Pith. sign in

REVIEW 2 major objections 3 minor 21 references

FinCacheServe claims that answer reuse, guarded by document versions, evidence fingerprints, tool fingerprints, model identity, and decoding configuration, can cut hosted RAG LLM calls by more than half while recording zero stale outputs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Dependency-checked answer reuse lets FinCacheServe skip about 53% of LLM calls in hosted SEC-derived RAG traces without serving any observed stale outputs.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Credible dependency-cache system with an unresolved gap between dependency freshness and semantic equivalence; worth refereeing, not desk rejecting. the 2 major comments →

arxiv 2607.26076 v1 pith:JQ7NDNIH submitted 2026-07-14 cs.IR cs.AI

FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents

Lingteng Zeng , Yifan Jin This is my paper
classification cs.IR cs.AI
keywords answer cachingretrieval-augmented generationdependency consistencycache invalidationmutable documentsLLM servingfinancial document QAGPU cost reduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that cached answers to repeated analytical questions over mutable corporate documents can be safely served without rerunning the LLM, provided each cached answer is indexed by query intent and guarded by fingerprints of the evidence, tools, and document versions it depends on. On a 2,230-request 7B-model trace it reports a 53.27% skip rate with zero observed dependency-stale outputs; on 32B operator-suite seeds it reports 53.31%, matching and exceeding safer baselines under the same freshness criterion. The reason to care is that an answer-level hit eliminates both prefill and decoding, making it fundamentally cheaper than KV-cache or prefix reuse. The paper separates the correctness gate from capacity management, showing that bounded memory can be handled without weakening the freshness contract.

Core claim

FinCacheServe's central claim is that a generated answer can be treated as a materialized serving object whose reuse is allowed only while the financial intent, source-document versions, evidence fingerprint, tool fingerprint, model identity, and decoding configuration recorded at creation time still match the current request and metadata state. Reuse is admitted through four gate families—signature, evidence, version, and tool/model—and a document-to-answer reverse index invalidates dependent entries when filings update. In the reported workloads, this contract removed more than half of LLM invocations while recording zero dependency-stale outputs, and a 2,376-probe near-collision suite acc

What carries the argument

The dependency-consistent answer-cache contract. Each cache entry stores a normalized request signature (company, period, query family, document scope, tool requirement), the versions of cited documents, an order-stable evidence fingerprint over chunk identifiers/hashes/versions, a tool fingerprint over tool name, version, inputs, and structured output, plus model identity and decoding configuration. The reuse gate admits a cached answer only if every recorded dependency still matches current metadata, and a document-to-answer reverse index propagates filing updates before later reads. This turns each answer into an invalidatable serving object and is what allows the model call to be bypasse

Load-bearing premise

Every semantically important change in the evidence or tooling surface must appear as a change in document version, evidence fingerprint, or tool fingerprint; if a pipeline updates content without altering those signals, the gate cannot see the change and the zero-stale claim no longer holds.

What would settle it

Take a cached answer, modify the underlying evidence chunk text without bumping the document version or chunk hash, and resubmit the same query; if the system serves the cached answer, the gate missed a semantic change. Likewise, change a ratio tool's output format without updating its tool fingerprint and observe whether a stale answer is accepted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Repeated analytical RAG traffic can be served from an answer cache instead of GPU generation, with reported skip rates near 53% on both 7B and 32B hosted workloads.
  • Cache hits complete within tight latency budgets, so dependency-fresh goodput at a 2-second SLO matches the skip rate and rises once generation-path requests fit the budget.
  • Utility-aware admission and eviction under bounded capacity preserves most of the reuse benefit, staying within about 0.6 percentage points of an offline oracle at 64-entry capacity.
  • Field-level ablations and near-collision probes indicate that dropping any guard (period, query family, tool hash, model, decoding config, or evidence) admits unsafe reuse, meaning the contract works as a joint condition.
  • The metadata plane scales to 100k entries and a persistent transactional backend while reporting zero dependency-stale serves in interleaved query/update stress.

