You are viewing a plain text version of this content. The canonical link for it is here.

Posted to user@spark.apache.org by abbim <ab...@amazon.com> on 2017/09/07 06:02:24 UTC

CSV write to S3 failing silently with partial completion

Hi all,
My team has been experiencing a recurring unpredictable bug where only a
partial write to CSV in S3 on one partition of our Dataset is performed. For
example, in a Dataset of 10 partitions written to CSV in S3, we might see 9
of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However, the
job does not exit with an error code.

This becomes problematic in the following ways:
1. When we copy the data to Redshift, we get a bad decrypt error on the
partial file, suggesting that the failure occurred at a weird byte in the
file. 
2. We lose data - sometimes as much as 10%.

We don't see this problem with parquet format, which we also use, but moving
all of our data to parquet is not currently feasible. We're using the Java
API with Spark 2.2 and Amazon EMR 5.8, code is a simple as this:
df.write().csv("s3://some-bucket/some_location"). We're experiencing the
issue 1-3x/week on a daily job and are unable to reliably reproduce the
problem. 

Any thoughts on why we might be seeing this and how to resolve?
Thanks in advance.



--
Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org

Re: CSV write to S3 failing silently with partial completion

Posted by Steve Loughran <st...@hortonworks.com>.

On 7 Sep 2017, at 18:36, Mcclintic, Abbi <ab...@amazon.com>> wrote:

Thanks all – couple notes below.

Generally all our partitions are of equal size (ie on a normal day in this particular case I see 10 equally sized partitions of 2.8 GB). We see the problem with repartitioning and without – in this example we are repartitioning to 10 but we also see the problem without any repartitioning when the default partition count is 200. We know that data loss is occurring because we have a final quality gate that counts the number of rows and halts the process if we see too large of a drop.

We have one use case where the data needs to be read on a local machine after processing and one use case of copying to redshift. Regarding the redshift copy, it gets a bit complicated owing to VPC and encryption requirements so we haven’t looked into using the JDBC driver yet.

My understanding was that Amazon EMR does not support s3a<https://aws.amazon.com/premiumsupport/knowledge-center/emr-file-system-s3/>, but it may be worth looking into.

1. No, it doesn't
2. You can't currently use s3a as a direct destination of work due to s3 not being consistent, not without a consistency layer on top (S3Guard, etc)

We may also try a combination of writing to HDFS combined with s3distcp.


+1

Re: CSV write to S3 failing silently with partial completion

Posted by "Mcclintic, Abbi" <ab...@amazon.com>.

Thanks all – couple notes below.

Generally all our partitions are of equal size (ie on a normal day in this particular case I see 10 equally sized partitions of 2.8 GB). We see the problem with repartitioning and without – in this example we are repartitioning to 10 but we also see the problem without any repartitioning when the default partition count is 200. We know that data loss is occurring because we have a final quality gate that counts the number of rows and halts the process if we see too large of a drop.

We have one use case where the data needs to be read on a local machine after processing and one use case of copying to redshift. Regarding the redshift copy, it gets a bit complicated owing to VPC and encryption requirements so we haven’t looked into using the JDBC driver yet.

My understanding was that Amazon EMR does not support s3a<https://aws.amazon.com/premiumsupport/knowledge-center/emr-file-system-s3/>, but it may be worth looking into. We may also try a combination of writing to HDFS combined with s3distcp.

Thanks,

Abbi

On 9/7/17, 7:50 AM, "Patrick Alwell" <pa...@hortonworks.com> wrote:

Sounds like an S3 bug. Can you replicate locally with HDFS?

Try using S3a protocol too; there is a jar you can leverage like so: spark-submit --packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.3 my_spark_program.py

EMR can sometimes be buggy. :/

You could also try leveraging EC2 nodes and manually creating a cluster with password less SSH.

But I feel your pain man, I’ve had weird issues with Redshift and EMR as well.

Let me know if you can or can’t replicate locally; and I can bring it up with our S3 team for the next release of HDP and we can file a bug with AWS.

-Pat

On 9/7/17, 2:59 AM, "JG Perrin" <jp...@lumeris.com> wrote:

Are you assuming that all partitions are of equal size? Did you try with more partitions (like repartitioning)? Does the error always happen with the last (or smaller) file? If you are sending to redshift, why not use the JDBC driver?

-----Original Message-----

From: abbim [mailto:abbim@amazon.com]

Sent: Thursday, September 07, 2017 1:02 AM

To: user@spark.apache.org

Subject: CSV write to S3 failing silently with partial completion

Hi all,

My team has been experiencing a recurring unpredictable bug where only a partial write to CSV in S3 on one partition of our Dataset is performed. For example, in a Dataset of 10 partitions written to CSV in S3, we might see 9 of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However, the job does not exit with an error code.

This becomes problematic in the following ways:

1. When we copy the data to Redshift, we get a bad decrypt error on the partial file, suggesting that the failure occurred at a weird byte in the file.

2. We lose data - sometimes as much as 10%.

df.write().csv("s3://some-bucket/some_location"). We're experiencing the issue 1-3x/week on a daily job and are unable to reliably reproduce the problem.

