Flink Checkpoint runs slow for low load stream

[Chakravarthy varaga]

Hi Stephan, All,

     I just got a chance to try if 1.1.3 fixes slow check pointing on FS backend. It seemed to have been fixed. Thanks for the fix.

     While testing this, with varying check point intervals, there seem to be Spikes of slow checkpoints every 30/40 seconds for an interval of 15 secs. The check point time lasts for about 300 ms as apposed to 10/20 ms.
     Basically 15 secs seem to be the nominal value so far. anything below this interval shoots the spikes too often. For us living with 15 sec recovery is do-able and eventually catch up on recovery !

Best Regards

CVP

On Tue, Oct 4, 2016 at 6:20 PM, Chakravarthy varaga <[hidden email]> wrote:

Thanks for your prompt response Stephan.

    I'd wait for Flink 1.1.3 !!!

Best Regards

Varaga

On Tue, Oct 4, 2016 at 5:36 PM, Stephan Ewen <[hidden email]> wrote:

The plan to release 1.1.3 is asap ;-)

Waiting for last backported patched to get in, then release testing and release.

If you want to test it today, you would need to manually build the release-1.1 branch.

Best,

Stephan

On Tue, Oct 4, 2016 at 5:46 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Gordon,

     Do I need to clone and build release-1.1 branch to test this?

     I currently use flinlk 1.1.2 runtime. When is the plan to release it in 1.1.3?

Best Regards

Varaga

On Tue, Oct 4, 2016 at 9:25 AM, Tzu-Li (Gordon) Tai <[hidden email]> wrote:

Hi,

Helping out here: this is the PR for async Kafka offset committing - https://github.com/apache/flink/pull/2574.

It has already been merged into the master and release-1.1 branches, so you can try out the changes now if you’d like.

The change should also be included in the 1.1.3 release, which the Flink community is discussing to release soon.

Will definitely be helpful if you can provide feedback afterwards!

Best Regards,

Gordon

On October 3, 2016 at 9:40:14 PM, Chakravarthy varaga ([hidden email]) wrote:

Hi Stephan,

    Is the Async kafka offset commit released in 1.3.1?

Varaga

On Wed, Sep 28, 2016 at 9:49 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stephan,

     That should be great. Let me know once the fix is done and the snapshot version to use, I'll check and revert then.

     Can you also share the JIRA that tracks the issue?

 

     With regards to offset commit issue, I'm not sure as to how to proceed here. Probably I'll use your fix first and see if the problem reoccurs.

Thanks much

Varaga

On Tue, Sep 27, 2016 at 7:46 PM, Stephan Ewen <[hidden email]> wrote:

@CVP

Flink stores in checkpoints in your case only the Kafka offsets (few bytes) and the custom state (e).

Here is an illustration of the checkpoint and what is stored (from the Flink docs).

https://ci.apache.org/projects/flink/flink-docs-master/internals/stream_checkpointing.html

I am quite puzzled why the offset committing problem occurs only for one input, and not for the other.

I am preparing a fix for 1.2, possibly going into 1.1.3 as well.

Could you try out a snapshot version to see if that fixes your problem?

Greetings,

Stephan

On Tue, Sep 27, 2016 at 2:24 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

     Thanks a million for your detailed explanation. I appreciate it.

     -  The zookeeper bundled with kafka 0.9.0.1 was used to start zookeeper. There is only 1 instance (standalone) of zookeeper running on my localhost (ubuntu 14.04)

     -  There is only 1 Kafka broker (version: 0.9.0.1 )

     With regards to Flink cluster there's only 1 JM & 2 TMs started with no HA. I presume this does not use zookeeper anyways as it runs as standalone cluster.

 

     BTW., The kafka connector version that I use is as suggested in the flink connectors page.
       <dependency>
              <groupId>org.apache.flink</groupId>
              <artifactId>flink-connector-kafka-0.9_2.10</artifactId>
              <version>1.1.1</version>
        </dependency>
 

     Do you see any issues with versions?

   

     1) Do you have benchmarks wrt., to checkpointing in flink?

     2) There isn't detailed explanation on what states are stored as part of the checkpointing process. For ex.,  If I have pipeline like source -> map -> keyBy -> map -> sink, my assumption on what's stored is:

         a) The source stream's custom watermarked records

         b) Intermediate states of each of the transformations in the pipeline

         c) Delta of Records stored from the previous sink

         d) Custom States (SayValueState as in my case) - Essentially this is what I bother about storing.

         e) All of my operators

      Is my understanding right?

     3) Is there a way in Flink to checkpoint only d) as stated above

     4) Can you apply checkpointing to only streams and certain operators (say I wish to store aggregated values part of the transformation)

Best Regards

CVP

On Mon, Sep 26, 2016 at 6:18 PM, Stephan Ewen <[hidden email]> wrote:

Thanks, the logs were very helpful!

TL:DR - The offset committing to ZooKeeper is very slow and prevents proper starting of checkpoints.

Here is what is happening in detail:

  - Between the point when the TaskManager receives the "trigger checkpoint" message and when the point when the KafkaSource actually starts the checkpoint is a long time (many seconds) - for one of the Kafka Inputs (the other is fine).

