Operational Data Analytics

• Operational Data Analytics is a disciplined framework that transforms heterogeneous data into actionable insights through monitoring, modeling, and feedback loops.
• It integrates processes such as data collection, transformation, visualization, and control to support both descriptive and predictive decision making across diverse applications.
• ODA is applied in fields from digital business transformation to HPC, using innovative methods like real-time analytics and anomaly detection to optimize performance and efficiency.

Operational Data Analytics (ODA) is the disciplined use of data and analytical practices to generate timely, high-quality information about core operations so that organizations can reduce uncertainty, improve operational efficiency, and enable new digital business models (Xu et al., 2022). In large-scale computing environments, the same term denotes the continuous collection, integration, and analysis of operational telemetry from facilities, hardware, software, and applications in order to improve efficiency, performance, reliability, and sustainability (Netti et al., 2019). Across these settings, ODA is not merely reporting: it combines monitoring, transformation, modeling, visualization, and, in some deployments, feedback and control. The literature consequently treats ODA as both an information-processing capability and an operational loop that links observation to action, spanning descriptive, diagnostic, predictive, prescriptive, and real-time analytics (Xu et al., 2022).

1. Conceptual scope and analytical role

ODA is defined in multiple but compatible ways across the literature. In digital business transformation, it is centered on producing “analytical information” about processes, customers, products, and markets, and on building the “analytical information processing capability” required to capture, transform, analyze, and visualize operational data in dynamic environments (Xu et al., 2022). In HPC operations, ODA is the analysis component within broader Monitoring and Operational Data Analytics, or MODA: monitoring continuously collects telemetry, whereas ODA extracts patterns, detects anomalies, forecasts, and turns measurements into decisions that can improve efficiency, throughput, reliability, and energy use (Boito et al., 2024). In operations management, a closely related formulation describes the paradigm as using data to create models that lead to decisions that create value (Mišić et al., 2019).

A persistent theme is that ODA addresses both uncertainty and equivocality. Under the Information Processing Theory framing, uncertainty is the absence of information, while equivocality refers to ambiguity or multiple interpretations; ODA reduces the former by producing rich operational insight and the latter by enabling collaboration, testing, and feedback around shared operational signals (Xu et al., 2022). This distinguishes ODA from conventional dashboards that merely expose raw telemetry. The same distinction appears in HPC practice, where ODA is explicitly contrasted with traditional monitoring by its use of online processing, predictive models, and actuation pathways rather than thresholded metric display alone (Netti et al., 2021).

The scope of ODA is broad. Representative domains in the literature include digital business transformation (Xu et al., 2022), industrial diagnostics over static and streaming data (Kharlamov et al., 2016), connected vehicle fleets (Ulm et al., 2019), HPC and data center operations (Netti et al., 2019), and operational timeseries replay for ISPs, content delivery networks, and video analytics (Kamarthi et al., 7 Jan 2026). This breadth has two implications. First, ODA is domain-specific in data semantics and control surfaces. Second, its recurrent abstractions—telemetry ingestion, cohorting, aggregation, model scoring, and feedback—are shared enough that frameworks, ontologies, and orchestration patterns recur across domains.

2. Data foundations, information processing, and control loops

The data substrate of ODA is heterogeneous. Typical inputs include time-series metrics such as power, temperature, utilization, fan speeds, network counters, and I/O rates; logs and events such as syslogs, scheduler events, and hardware errors; static metadata describing topology, users, jobs, products, or assets; and, in newer systems, images, tables, and documents (Vargis et al., 2024). Industrial OBDA scenarios additionally combine streaming sensor windows with archived traces and static ontological facts, while connected-vehicle platforms integrate CAN bus signals, assignment metadata, and cloud-side aggregates (Kharlamov et al., 2016).

