Argus: A Real-EKS Study of When Predicting Spot Interruptions Beats Simple Checkpointing

Angshuman Chakravertty; MD Rayyan

cs.LGSep 30, 2026

Argus: A Real-EKS Study of When Predicting Spot Interruptions Beats Simple Checkpointing

Authors: Angshuman Chakravertty, MD Rayyan

Organizations: SVKM’s NMIMS (Deemed to be University)

Abstract

Elastic Compute Cloud (EC2) Spot is 60% to 90% cheaper than On-Demand but can be reclaimed on just a 2-minute notice; for expensive multi-node training this loss can be severe, with one reclaim costing hours of synchronous progress. We build Argus, a Kubernetes operator, and ask empirically, on a CIFAR-10 testbed, when predicting interruptions beats simple checkpointing. Argus on real EKS survives a real Spot drain with a graceful SIGTERM checkpoint, resuming from epoch 8 and losing only the in-progress epoch. Alongside, we further find that in an 80-trial benchmark, the reactive-on-notice degrades toward no protection once interruption outpaces the fixed 2-minute notice, and predictive wasted compute is driven to zero, but with an oversized fixed lead it over-migrates so severely that at the fastest rate only one of five runs completes, while periodic is a strong ML-free baseline. A lead-time sweep turns the lead prediction into a guideline where a small lead suffices for zero waste, but excess lead is wasteful. The predictor built is advisory (a proxy label); real interruption labels and large-model-scale validation are future work.

Figures & tables

Figure 1: Argus architecture. Two triggers (predictive, when the risk score crosses the threshold; reactive, the real 2-minute Spot warning via NTH) drive one SIGTERM, checkpoint, reschedule, resume flow.

Arm	Compl.	Wasted (s)	Makespan (s)	Ckpts
No-protection	1.0	$202.3\pm 141.1$	361.8	0.0
Reactive-on-notice	1.0	$169.4\pm 105.8$	328.0	4.8
Periodic	1.0	$4.3\pm 3.4$	158.8	24.0
Predictive (Argus)	0.2	$\mathbf{0.0\pm 0.0}$	353.2	221.6

Table 1: Results at the fastest tested rate (MTBF = 120 s, 5 repetitions).

Figure 2: Lead-time sensitivity. Wasted compute stays at zero across all leads, while makespan and checkpoint count climb once the lead exceeds the interruption interval.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Time (UTC)	Event
12:26:18	NTH receives EC2 Spot Instance Interruption Warning
12:26:19	Drain evicts pod cifar10-test (graceful SIGTERM)
12:26:19	SIGTERM handler writes checkpoint (epoch 8) to S3
12:26:25	Node cordoned and drained in 10 s
$\sim$ 75 s later	Replacement pod on healthy node: “Resuming from epoch 8”

Appendix

Table 2: Real-EKS Spot interruption timeline.

Figure 3: The 3.8 MB latest_checkpoint.pt written to the real S3 bucket during the Spot drain (Table 2 ), direct evidence the checkpoint survived the interruption, not a simulation.

Lead (s)	Makespan (s)	Checkpoints
2	73.1	5.4
5	76.2	7.4
10	78.0	8.8
15	83.3	12.6
20	87.0	15.2
30	96.7	21.8

Appendix

Table 3: Lead-time sensitivity (predictive arm, 20 s mean interruption interval, 5 repetitions). Periodic baseline: 4.3 s wasted, 78.1 s makespan, 29.0 checkpoints.

Figure 4: Wasted compute by arm across all four tested interruption rates.

Figure 5: Argus observability: predicted risk crosses the trigger threshold and a proactive checkpoint fires at that instant (real operator, real S3/SQS).

Explore similar work

Learning to Remember: Attentive Reinforcement Learning for Edge Serverless Autoscaling
Mar 21, 2026Faraz Shaikh, Gianluca Reali, Mauro FemminellaEdge Platforms Rate Scaling
Department of Engineering, University of Perugia, Perugia, Italy · Consorzio Nazionale Interuniversitario per le Telecomunicazioni (CNIT), 43124 Parma, Italy
Spotlight: Synergizing Seed Exploration and Spot GPUs for DiT RL Post-Training
Jun 17, 2026Ruiqi Lai, Dakai An, Wei Gao +6Diffusion-Based Reinforcement Learning Methods Reinforcement Learning Post-Training
NTU Singapore · Hong Kong University of Science and Technology · Alibaba Group
ARGOS: Reinforcement Learning-Driven Multidimensional Elasticity for Service Orchestration in the Computing Continuum
Sep 29, 2026Javier Mateos-Bravo, Sergio Laso, Juan Luis Herrera +3Rate Scaling Orchestration
Department of Computer Science and Telematics Engineering, University of Extremadura, Spain · Global Process and Product Improvement S.L., Spain · Department of Mechatronics, University of Prishtina, Kosova +2

Mar 21, 2026, cs.DCJ/K move · Enter open · S save

Learning to Remember: Attentive Reinforcement Learning for Edge Serverless Autoscaling

Faraz Shaikh, Gianluca Reali, Mauro Femminella

Department of Engineering, University of Perugia, Perugia, Italy · Consorzio Nazionale Interuniversitario per le Telecomunicazioni (CNIT), 43124 Parma, Italy

Edge Platforms Rate Scaling Cloud Workload Post-Deployment Adaptation Edge+1

Open paper AskPDF arXiv

In edge computing, the stochastic and bursty nature of serverless workloads challenges autonomous resource orchestration. Traditional reactive controllers, such as the Kubernetes Horizontal Pod Autoscaler (HPA), suffer from reaction latency, leading to Service Level Objective (SLO) violations during traffic spikes and resource flapping during ramp-downs. While Deep Reinforcement Learning (DRL) offers a pathway toward proactive management, standard agents suffer from \textit{temporal blindness}, an inability to exploit the recent temporal context in non-Markovian edge environments. To bridge this gap, we propose a stability-aware autoscaling framework unifying short-horizon temporal context and control via an Attention-Enhanced Double-Stacked LSTM architecture integrated within a Proximal Policy Optimization (PPO) agent. Unlike shallow recurrent models, our approach employs a learned attention mechanism that weights recent historical states non-uniformly, suppressing high-frequency jitter while preserving the trend that precedes demand shifts. We validate the framework on two independent Kubernetes clusters using real-world Azure Functions traces. Against the single-layer LSTM ablation and the static HPA baseline, our approach reduces P90 latency by

\approx

67%, and holds average latency within the 50ms hard SLO for 98.8% of the run against 49.6% and 43.5% respectively. Against Kubernetes Event-Driven Autoscaling (KEDA), it matches latency performance at 75% fewer replica-steps and 59% less churn, with P90 hard-SLO violation bursts of at most 5 consecutive intervals against up to 24 for KEDA. These results indicate that mitigating temporal blindness through deep attentive memory improves the reliability and stability of Kubernetes autoscaling under bursty edge workloads.

Abstract

Explore similar work

Learning to Remember: Attentive Reinforcement Learning for Edge Serverless Autoscaling

Spotlight: Synergizing Seed Exploration and Spot GPUs for DiT RL Post-Training

ARGOS: Reinforcement Learning-Driven Multidimensional Elasticity for Service Orchestration in the Computing Continuum