Traffic accidents pose a significant societal challenge, with many fatalities being avoidable through timely emergency response. We introduce IMPACT (Integrated Multimodal Pipeline for Rapid Accident Causality Tracking), a scalable AI framework designed for autonomous, rapid traffic incident analysis using existing urban CCTV infrastructure. IMPACT combines a low-latency CPU-based vision module for real-time key-frame filtering (24 FPS) with the causal reasoning capabilities of MLLMs, reducing costly MLLM calls by over 92% compared to naive sparse sampling. We further present TRACE10K, a dataset featuring three-tier textual annotations that describe accident dynamics at the frame-sequence level.

Traditionally, query optimizers rely on cost models to choose the best execution plan from several candidates, making precise cost estimates critical for efficient query execution. In recent years, cost models based on machine learning have been proposed to overcome the weaknesses of traditional cost models. While these models have been shown to provide better prediction accuracy, only limited efforts have been made to investigate how well Learned Cost Models (LCMs) actually perform in query optimization and how they affect overall query performance. In this paper, we address this by a systematic study evaluating LCMs on three of the core query optimization tasks: join ordering, access path selection, and physical operator selection . In our study, we compare seven state-of-the-art LCMs to a traditional cost model and, surprisingly, find that the traditional model often still outperforms LCMs in these tasks. We conclude by highlighting major takeaways and recommendations to guide future research toward making LCMs more effective for query optimization.

Learned Cost Model s (LCMs) have shown superior results over traditional database cost models as they can significantly improve the accuracy of cost predictions. However, LCMs still fail for some query plans, as prediction errors can be large in the tail. Unfortunately, recent LCMs are based on complex deep neural models, and thus, there is no easy way to understand where this accuracy drop is rooted, which critically prevents systematic troubleshooting. In this demo paper, we present the very first approach for opening the black box by bringing AI explainability approaches to LCMs. As a core contribution, we developed new explanation techniques that extend existing methods that are available for the general explainability of AI models and adapt them significantly to be usable for LCMs. In our demo, we provide an interactive tool to showcase how explainability for LCMs works. We believe this is a first step for making LCMs debuggable and thus paving the road for new approaches for systematically fixing problems in LCMs.

Learned cost models (LCMs) have recently gained traction as a promising alternative to traditional cost estimation techniques in data management, offering improved accuracy by capturing complex interactions between queries, data, and runtime behavior. While initially developed for batch systems, LCMs are now increasingly applied to stream processing as well, where real-time demands pose new challenges. This tutorial presents the first unified overview of LCMs across both batch and stream processing systems, examining their role as essential components in modern query optimizers. We explore key aspects of LCM design—including input representations and model architectures—and highlight how these models deal with query optimization tasks.

This paper introduces ZEROTuNE, a novel cost model for parallel and distributed stream processing that can be used to effectively set initial parallelism degrees of streaming queries. Unlike existing models, which rely majorly on online learning statistics that are non-transferable, context-specific, and require extensive training, ZEROTuNE proposes data-efficient zero-shot learning techniques that enable very accurate cost predictions without having observed any query deployment. To overcome these challenges, we propose ZEROTuNE, a graph neural network architecture that can learn from the structural complexity of parallel distributed stream processing systems, enabling them to adapt to unseen workloads and hardware configurations. In our experiments, we show when integrating ZEROTuNE in a distributed streaming system such as Apache Flink, we can accurately set the degree of parallelism, showing an average speed-up of around 5× in comparison to existing approaches.

In this work, we present Costream, a novel learned cost model for Distributed Stream Processing Systems that provides accurate predictions of the execution costs of a streaming query in an edge-cloud environment. The cost model can be used to find an initial placement of operators across heterogeneous hardware, which is particularly important in these environments. In our evaluation, we demonstrate that Costream can produce highly accurate cost estimates for the initial operator placement and even generalize to unseen placements, queries, and hardware. When using Costream to optimize the placements of streaming operators, a median speedup of around 21 × can be achieved compared to baselines.

In this paper, we present a novel pipelined GPU join that accelerates the performance of distributed DBMSs by leveraging GPU resources on fast networks. A key insight is that we enable pipelined join execution by overlapping the network shuffling with the build and probe phases, thereby significantly reducing the GPU idle time. To demonstrate this, we propose novel algorithms for distributed pipelined GPU joins with RDMA and GPUDirect for both arbitrarily large probe- and build-side tables. In our evaluation, we show our pipelined distributed GPU join can reduce the overall runtime of a full query by up to 6× against a state-of-the-art CPU-only join.

Deep learning (DL) inference has become an essential building block in modern intelligent applications. Due to the high computational intensity of DL, it is critical to scale DL inference serving systems in response to fluctuating workloads to achieve resource efficiency. Meanwhile, intelligent applications often require strict service level agreements (SLAs), which need to be guaranteed when the system is scaled. The problem is complex and has been tackled only in simple scenarios so far.This paper describes FA2, a fast and accurate autoscaler concept for DL inference serving systems. In contrast to related works, FA2 adopts a general, contrived two-phase approach. Specifically, it starts by capturing the autoscaling challenges in a comprehensive graph-based model. Then, FA2 applies targeted graph transformation and makes autoscaling decisions with an efficient algorithm based on dynamic programming. We implemented FA2 and built and evaluated a prototype. Compared with state-of-the-art autoscaling solutions, our experiments showed FA2 to achieve significant resource reduction (19% under CPUs and 25% under GPUs, on average) in combination with low SLA violations (less than 1.5%). FA2 performed close to the theoretical optimum, matching exactly the optimal decisions (with the least required resources) in 96.8% of all the cases in our evaluation.

