{"id":"9bb87572-d8be-485f-924b-89de25e1499c","arxiv_id":"2501.11369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multidimensional elasticity framework uses a Q-learning orchestrator to adapt both resources and analytics requirements (coverage, sample, freshness) in computing continuum systems.","lead":"This paper proposes a framework that adjusts both infrastructure resources and data analytics settings, such as coverage, sample size, and freshness, to keep IoT analytics running smoothly across edge and cloud nodes. The framework uses a reinforcement learning orchestrator to make these adjustments automatically, and it is tested in a small smart-city testbed and a simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RL orchestrator's implemented action space omits coverage (§III-D3b), so the abstract's claim that the framework adaptively modulates coverage, sample, and freshness is not demonstrated; only sample and freshness are RL-adaptable.","rationale":"The reader's weakest_assumption (identical nodes, binary CPU utilization) is a real modeling simplification, but it concerns generalization to heterogeneous deployments and can be relaxed without changing the framework's structure. The coverage omission is more direct: §III-D3b explicitly states that the core decision-making component does not include a dimension that the abstract and introduction list as one of the three data analytics requirements the framework modulates. This is an internal consistency issue rather than a question of external consensus, so it is more load-bearing for the central claim. The paper does provide useful evidence: the testbed impact measurements (Fig. 3) and the RL SLO-fulfillment study (Table I) support adaptive management of sample, freshness, resources, and cost, and the balanced reward row shows a plausible trade-off. Those results are not invalidated; they just do not support coverage adaptation. Therefore the appropriate verdict remains CONDITIONAL: the authors should either add coverage to the RL agent and evaluate it, or explicitly scope the claim to sample, freshness, and resource/cost dimensions, with coverage as a manually controlled parameter. This does not change the reader's verdict, hence UNCHANGED.","tokens_in":7234,"tokens_out":3520,"duration_ms":37701,"concrete_test":"Extend the RL agent as the paper's own extension sketch suggests (§III-D3b): add a discrete coverage component to the state and action spaces (e.g., number of active edge nodes) and a coverage SLO with an associated reward weight r_cov, then rerun the Table I experiments with a balanced reward structure. If the learned policy cannot satisfy the coverage SLO or the balanced trade-off across CPU/freshness/sample/cost/coverage degrades, the omission is load-bearing; if it can, the authors should update the paper to include coverage in the implemented action space and evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract and §I) is that the orchestrator can dynamically adjust data analytics requirements, 'including coverage, sample, and freshness.' However, the RL-based predictor's state and action spaces encode only 'freshness (data collection frequency) and sample size' (§III-D3b), and the paper explicitly states: 'in this instantiation of the QL agent, we do not consider analytics coverage.' Consequently, the orchestrated adaptive loop cannot change coverage, and the evaluation reflects this: Table I reports SLO fulfillment only for CPU load, freshness, sample, and cost, with no coverage SLO, and the adaptation example in Fig. 4 adjusts sample and freshness, not coverage. Coverage appears in Fig. 3a solely as a manually varied testbed parameter, not as an orchestrator action. Thus the 'multidimensional' elasticity claim overstates what is implemented: the framework adapts two analytics requirements (sample, freshness) plus resources/cost, while coverage remains a manually set parameter. If the central contribution is specifically multidimensional adaptation of analytics requirements, this gap is material.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multidimensional elasticity framework for data analytics in the computing continuum. It distinguishes data analytics requirements (coverage, sample, freshness, response time, cost) from infrastructure dimensions (computing nodes, network, data) and presents an orchestrator whose decision-making is driven by a Q-learning agent. The agent is intended to adapt both resource allocations and analytics configurations to meet SLOs. The evaluation has two parts: testbed experiments on a head node and three edge nodes measuring the CPU impact of varying coverage, sample, and freshness; and a simulation study reporting SLO fulfillment ratios under different reward configurations for the Q-learning predictor.","tokens_in":7476,"tokens_out":2805,"duration_ms":31840,"significance":"If fully realized, the framework would make a useful contribution by letting continuum systems automatically trade analytics quality (coverage, sample, freshness) for resource headroom and cost. The paper provides a clear categorization of analytics-related elasticity dimensions, a concrete Q-learning orchestrator design, and an explicit reward formulation tied to SLOs. The testbed measurements give a qualitative demonstration that analytics requirements affect CPU usage, and