{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp11_mechanism_context_external20_20260804","research_id":"eoa_inverse_innovation_exp11_external_scrutiny_20260804","cell_id":"invariant_mode_decomposition_design__computer_science","opaque_id":"invariant_mode_decomposition_design__computer_science__C","search_lanes":{"direct_problem":{"queries":["microservice autoscaling oscillation coupled services retry storm queues control research","dependency-aware autoscaling microservices coordinated scaling research paper","microservice autoscaling \"eigenvalue\" stability controller"],"source_ids":["S1","S4","S5","S6"],"no_result_note":null},"closest_prior_art":{"queries":["\"dynamic mode decomposition\" microservices autoscaling","eigenmodes state transition operator cloud service autoscaling control","microservice \"dynamic mode decomposition\" telemetry","autoscaling \"spectral gap\" microservices control"],"source_ids":["S3","S4","S5","S6","S7","S8"],"no_result_note":"No retained empirical source implemented the complete package of local transition-operator estimation, unstable-eigenmode selection, mode-to-service control mapping, residual and drift checks, and reversible microservice actuation."},"historical_terminology":{"queries":["autonomic computing multi-tier application MIMO control resource allocation eigenvalue stability","feedback control scheduling web server oscillation admission control 2005 autoscaling","multi-tier internet service joint resource provisioning feedback controller queues 2008","dynamic provisioning multi-tier applications model predictive control coupled"],"source_ids":["S3","S7"],"no_result_note":null},"products_practices_standards":{"queries":["site:kubernetes.io autoscaling oscillation stabilization window HPA official","site:aws.amazon.com builders library retries cascading failures overload official","site:cloud.google.com architecture autoscaling cascading failure retry storm official","site:sre.google/sre-book cascading failures retries queue overload"],"source_ids":["S1","S2"],"no_result_note":null},"non_english_regional":{"queries":["微服务 自动扩缩容 震荡 级联 故障 服务依赖","微服务 弹性伸缩 振荡 耦合 队列 重试风暴","マイクロサービス オートスケーリング 振動 依存関係","Microservices automatische Skalierung Schwingung Abhängigkeiten Regelung"],"source_ids":["S2","S3"],"no_result_note":"Chinese, Japanese, and German terminology searches found regional guidance and research on retry storms, coordinated overload control, and autoscaler stabilization, but no substantiated regional implementation of eigenmode-directed microservice control."},"composition_subproblems":{"queries":["cloud autoscaling \"system identification\" eigenvalues MIMO","microservice \"dynamic mode decomposition\" telemetry","microservice autoscaling \"eigenvalue\" stability controller","AutoControl online model estimator MIMO controller multi-tier application 2009"],"source_ids":["S5","S6","S7","S8"],"no_result_note":"The component methods are documented separately: online MIMO modeling and control, dynamic-mode estimation with control inputs, dependency learning, coordinated scaling, and oscillation mitigation. Their proposed composition was not found as an evaluated microservice system."}},"sources":[{"source_id":"S1","title":"Addressing Cascading Failures","url":"https://sre.google/sre-book/addressing-cascading-failures/","publisher":"Google Site Reliability Engineering","date_or_year":"2016","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["Cascading failures arise through positive feedback among overloaded components.","Queue growth, missed deadlines, retries, resource exhaustion, health actions, and load redistribution can amplify one another.","Retries can destabilize a service and have contributed to cascading failures.","Service owners and SRE teams are operational actors; load testing, gradual canaries, load shedding, retry budgets, and rollback-oriented incident practices are recognized safeguards."]},{"source_id":"S2","title":"HorizontalPodAutoscaler v2 API reference","url":"https://kubernetes.io/docs/reference/kubernetes-api/autoscaling/horizontal-pod-autoscaler-v2/","publisher":"Kubernetes","date_or_year":"2026","source_type":"OFFICIAL_STANDARD","language":"English","claims_supported":["A mainstream autoscaler calculates desired replicas from configured metrics and applies scaling rules independently to a target resource.","Scaling velocity limits, metric tolerance, and stabilization windows are official mechanisms for preventing flapping.","The default scale-down behavior includes a 300-second stabilization window, showing that oscillatory scaling is an acknowledged operational concern."]