Pith. sign in

REVIEW 2 major objections 5 minor 51 references

DaiSy: A Library for Scalable Data Series Similarity Search

T0 review · 2 major / 5 minor · reviewed 2026-07-13 · grok-4.5

Pith's one-line read DaiSy unifies exact data-series and vector similarity search across disk, memory, GPU, and distributed settings in one library.

desk verdict Solid systems library paper: first unified multi-backend exact iSAX search package with real code; novelty is integration, not new algorithms. read the letter →

arxiv 2603.27719 v2 pith:U5DLUQFT submitted 2026-03-29 cs.DB

classification cs.DB
keywords dataseriesexactsimilaritysearchiSAXindexvectorGPUaccelerationdistributedlibrarydesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Exact nearest-neighbor search over huge collections of data series (and high-dimensional vectors) is required by many analytics tasks, yet the best algorithms live in separate codebases, each locked to one hardware setting. This paper presents DaiSy, an open library that folds four leading exact iSAX-based algorithms into a single C++ core with matching Python bindings. The same conceptual pipeline—index build plus exact k-NN search—runs on disk when data do not fit in RAM, entirely in memory on multi-core CPUs, with GPU offload, or across distributed nodes. Because the methods work on raw vectors as well as series, the library also supplies ground-truth answers for approximate vector-search workloads. Users select an algorithm by data size and available hardware; default hyperparameters come from the original papers, and both L2-squared and DTW distances are supported with early-abort lower bounds. The result is a practical, extensible stack that removes the need to re-implement or re-wire specialized systems for each new environment.

What carries the argument

A layered C++ architecture whose core primitives (distance computer with lower bounds, adapter-based data streams, shared iSAX index service) are reused by interchangeable execution models—ParIS+ (disk), MESSI (in-memory), SING (GPU), Odyssey (distributed)—all exposed through the same buildIndex/searchIndex API in C++ and Python.

What would settle it

A head-to-head benchmark, on the same hardware and datasets used in the original papers, that shows either incorrect nearest neighbors or substantially worse wall-clock time for any of the four back-ends relative to the original implementations.

Watch

Extended reading notes

Core claim

DaiSy is the first library that delivers exact similarity search for data series and vectors under a single coherent interface while spanning disk-based, in-memory, GPU-accelerated, and distributed execution, by integrating one state-of-the-art iSAX algorithm for each setting and guaranteeing that every returned answer is exact.

Load-bearing premise

The four chosen algorithms remain the right representatives of their hardware regimes, and the library re-implementation preserves both their exactness guarantees and their published performance without new regressions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper presents DaiSy, an open-source C++/Python library that unifies four previously published exact iSAX-based similarity search algorithms—ParIS+ (disk), MESSI (in-memory), SING (GPU), and Odyssey (distributed)—under a layered architecture with shared distance primitives, an iSAX index service, and interchangeable execution backends. Exactness and SOTA status are inherited by reference to the original papers rather than re-proved. The library also supports L2-squared and DTW, z-normalized and non-normalized data, brute-force baselines, and is positioned as applicable to high-dimensional vector search. The only new empirical result is an in-memory comparison of DaiSy-MESSI against FAISS-IndexFlat on Deep100M and Seismic100M (Figure 3). The contribution is primarily systems/integration: a single coherent API and open artifact spanning multiple execution environments.

Significance. If the artifact faithfully packages the four engines under a usable dual-language interface, DaiSy fills a genuine gap: exact data-series (and vector) search tooling has been fragmented across specialized codebases. Open-sourcing a multi-environment library with C++ and Python bindings is a concrete community contribution that can lower adoption cost for exact search, ground-truth generation for approximate methods, and multi-hardware experimentation. Strengths that should be credited include the public GitHub release, the explicit layered design (core primitives / index service / execution models), dual-language demos and benchmarks, and the clear algorithm-selection decision tree (Figure 2). The work is incremental relative to the four prior systems papers, but the integration claim is real and useful for a systems/library venue.