Where Pith is reading between the lines

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

  • The contract generalizes beyond finance: any RAG domain with mutable sources and stable query families—legal, medical, code documentation—could use the same signature, evidence, version, tool, and model gate.
  • The zero-stale claim is bounded by the completeness of the signal set; a production pipeline that re-embeds chunk text without bumping the version or chunk hash would be invisible to the gate, so a content-addressed evidence ID would be a natural hardening.
  • The near-collision probes are deliberate constructions; a field study comparing gate decisions against human-labeled answer freshness on real filing diffs would test whether version bumps and fingerprints capture every semantically relevant change.
  • The energy numbers are derived from trace timing and board-power assumptions, so they indicate relative ordering rather than absolute wattage; direct meter readings would be needed to confirm the 44% Wh reduction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. FinCacheServe proposes a serving-layer answer cache for RAG over mutable enterprise documents. Each cached answer is treated as a materialized object keyed by a request signature (company, reporting period, query family, document scope, tool requirement, model identity, decoding configuration) and guarded by source-document versions, evidence fingerprints, and tool fingerprints. On a hit, if all gates pass, the LLM is skipped. The paper evaluates on SEC-derived workloads served with vLLM and Qwen2.5 models, reporting a 53.27% skip rate on a 2,230-request 7B hosted trace and 53.31% across three hosted 32B operator-suite seeds, with zero observed dependency-stale outputs. It also reports capacity-limited cache-management replays close to an offline oracle, a 2,376-probe controlled near-collision suite, metadata-backend and interleaved update stress, and energy sensitivity estimates.

Significance. If the results hold, the paper makes a practical contribution: dependency-consistent answer reuse can remove a large fraction of hosted LLM calls for repeated analytical RAG traffic while maintaining a precise freshness contract for mutable evidence. The paper's strengths are the breadth of controlled experiments: hosted 7B and 32B traces with confidence intervals, field-level gate ablations, a sizable adversarial probe suite, capacity/update/policy-history replays, transactional metadata validation, and a human audit. The authors are also transparent that energy numbers are estimates and that dependency freshness is not equivalent to factual correctness. The main weakness is that the headline reuse contract identifies "intent" with a coarse query-family bucket, and the paper does not demonstrate that within-bucket requests are semantically equivalent; this gap directly affects the meaning of the central skip-rate claim.

major comments (2)
  1. [§4.1, §4.3, Table 10, §7.12] The reuse contract in Eq. (2) uses s(r) ≃ s(c), where s contains query family but not the normalized question. The evidence gate (Eq. 4) only requires the same evidence fingerprint. Consequently, two distinct questions in the same family with the same company/period/scope/tool/model/decode and overlapping retrieved evidence can be admitted even though one answer is not a response to the other. The paper's own consistency definition (Section 3) explicitly excludes factual correctness, so "zero dependency-stale outputs" does not establish that the served output is responsive. The 2,376-probe suite (Table 10) tests wrong query family, wrong period, wrong scope, and changed tool/model, but it does not include within-family question substitution. The normalizer audit (Table 12) checks metadata-field consistency, not semantic equivalence of the questions. The abstract and §4.1 promise reuse fo
  2. [Data availability / §7.4 / §7.10] The central empirical results depend on non-public hosted traces and custom code. The capacity replay, interleaved metadata stress, backend validation, and SLO/energy replays are all computed from "archived traces" that are not released; the Data and Code availability statements only say a package "may be prepared for a future venue or archival release." Without the traces, the gate implementation, or the policy code, the exact skip rates and stale counts in Tables 3, 5, and 6 cannot be independently verified. For a systems paper whose claims are empirical measurements, this is a reproducibility gap that should be addressed before acceptance, either by releasing the artifacts or by providing a sufficiently detailed and executable specification of the workload reconstruction and replay procedures.
minor comments (3)
  1. [Section 9] The validity-scope discussion is helpful, but it should explicitly give a failure-mode example: if an extraction pipeline refreshes chunk text without bumping the document version, or a tool changes its output format without updating its fingerprint, the gate cannot observe the change. The current wording implies the guarantee is relative to recorded dependencies, but adding this concrete caveat would prevent over-reading of "zero observed dependency-stale outputs."
  2. [§6.4] Please define "provider-call closure" at first use. It appears in the metrics list but is not explained until the hosted-accounting discussion, and the exact set of calls included (LLM only, or also retrieval/tool calls) should be stated explicitly.
  3. [Throughout] Minor typographical issues: "F AISS" appears in Table 2 and §5; several words use unusual ligatures (e.g., "Efficient", "insufficient", "traffic"); Figure 2 uses "FCS" without defining the abbreviation. These should be cleaned up.

