pith. sign in

arxiv: 2606.21430 · v1 · pith:44RGWNWPnew · submitted 2026-06-19 · 💻 cs.DC

rush: Scalable Asynchronous Distributed Computing via Shared State in R

Marc Becker , Bernd Bischl This is my paper

Pith reviewed 2026-06-26 12:59 UTC · model grok-4.3

classification 💻 cs.DC
keywords asynchronous computingdistributed systemsR packageRedisBayesian optimizationhyperparameter optimizationshared statemlr3
0
0 comments X

The pith

rush provides a Redis-based shared-state layer enabling asynchronous decentralized parallel algorithms in R.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing R packages for parallel computing rely on centralized controller-worker models that cannot support patterns where workers communicate through shared state. rush fills this gap with a high-level API for task lifecycle management built on a Redis key-value store. Workers independently read and write task data, execute their loops, and benefit from claimed sub-millisecond overhead, automatic lost-worker detection, and caching. The package is shown powering asynchronous decentralized Bayesian optimization for LightGBM hyperparameter tuning across four datasets on 448 workers, and it integrates optionally with the mlr3 ecosystem.

Core claim

rush is an R package supplying a shared-state coordination layer for asynchronously parallelized iterative algorithms; it uses a Redis database as a shared key-value store so that workers read and write task data through the database and execute their own independent loops, delivering a high-level API for task management with sub-millisecond per-task overhead, robust error handling including automatic lost-worker detection, and efficient caching.

What carries the argument

The Redis-based shared key-value store that lets workers coordinate tasks and lifecycle without a central controller.

If this is right

  • Asynchronous decentralized Bayesian optimization becomes directly usable inside R for hyperparameter tuning.
  • The mlr3 ecosystem gains native support for async optimization through bbotk and mlr3tuning.
  • Task management with automatic worker-failure handling and caching works at scale up to at least 448 workers on LightGBM datasets.
  • Other iterative algorithms in statistics and machine learning that rely on shared state can be implemented without a central dispatcher.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same shared-state pattern could be ported to other languages whose parallel packages remain centralized.
  • Performance under heterogeneous or high-latency networks would need separate validation beyond the reported benchmark.
  • The caching layer may reduce redundant evaluations in longer-running black-box optimization loops.

Load-bearing premise

A Redis key-value store can deliver sub-millisecond overhead and reliable error handling for lost workers under real network conditions and failures without additional mechanisms.

What would settle it

A benchmark of the described asynchronous Bayesian optimization run that measures average per-task overhead above one millisecond or fails to detect lost workers automatically.

Figures

Figures reproduced from arXiv: 2606.21430 by Bernd Bischl, Marc Becker.

Figure 1
Figure 1. Figure 1: Centralized (left) and decentralized (right) networks. Octagons denote processes, including the main process M and workers W. In the centralized network, M proposes tasks to the workers and processes the returned computations. In the decentralized network, each worker W independently proposes tasks, evaluates them, and exchanges task information via a shared Redis database (red rectangle). also our primary… view at source ↗
Figure 2
Figure 2. Figure 2: The communication flow between a worker and the Redis database in a rush network. The oc￾tagon represents a worker and the rectangle represents the Redis database. In each iteration, the worker reads shared state via $fetch_tasks_with_state(), registers a new task via $push_running_tasks(), computes the result, and writes it back via $finish_tasks(). to the Redis API and exposes the command set through a l… view at source ↗
Figure 3
Figure 3. Figure 3: Runtime of fetching finished tasks with (blue) and without (red) caching for different numbers of parameters as a function of the total number of tasks in the database. The payload size is 1 in this benchmark. Without caching, all tasks are retrieved from the database on each call. With caching, only one new task is retrieved from the database; all previously fetched tasks are served from the worker’s cach… view at source ↗
Figure 4
Figure 4. Figure 4: CPU utilization for short-running and long-running objective functions. Bars represent the proposal phase (red) and evaluation phase (blue) of the BO loop. Bar height indicates CPU utilization on a 10-core machine: during proposal, CL uses a single core, whereas during evaluation it uses all 10 cores. The proposal phase grows over iterations as surrogate fitting becomes more expensive with an increasing nu… view at source ↗
Figure 5
Figure 5. Figure 5: Classification error over wall-clock time of the three algorithms [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
read the original abstract

