How to Turn AIOps from Hype into Reality: A Practical Roadmap

AIOps is a hot concept, but how can it be deployed? In Pei Dan’s keynote at GOPS Shanghai, he presents a technical roadmap for AIOps implementation and a strategic roadmap that leverages community collaboration.

Pei Dan, a veteran operations researcher who worked at AT&T since 2005, shares his experience of applying machine‑learning and AI techniques to operations problems and notes the growing interest in intelligent operations (AIOps) over the past two years.

He emphasizes that many operators still hesitate because they wonder how AIOps can be applied to their own scenarios. Instead of focusing on specific case studies, he proposes a universal technical and strategic roadmap that he will continue to refine over the next decade.

Operations will become increasingly critical as the number of devices worldwide is expected to reach 50‑100 billion by 2020, spanning the internet, finance, IoT, manufacturing, telecom, power grids, and government services. The sheer scale and complexity of hardware and software make reliable, high‑speed, safe operation essential.

Traditional operations face three main pain points: detecting unexpected failures, stopping damage, and repairing or avoiding future incidents. Human‑driven analysis is slow, inaccurate, and cannot keep up with the massive, fast‑changing systems.

Pei Dan argues that the solution lies in leveraging the massive amount of monitoring data already collected. Machine‑learning excels at processing large‑scale, high‑velocity, heterogeneous data, making it a natural fit for operations automation.

The envisioned AIOps workflow consists of a monitoring layer that feeds data to an AIOps engine, which then provides decision suggestions to a small group of experts. The experts approve the actions, and automated scripts execute fault mitigation, repair, or avoidance.

Key engine modules include:

Anomaly detection : raises alerts for potential failures.

Anomaly localization : offers immediate mitigation advice.

Root‑cause analysis : identifies the underlying cause for repair.

Anomaly prediction : forecasts performance bottlenecks, capacity shortages, or failures to enable proactive avoidance.

He illustrates the challenges of anomaly detection: static thresholds cause many false positives/negatives, algorithms lack clear applicability, data may be missing, and labeling is scarce. To address these, he proposes a “butcher‑the‑cow” methodology that decomposes the problem into AI‑friendly sub‑tasks.

First, an unsupervised anomaly detector acts as a coarse filter. Operators then label a few critical cases, and a supervised model learns from these labels, automatically extending the knowledge to similar patterns. When a KPI’s pattern changes dramatically, transfer learning adapts the model parameters.

For millions of KPIs, he suggests clustering them into a limited number of groups, selecting a representative algorithm per group, and fine‑tuning it for each individual series.

By breaking down the overall fault‑detection problem into well‑defined AI problems—unsupervised detection, semi‑supervised labeling, transfer learning, and KPI clustering—the system becomes tractable.

Beyond detection, the roadmap includes fault localization (identifying the affected dimension or component), root‑cause analysis (building a fault‑propagation tree from module call graphs and configuration data), and prediction (capacity, performance, and failure forecasting).

He stresses that successful AIOps requires a community effort: industrial practitioners provide real‑world data and scenarios, while algorithm researchers contribute robust methods. A dedicated AIOps Challenge platform is proposed to gather community contributions and benchmark solutions.

The overall AIOps architecture integrates the monitoring data, the AI engine, expert validation, and automated remediation, forming a closed loop that continuously learns from operations experience.

In summary, the AIOps technical roadmap consists of anomaly detection, localization, root‑cause analysis, and prediction, each broken down into AI‑solvable sub‑problems, while the strategic roadmap calls for a community‑driven challenge to create benchmark data and accelerate research.