},{"source_id":"S3","title":"Overload Control for Scaling WeChat Microservices","url":"https://arxiv.org/abs/1806.04075","publisher":"arXiv; research from WeChat/Tencent and collaborating institutions","date_or_year":"2018","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Service-specific overload control can harm the whole system because of intricate service dependencies.","DAGOR performs collaborative load shedding among related microservices rather than relying only on isolated service action.","The reported system-centric control had been used in the WeChat backend for five years, demonstrating an identifiable production adopter and established coordinated overload intervention."]},{"source_id":"S4","title":"PBScaler: A Bottleneck-aware Autoscaling Framework for Microservice-based Applications","url":"https://arxiv.org/abs/2303.14620","publisher":"arXiv","date_or_year":"2023","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Performance anomalies propagate through microservice interactions and can obscure the actual bottleneck.","Repeated online autoscaling optimization attempts can cause replica and latency oscillations and performance degradation.","PBScaler uses topology-aware bottleneck ranking and offline optimization to reduce unnecessary scaling, and was evaluated on microservice benchmarks."]},{"source_id":"S5","title":"DeepScaler: Holistic Autoscaling for Microservices Based on Spatiotemporal GNN with Adaptive Graph Learning","url":"https://arxiv.org/abs/2309.00859","publisher":"arXiv; ASE research paper","date_or_year":"2023","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Complex, time-varying service dependencies can cause cascading effects during resource allocation.","DeepScaler learns latent service dependencies and spatiotemporal behavior from graph-structured telemetry.","It simultaneously reconfigures interacting services and reports lower SLA violations and costs than evaluated alternatives."]},{"source_id":"S6","title":"GraphPilot: A Temporal Graph Actor-Critic Autoscaler Reducing Degradation of Resource Oscillation in Microservice","url":"https://scholars.lib.ntu.edu.tw/entities/publication/c065f4f0-3271-4381-807d-004d81c6d1fd","publisher":"ACM and National Taiwan University","date_or_year":"2025","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Repeated up/down resource scaling is treated as a specific microservice performance and cost problem.","GraphPilot attributes such oscillation to delayed cross-service dependency propagation under abrupt workload changes or prediction errors.","The evaluated intervention combines temporal dependency graphs with coordinated proactive scaling, making it the closest same-problem rival found."]},{"source_id":"S7","title":"Automated Control of Multiple Virtualized Resources","url":"https://research.google/pubs/automated-control-of-multiple-virtualized-resources/","publisher":"EuroSys/ACM; hosted by Google Research","date_or_year":"2009","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["AutoControl is historical prior art for online estimation and multi-input, multi-output control of coupled computing resources.","It maps application-performance observations and multiple resource allocations into coordinated control decisions across multi-tier applications and nodes.","Experiments used RUBiS, TPC-W, and production-trace-driven workloads, showing a testable predecessor to modern coordinated autoscaling even though it did not prioritize eigenmodes."]},{"source_id":"S8","title":"Dynamic Mode Decomposition with Control","url":"https://arxiv.org/abs/1409.6358","publisher":"arXiv; journal version in SIAM Journal on Applied Dynamical Systems","date_or_year":"2014/2016","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["DMD with control estimates a low-order state-transition operator and input map from state and actuation snapshots.","The method extracts dynamic modes and can support controller design without known governing equations.","The paper warns that ordinary DMD confounds internal dynamics with external forcing, directly supporting the need to include control inputs and to test model validity."]}],"problem_evidence":{"status":"SUPPORTED","finding":"Independent official guidance, platform documentation, and primary research agree that queues, retries, dependencies, resource controls, and delayed scaling decisions can form positive-feedback cascades or repeated scale-up/scale-down behavior. The evidence supports the operational problem, although not every incident will have a persistent modal structure.","source_ids":["S1","S2","S3","S4","S5","S6"],"uncertainty":"The sources establish several causal classes—overload, retry amplification, delayed dependency propagation, and optimization-induced oscillation—but do not establish that one locally linear coupled-mode model explains most incidents."