Many algorithms in statistics and machine learning can be parallelized in an asynchronous manner where workers need to communicate through shared state rather than execute independent tasks dispatched by a central controller. Especially in modern hyperparameter optimization and parallel black-box optimization with expensive objectives, this decentralized approach has become widespread, and several Python frameworks adopt it (e.g., Optuna, DeepHyper, and Hyperopt). However, all popular R packages for parallel computing follow a centralized controller-worker architecture that does not support this pattern. We present rush, an R package that provides a shared-state coordination layer for asynchronously parallelized iterative algorithms. rush uses a Redis database as a shared key-value store: workers read and write task data through the database and independently execute their own loops. The package provides a high-level API for managing tasks and their lifecycle, featuring sub-millisecond per-task overhead, robust error handling with automatic detection of lost workers, and efficient caching. rush optionally integrates with the mlr3 ecosystem, powering asynchronous optimization in the bbotk and mlr3tuning packages. We demonstrate the practical utility of rush by implementing asynchronous decentralized Bayesian optimization (ADBO) and benchmarking it on hyperparameter optimization of LightGBM across four datasets using 448 workers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces the rush R package for scalable asynchronous distributed computing using a Redis shared state for task management in iterative algorithms, particularly for hyperparameter optimization. It claims sub-millisecond per-task overhead, robust error handling including automatic lost worker detection, and efficient caching, demonstrated through an implementation of asynchronous decentralized Bayesian optimization on LightGBM hyperparameter tuning with 448 workers across four datasets, integrated with the mlr3 ecosystem.

Significance. If the performance and robustness claims hold, rush would fill a gap in R's parallel computing tools by supporting decentralized asynchronous patterns used in modern ML optimization frameworks like Optuna, potentially benefiting the mlr3 ecosystem. The high-level API and practical demonstration are strengths, but the lack of quantitative results limits evaluation of its significance.

major comments (2)
  1. [Abstract] Abstract: The claims of 'sub-millisecond per-task overhead' and 'robust error handling with automatic detection of lost workers' lack any supporting quantitative benchmark data, implementation details on Redis operations (such as SET/GET patterns, EXPIRE, WATCH/MULTI, or client-side recovery logic), or measurements under network conditions, making it impossible to verify if these are achieved or merely asserted.
  2. [Abstract] Abstract: The demonstration of asynchronous decentralized Bayesian optimization is described as using 448 workers across four datasets, but no results, timings, error rates, or comparisons are provided, so the practical utility cannot be assessed from the given information.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their comments. We agree that the abstract would be strengthened by direct references to supporting data and results, and we will revise it in the next version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claims of 'sub-millisecond per-task overhead' and 'robust error handling with automatic detection of lost workers' lack any supporting quantitative benchmark data, implementation details on Redis operations (such as SET/GET patterns, EXPIRE, WATCH/MULTI, or client-side recovery logic), or measurements under network conditions, making it impossible to verify if these are achieved or merely asserted.

    Authors: We acknowledge that the abstract presents these performance and robustness claims without inline quantitative support or implementation specifics. The full manuscript contains benchmark results and Redis usage details in later sections, but the abstract does not reference them. We will revise the abstract to incorporate key quantitative findings and a concise description of the Redis operations and recovery logic. revision: yes

  2. Referee: [Abstract] Abstract: The demonstration of asynchronous decentralized Bayesian optimization is described as using 448 workers across four datasets, but no results, timings, error rates, or comparisons are provided, so the practical utility cannot be assessed from the given information.

    Authors: We agree that the abstract summarizes the experimental setup but omits outcome metrics. The manuscript's experimental section reports the results of the ADBO runs. We will revise the abstract to include a brief summary of the key timings, error rates, and performance observations from those experiments. revision: yes

Circularity Check

0 steps flagged

No circularity: software package description with empirical benchmarks only

full rationale

The manuscript presents the rush R package for shared-state asynchronous coordination via Redis. No derivation chain, equations, fitted parameters, or predictions exist. Claims rest on API design, implementation details, and runtime benchmarks (e.g., sub-millisecond overhead on 448 workers for LightGBM tuning). These are direct measurements, not reductions to self-referential inputs. No self-citation load-bearing steps or uniqueness theorems are invoked. The paper is self-contained as an engineering artifact against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a software package paper. No mathematical free parameters, axioms, or invented entities are introduced or required by the central claim.

