Cloud Native, Data & InfrastructureCloud Native, Data & Infrastructure
University showcase15min
INTERMEDIATE

What a Failing Bearing Can Teach Us About CI Pipelines

This talk reframes CI regression detection as predictive maintenance: predicting when failures will occur rather than reacting to alerts. It presents PdMLabs, an open-source system tested on 174 annotated CI time series, and explains ground truth creation, time-series validation, and imbalance-aware evaluation metrics.

talk.summaryAiDisclaimer

Anastasios Papadopoulos
Anastasios PapadopoulosData & Web Science Lab, Aristotle University of Thessaloniki

talkDetail.whenAndWhere

Saturday, November 21, 15:10-15:25
Room 3 - Alexandros
talks.description
A vibrating bearing on a factory floor, a slow database query, and a regression sneaking through your CI have more in common than you would think. Each is an event of interest hiding in a time series. Detect it early, and you save money. Miss it, and you pay in downtime, mechanical or reputational.

Agents now write a growing share of the code that reaches production, and a CI rule built on static thresholds may not be enough to catch what slips through. MongoDB has published why that approach stopped working for them. Alerting is reactive by construction.

Predictive maintenance has worked on this same shape of problem for years, but asks a different question: not "did something change?" but "how long until this breaks?"

In this talk, Anastasios will walk through that field's four problem families and what each means for a CI build pipeline. He will then showcase PdMLabs, an open-source system for assessing whether early signs of events of interest are detectable at all, on 174 CI time series annotated by 11 Mozilla performance engineers.

Attendees will learn how to frame regression detection as prediction rather than alerting, how to build a ground truth when annotators disagree, why time-ordered data breaks ordinary validation, and how to evaluate detectors fairly with imbalance-aware metrics instead of misleading accuracy.

Two communities have been solving one problem without talking to each other. Perhaps this is a first step forward.
anomaly detection
predictive maintenance
regression detection
time series
talks.speakers
Anastasios Papadopoulos

Anastasios Papadopoulos

Data & Web Science Lab, Aristotle University of Thessaloniki

Greece

Anastasios Papadopoulos is a PhD candidate at the Data and Web Science Lab of Aristotle University of Thessaloniki. He develops PdMLabs, an open-source system that tests whether early signs of events of interest, such as mechanical failures or software performance regressions, can be detected in monitoring data. His research focuses on predictive maintenance, automated machine learning, and turning time-series telemetry into actionable reliability. He holds an MSc in Data and Web Science from Aristotle University of Thessaloniki, where his thesis won the best dissertation award. Before his PhD, Anastasios spent two years as a full-stack software engineer, using MERN and GraphQL to build microservices for Erasmus Without Paper, the network European universities use to exchange student data. He is also a Huawei Seeds for the Future awardee.