Went through Datadog's MLE loop in February 2026. Sharing this specifically because the ML interview meta at observability companies is different from what you'll read in generic "MLE interview prep" guides, which are all written assuming you're interviewing at a consumer recommendation system.
Datadog's ML problems are fundamentally about time series, anomaly detection, and forecasting at scale. If your ML background is NLP or computer vision, that's fine, but you need to be able to pivot to these domains convincingly.
Loop structure:
Phone screen: 60 min. One coding problem (implement a basic sliding window anomaly detector in Python, nothing fancy, they want to see if you think about edge cases) and then a 20-minute discussion about a project on my resume where I deployed a model to prod.
On-site, round 1: ML breadth. Classic stuff but framed toward observability. They asked: how would you detect anomalies in a multivariate time series where sensors have different latencies and missing values are common. Bayesian methods, robust statistics, simple threshold-based approaches and when to use each. They didn't want a single right answer, they wanted to see how I think through tradeoffs.
On-site, round 2: ML systems. Design a training pipeline for a model that needs to process billions of metric data points per day with near-real-time feedback on model quality. This is basically an MLOps question. Know feature stores, data validation, model monitoring, drift detection.
On-site, round 3: Coding. Implement a simple changepoint detection algorithm. They specified the algorithm so it wasn't open-ended, but I had to code it cleanly in Python and explain the complexity. Felt like a medium-hard DSA problem if you're not familiar with the specific algorithm.
On-site, round 4: Behavioral.
What I'd tell someone prepping: spend serious time on time series forecasting (Prophet, statistical models, not just transformer-based), anomaly detection algorithms (isolation forest, CUSUM, Bayesian changepoint), and MLOps concepts at scale. The actual ML math was secondary to showing you understand the systems context.