RecSys Multi-Stage Funnel & 50ms SLA

The Multi-Stage Recommendation Funnel is a classic industrial cascading architecture that refines global item candidates ($10^7 \sim 10^8$) down to final user recommendations within a strict 50ms Latency SLA; the standard pipeline enforces a strict latency budget across 4 stages: 1) Multi-Channel Retrieval (10ms budget, filtering down to $\sim 10,000$ items via ANN/collaborative filtering); 2) Pre-Ranking (5ms budget, lightweight scoring reducing to $\sim 1,000$ items); 3) Heavy Ranking (30ms budget, deep multi-task models predicting CTR/CVR for $\sim 100$ items); 4) Re-Ranking (5ms budget, diversity DPP, fatigue rules, yielding Top 10).

E-commerce feeds (Amazon/Taobao), short-video feeds (TikTok), visual discovery feeds (Pinterest/Xiaohongshu), and streaming media homepages (Netflix).

Heavy multi-task ranking networks cannot forward-pass millions of items within a 50ms SLA; the cascading funnel balances computational depth against candidate size at every transition, maximizing relevance under hard latency constraints.

Diagram the 4-stage funnel architecture and explain why candidate sizes exhibit exponential decay: $10^8 \to 10^4 \to 10^3 \to 10^2 \to 10^1$?

Break down the latency budget across Network RTT, Feature Store fetching, and Model forward passes under a 50ms SLA?

Why does the retrieval layer optimize for Recall@K while heavy ranking optimizes for fidelity metrics like GAUC and NDCG?

How to implement graceful load shedding (e.g., pruning pre-ranking output from 1000 to 300) during peak traffic spikes to preserve SLA?

Analyze the 'retrieval ceiling dictates ranking quality' dependency and how to monitor multi-channel cross-coverage ratios?