major comments (2)
  1. [§4.3, Figure 3] §4.3 and Figure 3 provide the only new empirical evidence, and only for the in-memory MESSI backend versus FAISS-IndexFlat. The central multi-environment claim (disk, GPU, distributed) is not accompanied by any timing, scalability, or correctness check for ParIS+, SING, or Odyssey inside DaiSy. For a library paper this need not be a full re-evaluation of every original paper, but at least one controlled result per backend—or an explicit fidelity check against the original codebases—would make the “integrates SOTA without regression” claim load-bearing rather than purely by citation.
  2. [§1, §3.1] §1 and §3.1 assert that all four algorithms “are guaranteed to always return the exact, correct answers” and that DaiSy inherits this property. Exactness is not re-proved and no regression-test methodology or differential check against the original implementations is described. Given that the library re-implements or re-packages these engines under a new adapter and distance layer (§3.1), a short statement of what was reimplemented versus wrapped, plus how exactness is validated in the released tests, is needed to support the correctness claim for users who will treat DaiSy as authoritative ground truth.
minor comments (5)
  1. [Figure 1] Figure 1 is rendered as a dense block of placeholder glyphs in the manuscript text and is effectively unreadable. A clean component diagram is essential for a library architecture paper.
  2. [§2.1] §2.1 positions FAISS, UCR Suite, TSSEARCH, AEON, and TSLEARN appropriately, but a one-sentence note on whether any other exact multi-backend data-series libraries exist (or that none do) would strengthen the “first library” claim.
  3. [§3.3, §4.4] §3.3 notes that algorithms are “more suited to Z-normalized data” because default iSAX breakpoints assume that; the API examples in §4.4 do not show how non-normalized mode is selected. A brief API note would help.
  4. [Title page] Author emails in the header appear truncated (e.g., “francescadelgaudio56 @gmail.com”, “manos.chatzaki @gmail.com”). Fix for the camera-ready version.
  5. [§5] §5 lists planned extensions (Bayesian hyperparameter tuning, subsequence search, streaming, early termination). These are fine as future work; ensure they are not read as current features of the released library.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: library integration claim is self-contained engineering, not a derived prediction forced by its inputs.

full rationale

DaiSy’s central claim is systems/integration work: packaging four previously published exact iSAX engines (ParIS+, MESSI, SING, Odyssey) under one C++/Python API across disk, in-memory, GPU, and distributed backends, with open-source release. Exactness and prior SOTA status are inherited by ordinary citation of peer-reviewed systems papers (§1, §3.1), which is standard for a library paper and does not reduce the existence claim to a fit or a definitional identity. The only new empirical result (Fig. 3, MESSI vs FAISS-IndexFlat on Deep100M/Seismic100M) is an external benchmark comparison, not a quantity fitted then re-presented as prediction. Hyperparameters default to values recommended in the original algorithm papers (§4.2) without being re-labeled as independent forecasts. No self-definitional loop, no fitted-input-as-prediction, no uniqueness theorem smuggled from the authors to forbid alternatives, and no renaming of a known empirical pattern as a first-principles derivation. Self-citation of the group’s prior engines is present but not load-bearing against the stated contribution once the artifact is public; per the analyzer rules this is ordinary systems scholarship, not circular construction. Score 0 with empty steps is the proportionate finding.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

As a systems/library paper the load-bearing content is engineering packaging of previously published algorithms. Free parameters are the algorithm hyper-parameters inherited from those papers; axioms are the domain guarantees of exactness and lower-bound correctness of iSAX-based methods; the principal invented entity is the DaiSy library itself.

