S3 checkpointing in AWS in Frankfurt

[Jonathan Share]

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Greg Hogan]

Hi Jonathan,

Which S3 storage class are you using? Do you have a breakdown of the S3 costs as storage / API calls / early deletes / data transfer?

Greg

On Wed, Nov 23, 2016 at 2:52 AM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Scott Kidder]

Hi Jonathan,

You might be better off creating a small Hadoop HDFS cluster just for the purpose of storing Flink checkpoint & savepoint data. Like you, I tried using S3 to persist Flink state, but encountered AWS SDK issues and felt like I was going down an ill-advised path. I then created a small 3-node HDFS cluster in the same region as my Flink hosts but distributed across 3 AZs. The checkpointing is very fast and, most importantly, just works.

Is there a firm requirement to use S3, or could you use HDFS instead?

Best,

--Scott Kidder

On Tue, Nov 22, 2016 at 11:52 PM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[rmetzger0]

Hi Jonathan,

have you tried using Amazon's latest EMR Hadoop distribution? Maybe they've fixed the issue in their for older Hadoop releases?

On Wed, Nov 23, 2016 at 4:38 PM, Scott Kidder <[hidden email]> wrote:

Hi Jonathan,

You might be better off creating a small Hadoop HDFS cluster just for the purpose of storing Flink checkpoint & savepoint data. Like you, I tried using S3 to persist Flink state, but encountered AWS SDK issues and felt like I was going down an ill-advised path. I then created a small 3-node HDFS cluster in the same region as my Flink hosts but distributed across 3 AZs. The checkpointing is very fast and, most importantly, just works.

Is there a firm requirement to use S3, or could you use HDFS instead?

Best,

--Scott Kidder

On Tue, Nov 22, 2016 at 11:52 PM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Foster, Craig]

I would suggest using EMRFS anyway, which is the way to access the S3 file system from EMR (using the same s3:// prefixes).  That said, you will run into the same shading issues in our build until the next release—which is coming up relatively shortly.

 

 

 

From: Robert Metzger <[hidden email]>
Reply-To: "[hidden email]" <[hidden email]>
Date: Wednesday, November 23, 2016 at 8:24 AM
To: "[hidden email]" <[hidden email]>
Subject: Re: S3 checkpointing in AWS in Frankfurt

 

Hi Jonathan,

 

have you tried using Amazon's latest EMR Hadoop distribution? Maybe they've fixed the issue in their for older Hadoop releases?

 

On Wed, Nov 23, 2016 at 4:38 PM, Scott Kidder <[hidden email]> wrote:

Hi Jonathan,

 

You might be better off creating a small Hadoop HDFS cluster just for the purpose of storing Flink checkpoint & savepoint data. Like you, I tried using S3 to persist Flink state, but encountered AWS SDK issues and felt like I was going down an ill-advised path. I then created a small 3-node HDFS cluster in the same region as my Flink hosts but distributed across 3 AZs. The checkpointing is very fast and, most importantly, just works.

 

Is there a firm requirement to use S3, or could you use HDFS instead?

 

Best,

 

--Scott Kidder

 

On Tue, Nov 22, 2016 at 11:52 PM, Jonathan Share <[hidden email]> wrote:

Hi,

 

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

 

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

 

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

 

Thanks in advance

Jonathan

 

 

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

 

 

[Jonathan Share]

Hi Greg,

Standard storage class, everything is on defaults, we've not done anything special with the bucket.

Cloud Watch only appears to give me total billing for S3 in general, I don't see a breakdown unless that's something I can configure somewhere.

Regards,

Jonathan

On 23 November 2016 at 16:29, Greg Hogan <[hidden email]> wrote:

Hi Jonathan,

Which S3 storage class are you using? Do you have a breakdown of the S3 costs as storage / API calls / early deletes / data transfer?

Greg

On Wed, Nov 23, 2016 at 2:52 AM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Jonathan Share]

Hi Scott,

Thanks for the suggestion, it sounds like you and I think alike, going over to hdfs sounds to me like the simplest solution.

There are no requirements to use S3, just another team member who is generally sceptical fearing that adding HDFS will add a new class of maintenance problems to our stack, and the project has a general goal of using managed services as much as possible so we wanted to try and make it work.

Regards,

Jonathan

On 23 November 2016 at 16:38, Scott Kidder <[hidden email]> wrote:

Hi Jonathan,

You might be better off creating a small Hadoop HDFS cluster just for the purpose of storing Flink checkpoint & savepoint data. Like you, I tried using S3 to persist Flink state, but encountered AWS SDK issues and felt like I was going down an ill-advised path. I then created a small 3-node HDFS cluster in the same region as my Flink hosts but distributed across 3 AZs. The checkpointing is very fast and, most importantly, just works.

Is there a firm requirement to use S3, or could you use HDFS instead?

Best,

--Scott Kidder

On Tue, Nov 22, 2016 at 11:52 PM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Jonathan Share]

We're not running on EMR (running Flink as a standalone cluster on Kubernetes on EC2). I assume that it's not possible to use EMRFS if not running on Amazon's EMR images.

On 23 November 2016 at 18:00, Foster, Craig <[hidden email]> wrote:

I would suggest using EMRFS anyway, which is the way to access the S3 file system from EMR (using the same s3:// prefixes).  That said, you will run into the same shading issues in our build until the next release—which is coming up relatively shortly.

