Lambda-Architecture-BigData-Pipeline-Curation

Architecture Overview

1. Ingest Layer

• Set up the Kafka cluster with monitoring tools in the form of Docker compose.
• Collect the house transation data as an event and send to Kafka topic.

2. Speed Layer

• Kafka producer & consumer.
• Streaming ETL that will sink to Big Query.
• Materialize the KTable of count based on the tumbling timewindow.

3. Batch Layer

• Train the Regression Model by machine learning API provided by Spark MLlib and persist into GCS.
• Use metrics to evaluate the performance of each Model.

4. Serving Layer

• Load the ML model inside SpringBoot application to provide prediction function.
• Provide interactive query based on Kafka Timewindow Streaming.

Limitation and Future Enhancement

• Integrate the Neural Network to solve regression problem by using deeplearning4j.
• Use Flink & Alink to achieve the unified online & offline machine learning, since the Spark MLlib only provide the streaming linear regression model for the regression problem.
• Embed the pre-processing and feature engineering into pipeline and make them automatically.