She Crashed a Raspberry Pi to Fix IoT Fault Tolerance?! [SIGMOD 2024] with Anastasiia Kozar скачать в хорошем качестве

She Crashed a Raspberry Pi to Fix IoT Fault Tolerance?! [SIGMOD 2024] with Anastasiia Kozar

What happens when billions of IoT devices start flooding the cloud…and your “fault tolerance” silently kills the edge? 🎯 In this episode of Disseminate, host Jack Waudby sits down with Anastasiia Kozar (PhD, TU Berlin) to unpack real-world stream processing for the Internet of Things, her work on NebulaStream, and why treating fault tolerance as a first-class citizen can be the difference between insights and outages. We cover: Why pushing all sensor data to the cloud backfires—and how edge/near-sensor compute saves the day The hidden costs of “default on” fault tolerance in heterogeneous IoT networks Upstream backup vs. active standby vs. checkpointing—when each makes sense A wild experiment: enabling fault tolerance on a 1GB Raspberry Pi and watching capacity collapse…then clawing it back with smarter placement & dynamic trimming Practical takeaways for engineers deploying Flink-like systems vs. IoT-native platforms such as NebulaStream How to estimate memory/network trade-offs, pick trimming intervals, and avoid OOM at the edge Research-to-practice advice for practitioners, researchers, and students 👩‍🔬 Guest: Anastasiia Kozar, Technical University of Berlin 🎙️ Host: Jack Waudby 🔬 Themes: IoT, Stream Processing, Edge Computing, Fault Tolerance, Databases, Systems Research Who should watch Industry practitioners building IoT/edge data pipelines Researchers translating SIGMOD-grade ideas into production systems Students learning how to bridge theory and systems reality Chapters 00:20 Intro & what this episode is about 00:48 Meet Anastasiia + her path into data management research 01:39 IoT in the wild: cow-behavior sensing in Bavaria 03:27 Defining IoT, stream processing engines, & fault tolerance 05:33 NebulaStream: processing on the way from sensor → edge → cloud 07:47 Why cloud-centric fault tolerance fails on constrained devices 10:26 The capacity cliff: turning on fault tolerance and killing the Pi 16:56 Making fault tolerance a first-class operator (placement & cost) 18:03 Fault tolerance strategies: upstream backup, active standby, checkpointing 23:21 Naïve vs. multi-objective placement (reliability vs. resources) 27:26 Dynamic trimming: trading network for memory to avoid OOM 29:57 Scaling experiments (Kubernetes, many “Pis”) & key results 46:25 NebulaStream today & research roadmap 47:28 Final takeaways for building reliable IoT pipelines #IoT #StreamProcessing #EdgeComputing #FaultTolerance #DistributedSystems #DataEngineering #Databases #NebulaStream #ApacheFlink #SIGMOD #ComputerScience #SystemsResearch #ResearchToPractice #RaspberryPi #CloudComputing