the authors are transparent about several modeling simplifications. However, the implemented RL agent omits coverage, the evaluation is preliminary (single small testbed, no baseline comparison, no statistical rigor), and the RL results largely reflect the hand-set reward structure. The significance as demonstrated is therefore moderate, although the architectural idea is promising.","major_comments":[{"comment":"The abstract and §I claim that the orchestrator dynamically adjusts data analytics requirements 'including coverage, sample, and freshness,' but the Q-learning agent's state and action spaces encode only freshness and sample size, and the paper explicitly states 'we do not consider analytics coverage.' Table I reports no coverage SLO, and Fig. 4 adapts only sample and freshness. Coverage appears in Fig. 3a only as a manually varied testbed parameter. Thus the central 'multidimensional' claim overstates what is implemented and evaluated; the framework currently adapts two analytics requirements plus resource/cost dimensions, not the three named requirements.","section":"§III-D3b, §IV-E, Abstract"},{"comment":"The reward weights in §III-D3c are hand-set to prioritize specific SLOs (e.g., rf = 10 with other weights set to 1), and Table I then reports that the prioritized SLO is fulfilled close to 100% of the time. This outcome is largely a consequence of the reward design rather than an independent prediction of the framework's effectiveness. The experiment demonstrates that the Q-learning agent can optimize a given reward, but it does not validate the framework's ability to balance requirements in a realistic setting. A baseline (e.g., a greedy policy, a random policy, or an offline heuristic) and a convergence analysis would make the decision-making claim more substantive.","section":"§III-D3c, Table I"},{"comment":"The simulation model is only partially specified: the overload transition probability is mentioned but not given, the data intensity distribution is not defined, and the mapping from capacity levels to cost is described only qualitatively. No error bars, confidence intervals, or multiple-seed results are reported for the SLO fulfillment ratios. As a result, the numbers in Table I cannot be reproduced or assessed for statistical significance, and the claimed superiority of the balanced reward configuration is not supported by any comparison to alternative policies.","section":"§IV-E"},{"comment":"The state space assumes that all compute nodes and application instances are identical and that CPU utilization is binary (overloaded or not, if any node is overloaded). This simplification prevents the agent from learning to target a bottleneck node, and the policy may not transfer to heterogeneous continuum deployments where nodes differ in capacity, cost, and location. The authors acknowledge this simplification, but the evaluation does not test robustness to node heterogeneity, which is central to the 'computing continuum' setting described in §I.","section":"§III-D3b, §IV"}],"minor_comments":[{"comment":"Response Time and Cost are listed as data analytics requirements, but they do not appear in the RL state/action spaces or in the evaluation; clarify how these requirements are meant to be handled by the orchestrator.","section":"§III-A"},{"comment":"The testbed measurements in Figs. 3a-c appear to be single runs; adding error bars or repeated-run statistics would make the reported trends more convincing.","section":"§IV-C"},{"comment":"The description of the Q-learning setup omits the learning rate, discount factor, exploration strategy, and convergence criterion; specifying these would improve reproducibility.","section":"§IV-E"},{"comment":"The paper says it uses 'the One simulator' to generate the person movement dataset, but reference [15] is titled 'Oppnets and rural areas: an opportunistic solution for remote communications'; please verify the citation and provide the correct reference for the ONE (Opportunistic Network Environment) simulator.","section":"References"},{"comment":"The reward structure notation '(rCP U: rf : rs : rc)' is not defined in the table caption; adding a one-line definition would improve readability.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper reads more like an architecture/position paper with a preliminary demonstration than a fully validated system. The gap between the abstract's claim of adaptively managing coverage, sample, and freshness and the implemented RL agent that omits coverage is significant and should be addressed either by extending the implementation or by carefully restating the contribution. The evaluation would also need stronger baselines and statistical grounding. The manuscript fits the journal's scope if revised to match claims to evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a modest but legitimate extension of the authors' prior quality-elasticity work: they add a Q-learning orchestrator that jointly adjusts resource allocation and two analytics requirements (sample size and freshness), with cost folded into the resource dimension. The testbed experiments usefully show that coverage, sample, and freshness all affect CPU usage in the expected directions, and the simulation demonstrates that hand-set reward weights can steer SLO priorities. The writing is clear and the limitations are mostly stated openly.