pith-pipeline@v0.9.1-grok · 5743 in / 1178 out tokens · 15638 ms · 2026-06-26T12:59:45.062689+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 25 canonical work pages · 2 internal anchors

  1. [1]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages =

    doi: 10.1145/3292500.3330701. [p3, 8] T. Barrett, M. Dowle, A. Srinivasan, J. Gorecki, M. Chirico, T. Hocking, B. Schwendinger, and I. Krylov. Data.T able: Extension of Data.Frame,

    work page doi:10.1145/3292500.3330701

  2. [2]

    URL https://doi.org/10.32614/CRAN.package.data. table. R package version 1.18.4. [p4] The R Journal Vol. XX/YY, AAAA 20ZZ ISSN 2073-4859 CONTRIBUTED RESEARCH ARTICLE 17 M. Becker, M. Lang, B. Bischl, and J. Richter. Mlr3tuning: Hyperparameter Optimization for ’mlr3’. Zenodo, July

    work page doi:10.32614/cran.package.data 2073

  3. [3]

    mlr3mbo: Bayesian Optimization in R

    URLhttps://doi.org/10.5281/zenodo.15852621. [p1, 12] M. Becker, J. Richter, M. Lang, L. Schneider, B. Bischl, P . Schratz, M. Binder, S. Fischer, M. Chirico, M. López-Ibáñez, L. Henry, and L. Burk. Bbotk: Black-Box Optimization Toolkit. Zenodo, June 2026a. URLhttps://doi.org/10.5281/zenodo.20669808. [p1] M. Becker, L. Schneider, M. Binder, L. Kotthoff, an...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.5281/zenodo.15852621

  4. [4]

    doi: 10.32614/RJ-2021-048

    ISSN 2073-4859. doi: 10.32614/RJ-2021-048. [p1, 2] J. Bergstra, D. Yamins, and D. Cox. Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures. InProceedings of the 30th International Conference on Machine Learning, pages 115–123. PMLR, Feb

    work page doi:10.32614/rj-2021-048 2073

  5. [5]

    Bischl, M

    ISSN 1942-4795. doi: 10.1002/widm.1484. [p1] B. Bischl, G. Casalicchio, T. Das, M. Feurer, S. Fischer, P . Gijsbers, S. Mukherjee, A. C. Müller, L. Németh, L. Oala, L. Purucker, S. Ravi, J. N. van Rijn, P . Singh, J. Vanschoren, J. van der Velde, and M. Wever. OpenML: Insights from 10 years and more than a thousand papers.Patterns, 6(7):101317,

    work page doi:10.1002/widm.1484 1942

  6. [6]

    doi: 10.1016/j.patter.2025.101317. [p15] X. Bouthillier, C. Tsirigotis, F. Corneau-Tremblay, T. Schweizer, L. Dong, P . Delaunay, F. Normandin, M. Bronzi, D. Suhubdy, R. Askari, M. Noukhovitch, C. Xue, S. Ortiz-Gagné, O. Breuleux, A. Bergeron, O. Bilaniuk, S. Bocco, H. Bertrand, G. Alain, D. Serdyuk, P . Henderson, P . Lamblin, and C. Beckham. Orion: Asyn...

    work page doi:10.1016/j.patter.2025.101317 2025

  7. [7]

    URL https: //doi.org/10.5281/zenodo.3478592. [p3] R. Carnell.Lhs: Latin Hypercube Samples,

    work page doi:10.5281/zenodo.3478592

  8. [8]

    R package version 1.2.1

    URL https://doi.org/10.32614/CRAN.package.lhs. R package version 1.2.1. [p9] R. Egelé, I. Guyon, V . Vishwanath, and P . Balaprakash. Asynchronous decentralized bayesian opti- mization for large scale hyperparameter optimization. In2023 IEEE 19th International Conference on E-Science (e-Science), pages 1–10. IEEE,

    work page doi:10.32614/cran.package.lhs

  9. [9]

