Category: apache Flink

Apache Flink – Flink vs Spark vs Hadoop Here is a comprehensive table, which shows the comparison between three most popular big data frameworks: Apache Flink, Apache Spark and Apache Hadoop. Apache Hadoop Apache Spark Apache Flink Year of Origin 2005 2009 2009 Place of Origin MapReduce (Google) Hadoop (Yahoo) University of California, Berkeley Technical University of Berlin Data Processing Engine Batch Batch Stream Processing Speed Slower than Spark and Flink 100x Faster than Hadoop Faster than spark Programming Languages Java, C, C++, Ruby, Groovy, Perl, Python Java, Scala, python and R Java and Scala Programming Model MapReduce Resilient distributed Datasets (RDD) Cyclic dataflows Data Transfer Batch Batch Pipelined and Batch Memory Management Disk Based JVM Managed Active Managed Latency Low Medium Low Throughput Medium High High Optimization Manual Manual Automatic API Low-level High-level High-level Streaming Support NA Spark Streaming Flink Streaming SQL Support Hive, Impala SparkSQL Table API and SQL Graph Support NA GraphX Gelly Machine Learning Support NA SparkML FlinkML

Apache Flink – Useful Resources The following resources contain additional information on Apache Flink. Please use them to get more in-depth knowledge on this. Useful Links on Apache Flink − Wikipedia Reference for Apache Flink − official Site of Apache Flink Useful Books on Apache Flink To enlist your site on this page, please drop an email to [email protected]

Apache Flink – Machine Learning Apache Flink”s Machine Learning library is called FlinkML. Since usage of machine learning has been increasing exponentially over the last 5 years, Flink community decided to add this machine learning APO also in its ecosystem. The list of contributors and algorithms are increasing in FlinkML. This API is not a part of binary distribution yet. Here is an example of linear regression using FlinkML − // LabeledVector is a feature vector with a label (class or real value) val trainingData: DataSet[LabeledVector] = … val testingData: DataSet[Vector] = … // Alternatively, a Splitter is used to break up a DataSet into training and testing data. val dataSet: DataSet[LabeledVector] = … val trainTestData: DataSet[TrainTestDataSet] = Splitter.trainTestSplit(dataSet) val trainingData: DataSet[LabeledVector] = trainTestData.training val testingData: DataSet[Vector] = trainTestData.testing.map(lv => lv.vector) val mlr = MultipleLinearRegression() .setStepsize(1.0) .setIterations(100) .setConvergenceThreshold(0.001) mlr.fit(trainingData) // The fitted model can now be used to make predictions val predictions: DataSet[LabeledVector] = mlr.predict(testingData) Inside flink-1.7.1/examples/batch/ path, you will find KMeans.jar file. Let us run this sample FlinkML example. This example program is run using the default point and the centroid data set. ./bin/flink run examples/batch/KMeans.jar –output Print

Discuss Apache Flink Apache Flink is an open source stream processing framework, which has both batch and stream processing capabilities. Apache Flink is very similar to Apache Spark, but it follows stream-first approach. It is also a part of Big Data tools list. This tutorial explains the basics of Flink Architecture Ecosystem and its APIs.

Apache Flink – Conclusion The comparison table that we saw in the previous chapter concludes the pointers pretty much. Apache Flink is the most suited framework for real-time processing and use cases. Its single engine system is unique which can process both batch and streaming data with different APIs like Dataset and DataStream. It does not mean Hadoop and Spark are out of the game, the selection of the most suited big data framework always depends and vary from use case to use case. There can be several use cases where a combination of Hadoop and Flink or Spark and Flink might be suited. Nevertheless, Flink is the best framework for real time processing currently. The growth of Apache Flink has been amazing and the number of contributors to its community is growing day by day. Happy Flinking!

