arXiv subjects

Otmar Ertl

Publications and source records attributed to Otmar Ertl.

18 recordsLinked to original sources

arXiv2026

BlobShuffle: Cost-Effective Repartitioning in Stream Processing Systems via Object Storage Exemplified with Kafka Streams

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.

cs.DC↗

arXiv2026

FluxSieve: Unifying Streaming and Analytical Data Planes for Scalable Cloud Observability

Despite many advances in query optimization, indexing techniques, and data storage, modern data platforms still face difficulties in delivering robust query performance under high concurrency and computationally intensive queries. This challenge is particularly pronounced in large-scale observability platforms handling high-volume, high-velocity data records. For instance, recurrent, expensive filtering queries at query time impose substantial computational and storage overheads in the analytical data plane. In this paper, we propose FluxSieve, a unified architecture that reconciles traditional pull-based query processing with push-based stream processing by embedding a lightweight in-stream precomputation and filtering layer directly into the data ingestion path. This avoids the complexity and operational burden of running queries in dedicated stream processing frameworks. Concretely, this work (i) introduces a foundational architecture that unifies streaming and analytical data planes via in-stream filtering and records enrichment, (ii) designs a scalable multi-pattern matching mechanism that supports concurrent evaluation and on-the-fly updates of filtering rules with minimal per-record overhead, (iii) demonstrates how to integrate this ingestion-time processing with two open-source analytical systems -- Apache Pinot as a Real-Time Online Analytical Processing (RTOLAP) engine and DuckDB as an embedded analytical database, and (iv) performs comprehensive experimental evaluation of our approach. Our evaluation across different systems, query types, and performance metrics shows up to orders-of-magnitude improvements in query performance at the cost of negligible additional storage and very low computational overhead.

cs.DB↗

arXiv2025

Should I Run My Cloud Benchmark on Black Friday?

Benchmarks and performance experiments are frequently conducted in cloud environments. However, their results are often treated with caution, as the presumed high variability of performance in the cloud raises concerns about reproducibility and credibility. In a recent study, we empirically quantified the impact of this variability on benchmarking results by repeatedly executing a stream processing application benchmark at different times of the day over several months. Our analysis confirms that performance variability is indeed observable at the application level, although it is less pronounced than often assumed. The larger scale of our study compared to related work allowed us to identify subtle daily and weekly performance patterns. We now extend this investigation by examining whether a major global event, such as Black Friday, affects the outcomes of performance benchmarks.

cs.SE↗

arXiv2025

When Should I Run My Application Benchmark?: Studying Cloud Performance Variability for the Case of Stream Processing Applications

Performance benchmarking is a common practice in software engineering, particularly when building large-scale, distributed, and data-intensive systems. While cloud environments offer several advantages for running benchmarks, it is often reported that benchmark results can vary significantly between repetitions -- making it difficult to draw reliable conclusions about real-world performance. In this paper, we empirically quantify the impact of cloud performance variability on benchmarking results, focusing on stream processing applications as a representative type of data-intensive, performance-critical system. In a longitudinal study spanning more than three months, we repeatedly executed an application benchmark used in research and development at Dynatrace. This allows us to assess various aspects of performance variability, particularly concerning temporal effects. With approximately 591 hours of experiments, deploying 789 Kubernetes clusters on AWS and executing 2366 benchmarks, this is likely the largest study of its kind and the only one addressing performance from an end-to-end, i.e., application benchmark perspective. Our study confirms that performance variability exists, but it is less pronounced than often assumed (coefficient of variation of < 3.7%). Unlike related studies, we find that performance does exhibit a daily and weekly pattern, although with only small variability (<= 2.5%). Re-using benchmarking infrastructure across multiple repetitions introduces only a slight reduction in result accuracy (<= 2.5 percentage points). These key observations hold consistently across different cloud regions and machine types with different processor architectures. We conclude that for engineers and researchers focused on detecting substantial performance differences (e.g., > 5%) in...

