Integrating Apache Spark and NiFi for Data Lakes
Download as PPTX, PDF25 likes11,037 views
This document discusses using Apache Spark and Apache NiFi together for data lakes. It outlines the goals of a data lake including having a central data repository, reducing costs, enabling easier discovery and prototyping. It also discusses what is needed for a Hadoop data lake, including automation of pipelines, governance, and interactive data discovery. The document then provides an example ingestion project and describes using Apache Spark for functions like cleansing, validating, and profiling data. It outlines using Apache NiFi for the pipeline design with drag and drop functionality. Finally, it demonstrates ingesting and preparing data, data self-service and transformation, data discovery, and operational monitoring capabilities.
Integrating Apache Spark and NiFi for Data Lakes
- 1. MAKING BIG DATA COME ALIVE Integrating Apache Spark And NiFi For Data Lakes Ron Bodkin Founder & President Scott Reisdorf R&D Architect
- 3. 3 • A central repository with trusted, consistent data • Reduce costs by offloading analytical systems and archiving cold data • Derive value quickly with easier discovery and prototyping • A laboratory for experimenting with new technologies and data Goals for a Data Lake
- 4. 4 • Automation of pipelines with metadata and performance tracking • Governance with clear distinction of roles and responsibilities • SLA tracking with alerts on failures or violations • Interactive data discovery and experimentation What’s Needed For A Hadoop Data Lake?
- 5. 5 Example Ingestion Project • 4000+ unique flat files and RDMS tables, plus a few streaming data feeds • Mix of incremental and snapshot data • Ingest into Hadoop (minimally HDFS and Hive tables) • Cleansing/encryption and data validation • Metadata capture Focus shifts over time from data ingestion to transformation then to analytics
- 6. 6 Design
- 7. 7 Apache Spark Functions • Cleanse • Validate • Profile • Wrangle
- 8. 8 Pipeline design with Apache • Visual drag-and-drop • Dozens of data connectors • 150+ pre-built transforms • Data lineage • Batch and Streaming • Extensible © 2016 Think Big, a Teradata Company 7/10/2016
- 9. 9 Role separation • IT Designers design models in NiFi • Register with framework • Integrated development process © 2016 Think Big, a Teradata Company 7/10/2016 Apache NiFi Think Big framework • Users configure new feeds • Based on common model • Generated and executed in NiFi register deploy
- 10. 1010 7/10/2016 © 2015 Think Big, a Teradata Company User features around org. roles Visual design Streaming and Batch Fully governed Integrated Best Practices Secure, modern architecture Design Approach Will be open source (Apache license)
- 11. 1111 Ingest and Prepare • UI-guided feed creation • Data protection • Data cleanse • Data validation • Data profiling • Powered by Apache Spark
- 12. Unpack and/or merge small files Put file HDFS Cleanse/Stand ardize Spark Data Profile Spark Metadata Validate Spark Data Ingest Model Metadata determines behavior of individual components Adds many Hadoop- specific higher-level NiFi processors Index Text Elasticsearch Merge / Dedupe Hive Compress & Archive Originals HDFS,S3 Extract Table JDBC Get File(s) Filesystem Message JMS/Kafka Other HTTP/REST, etc. Data policies 12
- 13. 1313 Data self-service and “wrangle” • Graphical SQL builder • 100+ transform functions • Machine learning • Publish and schedule • Powered by Apache Spark
- 14. 1414 Data Discovery • Google-like searching • Extensible metadata • Data profile • Data sampling
- 15. 1515 Operations • Dashboard • Health Monitoring • Data Confidence • SLA enforcement • Alerts • Performance reports
- 16. 16 • Powerful search capabilities for users against data (think Google-like searching) • NiFi processor extracts source data from Hadoop table for indexing in ElasticSearch • Incremental updates during ingest ElasticSearch – Full Text Indexing Data Lake select id,user,tweet from twitter_feed extract JSON
- 17. 17 Demo
- 18. 1818
Editor's Notes
- #13: Notice that we delegate processing to the Spark and Hadoop cluster for much of our work