Complex Event Processing (CEP) is a powerful paradigm for scalable data management that is employed in many real-world scenarios such as detecting credit card fraud in banks. The so-called complex events are expressed using a specification language that is typically implemented and executed on a specific runtime system. While the tight coupling of these two components has been regarded as the key for supporting CEP at high performance, such dependencies pose several inherent challenges as follows. (1) Application development atop a CEP system requires extensive knowledge of how the runtime system operates, which is typically highly complex in nature. (2) The specification language dependence requires the need of domain experts and further restricts and steepens the learning curve for application developers.In this paper, we propose CEPless, a scalable data management system that decouples the specification from the runtime system by building on the principles of serverless computing. CEPless provides "operator as a service" and offers flexibility by enabling the development of CEP application in any specification language while abstracting away the complexity of the CEP runtime system. As part of CEPless, we designed and evaluated novel mechanisms for in-memory processing and batching that enable the stateful processing of CEP operators even under high rates of ingested events. Our evaluation demonstrates that CEPless can be easily integrated into existing CEP systems like Apache Flink while attaining similar throughput under high scale of events (up to 100K events per second) and dynamic operator update in ~238 ms.

Publish/subscribe systems have to react fast on changes in their environment while handling many events with low end-to-end latency and high throughput. Moving the broker functionality of publish/subscribe systems to the underlying network layer reduces the path length of events and, in addition, forwarding benefits from powerful and programmable hardware. So far attempts of underlay publish/subscribe depend on a specific API of the network devices, e. g., the OpenFlow protocol, which have restrictions in dealing with dynamic devices and corresponding changes in the introduced attribute names for matching and filtering events.In this work, we focus on the next generation of network devices, which are envisioned to provide reconfigurable hardware components, specified by the open P4 description language. We introduce two new approaches that enable a flexible and generic attribute/value encoding, understandable by P4-capable packet processors, to benefit from the performance properties of hardware. Furthermore, the proposed approaches reduce the effort in encoding and decoding event messages.

Emerging network architectures like Information-Centric Networking (ICN)offer simplicity in the data plane by addressing named data. Such flexibility opens up the possibility to move data processing inside network elements for high-performance computation, known as in-network processing. However, existing ICN architectures are limited in terms of (i)in-network processing and (ii)data plane programming abstractions. Such architectures can benefit from Complex Event Processing (CEP), an in-network processing paradigm to efficiently process data inside the data plane. Yet, it is extremely challenging to integrate CEP because the current communication model of ICN is limited to consumer-initiated interaction that comes with significant overhead in number of requests to process continuous data streams. In contrast, a change to producer-initiated interaction, as favored by CEP, imposes severe limitations for request-reply interactions. In this paper, we propose an in-network CEP architecture, INETCEP that supports unified interaction patterns (consumer- and producer-initiated). In addition, we provide a CEP query language and facilitate CEP operations while increasing the range of applications that can be supported by ICN. We provide an open source implementation and evaluation of INETCEP over an ICN architecture, Named Function Networking, and two applications: energy forecasting in smart homes and a disaster scenario.

Countless novel approaches to communication protocols, overlay networks, and distributed middleware are published every year, yet the adoption of such novel findings in the global Internet landscape progresses at a slow pace. Many of such new communication mechanisms excel (only) under specific deployment conditions, while user mobility and application usage patterns lead to dynamic operation conditions. This mismatch is one reason that makes a wide deployment of new specialized mechanisms particularly hard as observed, for example, for multipath transport protocol extensions until the emergence of multipath transmission control protocol (TCP). This paper formalizes the concept of Transitions, i.e., a method to instrumentalize adaptivity at runtime in communication systems. It allows to exchange communication mechanisms in a running system to optimize the communication quality. In the following, we describe the building blocks required to: 1) capture the features and relations within a communication system and 2) express and optimize the decision making process in such a system. We show how this concept maps intuitively to the Internet model which makes a protocol-independent deployment of applications feasible in the future Internet.

Crowd sensing exploits users' smart devices and human mobility to collect information on a large scale. To realize a crowd sensing campaign, sensing tasks with spatio-temporal requirements are distributed to the devices that can provide the requested information. Typically, the distribution of sensing tasks relies on a centralized communication infrastructure such as cloud servers. However, such an approach will be unsuitable if access to communication infrastructure is restricted, for example in disaster relief scenarios. To fill this gap, we propose a distributed context-aware framework for disseminating sensing tasks, based on the Named Data Networking (NDN) paradigm. By adding context attributes to Interest packets, we allow a device to utilize this information to make forwarding decisions autonomously, thus guiding the sensing tasks towards the suitable sensing devices. Through intensive evaluation, we show that our framework achieves a timely delivery of sensing tasks, while keeping the communication overhead to a minimum compared to pure geo forwarding and flooding approaches.

In comparison to professional productions, UGV often suffers from quality degradations such as camera shakes or harmful occlusions. Quick and reliable algorithms are needed to detect and assess the impact of these quality degradations. Recently, researchers leverage data from auxiliary sensors such as gyroscope readings or accelerometer values for quality assessment. In comparison to video-based approaches, auxiliary sensor-based ones achieve results more quickly but often lack precision.