Apache Flink – Quick Guide Apache Flink – Big Data Platform The advancement of data in the last 10 years has been enormous; this gave rise to a term ”Big Data”. There is no fixed size of data, which you can call as big data; any data that your traditional system (RDBMS) is not able to handle is Big Data. This Big Data can be in structured, semi-structured or un-structured format. Initially, there were three dimensions to data − Volume, Velocity, Variety. The dimensions have now gone beyond just the three Vs. We have now added other Vs − Veracity, Validity, Vulnerability, Value, Variability, etc. Big Data led to the emergence of multiple tools and frameworks that help in the storage and processing of data. There are a few popular big data frameworks such as Hadoop, Spark, Hive, Pig, Storm and Zookeeper. It also gave opportunity to create Next Gen products in multiple domains like Healthcare, Finance, Retail, E-Commerce and more. Whether it is an MNC or a start-up, everyone is leveraging Big Data to store and process it and take smarter decisions. Apache Flink – Batch vs Real-time Processing In terms of Big Data, there are two types of processing − Batch Processing Real-time Processing Processing based on the data collected over time is called Batch Processing. For example, a bank manager wants to process past one-month data (collected over time) to know the number of cheques that got cancelled in the past 1 month. Processing based on immediate data for instant result is called Real-time Processing. For example, a bank manager getting a fraud alert immediately after a fraud transaction (instant result) has occurred. The table given below lists down the differences between Batch and Real-Time Processing − Batch Processing Real-Time Processing Static Files Event Streams Processed Periodically in minute, hour, day etc. Processed immediately nanoseconds Past data on disk storage In Memory Storage Example − Bill Generation Example − ATM Transaction Alert These days, real-time processing is being used a lot in every organization. Use cases like fraud detection, real-time alerts in healthcare and network attack alert require real-time processing of instant data; a delay of even few milliseconds can have a huge impact. An ideal tool for such real time use cases would be the one, which can input data as stream and not batch. Apache Flink is that real-time processing tool. Apache Flink – Introduction Apache Flink is a real-time processing framework which can process streaming data. It is an open source stream processing framework for high-performance, scalable, and accurate real-time applications. It has true streaming model and does not take input data as batch or micro-batches. Apache Flink was founded by Data Artisans company and is now developed under Apache License by Apache Flink Community. This community has over 479 contributors and 15500 + commits so far. Ecosystem on Apache Flink The diagram given below shows the different layers of Apache Flink Ecosystem − Storage Apache Flink has multiple options from where it can Read/Write data. Below is a basic storage list − HDFS (Hadoop Distributed File System) Local File System S3 RDBMS (MySQL, Oracle, MS SQL etc.) MongoDB HBase Apache Kafka Apache Flume Deploy You can deploy Apache Fink in local mode, cluster mode or on cloud. Cluster mode can be standalone, YARN, MESOS. On cloud, Flink can be deployed on AWS or GCP. Kernel This is the runtime layer, which provides distributed processing, fault tolerance, reliability, native iterative processing capability and more. APIs & Libraries This is the top layer and most important layer of Apache Flink. It has Dataset API, which takes care of batch processing, and Datastream API, which takes care of stream processing. There are other libraries like Flink ML (for machine learning), Gelly (for graph processing ), Tables for SQL. This layer provides diverse capabilities to Apache Flink. Apache Flink – Architecture Apache Flink works on Kappa architecture. Kappa architecture has a single processor – stream, which treats all input as stream and the streaming engine processes the data in real-time. Batch data in kappa architecture is a special case of streaming. The following diagram shows the Apache Flink Architecture. The key idea in Kappa architecture is to handle both batch and real-time data through a single stream processing engine. Most big data framework works on Lambda architecture, which has separate processors for batch and streaming data. In Lambda architecture, you have separate codebases for batch and stream views. For querying and getting the result, the codebases need to be merged. Not maintaining separate codebases/views and merging them is a pain, but Kappa architecture solves this issue as it has only one view − real-time, hence merging of codebase is not required. That does not mean Kappa architecture replaces Lambda architecture, it completely depends on the use-case and the application that decides which architecture would be preferable. The following diagram shows Apache Flink job execution architecture. Program It is a piece of code, which you run on the Flink Cluster. Client It is responsible for taking code (program) and constructing job dataflow graph, then passing it to JobManager. It also retrieves the Job results. JobManager After receiving the Job Dataflow Graph from Client, it is responsible for creating the execution graph. It assigns the job to TaskManagers in the cluster and supervises the execution of the job. TaskManager It is responsible for executing all the tasks that have been assigned by JobManager. All the TaskManagers run the tasks in their separate slots in specified parallelism. It is responsible to send the status of the tasks to JobManager. Features of Apache Flink The features of Apache Flink are as follows − It has a streaming processor, which can run both batch and stream programs. It can process data at lightning fast speed. APIs available in Java, Scala and Python. Provides APIs for all the common operations, which is very easy for programmers to use. Processes data in low latency (nanoseconds) and high throughput. Its fault tolerant. If a node,

