Spark DataSets and multiple write(.) calls

classic Classic list List threaded Threaded
7 messages Options
Reply | Threaded
Open this post in threaded view
|

Spark DataSets and multiple write(.) calls

Rico B.
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]

Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

Magnus Nilsson

Magnus Nilsson

9:43 AM (0 minutes ago)
to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]

Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

Rico B.

Thanks for your advise. But I'm using Batch processing. Does anyone have a solution for the batch processing case?

Best,

Rico.


Am 19.11.2018 um 09:43 schrieb Magnus Nilsson:

Magnus Nilsson

9:43 AM (0 minutes ago)

to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]



Virenfrei. www.avast.com


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]

pEpkey.asc (1K) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

Vadim Semenov-2
You can use checkpointing, in this case Spark will write out an rdd to whatever destination you specify, and then the RDD can be reused from the checkpointed state avoiding recomputing.

On Mon, Nov 19, 2018 at 7:51 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Thanks for your advise. But I'm using Batch processing. Does anyone have a solution for the batch processing case?

Best,

Rico.


Am 19.11.2018 um 09:43 schrieb Magnus Nilsson:

Magnus Nilsson

9:43 AM (0 minutes ago)

to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]



Virenfrei. www.avast.com

---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]


--
Sent from my iPhone
Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

Rico B.

Hi!

Thanks Vadim for your answer. But this would be like caching the dataset, right? Or is checkpointing faster then persisting to memory or disk?

I attach a pdf of my dataflow program. If I could compute the output of outputs 1-5 in parallel the output of flatmap1 and groupBy could be reused, avoiding to write to disk (at least until the grouping).

Any other ideas or proposals?

Best,

Rico.


Am 19.11.2018 um 19:12 schrieb Vadim Semenov:
You can use checkpointing, in this case Spark will write out an rdd to whatever destination you specify, and then the RDD can be reused from the checkpointed state avoiding recomputing.

On Mon, Nov 19, 2018 at 7:51 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Thanks for your advise. But I'm using Batch processing. Does anyone have a solution for the batch processing case?

Best,

Rico.


Am 19.11.2018 um 09:43 schrieb Magnus Nilsson:

Magnus Nilsson

9:43 AM (0 minutes ago)

to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]



Virenfrei. www.avast.com

---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]


--
Sent from my iPhone



---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]

ExecutionDAG.pdf (62K) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

smikesh

You can also cache the data frame on disk, if it does not fit into memory. 
An alternative would be to write out data frame as parquet and then read it, you can check if in this case the whole pipeline works faster as with the standard cache. 

Best,
Michael


On Tue, Nov 20, 2018 at 9:14 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Hi!

Thanks Vadim for your answer. But this would be like caching the dataset, right? Or is checkpointing faster then persisting to memory or disk?

I attach a pdf of my dataflow program. If I could compute the output of outputs 1-5 in parallel the output of flatmap1 and groupBy could be reused, avoiding to write to disk (at least until the grouping).

Any other ideas or proposals?

Best,

Rico.


Am 19.11.2018 um 19:12 schrieb Vadim Semenov:
You can use checkpointing, in this case Spark will write out an rdd to whatever destination you specify, and then the RDD can be reused from the checkpointed state avoiding recomputing.

On Mon, Nov 19, 2018 at 7:51 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Thanks for your advise. But I'm using Batch processing. Does anyone have a solution for the batch processing case?

Best,

Rico.


Am 19.11.2018 um 09:43 schrieb Magnus Nilsson:

Magnus Nilsson

9:43 AM (0 minutes ago)

to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]



Virenfrei. www.avast.com

---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]


--
Sent from my iPhone


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]
Reply | Threaded
Open this post in threaded view
|

Re: Spark DataSets and multiple write(.) calls

Gourav Sengupta
Hi,

 this is interesting, can you please share the code for this and if possible the source schema and it will be great if you could kindly share a sample file.


Regards,
Gourav Sengupta

On Tue, Nov 20, 2018 at 9:50 AM Michael Shtelma <[hidden email]> wrote:

You can also cache the data frame on disk, if it does not fit into memory. 
An alternative would be to write out data frame as parquet and then read it, you can check if in this case the whole pipeline works faster as with the standard cache. 

Best,
Michael


On Tue, Nov 20, 2018 at 9:14 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Hi!

Thanks Vadim for your answer. But this would be like caching the dataset, right? Or is checkpointing faster then persisting to memory or disk?

I attach a pdf of my dataflow program. If I could compute the output of outputs 1-5 in parallel the output of flatmap1 and groupBy could be reused, avoiding to write to disk (at least until the grouping).

Any other ideas or proposals?

Best,

Rico.


Am 19.11.2018 um 19:12 schrieb Vadim Semenov:
You can use checkpointing, in this case Spark will write out an rdd to whatever destination you specify, and then the RDD can be reused from the checkpointed state avoiding recomputing.

On Mon, Nov 19, 2018 at 7:51 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:

Thanks for your advise. But I'm using Batch processing. Does anyone have a solution for the batch processing case?

Best,

Rico.


Am 19.11.2018 um 09:43 schrieb Magnus Nilsson:

Magnus Nilsson

9:43 AM (0 minutes ago)

to info
I had the same requirements. As far as I know the only way is to extend the foreachwriter, cache the microbatch result and write to each output.


Unfortunately it seems as if you have to make a new connection "per batch" instead of creating one long lasting connections for the pipeline as such. Ie you might have to implement some sort of connection pooling by yourself depending on sink. 

Regards,

Magnus


On Mon, Nov 19, 2018 at 9:13 AM Dipl.-Inf. Rico Bergmann <[hidden email]> wrote:
Hi!

I have a SparkSQL programm, having one input and 6 ouputs (write). When
executing this programm every call to write(.) executes the plan. My
problem is, that I want all these writes to happen in parallel (inside
one execution plan), because all writes have a common and compute
intensive subpart, that can be shared by all plans. Is there a
possibility to do this? (Caching is not a solution because the input
dataset is way to large...)

Hoping for advises ...

Best, Rico B.


---
Diese E-Mail wurde von Avast Antivirus-Software auf Viren geprüft.
https://www.avast.com/antivirus


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]



Virenfrei. www.avast.com

---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]


--
Sent from my iPhone


---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]