We present Apache Flink 2.0, an evolution of the popular stream processing system's architecture that decouples computation from state management. Flink 2.0 relies on a remote distributed file system (DFS) for primary state storage and uses local disks as a secondary cache, with state updates streamed continuously and directly to the DFS. To address the latency implications of remote storage, Flink 2.0 incorporates an asynchronous runtime execution model. Furthermore, Flink 2.0 introduces ForSt, a novel state store featuring a unified file system that enables faster and lightweight checkpointing, recovery, and reconfiguration with minimal intrusion to the existing Flink runtime architecture. Using a comprehensive set of Nexmark benchmarks and a large-scale stateful production workload, we evaluate Flink 2.0's large-state processing, checkpointing, and recovery mechanisms. Our results show significant performance improvements and reduced resource utilization compared to the baseline Flink 1.20 implementation. Specifically, we observe up to 94% reduction in checkpoint duration, up to 49× faster recovery after failures or a rescaling operation, and up to 50% cost savings.

There is an evident need to complement embedded critical control logic with AI inference, but today's AI-capable hardware, software, and processes are primarily targeted towards the needs of cloud-centric actors. Telecom and defense airspace industries, which make heavy use of specialized hardware, face the challenge of manually hand-tuning AI workloads and hardware, presenting an unprecedented cost and complexity due to the diversity and sheer number of deployed instances. Furthermore, embedded AI functionality must not adversely affect real-time and safety requirements of the critical business logic. To address this, end-to-end AI pipelines for critical platforms are needed to automate the adaption of networks to fit into resource-constrained devices under critical and real-time constraints, while remaining interoperable with de-facto standard AI tools and frameworks used in the cloud. We present two industrial applications where such solutions are needed to bring AI to critical and resource-constrained hardware, and a generalized end-to-end AI pipeline that addresses these needs. Crucial steps to realize it are taken in the industry-academia collaborative FASTER-AI project.

Failure transparency enables users to reason about distributed systems at a higher level of abstraction, where complex failure-handling logic is hidden. This is especially true for stateful dataflow systems, which are the backbone of many cloud applications. In particular, this paper focuses on proving failure transparency in Apache Flink, a popular stateful dataflow system. Even though failure transparency is a critical aspect of Apache Flink, to date it has not been formally proven. Showing that the failure transparency mechanism is correct, however, is challenging due to the complexity of the mechanism itself. Nevertheless, this complexity can be effectively hidden behind a failure transparent programming interface. To show that Apache Flink is failure transparent, we model it in small-step operational semantics. Next, we provide a novel definition of failure transparency based on observational explainability, a concept which relates executions according to their observations. Finally, we provide a formal proof of failure transparency for the implementation model; i.e., we prove that the failure-free model correctly abstracts from the failure-related details of the implementation model. We also show liveness of the implementation model under a fair execution assumption. These results are a first step towards a verified stack for stateful dataflow systems.

Stream processing in the last decade has seen broad adoption in both commercial and research settings. One key element for this success is the ability of modern stream processors to handle failures while ensuring exactly-once processing guarantees. At the moment of writing, virtually all stream processors that guarantee exactly-once processing implement a variant of Apache Flink's coordinated checkpoints - an extension of the original Chandy-Lamport checkpoints from 1985. However, the reasons behind this prevalence of the coordinated approach remain anecdotal, as reported by practitioners of the stream processing community. At the same time, common checkpointing approaches, such as the uncoordinated and the communication-induced ones, remain largely unexplored. This paper is the first to address this gap by i) shedding light on why practitioners have favored the coordinated approach and ii) investigating whether there are viable alternatives. To this end, we implement three checkpointing approaches that we surveyed and adapted for the distinct needs of streaming dataflows. Our analysis shows that the coordinated approach outperforms the uncoordinated and communication-induced protocols under uniformly distributed workloads. To our surprise, however, the uncoordinated approach is not only competitive to the coordinated one in uniformly distributed workloads, but it also outperforms the coordinated approach in skewed workloads. We conclude that rather than blindly employing coordinated checkpointing, research should focus on optimizing the very promising uncoordinated approach, as it can address issues with skew and support prevalent cyclic queries. We believe that our findings can trigger further research into checkpointing mechanisms.

Abstract Stream processing has been an active research field for more than 20 years, but it is now witnessing its prime time due to recent successful efforts by the research community and numerous worldwide open-source communities. This survey provides a comprehensive overview of fundamental aspects of stream processing systems and their evolution in the functional areas of out-of-order data management, state management, fault tolerance, high availability, load management, elasticity, and reconfiguration. We review noteworthy past research findings, outline the similarities and differences between the first (’00–’10) and second (’11–’23) generation of stream processing systems, and discuss future trends and open problems.

