Pith. sign in

REVIEW 3 major objections 6 minor 67 references

AutoSLO: Practical Latency SLOs on Cloud Data Warehouses -- Extended Version

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read AutoSLO meets query latency targets on multi-cluster cloud warehouses while cutting cost about a quarter versus the next-best option.

desk verdict Solid multi-timescale systems paper that actually takes explicit latency SLOs as inputs; the 26.4% cost claim is real but rests on a thin end-to-end sample. read the letter →

arxiv 2607.11770 v1 pith:W6FFFEJZ submitted 2026-07-13 cs.DB

classification cs.DB
keywords clouddatawarehouseslatencySLOsmulti-clustermanagementqueryroutingautoscalingworkloadforecastingconcurrency-awareprediction
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

Cloud data warehouses let organizations run many compute clusters over the same data, but today's systems still force administrators to guess cluster sizes and routing rules rather than state the latency each query must meet. AutoSLO treats those latency targets as first-class inputs and manages clusters across three timescales: daily planning from workload history, minute-scale reactive spin-up when the forecast is wrong, and per-query routing that accounts for concurrent interference. On realistic Redbench workloads the system holds a 10% violation target and spends a mean of 26.4% less than the best competing baseline in each scenario. Component tests show the router and autoscaler each cut violation rates sharply relative to simpler alternatives, and a single day of history already yields most of the planning benefit. The practical claim is that explicit SLOs plus multi-timescale control can replace static isolation and qualitative sliders with cheaper, target-aware operation.

What carries the argument

The Query Router's lexicographic placement rule: for each arriving query it uses Iconq+ predictions to choose the active cluster that minimizes the increase in projected SLO violations, then the increase in billed cost, thereby trading interference risk against resource cost in real time.

What would settle it

Rerun the four Redbench day/SLO scenarios with the same Policy Tuner configurations but replace Iconq+ with a concurrency-unaware predictor; if the measured violation rate then exceeds the 0.1 target or the cost advantage over RAIS-Balanced disappears, the central claim fails.

Watch

Extended reading notes

Core claim

A three-timescale controller—proactive Policy Tuner, reactive Autoscaler, and concurrency-aware Query Router—can meet specified query latency SLOs on multi-cluster cloud data warehouses while reducing infrastructure cost by a mean of 26.4% versus the per-scenario next-best baseline on realistic Redbench workloads.

Load-bearing premise

That the concurrency-aware latency model is accurate enough under real mixes that routing and short-horizon simulation decisions based on its predictions actually improve measured SLO outcomes, not merely predicted ones.

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

3 major / 6 minor

Summary. The paper presents AutoSLO, a three-timescale framework for meeting query-level latency SLOs cost-efficiently on multi-cluster cloud data warehouses (e.g., Redshift Serverless). It combines a periodic Policy Tuner that plans proactive spinups and tunes Autoscaler knobs from history-derived forecasts and batch simulation; an SLO-aware Autoscaler that triggers and sizes reactive spinups via short-horizon simulation; and an online Query Router that places each query using Iconq+, a concurrency- and size-aware latency predictor, under a lexicographic BestWithTarget objective (meet violation-rate target τ first, then minimize cost). On four Redbench scenarios (two Redset-mapped days × two SLO strictness levels κ), AutoSLO meets τ=0.1 and reports a mean 26.4% cost reduction versus the per-scenario next-best baseline (49.6% versus RAIS-Balanced). Component ablations on synthetic BaseWorkload report large relative violation-rate reductions for the Router, Spinup Size Selector, Trigger, and Policy Tuner, with efficiency microbenchmarks matching each component’s timescale.

Significance. If the end-to-end result holds under broader evaluation, this is a practically important contribution: cloud warehouses already expose multi-cluster isolation, but administrators still cannot specify latency SLOs directly, and prior systems (RAIS, Auto-WLM, BRAD, ResTune, WiseDB, etc.) lack the full plan/adjust/react stack with concurrency-aware placement. The problem formulation (Definitions 1–3, BestWithTarget), the three-timescale architecture, and the carefully specified algorithms (Route, MaybeSpinup, FindBestSpinupSize, ScheduleSpinups, FindGoodSpinupTime) with complexity bounds are clear and reusable. Strengths include open code, live Redshift Serverless measurements of client-side latency and billing, and systematic component ablations. The work is a strong systems paper candidate for a top data-management venue, contingent on more robust evidence that the headline cost/SLO tradeoff is not an artifact of four compressed single runs and of Iconq+ accuracy on the evaluation mix.

