{"id":"a7ce06a6-fbe0-4b8d-be41-b9b68d7cfa26","arxiv_id":"2606.30563","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Global-view greedy placement stays within a few percent of BLP optimum to 10k nodes; aggregated-view local decisions outperform centralized policies under client mobility because data placement dominates scheduling.","lead":"The paper jointly optimizes serverless function scheduling and data-replica placement under strong and eventual consistency on edge-cloud trees. A global greedy heuristic tracks the exact optimum at large scale while a fully local reactive protocol keeps client latency stable when users move.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged static-access assumption.","rationale":"The strongest claim is comparative and scoped to the paper's own model and simulation. Inside that scope the evidence is careful: BLP vs. GV scaling and quality (Figs. 3–4), client-observed latency under static and mobile workloads (Fig. 5), and the CO/CD baselines that isolate data placement. The only assumption that, if violated, would invalidate both the BLP correctness guarantees and the force/always-copy heuristics is precisely the static access knowledge already flagged by the reader. No further soft spot (e.g., an unstated bound on tree depth, an untested regime of α, or a confounding of gather-scatter cost) rises to the same load-bearing level. Consequently the CONDITIONAL verdict, high confidence, and low correctness-risk assessment stand; the concrete test above is simply a direct falsification of that shared weakest assumption.","tokens_in":27017,"tokens_out":537,"duration_ms":5879,"concrete_test":"Re-run the RQ2 collection sweep (Fig. 4) and the RQ4 mobility experiment (Fig. 5b) after randomly re-assigning 20 % of each function's access set every 30 s of simulated time (matching the mobility interval). If GV's T-diff to BLP exceeds ~15 % or AV loses its flat latency profile under the same mobility pattern, the static-access assumption is load-bearing for the strongest claim; otherwise the claim remains robust inside the model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central comparative claims (GV tracks BLP within a few percent up to 10^4 nodes; AV yields lower/stabler client latency under mobility; data placement dominates scheduling) are internally consistent with the model, the BLP formulation (§4), the two heuristics (§5–6), and the simulation campaign (§7). The reader's weakest assumption—static a-priori knowledge of function-to-collection access sets and read/write classification (§3.2, §7.1)—is correctly identified as the main external limitation; the authors themselves list it as future work (§8). No additional load-bearing internal inconsistency, hidden scaling assumption, or evaluation artifact that would overturn the strongest claim was found. The idealized tree topology, pure-simulation setting, and lack of released code are standard genre limitations already reflected in the CONDITIONAL verdict; they do not introduce a new correctness risk that the reported numbers fail to support inside the stated model.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies the joint problem of serverless function scheduling and data-replica placement in a hierarchical edge–cloud continuum under two consistency models (strong single-leader replication and eventual replication). It formulates the problem as a Binary Linear Program that jointly optimizes placement, leader election, scheduling, and ER fetch routing under a latency–storage trade-off α; proposes a Global-View (GV) greedy heuristic that builds a minimal feasible baseline then expands replicas by the same trade-off; and proposes an Aggregated-View (AV) decentralized protocol that places SR leaders/followers via a virtual force model and ER replicas via always-copy, with hop-by-hop scheduling from local reachability state. Evaluation on synthetic three-tier trees shows: BLP does not scale past a few hundred nodes; GV stays close to the BLP objective while scaling beyond 10^4 nodes; under static and mobile client workloads in a discrete-event simulator, AV often yields lower and more stable client-observed latency than snapshot-based centralized policies; and naive baselines that keep data centralized confirm that data placement dominates pure function scheduling.","tokens_in":27276,"tokens_out":1436,"duration_ms":27894,"significance":"The joint treatment of function scheduling and data placement under heterogeneous consistency is a genuine and under-addressed systems problem for stateful FaaS at the edge. Strengths include a correctly linearized BLP (McCormick envelopes for miss indicators), a GV heuristic that reuses the same objective trade-off and is shown to track the solver on shared instances, and an AV design that cleanly separates proactive SR force equilibrium from reactive ER caching. The evaluation structure (scalability of exact vs greedy, analytic approximation quality, then client-observed latency under static and mobile demand) is coherent and supports the main comparative claims inside the stated model. The progressive information-scope design (full snapshot → local aggregates) is a useful contribution even if the topology and access model are idealized.","major_comments":[{"comment":"Abstract and §7.3 / Fig. 4: the claim that GV stays “within a few percent of the optimum” is not uniformly supported. At α=1 under ER, GV produces the same placement as Cloud-only (no greedy expansion) and incurs roughly 12% higher T than BLP at small |C|; larger gaps and wide IQR appear under MIX at small scale. The abstract and conclusions should qualify the approximation quality by α and consistency mix (the