One of the most important issues in distributed data stream processing systems is using operator migration to handle highly variable workloads cost-efficiently and adapt to the needs at any given time on demand. Operator migration is a complex process involving changes in the state and stream management of a running query, typically without any data loss, and with as little disruption to the execution as possible. This tutorial aims to introduce operator migration, explain the core elements of operator migration, and provide the reader with a good understanding of the design alternatives used in existing solutions. We developed a conceptual model to explain the fundamentals of operator migration and introduce a unified terminology, leading to a taxonomy of existing solutions. The conceptual model separates mechanisms, i.e., how to migrate, and policy, i.e., when to migrate. This separation is further applied to structure the description of existing solutions, offering the reader an algorithmic perspective on various design alternatives. To enhance our understanding of the impact of various design alternatives on migration mechanisms, we also conducted an empirical study that provides quantitative insights. The operator downtime for the naïve migration approach is almost 20 times longer than when applying an incremental checkpoint-based approach.

中文翻译：

---

迁移还是不迁移：分布式流处理算子迁移分析

分布式数据流处理系统中最重要的问题之一是使用操作员迁移来经济高效地处理高度可变的工作负载，并按需适应任何给定时间的需求。操作员迁移是一个复杂的过程，涉及正在运行的查询的状态和流管理的变化，通常不会丢失任何数据，并且对执行的干扰尽可能小。本教程旨在介绍算子迁移，解释算子迁移的核心要素，并让读者更好地理解现有解决方案中使用的设计替代方案。我们开发了一个概念模型来解释运营商迁移的基本原理并引入统一的术语，从而对现有解决方案进行分类。概念模型将机制（即如何迁移）和策略（即何时迁移）分开。这种分离进一步应用于构建现有解决方案的描述，为读者提供各种设计方案的算法视角。为了加深我们对各种设计方案对迁移机制影响的理解，我们还进行了一项提供定量见解的实证研究。与应用基于增量检查点的方法相比，简单迁移方法的操作员停机时间几乎长 20 倍。