\n\nThe main problem is the coverage gap. The abstract and introduction claim the framework adaptively modulates coverage, sample, and freshness, but the RL agent's state and action spaces, described in Section III-D3b, only include freshness and sample size. The paper explicitly says \"in this instantiation of the QL agent, we do not consider analytics coverage.\" Coverage appears only as a manually varied parameter in the testbed, never as an orchestrator action. So the central \"multidimensional\" claim overstates what is implemented. This is not a nitpick; it is the advertised contribution.\n\nThe evaluation is otherwise thin but not dishonest. The testbed has one head node and three edge nodes with no repetition or error bars, so the CPU curves are qualitative. The RL simulation lacks baselines (random or heuristic policies), and Table I largely reflects the reward weights rather than an independent finding. The system model also assumes identical nodes and binary overload state, which limits transferability. These are weaknesses, but they are the usual kind of proof-of-concept gaps, not fatal flaws.\n\nThe paper would benefit from aligning the claims with the actual implementation, adding baselines and basic statistics, and either extending the agent to handle coverage or explicitly scoping the contribution to sample and freshness. I would send it to peer review, because the framework is plausible, the gap is fixable in revision, and the subfield could use more concrete orchestrator designs. I would not cite it in my own work until the coverage issue is resolved.","headline":"A clearly written elasticity framework with a real gap: the RL orchestrator adapts sample and freshness but not coverage, despite the abstract's multidimensional claim.","tokens_in":7996,"tokens_out":1608,"would_cite":false,"duration_ms":18540,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multidimensional elasticity framework lets an RL-based orchestrator adapt both infrastructure resources and data analytics requirements (coverage, sample, freshness) so analytics can continue even when resources cannot scale.","keywords":["multidimensional elasticity","computing continuum","Q-learning orchestrator","data analytics requirements","SLO fulfillment","edge computing","resource adaptation","smart cities"],"falsifier":"Run the trained Q-learning policy on a deployment where nodes have unequal processing capacities and uneven data sources; if SLO fulfillment degrades significantly below the reported 95–98% balanced ratios, or the agent repeatedly overloads the slower nodes, the identical-node simplification is falsified. Alternatively, remove a node during operation and observe whether the policy recovers without retraining.","tokens_in":7048,"feed_emoji":"⚙️","tokens_out":3451,"duration_ms":33422,"temperature":0.7,"pith_summary":"The paper proposes a management framework for computing continuum systems (IoT-edge-cloud) that treats data analytics requirements—coverage, sample, and freshness—as elastic dimensions alongside CPU, memory, bandwidth, and cost. The central claim is that an orchestrator driven by Q-learning can, at runtime, choose adjustments to both the analytics configuration and the allocated compute capacity so that service-level objectives are met even when the infrastructure cannot simply scale up. If correct, this gives operators a way to keep analytics useful during resource bottlenecks by deliberately degrading less-critical analytics qualities rather than letting the system fail. Experiments on a smart-city heatmap testbed and simulations of a four-node environment show the impact of these requirements on CPU usage and that the RL agent can prioritize whichever SLO the operator weights most heavily.","feed_headline":"When compute runs short, this orchestrator trims analytics scope","feed_subtitle":"A Q-learning controller adjusts freshness, coverage and sample size to keep service-level objectives within reach.","key_machinery":"The load-bearing mechanism is a tabular Q-learning agent embedded in an orchestrator with two monitors: the Architecture Monitor tracks resource dimensions and the Analytics Monitor tracks consumer requirements. The state space is a discretized combination of a binary CPU-utilization flag (overloaded or not), the allocated compute capacity level, and the analytics configuration levels (freshness and sample size); the action space likewise selects a capacity level and an analytics configuration. The reward function is the sum of fixed positive or negative contributions for each SLO (CPU, cost, freshness, sample), making the learned policy directly reflect operator priorities. The framework also spells out explicit relationships between requirements and architecture dimensions (coverage maps to active node count, sample to data volume, freshness to refresh rate), which are used to explain the measured CPU effects and guide the adaptation space.","core_discovery":"The paper's contribution is a framework in which elasticity applies not only to infrastructure (CPU, memory, bandwidth) but also to the data analytics application's own requirements. The discovery is that a Q-learning agent can learn a policy that interleaves two kinds of actions: scaling resources (changing the number of application instances or capacity) and degrading