Apache Apex: next gen Big Data analytics

Below is a compilation of my notes taken during the presentation of Apache Apex by Thomas Weise from DataTorrent, the company behind Apex.

Introduction

Apache Apex is an in-memory distributed parallel stream processing engine, like Flink or Storm. However, it is built with native Hadoop integration in mind:

• Yarn is used for resource managing and ordonnancing
• HDFS is used to store persistant states

Application development model

• A stream is a sequence of tuples

• An operator:

  • takes one or more input streams as input
  • performs custom computation on the tuples (logic is in Java)
  • emits one or more input streams
  • has many parallel instances, each single threaded
  • uses the DAG model to optimize computation
• An application is a suite of operators

Development process

A typical WordCount setup with Apex looks like this:

• Apache Kafka brings the data

• The Apex application processes through the following operators:

  • Kafka input
  • Parser
  • Filter
  • Word counter
  • JDBC output
• End data is written in DB

The development process goes like this:

• Take an operator from existing libraries or implement a custom logic
• Connect the operators to form an application
• Configure the operators properties
• Configure scaling & platform attributes
• Test functionalities, performance and iterate

Operator libraries

Apex provides very complete operator libraries through Apache Apex Malhar:

• Messaging (Kafka, ActiveMQ, …)
• NoSQL (HBase, Cassandra, MongoDB, Redis, CouchDB, …)
• RDBMS (JDBC, MySQL, …)
• FileSystem (HDFS / Hive, …)

Notes

• Apex uses Apache BEAM for the job implementation so he enjoys its multiple benfits:

  • dynamic partition at runtime
  • load-balancing between operators
  • windowing
  • …
• For fault tolerance, the operators states are checkpointed and persisted

• Apex processing guarentees for

  • at least once
  • at most once
  • exactly once