major comments (3)
  1. Section 8.2 and Figure 6: The central claim (meets τ=0.1 with mean 26.4% cost reduction vs next-best baseline) rests on four single executions of one Redbench-mapped SELECT workload (provisioned cluster 157), with interarrival times compressed 6×. There are no multi-run variance estimates, error bars, hold-out days beyond the two Mondays, or alternate Redset clusters. For a load-bearing empirical claim, the manuscript needs either additional days/clusters, repeated runs with reported variance, or a clear sensitivity analysis of the compression factor and history window. Without that, the 26.4%/49.6% numbers are not yet shown to be robust to ordinary day-to-day or run-to-run variation.
  2. Sections 4–6 and 8.3: Iconq+ drives both Route (Algorithm 1) and FindBestSpinupSize (Algorithm 3). Section 8.3 reports Q-error percentiles only on BaseWorkload folds (and inference latency under a stress scenario), not calibration of predicted versus realized SLO violations under the Redbench-mapped mix used for the headline result. If Iconq+ systematically under- or over-predicts contention on those templates, the lexicographic argmin and the short-horizon spinup simulations can select configurations that look good under predicted latencies yet fail to deliver the reported real-system VR/cost tradeoff. A calibration plot or predicted-vs-realized violation-rate comparison on Redbench (or a hold-out production-like mix) is needed to support transfer of the component gains to the end-to-end claim.
  3. Tables 2–4 versus Section 8.2: Component VR reductions (Router 47.8%, Spinup Size Selector 93.7%, Trigger 54.6%) are measured on synthetic BaseWorkload Poisson mixes with forced or controlled scaling, while the end-to-end claim uses Redbench. The paper does not show that the same relative gains hold when the Router and Autoscaler operate on the production-like arrival process and template mix. At least one component ablation (e.g., Router or Spinup Size Selector) on a Redbench day would close this gap between the large relative numbers and the headline scenarios.
minor comments (6)
  1. Section 8.2 footnote and RAIS baselines: The observation that RAIS bills the full 128 RPU regardless of slider setting is important for interpreting cost; it should be elevated into the main text and, if possible, confirmed with the vendor or additional slider settings so readers can judge whether the comparison is fair.
  2. Figure 6: Axes and units (cost in dollars, VR on [0,1]) are readable, but adding the numeric VR/cost values next to each marker (or a small table) would make the 26.4% mean easier to verify without visual estimation.
  3. Section 8.1.3: SLO construction as κ·ℓ_baseline(t) from a closed-loop 16-RPU run is reasonable but should state whether baseline latencies include cold-start effects and how template-level max latency interacts with multi-query interference when κ is small.
  4. Table 1 and Section 9: Related-work coverage is thorough; a short note on how AutoSLO would compose with engine selection (BRAD) would help position multi-cluster same-engine routing versus multi-engine blueprints.
  5. Notation: BestWithTarget is defined in Section 2 and reused in Algorithms 3–4; a brief restatement near Algorithm 3 would help readers who jump to the Autoscaler.
  6. Typos/clarity: Abstract and intro use “AutoSLO” consistently; ensure “Iconq+” vs “Iconq” is uniform in figure captions (Figure 7). Section 8.2 has a duplicated subsection number (“8.2 End-to-end SLO Adherence: Findings” after 8.2).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical systems paper; cost and SLO violation rates are measured on live Redshift executions, not derived from fitted constants that redefine the target.

full rationale

AutoSLO’s load-bearing claims are end-to-end and component measurements of client-side latency L(q) versus administrator-supplied SLOs and of vendor billing timers (Definition 3, BestWithTarget, Sections 8.2–8.7). Iconq+ is trained on separate BaseWorkload runs (Section 8.3) and then used only as a decision aid inside Route and FindBestSpinupSize; evaluation metrics remain observed outcomes, not model outputs. Policy Tuner forecasts and Batch Simulator configurations are optimized offline and then executed on real clusters; reported VR/cost are from those executions, not from the simulation objective. Self-citations of Iconq, Stage, BRAD, and RAIS supply background tooling and baselines; none is a uniqueness theorem or ansatz that forces the 26.4% cost claim. No equation equates a fitted parameter with a claimed prediction of the same quantity. The derivation chain is therefore self-contained against external benchmarks (Redbench/Redset-mapped TPC-DS on Amazon Redshift Serverless).