Circularity Check

1 steps flagged

Zero dependency-stale claim is the gate's own contract by construction; headline skip rates are empirical and externally benchmarked.

specific steps
  1. self definitional [§3 (Consistency target, Eqs. 2–6) and §4.3 Reuse gate; abstract claim of zero dependency-stale outputs]
    "A served output is dependency-stale when the answer or prompt uses a cached dependency whose recorded source-document version, evidence fingerprint, or tool fingerprint conflicts with the current metadata state for the request. A cached answer can be served at τ when all predicates below hold: s(r) ≃ s(c), (2) ∀d ∈ Dc, vτ(d) = V (c, d), (3) hE(r) = hE(c), (4) hT (r) = hT (c), (5) m = m(c), θ = θ(c). (6)"

    The headline safety result 'zero observed dependency-stale outputs' is entailed by the reuse gate's own admission condition. Dependency-stale is defined as a served output whose recorded source-document versions, evidence fingerprint, or tool fingerprint conflicts with the current metadata state (§3), while the gate admits an answer only when s(r) ≃ s(c), every recorded version equals the current version, and hE, hT match (Eqs. 2–6). A FinCacheServe-served answer therefore cannot be dependency-stale by construction (absent metadata-plane races); the 0-stale count restates the gate contract rather than testing a freshness prediction. The paper discloses this ('It evaluates consistency with recorded dependencies; factual answer accuracy is evaluated separately'), so the circularity is mild a

full rationale

Central claims are empirical, not derived. The 53.27% (1,188/2,230) 7B and 53.31% (290/544) 32B skip rates are measured counts from hosted vLLM runs, compared under one dependency-stale criterion against reimplemented external baselines (versioned semantic 38.97%, grounded-style 22.43%, entity-period 10.48%, exact+doc-invalidation 2.57%); the skip-rate arithmetic stands alone. There is no load-bearing self-citation: no reference overlaps the authors' names (Zeng/Jin), and no 'uniqueness theorem' or prior-work ansatz is imported — the design is assembled from cited external systems and benchmarks (vLLM, SGLang, FAISS, Qwen2.5, FinQA, FinanceBench, HTTP caching, materialized views, provenance). The one self-definitional element is the consistency metric: dependency-stale is defined as a conflict with the recorded versions/evidence/tool fingerprints that the reuse gate (Eqs. 2–6) enforces before serving, so 'zero dependency-stale' is largely the gate contract restated; the paper discloses this in §3 and §9 ('The consistency claim is dependency freshness with respect to recorded source versions, evidence fingerprints, and tool fingerprints'). The metric retains discriminating power (unsafe controls register 46–134 stale serves), which is why the score stays low. The capacity-replay CacheOpt utility is explicitly labeled an upper-bound target using trace-wide statistics ('to establish an upper-bound policy target short of Belady'), not a deployable prediction, and its λ weights are not shown to be fitted to the headline traces; the online variant is history-based. The skeptical concern that same-family distinct questions can collide (current vs. quick ratio under identical evidence) is a completeness/validity threat about the guarded signal set, explicitly scoped out by the paper's own consistency definition, and is a correctness risk rather than circularity. Overall: one mild, disclosed self-definitional element; central performance claims self-contained. Score 2.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

