Storage options for RocksDBStateBackend

[ayush]

Hello, 

I had a few questions regarding checkpoint storage options using 
RocksDBStateBackend. In the flink 1.2 documentation, it is the recommended state 
backend due to it's ability to store large states and asynchronous snapshotting. 
For high availabilty it seems HDFS is the recommended store for state backend 
data. In AWS deployment section, it is also mentioned that s3 can be used for 
storing state backend data. 

We don't want to depend on a hadoop cluster for flink deployment, so I had 
following questions: 

1. Can we use any storage backend supported by flink for storing RocksDB 
StateBackend data with file urls: there are quite a few supported as mentioned here: 
https://ci.apache.org/projects/flink/flink-docs-release-1.3/internals/filesystems.html
and here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md

2. Is there some work already done to support Windows Azure Blob Storage for 
storing State backend data? There are some docs here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md
can we utilize this for that? 

3. If utilizing S3 for state backend, is there any performance impact? 

4. For high availability can we use a NFS volume for state backend, with 
"file://" urls? Will there be any performance impact? 

PS: I posted this email earlier via nabble, but it's not showing up in apache archive. So sending again. Apologies if it results in multiple threads.

-- Ayush

[Till Rohrmann]

Hi Ayush,

you’re right that RocksDB is the recommend state backend because of the above-mentioned reasons. In order to make the recovery properly work, you have to configure a shared directory for the checkpoint data via state.backend.fs.checkpointdir. You can basically configure any file system which is supported by Hadoop (no HDFS required). The reason is that we use Hadoop to bridge between different file systems. The only thing you have to make sure is that you have the respective file system implementation in your class path.

I think you can access Windows Azure Blob Storage via Hadoop [1] similarly to access S3, for example.

If you use S3 to store your checkpoint data, then you will benefit from all the advantages of S3 but also suffer from its drawbacks (e.g. that list operations are more costly). But these are not specific to Flink.

A URL like file:// usually indicates a local file. Thus, if your Flink cluster is not running on a single machine, then this won’t work.

[1] https://hadoop.apache.org/docs/stable/hadoop-azure/index.html

Cheers,
Till

​

On Thu, May 11, 2017 at 10:41 AM, Ayush Goyal <[hidden email]> wrote:

Hello, 

I had a few questions regarding checkpoint storage options using 
RocksDBStateBackend. In the flink 1.2 documentation, it is the recommended state 
backend due to it's ability to store large states and asynchronous snapshotting. 
For high availabilty it seems HDFS is the recommended store for state backend 
data. In AWS deployment section, it is also mentioned that s3 can be used for 
storing state backend data. 

We don't want to depend on a hadoop cluster for flink deployment, so I had 
following questions: 

1. Can we use any storage backend supported by flink for storing RocksDB 
StateBackend data with file urls: there are quite a few supported as mentioned here: 
https://ci.apache.org/projects/flink/flink-docs-release-1.3/internals/filesystems.html
and here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md

2. Is there some work already done to support Windows Azure Blob Storage for 
storing State backend data? There are some docs here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md
can we utilize this for that? 

3. If utilizing S3 for state backend, is there any performance impact? 

4. For high availability can we use a NFS volume for state backend, with 
"file://" urls? Will there be any performance impact? 

PS: I posted this email earlier via nabble, but it's not showing up in apache archive. So sending again. Apologies if it results in multiple threads.

-- Ayush

[Stephan Ewen]

Small addition to Till's comment:

In the case where file:// points to a mounted distributed file system (NFS, MapRFs, ...), then it actually works. The important thing is that the filesystem where the checkpoints go is replicated (fault tolerant) and accessible from all nodes.

On Thu, May 11, 2017 at 2:16 PM, Till Rohrmann <[hidden email]> wrote:

Hi Ayush,

you’re right that RocksDB is the recommend state backend because of the above-mentioned reasons. In order to make the recovery properly work, you have to configure a shared directory for the checkpoint data via state.backend.fs.checkpointdir. You can basically configure any file system which is supported by Hadoop (no HDFS required). The reason is that we use Hadoop to bridge between different file systems. The only thing you have to make sure is that you have the respective file system implementation in your class path.