few-percent regime is mainly α≪1 or SR/MIX after expansion), rather than stating it as a blanket result.","section":"Abstract, §7.3, Fig. 4"},{"comment":"§6.2.1 and §7.1–7.5: AV’s force model introduces free parameters μ, τ_L, τ_R, and decay γ (plus the shared 32 s epoch/window W) with no sensitivity or robustness study. The central mobility claim—that AV maintains low, stable client latency while BLP/GV spike (Fig. 5b)—depends on these choices relative to the mobility interval (30 s) and W. At minimum, report the concrete values used, justify them, and show that the qualitative ranking is stable under moderate perturbation; otherwise the RQ4 advantage is hard to reproduce or generalize.","section":"§6.2.1, §7.1, Fig. 5b"},{"comment":"§7.4–7.5: “BLP” in the simulator is a time-budgeted incumbent (0.2 s / 0.5 s per iteration) plus gather–scatter delay, not the optimal BLP of §4. The text correctly notes timeouts but still frames results as AV vs centralized/optimal policies. Please separate (i) analytic optimality gap of GV vs true BLP (RQ1–2) from (ii) operational client latency of time-limited re-planning vs AV (RQ3–4), and avoid language that equates the short-timeout incumbent with the §4 optimum when interpreting mobility spikes.","section":"§7.4, §7.5"}],"minor_comments":[{"comment":"§3.2 / §8: static a-priori knowledge of function–collection access sets and read/write labels is load-bearing for BLP, GV, and AV correctness. It is acknowledged as future work; a short explicit threat-to-validity paragraph in §7 would help readers bound the claims.","section":"§3.2, §8"},{"comment":"Release or archive the scenario generator, BLP model, heuristics, and simulator (or a detailed artifact appendix). Without it, reproducing Fig. 3–5 and the force-parameter settings is difficult.","section":"§7"},{"comment":"Fig. 3–5: axis labels and legend text are hard to read in the manuscript rendering; enlarge fonts and spell out |N|, |F|, |C| in captions. Fig. 5 y-axis “Latency (s)” with a 10^0-style scale needs a clear unit and whether the series is median-per-bin.","section":"Fig. 3, Fig. 4, Fig. 5"},{"comment":"Abstract and early pages contain many missing spaces after punctuation (e.g., “Serverlesscomputing”, “BinaryLinearProgram”). Clean typesetting before camera-ready.","section":"Abstract, §1"},{"comment":"§4.2: state explicitly whether T_max / S_max normalization constants are recomputed per instance and whether α is dimensionless after normalization (it is, but a sentence would prevent mis-tuning by readers).","section":"§4.2, Eq. (6)"},{"comment":"Related work: a short table mapping prior systems (Cloudburst, FReD/Enoki, FaDO, etc.) onto {schedules functions?, places data?, replication?, consistency in the optimizer?} would sharpen the gap claim in §2.","section":"§2"}],"recommendation":"minor_revision","confidential_remarks":"Solid systems-modeling paper with a clean BLP and two well-motivated heuristics; suitable for a DC/edge journal after tightening overstated “few percent” language and documenting AV parameters. No evidence of circular evaluation. Pure simulation and no public artifact are the main residual risks for impact, not correctness. I would not require a full real-system deployment for acceptance if the parameter and claim-qualification issues are fixed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a solid systems paper that finally treats the joint placement problem under heterogeneous consistency the way people actually need it. The new pieces are real: a BLP that encodes SR co-location (leader for writes, any replica for reads) and ER miss/routing with McCormick linearization, a Global-View greedy that expands replicas under the same latency-storage objective and stays within a few percent of the solver out to 10^4 nodes, and an Aggregated-View protocol that settles SR leaders by elastic/buoyancy force balance while ER just always-copies. The mobility experiment is the payoff—centralized policies spike on stale windows while AV stays flat—and the claim that data placement dominates scheduling is backed by the CO/CD baselines.\n\nWhat they do well: the model is carefully scoped (single SR collection per function, tree topology, cloud as permanent root copy), the evaluation reports medians/IQRs over seeds, and they are honest about the free parameters (α, τ, μ, γ, W). The force model is dimensionally consistent and the always-copy + hierarchical eviction is simple enough to implement. Related work is fair; they correctly note that prior systems either ignore data, ignore consistency distinctions, or leave replication to the operator.\n\nSoft spots are the usual genre ones, not hidden flaws. Access patterns are assumed known a priori and fixed—exactly the assumption the authors flag as future work. Pure simulation on a balanced tree, no code or traces released, and the force thresholds are hand-tuned. None of that overturns the internal claims. The BLP is correctly formulated; the scaling numbers and mobility curves are consistent with the model.