  - The only way this delayed can be introduced is if another checkpoint related operation (such as trigger() or notifyComplete() ) is still in progress when the checkpoint is started. Flink does not perform concurrent checkpoint operations on a single operator, to ease the concurrency model for users.

  - The operation that is still in progress must be the committing of the offsets (to ZooKeeper or Kafka). That also explains why this only happens once one side receives the first record. Before that, there is nothing to commit.

What Flink should fix:

  - The KafkaConsumer should run the commit operations asynchronously, to not block the "notifyCheckpointComplete()" method.

What you can fix:

  - Have a look at your Kafka/ZooKeeper setup. One Kafka Input works well, the other does not. Do they go against different sets of brokers, or different ZooKeepers? Is the metadata for one input bad?

  - In the next Flink version, you may opt-out of committing offsets to Kafka/ZooKeeper all together. It is not important for Flink's checkpoints anyways.

Greetings,

Stephan

On Mon, Sep 26, 2016 at 5:13 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

    Please find my responses below.

    - What source are you using for the slow input?

     [CVP] - Both stream as pointed out in my first mail, are Kafka Streams

  - How large is the state that you are checkpointing?

     [CVP] - I have enabled checkpointing on the StreamEnvironment as below.
         final StreamExecutionEnvironment streamEnv = StreamExecutionEnvironment.getExecutionEnvironment();
         streamEnv.setStateBackend(new FsStateBackend("file:///tmp/flink/checkpoints"));
         streamEnv.enableCheckpointing(10000);

      In terms of the state stored, the KS1 stream has payload of 100K events/second, while KS2 have about 1 event / 10 minutes... basically the operators perform flatmaps on 8 fields of tuple (all fields are primitives). If you look at the states' sizes in dashboard they are in Kb...

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?
    [CVP] -There are no back pressure atleast from the sample computation in the flink dashboard. 100K/second is low load for flink's benchmarks. I could not quite get the barriers vs snapshot state. I have attached the Task Manager log (DEBUG) info if that will interest you.
 

     I have attached the checkpoints times' as .png from the dashboard. Basically if you look at checkpoint IDs 28 & 29 &30- you'd see that the checkpoints take more than a minute in each case. Before these checkpoints, the KS2 stream did not have any events. As soon as an event(should be in bytes) was generated, the checkpoints went slow and subsequently a minute more for every checkpoint thereafter.

   This log was collected from the standalone flink cluster with 1 job manager & 2 TMs. 1 TM was running this application with checkpointing (parallelism=1)
 

    Please let me know if you need further info.,

    

On Fri, Sep 23, 2016 at 6:26 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

Let's try to figure that one out. Can you give us a bit more information?

  - What source are you using for the slow input?

  - How large is the state that you are checkpointing?

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?

Greetings,

Stephan

On Fri, Sep 23, 2016 at 3:35 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

I'm not so much familiar with the internals of the checkpointing system, but maybe Stephan (in CC) has an idea what's going on here.

Best, Fabian

2016-09-23 11:33 GMT+02:00 Chakravarthy varaga <[hidden email]>:

Hi Aljoscha & Fabian,

    I have a stream application that has 2 stream source as below.

     KeyedStream<String, String> ks1 = ds1.keyBy("*") ;
     KeyedStream<Tuple2<String, V>, String> ks2 = ds2.flatMap(split T into k-v pairs).keyBy(0);

     ks1.connect(ks2).flatMap(X);
     //X is a CoFlatMapFunction that inserts and removes elements from ks2 into a key-value state member. Elements from ks1 are matched against that state. the CoFlatMapFunction operator maintains ValueState<Tuple2<Long, Long>>;

     //ks1 is streaming about 100K events/sec from kafka topic

     //ks2 is streaming about 1 event every 10 minutes... Precisely when the 1st event is consumed from this stream, checkpoint takes 2 minutes straight away.

    The version of flink is 1.1.2.

I tried to use checkpoint every 10 Secs using a FsStateBackend... What I notice is that the checkpoint duration is almost 2 minutes for many cases, while for the other cases it varies from 100 ms to 1.5 minutes frequently. I'm attaching the snapshot of the dashboard for your reference.

     Is this an issue with flink checkpointing?

 Best Regards

CVP

[Fabian Hueske-2]

Hi CVP,

we recently release Flink 1.1.4, i.e., the next bugfix release of the 1.1.x series with major robustness improvements [1].

You might want to give 1.1.4 a try as well.

Best, Fabian

[1] http://flink.apache.org/news/2016/12/21/release-1.1.4.html

2017-01-04 16:51 GMT+01:00 Chakravarthy varaga <[hidden email]>:

Hi Stephan, All,

     I just got a chance to try if 1.1.3 fixes slow check pointing on FS backend. It seemed to have been fixed. Thanks for the fix.

     While testing this, with varying check point intervals, there seem to be Spikes of slow checkpoints every 30/40 seconds for an interval of 15 secs. The check point time lasts for about 300 ms as apposed to 10/20 ms.
     Basically 15 secs seem to be the nominal value so far. anything below this interval shoots the spikes too often. For us living with 15 sec recovery is do-able and eventually catch up on recovery !

