← Uber Interview Insights

Uber·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Jun 2026

Summary

ML system design round at Uber for an MLE role. The whole thing was a deep dive into building a real-time feed ranking system, covering everything from candidate generation to serving infrastructure. Pretty intense scope for a single session.

Questions Asked (7)

Q1

Design a machine-learning-based feed ranking system for a consumer product that ranks candidate posts for a user's home feed in real time.

System DesignTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This is basically a full system design question disguised as an ML question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product context and success metrics, then outline a two-stage ranking architecture (candidate generation and ranking) with real-time feature serving. Discuss model choices, training pipeline, and trade-offs around latency, freshness, and business objectives.

Pro tip: Emphasize how you would handle the cold-start problem and feedback loops, and propose an online evaluation framework (e.g., interleaving or A/B tests) to measure long-term user satisfaction, not just clicks.

1. Clarify Requirements and Metrics

Ask about the product (e.g., Uber's feed), user scale, latency constraints, and define success metrics like CTR, engagement time, and retention. Align with business goals such as driver utilization or rider conversion.

2. Design the Ranking Pipeline

Propose a two-stage system: candidate generation (e.g., from followed accounts, trending, or geo-based) and ranking with a ML model. Describe feature engineering (user, post, context) and model choice (e.g., gradient boosted trees or deep neural networks).

3. Address Real-Time Serving and Scalability

Explain how to serve features and model predictions with low latency using a feature store, caching, and precomputed embeddings. Discuss trade-offs between model complexity and inference speed.

4. Outline Training and Evaluation

Describe offline training with historical logs, handling biases, and online evaluation via A/B tests. Include metrics for both short-term (CTR) and long-term (retention) impact.

5. Discuss Trade-offs and Iteration

Highlight trade-offs: relevance vs. diversity, freshness vs. quality, and exploration vs. exploitation. Propose a roadmap for continuous improvement and monitoring.

Key Points to Mention

  • Two-stage architecture: candidate generation and ranking
  • Feature engineering: user, post, context, and real-time signals
  • Model selection: trade-offs between tree-based and deep learning models
  • Real-time serving: feature store, low-latency inference, caching
  • Evaluation: offline metrics, online A/B testing, and counterfactual methods
  • Cold-start and feedback loops: exploration strategies and debiasing

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

How would you handle cold start for new users and new content items in a feed ranking system?

System DesignAdaptability & Ambiguity
Author's notes

Cold start for items is easier to talk about than cold start for users, and I think I leaned too hard on the item side.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem scope—what type of feed, what data is available for new users/items, and business metrics. Then structure your answer around two parallel tracks: new users and new items, covering data collection, modeling techniques, and evaluation. Emphasize a hybrid approach that combines content-based methods with rapid online learning to mitigate cold start.

Pro tip: At Uber, cold start is often addressed by leveraging cross-domain signals (e.g., Uber Eats behavior for Uber rides) and using contextual bandits to explore-exploit new items/users. Mentioning these shows you understand real-world constraints and scalable solutions.

1. Clarify the problem and constraints

Ask about the feed type (e.g., restaurant recommendations, driver matching), available data (user demographics, item metadata), and business objectives (engagement, conversion). This ensures your answer is tailored and shows you can handle ambiguity.

2. Address new user cold start

For new users, use onboarding signals (e.g., sign-up info, location, device) and contextual features to make initial recommendations. Employ exploration strategies like epsilon-greedy or Thompson sampling to gather preferences quickly.

3. Address new item cold start

For new items, rely on content-based features (e.g., text, images, metadata) and transfer learning from similar items. Use attribute-based embeddings and consider a multi-armed bandit approach to allocate impressions and learn item quality.

4. Design a unified modeling approach

Propose a hybrid model that combines collaborative filtering with content-based methods, and incorporates online learning to update embeddings as new data arrives. Consider a two-tower architecture where one tower handles user features and the other item features, enabling generalization to unseen entities.