cs.SE↗

arXiv2024

UltraLogLog: A Practical and More Space-Efficient Alternative to HyperLogLog for Approximate Distinct Counting

Since its invention HyperLogLog has become the standard algorithm for approximate distinct counting. Due to its space efficiency and suitability for distributed systems, it is widely used and also implemented in numerous databases. This work presents UltraLogLog, which shares the same practical properties as HyperLogLog. It is commutative, idempotent, mergeable, and has a fast guaranteed constant-time insert operation. At the same time, it requires 28% less space to encode the same amount of distinct count information, which can be extracted using the maximum likelihood method. Alternatively, a simpler and faster estimator is proposed, which still achieves a space reduction of 24%, but at an estimation speed comparable to that of HyperLogLog. In a non-distributed setting where martingale estimation can be used, UltraLogLog is able to reduce space by 17%. Moreover, its smaller entropy and its 8-bit registers lead to better compaction when using standard compression algorithms. All this is verified by experimental results that are in perfect agreement with the theoretical analysis which also outlines potential for even more space-efficient data structures. A production-ready Java implementation of UltraLogLog has been released as part of the open-source Hash4j library.

cs.DS↗

arXiv2024

A Comprehensive Benchmarking Analysis of Fault Recovery in Stream Processing Frameworks

Nowadays, several software systems rely on stream processing architectures to deliver scalable performance and handle large volumes of data in near real-time. Stream processing frameworks facilitate scalable computing by distributing the application's execution across multiple machines. Despite performance being extensively studied, the measurement of fault tolerance-a key feature offered by stream processing frameworks-has still not been measured properly with updated and comprehensive testbeds. Moreover, the impact that fault recovery can have on performance is mostly ignored. This paper provides a comprehensive analysis of fault recovery performance, stability, and recovery time in a cloud-native environment with modern open-source frameworks, namely Flink, Kafka Streams, and Spark Structured Streaming. Our benchmarking analysis is inspired by chaos engineering to inject failures. Generally, our results indicate that much has changed compared to previous studies on fault recovery in distributed stream processing. In particular, the results indicate that Flink is the most stable and has one of the best fault recovery. Moreover, Kafka Streams shows performance instabilities after failures, which is due to its current rebalancing strategy that can be suboptimal in terms of load balancing. Spark Structured Streaming shows suitable fault recovery performance and stability, but with higher event latency. Our study intends to (i) help industry practitioners in choosing the most suitable stream processing framework for efficient and reliable executions of data-intensive applications; (ii) support researchers in applying and extending our research method as well as our benchmark; (iii) identify, prevent, and assist in solving potential issues in production deployments.

cs.DC↗

arXiv2024

High-level Stream Processing: A Complementary Analysis of Fault Recovery

Parallel computing is very important to accelerate the performance of software systems. Additionally, considering that a recurring challenge is to process high data volumes continuously, stream processing emerged as a paradigm and software architectural style. Several software systems rely on stream processing to deliver scalable performance, whereas open-source frameworks provide coding abstraction and high-level parallel computing. Although stream processing's performance is being extensively studied, the measurement of fault tolerance--a key abstraction offered by stream processing frameworks--has still not been adequately measured with comprehensive testbeds. In this work, we extend the previous fault recovery measurements with an exploratory analysis of the configuration space, additional experimental measurements, and analysis of improvement opportunities. We focus on robust deployment setups inspired by requirements for near real-time analytics of a large cloud observability platform. The results indicate significant potential for improving fault recovery and performance. However, these improvements entail grappling with configuration complexities, particularly in identifying and selecting the configurations to be fine-tuned and determining the appropriate values for them. Therefore, new abstractions for transparent configuration tuning are also needed for large-scale industry setups. We believe that more software engineering efforts are needed to provide insights into potential abstractions and how to achieve them. The stream processing community and industry practitioners could also benefit from more interactions with the high-level parallel programming community, whose expertise and insights on making parallel programming more productive and efficient could be extended.