},"adopter_evidence":{"status":"SUPPORTED","finding":"Platform/SRE teams, service owners, data-center administrators, and application operators are identifiable users and authorizers. Google assigns cascading-failure prevention to service owners and SREs; Kubernetes exposes operator-configured scaling behavior; DAGOR reports multi-year production adoption at WeChat.","source_ids":["S1","S2","S3","S7"],"uncertainty":"Joint authorization boundaries between a central platform team and independently owned services remain organization-specific."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"All major ingredients have precedents, but not as one evaluated package. DMDc supplies input-aware modal system identification; AutoControl supplies online MIMO resource modeling and control; DeepScaler and GraphPilot supply dependency-aware coordinated scaling; PBScaler supplies topology-aware bottleneck control; Google and Kubernetes supply gradual-change and stabilization safeguards.","source_ids":["S1","S2","S4","S5","S6","S7","S8"],"uncertainty":"No retained source demonstrates that stable, separated eigenmodes can be estimated from production microservice telemetry quickly enough, mapped reliably to safe controls, and retained after intervention."},"prior_art":{"disposition":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"GraphPilot temporal dependency autoscaler","source_ids":["S6"],"same_problem":true,"same_causal_lever":false,"overlap":"Targets resource oscillation caused by delayed cross-service dependency propagation and makes coordinated proactive scaling decisions.","remaining_difference":"Uses temporal graph embeddings and actor-critic control rather than estimating a local transition operator, selecting growing eigenmodes, checking modal residuals and drift, and intervening in modal coordinates."},{"name":"PBScaler bottleneck-aware autoscaling","source_ids":["S4"],"same_problem":true,"same_causal_lever":false,"overlap":"Recognizes interaction-propagated anomalies and optimization-induced replica oscillation, then reduces unnecessary scaling through topology-aware bottleneck selection.","remaining_difference":"Ranks bottleneck services with a random-walk method and offline optimization; it does not identify persistent or amplified cross-service dynamic modes."},{"name":"DeepScaler holistic coordinated autoscaling","source_ids":["S5"],"same_problem":true,"same_causal_lever":false,"overlap":"Learns changing service dependencies and simultaneously reallocates resources across interacting services to avoid cascading effects.","remaining_difference":"Predicts resource needs with adaptive graph learning and a spatiotemporal GNN rather than using eigenvalues and eigenvectors of an interpretable local state-transition model."},{"name":"AutoControl online MIMO resource controller","source_ids":["S7"],"same_problem":false,"same_causal_lever":false,"overlap":"Historically established online black-box estimation and coordinated multi-input, multi-output control for multi-tier computing resources under changing workload.","remaining_difference":"Targets application SLO regulation and bottleneck allocation, not hidden autoscaling/retry modes; it does not use modal gain, spectral separation, mode drift, or modal intervention mapping."},{"name":"DAGOR collaborative overload control","source_ids":["S3"],"same_problem":true,"same_causal_lever":false,"overlap":"Treats isolated service controls as potentially system-harming and coordinates load shedding across dependent microservices in production.","remaining_difference":"Uses distributed overload signaling and business-priority-aware shedding rather than learned dynamic modes or modal gain estimates."},{"name":"Dynamic Mode Decomposition with Control","source_ids":["S8"],"same_problem":false,"same_causal_lever":true,"overlap":"Provides the proposed mathematical core: estimate a transition operator and input map from snapshots, extract modes, and use the resulting low-order model for control design.","remaining_difference":"Is a domain-general method, not a microservice autoscaling implementation, and does not supply topology validity contracts, service-control mappings, protected SLO outcomes, or production rollback governance."