Zero-Cost Monitoring Systems

• Zero-Cost Monitoring is a design principle that reduces recurring observation expenses by reusing existing signals, commodity hardware, and open-source software.
• It is applied in domains such as network telemetry, datacenter sampling, deep-learning process supervision, GNSS interference detection, and neural architecture search.
• Research focuses on optimizing sampling strategies, constant memory techniques, and proxy approaches to balance cost savings with effective system observability.

Zero-Cost Monitoring denotes a class of monitoring designs in which an otherwise dominant cost center is eliminated or reduced to a negligible constant by reusing existing signals, commodity hardware, open-source software, operating-system primitives, or one-shot initialization statistics. In the cited literature, the term spans community-network observability, sampled datacenter telemetry, feature-drift detection in data streams, GNSS interference surveillance with low-cost commercial off-the-shelf receivers, process-level supervision of long-running deep-learning jobs, and zero-cost proxy methods that monitor or score candidate neural architectures without full training (Arcia-Moret et al., 2021, Yaseen et al., 2021, Conde et al., 2022, Kriezis et al., 17 Sep 2025, Zhang, 7 Apr 2026, Dong et al., 2024, Akhauri et al., 2022).

1. Cost dimensions and semantic scope

The phrase zero-cost is not used uniformly across the literature. In Octopus, it refers to zero additional licence cost, because all code is open, it relies on commodity hardware, and it uses volunteer/decentralized resources (Arcia-Moret et al., 2021). In the datacenter-monitoring literature, the emphasis is on reducing collection, storage and analysis costs by controlling how much and which measurements are collected (Yaseen et al., 2021). In Feature Monitoring (FM), the claim is a small and constant memory footprint and a small computational cost in streaming applications, achieved with Exponential Moving Histograms (Conde et al., 2022). In Deep Researcher Agent, Zero-Cost Monitoring means zero LLM API costs during model training, because monitoring relies solely on process-level checks and log file reads (Zhang, 7 Apr 2026). In GNSS interference monitoring, the phrase denotes zero-cost in terms of additional hardware: no specialized or proprietary hardware is required beyond low-cost COTS receivers, a Raspberry Pi, and a standard antenna (Kriezis et al., 17 Sep 2025). In NAS, zero-cost proxies eliminate the need to train each candidate architecture, instead scoring it at initialization or after one forward+backward pass (Dong et al., 2024, Akhauri et al., 2022).

Domain	Zero-cost target	Mechanism
Community networks	zero additional licence cost	open-source stack on commodity hardware
Datacenter telemetry	collection, storage, analysis cost	sampling-rate control near the Nyquist rate
Data streams	constant memory and small computational cost	Exponential Moving Histograms
Autonomous experimentation	zero LLM API costs during training	OS probes and log-file tails
GNSS RFI monitoring	zero-cost in terms of additional hardware	low-cost COTS receivers and calibration
NAS	no full training of candidates	zero-cost proxies at initialization

A common misconception is that zero-cost means the absence of any resource consumption. The cited systems do not make that claim. They instead remove a specific recurring expense: licensing, specialized hardware, repeated LLM polling, raw-data retention, or repeated model training. This suggests that zero-cost monitoring is better understood as a design constraint on the marginal cost of observability than as a literal claim of zero compute, zero hardware, or zero operational complexity.

2. Open-source and sampled observability in networks

Octopus is a proof-of-concept architecture for stream network monitoring in DIY and community networks. Its core modules are the Octopus Head (OH), which is the web application entry point and manages user roles, test-definition UI, schedule management, and visualization dashboards; Octopus Tentacles (OT), which are lightweight agents installed at remote vantage points and load test modules dynamically from the central code-repository; a Data Repository for raw and processed traces; a Git-style Code Repository for probe definitions, test scripts, and visualization templates; a Stream-Processing & Caching Layer implementing the “retrieve data only when needed” principle; and a Distributed Task Processor in which Celery and Redis handle asynchronous job dispatch to worker processes (Arcia-Moret et al., 2021). The end-to-end flow is likewise explicit: a probe is scheduled via the OH web UI, enqueued in Celery/Redis, executed by an available OT worker after a “git pull,” stored in the central database or cloud store, post-processed into compact summaries, rendered into charts such as heatmaps or time-series, cached, and then displayed in a browser (Arcia-Moret et al., 2021).