I think you can access Windows Azure Blob Storage via Hadoop [1] similarly to access S3, for example.

If you use S3 to store your checkpoint data, then you will benefit from all the advantages of S3 but also suffer from its drawbacks (e.g. that list operations are more costly). But these are not specific to Flink.

A URL like file:// usually indicates a local file. Thus, if your Flink cluster is not running on a single machine, then this won’t work.

[1] https://hadoop.apache.org/docs/stable/hadoop-azure/index.html

Cheers,
Till

​

On Thu, May 11, 2017 at 10:41 AM, Ayush Goyal <[hidden email]> wrote:

Hello, 

I had a few questions regarding checkpoint storage options using 
RocksDBStateBackend. In the flink 1.2 documentation, it is the recommended state 
backend due to it's ability to store large states and asynchronous snapshotting. 
For high availabilty it seems HDFS is the recommended store for state backend 
data. In AWS deployment section, it is also mentioned that s3 can be used for 
storing state backend data. 

We don't want to depend on a hadoop cluster for flink deployment, so I had 
following questions: 

1. Can we use any storage backend supported by flink for storing RocksDB 
StateBackend data with file urls: there are quite a few supported as mentioned here: 
https://ci.apache.org/projects/flink/flink-docs-release-1.3/internals/filesystems.html
and here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md

2. Is there some work already done to support Windows Azure Blob Storage for 
storing State backend data? There are some docs here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md
can we utilize this for that? 

3. If utilizing S3 for state backend, is there any performance impact? 

4. For high availability can we use a NFS volume for state backend, with 
"file://" urls? Will there be any performance impact? 

PS: I posted this email earlier via nabble, but it's not showing up in apache archive. So sending again. Apologies if it results in multiple threads.

-- Ayush

[ayush]

Till and Stephan, thanks for your clarification.

@Till One more question, from what I have read about the checkpointing [1], the list operations don't seem likely to be performed frequently, so storing state backend on s3 shouldn't have any severe impact on flink performance. Is this assumption right?

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.2/internals/stream_checkpointing.html

-- Ayush

On Fri, May 12, 2017 at 1:05 AM Stephan Ewen <[hidden email]> wrote:

Small addition to Till's comment:

In the case where file:// points to a mounted distributed file system (NFS, MapRFs, ...), then it actually works. The important thing is that the filesystem where the checkpoints go is replicated (fault tolerant) and accessible from all nodes.

On Thu, May 11, 2017 at 2:16 PM, Till Rohrmann <[hidden email]> wrote:

Hi Ayush,

you’re right that RocksDB is the recommend state backend because of the above-mentioned reasons. In order to make the recovery properly work, you have to configure a shared directory for the checkpoint data via state.backend.fs.checkpointdir. You can basically configure any file system which is supported by Hadoop (no HDFS required). The reason is that we use Hadoop to bridge between different file systems. The only thing you have to make sure is that you have the respective file system implementation in your class path.

I think you can access Windows Azure Blob Storage via Hadoop [1] similarly to access S3, for example.

If you use S3 to store your checkpoint data, then you will benefit from all the advantages of S3 but also suffer from its drawbacks (e.g. that list operations are more costly). But these are not specific to Flink.

A URL like file:// usually indicates a local file. Thus, if your Flink cluster is not running on a single machine, then this won’t work.

[1] https://hadoop.apache.org/docs/stable/hadoop-azure/index.html

Cheers,
Till

​

On Thu, May 11, 2017 at 10:41 AM, Ayush Goyal <[hidden email]> wrote:

Hello, 

I had a few questions regarding checkpoint storage options using 
RocksDBStateBackend. In the flink 1.2 documentation, it is the recommended state 
backend due to it's ability to store large states and asynchronous snapshotting. 
For high availabilty it seems HDFS is the recommended store for state backend 
data. In AWS deployment section, it is also mentioned that s3 can be used for 
storing state backend data. 

