Integration with Cloud Infrastructures
- Spark Streaming and Object Storage
- Committing work into cloud storage safely and fast.
- Further Reading
All major cloud providers offer persistent data storage in object stores.
These are not classic “POSIX” file systems.
In order to store hundreds of petabytes of data without any single points of failure,
object stores replace the classic file system directory tree
with a simpler model of
object-name => data. To enable remote access, operations
on objects are usually offered as (slow) HTTP REST operations.
Spark can read and write data in object stores through filesystem connectors implemented in Hadoop or provided by the infrastructure suppliers themselves. These connectors make the object stores look almost like file systems, with directories and files and the classic operations on them such as list, delete and rename.
Important: Cloud Object Stores are Not Real Filesystems
While the stores appear to be filesystems, underneath they are still object stores, and the difference is significant
They cannot be used as a direct replacement for a cluster filesystem such as HDFS except where this is explicitly stated.
Key differences are:
- Changes to stored objects may not be immediately visible, both in directory listings and actual data access.
- The means by which directories are emulated may make working with them slow.
- Rename operations may be very slow and, on failure, leave the store in an unknown state.
- Seeking within a file may require new HTTP calls, hurting performance.
How does this affect Spark?
- Reading and writing data can be significantly slower than working with a normal filesystem.
- Some directory structures may be very inefficient to scan during query split calculation.
- The output of work may not be immediately visible to a follow-on query.
- The rename-based algorithm by which Spark normally commits work when saving an RDD, DataFrame or Dataset is potentially both slow and unreliable.
For these reasons, it is not always safe to use an object store as a direct destination of queries, or as an intermediate store in a chain of queries. Consult the documentation of the object store and its connector to determine which uses are considered safe.
In particular: without some form of consistency layer, Amazon S3 cannot be safely used as the direct destination of work with the normal rename-based committer.
With the relevant libraries on the classpath and Spark configured with valid credentials,
objects can be read or written by using their URLs as the path to data.
sparkContext.textFile("s3a://landsat-pds/scene_list.gz") will create
an RDD of the file
scene_list.gz stored in S3, using the s3a connector.
To add the relevant libraries to an application’s classpath, include the
module and its dependencies.
In Maven, add the following to the
pom.xml file, assuming
is set to the chosen version of Spark:
Commercial products based on Apache Spark generally directly set up the classpath for talking to cloud infrastructures, in which case this module may not be needed.
Spark jobs must authenticate with the object stores to access data within them.
- When Spark is running in a cloud infrastructure, the credentials are usually automatically set up.
AWS_SESSION_TOKENenvironment variables and sets the associated authentication options for the
s3aconnectors to Amazon S3.
- In a Hadoop cluster, settings may be set in the
- Authentication details may be manually added to the Spark configuration in
- Alternatively, they can be programmatically set in the
SparkConfinstance used to configure the application’s
Important: never check authentication secrets into source code repositories, especially public ones
Consult the Hadoop documentation for the relevant configuration and security options.
Each cloud connector has its own set of configuration parameters, again, consult the relevant documentation.
Recommended settings for writing to object stores
For object stores whose consistency model means that rename-based commits are safe
FileOutputCommitter v2 algorithm for performance; v1 for safety.
This does less renaming at the end of a job than the “version 1” algorithm.
As it still uses
rename() to commit files, it is unsafe to use
when the object store does not have consistent metadata/listings.
The committer can also be set to ignore failures when cleaning up temporary files; this reduces the risk that a transient network problem is escalated into a job failure:
The original v1 commit algorithm renames the output of successful tasks to a job attempt directory, and then renames all the files in that directory into the final destination during the job commit phase:
The slow performance of mimicked renames on Amazon S3 makes this algorithm very, very slow. The recommended solution to this is switch to an S3 “Zero Rename” committer (see below).
For reference, here are the performance and safety characteristics of different stores and connectors when renaming directories:
|Store||Connector||Directory Rename Safety||Rename Performance|
|Azure Datalake Gen 2||abfs||Safe||O(1)|
|Google Cloud Storage||gs||Safe||O(1)|
As storing temporary files can run up charges; delete
"_temporary" on a regular basis.
Parquet I/O Settings
For optimal performance when working with Parquet data use the following settings:
spark.hadoop.parquet.enable.summary-metadata false spark.sql.parquet.mergeSchema false spark.sql.parquet.filterPushdown true spark.sql.hive.metastorePartitionPruning true
These minimise the amount of data read during queries.
ORC I/O Settings
For best performance when working with ORC data, use these settings:
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
Again, these minimise the amount of data read during queries.
Spark Streaming and Object Storage
Spark Streaming can monitor files added to object stores, by
FileInputDStream to monitor a path in the store through a call to
The time to scan for new files is proportional to the number of files under the path, not the number of new files, so it can become a slow operation. The size of the window needs to be set to handle this.
Files only appear in an object store once they are completely written; there is no need for a workflow 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.
Streams should only be checkpointed to a store implementing a fast and atomic
rename()operation. Otherwise the checkpointing may be slow and potentially unreliable.
Committing work into cloud storage safely and fast.
As covered earlier, commit-by-rename is dangerous on any object store which exhibits eventual consistency (example: S3), and often slower than classic filesystem renames.
Some object store connectors provide custom committers to commit tasks and jobs without using rename. In versions of Spark built with Hadoop 3.1 or later, the S3A connector for AWS S3 is such a committer.
Instead of writing data to a temporary directory on the store for renaming, these committers write the files to the final destination, but do not issue the final POST command to make a large “multi-part” upload visible. Those operations are postponed until the job commit itself. As a result, task and job commit are much faster, and task failures do not affect the result.
To switch to the S3A committers, use a version of Spark was built with Hadoop 3.1 or later, and switch the committers through the following options.
spark.hadoop.fs.s3a.committer.name directory spark.sql.sources.commitProtocolClass org.apache.spark.internal.io.cloud.PathOutputCommitProtocol spark.sql.parquet.output.committer.class org.apache.spark.internal.io.cloud.BindingParquetOutputCommitter
It has been tested with the most common formats supported by Spark.
More details on these committers can be found in the latest Hadoop documentation.
Here is the documentation on the standard connectors both from Apache and the cloud providers.
- OpenStack Swift.
- Azure Blob Storage and Azure Datalake Gen 2.
- Azure Data Lake Gen 1.
- Hadoop-AWS module (Hadoop 3.x).
- Amazon S3 via S3A and S3N (Hadoop 2.x).
- Amazon EMR File System (EMRFS). From Amazon.
- Google Cloud Storage Connector for Spark and Hadoop. From Google.
- The Azure Blob Filesystem driver (ABFS)
- IBM Cloud Object Storage connector for Apache Spark: Stocator, IBM Object Storage. From IBM.