Octopus adopts a “no-fee, easily extensible” model: there are no licensing fees or special appliances; OH and OT can run on repurposed low-end servers or even Raspberry Pi–class devices; and new probes are added via a Git-style code-repository pull rather than a paid SDK or proprietary binaries (Arcia-Moret et al., 2021). The paper does not include any new algorithms or protocol designs, and it does not publish any explicit mathematical models or formulae for latency, throughput, packet-loss rate or system overhead. Its performance evaluation is limited to a single screenshot showing a 2.5-month heatmap of server RTTs in 10-minute bins, with no detailed experimental setup, no tabulated metrics, and no quantitative resource-utilization numbers (Arcia-Moret et al., 2021). That limitation is material: the contribution is architectural rather than algorithmic.

A complementary network-monitoring line reframes telemetry as sampled time-series. The central result is the Nyquist–Shannon bound,

$f_s \ge 2 f_{\max},$

used to argue that perfect reconstruction of a band-limited network metric requires sampling at twice its highest nonzero frequency component (Yaseen et al., 2021). The same work proposes a simple cost decomposition,

$$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$

with collection, storage, and analysis all increasing with sampling rate, while reconstruction error falls toward zero near the Nyquist rate (Yaseen et al., 2021). On 14 distinct metrics over 1,613 device–metric traces, 89% of traces were sampled at rates above $2f_{\max}$; the median oversampling ratio was approximately $10\times$, the 20th percentile approximately $100\times$, and the 5th percentile approximately $1{,}000\times$; CPU temperature sampled every $1\,\mathrm{s}$ corresponded to about $500\times$ oversampling; and dropping sampling to $2f_{\max}$ reduced data volume by 99.8% (Yaseen et al., 2021). The paper further reports that adaptive sampling yielded an additional $\sim 2$–$$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$0 seasonal saving with zero loss of detection risk (Yaseen et al., 2021). Here, zero-cost is approximate rather than absolute: the method seeks a cost-vs-quality sweet spot, not literal elimination of monitoring infrastructure.

3. Constant-footprint feature monitoring for data streams

Feature Monitoring (FM) addresses automated real-time stream processing systems by monitoring all features that are used by the system, with a small and constant memory footprint and a small computational cost (Conde et al., 2022). The key data structure is the Exponential Moving Histogram (EMH). For each feature $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$1 with $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$2 bins, if an incoming event falls into bin $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$3, the update rule is

$$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$4

or, in vector form,

$$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$5

where $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$6 is determined by the event-based half-life (Conde et al., 2022). Because the system stores only $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$7 counters per feature and each event entails a constant-size histogram update, both per-event time and memory are treated as $$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$8 when the number of features and bins are fixed constants (Conde et al., 2022).

FM never retains raw samples. Instead, it computes a reference histogram once in a batch pass over training data, maintains a streaming histogram online, estimates the null distribution of feature-wise divergences from sampled reference-period time steps, and then computes per-feature p-values at each monitoring period (Conde et al., 2022). Global drift is declared via a Holm–Bonferroni procedure: sort p-values, compute adjusted signals

$$C_{\mathrm{tot}}(f_s)=C_{\mathrm{coll}}(f_s)+C_{\mathrm{store}}(f_s)+C_{\mathrm{anal}}(f_s),$$9

and reject the global null if any $2f_{\max}$0 (Conde et al., 2022). The system also reports all features whose adjusted signal falls below the threshold, ordered by ascending signal, thereby providing an interpretable feature ranking for root-cause inspection (Conde et al., 2022).

