Pith. sign in

REVIEW 3 major objections 6 minor 31 references

Data Replication Meets Function Scheduling in the Edge-Cloud Continuum

T0 review · 3 major / 6 minor · reviewed 2026-07-15 · grok-4.5

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

Pith's one-line read Where you put the data matters more than where you run the functions in edge serverless.

desk verdict 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. read the letter →

arxiv 2606.30563 v2 pith:R6RXRWXG submitted 2026-06-29 cs.DC

Matteo Cenzato , Dario d'Abate , Arianna Dragoni , Matteo Briscini , Alessandro Margara This is my paper · ORCID
classification cs.DC
keywords Function-as-a-Serviceedge-cloudcontinuumfunctionschedulingdataplacementreplicationconsistencymodelsserverless
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 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.

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 (3)
  1. [Abstract, §7.3, Fig. 4] 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.
  2. [§6.2.1, §7.1, Fig. 5b] §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.
  3. [§7.4, §7.5] §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.
minor comments (6)
  1. [§3.2, §8] §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.
  2. [§7] 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.
  3. [Fig. 3, Fig. 4, Fig. 5] 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.
  4. [Abstract, §1] Abstract and early pages contain many missing spaces after punctuation (e.g., “Serverlesscomputing”, “BinaryLinearProgram”). Clean typesetting before camera-ready.
  5. [§4.2, Eq. (6)] §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).
  6. [§2] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: BLP, GV, and AV are independently defined algorithms evaluated on generated instances; outcomes are not forced by construction from fitted inputs.

full rationale

This is a systems optimization paper, not a first-principles prediction paper. The derivation chain is: (1) formalize joint placement under SR/ER as a BLP with an explicit weighted latency–storage objective (§4, Eq. 6); (2) define a Global-View greedy that approximates the same objective from a full snapshot (§5); (3) define an Aggregated-View protocol that places/schedules from local reachability and force/always-copy rules (§6); (4) run all three plus naive baselines on synthetic tree instances and report measured T, S, solve time, and client-observed latency (§7). None of the central claims reduces to its inputs by definition. Free parameters (α, τ_L, τ_R, μ, γ, W) are author-chosen and either swept or fixed; they are not fitted to a subset of the reported superiority metrics and then re-presented as predictions. Related-work citations (including the external spring analogy of Pietzuch et al.) supply motivation or analogy, not load-bearing uniqueness theorems from overlapping authors. The evaluation simply executes the stated algorithms and records outcomes; that is ordinary empirical comparison, not circularity. Score 0 is therefore the correct finding.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claims rest on a standard hierarchical-tree model of the continuum, two classical consistency protocols, and a handful of free parameters that control the latency-storage trade-off and the force thresholds. No new physical entities are postulated; the 'virtual forces' are purely algorithmic constructs. The largest modeling assumptions are static access patterns and the absence of compute or failure dynamics.

free parameters (5)
  • α (latency-storage trade-off) = 1 / 0.1 / 0.01
    Scalar that multiplies normalized storage cost in the BLP and GV objective; varied over {1,0.1,0.01} but chosen by hand.
  • τ_L, τ_R (hysteresis thresholds)
    Prevent oscillation of leader and follower decisions in the force model; set by authors.
  • μ (buoyancy scale)
    Aligns storage-pressure force with elastic force units; free constant.
  • γ (exponential decay rate)
    Traffic-statistic forgetting factor; free.
  • W (sliding-window length) = 32 s
    32 s observation window for centralized rate estimation; chosen so that ≥70 % of clients emit at least one request.
assumptions (5)
  • domain assumption Infrastructure is a rooted tree; latency and bandwidth are path-additive / path-min.
    §3.1; standard continuum model but excludes general graphs and multi-path routing.
  • domain assumption Each function accesses at most one SR collection; writes must co-locate with the single leader.
    §3.3; motivated by avoiding cross-collection 2PC but restricts the application model.
  • domain assumption A permanent copy of every collection always exists at the cloud root.
    §3.3 fault-tolerance baseline; simplifies recovery but is an idealized assumption.
  • ad hoc to paper Function-to-collection access sets and read/write labels are known statically.
    §3.2; required for both BLP and heuristics; not discovered at runtime.
  • domain assumption Cold-start and per-node compute capacity are ignored; only data-access latency matters.
    §3.2; common simplification but may not hold under heavy load.
invented entities (1)
  • Virtual elastic and buoyancy forces for SR leader/follower placement
    purpose: Provide a local, continuous decision rule that approximates the global latency-storage trade-off without a snapshot.
    Purely algorithmic analogy (spring + fluid expulsion); no claim of physical reality; independent_evidence false because the forces exist only inside the protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data Replication Meets Function Scheduling in the Edge-Cloud Continuum." pith.science (2026). https://pith.science/paper/R6RXRWXG

