← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
Apr 2026

Summary

Meta DS interview focused entirely on one deep system design question about replacing a legacy ads ranking model. The whole thing was a multi-part beast covering offline eval, online experimentation, and post-launch monitoring. Felt more like a product/ML systems design loop than a typical DS screen.

Questions Asked (5)

Q1

You have a legacy ads ranking model and a new challenger model built by another team. Walk through your complete plan to validate and replace the old model, from offline checks all the way to production rollout.

A/B Testing & ExperimentationSystem DesignProduct Analytics & Metrics
Author's notes

This question is massive and I underestimated how structured they wanted the answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer as a staged validation funnel: offline evaluation, shadow/canary testing, then a controlled online A/B test with guardrail metrics, and finally a gradual rollout with monitoring. Emphasize that the goal is not just to prove the new model is better, but to ensure it doesn't regress any critical business or user experience metrics.

Pro tip: Always define the primary success metric and guardrail metrics upfront, and pre-register your analysis plan to avoid p-hacking. Also, consider the cost of switching: even if the new model wins, a phased rollout with a kill switch is essential to manage risk.

1. Offline Validation

Compare the challenger and legacy models on a held-out dataset using ranking metrics (e.g., NDCG, MAP) and business proxies (e.g., predicted CTR). Check for calibration, bias, and latency constraints.

2. Shadow Deployment

Run the new model in parallel with the old one in production without affecting user experience. Log predictions and compare them to the legacy model to catch discrepancies and ensure stability.

3. Online A/B Test

Design a randomized controlled experiment with a clear hypothesis, primary metric (e.g., CTR, revenue), and guardrail metrics (e.g., user satisfaction, latency). Ensure sufficient power and run for a pre-determined duration.

4. Gradual Rollout

If the A/B test shows a statistically significant win without guardrail regressions, roll out incrementally (e.g., 1%, 5%, 20%, 100%) while monitoring key metrics and having a rollback plan.

5. Post-Launch Monitoring

After full rollout, continue to monitor model performance, data drift, and business metrics. Set up alerts for anomalies and schedule periodic re-evaluation.

Key Points to Mention

  • Define primary and guardrail metrics before testing to avoid bias.
  • Use offline metrics like NDCG and calibration to filter out bad models early.
  • Shadow deployment helps catch production-only issues without user impact.
  • A/B test must be randomized, controlled, and sufficiently powered.
  • Gradual rollout with a kill switch minimizes risk.
  • Monitor for data drift and model degradation post-launch.

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

Q2

How would you use offline logs from the old model to estimate how the new model would have performed, and what are the assumptions and limitations of that approach?

A/B Testing & ExperimentationTechnical Trade-offsProduct Analytics & Metrics
Author's notes

Basically the counterfactual eval sub-question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as an off-policy evaluation (OPE) task, where you use logged data from the old model to estimate the new model's performance. Then, describe a specific method like inverse propensity scoring (IPS) or doubly robust estimation, and clearly state the assumptions (e.g., positivity, unconfoundedness) and limitations (e.g., high variance, support mismatch).

Pro tip: Emphasize that while OPE provides a quick estimate, it is not a substitute for a live A/B test; always validate with a small-scale online experiment before full deployment.

1. Define the estimand and data

Clarify what metric you want to estimate (e.g., CTR, conversion rate) and identify the logged data: actions taken by the old model, context features, and rewards observed.

2. Choose an OPE method

Select an appropriate estimator such as IPS, SNIPS, or doubly robust, based on the logging policy and available data. Explain how it reweights logged rewards to estimate the new model's performance.

3. State assumptions

List key assumptions: positivity (all actions the new model might take have non-zero probability under the old model), unconfoundedness (no unobserved confounders), and that the logged data is representative.

4. Discuss limitations and diagnostics

Acknowledge limitations like high variance when propensities are small, support mismatch, and potential bias if assumptions are violated. Mention diagnostics like effective sample size and propensity distribution checks.

5. Recommend validation and next steps

Suggest validating OPE estimates with a small online A/B test or a shadow deployment, and iterating if discrepancies arise.

Key Points to Mention

  • Off-policy evaluation (OPE) techniques: IPS, SNIPS, doubly robust
  • Importance of propensity scores and the positivity assumption
  • Unconfoundedness assumption and its potential violations
  • Variance-bias trade-off and effective sample size
  • Support mismatch between old and new model action distributions
  • Need for online validation (A/B test) to confirm OPE estimates

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

Q3

Your experiment shows CTR dropped about 1.5% but revenue per thousand impressions went up 3% and advertiser return on ad spend improved. How do you decide whether to ship the new model?

Product Analytics & MetricsA/B Testing & ExperimentationPricing & Monetization
Author's notes

This was the part I actually liked.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and the trade-offs between user experience (CTR) and monetization (RPM, ROAS). Then evaluate statistical significance and practical significance of each metric, and consider long-term effects and potential guardrail metrics before making a recommendation.

Pro tip: Always tie the decision back to the company's north star metric and consider running a follow-up experiment to measure long-term effects, as short-term gains may not persist.