Best Regards

CVP

On Tue, Oct 4, 2016 at 6:20 PM, Chakravarthy varaga <[hidden email]> wrote:

Thanks for your prompt response Stephan.

    I'd wait for Flink 1.1.3 !!!

Best Regards

Varaga

On Tue, Oct 4, 2016 at 5:36 PM, Stephan Ewen <[hidden email]> wrote:

The plan to release 1.1.3 is asap ;-)

Waiting for last backported patched to get in, then release testing and release.

If you want to test it today, you would need to manually build the release-1.1 branch.

Best,

Stephan

On Tue, Oct 4, 2016 at 5:46 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Gordon,

     Do I need to clone and build release-1.1 branch to test this?

     I currently use flinlk 1.1.2 runtime. When is the plan to release it in 1.1.3?

Best Regards

Varaga

On Tue, Oct 4, 2016 at 9:25 AM, Tzu-Li (Gordon) Tai <[hidden email]> wrote:

Hi,

Helping out here: this is the PR for async Kafka offset committing - https://github.com/apache/flink/pull/2574.

It has already been merged into the master and release-1.1 branches, so you can try out the changes now if you’d like.

The change should also be included in the 1.1.3 release, which the Flink community is discussing to release soon.

Will definitely be helpful if you can provide feedback afterwards!

Best Regards,

Gordon

On October 3, 2016 at 9:40:14 PM, Chakravarthy varaga ([hidden email]) wrote:

Hi Stephan,

    Is the Async kafka offset commit released in 1.3.1?

Varaga

On Wed, Sep 28, 2016 at 9:49 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stephan,

     That should be great. Let me know once the fix is done and the snapshot version to use, I'll check and revert then.

     Can you also share the JIRA that tracks the issue?

 

     With regards to offset commit issue, I'm not sure as to how to proceed here. Probably I'll use your fix first and see if the problem reoccurs.

Thanks much

Varaga

On Tue, Sep 27, 2016 at 7:46 PM, Stephan Ewen <[hidden email]> wrote:

@CVP

Flink stores in checkpoints in your case only the Kafka offsets (few bytes) and the custom state (e).

Here is an illustration of the checkpoint and what is stored (from the Flink docs).

https://ci.apache.org/projects/flink/flink-docs-master/internals/stream_checkpointing.html

I am quite puzzled why the offset committing problem occurs only for one input, and not for the other.

I am preparing a fix for 1.2, possibly going into 1.1.3 as well.

Could you try out a snapshot version to see if that fixes your problem?

Greetings,

Stephan

On Tue, Sep 27, 2016 at 2:24 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

     Thanks a million for your detailed explanation. I appreciate it.

     -  The zookeeper bundled with kafka 0.9.0.1 was used to start zookeeper. There is only 1 instance (standalone) of zookeeper running on my localhost (ubuntu 14.04)

     -  There is only 1 Kafka broker (version: 0.9.0.1 )

     With regards to Flink cluster there's only 1 JM & 2 TMs started with no HA. I presume this does not use zookeeper anyways as it runs as standalone cluster.

 

     BTW., The kafka connector version that I use is as suggested in the flink connectors page.
       <dependency>
              <groupId>org.apache.flink</groupId>
              <artifactId>flink-connector-kafka-0.9_2.10</artifactId>
              <version>1.1.1</version>
        </dependency>
 

     Do you see any issues with versions?

   

     1) Do you have benchmarks wrt., to checkpointing in flink?

     2) There isn't detailed explanation on what states are stored as part of the checkpointing process. For ex.,  If I have pipeline like source -> map -> keyBy -> map -> sink, my assumption on what's stored is:

         a) The source stream's custom watermarked records

         b) Intermediate states of each of the transformations in the pipeline

         c) Delta of Records stored from the previous sink

         d) Custom States (SayValueState as in my case) - Essentially this is what I bother about storing.

         e) All of my operators

      Is my understanding right?

     3) Is there a way in Flink to checkpoint only d) as stated above

     4) Can you apply checkpointing to only streams and certain operators (say I wish to store aggregated values part of the transformation)

Best Regards

CVP

On Mon, Sep 26, 2016 at 6:18 PM, Stephan Ewen <[hidden email]> wrote:

Thanks, the logs were very helpful!

TL:DR - The offset committing to ZooKeeper is very slow and prevents proper starting of checkpoints.

Here is what is happening in detail:

  - Between the point when the TaskManager receives the "trigger checkpoint" message and when the point when the KafkaSource actually starts the checkpoint is a long time (many seconds) - for one of the Kafka Inputs (the other is fine).

  - The only way this delayed can be introduced is if another checkpoint related operation (such as trigger() or notifyComplete() ) is still in progress when the checkpoint is started. Flink does not perform concurrent checkpoint operations on a single operator, to ease the concurrency model for users.

  - The operation that is still in progress must be the committing of the offsets (to ZooKeeper or Kafka). That also explains why this only happens once one side receives the first record. Before that, there is nothing to commit.