The reported experiments illustrate the practical meaning of “for free.” Larger half-lives yielded fewer, longer alarms; at least 100 bins stabilized alarm counts; KS and Wasserstein divergences were more sensitive on numeric features than Jensen–Shannon; and the method detected anomalies that were not directly related to a single tracked feature, including a sudden spike in card registrations and an uptick in missing billing ZIP codes months before manual detection (Conde et al., 2022). In a public-dataset injection experiment, converting 10% of transaction amounts from dollars to cents over a 1-month window caused clear dips in p-values and crossings of the $2f_{\max}$1 alarm threshold across raw and derived amount features (Conde et al., 2022). The paper states that in all experiments FM added no perceptible latency or memory pressure, even at half-million events/day, because its only per-event cost is the constant-size EMH update (Conde et al., 2022).

4. Process-level monitoring in autonomous experimentation

Deep Researcher Agent introduces Zero-Cost Monitoring as part of an Execute→Monitor→Reflect loop for autonomous deep-learning experimentation (Zhang, 7 Apr 2026). The motivating observation is that model training typically occupies 90–99% of an experiment’s wall-clock time, so repeatedly calling an LLM merely to ask whether a job is still running is wasteful (Zhang, 7 Apr 2026). The monitor therefore runs entirely outside the LLM. Given the process identifier returned by the training launch, it wakes up every $2f_{\max}$2 seconds, with default $2f_{\max}$3, performs kill -0 $2f_{\max}$4 to test liveness, runs nvidia-smi to check that the process is actually driving GPU utilization above $2f_{\max}$5, reads the last $2f_{\max}$6 lines of the training log with default $2f_{\max}$7, appends those lines to a lightweight local buffer, and repeats until the process exits or crashes (Zhang, 7 Apr 2026). Only after training ends are the accumulated log snippets handed to the Reflect phase.

This design removes monitoring-time API usage entirely. Table 4 reports that the monitor itself performs 0 calls, 0 tokens, and costs \$2f_{\max}$80.50 just for monitoring (<a href="/papers/2604.05854" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Zhang, 7 Apr 2026</a>). The overall 24-hour cycle cost is reported as \$2f_{\max}$91.60 for a conventional design, a 10–20× reduction (<a href="/papers/2604.05854" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Zhang, 7 Apr 2026</a>). Over 30 days on 4 projects, the total LLM cost is given as \$10\times$048.00 conventionally; sustained deployments spanned 30+ days and completed 500+ experiment cycles across four concurrent research projects; and one project achieved a 52% improvement over baseline metrics through 200+ automated experiments (Zhang, 7 Apr 2026). The architecture also constrains memory growth by keeping the Tier-2 Memory Log bounded at at most 5,000 characters (Zhang, 7 Apr 2026).

The limitations are equally explicit. The approach depends on logs being written at regular intervals; regex-based metric extraction in Reflect can miss non-standard log formats; and a silent hang of a GPU job that neither crashes nor logs new lines may go undetected until process exit (Zhang, 7 Apr 2026). These caveats delimit what zero-cost means in this setting: monitoring is cheap because it is externalized to operating-system primitives, not because the monitoring problem itself disappears.

5. Low-cost COTS monitoring of GNSS interference

Kriezis et al. present a methodology for detecting and classifying GNSS radio-frequency interference events using low-cost commercial off-the-shelf receivers, specifically a u-blox F9P with Raspberry Pi, standard antenna, and free observables (Kriezis et al., 17 Sep 2025). The method combines carrier-to-noise ratio ($10\times$1) measurements with a calibrated received-power metric. Under nominal conditions, $10\times$2 varies slowly with satellite elevation and front-end noise figure; during jamming, the noise density increases and $10\times$3 degrades; in spoofing, an attacker often balances signal power and noise so $10\times$4 stays near nominal (Kriezis et al., 17 Sep 2025). To minimize natural variability, the method focuses on SBAS (GEO) satellites at fixed azimuth/elevation, yielding a tight $10\times$5 distribution over months with $10\times$6-Hz (Kriezis et al., 17 Sep 2025).