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central empirical claim rests on domain models of cloud warehouse billing and spinup delay, synthetic SLO construction from isolated baselines, history-based workload recurrence for planning, and the adequacy of Iconq+ as a decision oracle. Free parameters are mostly operational knobs and evaluation targets rather than physics-style fitted constants; invented entities are system components and the Iconq+ model variant.

free parameters (5)
  • τ_target (SLO violation rate target) = 0.1 in main experiments
    Lexicographic objective BestWithTarget uses τ (default 0.1 in end-to-end evals); choice directly shapes which policy wins.
  • κ (SLO strictness multiplier) = 2 or 4
    SLO(q)=κ·ℓ_baseline(template); κ∈{2,4} defines difficulty and is chosen by experimenters, not measured from apps.
  • Autoscaler knobs (τ_trigger, θ, μ, w, T_idle, T_min_lifetime) = tuned per day/history via Policy Tuner
    Trigger and spinup-size behavior depend on these; Policy Tuner searches them but they remain free configuration parameters of the method.
  • Forecast decay λ and history window = λ∈(0,1); evals use 0–30 days
    Workload Forecaster weights past same-hour-of-week bins by λ^{k-1}; λ and history length (1/7/30 days) affect planned spinups.
  • Interarrival compression factor (Redbench) =
    Day-long traces compressed 6× so a day runs in 4 hours; changes concurrency dynamics relative to real time.
assumptions (6)
  • domain assumption Usage-based billing with idle pause after m seconds (often m=60) and constant price × size × active time.
    Section 2 cost model; routing and spinup decisions optimize against this timer model.
  • domain assumption New clusters become usable only after spinup delay δ (and teardown after ζ); preparation backlog must be simulated.
    Definitions 2 and Algorithm 3; without non-zero δ, pure reactive scaling would suffice.
  • domain assumption Enterprise warehouse workloads have enough recurring structure that history-derived forecasts support useful proactive spinups.
    Section 1 and Policy Tuner design; Table 5 shows benefit from 1 day of history under this assumption.
  • domain assumption Redbench join-matching of Redset cluster 157 onto TPC-DS SF1000 yields a realistic executable proxy for production SELECT load.
    Section 8.2 end-to-end setup; central cost claim is measured only on this proxy.
  • ad hoc to paper Client-side latency L(q) vs application-provided SLO(q) is the right success criterion; τ is soft (lexicographic), not a hard constraint.
    Definition 3 BestWithTarget; comparison policy is paper-defined rather than vendor SLA standard.
  • ad hoc to paper Standard event-driven simulation with predicted completions approximates short-term autoscaling outcomes well enough to choose spinup size.
    Workload Simulator / Batch Simulator used in Algorithms 3–4.
invented entities (3)
  • AutoSLO three-component framework (Policy Tuner, Autoscaler, Query Router)
    purpose: Jointly plan, adjust, and react for SLO-aware multi-cluster management.
    Core system architecture of the paper; evaluated as a whole and per component.
  • Iconq+ latency predictor
    purpose: Extend Iconq with cluster-size features, censored training, and incremental inference for routing and simulation.
    Section 4; decision quality of Router and Autoscaler depends on this model variant.
  • BestWithTarget lexicographic objective
    purpose: Rank policies by excess violation rate above τ then by cost.
    Definition 3; defines what 'best' means throughout selection and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AutoSLO: Practical Latency SLOs on Cloud Data Warehouses -- Extended Version." pith.science (2026). https://pith.science/paper/W6FFFEJZ

@misc{pith2026260711770,
  author       = {Pith},
  title        = {Pith review of: AutoSLO: Practical Latency SLOs on Cloud Data Warehouses -- Extended Version},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W6FFFEJZ}},
  note         = {Machine review of arXiv:2607.11770}
}
read the original abstract

Modern cloud data warehouses decouple compute from storage, making it easy for organizations to access the same underlying data with multiple compute clusters. This flexibility is often used for performance isolation among diverse workloads, so that each workload meets its latency service-level objective (SLO) more reliably. For example, interactive dashboards, ad hoc analysis, and batch jobs can each run on separate clusters. However, this dedicated-cluster approach requires each compute cluster to be continuously scaled to adapt to workload evolution, with over-provisioning wasting resources and under-provisioning risking SLO violations. We present AutoSLO, a latency-SLO-aware workload management framework for multi-cluster cloud data warehouses. AutoSLO operates across three timescales through three key components. First, a periodic Policy Tuner plans proactive cluster scaling actions and tunes configuration parameters, using simulations of history-derived workload forecasts. Second, an SLO-aware reactive Autoscaler adjusts the active cluster set when recent workload behavior deviates from the forecast. Third, an online Query Router reacts to live load when placing each query, using a concurrency-aware latency predictor to avoid SLO violations. On realistic Redbench workloads, AutoSLO successfully meets latency SLOs of varying strictness, reducing cost by a mean of 26.4% compared to the per-scenario next-best baseline. Component-level evaluations show that the Query Router and Autoscaler respectively reduce SLO violation rates by a mean of 47.8% and 93.7%, relative to their corresponding alternatives. Finally, we show that the Policy Tuner can reduce the SLO violation rate by a mean of 44.6% using a single day of workload history, and that each component is efficient given its intended operating timescale.