5. Evaluate and iterate

Define offline metrics (e.g., recall@k, NDCG) and online metrics (CTR, conversion). Use A/B testing to measure the impact of cold-start strategies, and set up monitoring to detect degradation. Iterate by incorporating feedback loops.

Key Points to Mention

  • Content-based filtering using item metadata and user demographics
  • Exploration-exploitation trade-off via multi-armed bandits or reinforcement learning
  • Transfer learning and cross-domain signals (e.g., from other Uber services)
  • Online learning and real-time feature updates
  • Hybrid models (e.g., two-tower neural networks) that handle both warm and cold entities
  • Evaluation metrics and A/B testing for cold-start scenarios

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

What online and offline metrics would you use to evaluate a feed ranking model, and how do they relate to the product's business goals?

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

Talked through NDCG and MRR for offline, then engagement rate and session depth for online.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product context and business goals (e.g., engagement, revenue, retention) for the feed. Then, structure your answer by separating offline metrics (ranking quality, diversity, calibration) and online metrics (CTR, conversion, session time, retention), and explicitly link each metric to business outcomes. Emphasize the importance of aligning offline and online metrics and using A/B tests to validate offline improvements.

Pro tip: Highlight the trade-offs between optimizing for short-term engagement vs. long-term user satisfaction and business metrics, and mention how you would guard against metric gaming (e.g., clickbait) by including satisfaction or quality metrics.

1. Clarify business goals and product context

Ask clarifying questions to understand the feed's purpose (e.g., Uber Eats restaurant feed, Uber Newsfeed) and the primary business objectives (e.g., orders, revenue, retention). This ensures your metrics are relevant.

2. Define offline metrics

List offline evaluation metrics such as NDCG, MAP, precision@k, recall@k, AUC, calibration, diversity, and novelty. Explain how they measure ranking quality and user relevance.

3. Define online metrics

List online metrics such as CTR, conversion rate, session time, retention, and revenue per user. Explain how they capture real user behavior and business impact.

4. Relate metrics to business goals

Map each metric to business outcomes: e.g., higher CTR may lead to more orders, better ranking diversity may improve long-term retention. Discuss trade-offs and potential conflicts.

5. Validate with A/B testing and monitor

Explain how offline improvements are validated via online A/B tests, and how to monitor for metric regressions or unintended consequences (e.g., decreased user satisfaction).

Key Points to Mention

  • Offline metrics: NDCG, MAP, precision@k, recall@k, AUC, calibration, diversity, novelty.
  • Online metrics: CTR, conversion rate, session time, retention, revenue per user, user satisfaction (e.g., thumbs up/down).
  • Business goals: engagement, revenue, retention, growth, user satisfaction.
  • Trade-offs: short-term engagement vs. long-term satisfaction, exploration vs. exploitation, diversity vs. relevance.
  • A/B testing: how to design experiments, measure statistical significance, and guard against novelty effects.
  • Guardrail metrics: ensure no harm to user experience or business metrics (e.g., unsubscribes, complaints).

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

Walk through the candidate generation and multi-stage ranking pipeline you would build, including filtering steps.

System DesignAlgorithms & Data Structures
Author's notes

Pretty comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem scope (e.g., Uber Eats restaurant recommendations) and then walk through the pipeline from candidate generation to final ranking, emphasizing filtering and efficiency. Use a funnel analogy to show how millions of items are reduced to a handful, and highlight trade-offs between recall and latency at each stage.

Pro tip: Quantify the funnel: mention that candidate generation might retrieve thousands of items, then filtering and ranking reduce to tens, and finally re-ranking to a few. This shows you understand scale and latency constraints at Uber.

1. Clarify requirements and constraints

Ask about the use case (e.g., Uber Eats home feed), scale (millions of users/items), latency budget (<100ms), and business metrics (CTR, conversion). This ensures your design is tailored.

2. Candidate generation

Describe multiple retrieval sources: e.g., collaborative filtering, content-based, trending, and geo-based. Use approximate nearest neighbor (ANN) for embeddings to fetch top-K candidates per source, then union them.