    [p1, 2, 8, 12] R

    doi: 10.1109/e-Science58273.2023.10254839. [p1, 2, 8, 12] R. Egelé, P . Balaprakash, G. M. Wiggins, and B. Eiffert. DeepHyper: A Python Package for Massively Parallel Hyperparameter Optimization in Machine Learning.Journal of Open Source Software, 10 (109):7975, May

    work page doi:10.1109/e-science58273.2023.10254839 2023

  10. [10]

    doi: 10.21105/joss.07975

    ISSN 2475-9066. doi: 10.21105/joss.07975. [p3] S. Fischer, J. Zobolas, R. Sonabend, M. Becker, M. Lang, M. Binder, L. Schneider, L. Burk, P . Schratz, B. C. Jaeger, S. A. Lauer, L. A. Kapsner, M. Mücke, Z. Wang, D. Pulatov, K. Ganz, H. Funk, L. Harutyunyan, P . Camilleri, P . Kopper, A. Bender, B. Zhou, N. German, L. Koers, A. Nazarova, and B. Bischl. Mlr...

    work page doi:10.21105/joss.07975

  11. [11]

    doi: 10.21105/joss.08331

    ISSN 2475-9066. doi: 10.21105/joss.08331. [p11] R. FitzJohn and R. Ashton.Rrq: Simple Redis Queue,

    work page doi:10.21105/joss.08331

  12. [12]

    package.lgr

    URL https://doi.org/10.32614/CRAN. package.lgr. R package version 0.5.2. [p6] C. Gao. Mirai: Minimalist Async Evaluation Framework for R. Zenodo, May

    work page doi:10.32614/cran

  13. [13]

    [p1, 2, 4] D

    URL https: //doi.org/10.5281/zenodo.7912722. [p1, 2, 4] D. Gelernter. Generative communication in Linda.ACM T ransactions on Programming Languages and Systems (TOPLAS), 7(1):80–112, Jan

    work page doi:10.5281/zenodo.7912722

  14. [14]

    doi: 10.1145/2363.2433

    ISSN 0164-0925. doi: 10.1145/2363.2433. [p1] D. Ginsbourger, R. Le Riche, and L. Carraro. Kriging Is Well-Suited to Parallelize Optimization. In Y. Tenne and C.-K. Goh, editors,Computational Intelligence in Expensive Optimization Problems, pages 131–162. Springer International Publishing, Berlin/Heidelberg,

    work page doi:10.1145/2363.2433

  15. [15]

    [p1, 11] The R Journal Vol

    doi: 10.1007/978-3-642-10701- 6_6. [p1, 11] The R Journal Vol. XX/YY, AAAA 20ZZ ISSN 2073-4859 CONTRIBUTED RESEARCH ARTICLE 18 D. Golovin, B. Solnik, S. Moitra, G. Kochanski, J. Karro, and D. Sculley. Google Vizier: A Service for Black-Box Optimization. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, K...

    work page doi:10.1007/978-3-642-10701- 2073

  16. [16]

    [p1, 3] T

    doi: 10.1145/3097983.3098043. [p1, 3] T. Head, K. Shi, Z. Vinícius, C. Schröder, N. Campos, S. Cereda, T. Fan, K. Shi, J. Schwabedal, M. Pak, F. Callaway, L. Estève, L. Besson, M. Cherti, K. Pfannschmidt, F. Linzberger, C. Cauet, A. Gut, A. Mueller, and A. Fabisch. Scikit-optimize. Zenodo, Mar

    work page doi:10.1145/3097983.3098043

  17. [17]

    URL https://papers.nips.cc/paper_ files/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html. [p11] W. M. Landau.Crew: A Distributed Worker Launcher Framework,

    2017

  18. [18]

    R package version 0.12.0

    URL https://doi.org/ 10.32614/CRAN.package.mlr3oml. R package version 0.12.0. [p11] M. Lang, B. Bischl, and D. Surmann. Batchtools: Tools for R to work on batch systems.The Journal of Open Source Software, 2(10):135, Feb

    work page doi:10.32614/cran.package.mlr3oml

  19. [19]

    doi: 10.21105/joss.00135

    ISSN 2475-9066. doi: 10.21105/joss.00135. [p1, 2] M. Lang, M. Binder, J. Richter, P . Schratz, F. Pfisterer, S. Coors, Q. Au, G. Casalicchio, L. Kotthoff, and B. Bischl. Mlr3: A modern object-oriented machine learning framework in R.Journal of Open Source Software, 4(44):1903, Dec

    work page doi:10.21105/joss.00135 1903

  20. [20]

    doi: 10.21105/joss.01903

    ISSN 2475-9066. doi: 10.21105/joss.01903. [p1] L. Li, K. Jamieson, A. Rostamizadeh, E. Gonina, J. Ben-tzur, M. Hardt, B. Recht, and A. Talwalkar. A System for Massively Parallel Hyperparameter Tuning.Proceedings of Machine Learning and Systems, 2:230–246, Mar

    work page doi:10.21105/joss.01903

  21. [21]