@misc{pith2026260630563,
  author       = {Pith},
  title        = {Pith review of: Data Replication Meets Function Scheduling in the Edge-Cloud Continuum},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6RXRWXG}},
  note         = {Machine review of arXiv:2606.30563}
}
abstract

Serverless computing is an appealing model for the edge-cloud continuum, but its stateless assumption breaks down once functions need persistent data: fetching state from a distant cloud store erases the latency benefit of running at the edge. Keeping data close means replicating it, and replication forces a placement decision that is coupled with where functions execute and with the consistency each application demands. We study this joint problem of function scheduling and data placement under two consistency models, strong and eventual replication. We first formulate it as a Binary Linear Program that yields the optimal placement for a given system snapshot, and use it as a reference point. Because the solver does not scale past a few hundred nodes, we add two heuristics with progressively less information: a Global-View greedy method that works from the same complete snapshot, and an Aggregated-View heuristic in which each node decides from locally observed demand alone. Across a range of system sizes the Global-View heuristic stays within a few percent of the optimum while scaling to over $10^4$ nodes. The Aggregated-View heuristic sacrifices some solution quality, but adapts continuously to each invocation. Under client mobility, centralized policies suffer from stale snapshots and recurring latency spikes, while the Aggregated-View maintains low and stable client-observed latency. Across all experiments, data placement proves more influential than function scheduling in determining the outcome.

Figures

Figures reproduced from arXiv: 2606.30563 by the authors.

Figure 1
Figure 1. Example of the local state maintained by each node (b) and its tight coupling with runtime operations (a). The replica store (𝑗 ) maps to collection persisted onto nodes, while the reachability table (𝜌𝑗 , 𝓁𝑗 ) dictates the known directions to data. Invocations drive function scheduling by querying this local state, while the resulting traffic metrics and data misses concurrently feed back into the data placement a… view at source ↗
Figure 2
Figure 2. Proactive data placement model for SR collections. (a) A replica is subject to opposing virtual forces: elastic forces pull it toward the read and write demand it serves, while a buoyant force expels it from nodes whose storage is nearly full. (b) Recursive evaluation for leader (★) delegation. Node 𝑗 weighs the elastic pull from a write-heavy child 𝑘 against opposing sibling forces and the tension 𝐹 𝑊 ext,𝑗 inherit… view at source ↗
Figure 3
Figure 3. Scalability of BLP and GV across problem dimensions. GV is consistently faster: 0.20 s (MIX), 0.08 s (SR), and be￾low 0.01 s (ER). Collections, by contrast, are harder to scale. Each additional collection adds || placement variables (𝑥𝑐,𝑗) together with storage and consistency constraints that couple it to every other collection on the same node. BLP exhibits high variance under MIX, with a median of ≈2 s but IQR p… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: 𝑇 and 𝑆 difference relative to BLP when scaling collections. 7.3. Solution Approximation To answer RQ2, we reuse the same problem instances as RQ1, but instead of their scaling behavior we now examine the quality of the placement each policy produces. We eval￾uate GV a…
Figure 5
Figure 5. Figure 5: Client-observed latency comparison across different access patterns. As 𝛼 decreases from 1 to 0.01, storage becomes cheaper in the objective, and all policies replicate more aggressively. At 𝛼 = 0.01, GV and AV both achieve latency within a few percentage points of BLP…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references

  1. [1]

    Resource provisioning and allocation in function- as-a-service edge-clouds

    Ascigil, O., Tasiopoulos, A.G., Phan, T.K., Sourlas, V., Psaras, I., Pavlou, G., 2022. Resource provisioning and allocation in function- as-a-service edge-clouds. TSC 15

