Incremental ETL Pipeline Scheduling for Near Real-Time Data Warehouses

We present our work based on an incremental ETL pipeline for on-demand data warehouse maintenance. Pipeline parallelism is exploited to concurrently execute a chain of maintenance jobs, each of which takes a batch of delta tuples extracted from source-local transactions with commit timestamps preceding the arrival time of an incoming warehouse query and calculates Ąnal deltas to bring relevant warehouse tables up-to-date. Each pipeline operator runs in a single, non-terminating thread to process one job at a time and re-initializes itself for a new one. However, to continuously perform incremental joins or maintain slowly changing dimension tables (SCD), the same staging tables or dimension tables can be concurrently accessed and updated by distinct pipeline operators which work on diferent jobs. Inconsistencies can arise without proper thread coordinations. In this paper, we proposed two types of consistency zones for SCD and incremental join to address this problem. Besides, we reviewed existing pipeline scheduling algorithms in our incremental ETL pipeline with consistency zones.