METL: a modern ETL pipeline with a dynamic mapping matrix

by   Christian Haase, et al.

Modern ETL streaming pipelines extract data from various sources and forward it to multiple consumers, such as data warehouses (DW) and analytical systems that leverage machine learning (ML). However, the increasing number of systems that are connected to such pipelines requires new solutions for data integration. The canonical (or common) data model (CDM) offers such an integration. It is particular useful for integrating microservice systems into ETL pipelines. (Villaca et al 2020, Oliveira et al 2019) However, a mapping to a CDM is complex. (Lemcke et al 2012) There are three complexity problems, namely the size of the required mapping matrix, the automation of updates of the matrix in response to changes in the extraction sources and the time efficiency of the mapping. In this paper, we present a new solution for these problems. More precisely, we present a new dynamic mapping matrix (DMM), which is based on permutation matrices that are obtained by block-partitioning the full mapping matrix. We show that the DMM can be used for automated updates in response to schema changes, for parallel computation in near real-time and for highly efficient compacting. For the solution, we draw on research into matrix partitioning (Quinn 2004) and dynamic networks (Haase et al 2021). The DMM has been implemented into an app called Message ETL (METL). METL is the key part of a new ETL streaming pipeline at EOS that conducts the transformation to a CDM. The ETL pipeline is based on Kafka-streams. It extracts data from more than 80 microservices with log-based Change Data Capture (CDC) with Debezium and loads the data to a DW and an ML platform. EOS is part of the Otto-Group, the second-largest e-commerce provider in Europe.


page 8

page 12

page 13

page 18


Testing the Robustness of AutoML Systems

Automated machine learning (AutoML) systems aim at finding the best mach...

Developing and Deploying Machine Learning Pipelines against Real-Time Image Streams from the PACS

Executing machine learning (ML) pipelines on radiology images is hard du...

FeedbackLogs: Recording and Incorporating Stakeholder Feedback into Machine Learning Pipelines

Even though machine learning (ML) pipelines affect an increasing array o...

Towards a Change Taxonomy for Machine Learning Systems

Machine Learning (ML) research publications commonly provide open-source...

Dynamic constant time parallel graph algorithms with sub-linear work

The paper proposes dynamic parallel algorithms for connectivity and bipa...

Dynamic Matrix Inverse: Improved Algorithms and Matching Conditional Lower Bounds

The dynamic matrix inverse problem is to maintain the inverse of a matri...

Auto-Validate by-History: Auto-Program Data Quality Constraints to Validate Recurring Data Pipelines

Data pipelines are widely employed in modern enterprises to power a vari...

Please sign up or login with your details

Forgot password? Click here to reset