobs 

that will read the updated dataset.

Thanks,

Kostas

On May 8, 2018, at 10:40 PM, Flavio Pompermaier <[hidden email]> wrote:

Thanks! Both solutions are reasonable but ehat abiut max state size (per key)?is there any suggested database/nosql store to use?

On Tue, 8 May 2018, 18:09 TechnoMage, <[hidden email]> wrote:

If you use a KeyedStream you can group records by key (city) and then use a RichFlatMap to aggregate state in a MapState or ListState per key.  You can then have that operator publish the updated results as a new aggregated record, or send it to a database or such as you see fit.

Michael

On May 8, 2018, at 4:22 AM, Flavio Pompermaier <[hidden email]> wrote:

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio

[Flavio Pompermaier]

Ok, thanks for the clarification Kostas. What about multiple jobs running at the same time?

On Wed, 9 May 2018, 14:39 Kostas Kloudas, <[hidden email]> wrote:

Hi Flavio,

Flink has no inherent limitations as far as state size is concerned, apart from the fact that the state associated to a single key

(not the total state) should fit in memory. For production use, it is also advised to use the RocksDB state backend, as this will

allow you to spill on disk when the state grows too large.

Now for recommended DB/no-sql store, there is no recommendation from my part. It depends on what you and your team are 

more familiar with. I suppose you are talking about sink, right? In this case, it also depends on what will optimize your batch jobs 

that will read the updated dataset.

Thanks,

Kostas

On May 8, 2018, at 10:40 PM, Flavio Pompermaier <[hidden email]> wrote:

Thanks! Both solutions are reasonable but ehat abiut max state size (per key)?is there any suggested database/nosql store to use?

On Tue, 8 May 2018, 18:09 TechnoMage, <[hidden email]> wrote:

If you use a KeyedStream you can group records by key (city) and then use a RichFlatMap to aggregate state in a MapState or ListState per key.  You can then have that operator publish the updated results as a new aggregated record, or send it to a database or such as you see fit.

Michael

On May 8, 2018, at 4:22 AM, Flavio Pompermaier <[hidden email]> wrote:

Hi all,

I'd like to introduce in our pipeline an efficient way to aggregate incoming data around an entity.

We have basically new incoming facts that are added (but also removed potentially) to an entity (by id). For example, when we receive a new name of a city we add this name to the known names of that city id (if the first field of the tuple is ADD, if it is DEL we remove it).

At the moment we use batch job to generate an initial version of the entities, another job that add facts to this initial version of the entities, and another one that merges the base and the computed data. This is somehow very inefficient in terms of speed and disk space (because every step requires to materialize the data on the disk).

I was wondering whether Flink could help here or not...there are a couple of requirements that make things very complicated:

• states could be potentially large (a lot of data related to an entity). Is there any limitation about the size of the states?
• data must be readable by a batch job. If I'm not wrong this could be easily solved flushing data periodically to an external sink (like HBase or similar)
• how to keep the long-running stream job up and run a batch job at the same time? Will this be possible after Flip-6?
• how to add ingest new data? Do I really need Kafka or can I just add new datasets to a staging HDFS dir (and move them to another dir once ingested)?

Best,

Flavio