    2022

  2. [2]

    NEP- TUNE: A comprehensive framework for managing serverless func- tions at the edge

    Baresi, L., Hu, D.Y.X., Quattrocchi, G., Terracciano, L., 2024. NEP- TUNE: A comprehensive framework for managing serverless func- tions at the edge. TAAS 19

    2024

  3. [3]

    Navigating the edge-cloud continuum: A state-of-practice survey

    Belcastro, L., Marozzo, F., Orsino, A., Talia, D., Trunfio, P., 2026. Navigating the edge-cloud continuum: A state-of-practice survey. IEEE Access 14

    2026

  4. [4]

    The computing continuum: Past, present, and future

    Bittencourt, L.F., Rodrigues-Filho, R., Spillner, J., De Turck, F., Santos, J., da Fonseca, N.L., Rana, O., Parashar, M., Foster, I., 2025. The computing continuum: Past, present, and future. Comp. Science Review 58

    2025

  5. [5]

    A decentralized framework for serverless edge computing in the internet of things

    Cicconetti, C., Conti, M., Passarella, A., 2021. A decentralized framework for serverless edge computing in the internet of things. TNSM 18

    2021

  6. [6]

    Cloud-native applica- tions

    Gannon, D., Barga, R., Sundaresan, N., 2017. Cloud-native applica- tions. Cloud Comp. 4

    2017

  7. [7]

    Cloud ProgrammingSimplified:ABerkeleyViewonServerlessComputing

    Jonas, E., Schleier-Smith, J., Sreekanti, V., Tsai, C.C., Khandelwal, A.,Pu,Q.,Shankar,V.,MenezesCarreira,J.,Krauth,K.,Yadwadkar, N., Gonzalez, J., Popa, R.A., Stoica, I., Patterson, D.A., 2019. Cloud ProgrammingSimplified:ABerkeleyViewonServerlessComputing. Technical Report UCB/EECS-2019-3

    2019

  8. [8]

    Designing Data-Intensive Applications: The Big Ideas Behind Reliable, Scalable, and Maintainable Systems

    Kleppmann, M., 2017. Designing Data-Intensive Applications: The Big Ideas Behind Reliable, Scalable, and Maintainable Systems. O’Reilly Media, Inc

    2017

Show all 31 references
  1. [9]

    Li, Q., Kraft, P., Cafarella, M., Çağatay Demiralp, Graefe, G., Kozyrakis, C., Stonebraker, M., Suresh, L., Yu, X., Zaharia, M.,

  2. [10]

    R3: Record-replay-retroaction for database-backed applica- tions. Proc. VLDB Endow. 16

  3. [11]

    Computability of global solutions to fac- torable nonconvex programs

    McCormick, G.P., 1976. Computability of global solutions to fac- torable nonconvex programs. Mathematical Programming 10

    1976

  4. [12]

    Function offloading and data migration for stateful serverless edge computing, in: ICPE, ACM

    Nardelli, M., Russo, G.R., 2024. Function offloading and data migration for stateful serverless edge computing, in: ICPE, ACM

    2024

  5. [13]

    Aserver- less real-time data analytics platform for edge computing

    Nastic,S.,Rausch,T.,Scekic,O.,Dustdar,S.,Gusev,M.,Koteska,B., Kostoska,M.,Jakimovski,B.,Ristov,S.,Prodan,R.,2017. Aserver- less real-time data analytics platform for edge computing. Internet Comp. 21

    2017

  6. [14]

    Enoki:StatefuldistributedFaaS from edge to cloud, in: MiddleWEdge, ACM

    Pfandzelter,T.,Bermbach,D.,2023. Enoki:StatefuldistributedFaaS from edge to cloud, in: MiddleWEdge, ACM

    2023

  7. [15]

    Pfandzelter,T.,Japke,N.,Schirmer,T.,Hasenburg,J.,Bermbach,D.,

  8. [16]

    Software: Practice and Experience 53

    Managing data replication and distribution in the fog with FReD. Software: Practice and Experience 53

  9. [17]