The calibrated received-power metric is obtained in four stages. First, the raw SPAN FFT bins are corrected for AGC using an empirical gain correction of $10\times$7 the SPAN PGA value:

$10\times$8

Second, temperature calibration normalizes measurements to a reference of $10\times$9. Third, the ten 500 kHz bins spanning GPS L1 C/A are combined using known PSD weights, giving

$100\times$0

Fourth, a laboratory setup with a controlled AWGN source and a spectrum analyzer maps processed SPAN output to true power density in dB-W/Hz, after which the antenna + LNA gain of $100\times$1 is subtracted to report power at the receiver input (Kriezis et al., 17 Sep 2025).

These observables define a two-dimensional detection space of $100\times$2 versus received power. A multivariate Gaussian fitted to months of SBAS data defines the nominal region’s $100\times$3-in-$100\times$4 false-alarm contour; a jamming path extends from the top edge of that contour down to the $100\times$5-Hz use-limit from RTCA/EUROCAE MOPS; the blockage region lies below nominal power with strongly reduced $100\times$6; the spoofing region lies to the right of the jamming path, where received power rises while $100\times$7 remains high; and combinations below the noise floor are assigned to an unrealistic/error region (Kriezis et al., 17 Sep 2025). Validation included controlled tests in Norway and real-world deployments in Northern Poland and the Southeastern Mediterranean. The false-alarm target was $100\times$8, approximately $100\times$9. Norway Day 1 jamming results gave sensitivity 99.8%, specificity 99.5%, and accuracy 99.6% for step RFI; sensitivity 89.1%, specificity 100%, and accuracy 94.7% for ramp RFI; and full-day accuracy 99.4%. Norway Day 2 gave RFI detection sensitivity/specificity 99.9% and accuracy 99.9%, while spoofing characterization reached sensitivity 42.1%, specificity 99.9%, and accuracy 78.2% (Kriezis et al., 17 Sep 2025). In this context, the zero-cost claim is explicitly restricted to additional hardware: the method is software-defined and calibration-driven, but not hardware-free.

6. Zero-cost proxy monitoring in neural architecture search

The NAS literature uses zero-cost in an adjacent but distinct sense: performance monitoring or scoring of candidate architectures without full training. ParZC formulates a candidate architecture as a directed acyclic graph with parameterized nodes $1{,}000\times$0, and extracts at each node the statistic set

$1{,}000\times$1

where $1{,}000\times$2 are randomly initialized weights, $1{,}000\times$3 are gradients on a single mini-batch, $1{,}000\times$4 are activations, and $1{,}000\times$5 is the diagonal of the Hessian or an approximation thereof (Dong et al., 2024). Standard zero-cost proxies map these statistics to node-wise scores and typically sum over nodes,

$1{,}000\times$6

ParZC instead parameterizes the proxy through a small differentiable network $1{,}000\times$7 that maps a stacked tensor of normalized proxy statistics to a predicted accuracy, using a Mixer Architecture with Bayesian Network (MABN) to model node-specific uncertainty and a differentiable Kendall surrogate, DiffKendall, to optimize rank correlation directly (Dong et al., 2024). Reported results on NAS-Bench-101, NAS-Bench-201, and NDS show ParZC outperforming several zero-shot baselines, with Spearman/Kendall of 83.2/63.7 on NB101-CIFAR10, 90.4/70.6 on NB201-CIFAR10, and 67.8/50.3 on NDS-DARTS, and search results on NAS-Bench-201 that reached 94.36 on CIFAR-10, 73.49 on CIFAR-100, and 46.34 on ImageNet-16 at a cost of 68 seconds (Dong et al., 2024). The framework requires a small calibration set of architectures, approximately 0.5–1% of the search space, and about 0.2 h of training on a single GPU (Dong et al., 2024).

