Hi, I'm Ines Panker.
I'm a senior staff engineer at Sleuth, a SaaS observability platform (the first engineer they hired, de-facto architect), and the person who just gave this talk.
Almost two decades of building software.
Multi-tenant architecture. Circuit breakers guarding 2-3 million third-party API calls a day across about 40 external APIs. Real-time anomaly detection for hundreds of clients.
Python Go Django Celery PostgreSQL Redis RabbitMQ Elasticsearch LLM agents Conference speaking
LinkedIn is the fastest way to reach me.
Did My Deploy Degrade Production? Detecting Bad Deploys with Math, Not AI
When you deploy some code, do you know its impact on your production stats, like memory usage, request duration, error rates? If you wanted to know if your code improved or deteriorated the speed of SQL queries, would you know where to get this data? In this talk, I walk through a real-world system we built to automatically detect pattern changes to metrics without having to set thresholds manually. (And with no LLMs.)
References for this talk 📚
- DBSCAN
- Elbow method (clustering)
- Gaussian kernel (RBF)
- Interquartile range
- k-nearest neighbors algorithm
- Kernel density estimation
- Kurtosis
- Moving average (rolling mean)
- Normal distribution
- Percentile
- PyOD python library
- Quartile Calculator
- scikit-learn NearestNeighbors
- scipy.signal.find_peaks
- Silverman's rule of thumb (bandwidth selection)
- Standard deviation
- Topographic prominence