}],"contrastive_claim_remaining":"Within a predeclared microservice dependency neighborhood and operating window, input-aware local operator estimation will reveal reproducible, spectrally separated growing or weakly damped modes whose coordinated suppression reduces replica oscillation, retry amplification, tail latency, and recovery time more than GraphPilot-style dependency control, PBScaler-style bottleneck control, or dependency-aware runbooks at comparable cost and error risk.","contrastive_claim_falsifier":"The claim fails if per-service variables plus recorded exogenous events predict propagation as well as the coupled model with no reproducible residual mode; if estimated modes or gains drift before action; if no separated controllable risky mode appears; or if mode-directed control does not improve pre-registered recovery, oscillation, latency, error, and cost outcomes versus the nearest rival.","confidence":"MODERATE","search_limitations":"The bounded search covered current and historical control terminology, official platform practice, regional Chinese/Japanese/German terms, and component combinations. Rapidly evolving 2025–2026 literature, abstract-only conference records, unindexed internal systems, patents, and proprietary incident data may contain closer work. A generic web page describing eigenvalue-based Kubernetes tuning was not retained because it offered no primary implementation or evaluation. This search cannot establish world novelty or routine absence."},"researchability_gates":{"externally_supported_problem":{"status":"PASS","rationale":"Official operations guidance and multiple independent empirical research sources document positive-feedback cascades, retry amplification, dependency propagation, autoscaler flapping, and resource oscillation.","source_ids":["S1","S2","S3","S4","S5","S6"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"Service owners, SRE/platform teams, Kubernetes operators, and data-center administrators are identifiable; WeChat provides a concrete production adopter of coordinated microservice overload control.","source_ids":["S1","S2","S3","S7"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"Close systems coordinate scaling or shedding using topology, learned graphs, bottleneck ranking, or MIMO control, but none of the retained sources evaluates unstable-mode identification plus modal actuation in microservices. The remaining claim has explicit comparative outcomes and falsifiers.","source_ids":["S4","S5","S6","S7","S8"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"The proposal can first be tested in shadow mode on synchronized incident and controlled-load traces, then in a low-traffic reversible canary against dependency-aware baselines. Existing benchmark evaluations, input-aware DMD, stabilization controls, and canary guidance make this a bounded experiment.","source_ids":["S1","S2","S4","S5","S6","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"The first step need not actuate production. A later canary can be jointly authorized, rate-limited, and automatically rolled back. Official guidance specifically recommends slow canaries and warns that remedial actions can amplify cascades; these are guardrails rather than a categorical stop.","source_ids":["S1","S2"]},"adequate_search_evidence":{"status":"PASS","rationale":"All six required lanes were searched adversarially using direct, historical, product/standard, multilingual/regional, and component-combination terminology. Exactly eight retained sources span Google, Kubernetes, ACM/EuroSys, NTU, and several independent research groups, with seven primary or official sources.","source_ids":["S1","S2","S3","S4","S5","S6","S7","S8"]}},"strict_success":true,"screen_survival":true,"remaining_research_value":"MODERATE","recommended_next_step":"Run a pre-registered two-stage experiment: first fit an input-aware local operator in shadow mode on one benchmark dependency neighborhood using synchronized queues, latency, retries, throughput, replicas, and recorded control actions; compare next-state prediction and incident explanation against per-service plus exogenous-event models. Proceed only if a risky mode is reproducible, spectrally separated, low-residual, and controllable. Then use an expiring low-traffic canary to compare mode-directed coordinated scaling/retry limiting with GraphPilot-style dependency control or PBScaler-style bottleneck control, measuring pod churn, retry amplification, p99 latency, error rate, recovery time, cost, residual error, and mode drift, with automatic restoration of the prior configuration at pre-registered limits.","world_novelty_boundary":"This bounded public-web review supports an adjacent-prior-art disposition and a distinct comparative experiment; it does not establish world novelty, patentability, freedom to operate, market size, routine absence from proprietary systems, or realized operational impact."}