Apache Spark vs Apache Flink
- 1. Apache Spark vs Apache Flink Two most contemporary general purpose data processing platform. AKASH SIHAGPh. No. : +91-7737111579 asihag70@gmail.com akash.sihag@infoobjects.com
- 2. Introduction Apache Spark is a fast and general engine for large-scale data processing. Apache Flink Apache Flink is an open source platform for distributed stream and batch data processing. Apache Spark
- 3. The Inception ● 2009 : At UC Berkeley's AMPLab by Matei Zaharia ● 2010 : Open-sourced under BSD license. ● 2013 : Donated to Apache Software Foundation and switched its license to Apache 2.0 ● 2014 : Top Level Apache Project Apache Flink ● 2010 : Started as a collaboration of of Technical University Berlin, Humboldt-Universität zu Berlin, and Hasso-Plattner-Institut Potsdam. ● 2014 : Apache Incubator. ● 2014(Dec) : Apache Top Level Project. Apache Spark
- 4. Overview Spark Apache link Components: ● Spark SQL : For SQL and unstructured data processing. ● Spark Streaming: For processing live streaming data. ● MLib : Machine Learning Algorithm. ● GraphX : Graph based processing. ● Spark Core : Its the base processing engine in Spark that works on the concept of RDD and all API’s resides on top of it. Deploy: ● Standalone : Included with Spark. ● Apache Mesos ● Apache YARN : Hadoop 2 resource manager. Spark Core
- 5. Overview Flink Apache link Components: ● DataStream API : For unbounded streams. ● DataSet API : Batch processing. ● Table API : For SQL like operations. ● CEP : Complex Event processing API. ● M L Library : For machine Learning algorithms. ● Gelly : Graph processing API. Deploy: ● Standalone : Included with Flink. ● Local : Single JVM ● Apache YARN : Hadoop 2 resource manager.
- 6. Deep Dive
- 7. Computing Paradigm ● Work on the abstraction of RDD i.e. Resilient distributed datasets. ● Supports in-memory computation. ● Lazy Evaluation (Transformation- action). ● DAG is generated for every Spark Job. ● Streams are processed as chunks of batches. Apache Flink ● Works on the abstraction of Cyclic Data Flows. ● Supports in-memory computation. ● Lazy Evaluation (Iterative- Transformation). ● Job Graph are generated. ● Batches are processed as streams. Apache Spark
- 8. Similarities Apache link● Both are data processing platforms. ● Similar kind of collection APIs. ● Leverages frameworks like AKKA, YARN. ● Since APIs are similar, code porting takes less efforts. ● Both provides stream and batch processing. ● Fault-Tolerant. ● APIs in JAVA and Scala.
- 9. Apache Spark Apache Flink ● Near real time stream processing. ● Batch and streaming transformations are possible. ● Limited window based operations. ● Catalyst Optimizer for SQL operations. ● Stateful Operation till v1.5 are not so efficient. Note: In Spark 1.6 stateful operations are drastically improved. ● Structured data source support is matured. Ex: HiveContext can be created directly via Spark SQL. ● More committer and third party APIs. ● Spark uses JAVA Heap memory allocation for cached data. Note: From Spark 1.5 spark started implementing off-heap memory allocation (Tungsten). ● ML algos are implemented via DAG ● Real time stream processing. ● Batch with streams operations are not possible and so operating on historic data with live streaming is not so great. ● Various flavours of window based operations based on triggers, record counts and events. ● Optimizer for streams as well as batches. ● Efficient stateful stream operations. ● Structured data support is not so matured and still only have Hadoop InputFormat API. ● Relatively new ecosystem. ● Flink implemented custom memory allocation from its inception. ● ML algos are implemented in native style. VS
- 10. Conclusion: Past ● Spark came first as a unified platform and lead the Big Data world. ● Flink took some time to come into existence. Present ● Spark due to its lead is now more mature and has a big community and API support. ● Flink improved the unified platform idea and is also capable of solving Spark’s limitations to some extent. Claims itself to be faster in stream as well as batch processing. Future ● As Spark has a very fast development cycle, it is supposed to improve itself over time. ● Flink proved itself better than Spark as far as abstraction is concerned but is still a newbie.
- 11. THANK YOU