spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Steve Loughran <ste...@hortonworks.com>
Subject Re: Remove Hadoop 1 support (Hadoop <2.2) for Spark 1.5?
Date Sat, 13 Jun 2015 08:39:24 GMT

> On 12 Jun 2015, at 17:12, Patrick Wendell <pwendell@gmail.com> wrote:
> 
>  For instance at Databricks we use
> the FileSystem library for talking to S3... every time we've tried to
> upgrade to Hadoop 2.X there have been significant regressions in
> performance and we've had to downgrade. That's purely anecdotal, but I
> think you have people out there using the Hadoop 1 bindings for whom
> upgrade would be a pain.

ah s3n. The unloved orphan FS, which has been fairly neglected as being non-strategic to anyone
but Amazon, who have a private fork. 

s3n broke in hadopo 2.4 where the upgraded Jets3t went in with some patch which swallowed
exceptions (nobody should ever do that) and as result would NPE on a seek(0) of a file of
length(0). HADOOP-10457. Fixed in Hadoop 2.5

Hadoop 2.6 has left S3n on maintenance out of fear of breaking more things, future work is
in s3a:,, which switched to the amazon awstoolkit JAR and moved the implementation to hadoop-aws
JAR. S3a promises: speed, partitioned upload, better auth. 

But: it's not ready for serious use in Hadoop 2.6, so don't try. You need the Hadoop 2.7 patches,
which are in ASF Hadoop 2.7, will be in HDP2.3, and have been picked up in CDH5.3. (HADOOP-11571).
For Spark, the fact that the block size is being returned as 0 in getFileStatus() could be
the killer.

Future work is going to improve performance and scale ( HADOOP-11694 )

Now, if spark is finding problems with s3a performance, tests for this would be great -complaints
on JIRAs too. There's not enough functional testing of analytics workloads against the object
stores, especially s3 and swift. If someone volunteers to add some optional test module for
object store testing, I'll help review it and suggest some tests to generate stress

That can be done without the leap to Hadoop 2 —though the proposed HADOOP-9565 work allowing
object stores to declare that they are and publish some of their consistency and atomicity
semantics will be Hadoop 2.8+. If you want your output committers to recognise when the destination
is an eventually constitent object store with O(n) directory rename and delete, that's where
the code will be.
Mime
View raw message