The preceding chapters covered the core algorithm modules — retrieval, ranking, and re-ranking. But a model that runs in a paper is not the same as a model you can deploy in a real setting — a gap nearly every recommender-system learner runs into. This part uses an end-to-end movie recommender project to string the scattered algorithms into a complete system that runs, serves, and deploys, answering the engineering question: how do you build a production-grade recommender from scratch?
Chapters
| Chapter | Topic | The Big Idea |
|---|---|---|
| 11.1 | Project Introduction and Goals | Clarify the gap between offline evaluation and online deployment; settle the technology choices and the learning path |
| 11.2 | System Architecture Design | Decouple offline from online; the classic funnel of retrieval → ranking → re-ranking |
| 11.3 | Offline Pipeline | Feature engineering, YoutubeDNN/DeepFM training, embedding generation, feature ingestion, and model deployment |
| 11.4 | Online Pipeline | Cold start (UCB), multi-route retrieval (Snake Merge), DeepFM ranking, diversity re-ranking |
| 11.5 | Frontend and Interaction | Five Vue 3 pages, Pinia state, search debouncing, a rating-driven data feedback loop |
| 11.6 | Deployment and Operations | Orchestrate five services with one Docker Compose command; health checks and troubleshooting |
What You Will Be Able to Do After This Part
- 🟢 Describe the responsibility boundary between the offline and online systems, and how the storage layer decouples them
- 🟢 Explain why the funnel architecture is necessary: light models filter candidates in retrieval, heavy models score precisely in ranking
- 🟡 Implement the training loop for YoutubeDNN retrieval and DeepFM ranking, and understand why item embeddings are precomputed
- 🟡 Design a cold-start strategy (UCB exploration + preferred genres + popular fallback) and multi-route retrieval fusion (Snake Merge)
- 🔴 Deploy a multi-container system with PostgreSQL/Redis/Elasticsearch/backend/frontend, and troubleshoot common problems
- 🟢 Complete the tiered practice problems in each section to consolidate the engineering essentials
Key Concepts
| Concept | Section | Relevance |
|---|---|---|
| Offline vs. online | 11.2 | The fundamental boundary of recommender engineering; the quality-vs-latency trade-off |
| Funnel architecture (retrieval → ranking → re-ranking) | 11.2 | The backbone of industrial recommendation |
| Item embedding precomputation | 11.3 | The prerequisite for millisecond-level online vector search |
| Cold start / UCB | 11.4 | The exploration-exploitation balance for new users with no behavior |
| Multi-route retrieval + Snake Merge | 11.4 | Fusion compensates for the coverage gaps of any single strategy |
| Data feedback loop | 11.5 | Frontend behavior feedback drives feature updates and recommendation improvements |
Prerequisites
- The three-stage pipeline mental model from 1.1, and an understanding of how retrieval/ranking/re-ranking divide the work
- The basics of the two-tower model in 2.3 (YoutubeDNN) and the ranking models in 3.x (DeepFM)
- Working knowledge of Python, basic neural networks, and SQL; familiarity with Docker basics is a plus
The code for this project lives in the
web_project/directory of thedatawhalechina/fun-recrepository — you can run it as you read.
Tips for This Part
- Get it running before nitpicking the details. Launch the full project with one Docker Compose command and build overall intuition first.
- Grasp the offline/online boundary. This is the core mental framework for engineered systems.
- Pay attention to the steps papers never mention: how features move across systems, how models update without downtime, and how cold start degrades gracefully.
Let's build it! 🛠️