1. Clarify Business Goals and Trade-offs

Understand the primary objective: is it maximizing revenue, advertiser value, or user engagement? Recognize that CTR drop may indicate user dissatisfaction, while RPM and ROAS improvements suggest better monetization.

2. Assess Statistical Significance and Effect Sizes

Check if the observed changes are statistically significant and calculate confidence intervals. Evaluate whether the effect sizes are practically meaningful for the business.

3. Consider Long-term and Guardrail Metrics

Look beyond short-term metrics: will the CTR drop lead to user churn or reduced engagement over time? Identify guardrail metrics (e.g., user retention, satisfaction) that should not degrade.

4. Evaluate Segment-level Impact

Analyze if the effects vary across user segments (e.g., new vs. existing users, demographics). A model might benefit some segments but harm others, affecting overall strategy.

5. Make a Recommendation and Plan Next Steps

Weigh the trade-offs: if revenue and ROAS gains outweigh CTR drop and no guardrails are violated, consider shipping with monitoring. Otherwise, iterate or run a longer experiment.

Key Points to Mention

  • Statistical significance and confidence intervals for each metric
  • Practical significance and business impact of the changes
  • Long-term effects and potential user experience degradation
  • Guardrail metrics such as user retention and satisfaction
  • Segment-level analysis to identify heterogeneous treatment effects
  • Alignment with company's north star metric and strategic priorities

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

Q4

What does your ramp schedule look like for rolling out the new model, and what would trigger a rollback at each stage?

A/B Testing & ExperimentationSystem Design
Author's notes

Talked through a staged ramp starting small and expanding gradually.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Outline a phased ramp schedule starting with a small canary or shadow deployment, then progressively increasing traffic while monitoring guardrail metrics. At each stage, define specific rollback triggers based on statistical significance, business metrics, and system health. Emphasize that rollback criteria should be pre-registered and automated where possible.

Pro tip: Mention that you would set up automated alerts and a kill switch, and that you'd pre-define the minimum detectable effect and required sample size to avoid peeking. Also, highlight the importance of having a rollback plan that includes model versioning and data pipeline reversibility.

1. Define success and guardrail metrics

Identify primary success metrics (e.g., CTR, conversion) and guardrail metrics (e.g., latency, error rate, user satisfaction) that must not degrade. Set thresholds for acceptable performance.

2. Design phased rollout

Start with a shadow deployment or 1% canary to validate model behavior without user impact. Then gradually increase to 5%, 10%, 25%, 50%, and 100% based on predefined criteria.

3. Establish rollback triggers per stage

For each stage, specify quantitative triggers: e.g., if guardrail metric degrades by >2% with statistical significance, or if error rate exceeds 0.1%, rollback immediately. Also include qualitative triggers like user complaints.

4. Monitor and automate

Set up real-time dashboards and automated alerts. Use sequential testing or Bayesian methods to allow continuous monitoring without inflating false positives. Automate rollback if triggers are hit.

5. Document and iterate

After each stage, document learnings and adjust thresholds if needed. Ensure rollback process is tested and fast (e.g., <5 minutes) to minimize impact.

Key Points to Mention

  • Phased rollout: shadow, canary, progressive traffic increase
  • Pre-registered metrics and thresholds to avoid p-hacking
  • Statistical significance and power analysis for sample size
  • Automated monitoring and kill switch
  • Rollback triggers: guardrail metric degradation, system errors, latency spikes
  • Model versioning and data pipeline reversibility for quick rollback

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

Q5

What are the main production risks after the new model is fully launched, and how would you monitor for them?

System DesignRoot Cause AnalysisTechnical Trade-offs
Author's notes

Feature freshness and model latency were the obvious ones I hit first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the ML production lifecycle: data, model, serving, and business impact. For each risk, describe a concrete monitoring signal and an actionable mitigation plan. Emphasize proactive detection and automated rollback to show operational maturity.

Pro tip: Tie every risk to a specific metric and alert threshold, and mention how you'd use shadow deployment or canary releases to catch issues before full impact. This demonstrates you think like an owner, not just a modeler.

1. Identify risk categories

Break down production risks into data drift, model degradation, infrastructure failures, and business metric shifts. This ensures comprehensive coverage.

2. Define monitoring metrics

For each risk category, specify quantitative metrics (e.g., PSI for drift, latency percentiles, CTR) and set alert thresholds based on historical baselines.

3. Establish detection mechanisms

Describe how you'd collect and analyze these metrics in real-time, using tools like dashboards, anomaly detection, and automated alerts.

4. Plan mitigation and rollback

Outline response strategies for each alert, such as retraining, fallback to previous model, or traffic shifting, and define ownership and escalation paths.

5. Iterate and learn

Emphasize post-mortems and continuous improvement of monitoring based on incidents and feedback loops.

Key Points to Mention

  • Data drift and concept drift detection (e.g., PSI, KL divergence)
  • Model performance monitoring (e.g., accuracy, AUC, calibration) with delayed labels
  • System health metrics (latency, throughput, error rates) and infrastructure alerts
  • Business KPI monitoring (e.g., CTR, conversion, revenue) and A/B test guardrails
  • Shadow deployment and canary releases for safe rollout
  • Automated rollback and fallback mechanisms

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