Figures

Figures reproduced from arXiv: 2607.11770 by the authors.

Figure 1
Figure 1. The architecture of AutoSLO. Solid lines indicate input/output flow, while dashed lines indicate control flow. above 𝜏. In particular, we compare policies with SLO violation rates 𝑉 = {𝑣 P1 , . . . , 𝑣 P𝑛 } and costs 𝐾 = {𝑘 P1 , . . . , 𝑘 P𝑛 } using: BestWithTarget(𝑉 , 𝐾, 𝜏) = lex min (𝑣,𝑘) ∈ {𝑉 ,𝐾} (max(0, 𝑣 − 𝜏), 𝑘) Problem Definition. We want to find the best policy under the lexicographic target-aware objective … view at source ↗
Figure 2
Figure 2. When predicting the latency of 𝑞2, Iconq+ ingests the interaction feature vectors of later neighbors (𝑞3 and 𝑞4) in chronological order, enabling incremental inference. 4 LATENCY PREDICTOR We start with Iconq [58], a recent LSTM-based concurrent query latency estimation model, but modify it to address the needs of AutoSLO. In particular, we need support for clusters of different sizes (in the Query Router), while we… view at source ↗
Figure 3
Figure 3. The Query Router balances SLOs and cost. The x￾axis is time. Each query is a rectangle and same-color dashed lines show when it will miss its SLO. Hasing indicates pre￾dicted latencies; a red outline marks an SLO miss. Thin blue lines show billed time; dark blue indicates the minimum. fact that later queries generally cannot make an earlier query faster, and cannot be smaller than its observed latency so far. For ea… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The Spinup Size Selector evaluates each scaling action at 𝑡𝑑 by simulating copies of the arrival window W𝑎 . It stops once 𝜇 queries (here 𝜇 = 3) issued after the new cluster is available (𝑡𝑑 +𝛿; queries outlined in black) finish executing. re-evaluation for 𝛿 seconds …
Figure 5
Figure 5. Figure 5: Candidate spinup time determination. Here, queries [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: End-to-end performance of AutoSLO. middle Monday (April 15, 2024) of the workload, to ensure non￾overlapping recent workload histories in the Policy Tuner. Then, we examine two different levels of SLO difficulty, for 𝜅 ∈ {2, 4}. Across scenarios, our SLO violation rate…
Figure 7
Figure 7. Figure 7: Iconq+ delivers superior accuracy and efficiency. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Efficiency evaluation of key AutoSLO algorithms. a single day of workload history to reduce the SLO violation rate by a mean of 44.6%, increasing cost by a mean of just 11.8%. With access to 30 days of history, the mean SLO violation rate reduction further improves to …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 15 canonical work pages

  1. [1]

    Mania Abdi, Samuel Ginzburg, Xiayue Charles Lin, Jose Faleiro, Gohar Irfan Chaudhry, Inigo Goiri, Ricardo Bianchini, Daniel S Berger, and Rodrigo Fon- seca. 2023. Palette Load Balancing: Locality Hints for Serverless Functions. In Proceedings of the Eighteenth European Conference on Computer Systems(Rome, Italy)(EuroSys ’23). Association for Computing Mac...

  2. [2]

    Mumtaz Ahmad, Ashraf Aboulnaga, Shivnath Babu, and Kamesh Munagala

  3. [3]

    In2008 IEEE 24th International Conference on Data Engineering

    QShuffler: Getting the Query Mix Right. In2008 IEEE 24th International Conference on Data Engineering. IEEE, Piscataway, NJ, USA, 1415–1417. https: //doi.org/10.1109/ICDE.2008.4497574

  4. [4]

    Amazon Web Services. 2026. Amazon Redshift. https://aws.amazon.com/ redshift/. Retrieved June 1, 2026

  5. [5]

    Amazon Web Services. 2026. Amazon Redshift Pricing. https://aws.amazon. com/redshift/pricing/. Retrieved March 17, 2026

  6. [6]

    Amazon Web Services. 2026. Datashares - Amazon Redshift. https://docs.aws. amazon.com/redshift/latest/mgmt/query-editor-v2-datashare-using.html. Re- trieved March 19, 2026

  7. [7]

    Amazon Web Services. 2026. Serverless Computing - AWS Lambda. https: //aws.amazon.com/lambda/. Retrieved March 20, 2026

  8. [8]

    Pankaj Arora, Surajit Chaudhuri, Sudipto Das, Junfeng Dong, Cyril George, Ajay Kalhan, Arnd Christian König, Willis Lang, Changsong Li, Feng Li, et al. 2023. Flexible Resource Allocation for Relational Database-as-a-Service.Proceedings of the VLDB Endowment16, 13 (2023), 4202–4215

Show all 67 references
  1. [9]

    Gohar Irfan Chaudhry, Esha Choukse, Íñigo Goiri, Rodrigo Fonseca, Adam Belay, and Ricardo Bianchini. 2025. Towards Resource-Efficient Compound AI Systems. InProceedings of the 2025 Workshop on Hot Topics in Operating Systems(Banff, AB, Canada)(HotOS ’25). Association for Compu...