3. Filtering and pre-ranking

Apply hard filters (availability, user preferences, business rules) and lightweight models (e.g., logistic regression) to prune candidates to a manageable set (e.g., hundreds).

4. Multi-stage ranking

Use a cascade of models: a light ranker (e.g., GBDT) to reduce to tens, then a heavy ranker (e.g., deep neural network) for final scoring. Optionally include a re-ranking stage for diversity or business objectives.

5. Evaluation and iteration

Discuss offline metrics (recall@K, NDCG) and online A/B testing. Mention monitoring for feedback loops and retraining frequency.

Key Points to Mention

  • Two-tower models for candidate generation and embedding-based retrieval
  • Approximate nearest neighbor (ANN) algorithms like HNSW or FAISS for scalability
  • Hard filters vs. soft filters and their impact on recall
  • Cascade ranking with progressively complex models to balance latency and accuracy
  • Feature engineering for ranking: user, item, context, and cross features
  • Online metrics: CTR, conversion rate, and business metrics like revenue

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

How would you ensure freshness and diversity in the ranked feed without sacrificing relevance?

Technical Trade-offsProduct Sense & Ideation
Author's notes

Freshness I handled okay by talking about time-decay features and recency boosts at re-ranking.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the problem as a multi-objective optimization where freshness and diversity are constraints or auxiliary objectives alongside relevance. Describe a concrete system design: candidate generation, ranking with a multi-task model, and re-ranking with diversification, then explain how you would measure and tune the trade-offs. Emphasize experimentation and business metrics to validate that relevance is not sacrificed.

Pro tip: Show that you understand the difference between short-term engagement metrics and long-term user satisfaction—freshness and diversity often improve retention even if they slightly reduce immediate CTR. Mention that you would use counterfactual or interleaving experiments to measure these long-term effects without hurting the user experience.

1. Clarify objectives and constraints

Define what freshness and diversity mean for Uber's feed (e.g., new restaurants, new cuisines, new drivers) and how relevance is measured (CTR, conversion, ratings). Establish the business goal: balancing immediate relevance with long-term user engagement and content ecosystem health.

2. Design a multi-objective ranking system

Propose a model that predicts relevance, freshness, and diversity scores, then combines them via a weighted sum or constrained optimization. Use a multi-task learning architecture to share representations and avoid conflicting objectives.

3. Apply re-ranking for diversity and freshness

After initial ranking, use a re-ranking algorithm (e.g., MMR, DPP, or sliding window) to inject fresh and diverse items while keeping the top results relevant. Set constraints like 'at least one new item in top 10' or 'no more than 2 items from the same cuisine'.

4. Measure and iterate with experiments

Define offline metrics (e.g., intra-list diversity, freshness rate, NDCG) and online metrics (CTR, retention, session length). Run A/B tests or interleaving experiments to tune weights and ensure relevance is not degraded.

5. Monitor and adapt to feedback loops

Implement monitoring for popularity bias and filter bubbles. Use exploration (e.g., epsilon-greedy) to gather feedback on fresh/diverse items and update the model periodically to reflect changing user preferences.

Key Points to Mention

  • Multi-objective optimization: balancing relevance, freshness, and diversity as competing objectives.
  • Multi-task learning or multi-head models to predict relevance, freshness, and diversity scores jointly.
  • Re-ranking techniques like Maximal Marginal Relevance (MMR) or Determinantal Point Processes (DPP) for diversification.
  • Offline metrics: NDCG, intra-list diversity, freshness rate, coverage; online metrics: CTR, conversion, retention, session length.
  • A/B testing and interleaving experiments to measure trade-offs and long-term effects.
  • Exploration strategies (e.g., epsilon-greedy, Thompson sampling) to avoid popularity bias and gather feedback on fresh/diverse content.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

How would you design the model training and serving architecture to meet real-time latency requirements at scale?

