[
https://issues.apache.org/jira/browse/SPARK-7481?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15389730#comment-15389730
]
Steve Loughran commented on SPARK-7481:
---------------------------------------
ps, latest s3a state
# [Object stores in
production|http://slideshare.net/HadoopSummit/hadoop-cloud-storage-object-store-integration-in-production]
# [Latest s3a
docs|https://github.com/apache/hadoop/blob/trunk/hadoop-tools/hadoop-aws/src/site/markdown/tools/hadoop-aws/index.md].
The options I'm going to recomment for working with ORC (or other data with
forward & backward seeks on read) are:
{code}
spark.hadoop.mapreduce.fileoutputcommitter.algorithm.version 2
spark.hadoop.fs.s3a.experimental.input.fadvise random
spark.hadoop.fs.s3a.readahead.range = 131072
spark.hadoop.fs.s3a.socket.send.buffer = 16384
spark.hadoop.fs.s3a.socket.recv.buffer = 16384
{code}
There are some other tunables (as are those ranges and buffers). fadvise=random
is fantastic on random access/positioned read; kills sequential scans though
things like CSV files. Use only when appropriate.
Spark will automatically get the speedups in S3A. What it will also need (and
which I haven't started on), is turning the spark code itself the way that
[~rajesh.balamohan] and [~ashutoshc] are doing for Hive: cache and re-use all
FileStatus results, use listFiles(recursive=true) for tree listing, move all
rename/deletes for cleanup off the query path, etc, etc.
This patch is just step 1: packaging and basic integration tests & hadoop-aws
regression testing —not the tuning which spark will need for maximum object
store perf (none of which will hurt HDFS performance, BTW)
> Add spark-cloud module to pull in aws+azure object store FS accessors; test
> integration
> ---------------------------------------------------------------------------------------
>
> Key: SPARK-7481
> URL: https://issues.apache.org/jira/browse/SPARK-7481
> Project: Spark
> Issue Type: Improvement
> Components: Build
> Affects Versions: 1.3.1
> Reporter: Steve Loughran
>
> To keep the s3n classpath right, to add s3a, swift & azure, the dependencies
> of spark in a 2.6+ profile need to add the relevant object store packages
> (hadoop-aws, hadoop-openstack, hadoop-azure)
> this adds more stuff to the client bundle, but will mean a single spark
> package can talk to all of the stores.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