  2. [10]

    Martínez

    Shuang Chen, Christina Delimitrou, and José F. Martínez. 2019. PARTIES: QoS-Aware Resource Partitioning for Multiple Interactive Services. InProceed- ings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems(Pro...

  3. [11]

    Wenyan Chen, Chengzhi Lu, Huanle Xu, Kejiang Ye, and Chengzhong Xu. 2025. Multiplexing Dynamic Deep Learning Workloads with SLO-awareness in GPU Clusters. InProceedings of the Twentieth European Conference on Computer Sys- tems(Rotterdam, Netherlands)(EuroSys ’25). Association...

  4. [12]

    1998.Improving WWW proxies performance with greedy- dual-size-frequency caching policy

    Ludmila Cherkasova. 1998.Improving WWW proxies performance with greedy- dual-size-frequency caching policy. Hewlett-Packard Laboratories, Palo Alto, CA, USA

  5. [13]

    Naughton

    Yun Chi, Hakan Hacígümüş, Wang-Pin Hsiung, and Jeffrey F. Naughton. 2013. Distribution-based Query Scheduling.Proc. VLDB Endow.6, 9 (jul 2013), 673–684. https://doi.org/10.14778/2536360.2536367

  6. [14]

    Yun Chi, Hyun Jin Moon, Hakan Hacigümüş, and Junichi Tatemura. 2011. SLA- tree: a framework for efficiently supporting SLA-based decisions in cloud com- puting. InProceedings of the 14th International Conference on Extending Database Technology(Uppsala, Sweden)(EDBT/ICDT ’11)....

  7. [15]

    Yun Chi, Hyun Jin Moon, and Hakan Hacigümüş. 2011. iCBS: incremental cost-based scheduling under piecewise linear SLAs.Proceedings of the VLDB Endowment4, 9 (2011), 563–574

  8. [16]

    Jones, Samuel Madden, and Hari Balakrishnan

    Carlo Curino, Evan P.C. Jones, Samuel Madden, and Hari Balakrishnan. 2011. Workload-aware database monitoring and consolidation. InProceedings of the 2011 ACM SIGMOD International Conference on Management of Data(Athens, Greece)(SIGMOD ’11). Association for Computing Machinery...

  9. [17]

    Narasayya, and Arnd Christian König

    Sudipto Das, Feng Li, Vivek R. Narasayya, and Arnd Christian König. 2016. Automated Demand-driven Resource Scaling in Relational Database-as-a-Service. InProceedings of the 2016 International Conference on Management of Data(San Francisco, California, USA)(SIGMOD ’16). Associa...

  10. [18]

    Christina Delimitrou and Christos Kozyrakis. 2013. Paragon: QoS-aware sched- uling for heterogeneous datacenters.SIGPLAN Not.48, 4 (March 2013), 77–88. https://doi.org/10.1145/2499368.2451125

  11. [19]

    Christina Delimitrou and Christos Kozyrakis. 2014. Quasar: resource-efficient and QoS-aware cluster management. InProceedings of the 19th International Conference on Architectural Support for Programming Languages and Operating Systems(Salt Lake City, Utah, USA)(ASPLOS ’14). A...

  12. [20]

    Alexander Fuerst and Prateek Sharma. 2021. FaasCache: keeping serverless computing alive with greedy-dual caching. InProceedings of the 26th ACM In- ternational Conference on Architectural Support for Programming Languages and Operating Systems(Virtual, USA)(ASPLOS ’21). Assoc...

  13. [21]

    Ke Hong, Xiuhong Li, Lufang Chen, Qiuli Mao, Guohao Dai, Xuefei Ning, Shengen Yan, Yun Liang, and Yu Wang. 2025. SOLA: Optimizing SLO Attainment for Large Language Model Serving with State-Aware Scheduling. InEighth Conference on Machine Learning and Systems. https://openrevie...

  14. [22]

    Cunchen Hu, Heyang Huang, Liangliang Xu, Xusheng Chen, Jiang Xu, Shuang Chen, Hao Feng, Chenxi Wang, Sa Wang, Yungang Bao, Ninghui Sun, and Yizhou Shan. 2024. Inference without Interference: Disaggregate LLM Inference for Mixed Downstream Workloads. arXiv:2401.11181 [cs.DC] ht...