What Flink should fix:

  - The KafkaConsumer should run the commit operations asynchronously, to not block the "notifyCheckpointComplete()" method.

What you can fix:

  - Have a look at your Kafka/ZooKeeper setup. One Kafka Input works well, the other does not. Do they go against different sets of brokers, or different ZooKeepers? Is the metadata for one input bad?

  - In the next Flink version, you may opt-out of committing offsets to Kafka/ZooKeeper all together. It is not important for Flink's checkpoints anyways.

Greetings,

Stephan

On Mon, Sep 26, 2016 at 5:13 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

    Please find my responses below.

    - What source are you using for the slow input?

     [CVP] - Both stream as pointed out in my first mail, are Kafka Streams

  - How large is the state that you are checkpointing?

     [CVP] - I have enabled checkpointing on the StreamEnvironment as below.
         final StreamExecutionEnvironment streamEnv = StreamExecutionEnvironment.getExecutionEnvironment();
         streamEnv.setStateBackend(new FsStateBackend("file:///tmp/flink/checkpoints"));
         streamEnv.enableCheckpointing(10000);

      In terms of the state stored, the KS1 stream has payload of 100K events/second, while KS2 have about 1 event / 10 minutes... basically the operators perform flatmaps on 8 fields of tuple (all fields are primitives). If you look at the states' sizes in dashboard they are in Kb...

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?
    [CVP] -There are no back pressure atleast from the sample computation in the flink dashboard. 100K/second is low load for flink's benchmarks. I could not quite get the barriers vs snapshot state. I have attached the Task Manager log (DEBUG) info if that will interest you.
 

     I have attached the checkpoints times' as .png from the dashboard. Basically if you look at checkpoint IDs 28 & 29 &30- you'd see that the checkpoints take more than a minute in each case. Before these checkpoints, the KS2 stream did not have any events. As soon as an event(should be in bytes) was generated, the checkpoints went slow and subsequently a minute more for every checkpoint thereafter.

   This log was collected from the standalone flink cluster with 1 job manager & 2 TMs. 1 TM was running this application with checkpointing (parallelism=1)
 

    Please let me know if you need further info.,

    

On Fri, Sep 23, 2016 at 6:26 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

Let's try to figure that one out. Can you give us a bit more information?

  - What source are you using for the slow input?

  - How large is the state that you are checkpointing?

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?

Greetings,

Stephan

On Fri, Sep 23, 2016 at 3:35 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

I'm not so much familiar with the internals of the checkpointing system, but maybe Stephan (in CC) has an idea what's going on here.

Best, Fabian

2016-09-23 11:33 GMT+02:00 Chakravarthy varaga <[hidden email]>:

Hi Aljoscha & Fabian,

    I have a stream application that has 2 stream source as below.

     KeyedStream<String, String> ks1 = ds1.keyBy("*") ;
     KeyedStream<Tuple2<String, V>, String> ks2 = ds2.flatMap(split T into k-v pairs).keyBy(0);

     ks1.connect(ks2).flatMap(X);
     //X is a CoFlatMapFunction that inserts and removes elements from ks2 into a key-value state member. Elements from ks1 are matched against that state. the CoFlatMapFunction operator maintains ValueState<Tuple2<Long, Long>>;

     //ks1 is streaming about 100K events/sec from kafka topic

     //ks2 is streaming about 1 event every 10 minutes... Precisely when the 1st event is consumed from this stream, checkpoint takes 2 minutes straight away.

    The version of flink is 1.1.2.

I tried to use checkpoint every 10 Secs using a FsStateBackend... What I notice is that the checkpoint duration is almost 2 minutes for many cases, while for the other cases it varies from 100 ms to 1.5 minutes frequently. I'm attaching the snapshot of the dashboard for your reference.

     Is this an issue with flink checkpointing?

 Best Regards

CVP

[Chakravarthy varaga]

BRILLIANT !!!

Checkpoint times are consistent with 1.1.4...

Thanks for your formidable support !

Best Regards

CVP

On Wed, Jan 4, 2017 at 5:33 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

we recently release Flink 1.1.4, i.e., the next bugfix release of the 1.1.x series with major robustness improvements [1].

You might want to give 1.1.4 a try as well.

Best, Fabian

[1] http://flink.apache.org/news/2016/12/21/release-1.1.4.html

2017-01-04 16:51 GMT+01:00 Chakravarthy varaga <[hidden email]>:

Hi Stephan, All,

     I just got a chance to try if 1.1.3 fixes slow check pointing on FS backend. It seemed to have been fixed. Thanks for the fix.

     While testing this, with varying check point intervals, there seem to be Spikes of slow checkpoints every 30/40 seconds for an interval of 15 secs. The check point time lasts for about 300 ms as apposed to 10/20 ms.
     Basically 15 secs seem to be the nominal value so far. anything below this interval shoots the spikes too often. For us living with 15 sec recovery is do-able and eventually catch up on recovery !

Best Regards

CVP

On Tue, Oct 4, 2016 at 6:20 PM, Chakravarthy varaga <[hidden email]> wrote:

Thanks for your prompt response Stephan.

