Github user srowen commented on a diff in the pull request:
https://github.com/apache/spark/pull/12004#discussion_r113697831
--- 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
+spark.hadoop.mapreduce.fileoutputcommitter.cleanup-failures.ignored true
+```
+
+There's also the option of skipping the cleanup of temporary files in the
output directory.
+Enabling this option eliminates a small delay caused by listing and
deleting any such files.
+
+```
+spark.hadoop.mapreduce.fileoutputcommitter.cleanup.skipped true
+```
+
+Bear in mind that storing temporary files can run up charges; Delete
+directories called `"_temporary"` on a regular basis to avoid this.
+
+
+### YARN Scheduler settings
+
+When running Spark in a YARN cluster running in EC2, turning off locality
avoids any delays
+waiting for the scheduler to find a node close to the data.
+
+```xml
+ <property>
+ <name>yarn.scheduler.capacity.node-locality-delay</name>
+ <value>0</value>
+ </property>
+```
+
+This has to be set in the YARN cluster configuration, not in the Spark
configuration.
+
+### Parquet I/O Settings
+
+For optimal performance when reading files saved in the Apache Parquet
format,
+read and write operations must be minimized, including generation of
summary metadata,
+and coalescing metadata from multiple files. The `filterPushdown` option
+enables the Parquet library to optimize data reads itself, potentially
saving bandwidth.
+
+```
+spark.hadoop.parquet.enable.summary-metadata false
+spark.sql.parquet.mergeSchema false
+spark.sql.parquet.filterPushdown true
+spark.sql.hive.metastorePartitionPruning true
+```
+
+### ORC I/O Settings
+
+For optimal performance when reading files saved in the Apache ORC format,
+read and write operations must be minimized. Here are the options to
achieve this.
+
+
+```
+spark.sql.orc.filterPushdown true
+spark.sql.orc.splits.include.file.footer true
+spark.sql.orc.cache.stripe.details.size 10000
+spark.sql.hive.metastorePartitionPruning true
+```
+
+The `filterPushdown` option enables the ORC library to optimize data reads
itself,
+potentially saving bandwidth.
+
+The `spark.sql.orc.splits.include.file.footer` option means that the ORC
file footer information,
+is passed around with the file information âso eliminating the need to
reread this data.
+
+## <a name="authenticating"></a>Authenticating with Object Stores
+
+Apart from the special case of public read-only data, all object stores
+require callers to authenticate themselves.
+To do this, the Spark context must be configured with the authentication
+details of the object store.
+
+1. In a YARN cluster, this may be done automatically in the
`core-site.xml` file.
+1. When Spark is running in cloud infrastructure (for example, on Amazon
EC2, Google Cloud or
+Microsoft Azure), the authentication details may be automatically derived
from information
+available to the VM.
+1. `spark-submit` automatically picks up the contents of `AWS_ACCESS_KEY`
and `AWS_SECRET_KEY`
+environment variables and sets the associated configuration parameters
for`s3n` and `s3a`
+to these values. This essentially propagates the values across the Spark
cluster.
+1. Authentication details may be manually added to the Spark configuration
+1. Alternatively, they can be programmatically added. *Important: never
put authentication
+secrets in source code. They will be compromised*.
+
+It is critical that the credentials used to access object stores are kept
secret. Not only can
+they be abused to run up compute charges, they can be used to read and
alter private data.
+
+1. If adding login details to a spark configuration file, do not share
this file, including
+attaching to bug reports or committing it to SCM repositories.
+1. Have different accounts for access to the storage for each application,
+each with access rights restricted to those object storage
buckets/containers used by the
+application.
+1. If the object store supports any form of session credential (e.g.
Amazon's STS), issue
+session credentials for the expected lifetime of the application.
+1. When using a version of Spark with with Hadoop 2.8+ libraries, consider
using Hadoop
+credential files to store secrets, referencing
+these files in the relevant ID/secret properties of the XML configuration
file.
+
+
+## <a name="object_stores"></a>Object stores and Their Library Dependencies
+
+The different object stores supported by Spark depend on specific Hadoop
versions,
+and require specific Hadoop JARs and dependent Java libraries on the
classpath.
+
+<table class="table">
+ <tr><th>Schema</th><th>Store</th><th>Details</th></tr>
+ <tr>
+ <td><code>s3a://</code></td>
+ <td>Amazon S3</a>
+ <td>
+ Recommended S3 client for Spark releases built on Apache Hadoop 2.7 or
later.
+ </td>
+ </tr>
+ <tr>
+ <td><code>s3n://</code></td>
+ <td>Amazon S3</a>
+ <td>
+ Deprected S3 client; only use for Spark releases built on Apache
Hadoop 2.6 or earlier.
+ </td>
+ </tr>
+ <tr>
+ <td><code>s3://</code></td>
+ <td>Amazon S3 on Amazon EMR</a>
+ <td>
+ Amazon's own S3 client; use only and exclusivley in Amazon EMR.
+ </td>
+ </tr>
+ <tr>
+ <td><code>wasb://</code></td>
+ <td>Azure Storage</a>
+ <td>
+ Client for Microsoft Azure Storage; since Hadoop 2.7.
+ </td>
+ </tr>
+ <tr>
+ <td><code>swift://</code></td>
+ <td>OpenStack Swift</a>
+ <td>
+ Client for OpenStack Swift object stores.
+ </td>
+ </tr>
+ <tr>
+ <td><code>gs://</code></td>
+ <td>Google Cloud Storage</a>
+ <td>
+ Google's client for their cloud object store.
+ </td>
+ </tr>
+</table>
+
+
+### <a name="working_with_amazon_s3"></a>Working with Amazon S3
+
+Amazon's S3 object store is probably the most widely used object store
âit is also the one
--- End diff --
"store -it" -> "store. It"
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