  15. [23]

    Yuwei Huang and Guoliang Li. 2024. Laser: Buffer-Aware Learned Query Sched- uling in Master-Standby Databases.Proc. VLDB Endow.18, 3 (Nov. 2024), 743–755. https://doi.org/10.14778/3712221.3712239

  16. [24]

    2019.Intel Xeon Gold 6230 CPU

    Intel Corporation. 2019.Intel Xeon Gold 6230 CPU. Intel Corporation. Retrieved July 31, 2024 from https://ark.intel.com/content/www/us/en/ark/ products/192437/intel-xeon-gold-6230-processor-27-5m-cache-2-10-ghz.html

  17. [25]

    Yadwadkar, and Christos Kozyrakis

    Kostis Kaffes, Neeraja J. Yadwadkar, and Christos Kozyrakis. 2022. Hermod: principled and practical scheduling for serverless functions. InProceedings of the 13th Symposium on Cloud Computing(San Francisco, California)(SoCC ’22). Association for Computing Machinery, New York, ...

  18. [26]

    Ferdi Kossmann, Ziniu Wu, Alex Turk, Nesime Tatbul, Lei Cao, and Samuel Madden. 2024. CascadeServe: Unlocking Model Cascades for Inference Serving. arXiv:2406.14424 [cs.DC] https://arxiv.org/abs/2406.14424

  19. [27]

    Tim Kraska, Tianyu Li, Samuel Madden, Markos Markakis, Amadou Ngom, Ziniu Wu, and Geoffrey X. Yu. 2023. Check Out the Big Brain on BRAD: Simplifying Cloud Data Processing with Learned Automated Data Meshes.Proc. VLDB Endow. 16, 11 (jul 2023), 3293–3301. https://doi.org/10.1477...

  20. [28]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. InProceedings of the 29th Symposium on Operating Systems Principl...

  21. [29]

    Patel, and Ajay Kalhan

    Willis Lang, Srinath Shankar, Jignesh M. Patel, and Ajay Kalhan. 2014. Towards Multi-Tenant Performance SLOs.IEEE Transactions on Knowledge and Data Engineering26, 6 (2014), 1447–1463. https://doi.org/10.1109/TKDE.2013.74

  22. [30]

    Philipp Leitner, Waldemar Hummer, Benjamin Satzger, Christian Inzinger, and Schahram Dustdar. 2012. Cost-efficient and application SLA-aware client side request scheduling in an infrastructure-as-a-service cloud. In2012 IEEE Fifth International Conference on Cloud Computing. I...

  23. [31]

    Yuhang Li, Rong Gu, Chengying Huan, Zhibin Wang, Renjie Yao, Chen Tian, and Guihai Chen. 2025. HotPrefix: Hotness-Aware KV Cache Scheduling for Efficient Prefix Sharing in LLM Inference Systems.Proc. ACM Manag. Data3, 4, Article 250 (Sept. 2025), 27 pages. https://doi.org/10.1...

  24. [32]

    Ziyang Liu, Hakan Hacıgümüş, Hyun Jin Moon, Yun Chi, and Wang-Pin Hsiung

  25. [33]

    InProceedings of the 16th International Conference on Extending Database Tech- nology(Genoa, Italy)(EDBT ’13)

    PMAX: tenant placement in multitenant databases for profit maximization. InProceedings of the 16th International Conference on Extending Database Tech- nology(Genoa, Italy)(EDBT ’13). Association for Computing Machinery, New York, NY, USA, 442–453. https://doi.org/10.1145/2452...

  26. [34]

    David Lo, Liqun Cheng, Rama Govindaraju, Parthasarathy Ranganathan, and Christos Kozyrakis. 2015. Heracles: improving resource efficiency at scale. SIGARCH Comput. Archit. News43, 3S (June 2015), 450–462. https://doi.org/10. 1145/2872887.2749475

  27. [35]

    Ryan Marcus and Olga Papaemmanouil. 2016. WiSeDB: a learning-based work- load management advisor for cloud databases.Proc. VLDB Endow.9, 10 (jun 2016), 780–791. https://doi.org/10.14778/2977797.2977804

  28. [36]

    Ryan Marcus and Olga Papaemmanouil. 2017. Releasing Cloud Databases for the Chains of Performance Prediction Models.. InCIDR

  29. [37]

    Markos Markakis. 2026. AutoSLO Code. https://github.com/mmarkakis/autoslo

  30. [38]

    Yadwadkar, and Christos Kozyrakis

    Daniel Mendoza, Francisco Romero, Qian Li, Neeraja J. Yadwadkar, and Christos Kozyrakis. 2021. Interference-Aware Scheduling for Inference Serving. InPro- ceedings of the 1st Workshop on Machine Learning and Systems(Online, United Kingdom)(EuroMLSys ’21). Association for Compu...