    I'd wait for Flink 1.1.3 !!!

Best Regards

Varaga

On Tue, Oct 4, 2016 at 5:36 PM, Stephan Ewen <[hidden email]> wrote:

The plan to release 1.1.3 is asap ;-)

Waiting for last backported patched to get in, then release testing and release.

If you want to test it today, you would need to manually build the release-1.1 branch.

Best,

Stephan

On Tue, Oct 4, 2016 at 5:46 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Gordon,

     Do I need to clone and build release-1.1 branch to test this?

     I currently use flinlk 1.1.2 runtime. When is the plan to release it in 1.1.3?

Best Regards

Varaga

On Tue, Oct 4, 2016 at 9:25 AM, Tzu-Li (Gordon) Tai <[hidden email]> wrote:

Hi,

Helping out here: this is the PR for async Kafka offset committing - https://github.com/apache/flink/pull/2574.

It has already been merged into the master and release-1.1 branches, so you can try out the changes now if you’d like.

The change should also be included in the 1.1.3 release, which the Flink community is discussing to release soon.

Will definitely be helpful if you can provide feedback afterwards!

Best Regards,

Gordon

On October 3, 2016 at 9:40:14 PM, Chakravarthy varaga ([hidden email]) wrote:

Hi Stephan,

    Is the Async kafka offset commit released in 1.3.1?

Varaga

On Wed, Sep 28, 2016 at 9:49 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stephan,

     That should be great. Let me know once the fix is done and the snapshot version to use, I'll check and revert then.

     Can you also share the JIRA that tracks the issue?

 

     With regards to offset commit issue, I'm not sure as to how to proceed here. Probably I'll use your fix first and see if the problem reoccurs.

Thanks much

Varaga

On Tue, Sep 27, 2016 at 7:46 PM, Stephan Ewen <[hidden email]> wrote:

@CVP

Flink stores in checkpoints in your case only the Kafka offsets (few bytes) and the custom state (e).

Here is an illustration of the checkpoint and what is stored (from the Flink docs).

https://ci.apache.org/projects/flink/flink-docs-master/internals/stream_checkpointing.html

I am quite puzzled why the offset committing problem occurs only for one input, and not for the other.

I am preparing a fix for 1.2, possibly going into 1.1.3 as well.

Could you try out a snapshot version to see if that fixes your problem?

Greetings,

Stephan

On Tue, Sep 27, 2016 at 2:24 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

     Thanks a million for your detailed explanation. I appreciate it.

     -  The zookeeper bundled with kafka 0.9.0.1 was used to start zookeeper. There is only 1 instance (standalone) of zookeeper running on my localhost (ubuntu 14.04)

     -  There is only 1 Kafka broker (version: 0.9.0.1 )

     With regards to Flink cluster there's only 1 JM & 2 TMs started with no HA. I presume this does not use zookeeper anyways as it runs as standalone cluster.

 

     BTW., The kafka connector version that I use is as suggested in the flink connectors page.
       <dependency>
              <groupId>org.apache.flink</groupId>
              <artifactId>flink-connector-kafka-0.9_2.10</artifactId>
              <version>1.1.1</version>
        </dependency>
 

     Do you see any issues with versions?

   

     1) Do you have benchmarks wrt., to checkpointing in flink?

     2) There isn't detailed explanation on what states are stored as part of the checkpointing process. For ex.,  If I have pipeline like source -> map -> keyBy -> map -> sink, my assumption on what's stored is:

         a) The source stream's custom watermarked records

         b) Intermediate states of each of the transformations in the pipeline

         c) Delta of Records stored from the previous sink

         d) Custom States (SayValueState as in my case) - Essentially this is what I bother about storing.

         e) All of my operators

      Is my understanding right?

     3) Is there a way in Flink to checkpoint only d) as stated above

     4) Can you apply checkpointing to only streams and certain operators (say I wish to store aggregated values part of the transformation)

Best Regards

CVP

On Mon, Sep 26, 2016 at 6:18 PM, Stephan Ewen <[hidden email]> wrote:

Thanks, the logs were very helpful!

TL:DR - The offset committing to ZooKeeper is very slow and prevents proper starting of checkpoints.

Here is what is happening in detail:

  - Between the point when the TaskManager receives the "trigger checkpoint" message and when the point when the KafkaSource actually starts the checkpoint is a long time (many seconds) - for one of the Kafka Inputs (the other is fine).

  - The only way this delayed can be introduced is if another checkpoint related operation (such as trigger() or notifyComplete() ) is still in progress when the checkpoint is started. Flink does not perform concurrent checkpoint operations on a single operator, to ease the concurrency model for users.

  - The operation that is still in progress must be the committing of the offsets (to ZooKeeper or Kafka). That also explains why this only happens once one side receives the first record. Before that, there is nothing to commit.

What Flink should fix:

  - The KafkaConsumer should run the commit operations asynchronously, to not block the "notifyCheckpointComplete()" method.

What you can fix:

  - Have a look at your Kafka/ZooKeeper setup. One Kafka Input works well, the other does not. Do they go against different sets of brokers, or different ZooKeepers? Is the metadata for one input bad?