Apache Flink – Libraries In this chapter, we will learn about the different libraries of Apache Flink. Complex Event Processing (CEP) FlinkCEP is an API in Apache Flink, which analyses event patterns on continuous streaming data. These events are near real time, which have high throughput and low latency. This API is used mostly on Sensor data, which come in real-time and are very complex to process. CEP analyses the pattern of the input stream and gives the result very soon. It has the ability to provide real-time notifications and alerts in case the event pattern is complex. FlinkCEP can connect to different kind of input sources and analyse patterns in them. This how a sample architecture with CEP looks like − Sensor data will be coming in from different sources, Kafka will act as a distributed messaging framework, which will distribute the streams to Apache Flink, and FlinkCEP will analyse the complex event patterns. You can write programs in Apache Flink for complex event processing using Pattern API. It allows you to decide the event patterns to detect from the continuous stream data. Below are some of the most commonly used CEP patterns − Begin It is used to define the starting state. The following program shows how it is defined in a Flink program − Pattern<Event, ?> next = start.next(“next”); Where It is used to define a filter condition in the current state. patternState.where(new FilterFunction <Event>() { @Override public boolean filter(Event value) throws Exception { } }); Next It is used to append a new pattern state and the matching event needed to pass the previous pattern. Pattern<Event, ?> next = start.next(“next”); FollowedBy It is used to append a new pattern state but here other events can occur b/w two matching events. Pattern<Event, ?> followedBy = start.followedBy(“next”); Gelly Apache Flink”s Graph API is Gelly. Gelly is used to perform graph analysis on Flink applications using a set of methods and utilities. You can analyse huge graphs using Apache Flink API in a distributed fashion with Gelly. There are other graph libraries also like Apache Giraph for the same purpose, but since Gelly is used on top of Apache Flink, it uses single API. This is very helpful from development and operation point of view. Let us run an example using Apache Flink API − Gelly. Firstly, you need to copy 2 Gelly jar files from opt directory of Apache Flink to its lib directory. Then run flink-gelly-examples jar. cp opt/flink-gelly* lib/ ./bin/flink run examples/gelly/flink-gelly-examples_*.jar Let us now run the PageRank example. PageRank computes a per-vertex score, which is the sum of PageRank scores transmitted over in-edges. Each vertex”s score is divided evenly among out-edges. High-scoring vertices are linked to by other high-scoring vertices. The result contains the vertex ID and the PageRank score. usage: flink run examples/flink-gelly-examples_<version>.jar –algorithm PageRank [algorithm options] –input <input> [input options] –output <output> [output options] ./bin/flink run examples/gelly/flink-gelly-examples_*.jar –algorithm PageRank –input CycleGraph –vertex_count 2 –output Print

Apache Flink – Setup/Installation Before the start with the setup/ installation of Apache Flink, let us check whether we have Java 8 installed in our system. Java – version We will now proceed by downloading Apache Flink. wget http://mirrors.estointernet.in/apache/flink/flink-1.7.1/flink-1.7.1-bin-scala_2.11.tgz Now, uncompress the tar file. tar -xzf flink-1.7.1-bin-scala_2.11.tgz Go to Flink”s home directory. cd flink-1.7.1/ Start the Flink Cluster. ./bin/start-cluster.sh Open the Mozilla browser and go to the below URL, it will open the Flink Web Dashboard. http://localhost:8081 This is how the User Interface of Apache Flink Dashboard looks like. Now the Flink cluster is up and running.