cs.DC↗

arXiv2024

JumpBackHash: Say Goodbye to the Modulo Operation to Distribute Keys Uniformly to Buckets

Introduction. Distributed data processing and storage systems require efficient methods to distribute keys across buckets. While simple and fast, the traditional modulo-based mapping is unstable when the number of buckets changes, leading to spikes in system resource utilization, such as network or database requests. Consistent hash algorithms minimize remappings but are either significantly slower, require floating-point arithmetic, or are based on a family of hash functions rarely available in standard libraries. This work introduces JumpBackHash, a consistent hash algorithm that overcomes those shortcomings. Methodology. JumpBackHash applies the concept of active indices borrowed from consistent weighted sampling, which inherently leads to consistency. It generates the active indices in reverse order, which avoids floating-point operations, enables the minimization of consumed random values and the use of a standard pseudorandom generator, and finally leads to a very efficient algorithm. Results. Theoretical analysis shows that JumpBackHash has an expected constant runtime. The expected value and the variance of the number of consumed random values perfectly agree with the experiments. Empirical tests also confirm the consistency. Conclusion. JumpBackHash offers a fast and efficient solution for uniformly distributing keys across buckets in distributed systems. Its simplicity, performance, and the availability of a production-ready Java implementation as part of the Hash4j open source library make it a viable replacement for the modulo-based approach for improving assignment and system stability.

cs.DS↗

arXiv2024

ShuffleBench: A Benchmark for Large-Scale Data Shuffling Operations with Distributed Stream Processing Frameworks

Distributed stream processing frameworks help building scalable and reliable applications that perform transformations and aggregations on continuous data streams. This paper introduces ShuffleBench, a novel benchmark to evaluate the performance of modern stream processing frameworks. In contrast to other benchmarks, it focuses on use cases where stream processing frameworks are mainly employed for shuffling (i.e., re-distributing) data records to perform state-local aggregations, while the actual aggregation logic is considered as black-box software components. ShuffleBench is inspired by requirements for near real-time analytics of a large cloud observability platform and takes up benchmarking metrics and methods for latency, throughput, and scalability established in the performance engineering research community. Although inspired by a real-world observability use case, it is highly configurable to allow domain-independent evaluations. ShuffleBench comes as a ready-to-use open-source software utilizing existing Kubernetes tooling and providing implementations for four state-of-the-art frameworks. Therefore, we expect ShuffleBench to be a valuable contribution to both industrial practitioners building stream processing applications and researchers working on new stream processing approaches. We complement this paper with an experimental performance evaluation that employs ShuffleBench with various configurations on Flink, Hazelcast, Kafka Streams, and Spark in a cloud-native environment. Our results show that Flink achieves the highest throughput while Hazelcast processes data streams with the lowest latency.

cs.SE↗

arXiv2024

ExaLogLog: Space-Efficient and Practical Approximate Distinct Counting up to the Exa-Scale

This work introduces ExaLogLog, a new data structure for approximate distinct counting, which has the same practical properties as the popular HyperLogLog algorithm. It is commutative, idempotent, mergeable, reducible, has a constant-time insert operation, and supports distinct counts up to the exa-scale. At the same time, as theoretically derived and experimentally verified, it requires 43% less space to achieve the same estimation error.

cs.DS↗

arXiv2021

SetSketch: Filling the Gap between MinHash and HyperLogLog

MinHash and HyperLogLog are sketching algorithms that have become indispensable for set summaries in big data applications. While HyperLogLog allows counting different elements with very little space, MinHash is suitable for the fast comparison of sets as it allows estimating the Jaccard similarity and other joint quantities. This work presents a new data structure called SetSketch that is able to continuously fill the gap between both use cases. Its commutative and idempotent insert operation and its mergeable state make it suitable for distributed environments. Fast, robust, and easy-to-implement estimators for cardinality and joint quantities, as well as the ability to use SetSketch for similarity search, enable versatile applications. The presented joint estimator can also be applied to other data structures such as MinHash, HyperLogLog, or HyperMinHash, where it even performs better than the corresponding state-of-the-art estimators in many cases.