Several architectural dichotomies organize this data flow. Wintermute distinguishes in-band analytics, sampled and consumed within a component at fine temporal scale with low-latency constraints, from out-of-band analytics over historical or asynchronous data with looser latency constraints; it also distinguishes online continuous analyses from on-demand invocations (Netti et al., 2019). The Dagstuhl autonomy-loop position paper frames end-to-end operations via MAPE-K—Monitor, Analyze, Plan, Execute over Knowledge—and extends that formalism to master–worker, fully decentralized, and hierarchical loop organizations (Boito et al., 2024). In business settings, the comparable architectural principle is “processing fit”: organizations should match analytical information processing capability to analytical information needs, with DataOps increasing capability by improving quality, speed, reliability, and business-value alignment (Xu et al., 2022).

DataOps supplies a general operational discipline for ODA. The literature characterizes it as an integrated and disciplined approach for delivering data analytical solutions and building analytics capabilities in a way that continuously accelerates output and improves quality. Its recurrent practices include business value orientation, testing and monitoring throughout the pipeline, continuous improvement, automation, orchestration across tools and teams, collaboration and communication, and cultural change (Xu et al., 2022). In practice, these principles appear in pipeline architectures that automate capture, transformation, validation, aggregation, and delivery; validate results at intermediate steps; and maintain governance, metadata, lineage, and access controls.

A complementary formalization appears in analytics-aware ontology-based access to static and streaming data. There, aggregate concepts are elevated to first-class ontology elements under closed-world semantics for attributes. The aggregate concept semantics are written as

$$\big(\theta_r(\mathrm{agg}\ F)\big)^I=\left\{a\in\Delta \mid \mathrm{agg}\{v\in D\mid (a,v)\in F^I\}\ \theta\ r\right\},$$

and conjunctive query answering remains tractable under the paper’s constraints (Kharlamov et al., 2016). This ontology-level treatment is designed for operational tasks such as turbine diagnostics, where declarative analytics must span live streams, archived windows, and static metadata.

3. Analytical paradigms and mathematical structure

The dominant functional taxonomy in ODA comprises descriptive, diagnostic, predictive, prescriptive, and real-time analytics (Xu et al., 2022). Descriptive ODA generates accurate, timely operational metrics and status reporting; diagnostic ODA identifies bottlenecks, root causes, and inefficiencies; predictive ODA forecasts demand, defect rates, delays, or anomalies; prescriptive ODA recommends process changes or allocations; and real-time ODA streams operational signals into dashboards or automated rules.

The literature shows that different ODA workloads favor different mathematical structures. For retrospective replay over operational timeseries, AHA formalizes the query as

$Q=\langle G, Alg, \theta, T\rangle,$

where cohorts $G$ and time window $T$ define the replay scope, and $Alg=\langle F,M,\theta\rangle$ separates feature extraction from model scoring (Kamarthi et al., 7 Jan 2026). Its central theoretical result is that exact replay is possible when the feature set contains only decomposable statistics. The replay store at epoch $t$ is

$Repl(t)=\{(a,F'(t,a)): a\in Leaf(t)\},$

and if all features in $F$ are decomposable, then storing sufficient leaf-level statistics guarantees exact recovery of features for any parent cohort and therefore guarantees 100% accuracy of downstream model outputs (Kamarthi et al., 7 Jan 2026). The same paper formalizes aggregation identities for sum, count, mean, variance, higher moments, and range, while treating quantiles through approximate decomposable sketches.

For near real-time anomaly detection in HPC environments, a lightweight unsupervised method uses an LSTM sequence autoencoder over a 4-step, 5-feature window sampled at 10-second cadence. With $X_t\in\mathbb{R}^{W\times F}$, reconstruction is

$\hat{X}_t=g_\theta(f_\theta(X_t)),$

and anomaly scores are derived from reconstruction error, with per-feature thresholds updated from the previous 4-hour interval as

$Q=\langle G, Alg, \theta, T\rangle,$0

The reported implementation uses as little as 4 hours of data and approximately 50 epochs per retraining cycle, retraining every four hours and achieving approximate accuracy of 96% on Taurus cluster data (Vargis et al., 2024).

For structural profiling rather than forecasting, Dimensional Data Analysis computes, for each entity or dimension, the triple $Q=\langle G, Alg, \theta, T\rangle,$1 and classifies it against ideal structural types: Identity, Authoritative, Organizational, and Vestigial (Gadepally et al., 2014). This yields a first-pass method for detecting vestigial fields, non-working sensors, formatting issues, and anomalous dimensional structure before advanced analytics are applied. The method is explicitly lightweight and is reported to take a fraction of ingest time on a 2.02M tweet dataset (Gadepally et al., 2014).