free parameters (1)
  • algorithm-specific hyper-parameters (iSAX cardinality, segment count, buffer sizes, etc.) = values recommended in the original algorithm papers
    Defaults are taken from the original publications of ParIS+, MESSI, SING and Odyssey; experienced users may override them. The paper does not re-tune them on new data.
assumptions (3)
  • domain assumption The four selected algorithms (ParIS+, MESSI, SING, Odyssey) always return exact nearest-neighbor answers under the stated distance measures.
    Stated repeatedly in §1 and §3; correctness is inherited rather than re-proved.
  • domain assumption iSAX lower bounds are valid lower bounds on L2 (and the library’s DTW lower bounds are valid), enabling safe pruning.
    Core primitives layer (§3.1) and distance-measures section (§3.2) rely on this standard property of the iSAX family.
  • domain assumption The chosen algorithms remain representative state-of-the-art for their respective execution environments.
    Selection justification in §1 cites earlier experimental surveys; the claim of superiority rests on those surveys.
invented entities (1)
  • DaiSy library (layered architecture with core primitives, index service, and interchangeable execution models) independent evidence
    purpose: Provide a single coherent framework and API that reuses the same index and distance abstractions across disk, memory, GPU and distributed back-ends.
    The architecture and public interfaces are the paper’s primary engineering contribution; independent evidence is the open-source repository itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DaiSy: A Library for Scalable Data Series Similarity Search." pith.science (2026). https://pith.science/paper/U5DLUQFT

@misc{pith2026260327719,
  author       = {Pith},
  title        = {Pith review of: DaiSy: A Library for Scalable Data Series Similarity Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5DLUQFT}},
  note         = {Machine review of arXiv:2603.27719}
}
read the original abstract

Exact similarity search over large collections of data series is a fundamental operation in modern applications, yet existing solutions are often fragmented, specialized, or tailored to specific execution environments. In this paper, we present DaiSy, a unified library for exact data series similarity search that integrates multiple state-of-the-art (iSAX-based) algorithms within a single, coherent framework. DaiSy is the first library to support exact similarity search across diverse execution environments, including implementations for disk-based, in-memory, GPU-accelerated, and distributed scalable similarity search. Although designed for data series, DaiSy is also directly applicable to exact similarity search over vector data, enabling its use in a broader range of applications. The library supports interfaces in both C++ and Python, enabling users to easily integrate its functionality into a variety of tasks. DaiSy is open-sourced and available at: https://github.com/MChatzakis/DaiSy.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith

  1. [1]

    Such domains typically generate large collections of data series that must be analyzed in order to extract meaningful knowledge [21, 43, 3, 38, 20, 34]

    INTRODUCTION Data series constitute one of the most popular data types, appearing across a wide range of applica- tion domains, including finance, astrophysics, neu- roscience, engineering, seismology, and many oth- ers [44, 28]. Such domains typically generate large collections of data series that must be analyzed in order to extract meaningful knowledge...

  2. [2]

    , pn}is a se- quence of points, where eachp i = (ui, ti) represents a valueu i at positiont i

    BACKGROUND Data Series.Adata seriesS={p 1, . . . , pn}is a se- quence of points, where eachp i = (ui, ti) represents a valueu i at positiont i. The valuenis itssize(or dimensionality). iSAX Summary.TheiSAX summary[36] discretizes a data series by dividing the x-axis into equal seg- ments (represented by their mean) and the y-axis into regions based on nor...

  3. [3]

    Foundational primitives provide distance evaluation, data access, and reusable lower bounds, while an index layer encapsulates the iSAX index as an internal service [4]

    SYSTEM DESIGN DaiSy’s architecture follows a layered design cen- tered on a C++ core that implements indexing, dis- tance computation, and search functionality, while exposing the same conceptual model through a Python interface. Foundational primitives provide distance evaluation, data access, and reusable lower bounds, while an index layer encapsulates ...

  4. [4]

    Though, the algo- rithms are more suited to Z-normalized data (we use the default iSAX breakpoints [36], which have been optimized for z-normalized data)

    and non z-normalized data. Though, the algo- rithms are more suited to Z-normalized data (we use the default iSAX breakpoints [36], which have been optimized for z-normalized data)

  5. [5]

    daisy/daisy.hpp

    USING DAISY 4.1. API and Tooling DaiSy exposes a simple API to enable easy use in both C++ and Python. •buildIndex.Initializes and builds the index structure for the selected algorithm. •searchIndex.Performs the top-k index search for the given queries, returning the most simi- lar data series from the index. Furthermore, reproducibility and systematic ev...

