Set partition number of Flink DataSet

[qi luo]

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

[Ken Krugler]

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[Ken Krugler]

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[Ken Krugler]

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files?

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable.

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[Ken Krugler]

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Hi Ken,

That looks awesome! I’ve implemented something similar to your bucketing sink, but using multiple internal writers rather than multiple internal output.

Besides this, I’m also curious whether Flink can achieve this like Spark: allow user to specify partition number in partitionBy() method (so no multiple output formats are needed). But this seems to need non-trivial changes in Flink core.

Thanks,

Qi

On Mar 15, 2019, at 2:36 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[Fabian Hueske-2]

Hi,

Flink works a bit differently than Spark.

By default, Flink uses pipelined shuffles which push results of the sender immediately to the receivers (btw. this is one of the building blocks for stream processing).

However, pipelined shuffles require that all receivers are online. Hence, there number of partitions determines the number of running tasks.

There is also a batch shuffle mode, but it needs to be explicitly enabled and AFAIK does not resolve the dependency of number of partitions and task parallelism.

However, the community is currently working on many improvements for batch processing, including scheduling and fault-tolerance.

Batched shuffles are an important building block for this and there might be better support for your use case in the future.

Best, Fabian

Am Fr., 15. März 2019 um 03:56 Uhr schrieb qi luo <[hidden email]>:

Hi Ken,

That looks awesome! I’ve implemented something similar to your bucketing sink, but using multiple internal writers rather than multiple internal output.

Besides this, I’m also curious whether Flink can achieve this like Spark: allow user to specify partition number in partitionBy() method (so no multiple output formats are needed). But this seems to need non-trivial changes in Flink core.

Thanks,

Qi

On Mar 15, 2019, at 2:36 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Hi Fabian,

I understand this is a by-design behavior, since Flink is firstly built for streaming. Supporting batch shuffle and custom partition number in Flink may be compelling in batch processing. 

Could you help explain a bit more on which works are needed to be done, so Flink can support custom partition numbers numbers? We would be willing to help improve this area.

Thanks,

Qi

On Mar 15, 2019, at 4:25 PM, Fabian Hueske <[hidden email]> wrote:

Hi,

Flink works a bit differently than Spark.

By default, Flink uses pipelined shuffles which push results of the sender immediately to the receivers (btw. this is one of the building blocks for stream processing).

However, pipelined shuffles require that all receivers are online. Hence, there number of partitions determines the number of running tasks.

There is also a batch shuffle mode, but it needs to be explicitly enabled and AFAIK does not resolve the dependency of number of partitions and task parallelism.

However, the community is currently working on many improvements for batch processing, including scheduling and fault-tolerance.

Batched shuffles are an important building block for this and there might be better support for your use case in the future.

Best, Fabian

Am Fr., 15. März 2019 um 03:56 Uhr schrieb qi luo <[hidden email]>:

Hi Ken,

That looks awesome! I’ve implemented something similar to your bucketing sink, but using multiple internal writers rather than multiple internal output.

Besides this, I’m also curious whether Flink can achieve this like Spark: allow user to specify partition number in partitionBy() method (so no multiple output formats are needed). But this seems to need non-trivial changes in Flink core.

Thanks,

Qi

On Mar 15, 2019, at 2:36 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[Fabian Hueske-2]

Hi,

I'm sorry but I'm only familiar with the high-level design but not with the implementation details and concrete roadmap for the involved components.

I think that FLINK-10288 [1] and FLINK-10429 [2] are related to partition handling.

Best,

Fabian

[1] https://issues.apache.org/jira/browse/FLINK-10288

[2] https://issues.apache.org/jira/browse/FLINK-10429

Am Fr., 15. März 2019 um 12:13 Uhr schrieb qi luo <[hidden email]>:

Hi Fabian,

I understand this is a by-design behavior, since Flink is firstly built for streaming. Supporting batch shuffle and custom partition number in Flink may be compelling in batch processing. 

Could you help explain a bit more on which works are needed to be done, so Flink can support custom partition numbers numbers? We would be willing to help improve this area.

Thanks,

Qi

On Mar 15, 2019, at 4:25 PM, Fabian Hueske <[hidden email]> wrote:

Hi,

Flink works a bit differently than Spark.

By default, Flink uses pipelined shuffles which push results of the sender immediately to the receivers (btw. this is one of the building blocks for stream processing).

However, pipelined shuffles require that all receivers are online. Hence, there number of partitions determines the number of running tasks.

There is also a batch shuffle mode, but it needs to be explicitly enabled and AFAIK does not resolve the dependency of number of partitions and task parallelism.

However, the community is currently working on many improvements for batch processing, including scheduling and fault-tolerance.

Batched shuffles are an important building block for this and there might be better support for your use case in the future.

Best, Fabian

Am Fr., 15. März 2019 um 03:56 Uhr schrieb qi luo <[hidden email]>:

Hi Ken,

That looks awesome! I’ve implemented something similar to your bucketing sink, but using multiple internal writers rather than multiple internal output.

Besides this, I’m also curious whether Flink can achieve this like Spark: allow user to specify partition number in partitionBy() method (so no multiple output formats are needed). But this seems to need non-trivial changes in Flink core.

Thanks,

Qi

On Mar 15, 2019, at 2:36 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

[qi luo]

Thank you Fabian! I will check these issues.

