TikTok Recommendation Engineer Interview: Feature Engineering, Model Serving, and RecSys Architecture

TikTok Recommendation Engineer Interview: Feature Engineering, Model Serving, and RecSys Architecture

One-sentence summary: TikTok's recommendation interview was the most "full-pipeline" recommendation interview I've ever had. From feature engineering to model training, from model serving to recommendation architecture — every环节 was covered. Interviewers didn't just focus on how well you do in one area; they wanted to see if you could connect the entire recommendation pipeline. This approach really tests your holistic understanding, but it also gives an advantage to those with hands-on experience.

Background: 3 Years of Recommendation System Experience, TikTok Recommendations

I studied Computer Science for both my bachelor's and master's degrees, then spent 3 years doing recommendation system development at a content platform. I've worked on every stage: recall, pre-ranking, ranking, and re-ranking. My tech stack was primarily Python + TensorFlow + PyTorch, with C++ for online serving. I've built feature platforms and optimized model inference — essentially full-pipeline recommendation experience.

I applied to TikTok's recommendation team because their recommendation system is top-tier globally, especially with the extreme real-time and personalization requirements of short-video recommendation. The technical challenges are significant. I received an interview notification about 4 days after submitting my resume.

1. Interview Process Recap

Round 1: Feature Engineering + Model Fundamentals (About 65 Minutes)

Round 1 was with an engineer working on the recommendation feature platform, starting with feature engineering questions.

"What categories of features are there in recommendation systems? How do you manage the feature lifecycle?" I covered four major categories: user features, item features, context features, and cross features, plus the lifecycle management of introduction → validation → deployment → retirement. The interviewer followed up: "How do you validate feature effectiveness before deployment?" I described a two-step validation process: offline evaluation (AUC, GAUC) and online A/B testing.

Then feature crossing: "How do you cross high-dimensional sparse features? What's the difference between DeepFM and DCN?" I explained that DeepFM uses FM for explicit crossing and DNN for implicit crossing, while DCN uses Cross Network for bounded-degree explicit crossing. The interviewer followed up: "Why can Cross Network do feature crossing? What's its mathematical form?" I wrote out the recursive formula for Cross Network and explained that each layer performs polynomial feature crossing.

For model fundamentals, the interviewer asked: "What loss functions are commonly used in recommendation systems? Why use LogLoss instead of AUC as the optimization objective?" I covered LogLoss, BPR Loss, Softmax Loss, etc., then explained that AUC is non-differentiable and can't be directly optimized. The interviewer followed up: "What if positive and negative samples are extremely imbalanced?" I discussed sampling strategies (undersampling, oversampling) and loss function adjustments (Focal Loss, class weights).

Round 1 also included a scenario question: "How do you predict watch time for short-video recommendation? How is it different from CTR prediction?" I explained the unique aspects of watch time prediction — continuous values, long-tail distribution, and biased observation (only clicked videos have watch time). The interviewer probed biased observation handling, and I covered IPW (Inverse Propensity Weighting) and ESMM (multi-task learning) approaches.

Round 2: Model Serving + Inference Optimization (About 70 Minutes)

Round 2 was with a recommendation engineering architecture expert, with questions more engineering-focused.

"What's the model serving architecture for recommendation systems? What steps does a request go through from arrival to result return?" I described the complete pipeline: request intake → feature retrieval → model inference → post-processing → result return. The interviewer followed up: "How do you optimize feature retrieval latency?" I covered feature pre-computation, feature caching (Redis), and feature merging optimization strategies.

Then model inference optimization: "How do you reduce inference latency for large models? What optimization methods have you used?" I covered two major directions: model compression (quantization, pruning, distillation) and inference optimization (operator fusion, batching, TensorRT acceleration). The interviewer followed up: "Does INT8 quantization significantly affect recommendation model performance? How do you evaluate it?" I explained that recommendation models are more sensitive to quantization than CV models because precision loss in embedding tables affects recall. Evaluation methods include offline AUC comparison + online A/B testing.

Round 2 also included a very practical question: "How do you update recommendation models? Full update or incremental update?" I covered the use cases for full updates (daily retraining) and incremental updates (online learning). The interviewer followed up: "What are the risks of incremental updates? How do you control them?" I discussed data distribution drift, model degradation, cold start issues, and using model monitoring and rollback mechanisms for risk control.

Then a design question: "Design a feature platform that supports feature registration, computation, storage, and serving, while ensuring online-offline consistency." I outlined a complete design: feature registration (schema management), feature computation (offline batch + real-time stream processing), feature storage (offline Hive + online Redis), and feature serving (unified API). The interviewer probed online-offline consistency guarantees, and I covered unified feature code (same code for offline and online) and feature snapshot alignment.

Round 3: Recommendation Architecture + Project Deep Dive (About 55 Minutes)

Round 3 was with the recommendation architecture lead, asking more macro-level questions.