  6. [6]

    In particular, we plan to in- corporate additional algorithms for exact similarity search [41, 35], including methods drawn from differ- ent data series summarization families [9]

    FUTURE WORK DaiSy serves as the foundation for several planned future efforts aimed at further improving and ex- tending the library. In particular, we plan to in- corporate additional algorithms for exact similarity search [41, 35], including methods drawn from differ- ent data series summarization families [9]. We also intend to introduce automatic hype...

  7. [7]

    Greece 2.0

    CONCLUSIONS We present DaiSy, a novel library for scalable exact similarity search on large collections of data series, as well as general high-dimensional vectors (e.g., deep embeddings). It supports efficient search algorithms across a wide range of hardware configurations, in- cluding disk-based, in-memory, GPU-accelerated, and distributed environments...

  8. [8]

    M. G. Al-Obeidallah, D. G. Al-Fraihat, A. M. Khasawneh, A. M. Saleh, and H. Addous. Empirical investigation of the impact of the adapter design pattern on software maintain- ability. InInternational Conference on Information Tech- nology, ICIT 2021, Amman, Jordan, July 14-15, 2021, pages 206–211. IEEE, 2021

Show all 51 references
  1. [9]

    Babenko and V

    A. Babenko and V. Lempitsky. Efficient indexing of billion- scale datasets of deep descriptors. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 2055–2063, 2016

  2. [10]

    Boniol and T

    P. Boniol and T. Palpanas. Series2Graph: Graph-based Sub- sequence Anomaly Detection for Time Series.PVLDB, 2020

  3. [11]

    Camerra, T

    A. Camerra, T. Palpanas, J. Shieh, and E. Keogh. isax 2.0: Indexing and mining one billion time series. In2010 IEEE international conference on data mining, pages 58–

  4. [12]

    Chatzakis, P

    M. Chatzakis, P. Fatourou, E. Kosmas, T. Palpanas, and B. Peng. Odyssey: A journey in the land of distributed data series similarity search.Proc. VLDB Endow., 16(5), 2023

  5. [13]

    Chatzakis, Y

    M. Chatzakis, Y. Papakonstantinou, and T. Palpanas. Darth: Declarative recall through early termination for ap- proximate nearest neighbor search.Proceedings of the ACM on Management of Data, 3(4), 2025

  6. [14]

    Daulton, M

    S. Daulton, M. Balandat, and E. Bakshy. Differen- tiable expected hypervolume improvement for parallel multi- objective bayesian optimization.Advances in Neural Infor- mation Processing Systems, 33, 2020

  7. [15]

    Douze, A

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.- E. Mazar´ e, M. Lomeli, L. Hosseini, and H. J´ egou. The faiss library.IEEE Transactions on Big Data, 2025

  8. [16]

    Echihabi, P

    K. Echihabi, P. Fatourou, K. Zoumpatianos, T. Palpanas, and H. Benbrahim. Hercules against data series similarity search.arXiv preprint arXiv:2212.13297, 2022

  9. [17]

    Echihabi and T

    K. Echihabi and T. Palpanas. Scalable analytics on large se- quence collections. In2022 23rd IEEE International Con- ference on Mobile Data Management (MDM). IEEE, 2022

  10. [18]

    Echihabi, T

    K. Echihabi, T. Tsandilas, A. Gogolou, A. Bezerianos, and T. Palpanas. Pros: data series progressive k-nn similarity search and classification with probabilistic quality guaran- tees.VLDBJ, 32(4), 2023

  11. [19]

    Echihabi, K

    K. Echihabi, K. Zoumpatianos, and T. Palpanas. New trends in high-d vector similarity search: ai-driven, progressive, and distributed.PVLDB, 14(12), 2021