These examples show that ODA is not reducible to one model family. Some deployments rely on exact sufficient statistics and OLAP roll-ups, some on sequence models and adaptive thresholds, some on Bayesian or influence-diagram reasoning, and some on lightweight structural signatures or ontology-level aggregates. A plausible implication is that the defining feature of ODA is less the specific algorithm than the operational placement of analytics in a loop that must remain timely, trustworthy, and actionable.

4. Systems, frameworks, and platforms

ODA has produced a diverse ecosystem of platforms, each oriented toward particular operational constraints. Some emphasize low-overhead online analytics within monitoring stacks; others emphasize exact replay, declarative access, edge orchestration, or natural-language interfaces.

System	Operational focus	Defining mechanism
Wintermute on DCDB (Netti et al., 2019)	Online and holistic HPC ODA	Operator plugins, sensor tree, in-band/out-of-band execution
OODIDA (Ulm et al., 2019)	Fleet-scale connected-vehicle ODA	Erlang/OTP orchestration with on-board/off-board JSON workers
AHA (Kamarthi et al., 7 Jan 2026)	Alternative history analysis	Exact replay from sufficient leaf statistics plus OLAP CUBE roll-ups
Analytics-aware OBDA (Kharlamov et al., 2016)	Industrial static+streaming analytics	Aggregate-aware ontology, STARQL, SQL$Q=\langle G, Alg, \theta, T\rangle,$2, and source/cost-aware optimization
VKG ODA chatbot (Khan et al., 27 Jun 2025)	Natural-language access to ODA data	Runtime Virtual Knowledge Graph plus NL-to-SPARQL
EPIC (Karimi et al., 29 Aug 2025)	Multi-modal HPC ODA via generative AI	Hierarchical multi-agent orchestration across IR, descriptive, and predictive analytics

Wintermute exemplifies the integration of ODA into a monitoring substrate. Its architecture consists of operator plugins, operators, blocks, a query engine, and an operator manager, with deployment possible either in Pushers on compute nodes or in Collect Agents on management nodes (Netti et al., 2019). It supports operator pipelines because outputs are exposed as regular sensors, and it uses a hierarchical sensor tree with block templates to instantiate analyses at large scale. Reported overhead is low: overall runtime overhead remained below 0.5% in synthetic query tests, Pusher CPU load per core peaked at 1.2%, and memory usage stayed below 25 MB (Netti et al., 2019).

OODIDA addresses a different constraint set: connected vehicles generating large telemetry volumes under intermittent connectivity. Its architecture separates cloud-side assignment handlers from client-side task handlers, with analytics workers running in arbitrary languages behind JSON interfaces. Experiments reported server-side iteration completion means of 2.7 ms, 5.3 ms, and 8.2 ms for 100 clients producing results immediately under 1, 5, and 10 concurrent tasks per client, respectively (Ulm et al., 2019). A later extension added active-code replacement, allowing user-defined Python modules to be swapped between iterations without restarts; reported averages were 20.3 ms for cloud deployment and 45.4 ms for client deployment, compared with 23.6 s + $Q=\langle G, Alg, \theta, T\rangle,$3 and 40.8 s + $Q=\langle G, Alg, \theta, T\rangle,$4 for standard redeployment (Ulm et al., 2019).

AHA is specialized for retrospective “what would have happened if …?” analysis over high-dimensional operational timeseries. Its late-binding design stores only sufficient statistics for observed leaf cohorts and derives parent-cohort features at query time using CUBE or GROUPING SETS over modern OLAP engines such as ClickHouse (Kamarthi et al., 7 Jan 2026). The paper reports 100% accuracy for decomposable-feature pipelines and up to 85× lower total cost of ownership than conventional solutions.