Apache Flink – Architecture Apache Flink works on Kappa architecture. Kappa architecture has a single processor – stream, which treats all input as stream and the streaming engine processes the data in real-time. Batch data in kappa architecture is a special case of streaming. The following diagram shows the Apache Flink Architecture. The key idea in Kappa architecture is to handle both batch and real-time data through a single stream processing engine. Most big data framework works on Lambda architecture, which has separate processors for batch and streaming data. In Lambda architecture, you have separate codebases for batch and stream views. For querying and getting the result, the codebases need to be merged. Not maintaining separate codebases/views and merging them is a pain, but Kappa architecture solves this issue as it has only one view − real-time, hence merging of codebase is not required. That does not mean Kappa architecture replaces Lambda architecture, it completely depends on the use-case and the application that decides which architecture would be preferable. The following diagram shows Apache Flink job execution architecture. Program It is a piece of code, which you run on the Flink Cluster. Client It is responsible for taking code (program) and constructing job dataflow graph, then passing it to JobManager. It also retrieves the Job results. JobManager After receiving the Job Dataflow Graph from Client, it is responsible for creating the execution graph. It assigns the job to TaskManagers in the cluster and supervises the execution of the job. TaskManager It is responsible for executing all the tasks that have been assigned by JobManager. All the TaskManagers run the tasks in their separate slots in specified parallelism. It is responsible to send the status of the tasks to JobManager. Features of Apache Flink The features of Apache Flink are as follows − It has a streaming processor, which can run both batch and stream programs. It can process data at lightning fast speed. APIs available in Java, Scala and Python. Provides APIs for all the common operations, which is very easy for programmers to use. Processes data in low latency (nanoseconds) and high throughput. Its fault tolerant. If a node, application or a hardware fails, it does not affect the cluster. Can easily integrate with Apache Hadoop, Apache MapReduce, Apache Spark, HBase and other big data tools. In-memory management can be customized for better computation. It is highly scalable and can scale upto thousands of node in a cluster. Windowing is very flexible in Apache Flink. Provides Graph Processing, Machine Learning, Complex Event Processing libraries.

Apache Flink – Table API and SQL Table API is a relational API with SQL like expression language. This API can do both batch and stream processing. It can be embedded with Java and Scala Dataset and Datastream APIs. You can create tables from existing Datasets and Datastreams or from external data sources. Through this relational API, you can perform operations like join, aggregate, select and filter. Whether the input is batch or stream, the semantics of the query remains the same. Here is a sample Table API program − // for batch programs use ExecutionEnvironment instead of StreamExecutionEnvironment val env = StreamExecutionEnvironment.getExecutionEnvironment // create a TableEnvironment val tableEnv = TableEnvironment.getTableEnvironment(env) // register a Table tableEnv.registerTable(“table1”, …) // or tableEnv.registerTableSource(“table2”, …) // or tableEnv.registerExternalCatalog(“extCat”, …) // register an output Table tableEnv.registerTableSink(“outputTable”, …); // create a Table from a Table API query val tapiResult = tableEnv.scan(“table1”).select(…) // Create a Table from a SQL query val sqlResult = tableEnv.sqlQuery(“SELECT … FROM table2 …”) // emit a Table API result Table to a TableSink, same for SQL result tapiResult.insertInto(“outputTable”) // execute env.execute()