Everything rests on completeness of dependency signals, representativeness of SEC-derived traces, and absence of fingerprint/version collisions. The only explicit chosen hyperparameters are the CacheOpt utility weights and board-power assumptions for energy; the headlined skip-rate numbers do not depend on them.

free parameters (2)
  • CacheOpt utility weights λ_m, λ_f, λ_u = not reported
    Equation (7) uses these coefficients to balance entry size, document fan-out, and update hazard; the admission/eviction results and oracle gap depend on them, but no values or tuning procedure are reported.
  • board-power assumption = 250/350/450/600 W
    The Wh comparisons in §7.7 and the abstract's 44.30% figure are linear in the assumed board power; the authors explicitly characterize these as estimates, not meter readings.
axioms (5)
  • domain assumption All dependency-relevant changes to source material are observed as document-version, evidence-fingerprint, or tool-fingerprint changes.
    The consistency contract is only as strong as these signals; §4.2 defines fingerprints over chunk IDs/hashes/versions and tool name/version/inputs/payload, but the paper does not prove that every semantic change forces a signal change.
  • domain assumption The request normalizer's query-family and financial-identity labels are mostly accurate; ambiguous rows are conservatively routed to generation.
    §7.12 audit shows 78.9% signature-safe and 87.7% deterministic query-family consistency; if labels are frequently wrong, reuse opportunity shrinks but not necessarily freshness.
  • domain assumption Metadata-plane linearization orders all concurrent queries and updates.
    §4.4 and §5 use a service-level lock around version-store updates and entry mutation; the stress tests measure the implementation but do not prove safety under all possible schedules.
  • domain assumption SEC-derived hosted traces are representative of mutable enterprise RAG serving.
    §9 says production traffic can differ in diversity, update cadence, access skew, tool composition, SLO distribution, and capacity; the headlined skip rates are workload-specific.
  • standard math Order-stable hashes and fingerprints are collision-free for the workload.
    The gate compares hashes; a hash collision or a fingerprint that omits a semantically relevant field would admit a stale answer. The paper does not provide cryptographic collision analysis.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents." pith.science (2026). https://pith.science/paper/JQ7NDNIH

@misc{pith2026260726076,
  author       = {Pith},
  title        = {Pith review of: FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JQ7NDNIH}},
  note         = {Machine review of arXiv:2607.26076}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Retrieval-augmented generation services over mutable enterprise documents repeatedly execute semantically equivalent analysis requests. Answer reuse can remove GPU-bound generation work, yet response caches require dependency consistency when filings, evidence chunks, and tool outputs change. FinCacheServe treats each generated answer as a serving object indexed by enterprise intent and guarded by document versions, evidence fingerprints, tool fingerprints, model identity, and decoding configuration. A vLLM implementation evaluates SEC-derived financial-document workloads with Qwen2.5 models. On a 2,230-request hosted 7B trace, FinCacheServe skips 53.27% of LLM calls with zero observed dependency-stale outputs. Across three hosted 32B operator-suite seeds, it skips 53.31% of 544 requests, compared with 38.97% for versioned semantic caching and 22.43% for grounded-style reuse. Capacity, backend, and SLO replays show oracle-bounded cache management, 100k-entry transactional metadata behavior, and 44.30% lower estimated Wh per dependency-fresh 2s-SLO success than versioned semantic caching.