  12. [20]

    Echihabi, K

    K. Echihabi, K. Zoumpatianos, T. Palpanas, and H. Ben- brahim. The Lernaean Hydra of Data Series Similarity Search: An Experimental Evaluation of the State of the Art. PVLDB, 2018

  13. [21]

    Echihabi, K

    K. Echihabi, K. Zoumpatianos, T. Palpanas, and H. Ben- brahim. Return of the lernaean hydra: Experimental eval- uation of data series approximate similarity search.Proc. VLDB Endow., 13(3), 2019

  14. [22]

    Fatourou, E

    P. Fatourou, E. Kosmas, T. Palpanas, and G. Paterakis. Fresh: a lock-free data series index. In2023 42nd In- ternational Symposium on Reliable Distributed Systems (SRDS), pages 209–220. IEEE, 2023

  15. [23]

    Fatourou, E

    P. Fatourou, E. Kosmas, T. Palpanas, and G. Paterakis. Fresh: A lock-free data series index. In42nd International Symposium on Reliable Distributed Systems, SRDS 2023, Marrakesh, Morocco, September 25-29, 2023, pages 209–

  16. [24]

    Folgado, M

    D. Folgado, M. Barandas, M. Antunes, M. L. Nunes, H. Liu, Y. Hartmann, T. Schultz, and H. Gamboa. Tssearch: Time series subsequence search library.SoftwareX, 18:101049, 2022

  17. [25]

    I. R. I. for Seismology with Artificial Intelligence. Seismic Data Access.http://ds.iris.edu/data/access/, 2018

  18. [26]

    Gogolou, T

    A. Gogolou, T. Tsandilas, T. Palpanas, and A. Bezerianos. Progressive similarity search on time series data. InBigVis 2019-2nd International Workshop on Big Data Visual Ex- ploration and Analytics, 2019

  19. [27]

    Huijse, P

    P. Huijse, P. A. Estevez, P. Protopapas, J. C. Principe, and P. Zegers. Computational intelligence challenges and ap- plications on large-scale astronomical time series databases. CIM, 2014

  20. [28]

    Kashino, G

    K. Kashino, G. Smith, and H. Murase. Time-series active search for quick retrieval of audio and video. InICASSP, 1999

  21. [29]

    Keogh and C

    E. Keogh and C. A. Ratanamahatana. Exact indexing of dy- namic time warping.Knowledge and information systems, 7(3):358–386, 2005

  22. [30]

    E. J. Keogh, K. Chakrabarti, S. Mehrotra, and M. J. Paz- zani. Locally adaptive dimensionality reduction for indexing large time series databases. In S. Mehrotra and T. K. Sellis, editors,SIGMOD, 2001

  23. [31]

    Kondylakis, N

    H. Kondylakis, N. Dayan, K. Zoumpatianos, and T. Pal- panas. Coconut palm: Static and streaming data series ex- ploration now in your palm. InSIGMOD, 2019

  24. [32]

    Lewis, E

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in Neural Information Pro- cessing Systems, 33:9459–9474, 2020

  25. [33]

    Linardi and T

    M. Linardi and T. Palpanas. Scalable, variable-length simi- larity search in data series: The ulisse approach.Proceedings of the VLDB Endowment, 11(13):2236–2248, 2018

  26. [34]

    Middlehurst, A

    M. Middlehurst, A. Ismail-Fawaz, A. Guillaume, C. Holder, D. Guijo-Rubio, G. Bulatova, L. Tsaprounis, L. Mentel, M. Walter, P. Sch ¨afer, and A. Bagnall. aeon: a python toolkit for learning from time series.Journal of Machine Learning Research, 2024

  27. [35]

    Palpanas

    T. Palpanas. The parallel and distributed future of data series mining. InHPCS, 2017

  28. [36]

    Palpanas

    T. Palpanas. Evolution of a data series index: The isax family of data series indexes: isax, isax2. 0, isax2+, ads, ads+, ads-full, paris, paris+, messi, dpisax, ulisse, coconut- trie/tree, coconut-lsm. InInformation Search, Integration, and Personalization: 13th International...

