Hi Vidya,
As to the choice of serializer:
* Flink provides two implementations that support state migration, AVRO
serializer, and Pojo serializer
* Pojo serializer happens to be one of the fastest available serializers
(faster than AVRO)
* If your record sticks to Pojo coding rules it is probably a good choice,
no extra serializer coding needed
* See here [1]
As to the extra big incremental checkpoints at the end of a time window:
* This is quite plausible,
* windowing uses the ‘namespace’ subkey of keyed state
* ideally incremental checkpoints only store changes made since the last
checkpoint, and
* on a window change many window instances (i.e. one per key and time
interval) disappear and are eventually recreated for the next time interval,
hence the bigger checkpoint
* serialization efforts depend on the choice of state backend:
* RocksDBStateBackend dominantly uses serializers when reading and
writing state but to a lesser extend for checkpoints
* FsStateBackend does not use serializers when reading and writing state
but dominantly during checkpoints
In order to improve your situation you need to take a closer look into
* The numbers (how many keys, how many active window instances
(globally/per key), how many events are collected per window instance)
* The specific implementation of the rollup/aggregation function
* There are setups that store all events and iterate whenever a window
result is needed (triggered)
* Other setups pre-aggregate incoming events and summarize only when a
window result is needed (triggered)
* This choice makes a big difference when it comes to state size
Hope this helps … feel free to get back with further questions 😊
Thias
[1]
https://flink.apache.org/news/2020/04/15/flink-serialization-tuning-vol-1.html#pojoserializer
From: Vidya Sagar Mula <[email protected]>
Sent: Dienstag, 8. März 2022 02:44
To: Yun Tang <[email protected]>
Cc: user <[email protected]>
Subject: Re: Incremental checkpointing & RocksDB Serialization
Hi Yun,
Thank you for the response.
1. You could tune your job to avoid backpressure. Maybe you can upgrade
your flink engine to at least flink-1.13 to know how to monitor the back
pressure status [1].
[VIDYA] - In the view of my organization, it's a very big activity to upgrade
to Flink version from our current one(1.11). I need to continue for my dev
activity with 1.11 only.
1. You can refer to [2] to know how to custom your serializer.
[VIDYA] - Thanks for providing me with the link references for custom
serializer. I am wondering, how is the serialization part in the incremental
checkpointing is different from Full checkpointing. My pipeline logic is same
for both Full checkpoint and Incremental checkpoint, except the checkpoint.type
variable change and some other env variables. But, the code pipeline logic
should be same for both types of checkpoints.
- Full checkpoint of pipeline is not taking considerably long time when
compared to incremental checkpointing at the end of the window. I see the
backpressure is High and CPU utilization is high with incremental
checkpointing. Thread dump shows the stack related to serialization. How is the
serialization part different between full checkpointing vs Incremental
checkpointing? I know, RocksDB library has some serializers for Incremental.
- While I am not writing custom serializer for my pipeline in case of Full
checkpointing, is it the general pattern to implement custom serializer in case
of Incremental?
- With respect with serializers for Full vs Incremental checkpointing, What's
the general usage pattern across the Flink community? If I write custom
serializer for Incremental, how does it go with Full checkpointing.
Please clarify.
Thanks,
Vidya.
[1]
https://nightlies.apache.org/flink/flink-docs-release-1.13/docs/ops/monitoring/back_pressure/
[2]
https://nightlies.apache.org/flink/flink-docs-release-1.13/docs/dev/datastream/fault-tolerance/custom_serialization/
On Sun, Mar 6, 2022 at 12:11 AM Yun Tang
<[email protected]<mailto:[email protected]>> wrote:
Hi Vidya,
1. You could tune your job to avoid backpressure. Maybe you can upgrade your
flink engine to at least flink-1.13 to know how to monitor the back pressure
status [1]
2. You can refer to [2] to know how to custom your serializer.
[1]
https://nightlies.apache.org/flink/flink-docs-release-1.13/docs/ops/monitoring/back_pressure/
[2]
https://nightlies.apache.org/flink/flink-docs-release-1.13/docs/dev/datastream/fault-tolerance/custom_serialization/
Best,
Yun Tang
________________________________
From: Vidya Sagar Mula <[email protected]<mailto:[email protected]>>
Sent: Sunday, March 6, 2022 4:16
To: Yun Tang <[email protected]<mailto:[email protected]>>
Cc: user <[email protected]<mailto:[email protected]>>
Subject: Re: Incremental checkpointing & RocksDB Serialization
Hi Yun Tang,
Thank you for the reply. I have follow up questions and need some more details.
Can you please clarify my inline questions?
> Why is the incremental checkpointing taking more time for the snapshot at the
> end of the window duration?
I guess that this is because the job is under back pressure on end of window.
You can expand the checkpoint details to see whether that the async duration of
each task is much slower than the e2e duration? If so, this caused the
checkpoint barrier stay in the channel longer.