On Mar 20, 2019, at 4:25 PM, Fabian Hueske <[hidden email]> wrote:

Hi,

I'm sorry but I'm only familiar with the high-level design but not with the implementation details and concrete roadmap for the involved components.

I think that FLINK-10288 [1] and FLINK-10429 [2] are related to partition handling.

Best,

Fabian

[1] https://issues.apache.org/jira/browse/FLINK-10288

[2] https://issues.apache.org/jira/browse/FLINK-10429

Am Fr., 15. März 2019 um 12:13 Uhr schrieb qi luo <[hidden email]>:

Hi Fabian,

I understand this is a by-design behavior, since Flink is firstly built for streaming. Supporting batch shuffle and custom partition number in Flink may be compelling in batch processing. 

Could you help explain a bit more on which works are needed to be done, so Flink can support custom partition numbers numbers? We would be willing to help improve this area.

Thanks,

Qi

On Mar 15, 2019, at 4:25 PM, Fabian Hueske <[hidden email]> wrote:

Hi,

Flink works a bit differently than Spark.

By default, Flink uses pipelined shuffles which push results of the sender immediately to the receivers (btw. this is one of the building blocks for stream processing).

However, pipelined shuffles require that all receivers are online. Hence, there number of partitions determines the number of running tasks.

There is also a batch shuffle mode, but it needs to be explicitly enabled and AFAIK does not resolve the dependency of number of partitions and task parallelism.

However, the community is currently working on many improvements for batch processing, including scheduling and fault-tolerance.

Batched shuffles are an important building block for this and there might be better support for your use case in the future.

Best, Fabian

Am Fr., 15. März 2019 um 03:56 Uhr schrieb qi luo <[hidden email]>:

Hi Ken,

That looks awesome! I’ve implemented something similar to your bucketing sink, but using multiple internal writers rather than multiple internal output.

Besides this, I’m also curious whether Flink can achieve this like Spark: allow user to specify partition number in partitionBy() method (so no multiple output formats are needed). But this seems to need non-trivial changes in Flink core.

Thanks,

Qi

On Mar 15, 2019, at 2:36 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

See https://github.com/ScaleUnlimited/flink-utils/, for a rough but working version of a bucketing sink.

— Ken

On Mar 13, 2019, at 7:46 PM, qi luo <[hidden email]> wrote:

Hi Ken,

Agree. I will try partitonBy() to reducer the number of parallel sinks, and may also try sortPartition() so each sink could write files one by one. Looking forward to your solution. :)

Thanks,

Qi

On Mar 14, 2019, at 2:54 AM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

On Mar 13, 2019, at 1:26 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Do you mean that I can create a batch sink which writes to N files? 

Correct.

That sounds viable, but since our data size is huge (billions of records & thousands of files), the performance may be unacceptable. 

The main issue with performance (actually memory usage) is how many OutputFormats do you need to have open at the same time.

If you partition by the same key that’s used to define buckets, then the max number is less, as each parallel instance of the sink only gets a unique subset of all possible bucket values.

I’m actually dealing with something similar now, so I might have a solution to share soon.

— Ken

I will check Blink and give it a try anyway.

Thank you,

Qi

On Mar 12, 2019, at 11:58 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

If I understand what you’re trying to do, then this sounds like a variation of a bucketing sink.

That typically uses a field value to create a directory path or a file name (though the filename case is only viable when the field is also what’s used to partition the data)

But I don’t believe Flink has built-in support for that, in batch mode (see BucketingSink for streaming).

Maybe Blink has added that? Hoping someone who knows that codebase can chime in here.

Otherwise you’ll need to create a custom sink to implement the desired behavior - though abusing a MapPartitionFunction would be easiest, I think.

— Ken

On Mar 12, 2019, at 2:28 AM, qi luo <[hidden email]> wrote:

Hi Ken,

Thanks for your reply. I may not make myself clear: our problem is not about reading but rather writing. 

We need to write to N files based on key partitioning. We have to use setParallelism() to set the output partition/file number, but when the partition number is too large (~100K), the parallelism would be too high. Is there any other way to achieve this?

Thanks,

Qi

On Mar 11, 2019, at 11:22 PM, Ken Krugler <[hidden email]> wrote:

Hi Qi,

I’m guessing you’re calling createInput() for each input file.

If so, then instead you want to do something like:

     Job job = Job.getInstance();

for each file…

FileInputFormat.addInputPath(job, new org.apache.hadoop.fs.Path(file path));

env.createInput(HadoopInputs.createHadoopInput(…, job)

Flink/Hadoop will take care of parallelizing the reads from the files, given the parallelism that you’re specifying.

— Ken

On Mar 11, 2019, at 5:42 AM, qi luo <[hidden email]> wrote:

Hi,

We’re trying to distribute batch input data to (N) HDFS files partitioning by hash using DataSet API. What I’m doing is like:

env.createInput(…)

      .partitionByHash(0)

      .setParallelism(N)

      .output(…)

This works well for small number of files. But when we need to distribute to large number of files (say 100K), the parallelism becomes too large and we could not afford that many TMs.

In spark we can write something like ‘rdd.partitionBy(N)’ and control the parallelism separately (using dynamic allocation). Is there anything similar in Flink or other way we can achieve similar result? Thank you!

Qi

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra

--------------------------

Ken Krugler
[phone]
http://www.scaleunlimited.com
Custom big data solutions & training
Flink, Solr, Hadoop, Cascading & Cassandra