  - In the next Flink version, you may opt-out of committing offsets to Kafka/ZooKeeper all together. It is not important for Flink's checkpoints anyways.

Greetings,

Stephan

On Mon, Sep 26, 2016 at 5:13 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

    Please find my responses below.

    - What source are you using for the slow input?

     [CVP] - Both stream as pointed out in my first mail, are Kafka Streams

  - How large is the state that you are checkpointing?

     [CVP] - I have enabled checkpointing on the StreamEnvironment as below.
         final StreamExecutionEnvironment streamEnv = StreamExecutionEnvironment.getExecutionEnvironment();
         streamEnv.setStateBackend(new FsStateBackend("file:///tmp/flink/checkpoints"));
         streamEnv.enableCheckpointing(10000);

      In terms of the state stored, the KS1 stream has payload of 100K events/second, while KS2 have about 1 event / 10 minutes... basically the operators perform flatmaps on 8 fields of tuple (all fields are primitives). If you look at the states' sizes in dashboard they are in Kb...

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?
    [CVP] -There are no back pressure atleast from the sample computation in the flink dashboard. 100K/second is low load for flink's benchmarks. I could not quite get the barriers vs snapshot state. I have attached the Task Manager log (DEBUG) info if that will interest you.
 

     I have attached the checkpoints times' as .png from the dashboard. Basically if you look at checkpoint IDs 28 & 29 &30- you'd see that the checkpoints take more than a minute in each case. Before these checkpoints, the KS2 stream did not have any events. As soon as an event(should be in bytes) was generated, the checkpoints went slow and subsequently a minute more for every checkpoint thereafter.

   This log was collected from the standalone flink cluster with 1 job manager & 2 TMs. 1 TM was running this application with checkpointing (parallelism=1)
 

    Please let me know if you need further info.,

    

On Fri, Sep 23, 2016 at 6:26 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

Let's try to figure that one out. Can you give us a bit more information?

  - What source are you using for the slow input?

  - How large is the state that you are checkpointing?

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?

Greetings,

Stephan

On Fri, Sep 23, 2016 at 3:35 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

I'm not so much familiar with the internals of the checkpointing system, but maybe Stephan (in CC) has an idea what's going on here.

Best, Fabian

2016-09-23 11:33 GMT+02:00 Chakravarthy varaga <[hidden email]>:

Hi Aljoscha & Fabian,

    I have a stream application that has 2 stream source as below.

     KeyedStream<String, String> ks1 = ds1.keyBy("*") ;
     KeyedStream<Tuple2<String, V>, String> ks2 = ds2.flatMap(split T into k-v pairs).keyBy(0);

     ks1.connect(ks2).flatMap(X);
     //X is a CoFlatMapFunction that inserts and removes elements from ks2 into a key-value state member. Elements from ks1 are matched against that state. the CoFlatMapFunction operator maintains ValueState<Tuple2<Long, Long>>;

     //ks1 is streaming about 100K events/sec from kafka topic

     //ks2 is streaming about 1 event every 10 minutes... Precisely when the 1st event is consumed from this stream, checkpoint takes 2 minutes straight away.

    The version of flink is 1.1.2.

I tried to use checkpoint every 10 Secs using a FsStateBackend... What I notice is that the checkpoint duration is almost 2 minutes for many cases, while for the other cases it varies from 100 ms to 1.5 minutes frequently. I'm attaching the snapshot of the dashboard for your reference.

     Is this an issue with flink checkpointing?

 Best Regards

CVP

[Yassine MARZOUGUI]

Hi all,

I am experiencing a similar problem but with HDFS as a source instead of Kafka. I have a streaming pipeline as follows:

1 - read a folder continuousely from HDFS

2 - filter duplicates (using keyby(x->x) and keeping a state per key indicating whether its is seen)

3 - schedule some future actions on the stream using ProcessFunction and processing time timers (elements are kept in a MapState)

4- write results back to HDFS using a BucketingSink.

I am using RocksdbStateBackend, and Flink 1.3-SNAPSHOT (Commit: 9fb074c).

Currenlty the source contain just one a file of 1GB, so that's the maximum state that the job might hold. I noticed that the backpressure on the operators #1 and #2 is High, and the split reader has only read 60 Mb out of 1Gb source source file. I suspect this is because the ProcessFunction is slow (on purpose). However looks like this affected the checkpoints which are failing after the timeout (which is set to 2 hours).

In the job manager logs I keep getting warnings : 

2017-04-23 19:32:38,827 WARN org.apache.flink.runtime.checkpoint.CheckpointCoordinator - Received late message for now expired checkpoint attempt 8 from 210769a077c67841d980776d8caece0a of job 6c7e44d205d738fc8a6cb4da181d2d86.

Is the high backpressure the cause for the checkpoints being too slow? If yes Is there a way to disbale the backpressure mechanism since the records will be buffered in the rocksdb state after all which is backed by the disk?

Any help is appreciated. Thank you.

Best,

Yassine

On Jan 5, 2017 12:25, "Chakravarthy varaga" <[hidden email]> wrote:

BRILLIANT !!!