<VIDYA> - Yes, I expanded the checkpoint details and noticed e2e duration is
much higher than async duration. Attaching the screenshot here(Checkpoint #59)
Can you give elaborate more on "checkpoint barrier stay in the channel longer."
What are the suggested ways to mitigate this issue? I am wondering how can this
be avoided as it is happening only at the end of the window.
> Do you suggest any change in the serializer type in the RocksDB? (Kryo vs
> Avro)
From our experience, kryo is not a good choice in most cases.
<VIDYA> - What are your recommendations on other serializers? I tried to change
it to Avro by enabling the flag "forceAvro" to TRUE in the Execution Config.
But, it RocksDB is still going picking KryoSerializer. This is because the
Transformation is KeyType is assigned as GenericType. I am not sure what
changes need to made to my class/pojo to take the Avro Serialzer.
Can you please suggest the way to change to other better serializers?
On Fri, Mar 4, 2022 at 2:06 AM Yun Tang
<[email protected]<mailto:[email protected]>> wrote:
Hi Vidya,
> Why is the incremental checkpointing taking more time for the snapshot at the
> end of the window duration?
I guess that this is because the job is under back pressure on end of window.
You can expand the checkpoint details to see whether that the async duration of
each task is much slower than the e2e duration? If so, this caused the
checkpoint barrier stay in the channel longer.
> Why is RocksDB serialization causing the CPU peak?
This is caused by the implementation of your serializer.
> Do you suggest any change in the serializer type in the RocksDB? (Kryo vs
> Avro)
From our experience, kryo is not a good choice in most cases.
Best
Yun Tang
________________________________
From: Vidya Sagar Mula <[email protected]<mailto:[email protected]>>
Sent: Friday, March 4, 2022 17:00
To: user <[email protected]<mailto:[email protected]>>
Subject: Incremental checkpointing & RocksDB Serialization
Hi,
I have a cluster that contains the Flink 1.11 version with AWS - S3 backend. I
am trying the incremental checkpointing on this set up. I have a pipeline with
a 10 mins window and incremental checkpointing happens every 2 mins.
Observation:
-------------
I am observing the long duration while taking the snapshot at the end of each
window, which means every last checkpoint of the window (almost all the times).
I am attaching the Flink UI, checkpoint history.
My set up details:
-------------------
Cluster: Cloud cluster with instance storage.
Memory : 20 GB,
Heap : 10 GB
Flink Managed Memory: 4.5 GB
Flink Version : 1.11
CPUs : 2
ROCKSDB_WRITE_BUFFER_SIZE: "2097152000" ## 2GB
ROCKSDB_BLOCK_CACHE_SIZE: "104857600" ## 100 Mb
ROCKSDB_BLOCK_SIZE: "5242880" ## 5 Mb
ROCKSDB_CHECKPOINT_TRANSFER_THREAD_NUM: 4
ROCKSDB_MAX_BACKGROUND_THREADS: 4
In the analysis, I noticed that the CPU utilization is peaking to almost 100%
at the time of issue. With further analysis with thread dumps at the time CPU
peak, it is showing RocksDB serialization related call trace. All the thread
samples are pointing to this stack.
Based on pipeline transformation class type, RocksDB is choosing Kryo
Serializer. I did try to change the serializer type, but that is not the focal
point I want to stress here.
I would like to understand the reason for high CPU utilization. I have tried to
increase the CPU cycles to 2 and 4. But, it did not give me any better results.
I have parallelism 2.
Please take a look at the below stack trace. Please suggest me why it is taking
a lot of CPU at the time of serialize/deserialize in the RocksDB?
########
Stack-1, Stack-2, Stack-3 are attached to this email.
Questions:
-----------
- Why is the incremental checkpointing taking more time for the snapshot at the
end of the window duration?
- Why is RocksDB serialization causing the CPU peak?
- Do you suggest any change in the serializer type in the RocksDB? (Kryo vs
Avro)
Thank you,
Diese Nachricht ist ausschliesslich für den Adressaten bestimmt und beinhaltet
unter Umständen vertrauliche Mitteilungen. Da die Vertraulichkeit von
e-Mail-Nachrichten nicht gewährleistet werden kann, übernehmen wir keine
Haftung für die Gewährung der Vertraulichkeit und Unversehrtheit dieser
Mitteilung. Bei irrtümlicher Zustellung bitten wir Sie um Benachrichtigung per
e-Mail und um Löschung dieser Nachricht sowie eventueller Anhänge. Jegliche
unberechtigte Verwendung oder Verbreitung dieser Informationen ist streng
verboten.
This message is intended only for the named recipient and may contain
confidential or privileged information. As the confidentiality of email
communication cannot be guaranteed, we do not accept any responsibility for the
confidentiality and the intactness of this message. If you have received it in
error, please advise the sender by return e-mail and delete this message and
any attachments. Any unauthorised use or dissemination of this information is
strictly prohibited.