"What's the overall architecture of a recommendation system? How do the modules collaborate?" I described the multi-stage funnel architecture: recall (multi-path recall) → pre-ranking (lightweight model) → ranking (complex model) → re-ranking (business rules + diversity). The interviewer probed the difference between recall and pre-ranking — I explained that recall quickly filters from massive candidates, while pre-ranking does initial sorting of recalled results, with different model complexity and latency requirements.

Then project deep dive: "What's the most challenging recommendation project you've worked on?" I described building a real-time feature platform from scratch — designing and implementing a platform supporting real-time feature computation and serving. The interviewer probed technical details: "How do you fuse real-time and offline features? How do you control latency?" I described a dual-stream Join approach — real-time and offline feature streams joined by user ID, with latency controlled under 50ms.

Round 3 also included an open-ended question: "What's the difference between short-video recommendation and text-image recommendation? Are there different technical challenges?" I covered three unique challenges for short-video recommendation: watch time prediction (continuous values), multimodal understanding (video + audio + text), and real-time feedback (sequential nature of completions, likes, comments). The interviewer was particularly interested in multimodal understanding, and I discussed video embedding extraction methods (CNN + Transformer) and multimodal fusion approaches.

Finally, career plans. I mentioned wanting to deepen my expertise in recommendation architecture, especially real-time recommendation and on-device intelligence. The interviewer said TikTok is also exploring on-device intelligence and we could discuss further.

2. Interview Questions Summary

1. Recommendation feature categories? Feature lifecycle management?

2. How to validate feature effectiveness before deployment?

3. How to cross high-dimensional sparse features? DeepFM vs. DCN?

4. Cross Network's mathematical form? Why can it do feature crossing?

5. Common loss functions in recommendation? Why not optimize AUC directly?

6. How to handle extreme positive-negative sample imbalance?

7. How to predict short-video watch time? How to handle biased observation?

8. Recommendation model serving architecture? How to optimize feature retrieval latency?

9. How to reduce large model inference latency? INT8 quantization impact on recommendation models?

10. How to update models? Risks of incremental updates?

11. Design a feature platform? How to ensure online-offline consistency?

12. Overall recommendation architecture? Difference between recall and pre-ranking?

13. How to fuse real-time and offline features?

14. Differences between short-video and text-image recommendation?

3. Key Takeaways

1. Understand the full recommendation pipeline. TikTok's interview covers every stage — features, models, serving, and architecture. If you've only done model training and don't understand feature engineering or model serving, you'll struggle. I recommend systematically studying the full recommendation pipeline.

2. Feature engineering is the core of recommendations. Many people think recommendations are just about tuning models, but in practice, 80% of time is spent on features. Feature-related questions make up a large portion of the interview — prepare feature classification, feature crossing, feature management, and online-offline consistency thoroughly.

3. Model serving skills are a plus. If you can not only train models but also deploy and optimize them, interviewers will value you highly. Inference optimization (quantization, TensorRT, batching) is especially in demand in industry.

4. Online-offline consistency is a high-frequency topic. Training-serving inconsistency is one of the most common issues in recommendation systems. Feature inconsistency, label inconsistency, and data distribution inconsistency — prepare the causes and solutions for each type.

5. Understand business scenario differences. Short-video, e-commerce, and news recommendation each have different challenges. If you can articulate the differences and coping strategies across scenarios, it demonstrates deep experience.

4. FAQ

Q: How important are deep learning fundamentals for recommendation interviews?

Moderately high. You don't need to derive backpropagation from scratch, but you should understand the principles and use cases of common models (DeepFM, DIN, DIEN, MMOE, etc.). I recommend going through classic recommendation system papers, focusing on each model's innovation and motivation.

Q: Can I transition to recommendations without prior experience?

Yes, but you need to study up. Core recommendation knowledge (feature engineering, ranking models, recall strategies) isn't complex, but requires systematic learning. Start with "Recommender Systems: The Textbook" and "Deep Learning for Recommender Systems," then build a hands-on recommendation project.

Q: Does TikTok's recommendation interview value algorithms or engineering more?

Both matter. Round 1 leans toward algorithms (features, models), Round 2 toward engineering (serving, optimization), and Round 3 toward architecture. Being strong in both makes you very competitive. If you only excel in one area, highlight your strengths on your resume.

Q: How should I design A/B tests for recommendation systems?

The core is traffic splitting strategy and metric selection. Split traffic using user ID hashing to ensure the same user is always in the same group. Metrics have three tiers: core metrics (CTR, watch time), auxiliary metrics (diversity, novelty), and guardrail metrics (coverage, fairness). Experiment duration is typically 1-2 weeks — watch out for novelty effects.

Q: What are the career prospects for recommendation engineering?

Excellent. Recommendation systems are one of the most critical technologies in tech — nearly every content and e-commerce platform needs them. Especially with the rise of large language models, "LLM + recommendation" fusion is a new hotspot — using LLMs for feature extraction, content understanding, and conversational recommendation.