Implementation Practice of Bilibili's Tag System: Evolution, Architecture, and Future Plans

The presentation introduces the Bilibili tag system, outlining its evolution from the 2021 project kickoff, through a 2022 system reconstruction, to a 2023 systematic build that connects data platforms with business applications.

The overall architecture is divided into three layers: tag production, crowd selection, and crowd application. Data sources such as DB tables, event logs, and AB test data are ingested via the big‑data platform (Hive, Iceberg) and processed with Spark and ClickHouse to generate tags and audiences.

Tag production supports metadata management, tag building, and governance, handling both discrete (enumerated) and continuous (numeric) tags. Continuous tags are stored in Iceberg for real‑time queries, while discrete tags are materialized in ClickHouse using BitMap structures, with custom sharding to improve write performance.

Crowd selection offers five creation methods—rule‑based tags, CSV import, Hive table import, HTTP links from the data platform, and DMP audience packages—and supports both ad‑hoc and scheduled updates. A DSL describes selection logic, which is translated into Iceberg or ClickHouse SQL and executed as DAG tasks.

Online services meet SLA requirements by providing high‑concurrency, high‑availability APIs, storing results in Redis using a KKV model with version control, and supporting features such as traffic control, audience replacement, and rapid rollback via split‑table mechanisms.

Stability and performance are enhanced by multi‑engine support, task queue scheduling, and optimization of Iceberg queries, while continuous tags can be materialized as discrete tags for faster ClickHouse computation when low latency is needed.

Future plans focus on tag and audience governance with visual heat‑maps and lineage, data quality assurance through DQC integration, audience effect evaluation via metric platforms, and further standardization and sharing of tags across business domains.

The session concludes with acknowledgments to the speaker, editors, and the DataFun community.