We don't want to depend on a hadoop cluster for flink deployment, so I had 
following questions: 

1. Can we use any storage backend supported by flink for storing RocksDB 
StateBackend data with file urls: there are quite a few supported as mentioned here: 
https://ci.apache.org/projects/flink/flink-docs-release-1.3/internals/filesystems.html
and here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md

2. Is there some work already done to support Windows Azure Blob Storage for 
storing State backend data? There are some docs here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md
can we utilize this for that? 

3. If utilizing S3 for state backend, is there any performance impact? 

4. For high availability can we use a NFS volume for state backend, with 
"file://" urls? Will there be any performance impact? 

PS: I posted this email earlier via nabble, but it's not showing up in apache archive. So sending again. Apologies if it results in multiple threads.

-- Ayush

[Jain, Ankit]

Also, I hope state & checkpointing writes to S3 happens async w/o impacting the actual job execution graph?

 

If so, will there still be a performance impact from using S3?

 

Thanks

Ankit

 

From: Ayush Goyal <[hidden email]>
Date: Thursday, May 11, 2017 at 11:21 PM
To: Stephan Ewen <[hidden email]>, Till Rohrmann <[hidden email]>
Cc: user <[hidden email]>
Subject: Re: Storage options for RocksDBStateBackend

 

Till and Stephan, thanks for your clarification.

 

@Till One more question, from what I have read about the checkpointing [1], the list operations don't seem likely to be performed frequently, so storing state backend on s3 shouldn't have any severe impact on flink performance. Is this assumption right?

 

[1] https://ci.apache.org/projects/flink/flink-docs-release-1.2/internals/stream_checkpointing.html

 

-- Ayush

 

On Fri, May 12, 2017 at 1:05 AM Stephan Ewen <[hidden email]> wrote:

Small addition to Till's comment:

 

In the case where file:// points to a mounted distributed file system (NFS, MapRFs, ...), then it actually works. The important thing is that the filesystem where the checkpoints go is replicated (fault tolerant) and accessible from all nodes.

 

On Thu, May 11, 2017 at 2:16 PM, Till Rohrmann <[hidden email]> wrote:

Hi Ayush,

you’re right that RocksDB is the recommend state backend because of the above-mentioned reasons. In order to make the recovery properly work, you have to configure a shared directory for the checkpoint data via state.backend.fs.checkpointdir. You can basically configure any file system which is supported by Hadoop (no HDFS required). The reason is that we use Hadoop to bridge between different file systems. The only thing you have to make sure is that you have the respective file system implementation in your class path.

I think you can access Windows Azure Blob Storage via Hadoop [1] similarly to access S3, for example.

If you use S3 to store your checkpoint data, then you will benefit from all the advantages of S3 but also suffer from its drawbacks (e.g. that list operations are more costly). But these are not specific to Flink.

A URL like file:// usually indicates a local file. Thus, if your Flink cluster is not running on a single machine, then this won’t work.

[1] https://hadoop.apache.org/docs/stable/hadoop-azure/index.html

Cheers,
Till

​

 

On Thu, May 11, 2017 at 10:41 AM, Ayush Goyal <[hidden email]> wrote:

Hello, 

I had a few questions regarding checkpoint storage options using 
RocksDBStateBackend. In the flink 1.2 documentation, it is the recommended state 
backend due to it's ability to store large states and asynchronous snapshotting. 
For high availabilty it seems HDFS is the recommended store for state backend 
data. In AWS deployment section, it is also mentioned that s3 can be used for 
storing state backend data. 

We don't want to depend on a hadoop cluster for flink deployment, so I had 
following questions: 

1. Can we use any storage backend supported by flink for storing RocksDB 
StateBackend data with file urls: there are quite a few supported as mentioned here: 
https://ci.apache.org/projects/flink/flink-docs-release-1.3/internals/filesystems.html
and here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md

2. Is there some work already done to support Windows Azure Blob Storage for 
storing State backend data? There are some docs here: 
https://github.com/apache/flink/blob/master/docs/dev/batch/connectors.md
can we utilize this for that? 

3. If utilizing S3 for state backend, is there any performance impact? 

4. For high availability can we use a NFS volume for state backend, with 
"file://" urls? Will there be any performance impact? 

 

PS: I posted this email earlier via nabble, but it's not showing up in apache archive. So sending again. Apologies if it results in multiple threads.

-- Ayush