Github user steveloughran commented on a diff in the pull request:
https://github.com/apache/spark/pull/12004#discussion_r113970275
--- Diff: docs/cloud-integration.md ---
@@ -0,0 +1,512 @@
+---
+layout: global
+displayTitle: Integration with Cloud Infrastructures
+title: Integration with Cloud Infrastructures
+description: Introduction to cloud storage support in Apache Spark
SPARK_VERSION_SHORT
+---
+<!---
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. See accompanying LICENSE file.
+-->
+
+* This will become a table of contents (this text will be scraped).
+{:toc}
+
+## <a name="introduction"></a>Introduction
+
+
+All the public cloud infrastructures, Amazon AWS, Microsoft Azure, Google
GCS and others offer
+persistent data storage systems, "object stores". These are not quite the
same as classic file
+systems: in order to scale to hundreds of Petabytes, without any single
points of failure
+or size limits, object stores, "blobstores", have a simpler model of `name
=> data`.
+
+Apache Spark can read or write data in object stores for data access.
+through filesystem connectors implemented in Apache Hadoop or provided by
third-parties.
+These libraries make the object stores look *almost* like filesystems,
with directories and
+operations on files (rename) and directories (create, rename, delete)
which mimic
+those of a classic filesystem. Because of this, Spark and Spark-based
applications
+can work with object stores, generally treating them as as if they were
slower-but-larger filesystems.
+
+With these connectors, Apache Spark supports object stores as the source
+of data for analysis, including Spark Streaming and DataFrames.
+
+
+## <a name="quick_start"></a>Quick Start
+
+Provided the relevant libraries are on the classpath, and Spark is
configured with your credentials,
+objects in an object store can be can be read or written through URLs
which uses the name of the
+object store client as the schema and the bucket/container as the hostname.
+
+
+### Dependencies
+
+The Spark application neeeds the relevant Hadoop libraries, which can
+be done by including the `spark-hadoop-cloud` module for the specific
version of spark used.
+
+The Spark application should include <code>hadoop-openstack</code>
dependency, which can
+be done by including the `spark-hadoop-cloud` module for the specific
version of spark used.
+For example, for Maven support, add the following to the
<code>pom.xml</code> file:
+
+{% highlight xml %}
+<dependencyManagement>
+ ...
+ <dependency>
+ <groupId>org.apache.spark</groupId>
+ <artifactId>spark-hadoop-cloud_2.11</artifactId>
+ <version>${spark.version}</version>
+ </dependency>
+ ...
+</dependencyManagement>
+{% endhighlight %}
+
+If using the Scala 2.10-compatible version of Spark, the artifact is of
course `spark-hadoop-cloud_2.10`.
+
+### Basic Use
+
+You can refer to data in an object store just as you would data in a
filesystem, by
+using a URL to the data in methods like `SparkContext.textFile()` to read
data,
+`saveAsTextFile()` to write it back.
+
+
+Because object stores are viewed by Spark as filesystems, object stores can
+be used as the source or destination of any spark work âbe it batch,
SQL, DataFrame,
+Streaming or something else.
+
+The steps to do so are as follows
+
+1. Use the full URI to refer to a bucket, including the prefix for the
client-side library
+to use. Example: `s3a://landsat-pds/scene_list.gz`
+1. Have the Spark context configured with the authentication details of
the object store.
+In a YARN cluster, this may also be done in the `core-site.xml` file.
+
+
+## <a name="output"></a>Object Stores as a substitute for HDFS
+
+As the examples show, you can write data to object stores. However, that
does not mean
+That they can be used as replacements for a cluster-wide filesystem.
+
+The full details are covered in [Cloud Object Stores are Not Real
Filesystems](#cloud_stores_are_not_filesystems).
+
+The brief summary is:
+
+| Object Store Connector | Replace HDFS? |
+|-----------------------------|--------------------|
+| `s3a://` `s3n://` from the ASF | No |
+| Amazon EMR `s3://` | Yes |
+| Microsoft Azure `wasb://` | Yes |
+| OpenStack `swift://` | No |
+
+It is possible to use any of the object stores as a destination of work,
i.e. use
+`saveAsTextFile()` or `save()` to save data there, but the commit process
may be slow
+and, unreliable in the presence of failures.
+
+It is faster and safer to use the cluster filesystem as the destination of
Spark jobs,
+using that data as the data for follow-on work. The final results can
+be persisted in to the object store using `distcp`.
+
+#### <a name="checkpointing"></a>Spark Streaming and object stores
+
+Spark Streaming can monitor files added to object stores, by
+creating a `FileInputDStream` DStream monitoring a path under a bucket
through
+`StreamingContext.textFileStream()`.
+
+
+1. The time to scan for new files is proportional to the number of files
+under the path ânot the number of *new* files, and that it can become a
slow operation.
+The size of the window needs to be set to handle this.
+
+1. Files only appear in an object store once they are completely written;
there
+is no need for a worklow of write-then-rename to ensure that files aren't
picked up
+while they are still being written. Applications can write straight to the
monitored directory.
+
+1. Streams should only be checkpointed to an object store considered
compatible with
+HDFS. Otherwise the checkpointing will be slow and potentially unreliable.
+
+### Recommended settings for writing to object stores
+
+Here are the settings to use when writing to object stores. This uses the
"version 2" algorithm
+for committing files âwhich does less renaming than the v1 algorithm.
Speculative execution is
+disabled to avoid multiple writers corrupting the output.
+
+```
+spark.speculation false
+spark.hadoop.mapreduce.fileoutputcommitter.algorithm.version 2
--- End diff --
That `FileOutputCommitter` code is a mess: I've stepped through it
repeatedly and never quite worked out what they do. As usual, the design comes
down to big Y! queries and things that went wrong. A big part of the design is
to handle the failure of entire MR jobs and their restart, trying to recover
all data already generated by the first one and committed into that first
attempt.
These were MR jobs with many workers taking hours, the probability of
failure is high. The faster you can get the work done, or the fewer executors
you need, the less frequent failures are. That whole scenario "new app instance
trying to recover incomplete work of previous instance" doesn't exist any more.
(What may still exist though is: old app instance still running due to some
network partition)
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
