Mixed Ranking Service Upgrade for E-commerce Recommendation System

Recommendation systems filter information to predict user preferences. The mixing stage combines results from multiple sources (ads, natural recommendations) to produce a final ranked list.

The current Taobao feed mixing has issues: inconsistent algorithm objectives between ad and natural results, and tight coupling between algorithm and engineering logic.

Solution: upgrade the mixing strategy service using the xrec framework, creating an independent service (xhuffle) that handles both ad and natural result mixing. Features include unified objective (user + commercial value), decoupled strategy from business chain, componentized workflow, full asynchronous concurrency.

Two integration schemes were evaluated:

Parallel call: split sorting stage into pre‑sorting and post‑sorting, invoking xhuffle in parallel with other services. This reduces RT impact but increases downstream card volume, causing redundant resource consumption.

Serial call: invoke xhuffle as a single serial module after pre‑sorting. This increases RT but avoids extra resource usage.

After analysis, the serial approach was selected. Optimizations include moving the unified value model call inside the service and relaxing timeout limits.

Results: xhuffle maintains average RT <30 ms (P99 <70 ms), timeout rate <0.5 %, overall system timeout <0.3 %, and improves both natural and ad metrics.

Future work: extend mixing to short video/live streams, incorporate category‑scattering rules, and build a generic mixing service for more scenarios.