Any thoughts on why we might be seeing this and how to resolve?

Thanks in advance.

Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

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

To unsubscribe e-mail: user-unsubscribe@spark.apache.org

______________________________________________________________________

This electronic transmission and any documents accompanying this electronic transmission contain confidential information belonging to the sender. This information may contain confidential health information that is legally privileged. The information is intended only for the use of the individual or entity named above. The authorized recipient of this transmission is prohibited from disclosing this information to any other party unless required to do so by law or regulation and is required to delete or destroy the information after its stated need has been fulfilled. If you are not the intended recipient, you are hereby notified that any disclosure, copying, distribution or the taking of any action in reliance on or regarding the contents of this electronically transmitted information is strictly prohibited. If you have received this E-mail in error, please notify the sender and delete this message immediately.

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

To unsubscribe e-mail: user-unsubscribe@spark.apache.org

Re: CSV write to S3 failing silently with partial completion

Posted by Patrick Alwell <pa...@hortonworks.com>.

Sounds like an S3 bug. Can you replicate locally with HDFS?

Try using S3a protocol too; there is a jar you can leverage like so: spark-submit --packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.3 my_spark_program.py

EMR can sometimes be buggy. :/

You could also try leveraging EC2 nodes and manually creating a cluster with password less SSH.

But I feel your pain man, I’ve had weird issues with Redshift and EMR as well.

Let me know if you can or can’t replicate locally; and I can bring it up with our S3 team for the next release of HDP and we can file a bug with AWS.

-Pat

On 9/7/17, 2:59 AM, "JG Perrin" <jp...@lumeris.com> wrote:

-----Original Message-----
From: abbim [mailto:abbim@amazon.com]
Sent: Thursday, September 07, 2017 1:02 AM
To: user@spark.apache.org
Subject: CSV write to S3 failing silently with partial completion

Hi all,
My team has been experiencing a recurring unpredictable bug where only a partial write to CSV in S3 on one partition of our Dataset is performed. For example, in a Dataset of 10 partitions written to CSV in S3, we might see 9 of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However, the job does not exit with an error code.

This becomes problematic in the following ways:
1. When we copy the data to Redshift, we get a bad decrypt error on the partial file, suggesting that the failure occurred at a weird byte in the file.
2. We lose data - sometimes as much as 10%.

We don't see this problem with parquet format, which we also use, but moving all of our data to parquet is not currently feasible. We're using the Java API with Spark 2.2 and Amazon EMR 5.8, code is a simple as this:
df.write().csv("s3://some-bucket/some_location"). We're experiencing the issue 1-3x/week on a daily job and are unable to reliably reproduce the problem.

Any thoughts on why we might be seeing this and how to resolve?
Thanks in advance.

--
Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org

______________________________________________________________________
This electronic transmission and any documents accompanying this electronic transmission contain confidential information belonging to the sender. This information may contain confidential health information that is legally privileged. The information is intended only for the use of the individual or entity named above. The authorized recipient of this transmission is prohibited from disclosing this information to any other party unless required to do so by law or regulation and is required to delete or destroy the information after its stated need has been fulfilled. If you are not the intended recipient, you are hereby notified that any disclosure, copying, distribution or the taking of any action in reliance on or regarding the contents of this electronically transmitted information is strictly prohibited. If you have received this E-mail in error, please notify the sender and delete this message immediately.

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org

RE: CSV write to S3 failing silently with partial completion

Posted by JG Perrin <jp...@lumeris.com>.

Are you assuming that all partitions are of equal size? Did you try with more partitions (like repartitioning)? Does the error always happen with the last (or smaller) file? If you are sending to redshift, why not use the JDBC driver?

-----Original Message-----
From: abbim [mailto:abbim@amazon.com] 
Sent: Thursday, September 07, 2017 1:02 AM
To: user@spark.apache.org
Subject: CSV write to S3 failing silently with partial completion

Hi all,
My team has been experiencing a recurring unpredictable bug where only a partial write to CSV in S3 on one partition of our Dataset is performed. For example, in a Dataset of 10 partitions written to CSV in S3, we might see 9 of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However, the job does not exit with an error code.

This becomes problematic in the following ways:
1. When we copy the data to Redshift, we get a bad decrypt error on the partial file, suggesting that the failure occurred at a weird byte in the file. 
2. We lose data - sometimes as much as 10%.

We don't see this problem with parquet format, which we also use, but moving all of our data to parquet is not currently feasible. We're using the Java API with Spark 2.2 and Amazon EMR 5.8, code is a simple as this:
df.write().csv("s3://some-bucket/some_location"). We're experiencing the issue 1-3x/week on a daily job and are unable to reliably reproduce the problem. 

Any thoughts on why we might be seeing this and how to resolve?
Thanks in advance.

