Debezium CDC | OOM

[Ayush Chauhan]

Hi, 

I am using flink cdc to stream CDC changes in an iceberg table. When I first run the flink job for a topic which has all the data for a table, it get out of heap memory as flink try to load all the data during my 15mins checkpointing interval. Right now, only solution I have is to pass -ytm 8192 -yjm 2048m for a table with 10M rows and then reduce it after flink has consumed all the data. Is there a way to tell flink cdc code to trigger checkpoint or throttle the consumption speed(I think backpressure should have handled this)?

--

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.

[Matthias]

Hi Ayush,

Which state backend have you configured [1]? Have you considered trying out RocksDB [2]? RocksDB might help with persisting at least keyed state.

Best,
Matthias

[1] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#choose-the-right-state-backend

[2] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#the-rocksdbstatebackend

On Thu, Apr 22, 2021 at 7:52 AM Ayush Chauhan <[hidden email]> wrote:

Hi, 

I am using flink cdc to stream CDC changes in an iceberg table. When I first run the flink job for a topic which has all the data for a table, it get out of heap memory as flink try to load all the data during my 15mins checkpointing interval. Right now, only solution I have is to pass -ytm 8192 -yjm 2048m for a table with 10M rows and then reduce it after flink has consumed all the data. Is there a way to tell flink cdc code to trigger checkpoint or throttle the consumption speed(I think backpressure should have handled this)?

--

 Ayush Chauhan
 Software Engineer | Data Platform 
   [phone] 

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.

[Ayush Chauhan]

Hi Matthias,

I am using RocksDB as a state backend. I think the iceberg sink is not able to propagate back pressure to the source which is resulting in OOM for my CDC pipeline. 

Please refer to this - https://github.com/apache/iceberg/issues/2504

On Thu, Apr 22, 2021 at 8:44 PM Matthias Pohl <[hidden email]> wrote:

Hi Ayush,

Which state backend have you configured [1]? Have you considered trying out RocksDB [2]? RocksDB might help with persisting at least keyed state.

Best,
Matthias

[1] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#choose-the-right-state-backend

[2] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#the-rocksdbstatebackend

On Thu, Apr 22, 2021 at 7:52 AM Ayush Chauhan <[hidden email]> wrote:

Hi, 

I am using flink cdc to stream CDC changes in an iceberg table. When I first run the flink job for a topic which has all the data for a table, it get out of heap memory as flink try to load all the data during my 15mins checkpointing interval. Right now, only solution I have is to pass -ytm 8192 -yjm 2048m for a table with 10M rows and then reduce it after flink has consumed all the data. Is there a way to tell flink cdc code to trigger checkpoint or throttle the consumption speed(I think backpressure should have handled this)?

--

 Ayush Chauhan
 Software Engineer | Data Platform 
   [phone] 

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.

--

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.

[Matthias]

Got it. Thanks for clarifying.

On Fri, Apr 23, 2021 at 6:36 AM Ayush Chauhan <[hidden email]> wrote:

Hi Matthias,

I am using RocksDB as a state backend. I think the iceberg sink is not able to propagate back pressure to the source which is resulting in OOM for my CDC pipeline. 

Please refer to this - https://github.com/apache/iceberg/issues/2504

On Thu, Apr 22, 2021 at 8:44 PM Matthias Pohl <[hidden email]> wrote:

Hi Ayush,

Which state backend have you configured [1]? Have you considered trying out RocksDB [2]? RocksDB might help with persisting at least keyed state.

Best,
Matthias

[1] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#choose-the-right-state-backend

[2] https://ci.apache.org/projects/flink/flink-docs-stable/ops/state/state_backends.html#the-rocksdbstatebackend

On Thu, Apr 22, 2021 at 7:52 AM Ayush Chauhan <[hidden email]> wrote:

Hi, 

I am using flink cdc to stream CDC changes in an iceberg table. When I first run the flink job for a topic which has all the data for a table, it get out of heap memory as flink try to load all the data during my 15mins checkpointing interval. Right now, only solution I have is to pass -ytm 8192 -yjm 2048m for a table with 10M rows and then reduce it after flink has consumed all the data. Is there a way to tell flink cdc code to trigger checkpoint or throttle the consumption speed(I think backpressure should have handled this)?

--

 Ayush Chauhan
 Software Engineer | Data Platform 
   [phone] 

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.

--

 Ayush Chauhan
 Software Engineer | Data Platform 
   [phone] 

This email is intended only for the person or the entity to whom it is addressed. If you are not the intended recipient, please delete this email and contact the sender.