VMware · Primly Community

VMware machine learning engineer interview: what they actually test (it's not what I expected)

ml_mike · 5 replies

Went through VMware's ML engineer interview for a role on their AIOps team. This is the team building anomaly detection and predictive analytics for infrastructure monitoring. I'll cut to it.

The process was four rounds: one phone screen and three onsite interviews (virtual).

Phone screen: 45 minutes. One ML concept question (explain precision vs. recall, when would you optimize one over the other for an infrastructure alert use case?) and one coding problem, medium difficulty, mostly data wrangling in Python. They want to know you can code, not just research.

Onsite round 1: ML depth. This is where it got interesting. They asked me to walk through how I'd build a time-series anomaly detection system for server metrics. Not a whiteboard exercise, a conversation. They pushed on: what features you'd engineer from raw CPU/memory telemetry, how you'd handle seasonality and regime changes, what algorithms you'd consider (they specifically asked about isolation forest and when LSTM-based approaches are worth the cost), how you'd evaluate the model when the definition of 'anomaly' is fuzzy, and how you'd handle class imbalance given that true anomalies are rare. This is a real domain problem for them, so they had a lot of opinions. Be ready to defend yours.

Onsite round 2: Coding. Two problems. One was a Python implementation question about writing an efficient sliding window algorithm for detecting metric spikes. The second was more of a data manipulation problem: given a pandas dataframe of events, compute per-entity rolling statistics. Nothing LeetCode-exotic, but you need to be comfortable with numpy/pandas.

Onsite round 3: System design for ML. Design a real-time anomaly detection pipeline at scale. Covered: feature engineering from streaming data, model training and re-training cadence, serving infrastructure, monitoring the model in production, how to handle concept drift when infrastructure patterns change. They care a lot about production ML, not just modeling.

No behavioral round separately, they wove behavioral questions into each interview.

Comp for senior MLE was in the 180-220k total comp range for remote (my offer was at the lower end of that). Not top-of-market but the work is legitimately interesting if you care about applied ML in infrastructure.

Main advice: know your time-series methods, be ready to talk about production model monitoring, and don't come in expecting classic NLP or recsys problems. The domain is infra. Study up.

5 replies

finance_faye

The anomaly detection focus makes sense given what their products do. Did they ask about specific frameworks like MLflow or anything about their internal stack? Trying to figure out how much tooling knowledge matters vs. core ML chops.

ml_mike

Mentioned MLflow in passing and they knew it but didn't quiz me on it. Core ML chops mattered way more. The one tooling thing they cared about was Spark for large-scale feature computation, and that only came up in the system design round. Bring it up if you know it, but don't worry if you don't.

sre_sol

Infrastructure anomaly detection is a genuinely hard problem. The false positive rate is brutal if you get it wrong and oncall engineers just start ignoring the alerts. Did they talk at all about how they handle the human-trust dimension of the model, not just the technical side?

consultant_cam

180-220k total comp for senior MLE remote. Good to have a data point. That tracks with what I've seen for enterprise infra companies in 2026, not FAANG but not the old VMware comp from five years ago either.

ops_omar

Concept drift in infrastructure is especially nasty because a 'normal' pattern can change overnight if someone migrates workloads. Sounds like they're asking the right questions in interviews at least.