Omni-Paxos is a system for state machine replication that is completely resilient to partial network partitions, a major source of service disruptions in recent years. Omni-Paxos achieves its resilience through a decoupled design that separates the execution and state of leader election from log replication. The leader election builds on the concept of quorum-connected servers, with the sole focus on connectivity. Additionally, by decoupling reconfiguration from log replication, Omni-Paxos provides flexible and parallel log migration that improves the performance and robustness of reconfiguration. Our evaluation showcases two benefits over state-of-the-art protocols: (1) guaranteed recovery in at most four election timeouts under extreme partial network partitions, and (2) up to 8x shorter reconfiguration periods with 46% less I/O at the leader.

Serverless applications spanning the cloud and edge require flexible programming frameworks for expressing compositions across the different levels of deployment. Another critical aspect for applications with state is failure resilience beyond the scope of a single dataflow graph that is the current standard in data streaming systems. This paper presents Portals, an interactive, stateful dataflow composition framework with strong end-to-end guarantees. Portals enables event-driven, resilient applications that span across dataflow graphs and serverless deployments. The demonstration exhibits three scenarios in our multi-dataflow streaming-based system: dynamically composing a stateful serverless application; an interactive cloud and edge serverless application; and a Portals browser playground.

Portals is a serverless, distributed programming model that blends the exactly-once processing guarantees of stateful dataflow streaming frameworks with the message-driven compositionality of actor frameworks. Decentralized applications in Portals can be built dynamically, scale on demand, and always satisfy strict atomic processing guarantees that are natively embedded in the framework's principal elements of computation, known as atomic streams. In this paper, we describe the capabilities of Portals and demonstrate its use in supporting several popular existing distributed programming paradigms and use-cases. We further introduce all programming model invariants and the corresponding system methods used to satisfy them.

Stream processing has been an active research field for more than 20 years, but it is now witnessing its prime time due to recent successful efforts by the research community and numerous worldwide open-source communities. The goal of this tutorial is threefold. First, we aim to review and highlight noteworthy past research findings, which were largely ignored until very recently. Second, we intend to underline the differences between early ('00-'10) and modern ('11-'18) streaming systems, and how those systems have evolved through the years. Most importantly, we wish to turn the attention of the database community to recent trends: streaming systems are no longer used only for classic stream processing workloads, namely window aggregates and joins. Instead, modern streaming systems are being increasingly used to deploy general event-driven applications in a scalable fashion, challenging the design decisions, architecture and intended use of existing stream processing systems.

Graph partitioning is an essential yet challenging task for massive graph analysis in distributed computing. Common graph partitioning methods scan the complete graph to obtain structural characteristics offline, before partitioning. However, the emerging need for low-latency, continuous graph analysis led to the development of online partitioning methods. Online methods ingest edges or vertices as a stream, making partitioning decisions on the fly based on partial knowledge of the graph. Prior studies have compared offline graph partitioning techniques across different systems. Yet, little effort has been put into investigating the characteristics of online graph partitioning strategies. In this work, we describe and categorize online graph partitioning techniques based on their assumptions, objectives and costs. Furthermore, we employ an experimental comparison across different applications and datasets, using a unified distributed runtime based on Apache Flink. Our experimental results showcase that model-dependent online partitioning techniques such as low-cut algorithms offer better performance for communication-intensive applications such as bulk synchronous iterative algorithms, albeit higher partitioning costs. Otherwise, model-agnostic techniques trade off data locality for lower partitioning costs and balanced workloads which is beneficial when executing data-parallel single-pass graph algorithms.

Stream processors are emerging in industry as an apparatus that drives analytical but also mission critical services handling the core of persistent application logic. Thus, apart from scalability and low-latency, a rising system need is first-class support for application state together with strong consistency guarantees, and adaptivity to cluster reconfigurations, software patches and partial failures. Although prior systems research has addressed some of these specific problems, the practical challenge lies on how such guarantees can be materialized in a transparent, non-intrusive manner that relieves the user from unnecessary constraints. Such needs served as the main design principles of state management in Apache Flink, an open source, scalable stream processor. We present Flink's core pipelined, in-flight mechanism which guarantees the creation of lightweight, consistent, distributed snapshots of application state, progressively, without impacting continuous execution. Consistent snapshots cover all needs for system reconfiguration, fault tolerance and version management through coarse grained rollback recovery. Application state is declared explicitly to the system, allowing efficient partitioning and transparent commits to persistent storage. We further present Flink's backend implementations and mechanisms for high availability, external state queries and output commit. Finally, we demonstrate how these mechanisms behave in practice with metrics and large-deployment insights exhibiting the low performance trade-offs of our approach and the general benefits of exploiting asynchrony in continuous, yet sustainable system deployments.

Aggregation queries on data streams are evaluated over evolving and often overlapping logical views called windows. While the aggregation of periodic windows were extensively studied in the past through the use of aggregate sharing techniques such as Panes and Pairs, little to no work has been put in optimizing the aggregation of very common, non-periodic windows. Typical examples of non-periodic windows are punctuations and sessions which can implement complex business logic and are often expressed as user-defined operators on platforms such as Google Dataflow or Apache Storm. The aggregation of such non-periodic or user-defined windows either falls back to expensive, best-effort aggregate sharing methods, or is not optimized at all.