EZNAS automates the discovery of zero-cost proxies through genetic programming over symbolic expression trees whose terminals are layer-wise statistics collected at initialization under real mini-batches, pure noise, and perturbed data, and whose internal nodes are differentiable tensor operations (Akhauri et al., 2022). Fitness is Kendall’s $1{,}000\times$8 between proxy scores and true test accuracies across held-out architectures; offspring are produced by crossover, mutation, and reproduction; and half of each generation is replaced by new random valid trees to maintain diversity (Akhauri et al., 2022). The final discovered proxy, EZNAS-A, computes the gradient of each convolutional weight tensor with respect to a fixed random loss on Gaussian noise, squares that gradient element-wise, averages within each layer, and then averages across layers:

$1{,}000\times$9

On NASBench-201, EZNAS-A achieved Spearman’s $1\,\mathrm{s}$0 and Kendall’s $1\,\mathrm{s}$1 on CIFAR-10, CIFAR-100, and ImageNet-16-120, exceeding several prior zero-cost proxies; it also generalized across NDS skeletons and NATS-Bench-SSS despite being evolved only on DARTS-CIFAR-10 (Akhauri et al., 2022). This suggests that, in NAS, zero-cost monitoring refers less to operational telemetry than to real-time performance estimation inside the search loop.

7. Limitations, misconceptions, and research directions

Across domains, the strongest limitation is semantic drift in the term zero-cost. Octopus removes licence fees and proprietary appliances but does not provide built-in per-packet capture or deep-packet inspection, lacks formal SLAs or synthetic workload generators beyond standard tools, and has unmeasured scalability limits at thousands of tentacles (Arcia-Moret et al., 2021). The datacenter-sampling framework shows that aggressive down-sampling can preserve fidelity near the Nyquist rate, but it must address irregular or packet-lossy sampling, unknown or time-varying $1\,\mathrm{s}$2, quantization noise, synchronization across signals, and reconstruction latency (Yaseen et al., 2021). FM has constant-memory and low-overhead properties only because the number of features and histogram bins are fixed, low-order constants, and because divergence reference models are precomputed rather than learned continuously online (Conde et al., 2022).

The same pattern appears in other domains. Deep Researcher Agent achieves zero LLM cost during training, but only if jobs expose stable process identifiers, write logs regularly, and do not fail silently while remaining live (Zhang, 7 Apr 2026). The GNSS method avoids specialized hardware, yet still depends on careful AGC correction, temperature calibration, PSD weighting, and laboratory power mapping; moreover, spoofing characterization sensitivity was only 42.1% in the Norway Day 2 experiment, which the authors attribute to receiver anti-spoof logic and weak signals (Kriezis et al., 17 Sep 2025). NAS zero-cost proxies remove the cost of training each candidate, but ParZC may mis-rank architectures when the search space is radically different from the calibration set, and EZNAS, like other zero-cost proxies, still struggles to perfectly rank the top 10% of architectures (Dong et al., 2024, Akhauri et al., 2022).

Current research directions are correspondingly domain-specific. Octopus mentions more advanced stream-processing engines such as Apache Flink or Spark Streaming, formal performance benchmarking, true zero-overhead inline packet taps, and user-defined anomaly detection plug-ins (Arcia-Moret et al., 2021). The datacenter work points toward adaptive sampling and aliasing detectors (Yaseen et al., 2021). FM motivates broader use of full feature-space monitoring in place of custom signals (Conde et al., 2022). Deep Researcher Agent proposes structured logging, distributed-training support, and hybrid schemes that fall back to a single LLM probe after prolonged log silence (Zhang, 7 Apr 2026). GNSS monitoring motivates dense networks around airports or critical sites (Kriezis et al., 17 Sep 2025). NAS work points toward graph-aware proxies, richer calibration, and extensions beyond image classification (Dong et al., 2024, Akhauri et al., 2022). Taken together, these directions indicate that zero-cost monitoring is not a single method but a recurring systems principle: remove the most expensive form of observation while preserving the specific observability guarantees required by the application.