More recent systems target semantic access and conversational interfaces. A unified ontology for ODA in HPC introduces a shared RDF/OWL model spanning two public datasets, M100 and F-DATA, and validates coverage with 36 competency questions (Khan et al., 8 Jul 2025). The VKG-based ODA chatbot avoids materializing a full telemetry KG; instead, it constructs a query-specific graph at runtime and combines that with NL-to-SPARQL. In the reported experiments, this approach achieved 92.5% accuracy compared to 25% with direct NoSQL queries, reduced average end-to-end latency from 20.36 s to 3.03 s, and kept VKG sizes under 179 MiB (Khan et al., 27 Jun 2025). EPIC generalizes the conversational paradigm further by combining a top-level LLM with specialized low-level agents for information retrieval, descriptive analytics, and predictive analytics, and reports up to 26% higher accuracy for fine-tuned smaller models over large state-of-the-art foundation models in descriptive analytics, alongside 19× savings in LLM operational costs compared to proprietary solutions (Karimi et al., 29 Aug 2025).

5. Representative applications and reported outcomes

In business transformation, ODA is presented as a mechanism for both operational efficiency and new business-model creation. Examples include manufacturing cases in which product sensor data processed through analytical algorithms extracted usage patterns that enabled market segmentation and informed new products and services; telecommunications cases in which social media and text analytics improved customer satisfaction and market position; a technology provider case in which service-request analytics improved efficiency and customer service; and a banking case in which analytical outputs became consultancy services provided to third parties (Xu et al., 2022). These examples are explicitly used to argue that ODA can affect both process optimization and revenue generation.

In HPC, the reported applications are especially rich. Wintermute’s case studies include node-level power prediction with average relative error 6.2% at 250 ms, job-behavior visualization via per-core metrics and job-level deciles, and out-of-band clustering that flagged outliers with posterior probability below $Q=\langle G, Alg, \theta, T\rangle,$5 (Netti et al., 2019). Production deployments described from design to deployment report predictive cooling control on DEEP-EST and job-data visualization on SuperMUC-NG. The cooling controller updates rack internal loop set temperature every minute according to forecasted component temperatures; with $Q=\langle G, Alg, \theta, T\rangle,$6, $Q=\langle G, Alg, \theta, T\rangle,$7, and $Q=\langle G, Alg, \theta, T\rangle,$8, the empirically fitted relationship showed approximately 50% lower primary flow at $Q=\langle G, Alg, \theta, T\rangle,$9 versus $G$0 on both CM and ESB racks (Netti et al., 2021). The same deployment maintained compute-node overhead below 1.1% during HPL (Netti et al., 2021).

Near real-time anomaly detection on Taurus cluster data illustrates the lightweight end of the spectrum. Using five node metrics, 10-second buckets, and an LSTM autoencoder with fewer than 68,000 trainable parameters, the method retrains every four hours and reports approximate accuracy of 96% (Vargis et al., 2024). Correlation-wise Smoothing addresses a different HPC use case: compact knowledge extraction from monitoring time series. On the released HPC-ODA collection, it is reported to yield the same performance as most state-of-the-art methods while producing signatures that are up to ten times smaller and up to ten times faster, and, in a cross-architecture application-classification experiment, achieved $G$1 with Random Forest using 20-block signatures across Skylake, KNL, and Rome nodes (Yu et al., 2020).

Operational replay and retrospective analysis also show substantial operational gains. AHA was benchmarked on production-scale video analytics data and other datasets; the paper reports 100% accuracy for decomposable-feature pipelines, up to 85× lower total cost of ownership than conventional methods, 73× lower cost than storing raw session data, and over 100× lower cost than approximate methods to attain a common goal of >95% accuracy (Kamarthi et al., 7 Jan 2026). In a production case study on a 30-day subset with 95,408,549 sessions, preprocessing per-minute aggregations over 45,454 cohorts ran 1.25× faster with AHA than baseline GROUP BY scans, and downstream regression analysis achieved identical model quality with $G$2 while enabling 6.2× faster query execution (Kamarthi et al., 7 Jan 2026).

Connected-vehicle ODA emphasizes deployment agility. OODIDA was designed because centralized processing is impractical when per-vehicle data volumes can exceed 20 GB/hour (Ulm et al., 2019). Its real-time orchestration is complemented by active-code replacement, which enables iterative A/B testing, on-the-fly algorithm modification, and ad hoc federated learning workflows, with consistency enforced by majority vote over code signatures (Ulm et al., 2019).

