Skip to main content

Optimizing Spark Job Performance With Apache Ignite (Part 1)

The Apache Ignite bookPortions of this article were taken from the book The Apache Ignite book. If it got you interested, check out the rest of the book for more helpful information. Spark Data frame is included in new book.

Apache Ignite offers several ways to improve a Spark job's performance: Ignite RDD, which represents an Ignite cache as a Spark RDD abstraction, and Ignite IGFS, an in-memory file system that can be transparently plugged into Spark deployments. Ignite RDD allows easily sharing states in-memory between different Spark jobs or applications. With Ignite in-memory shares RDDs, any Spark job can put some data into an Ignite cache that other Spark jobs can access later. Ignite RDD is implemented as a view over the Ignite distributed cache, which can be deployed either within the Spark job execution process or on a Spark worker.


Before we move on to more advanced topics, let's have a look at the history of Spark and what kinds of problems can be solved by Ignite RDDs.
Apache Spark was invented by AMPLab for fast computation. It was built on top of Hadoop MapReduce and extends the MapReduce model to efficiently use more types of operations, such as interactive queries and stream processing.
The main difference between Spark and Hadoop MapReduce is that during execution, Spark tries to keep data in memory, whereas Hadoop MapReduce shuffles data into and out of disk. Hadoop MapReduce takes significant time to write intermediate data to disk and read it back. The elimination of these redundant disk operations makes Spark magnitudes faster. Spark can store data (intermediately) into memory without any I/O, so you can keep operating on the same data very quickly.
In order to store data into memory, Spark provides special dataset named Spark RDD. Spark RDD stands for Spark Resilient Distributed Dataset. Spark RDD has fundamental components of the Apache Spark large-scale data processing framework. The following illustration shows iterative operations on Spark RDD.
Note that the above figure is obtained from the Spark documentation. Spark RDD is an immutable, fault-tolerant distributed collection of data elements. You can imagine Spark RDD as a Hadoop HDFS in memory. Spark RDD supports two types of operations:
  1. Transformations, which create a new dataset from existing one
  2. Actions, which returns a value by performing a computation on the RDD (as shown in the next figure)
Spark RDD is created through the use of Spark transformation functions. Spark transformation functions can create Spark RDDs from various sources, such as text files. In addition to creating Spark RDDs from the text files, Spark RDDs may be created from external storage such as RDBMS, HBase, Cassandra, or any other data source compatible with Hadoop input format.
Most of the time, Spark RDDs are transformed from one RDD to another new Spark RDD in order to prepare the dataset for future processing. Let's consider the following data transformations steps in Spark:
  1. Load a text file with airline names and arrival times for any airport in RDD1.
  2. Load a text file with airline names and flight delay information for any airport into RDD2.
  3. Join RDD1 and RDD2 by airline names to get RDD3.
  4. Map on RDD3 to get a nice report for each airline as RDD4.
  5. Save RDD4 to file.
  6. Map RDD2 to extract the information of flight delay for certain airlines to get RDD5.
  7. Aggregate the RDD5 to get a count of how many flights are delayed for each airline as RDD6.
  8. Save the RDD6 into HDFS.
Spark RDDs are utilized to perform computations on an RDD dataset through Spark actions such as count or reduce. But there is a single problem with the Spark RDD: Spark RDD can't share between Spark Jobs or SparkContext because Spark RDD is bound to a Spark application. With native Spark distribution, the only way to share RDDs between different Spark jobs is to write the dataset into HDFS or somewhere in the file system and then pull the RDDs within the other jobs. However, the same functionality can be achieved by using Alluxio (formerly Tachyon) or Apache Ignite.
Apache Ignite's memory-centric architecture enables RDD sharing in a very efficient and effective way. Apache Ignite provides IgniteContext and IgniteRDD to share RDDs between Spark applications.
  1. IgniteContext: IgniteContext is the main entry point to the Spark-Ignite integration. To create an instance of an Ignite context, a user must provide an instance of SparkContext and a closure creating IgniteConfiguration (configuration factory). Ignite context will make sure that server or client Ignite nodes exist in all involved job instances. Alternatively, a path to an XML configuration file can be passed to IgniteContext constructor, which will be used to nodes being started.
  2. IgniteRDD: IgniteRDD is an implementation of Spark RDD abstraction representing a live view of Ignite cache. IgniteRDD is not immutable; all changes in the Ignite cache (regardless of whether they were caused by another RDD or by external changes in cache) will be visible to RDD users immediately. IgniteRDD utilizes the partitioned nature of Ignite caches and provides partitioning information to Spark executor. A number of partitions in IgniteRDD equals the number of partitions in the underlying Ignite cache. IgniteRDD also provides affinity information to Spark via getPrefferredLocations so that RDD computations use data locality.
In the next part of this series, we are going to install Apache Spark and do the following:
  1. Run the wordcount example to verify the Spark installation.
  2. Configure Apache Ignite to share RDDs between Spark applications.
  3. Run Spark applications through Spark Shell to use Ignite RDD.
  4. Develop a Scala Spark application to put some Ignite RDD into the Ignite cluster and pull them from another Scala Spark application.

Comments

Popular posts from this blog

Send e-mail with attachment through OSB

Oracle Service Bus (OSB) contains a good collection of adapter to integrate with any legacy application, including ftp, email, MQ, tuxedo. However e-mail still recognize as a stable protocol to integrate with any application asynchronously. Send e-mail with attachment is a common task of any business process. Inbound e-mail adapter which, integrated with OSB support attachment but outbound adapter doesn't. This post is all about sending attachment though JavaCallout action. There are two ways to handle attachment in OSB: 1) Use JavaCallout action to pass the binary data for further manipulation. It means write down a small java library which will get the attachment and send the e-mail. 2) Use integrated outbound e-mail adapter to send attachment, here you have to add a custom variable named attachment and assign the binary data to the body of the attachment variable. First option is very common and easy to implement through javax.mail api, however a much more developer manage t

Tip: SQL client for Apache Ignite cache

A new SQL client configuration described in  The Apache Ignite book . If it got you interested, check out the rest of the book for more helpful information. Apache Ignite provides SQL queries execution on the caches, SQL syntax is an ANSI-99 compliant. Therefore, you can execute SQL queries against any caches from any SQL client which supports JDBC thin client. This section is for those, who feels comfortable with SQL rather than execute a bunch of code to retrieve data from the cache. Apache Ignite out of the box shipped with JDBC driver that allows you to connect to Ignite caches and retrieve distributed data from the cache using standard SQL queries. Rest of the section of this chapter will describe how to connect SQL IDE (Integrated Development Environment) to Ignite cache and executes some SQL queries to play with the data. SQL IDE or SQL editor can simplify the development process and allow you to get productive much quicker. Most database vendors have their own front-en

Load balancing and fail over with scheduler

Every programmer at least develop one Scheduler or Job in their life time of programming. Nowadays writing or developing scheduler to get you job done is very simple, but when you are thinking about high availability or load balancing your scheduler or job it getting some tricky. Even more when you have a few instance of your scheduler but only one can be run at a time also need some tricks to done. A long time ago i used some data base table lock to achieved such a functionality as leader election. Around 2010 when Zookeeper comes into play, i always preferred to use Zookeeper to bring high availability and scalability. For using Zookeeper you have to need Zookeeper cluster with minimum 3 nodes and maintain the cluster. Our new customer denied to use such a open source product in their environment and i was definitely need to find something alternative. Definitely Quartz was the next choose. Quartz makes developing scheduler easy and simple. Quartz clustering feature brings the HA and