\n\nThis is for people building or evaluating stateful FaaS on the continuum. If you care about consistency-aware placement or decentralized edge schedulers, read it. It deserves a serious referee; I would not desk-reject. Engage with it, cite the formulation and the mobility result, and push them on runtime access discovery and an open artifact.","headline":"Clean joint formulation of function scheduling + data placement under SR/ER, with a usable exact baseline, a scalable greedy, and a decentralized force/always-copy protocol that actually wins under mobility.","tokens_in":27869,"tokens_out":509,"would_cite":true,"duration_ms":6116,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Where you put the data matters more than where you run the functions in edge serverless.","keywords":["Function-as-a-Service","edge-cloud continuum","function scheduling","data placement","replication","consistency models","serverless"],"falsifier":"Re-run the mobility experiments after making each function's access set change at runtime (or be discovered only from observed misses); if the Aggregated-View's latency advantage over the Global-View and BLP disappears or storage thrashing rises sharply, the central claim about continuous local adaptation collapses.","tokens_in":27919,"feed_emoji":"📡","tokens_out":866,"duration_ms":8508,"temperature":0.7,"pith_summary":"Serverless at the edge only wins if state lives near the functions that use it. Once you introduce replication and different consistency rules, function placement and data placement become one coupled problem. This paper formalizes that joint decision under strong and eventual consistency, gives an exact Binary Linear Program as a yardstick, then shows a global greedy method that stays within a few percent of the optimum out to more than ten thousand nodes, and a fully local Aggregated-View method that each node can run from demand it already sees. Under client mobility the local method keeps latency low and flat while centralized re-plans spike on stale snapshots. The consistent experimental lesson is that data placement, not function scheduling, dominates the outcome.","feed_headline":"Data placement beats function scheduling at the edge","feed_subtitle":"Local replica decisions keep latency flat under mobility; global re-plans spike on stale snapshots","key_machinery":"Three nested decision procedures of decreasing information: a Binary Linear Program that jointly places replicas and schedules invocations under SR/ER constraints; a Global-View greedy that first places a minimal feasible baseline then expands replicas by rate-weighted latency savings minus storage cost; and an Aggregated-View protocol that treats SR leaders as force-equilibrium barycenters of write traffic and ER replicas as always-copy caches driven by local misses.","core_discovery":"Joint function scheduling and data-replica placement under strong and eventual consistency can be solved optimally by a Binary Linear Program for a system snapshot; a Global-View greedy heuristic that uses the same snapshot stays within a few percent of that optimum while scaling past 10^4 nodes; an Aggregated-View heuristic that uses only local demand trades some optimality for continuous per-invocation adaptation and, under client mobility, produces lower and more stable client-observed latency than the centralized policies.","pith_inferences":["If access patterns must be learned online, the Aggregated-View's always-copy and force rules already supply a natural exploration signal that the offline BLP and Global-View lack.","The same force-equilibrium idea for leaders could be reused for other single-writer resources (leader-based consensus groups, primary shards) in hierarchical edge topologies.","Storage-pressure buoyancy as a continuous force may be more portable to non-tree topologies than the hop-count LCA scheduling used here."],"forward_implications":["Edge FaaS platforms should treat data-replica placement as a first-class, continuously updated decision rather than a static configuration or pure compute-scheduling problem.","Centralized re-planning loops become liability under mobility once the gather-scatter cycle is comparable to client movement intervals; local reaction wins.","Strong-consistency collections need a single-leader force model; eventual-consistency collections can be treated as simple caches, so mixed workloads can run both mechanisms side by side.","Naive baselines that keep data in the cloud while pushing compute to the edge produce substantially higher latency than any of the joint policies, confirming that co-location of state is the dominant lever."],"fun_headline_variants":["Data placement outweighs function scheduling at the edge","Local replica choices keep latency flat under mobility","Global-View stays near optimal past 10k nodes","Aggregated-View adapts from local demand alone","Replica placement beats function scheduling in continuum"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Which collections each function reads or writes, and whether it is read-only, is known up front and stays fixed for the whole planning or simulation window.","fun_headline_variants_meta":{"raw":{"variants":["Data placement outweighs function scheduling at the edge","Local replica choices keep latency flat under mobility","Global-View stays near optimal past 10k nodes","Aggregated-View adapts from local demand alone","Replica placement beats function scheduling in continuum"]},"model":"grok-4.5","effort":"low","cost_usd":0.003872,"raw_usage":{"total_tokens":1216,"prompt_tokens":808,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":38720000,"prompt_tokens_details":{"text_tokens":808,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":335,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":808,"tokens_out":73,"duration_ms":4424,"temperature":1.0,"reasoning_tokens":335,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T10:21:50.153812+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the mobility experiments after making each function's access set change at runtime (or be discovered only from observed misses); if the Aggregated-View's latency advantage over the Global-View and BLP disappears or storage thrashing rises sharply, the central claim about continuous local adaptation collapses.","supporting_citations":[],"review_version":2}