or restoring analytics requirements (freshness and sample size; coverage is left out of the RL instantiation). Under each state defined by CPU utilization level, allocated capacity, and analytics configuration, the agent receives a reward for each SLO met; tuning reward weights steers the resulting behavior, for example prioritizing freshness at the expense of cost. The evaluation shows the agent can maintain near-100% fulfillment of a prioritized SLO when that SLO's reward weight is dominant, while a balanced reward weighting yields roughly 95–98% fulfillment across CPU, freshness, sample, and cost SLOs.","pith_inferences":["The same reward-shaping approach could naturally extend to multi-objective or constrained RL to handle heterogeneous nodes, which the paper's identical-node assumption currently prevents.","A stress test with unequal node capacities and uneven data-source distributions would likely expose the binary CPU-utilization state's limitation, as the agent cannot pinpoint the bottleneck node.","Comparing the learned policy to a simple threshold-based heuristic would clarify how much of the benefit comes from RL rather than from the elasticity framework itself, a comparison the paper does not report."],"forward_implications":["Operators can encode SLO priorities as reward weights and obtain a runtime controller that automatically trades analytics quality for resource headroom, without needing an explicit model of the environment.","Because coverage, sample, and freshness each have distinct CPU load signatures (stepwise, gradual, and periodic-spiky), a controller that only scales resources cannot match the flexibility of also adapting these analytics requirements.","The Q-learning agent converges on a coarse discretized state and action space, suggesting that practical deployments do not require high-resolution state representations to achieve balanced SLO fulfillment around 95–98%.","The current RL instantiation omits coverage, so the framework as evaluated would need state and action extensions before it can manage geo-distributed deployments where coverage is a primary lever."],"supporting_citations":[{"why":"Defines elasticity along the quality, resources, and costs dimensions that this framework extends.","marker":"[7]"},{"why":"The authors' previous work on quality-related elasticity, which this paper goes beyond by adding resource and cost dimensions.","marker":"[8]"},{"why":"Supplies the traditional elasticity dimension concept that the multi-dimensional approach builds on.","marker":"[9]"},{"why":"Provides the ONE simulator used to generate the person movement dataset for the smart city heatmap case study.","marker":"[15]"}],"fun_headline_variants":["Q-learning orchestrator tunes compute and analytics freshness to hit SLOs","Orchestrator learns to trade compute for analytics freshness","Q-learning agents adjust data quality to keep compute SLOs","Adaptive orchestrator: shrink analytics to fit compute limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learning agent assumes all compute nodes are identical in capacity and allocation and treats the whole system as over-utilized if any single node is overloaded, so the learned policy may fail on heterogeneous real-world infrastructures.","fun_headline_variants_meta":{"raw":{"variants":["Q-learning orchestrator tunes compute and analytics freshness to hit SLOs","Orchestrator learns to trade compute for analytics freshness","Q-learning agents adjust data quality to keep compute SLOs","Adaptive orchestrator: shrink analytics to fit compute limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001138,"raw_usage":{"total_tokens":4693,"prompt_tokens":883,"completion_tokens":3810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":3739}},"tokens_in":499,"tokens_out":3810,"duration_ms":26852,"temperature":1.0,"reasoning_tokens":3739,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:19:15.778846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained Q-learning policy on a deployment where nodes have unequal processing capacities and uneven data sources; if SLO fulfillment degrades significantly below the reported 95–98% balanced ratios, or the agent repeatedly overloads the slower nodes, the identical-node simplification is falsified. Alternatively, remove a node during operation and observe whether the policy recovers without retraining.","supporting_citations":[{"cited_title":"Principles of elastic processes,","cited_arxiv_id":null,"evidence_quote":"Defines elasticity along the quality, resources, and costs dimensions that this framework extends."},{"cited_title":"Elastic data analytics for the cloud-to-things continuum,","cited_arxiv_id":null,"evidence_quote":"The authors' previous work on quality-related elasticity, which this paper goes beyond by adding resource and cost dimensions."},{"cited_title":"Decent: A decentralized configurator for controlling elasticity in dynamic edge networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the traditional elasticity dimension concept that the multi-dimensional approach builds on."},{"cited_title":"Oppnets and rural areas: an opportunistic solution for remote communications,","cited_arxiv_id":null,"evidence_quote":"Provides the ONE simulator used to generate the person movement dataset for the smart city heatmap case study."}],"review_version":1}