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

21 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    In: Proceedings of the 29th ACM Symposium on Operating Systems Princi- ples, pp

    Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C.H., Gonzalez, J.E., Zhang, H., Stoica, I.: Efficient memory management for large language model serving with Page- dAttention. In: Proceedings of the 29th ACM Symposium on Operating Systems Princi- ples, pp. 611–626 (2023). https://doi.org/10. 1145/3600006.3613165

    arXiv 2023

  2. [2]

    https:// doi.org/10.48550/arXiv.2312.07104

    Zheng, L., Yin, L., Xie, Z., Sun, C., Huang, J., Yu, C.H., Cao, S., Kozyrakis, C., Sto- ica, I., Gonzalez, J.E., Barrett, C., Sheng, Y.: SGLang: Efficient Execution of Structured Language Model Programs (2023). https:// doi.org/10.48550/arXiv.2312.07104

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.07104 2023

  3. [3]

    https://doi

    Jin, C., Zhang, Z., Jiang, X., Liu, F., Liu, X., Liu, X., Jin, X.: RAGCache: Efficient Knowledge Caching for Retrieval- Augmented Generation (2024). https://doi. org/10.48550/arXiv.2404.12457

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.12457 2024

  4. [4]

    https://doi.org/10.48550/arXiv.2405.16444

    Yao, J., Li, H., Liu, Y., Ray, S., Cheng, Y., Zhang, Q., Du, K., Lu, S., Jiang, J.: CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion (2024). https://doi.org/10.48550/arXiv.2405.16444

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.16444 2024

  5. [5]

    https://doi.org/10.48550/arXiv.2511.03475

    Jiang, Y., Huang, Y., Cheng, L., Deng, C., Sun, X., Mai, L.: ContextPilot: Fast Long- Context Inference via Context Reuse (2025). https://doi.org/10.48550/arXiv.2511.03475

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.03475 2025

  6. [6]

    Accepted at SOSP 2025 BigMem Workshop (2025)

    Feng, S., Li, H., Du, K., Gu, Z., Liu, Y., Yao, J., Ray, S., Shen, S., Cheng, Y., Anantha- narayanan, G., Jiang, J.: AdaptCache: KV Cache Native Storage Hierarchy for Low- Delay and High-Quality Language Model Serving. Accepted at SOSP 2025 BigMem Workshop (2025). https://doi.org/10.48550/ arXiv.2509.00105

    work page doi:10.48550/arxiv.2509.00105 2025

  7. [7]

    IEEE Transactions on Big Data 7(3), 535–547 (2021) https://doi.org/10.1109/TBDATA

    Johnson, J., Douze, M., Jégou, H.: Billion- scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535–547 (2021) https://doi.org/10.1109/TBDATA. 2019.2921572

    arXiv 2021

  8. [8]

    https://doi.org/10.48550/ arXiv.2412.15115 15

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., et al.: Qwen2.5 Techni- cal Report (2024). https://doi.org/10.48550/ arXiv.2412.15115 15

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2024

  9. [9]

    In: Proceedings of the 2021 Conference on Empirical Meth- ods in Natural Language Processing, pp

    Chen, Z., Chen, W., Smiley, C., Shah, S., Borova, I., Langdon, D., Moussa, R., Beane, M., Huang, T.-H., Routledge, B., Wang, W.Y.: FinQA: A dataset of numerical rea- soning over financial data. In: Proceedings of the 2021 Conference on Empirical Meth- ods in Natural Language Processing, pp. 3697–3711 (2021). https://doi.org/10.18653/ v1/2021.emnlp-main.300

    2021

  10. [10]

    https://doi.org/10.48550/ arXiv.2311.11944

    Islam, P., Kannappan, A., Kiela, D., Qian, R., Scherrer, N., Vidgen, B.: FinanceBench: A New Benchmark for Financial Question Answering (2023). https://doi.org/10.48550/ arXiv.2311.11944

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.11944 2023

  11. [11]

    Securities and Exchange Commission: EDGAR Application Programming Interfaces (APIs)

    U.S. Securities and Exchange Commission: EDGAR Application Programming Interfaces (APIs). https://www.sec.gov/search-filings/ edgar-application-programming-interfaces . Last reviewed or updated April 8, 2025; accessed June 29, 2026 (2025)

    2025

  12. [12]

    Accepted to IEEE INFOCOM 2026 (2025)