--
Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org

______________________________________________________________________
This electronic transmission and any documents accompanying this electronic transmission contain confidential information belonging to the sender.  This information may contain confidential health information that is legally privileged.  The information is intended only for the use of the individual or entity named above.  The authorized recipient of this transmission is prohibited from disclosing this information to any other party unless required to do so by law or regulation and is required to delete or destroy the information after its stated need has been fulfilled.  If you are not the intended recipient, you are hereby notified that any disclosure, copying, distribution or the taking of any action in reliance on or regarding the contents of this electronically transmitted information is strictly prohibited.  If you have received this E-mail in error, please notify the sender and delete this message immediately.

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org

Re: CSV write to S3 failing silently with partial completion

Posted by "Mcclintic, Abbi" <ab...@amazon.com>.

Hi folks,
We appear to have mitigated the issue by including the following configurations to our jobs, with significant improvement in S3 consistency with CSV and JSON (which turned out to be worse than CSV initially):

spark.speculation=false
spark.hadoop.mapreduce.fileoutputcommitter.algorithm.version=1

Still not really sure of the root cause, but this has at least stopped the bleeding for my team and so far hasn’t caused any large degradation in runtime for our jobs.

I’ve looked into the spark-redshift connector but I don’t think it supports client side encryption which is a requirement for our data and wouldn’t solve the problem for our data used outside of Redshift.

Hope that helps someone else out if you hit the same issue.

-Abbi


From: Gourav Sengupta <go...@gmail.com>
Date: Monday, September 11, 2017 at 6:32 AM
To: "Mcclintic, Abbi" <ab...@amazon.com>
Cc: user <us...@spark.apache.org>
Subject: Re: CSV write to S3 failing silently with partial completion

Hi,

Can you please let me know the following:
1. Why are you using JAVA?
2. The way you are creating the SPARK cluster
3. The way you are initiating SPARK session or context
4. Are you able to query the data that is written to S3 using a SPARK dataframe and validate that the number of rows in the source are same as the ones written to target?
5. how are you loading the data to Redshift (cluster size, version, command, compression, command, manifest file)
6. using Redshift JDBC (https://github.com/databricks/spark-redshift) you will have to play around with it a bit to understand how it works (be careful that it does not drop the table at target Redshift database)

Regards,
Gourav

On Thu, Sep 7, 2017 at 7:02 AM, abbim <ab...@amazon.com>> wrote:
Hi all,
My team has been experiencing a recurring unpredictable bug where only a
partial write to CSV in S3 on one partition of our Dataset is performed. For
example, in a Dataset of 10 partitions written to CSV in S3, we might see 9
of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However, the
job does not exit with an error code.

This becomes problematic in the following ways:
1. When we copy the data to Redshift, we get a bad decrypt error on the
partial file, suggesting that the failure occurred at a weird byte in the
file.
2. We lose data - sometimes as much as 10%.

We don't see this problem with parquet format, which we also use, but moving
all of our data to parquet is not currently feasible. We're using the Java
API with Spark 2.2 and Amazon EMR 5.8, code is a simple as this:
df.write().csv("s3://some-bucket/some_location"). We're experiencing the
issue 1-3x/week on a daily job and are unable to reliably reproduce the
problem.

Any thoughts on why we might be seeing this and how to resolve?
Thanks in advance.



--
Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/

---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org<ma...@spark.apache.org>

Re: CSV write to S3 failing silently with partial completion

Posted by Gourav Sengupta <go...@gmail.com>.

Hi,

Can you please let me know the following:
1. Why are you using JAVA?
2. The way you are creating the SPARK cluster
3. The way you are initiating SPARK session or context
4. Are you able to query the data that is written to S3 using a SPARK
dataframe and validate that the number of rows in the source are same as
the ones written to target?
5. how are you loading the data to Redshift (cluster size, version,
command, compression, command, manifest file)
6. using Redshift JDBC (https://github.com/databricks/spark-redshift) you
will have to play around with it a bit to understand how it works (be
careful that it does not drop the table at target Redshift database)

Regards,
Gourav

On Thu, Sep 7, 2017 at 7:02 AM, abbim <ab...@amazon.com> wrote:

> Hi all,
> My team has been experiencing a recurring unpredictable bug where only a
> partial write to CSV in S3 on one partition of our Dataset is performed.
> For
> example, in a Dataset of 10 partitions written to CSV in S3, we might see 9
> of the partitions as 2.8 GB in size, but one of them as 1.6 GB. However,
> the
> job does not exit with an error code.
>
> This becomes problematic in the following ways:
> 1. When we copy the data to Redshift, we get a bad decrypt error on the
> partial file, suggesting that the failure occurred at a weird byte in the
> file.
> 2. We lose data - sometimes as much as 10%.
>
> We don't see this problem with parquet format, which we also use, but
> moving
> all of our data to parquet is not currently feasible. We're using the Java
> API with Spark 2.2 and Amazon EMR 5.8, code is a simple as this:
> df.write().csv("s3://some-bucket/some_location"). We're experiencing the
> issue 1-3x/week on a daily job and are unable to reliably reproduce the
> problem.
>
> Any thoughts on why we might be seeing this and how to resolve?
> Thanks in advance.
>
>
>
> --
> Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/
>
> ---------------------------------------------------------------------
> To unsubscribe e-mail: user-unsubscribe@spark.apache.org
>
>