 

 

 

From: Robert Metzger <[hidden email]>
Reply-To: "[hidden email]" <[hidden email]>
Date: Wednesday, November 23, 2016 at 8:24 AM
To: "[hidden email]" <[hidden email]>
Subject: Re: S3 checkpointing in AWS in Frankfurt

 

Hi Jonathan,

 

have you tried using Amazon's latest EMR Hadoop distribution? Maybe they've fixed the issue in their for older Hadoop releases?

 

On Wed, Nov 23, 2016 at 4:38 PM, Scott Kidder <[hidden email]> wrote:

Hi Jonathan,

 

You might be better off creating a small Hadoop HDFS cluster just for the purpose of storing Flink checkpoint & savepoint data. Like you, I tried using S3 to persist Flink state, but encountered AWS SDK issues and felt like I was going down an ill-advised path. I then created a small 3-node HDFS cluster in the same region as my Flink hosts but distributed across 3 AZs. The checkpointing is very fast and, most importantly, just works.

 

Is there a firm requirement to use S3, or could you use HDFS instead?

 

Best,

 

--Scott Kidder

 

On Tue, Nov 22, 2016 at 11:52 PM, Jonathan Share <[hidden email]> wrote:

Hi,

 

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

 

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

 

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

 

Thanks in advance

Jonathan

 

 

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

 

 

[Greg Hogan]

EMRFS looks to *add* cost (and consistency).

Storing an object to S3 costs "$0.005 per 1,000 requests", so $0.432/day at 1 Hz. Is the number of checkpoint files simply parallelism * number of operators? That could add up quickly.

Is the recommendation to run HDFS on EBS?

On Wed, Nov 23, 2016 at 12:40 PM, Jonathan Share <[hidden email]> wrote:

Hi Greg,

Standard storage class, everything is on defaults, we've not done anything special with the bucket.

Cloud Watch only appears to give me total billing for S3 in general, I don't see a breakdown unless that's something I can configure somewhere.

Regards,

Jonathan

On 23 November 2016 at 16:29, Greg Hogan <[hidden email]> wrote:

Hi Jonathan,

Which S3 storage class are you using? Do you have a breakdown of the S3 costs as storage / API calls / early deletes / data transfer?

Greg

On Wed, Nov 23, 2016 at 2:52 AM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324

[Stephan Ewen]

We have been looking for a while for some way to decouple the S3 filesystem support from Hadoop.

Does anyone know a good S3 connector library that works independent of Hadoop and EMRFS?

Best,

Stephan

On Wed, Nov 23, 2016 at 7:57 PM, Greg Hogan <[hidden email]> wrote:

EMRFS looks to *add* cost (and consistency).

Storing an object to S3 costs "$0.005 per 1,000 requests", so $0.432/day at 1 Hz. Is the number of checkpoint files simply parallelism * number of operators? That could add up quickly.

Is the recommendation to run HDFS on EBS?

On Wed, Nov 23, 2016 at 12:40 PM, Jonathan Share <[hidden email]> wrote:

Hi Greg,

Standard storage class, everything is on defaults, we've not done anything special with the bucket.

Cloud Watch only appears to give me total billing for S3 in general, I don't see a breakdown unless that's something I can configure somewhere.

Regards,

Jonathan

On 23 November 2016 at 16:29, Greg Hogan <[hidden email]> wrote:

Hi Jonathan,

Which S3 storage class are you using? Do you have a breakdown of the S3 costs as storage / API calls / early deletes / data transfer?

Greg

On Wed, Nov 23, 2016 at 2:52 AM, Jonathan Share <[hidden email]> wrote:

Hi,

I'm interested in hearing if anyone else has experience with using Amazon S3 as a state backend in the Frankfurt region. For political reasons we've been asked to keep all European data in Amazon's Frankfurt region. This causes a problem as the S3 endpoint in Frankfurt requires the use of AWS Signature Version 4 "This new Region supports only Signature Version 4" [1] and this doesn't appear to work with the Hadoop version that Flink is built against [2].

After some hacking we have managed to create a docker image with a build of Flink 1.2 master, copying over jar files from the hadoop 3.0.0-alpha1 package and this appears to work, for the most part but we still suffer from some classpath problems (conflicts between AWS API used in hadoop and those we want to use in out streams for interacting with Kinesis) and the whole thing feels a little fragile. Has anyone else tried this? Is there a simpler solution?

As a follow-up question, we saw that with checkpointing on three relatively simple streams set to 1 second, our S3 costs were higher than the EC2 costs for our entire infrastructure. This seems slightly disproportionate. For now we have reduced checkpointing interval to 10 seconds and that has greatly improved the cost projections graphed via Amazon Cloud Watch, but I'm interested in hearing other peoples experience with this. Is that the kind of billing level we can expect or is this a symptom of a mis-configuration? Is this a setup others are using? As we are using Kinesis as the source for all streams I don't see a huge risk with larger checkpoint intervals and our Sinks are designed to mostly tolerate duplicates (some improvements can be made). 

Thanks in advance

Jonathan

[1] https://aws.amazon.com/blogs/aws/aws-region-germany/

[2] https://issues.apache.org/jira/browse/HADOOP-13324