cs.DS↗

arXiv2021

Estimation from Partially Sampled Distributed Traces

Sampling is often a necessary evil to reduce the processing and storage costs of distributed tracing. In this work, we describe a scalable and adaptive sampling approach that can preserve events of interest better than the widely used head-based sampling approach. Sampling rates can be chosen individually and independently for every span, allowing to take span attributes and local resource constraints into account. The resulting traces are often only partially and not completely sampled which complicates statistical analysis. To exploit the given information, an unbiased estimation algorithm is presented. Even though it does not need to know whether the traces are complete, it reduces the estimation error in many cases compared to considering only complete traces.

cs.DS↗

arXiv2020

ProbMinHash -- A Class of Locality-Sensitive Hash Algorithms for the (Probability) Jaccard Similarity

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.

cs.DS↗

arXiv2019

Computing Extremely Accurate Quantiles Using t-Digests

We present on-line algorithms for computing approximations of rank-based statistics that give high accuracy, particularly near the tails of a distribution, with very small sketches. Notably, the method allows a quantile $q$ to be computed with an accuracy relative to $\max(q, 1-q)$ rather than absolute accuracy as with most other methods. This new algorithm is robust with respect to skewed distributions or ordered datasets and allows separately computed summaries to be combined with no loss in accuracy. An open-source Java implementation of this algorithm is available from the author. Independent implementations in Go and Python are also available.

stat.CO↗

arXiv2018

BagMinHash - Minwise Hashing Algorithm for Weighted Sets

Minwise hashing has become a standard tool to calculate signatures which allow direct estimation of Jaccard similarities. While very efficient algorithms already exist for the unweighted case, the calculation of signatures for weighted sets is still a time consuming task. BagMinHash is a new algorithm that can be orders of magnitude faster than current state of the art without any particular restrictions or assumptions on weights or data dimensionality. Applied to the special case of unweighted sets, it represents the first efficient algorithm producing independent signature components. A series of tests finally verifies the new algorithm and also reveals limitations of other approaches published in the recent past.

cs.DS↗

arXiv2017

New Cardinality Estimation Methods for HyperLogLog Sketches

This work presents new cardinality estimation methods for data sets recorded by HyperLogLog sketches. A simple derivation of the original estimator was found, that also gives insight how to correct its deficiencies. The result is an improved estimator that is unbiased over the full cardinality range, is easy computable, and does not rely on empirically determined data as previous approaches. Based on the maximum likelihood principle a second unbiased estimation method is presented which can also be extended to estimate cardinalities of union, intersection, or relative complements of two sets that are both represented as HyperLogLog sketches. Experimental results show that this approach is more precise than the conventional technique using the inclusion-exclusion principle.

cs.DS↗

arXiv2017

SuperMinHash - A New Minwise Hashing Algorithm for Jaccard Similarity Estimation

This paper presents a new algorithm for calculating hash signatures of sets which can be directly used for Jaccard similarity estimation. The new approach is an improvement over the MinHash algorithm, because it has a better runtime behavior and the resulting signatures allow a more precise estimation of the Jaccard index.

cs.DS↗

arXiv2017

New cardinality estimation algorithms for HyperLogLog sketches

This paper presents new methods to estimate the cardinalities of data sets recorded by HyperLogLog sketches. A theoretically motivated extension to the original estimator is presented that eliminates the bias for small and large cardinalities. Based on the maximum likelihood principle a second unbiased method is derived together with a robust and efficient numerical algorithm to calculate the estimate. The maximum likelihood approach can also be applied to more than a single HyperLogLog sketch. In particular, it is shown that it gives more precise cardinality estimates for union, intersection, or relative complements of two sets that are both represented by HyperLogLog sketches compared to the conventional technique using the inclusion-exclusion principle. All the new methods are demonstrated and verified by extensive simulations.

cs.DS↗