Papers Comparison

Common backbone: the MAPE-K loop¶

Almost every paper in this review innovates in Analyze (a better forecaster) and reuses or hand-waves PLAN and EXECUTE. A minority build the whole loop.

Taxonomy¶

Pure forecasters¶

Analyze only; hand off to an external scaler.

Contribution is forecast accuracy/efficiency, not actuation.

• DAF (probabilistic job-arrival)

• FELT (one model, all containers)

• Fremer (frequency-domain)

• WGAN-gp (adversarial; thin “1 job = 1 VM” scaler).

Then “a standard HPA does the rest”.

End-to-end frameworks¶

Implement the full MAPE-K loop.

• MV-Transformer (scales VMs, live)

• AdaptiveAutoScaling (cost-optimizing controller, simulator)

• PredictiveAutoscaling (feeds KEDA, live, scales ingress)

• PredictiveK8s (Informer + simple rule, log-replay)

• CATScaler (+ LightGBM scaler, live pods)

• InformerAutoScale (Informer + own manager, live).

These papers carry all four MAPE steps and loop every 30–60 s.

Performance predictors¶

Predict slowdown, not workload.

• CloudFormer: VM degradation ratio from black-box host metrics; feeds placement/migration/scaling.

Papers notable for a special technique¶

The technique is often the whole reason to read it.

Comparison table¶

CloudFormer predicts degradation of an ongoing run (to pre-empt a QoS violation), not a fixed time-ahead forecast.

Model Training¶

Every paper in this review trains its model from scratch on cloud workload/trace data. None uses a pretrained time-series foundation model: no borrowed backbone, no fine-tuning, no zero-shot transfer. The contribution is always a task-specific architecture fit directly to the traces.

Two “false friends” worth flagging — wording that looks like pretraining but isn’t:

• Fremer uses the words “zero-shot” and “foundational backbone”, but these describe Fremer itself being trained once and transferring across datasets (via channel independence), and its potential as a future shared backbone — not the use of any external pretrained model.

• DAF calls itself a “hybrid” model, but “hybrid” means combining decomposition + diffusion + exogenous attention — not mixing pretrained and from-scratch training.

Other notables:

• WGAN-gp is the odd one out: trained adversarially, per workload (it does not generalize across traces), and with MADGRAD rather than Adam — an ablation shows that optimizer choice is the key accuracy driver.

• The closest thing to “train once, then serve” is the deploy pattern in CATScaler and PredictiveAutoscaling: train offline, serve, then periodically retrain (e.g. a Kubernetes CronJob) to fight distribution shift — still from-scratch retraining, not foundation-model fine-tuning.

MAPE Stage-by-stage¶

Monitor¶

No novelties, but richness varies. The table below lists the metrics each paper actually feeds the model and where they come from.

Takeaways:

• Prometheus recurs (PredictiveAutoscaling, CATScaler, and architecturally PredictiveK8s); InformerAutoScale instead reads Metrics Server + cAdvisor.

• Most “univariate” papers are workload-only (request/job-arrival rate); the multivariate ones add resource counters (CPU, memory, ...).

• CloudFormer is the outlier — black-box host hardware counters (206), not workload at all.

Analyze¶

• Plain transformer, made multivariate/looped: MV-Transformer, AdaptiveAutoScaling.

• Efficient long-sequence (Informer): PredictiveK8s, InformerAutoScale.

• Generative/probabilistic (diffusion): DAF, the only one with uncertainty.

• Adversarial (GAN): WGAN-gp, the only one with a critic on forecast realism.

• Frequency-domain (FFT): Fremer, the only one forecasting a spectrum.

• Convolution-augmented: CATScaler, local-spike + global-trend.

• Similarity-aware shared model: FELT, one model for thousands.

• Performance predictor: CloudFormer, predicts slowdown, not workload.

Plan¶

Where approaches diverge most sharply:

• Trivial / assumed

  • WGAN-gp (1 job = 1 VM)

  • PredictiveK8s

  • MV-Transformer

  • InformerAutoScale (pods = load / capacity)

  • Fremer

  • DAF (linear, left to HPA).

• Designed controllers

  • AdaptiveAutoScaling (α/β/γ cost optimization + damping cap)

  • CATScaler (LightGBM learns the nonlinear forecast→pod mapping, the most sophisticated Plan)

  • PredictiveAutoscaling (RPS → KEDA target).

• “Anti-flapping” logic

  • Cooldown / stabilization window: MV-Transformer’s CPT (Cool-down Period Time), CATScaler’s cooldown.

  • Reactive rescaling guard (RRS): only apply a scale-in if the lower demand persists / passes a check, rather than reacting to a single dip. MV-Transformer, InformerAutoScale (on a 60 s cadence).

  • Queue + cooldown: CATScaler runs scale requests through a FIFO queue plus a cooldown so bursts of decisions collapse into one.

  • Damping cap: AdaptiveAutoScaling limits how much capacity can change per step $|c_{t+1} − c_t| ≤ \delta_{max}$, so it can’t swing wildly.

• Not implemented

  • DAF

  • FELT

  • CloudFormer

Execute¶

• Live K8s/Docker

  • PredictiveAutoscaling (KEDA → HPA)

  • CATScaler (pod replicas)

  • InformerAutoScale & MV-Transformer (Docker Desktop + K8s)

• Simulator / log-replay

  • AdaptiveAutoScaling (discrete-time sim)

  • PredictiveK8s (offline replay)

• Real cloud VMs

  • WGAN-gp (Google Cloud e2-medium)

• No actuation (forecast-only)

  • DAF

  • FELT

  • Fremer (except 24h HPA demo)

  • CloudFormer

In one sentence each¶

• MV-Transformer — multivariate forecaster, full loop, scales real VMs.

• AdaptiveAutoScaling — forecaster + cost-optimizing controller, in simulation.

• PredictiveAutoscaling — scales the ingress controller via KEDA, live.

• PredictiveK8s — first to bring Informer to K8s autoscaling.

• DAF — probabilistic forecast with a confidence band.

• FELT — one tiny model for thousands of containers via similarity-aware attention.

• Fremer — frequency-domain, ~12x smaller, multi-period workloads.

• WGAN-gp — adversarially-trained, fast, bursty-traffic forecaster.

• CATScaler — convolution-augmented forecast + LightGBM pod model, full live loop.

• InformerAutoScale — Informer forecaster + own actuator, ~10x fewer pods than reactive.

• CloudFormer — predicts slowdown from black-box host metrics, not workload.

Acronyms & short names¶

Models / methods named in this review¶

Domain & infrastructure terms¶