  29. [37]

    B. Peng, P. Fatourou, and T. Palpanas. Paris: The next destination for fast data series indexing and query answer- ing. In2018 IEEE International Conference on Big Data (Big Data). IEEE, 2018

  30. [38]

    B. Peng, P. Fatourou, and T. Palpanas. Messi: In-memory data series indexing. In2020 IEEE 36th International Con- ference on Data Engineering (ICDE). IEEE, 2020

  31. [39]

    B. Peng, P. Fatourou, and T. Palpanas. Sing: Sequence in- dexing using gpus. In2021 IEEE 37th International Con- ference on Data Engineering (ICDE). IEEE, 2021

  32. [40]

    Rakthanmanon, B

    T. Rakthanmanon, B. Campana, A. Mueen, G. Batista, B. Westover, Q. Zhu, J. Zakaria, and E. Keogh. Search- ing and mining trillions of time series subsequences under dynamic time warping. InSIGKDD, 2012

  33. [41]

    U. Raza, A. Camerra, A. L. Murphy, T. Palpanas, and G. P. Picco. Practical data prediction for real-world wireless sen- sor networks.TKDE, 2015

  34. [42]

    Sch ¨afer, J

    P. Sch ¨afer, J. Brand, U. Leser, B. Peng, and T. Palpanas. Fast and exact similarity search in less than a blink of an eye. In41st IEEE International Conference on Data Engi- neering, ICDE 2025, Hong Kong, May 19-23, 2025, pages 2464–2477. IEEE, 2025

  35. [43]

    Shieh and E

    J. Shieh and E. J. Keogh.isax: indexing and mining ter- abyte sized time series. InProceedings of the 14th ACM SIGKDD International Conference on Knowledge Discov- ery and Data Mining, Las Vegas, Nevada, USA, August 24-27, 2008, pages 623–631. ACM, 2008

  36. [44]

    Tavenard, J

    R. Tavenard, J. Faouzi, G. Vandewiele, F. Divo, G. Androz, C. Holtz, M. Payne, R. Yurchak, M. Rußwurm, K. Kolar, et al. Tslearn, a machine learning toolkit for time series data.Journal of machine learning research, 21(118):1–6, 2020

  37. [45]

    Q. Wang, S. Whitmarsh, V. Navarro, and T. Palpanas. iEDeaL: A Deep Learning Framework for Detecting Highly Imbalanced Interictal Epileptiform Discharges.PVLDB, 16(2), 2023

  38. [46]

    Y. Wang, P. Wang, J. Pei, W. Wang, and S. Huang. A data-adaptive and dynamic segmentation index for whole matching on time series.PVLDB, 6(10), 2013

  39. [47]

    Z. Wang, Q. Wang, P. Wang, T. Palpanas, and W. Wang. Dumpyos: A data-adaptive multi-ary index for scalable data series similarity search.VLDB J., 2024

  40. [48]

    Z. Wang, Q. Wang, P. Wang, T. Palpanas, and W. Wang. Dumpyos: A data-adaptive multi-ary index for scalable data series similarity search.VLDB J., 33(6):1887–1911, 2024

  41. [49]

    D. E. Yagoubi, R. Akbarinia, F. Masseglia, and T. Palpanas. Dpisax: Massively distributed partitioned isax. InIEEE International Conference on Data Mining, ICDM. IEEE Computer Society, 2017

  42. [50]

    Ye and E

    L. Ye and E. Keogh. Time series shapelets: a new primitive for data mining. InSIGKDD. ACM, 2009

  43. [51]

    Zoumpatianos and T

    K. Zoumpatianos and T. Palpanas. Data series management: Fulfilling the need for big sequence analytics. InICDE, 2018

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.