LOG.io: Unified Rollback Recovery and Data Lineage Capture for Distributed Data Pipelines

TL;DR

LOG.io offers a unified approach for rollback recovery and data lineage capture in distributed data pipelines, supporting non-deterministic operators and dynamic scaling, with performance comparable or superior to existing methods under certain conditions.

Contribution

It introduces LOG.io, a novel log-based system that enhances recovery and lineage capture in serverless distributed pipelines, accommodating complex operator behaviors and scaling.

Findings

LOG.io performs as well as ABS during normal processing with moderate throughput.

LOG.io outperforms ABS during recovery when stragglers are present.

Data lineage capture overhead is less than 1.5% in all experiments.

Abstract

This paper introduces LOG.io, a comprehensive solution designed for correct rollback recovery and fine-grain data lineage capture in distributed data pipelines. It is tailored for serverless scalable architectures and uses a log-based rollback recovery protocol. LOG.io supports a general programming model, accommodating non-deterministic operators, interactions with external systems, and arbitrary custom code. It is non-blocking, allowing failed operators to recover independently without interrupting other active operators, thereby leveraging data parallelization, and it facilitates dynamic scaling of operators during pipeline execution. Performance evaluations, conducted within the SAP Data Intelligence system, compare LOG.io with the Asynchronous Barrier Snapshotting (ABS) protocol, originally implemented in Flink. Our experiments show that when there are straggler operators in a data pipeline and the throughput of events is moderate (e.g., 1 event every 100 ms), LOG.io performs as well as ABS during normal processing and outperforms ABS during recovery. Otherwise, ABS performs better than LOG.io for both normal processing and recovery. However, we show that in these cases, data parallelization can largely reduce the overhead of LOG.io while ABS does not improve. Finally, we show that the overhead of data lineage capture, at the granularity of the event and between any two operators in a pipeline, is marginal, with less than 1.5% in all our experiments.