Towards Latency Guarantee For Distributed Stream Processing
COM1 Level 2
SR6, COM1-02-03
Abstract:
Latency guarantees are critical for real-time stream processing applications, yet remain challenging to achieve under dynamic workloads. This thesis introduces two systems—StreamSwitch and Sluice—that address this challenge at different levels of granularity.
StreamSwitch provides operator-level latency guarantees through a lightweight control plane that unifies dynamic scaling and load balancing. It leverages a novel stream switch abstraction and employs ground-truth latency estimation via network calculus to guide timely scaling decisions through an Examine-Diagnose-Treat (EDT) procedure. Integrated with Apache Flink and Samza, StreamSwitch achieves efficient, low-overhead control in localized bottleneck scenarios.
Sluice targets end-to-end latency guarantees in complex dataflow topologies by decomposing overall latency into intrinsic(controllable) and extrinsic (disruptive) components. It employs a fine-grained Latency Estimation Model (LEM) and an adaptive scaling controller to maintain latency bounds dynamically. Implemented on Apache Flink, Sluice consistently achieves over 99% success in meeting latency targets under highly dynamic workloads, outperforming state-of-the-art solutions in resource efficiency and robustness.
Together, these systems reflect a common design philosophy: precise latency modeling, fine-grained control policy, and system-level optimization.