    URL https://proceedings.mlsys.org/paper_files/paper/2020/ hash/a06f20b349c6cf09a6b171c71b88bbfc-Abstract.html. [p1] R. Liaw, E. Liang, R. Nishihara, P . Moritz, J. E. Gonzalez, and I. Stoica. Tune: A Research Platform for Distributed Model Selection and Training, July

    2020

  22. [22]

    Tune: A Research Platform for Distributed Model Selection and Training

    URL https://doi.org/10.48550/arXiv. 1807.05118. Comment: 8 Pages, Presented at the 2018 ICML AutoML workshop. [p3] M. Lindauer, K. Eggensperger, M. Feurer, A. Biedenkapp, D. Deng, C. Benjamins, T. Ruhkopf, R. Sass, and F. Hutter. SMAC3: A Versatile Bayesian Optimization Package for Hyperparame- ter Optimization.Journal of Machine Learning Research, 23(54):1–9,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2018

  23. [23]

    doi: 10.1609/aimag.v7i2.537. [p1] M. Olson, E. Santorella, L. C. Tiao, S. Cakmak, M. Garrard, S. Daulton, Z. J. Lin, S. Ament, B. Beckerman, E. Onofrey, P . Igusti, C. Lara, B. Letham, C. Cardoso, S. S. Shen, A. C. Lin, M. Grange, E. Kashtelyan, D. Eriksson, M. Balandat, and E. Bakshy. Ax: A Platform for Adaptive Experimentation. In Proceedings of the Fou...

    work page doi:10.1609/aimag.v7i2.537

  24. [24]

    [p3] The R Journal Vol

    URL https://github.com/keras-team/keras-tuner. [p3] The R Journal Vol. XX/YY, AAAA 20ZZ ISSN 2073-4859 CONTRIBUTED RESEARCH ARTICLE 19 R Special Interest Group on Databases (R-SIG-DB), H. Wickham, and K. Müller.DBI: R Database Interface,

    2073

  25. [25]

    doi: 10.25080/Majora-7b98e3ed-013. [p3] A. J. Rossini, L. Tierney, and N. Li. Simple Parallel Statistical Computing in R.Journal of Computational and Graphical Statistics, 16(2):399–420,

    work page doi:10.25080/majora-7b98e3ed-013

  26. [26]

    doi: 10.1198/106186007X178979

    ISSN 1061-8600. doi: 10.1198/106186007X178979. [p2] D. Salinas, M. Seeger, A. Klein, V . Perrone, M. Wistuba, and C. Archambeau. Syne Tune: A Library for Large Scale Hyperparameter Tuning and Reproducible Research. InProceedings of the First International Conference on Automated Machine Learning, pages 16/1–23. PMLR, Sept

    work page doi:10.1198/106186007x178979

  27. [27]

    R package version 1.4.0

    URL https://doi.org/10.32614/CRAN.package.scales. R package version 1.4.0. [p10] C. O. Wilke.Cowplot: Streamlined Plot Theme and Plot Annotations for ’Ggplot2’,

    work page doi:10.32614/cran.package.scales

  28. [28]

    Gupta, Vittorio Orlandi, Chia-Rui Chang, Tianyu Wang, Marco Morucci, Pritam Dey, Thomas J

    ISSN 1548-7660. doi: 10.18637/jss. v077.i01. [p8] Marc Becker Ludwig-Maximilians-Universität München Department of Statistics Munich, Germany Munich Center for Machine Learning (MCML) Munich, Germany ORCiD: 0000-0002-8115-0400 marc.becker@stat.uni-muenchen.de Bernd Bischl Ludwig-Maximilians-Universität München Department of Statistics Munich, Germany Muni...

    work page doi:10.18637/jss 2073