    Network-aware operator placement for stream- processing systems, in: ICDE, IEEE

    Pietzuch, P., Ledlie, J., Shneidman, J., Roussopoulos, M., Welsh, M., Seltzer, M., 2006. Network-aware operator placement for stream- processing systems, in: ICDE, IEEE

    2006

  10. [18]

    Puliafito, C., Cicconetti, C., Conti, M., Mingozzi, E., Passarella, A.,

  11. [19]

    remote state allocation for micro-services in the cloud-edge continuum

    Balancing local vs. remote state allocation for micro-services in the cloud-edge continuum. Perv. and Mobile Comp. 93

  12. [20]

    Edgefaasbench:Benchmarkingedgedevicesusingserverlesscomput- ing, in: 2022 IEEE EDGE

    Rajput, K.R., Kulkarni, C.D., Cho, B., Wang, W., Kim, I.K., 2022. Edgefaasbench:Benchmarkingedgedevicesusingserverlesscomput- ing, in: 2022 IEEE EDGE

    2022

  13. [21]

    Optimized container scheduling for data-intensive serverless edge computing

    Rausch, T., Rashed, A., Dustdar, S., 2021. Optimized container scheduling for data-intensive serverless edge computing. FGCS 114

    2021

  14. [22]

    Conflict- free replicated data types, in: SSS, Springer-Verlag, Berlin, Heidel- berg

    Shapiro,M.,Preguiça,N.,Baquero,C.,Zawirski,M.,2011. Conflict- free replicated data types, in: SSS, Springer-Verlag, Berlin, Heidel- berg

    2011

  15. [23]

    Edge computing: Vision and challenges

    Shi, W., Cao, J., Zhang, Q., Li, Y., Xu, L., 2016. Edge computing: Vision and challenges. IoT Jour. 3

    2016

  16. [24]

    FaDO: FaaS functions and data orchestrator for multiple serverless edge-cloud clusters, in: ICFEC, IEEE

    Smith, C.P., Jindal, A., Chadha, M., Gerndt, M., Benedict, S., 2022. FaDO: FaaS functions and data orchestrator for multiple serverless edge-cloud clusters, in: ICFEC, IEEE

    2022

  17. [25]

    Cloudburst:Statefulfunctions- as-a-service

    Sreekanti, V., Wu, C., Lin, X.C., Schleier-Smith, J., Gonzalez, J.E., Hellerstein,J.M.,Tumanov,A.,2020. Cloudburst:Statefulfunctions- as-a-service. VLDB 13

    2020

  18. [26]

    Lambdata:Optimizingserverlesscomputing by making data intents explicit, in: CLOUD, IEEE

    Tang,Y.,Yang,J.,2020. Lambdata:Optimizingserverlesscomputing by making data intents explicit, in: CLOUD, IEEE

    2020

  19. [27]

    Energy- efficient resource management for real-time applications in faas edge computing platforms, in: UCC, ACM

    Vahabi, S., Righetti, F., Vallati, C., Tonellotto, N., 2023. Energy- efficient resource management for real-time applications in faas edge computing platforms, in: UCC, ACM

    2023

  20. [28]

    Low-latency state management for real-time tasks in edge serverless, in: ISPA, IEEE

    Wen, Y., Xu, G., Wang, J., Hao, W., 2024. Low-latency state management for real-time tasks in edge serverless, in: ISPA, IEEE

    2024

  21. [29]

    Anna:Akvsforany scale, in: ICDE

    Wu,C.,Faleiro,J.,Lin,Y.,Hellerstein,J.,2018. Anna:Akvsforany scale, in: ICDE

    2018

  22. [30]

    Stateful serverless application placement in MEC with function and state dependencies

    Xu,Z.,Zhou,L.,Liang,W.,Xia,Q.,Xu,W.,Ren,W.,Ren,H.,Zhou, P., 2023. Stateful serverless application placement in MEC with function and state dependencies. ToC 72

    2023

  23. [31]

    Cenzato et al.:Preprint submitted to ElsevierPage 18 of 18

    Yu,M.,Cao,T.,Wang,W.,Chen,R.,2023.Followingthedata,notthe function: Rethinking function orchestration in serverless computing, in: NSDI, USENIX Association. Cenzato et al.:Preprint submitted to ElsevierPage 18 of 18

    2023

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.