Checkpoint times are consistent with 1.1.4...

Thanks for your formidable support !

Best Regards

CVP

On Wed, Jan 4, 2017 at 5:33 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

we recently release Flink 1.1.4, i.e., the next bugfix release of the 1.1.x series with major robustness improvements [1].

You might want to give 1.1.4 a try as well.

Best, Fabian

[1] http://flink.apache.org/news/2016/12/21/release-1.1.4.html

2017-01-04 16:51 GMT+01:00 Chakravarthy varaga <[hidden email]>:

Hi Stephan, All,

     I just got a chance to try if 1.1.3 fixes slow check pointing on FS backend. It seemed to have been fixed. Thanks for the fix.

     While testing this, with varying check point intervals, there seem to be Spikes of slow checkpoints every 30/40 seconds for an interval of 15 secs. The check point time lasts for about 300 ms as apposed to 10/20 ms.
     Basically 15 secs seem to be the nominal value so far. anything below this interval shoots the spikes too often. For us living with 15 sec recovery is do-able and eventually catch up on recovery !

Best Regards

CVP

On Tue, Oct 4, 2016 at 6:20 PM, Chakravarthy varaga <[hidden email]> wrote:

Thanks for your prompt response Stephan.

    I'd wait for Flink 1.1.3 !!!

Best Regards

Varaga

On Tue, Oct 4, 2016 at 5:36 PM, Stephan Ewen <[hidden email]> wrote:

The plan to release 1.1.3 is asap ;-)

Waiting for last backported patched to get in, then release testing and release.

If you want to test it today, you would need to manually build the release-1.1 branch.

Best,

Stephan

On Tue, Oct 4, 2016 at 5:46 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Gordon,

     Do I need to clone and build release-1.1 branch to test this?

     I currently use flinlk 1.1.2 runtime. When is the plan to release it in 1.1.3?

Best Regards

Varaga

On Tue, Oct 4, 2016 at 9:25 AM, Tzu-Li (Gordon) Tai <[hidden email]> wrote:

Hi,

Helping out here: this is the PR for async Kafka offset committing - https://github.com/apache/flink/pull/2574.

It has already been merged into the master and release-1.1 branches, so you can try out the changes now if you’d like.

The change should also be included in the 1.1.3 release, which the Flink community is discussing to release soon.

Will definitely be helpful if you can provide feedback afterwards!

Best Regards,

Gordon

On October 3, 2016 at 9:40:14 PM, Chakravarthy varaga ([hidden email]) wrote:

Hi Stephan,

    Is the Async kafka offset commit released in 1.3.1?

Varaga

On Wed, Sep 28, 2016 at 9:49 AM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stephan,

     That should be great. Let me know once the fix is done and the snapshot version to use, I'll check and revert then.

     Can you also share the JIRA that tracks the issue?

 

     With regards to offset commit issue, I'm not sure as to how to proceed here. Probably I'll use your fix first and see if the problem reoccurs.

Thanks much

Varaga

On Tue, Sep 27, 2016 at 7:46 PM, Stephan Ewen <[hidden email]> wrote:

@CVP

Flink stores in checkpoints in your case only the Kafka offsets (few bytes) and the custom state (e).

Here is an illustration of the checkpoint and what is stored (from the Flink docs).

https://ci.apache.org/projects/flink/flink-docs-master/internals/stream_checkpointing.html

I am quite puzzled why the offset committing problem occurs only for one input, and not for the other.

I am preparing a fix for 1.2, possibly going into 1.1.3 as well.

Could you try out a snapshot version to see if that fixes your problem?

Greetings,

Stephan

On Tue, Sep 27, 2016 at 2:24 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

     Thanks a million for your detailed explanation. I appreciate it.

     -  The zookeeper bundled with kafka 0.9.0.1 was used to start zookeeper. There is only 1 instance (standalone) of zookeeper running on my localhost (ubuntu 14.04)

     -  There is only 1 Kafka broker (version: 0.9.0.1 )

     With regards to Flink cluster there's only 1 JM & 2 TMs started with no HA. I presume this does not use zookeeper anyways as it runs as standalone cluster.

 

     BTW., The kafka connector version that I use is as suggested in the flink connectors page.
       <dependency>
              <groupId>org.apache.flink</groupId>
              <artifactId>flink-connector-kafka-0.9_2.10</artifactId>
              <version>1.1.1</version>
        </dependency>
 

     Do you see any issues with versions?

   

     1) Do you have benchmarks wrt., to checkpointing in flink?

     2) There isn't detailed explanation on what states are stored as part of the checkpointing process. For ex.,  If I have pipeline like source -> map -> keyBy -> map -> sink, my assumption on what's stored is:

         a) The source stream's custom watermarked records

         b) Intermediate states of each of the transformations in the pipeline

         c) Delta of Records stored from the previous sink

         d) Custom States (SayValueState as in my case) - Essentially this is what I bother about storing.

         e) All of my operators

      Is my understanding right?

