Design and Implementation of the Optimus Tag Management Platform for Xianyu Feeds

Background: Xianyu product feeds use tag cards to display product attributes. Existing process requires manual changes across many apps, leading to slow experiments and high cost.

Design ideas: Build a centralized operation platform for tag configuration, support AB experiments per scene, and enable targeting by time, version, and category.

System: The Optimus (擎天柱) platform consists of three layers—tags, scenes, and scene experiments—allowing multiple experiment configurations per scene.

Operation console: Provides tag list, scene list, and experiment list, with functions for tag style configuration, scene experiment conditions, and pre‑release isolation with change‑free audit.

Client integration: Uses a rich‑client pattern where HSF remote calls fallback to Diamond configuration data. Example configuration class:

@Configuration
public class SenseLabelConfig {
    @HSFConsumer(serviceVersion = "${hsf.version}", clientTimeout = 200)
    public SenseLabelReadService senseLabelReadService;

    @Bean(name = "senseLabelReadClient", initMethod = "init")
    public SenseLabelReadClient senseLabelReadClient() {
        SenseLabelReadClient senseLabelReadClient = new SenseLabelReadClient();
        senseLabelReadClient.setSenseLabelReadService(senseLabelReadService);
        return senseLabelReadClient;
    }
}

API usage: queryItemSenseLabels returns label data for items.

IdleResultDO
> queryItemSenseLabels(
        MtopInfDO mtopInfo,
        String senseId,
        List
itemReqParams,
        SenseLabelExtraParams extraParams);

Performance: By allowing callers to pass serialized item data, interface latency dropped from 120 ms to 15 ms.

Concurrent tag parsing: Labels may come from MySQL, search, Tair, HSF, etc. The platform uses QLExpress rule engine to decouple logic from code, enabling parallel data fetching and rule execution.

Key components: LabelHandlerManager (parallel processing), LabelDataHandler (data fetch), ExpressRunner (rule execution).

Client protocol: Tags are transformed into a client‑recognizable protocol defined in the “Xianyu Tag Specification”, mapping tag regions (A‑E) to UI slots.

Results: Deployed on Xianyu home and search feeds, achieving rapid tag configuration (minutes vs days), multi‑dimensional experiments, and positive uplift in pCTR and pCVR.

Future work: Expand tag coverage, add UI template configuration, integrate algorithm platform for smarter personalization.