Data Stream Processing (DSP) applications analyze data flows in near real-time by means of operators, which process and transform incoming data. Operators handle high data rates running parallel replicas across multiple processors and hosts. To guarantee consistent performance without wasting resources in the face of variable workloads, auto-scaling techniques have been studied to adapt operator parallelism at run-time. However, most of the effort has been spent under the assumption of homogeneous computing infrastructures, neglecting the complexity of modern environments.

We consider the problem of deciding both how many operator replicas should be executed and which types of computing nodes should be acquired. We devise heterogeneity-aware policies by means of a two-layered hierarchy of controllers. While application-level components steer the adaptation process for whole applications, aiming to guarantee user-specified requirements, lower-layer components control auto-scaling of single operators. We tackle the fundamental challenge of performance and workload uncertainty, exploiting Bayesian optimization (BO) and reinforcement learning (RL) to devise policies. The evaluation shows that our approach is able to meet users’ requirements in terms of response time and adaptation overhead, while minimizing the cost due to resource usage, outperforming state-of-the-art baselines. We also demonstrate how partial model information is exploited to reduce training time for learning-based controllers.

数据流处理 (DSP) 应用程序通过运算器近乎实时地分析数据流，运算器处理和转换传入的数据。操作员可以处理跨多个处理器和主机运行并行副本的高数据速率。为了在面对可变工作负载时保证一致的性能而不浪费资源，人们研究了自动缩放技术来适应运行时的算子并行性。然而，大部分工作都是在同质计算基础设施的假设下进行的，忽略了现代环境的复杂性。

我们考虑决定应该执行多少个算子副本以及应该获取哪种类型的计算节点的问题。我们通过两层控制器层次结构来设计异构感知策略。应用程序级组件引导整个应用程序的适配过程，旨在保证用户指定的需求，而较低层组件则控制单个算子的自动扩展。我们利用贝叶斯优化 (BO) 和强化学习 (RL) 来制定策略，解决性能和工作负载不确定性的根本挑战。评估表明，我们的方法能够满足用户在响应时间和适应开销方面的要求，同时最大限度地降低资源使用成本，优于最先进的基线。我们还演示了如何利用部分模型信息来减少基于学习的控制器的训练时间。