System DesignTechnical Trade-offs
Author's notes

This is where I ran out of time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the use case and latency SLA, then propose a decoupled architecture with an offline training pipeline and an online serving layer optimized for low-latency inference. Discuss trade-offs between model complexity, accuracy, and latency, and explain how you would scale each component horizontally.

Pro tip: Emphasize the importance of feature consistency between training and serving to avoid training-serving skew, and mention how you would monitor and handle model drift in production.

1. Clarify Requirements

Ask about the specific application, expected QPS, latency SLA (e.g., p99 < 100ms), and scale (e.g., millions of requests per second). This ensures your design is grounded in real constraints.

2. Design Training Pipeline

Propose a scalable offline training pipeline using distributed frameworks (e.g., TensorFlow, PyTorch) with data parallelism and periodic retraining. Highlight the need for feature stores to ensure consistency.

3. Design Serving Architecture

Outline a low-latency serving layer with model inference optimized via techniques like quantization, pruning, and hardware acceleration (GPUs/TPUs). Use a microservices architecture with load balancing and auto-scaling.

4. Address Trade-offs

Discuss trade-offs between model size and latency, batch vs. real-time inference, and consistency vs. availability. Explain how you would choose the right balance for Uber's use case.

5. Ensure Reliability and Monitoring

Describe how to monitor latency, throughput, and model performance in production, and how to handle failures (e.g., fallback models, circuit breakers). Mention A/B testing and canary deployments.

Key Points to Mention

  • Feature store for training-serving consistency
  • Model optimization techniques (quantization, pruning, distillation)
  • Horizontal scaling and auto-scaling of serving instances
  • Caching and pre-computation of features
  • Latency monitoring and alerting
  • Trade-offs between model accuracy and inference speed

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q7

How would you set up experimentation and monitoring for a feed ranking system, and what failure modes would you watch for?

A/B Testing & ExperimentationRoot Cause Analysis
Author's notes

Standard A/B testing setup, talked about holdout groups and metric sensitivity.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer by first outlining a robust experimentation framework (e.g., A/B testing with guardrail metrics), then detail monitoring strategies (real-time dashboards, anomaly detection), and finally enumerate failure modes (e.g., feedback loops, position bias) with mitigation plans. Emphasize Uber's scale and need for low-latency, reliable systems.

Pro tip: Highlight the importance of defining clear success metrics upfront and using interleaving or switchback experiments when A/B testing is infeasible due to network effects or interference. Also, mention the need for automated rollback mechanisms to quickly revert to a safe model if anomalies are detected.

1. Define Objectives and Metrics

Clearly state the goal of the feed ranking system (e.g., increase user engagement, optimize for long-term value) and select primary and guardrail metrics (e.g., CTR, completion rate, user retention, latency).

2. Design Experimentation Setup

Choose appropriate experiment design (A/B test, switchback, interleaving) considering Uber's marketplace dynamics. Ensure randomization unit (user, session) and sample size calculations account for network effects.

3. Implement Monitoring Infrastructure

Set up real-time monitoring for model performance, system health (latency, error rates), and business metrics. Use dashboards, alerts, and anomaly detection to catch issues early.

4. Identify and Mitigate Failure Modes

Enumerate potential failure modes such as feedback loops, position bias, data drift, and system failures. Describe detection methods and mitigation strategies (e.g., randomization, debiasing techniques, fallback models).

5. Iterate and Learn

Establish a process for analyzing experiment results, learning from failures, and continuously improving the system. Include post-mortems and automated retraining pipelines.

Key Points to Mention

  • A/B testing with guardrail metrics to avoid negative side effects
  • Handling network effects and interference through switchback or cluster randomization
  • Real-time monitoring and anomaly detection for model performance and system health
  • Common failure modes: feedback loops, position bias, data drift, and cold start
  • Mitigation strategies: debiasing, exploration, fallback models, and automated rollback
  • Uber's scale and need for low-latency, reliable experimentation infrastructure

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.