Meta · Primly Community

Meta machine learning engineer interview: what the loop tests and what you can skip

ml_mike · 4 replies

Did the Meta MLE loop for an E5 role on the Feed Ranking team. Going to be blunt about what matters and what's overrated in the prep guides.

The loop (E5 MLE): Recruiter screen Technical phone screen: either coding or ML design, varies by team Onsite: coding round x1, ML system design x1, ML depth/theory x1, Jedi behavioral x1

Some people report getting two coding rounds at onsite. I got one. It was a medium-difficulty array/hash map problem, nothing exotic. Don't neglect coding prep -- I've seen people fail MLE loops at Meta specifically because they treated coding as an afterthought.

ML system design: This is the main differentiator round. They'll give you a classic recommendation or ranking problem: design a feed ranking system, design a content moderation classifier, design an ads relevance model. The framework they want: Problem formulation: what are you optimizing for, what's the business objective, how do they connect Data: what training data, how do you get labels, imbalanced classes Features: what signals, how do you handle real-time features vs. batch Model choice and why: start simple (logistic regression), know when to go deeper (two-tower, transformer) Training: loss function, evaluation metrics (AUC, NDCG, precision@k) Serving: latency constraints, how you handle a 2-stage retrieval + ranking pipeline Monitoring: how you detect model drift, feedback loops

You don't need to be an ML research scientist. You need to be someone who can take a vague business problem and spec out an end-to-end ML pipeline that could plausibly ship.

ML theory round: Lighter than you'd expect. Questions I got: explain gradient descent, when would you use L1 vs L2 regularization, what's the bias-variance tradeoff in your own words. One question about attention mechanisms that was more conceptual than mathematical. No deriving backprop from scratch.

What I'd skip: Coding LeetCode hards. Not worth the time. A handful of mediums in relevant categories (arrays, hash maps, trees) is enough. Meta MLE loop isn't testing LC-hard competence the way some other companies do.

TC note (won't give exact numbers): E5 MLE comp at Meta is solid. RSUs are the story -- they refresh aggressively and the stock has been strong.

4 replies

ds_dmitri

Really good framework on the ML system design. The two-stage retrieval + ranking pipeline is the one I see trip people up -- they design a single ranker and the interviewer has to drag out of them why you might want a fast candidate retrieval pass before the expensive scorer.

analyst_ana

For someone transitioning from DS to MLE, how much coding prep is realistic? I'm decent at Python/SQL but haven't done much algorithmic work.

ml_mike

Honest answer: probably 4-6 weeks of focused LC practice on mediums. Focus on arrays, hash maps, BFS/DFS. If you can reliably solve medium-difficulty problems in 25-30 minutes and explain your approach clearly, you're probably okay for E4/E5 MLE. Where DS-to-MLE candidates usually get dinged is on the ML system design -- make sure you have the full pipeline answer, not just 'I'd build a model.'

ae_andre

The monitoring and feedback loop question is where a lot of MLE candidates fall apart. Everyone knows how to train a model. Very few people have thought carefully about what happens when the model is live and the world changes. Good signal on experienced candidates.