Otmar Ertl

Sören Henning, Otmar Ertl, Adriano Vogel

Shuffling or repartitioning data streams is an essential operation of state-of-the-art stream processing frameworks to support stateful workloads in a large-scale, distributed setting. In today's cloud deployments, however, shuffling can become a major cost driver due to substantial network traffic across multiple availability zones (AZs) as well as an operational burden when operating a high-throughput, strongly consistent messaging backbone at scale. We present BlobShuffle, a novel approach to cost-effective shuffling for stream processing systems that leverages cloud object storage as an intermediate exchange layer. Instead of sending all shuffled records directly, BlobShuffle groups records into batches, stores these batches in cloud object storage, and forwards only compact notifications. Downstream operators use these notifications to retrieve the relevant batches and extract the corresponding records. BlobShuffle balances cost efficiency and latency through configurable batching and a distributed caching mechanism. BlobShuffle is implemented as an add-on for Kafka Streams that requires only minimal code changes to existing applications, leaves Kafka and the underlying infrastructure unmodified, and preserves Kafka Streams' consistency and correctness guarantees. In a large-scale experimental evaluation on a Kubernetes-based AWS deployment, we show that BlobShuffle can reduce shuffling costs by more than 40x compared to native Kafka Streams shuffling while keeping the 95th percentile shuffle latency below 2 seconds. Moreover, it scales to processing more than 2 GiB/s without encountering a scalability limit in our experiments, indicating that BlobShuffle can economically support shuffle-intensive workloads at large scale.

Otmar Ertl

The probability Jaccard similarity was recently proposed as a natural generalization of the Jaccard similarity to measure the proximity of sets whose elements are associated with relative frequencies or probabilities. In combination with a hash algorithm that maps those weighted sets to compact signatures which allow fast estimation of pairwise similarities, it constitutes a valuable method for big data applications such as near-duplicate detection, nearest neighbor search, or clustering. This paper introduces a class of one-pass locality-sensitive hash algorithms that are orders of magnitude faster than the original approach. The performance gain is achieved by calculating signature components not independently, but collectively. Four different algorithms are proposed based on this idea. Two of them are statistically equivalent to the original approach and can be used as drop-in replacements. The other two may even improve the estimation error by introducing statistical dependence between signature components. Moreover, the presented techniques can be specialized for the conventional Jaccard similarity, resulting in highly efficient algorithms that outperform traditional minwise hashing and that are able to compete with the state of the art.