  31. [39]

    Vivek Narasayya, Sudipto Das, Manoj Syamala, Badrish Chandramuli, and Surajit Chaudhuri. 2013. SQLVM: Performance Isolation in Multi-Tenant Relational Database-as-a-Service. InCIDR

  32. [40]

    Vivek Narasayya, Sudipto Das, Manoj Syamala, Surajit Chaudhuri, Feng Li, and Hyunjung Park. 2013. A demonstration of SQLVM: performance isolation in multi-tenant relational database-as-a-service. InProceedings of the 2013 ACM SIGMOD International Conference on Management of Da...

  33. [41]

    Vikram Nathan, Vikramank Singh, Zhengchun Liu, Mohammad Rahman, An- dreas Kipf, Dominik Horn, Davide Pagano, Gaurav Saxena, Balakrishnan Narayanaswamy, and Tim Kraska. 2024. Intelligent Scaling in Amazon Red- shift. InCompanion of the 2024 International Conference on Managemen...

  34. [42]

    Jennifer Ortiz, Victor Teixeira De Almeida, and Magdalena Balazinska. 2015. Changing the Face of Database Cloud Services with Personalized Service Level Agreements. InCIDR

  35. [43]

    Jennifer Ortiz, Brendan Lee, and Magdalena Balazinska. 2016. PerfEnforce Demonstration: Data Analytics with Performance Guarantees. InProceedings of the 2016 International Conference on Management of Data(San Francisco, California, USA)(SIGMOD ’16). Association for Computing M...

  36. [44]

    Jennifer Ortiz, Brendan Lee, Magdalena Balazinska, Johannes Gehrke, and Joseph L Hellerstein. 2018. SLAOrchestrator: Reducing the Cost of Performance SLAs for Cloud Data Analytics. In2018 USENIX Annual Technical Conference (USENIX ATC 18). USENIX Association, Berkeley, CA, USA...

  37. [45]

    Tirthak Patel and Devesh Tiwari. 2020. CLITE: Efficient and QoS-Aware Co- Location of Multiple Latency-Critical Jobs for Warehouse Scale Computers. In 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, Piscataway, NJ, USA, 193–206. https:...

  38. [46]

    Meikel Poess, Bryan Smith, Lubor Kollar, and Paul Larson. 2002. TPC-DS, taking decision support benchmarking to the next level. InProceedings of the 2002 ACM SIGMOD International Conference on Management of Data(Madison, Wiscon- sin)(SIGMOD ’02). Association for Computing Mach...

  39. [47]

    Uwe Rohm, Klemens Bohm, and H-J Schek. 2001. Cache-aware query routing in a cluster of databases. InProceedings 17th International Conference on Data Engineering. IEEE, IEEE, Piscataway, NJ, USA, 641–650

  40. [48]

    Francisco Romero and Christina Delimitrou. 2018. Mage: online and interference- aware scheduling for multi-scale heterogeneous systems. InProceedings of the 27th International Conference on Parallel Architectures and Compilation Techniques (Limassol, Cyprus)(PACT ’18). Associa...

  41. [49]

    Ibrahim Sabek, Tenzin Samten Ukyab, and Tim Kraska. 2022. LSched: A Workload- Aware Learned Query Scheduler for Analytical Database Systems. InProceedings of the 2022 International Conference on Management of Data(Philadelphia, PA, USA)(SIGMOD ’22). Association for Computing M...

  42. [50]

    Sherif Sakr and Anna Liu. 2012. SLA-based and consumer-centric dynamic provisioning for cloud databases. In2012 IEEE Fifth International Conference on Cloud Computing. IEEE, IEEE, Piscataway, NJ, USA, 360–367. 15

  43. [51]

    Gaurav Saxena, Mohammad Rahman, Naresh Chainani, Chunbin Lin, George Caragea, Fahim Chowdhury, Ryan Marcus, Tim Kraska, Ippokratis Pandis, and Balakrishnan (Murali) Narayanaswamy. 2023. Auto-WLM: Machine Learning Enhanced Workload Management in Amazon Redshift. InCompanion of ...