6. Evaluation, governance, misconceptions, and research frontiers

The literature does not converge on a single KPI catalog for ODA. In DataOps-enabled business settings, the emphasized measures are quality of analytics outputs, speed or cycle time to insights, pipeline reliability, and organizational outcomes such as operational efficiency, innovativeness, and new business models (Xu et al., 2022). In HPC and data-center settings, additional concerns include predictive accuracy, AUROC, precision/recall, F1, calibration, drift detection latency, time-to-detect anomalies, MTBF, MTTR, PUE, and transferability across systems (Jakobsche et al., 2022). Some papers are explicit about what they do not provide: the DataOps framework does not present formal equations or a KPI catalog (Xu et al., 2022), and the autonomy-loop position paper introduces no mathematical equations for metrics or decision rules (Boito et al., 2024).

Several misconceptions are addressed, implicitly or explicitly, by the surveyed work. One is that ODA is equivalent to machine learning. The literature contradicts this: exact replay via decomposable sufficient statistics (Kamarthi et al., 7 Jan 2026), ontology-level aggregate semantics (Kharlamov et al., 2016), structural dimensional analysis (Gadepally et al., 2014), and decision-analytic influence diagrams and value-of-information reasoning (Chavez, 2013) are all treated as core ODA mechanisms even when no learned model is involved. Another misconception is that materialized semantic layers are always preferable. The unified ontology work shows that knowledge graphs improve interoperability and query expressiveness, but it also reports substantial storage overhead; for one day of one M100 metric, the baseline KG was 1,074.89 MiB, the unified ontology with URIs reduced this to 657.36 MiB, and blank nodes reduced it further to 481.00 MiB, still far above the 2.77 MiB NoSQL baseline (Khan et al., 8 Jul 2025). The VKG chatbot responds to precisely this trade-off by constructing graphs only at query time (Khan et al., 27 Jun 2025).

Governance and trust recur as limiting factors. The MODA/QCS position paper emphasizes explainability, uncertainty quantification, concept drift handling, FAIR data, and privacy-aware sharing as central but currently lacking elements in production deployments (Jakobsche et al., 2022). The autonomy-loop literature highlights lack of telemetry interfaces, feedback hooks, auditable execution, and policy guardrails as reasons why most operational feedback remains human-in-the-loop (Boito et al., 2024). Organizationally, ODA deployment requires clear ownership of data quality, pipeline health, and analytic products, as well as governance and metadata sufficient for trust, reuse, and regulated environments (Xu et al., 2022).

Current research directions therefore cluster around empirical validation, interoperability, autonomy, and natural-language access. The DataOps framework explicitly calls for case studies, expert interviews, and deeper exploration of the technologies, processes, and people required for implementation (Xu et al., 2022). MODA researchers advocate an Open-Data Challenge to standardize evaluation and portability studies across sites (Jakobsche et al., 2022). The autonomy-loop community prioritizes reusable interfaces, open datasets, and significant use cases such as scheduler runtime extension, adaptive storage QoS, and OST remapping (Boito et al., 2024). Semantic ODA research is moving toward unified ontologies, scalable knowledge-graph construction, and chatbot interfaces (Khan et al., 8 Jul 2025). Generative-AI systems such as EPIC indicate another frontier: dynamic, multimodal, agent-orchestrated ODA over text, images, tables, telemetry, and predictive models (Karimi et al., 29 Aug 2025).

Taken together, the literature presents ODA as an operational discipline rather than a single technology stack. Its enduring structure is an end-to-end loop: collect and curate heterogeneous operational data, transform them into analytically useful representations, apply descriptive through prescriptive methods appropriate to the workload, and feed the resulting knowledge back into visualization, planning, or automated control. What varies by domain is the balance among latency, exactness, semantic richness, storage cost, and governance. This suggests that the most mature conception of ODA is not a monolithic framework, but a family of rigorously engineered analytic feedback systems adapted to specific operational environments.