    Liu, X., Atalar, B., Dai, X., Zuo, J., Wang, S., Lui, J.C.S., Chen, W., Joe-Wong, C.: Seman- tic Caching for Low-Cost LLM Serving: From Offline Learning to Online Adaptation. Accepted to IEEE INFOCOM 2026 (2025). https://doi.org/10.48550/arXiv.2508.07675

    doi:10.48550/arxiv.2508.07675 2026

  13. [13]

    https://doi.org/10.48550/ arXiv.2602.13165

    Singh, A.K., Wang, H., Attaluri, L.N.S., Chiam, T., Zhu, W.: Asynchronous Verified Semantic Caching for Tiered LLM Archi- tectures (2026). https://doi.org/10.48550/ arXiv.2602.13165

    doi:10.48550/arxiv.2602.13165 2026

  14. [14]

    Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer?

    Shah, S.H.: Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer? (2026). https: //doi.org/10.48550/arXiv.2605.27494

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.27494 2026

  15. [15]

    Future Generation Computer Systems 170, 107822 (2025) https: //doi.org/10.1016/j.future.2025.107822

    Haqiq, K., Jahan, M.V., Farimani, S.A., Masoom, S.M.F.: MinCache: A hybrid cache system for efficient chatbots with hierarchical embedding matching and LLM. Future Generation Computer Systems 170, 107822 (2025) https: //doi.org/10.1016/j.future.2025.107822

    arXiv 2025

  16. [16]

    RFC 9111 (2022)

    Nottingham, M., Fielding, R.T., Reschke, J.: HTTP Caching. RFC 9111 (2022). https:// doi.org/10.17487/RFC9111 . https://www. rfc-editor.org/rfc/rfc9111

    work page doi:10.17487/rfc9111 2022

  17. [17]

    IEEE Data Engineering Bulletin 18(2), 3–18 (1995)

    Gupta, A., Mumick, I.S.: Maintenance of materialized views: Problems, techniques, and applications. IEEE Data Engineering Bulletin 18(2), 3–18 (1995)

    1995

  18. [18]

    In: Proceedings of the 2001 ACM SIGMOD International Conference on Management of Data, pp

    Mistry, H., Roy, P., Sudarshan, S., Ramam- ritham, K.: Materialized view selection and maintenance using multi-query optimization. In: Proceedings of the 2001 ACM SIGMOD International Conference on Management of Data, pp. 307–318 (2001). https://doi.org/ 10.1145/375663.375706

    arXiv 2001

  19. [19]

    In: Database Theory – ICDT 2001, pp

    Buneman, P., Khanna, S., Tan, W.-C.: Why and where: A characterization of data prove- nance. In: Database Theory – ICDT 2001, pp. 316–330 (2001). https://doi.org/10.1007/ 3-540-44503-X_20

    2001

  20. [20]

    In: Proceedings of Twenty- First ACM SIGOPS Symposium on Operat- ing Systems Principles, pp

    DeCandia, G., Hastorun, D., Jampani, M., Kakulapati, G., Lakshman, A., Pilchin, A., Sivasubramanian, S., Vosshall, P., Vogels, W.: Dynamo: Amazon’s highly available key-value store. In: Proceedings of Twenty- First ACM SIGOPS Symposium on Operat- ing Systems Principles, pp. 205–220 (2007). https://doi.org/10.1145/1294261.1294281

    arXiv 2007

  21. [21]

    In: Proceedings of the VLDB Endow- ment, vol

    Cooper, B.F., Ramakrishnan, R., Srivastava, U., Silberstein, A., Bohannon, P., Jacobsen, H.-A., Puz, N., Weaver, D., Yerneni, R.: PNUTS: Yahoo!’s hosted data serving plat- form. In: Proceedings of the VLDB Endow- ment, vol. 1, pp. 1277–1288 (2008). https: //doi.org/10.14778/1454159.1454167 16

    arXiv 2008

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.