  44. [52]

    Snowflake. 2026. Multi-cluster warehouses | Snowflake Documentation. https: //docs.snowflake.com/en/user-guide/warehouses-multicluster. Retrieved March 19, 2026

  45. [53]

    Snowflake. 2026. Pricing Options. https://www.snowflake.com/en/pricing- options/. Retrieved March 17, 2026

  46. [54]

    Snowflake. 2026. Snowflake AI Data Cloud. https://www.snowflake.com/en/. Retrieved June 1, 2026

  47. [55]

    Elmore, Michael Stonebraker, and David DeWitt

    Rebecca Taft, Willis Lang, Jennie Duggan, Aaron J. Elmore, Michael Stonebraker, and David DeWitt. 2016. STeP: Scalable Tenant Placement for Managing Database- as-a-Service Deployments. InProceedings of the Seventh ACM Symposium on Cloud Computing(Santa Clara, CA, USA)(SoCC ’16...

  48. [56]

    Tanenbaum and D

    A.S. Tanenbaum and D. Wetherall. 2011.Computer Networks. Pearson Prentice Hall. https://books.google.com/books?id=T7lGswEACAAJ

  49. [57]

    Alexander van Renen, Dominik Horn, Pascal Pfeil, Kapil Vaidya, Wenjian Dong, Murali Narayanaswamy, Zhengchun Liu, Gaurav Saxena, Andreas Kipf, and Tim Kraska. 2024. Why TPC is Not Enough: An Analysis of the Amazon Redshift Fleet.Proc. VLDB Endow.17, 11 (July 2024), 3694–3706. ...

  50. [58]

    Johannes Wehrstein, Roman Heinrich, Mihail Stoian, Skander Krid, Martin Stemmer, Andreas Kipf, Carsten Binnig, and Muhammad El-Hindi. 2025. Red- bench: Workload Synthesis From Cloud Traces. arXiv:2511.13059 [cs.DB] https://arxiv.org/abs/2511.13059

  51. [59]

    Ziniu Wu, Ryan Marcus, Zhengchun Liu, Parimarjan Negi, Vikram Nathan, Pascal Pfeil, Gaurav Saxena, Mohammad Rahman, Balakrishnan Narayanaswamy, and Tim Kraska. 2024. Stage: Query Execution Time Prediction in Amazon Redshift. In Companion of the 2024 International Conference on...

  52. [60]

    Ziniu Wu, Markos Markakis, Chunwei Liu, Peter Baile Chen, Balakrishnan Narayanaswamy, Tim Kraska, and Samuel Madden. 2025. Improving DBMS Scheduling Decisions with Fine-grained Performance Prediction on Concurrent Queries – Extended. arXiv:2501.16256 [cs.DB] https://arxiv.org/...

  53. [61]

    Pengcheng Xiong, Yun Chi, Shenghuo Zhu, Hyun Jin Moon, Calton Pu, and Hakan Hacigümüş. 2011. Intelligent management of virtualized resources for database systems in cloud environment. In2011 IEEE 27th International Conference on Data Engineering. IEEE, Piscataway, NJ, USA, 87–...

  54. [62]

    Pengcheng Xiong, Yun Chi, Shenghuo Zhu, Junichi Tatemura, Calton Pu, and Hakan HacigümüŞ. 2011. ActiveSLA: a profit-oriented admission con- trol framework for database-as-a-service providers. InProceedings of the 2nd ACM Symposium on Cloud Computing(Cascais, Portugal)(SOCC ’11...

  55. [63]

    Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung- Gon Chun. 2022. Orca: A distributed serving system for {Transformer-Based} generative models. In16th USENIX symposium on operating systems design and implementation (OSDI 22). USENIX Association, Berkeley, C...

  56. [64]

    Yu, Ziniu Wu, Ferdi Kossmann, Tianyu Li, Markos Markakis, Amadou Ngom, Samuel Madden, and Tim Kraska

    Geoffrey X. Yu, Ziniu Wu, Ferdi Kossmann, Tianyu Li, Markos Markakis, Amadou Ngom, Samuel Madden, and Tim Kraska. 2024. Blueprinting the Cloud: Unifying and Automatically Optimizing Cloud Data Infrastructures with BRAD.Proc. VLDB Endow.17, 11 (Aug 2024), 3629–3643. https://doi...

  57. [65]

    Xinyi Zhang, Hong Wu, Zhuo Chang, Shuowei Jin, Jian Tan, Feifei Li, Tieying Zhang, and Bin Cui. 2021. ResTune: Resource Oriented Tuning Boosted by Meta- Learning for Cloud Databases. InProceedings of the 2021 International Conference on Management of Data(Virtual Event, China)...

  58. [66]

    Liang Zhao, Sherif Sakr, and Anna Liu. 2013. A framework for consumer-centric SLA management of cloud-hosted databases.IEEE Transactions on Services Computing8, 4 (2013), 534–549

  59. [67]

    Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xuanzhe Liu, Xin Jin, and Hao Zhang. 2024. DistServe: disaggregating prefill and decoding for goodput-optimized large language model serving. InProceedings of the 18th USENIX Conference on Operating Systems Design and...

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.