     3) Is there a way in Flink to checkpoint only d) as stated above

     4) Can you apply checkpointing to only streams and certain operators (say I wish to store aggregated values part of the transformation)

Best Regards

CVP

On Mon, Sep 26, 2016 at 6:18 PM, Stephan Ewen <[hidden email]> wrote:

Thanks, the logs were very helpful!

TL:DR - The offset committing to ZooKeeper is very slow and prevents proper starting of checkpoints.

Here is what is happening in detail:

  - Between the point when the TaskManager receives the "trigger checkpoint" message and when the point when the KafkaSource actually starts the checkpoint is a long time (many seconds) - for one of the Kafka Inputs (the other is fine).

  - The only way this delayed can be introduced is if another checkpoint related operation (such as trigger() or notifyComplete() ) is still in progress when the checkpoint is started. Flink does not perform concurrent checkpoint operations on a single operator, to ease the concurrency model for users.

  - The operation that is still in progress must be the committing of the offsets (to ZooKeeper or Kafka). That also explains why this only happens once one side receives the first record. Before that, there is nothing to commit.

What Flink should fix:

  - The KafkaConsumer should run the commit operations asynchronously, to not block the "notifyCheckpointComplete()" method.

What you can fix:

  - Have a look at your Kafka/ZooKeeper setup. One Kafka Input works well, the other does not. Do they go against different sets of brokers, or different ZooKeepers? Is the metadata for one input bad?

  - In the next Flink version, you may opt-out of committing offsets to Kafka/ZooKeeper all together. It is not important for Flink's checkpoints anyways.

Greetings,

Stephan

On Mon, Sep 26, 2016 at 5:13 PM, Chakravarthy varaga <[hidden email]> wrote:

Hi Stefan,

    Please find my responses below.

    - What source are you using for the slow input?

     [CVP] - Both stream as pointed out in my first mail, are Kafka Streams

  - How large is the state that you are checkpointing?

     [CVP] - I have enabled checkpointing on the StreamEnvironment as below.
         final StreamExecutionEnvironment streamEnv = StreamExecutionEnvironment.getExecutionEnvironment();
         streamEnv.setStateBackend(new FsStateBackend("file:///tmp/flink/checkpoints"));
         streamEnv.enableCheckpointing(10000);

      In terms of the state stored, the KS1 stream has payload of 100K events/second, while KS2 have about 1 event / 10 minutes... basically the operators perform flatmaps on 8 fields of tuple (all fields are primitives). If you look at the states' sizes in dashboard they are in Kb...

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?
    [CVP] -There are no back pressure atleast from the sample computation in the flink dashboard. 100K/second is low load for flink's benchmarks. I could not quite get the barriers vs snapshot state. I have attached the Task Manager log (DEBUG) info if that will interest you.
 

     I have attached the checkpoints times' as .png from the dashboard. Basically if you look at checkpoint IDs 28 & 29 &30- you'd see that the checkpoints take more than a minute in each case. Before these checkpoints, the KS2 stream did not have any events. As soon as an event(should be in bytes) was generated, the checkpoints went slow and subsequently a minute more for every checkpoint thereafter.

   This log was collected from the standalone flink cluster with 1 job manager & 2 TMs. 1 TM was running this application with checkpointing (parallelism=1)
 

    Please let me know if you need further info.,

    

On Fri, Sep 23, 2016 at 6:26 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

Let's try to figure that one out. Can you give us a bit more information?

  - What source are you using for the slow input?

  - How large is the state that you are checkpointing?

  - Can you try to see in the log if actually the state snapshot takes that long, or if it simply takes long for the checkpoint barriers to travel through the stream due to a lot of backpressure?

Greetings,

Stephan

On Fri, Sep 23, 2016 at 3:35 PM, Fabian Hueske <[hidden email]> wrote:

Hi CVP,

I'm not so much familiar with the internals of the checkpointing system, but maybe Stephan (in CC) has an idea what's going on here.

Best, Fabian

2016-09-23 11:33 GMT+02:00 Chakravarthy varaga <[hidden email]>:

Hi Aljoscha & Fabian,

    I have a stream application that has 2 stream source as below.

     KeyedStream<String, String> ks1 = ds1.keyBy("*") ;
     KeyedStream<Tuple2<String, V>, String> ks2 = ds2.flatMap(split T into k-v pairs).keyBy(0);

     ks1.connect(ks2).flatMap(X);
     //X is a CoFlatMapFunction that inserts and removes elements from ks2 into a key-value state member. Elements from ks1 are matched against that state. the CoFlatMapFunction operator maintains ValueState<Tuple2<Long, Long>>;

     //ks1 is streaming about 100K events/sec from kafka topic

     //ks2 is streaming about 1 event every 10 minutes... Precisely when the 1st event is consumed from this stream, checkpoint takes 2 minutes straight away.

    The version of flink is 1.1.2.

I tried to use checkpoint every 10 Secs using a FsStateBackend... What I notice is that the checkpoint duration is almost 2 minutes for many cases, while for the other cases it varies from 100 ms to 1.5 minutes frequently. I'm attaching the snapshot of the dashboard for your reference.

     Is this an issue with flink checkpointing?

 Best Regards

CVP