Provides introductions and tutorials on Bilibili-related technologies.
The project extracts FFmpeg’s demuxing logic into a lightweight WebAssembly module that feeds container‑agnostic video packets to WebCodecs, enabling fast, low‑cost frame extraction across many formats and cutting cover‑generation latency by ~40% while reducing container‑related failures by ~72%.
Bilibili replaced a restrictive third‑party video editor with a self‑developed engine, redesigning architecture for extensibility, manageability and controllable rollout, refactoring hundreds of API calls, enabling draft migration, adding observability, and achieving lower crash rates, faster timelines and stable conversion gains while continuing AI‑assisted feature expansion.
Bilibili migrated its massive user‑behavior, commercial AI training, and database synchronization pipelines from Hive and Kafka to an Iceberg‑based streaming‑batch architecture, using Flink and the Magnus optimizer to achieve minute‑level freshness, reduce CPU and memory usage by about 20‑22 %, save roughly 3.55 M CNY annually, and dramatically improve query latency and join performance.
DNTextSpotter is an arbitrary-shaped scene text spotting model using the DETR architecture with an improved denoising training scheme that adds noise to Bézier control points and employs mask‑sliding character queries, achieving significant benchmark gains without extra inference cost and enabling robust text recognition in challenging environments.
Bilibili launched a crowdsourced testing platform that mobilizes its engaged head‑users to run product tests across diverse devices, addressing limited professional resources and scenario mismatches, while mitigating information leakage, security, and compliance risks through confidentiality agreements, whitelist access, and an intelligent management mini‑program that tracks recruitment, feedback quality, and incentives.
Bilibili’s infrastructure team created a lightweight, multi‑layered disaster‑recovery drill platform—combining an atomic fault library, scenario catalogs, chaos‑experiment orchestration, real‑time observation, and a product‑level interface—backed by standardized governance and CI‑integrated automation, cutting drill preparation from weeks to days and boosting weekly resilience testing across the organization.
The article traces Bilibili’s customer‑service seat scheduling system from its initial balanced‑distribution algorithm and Redis‑based priority queues for live chat and ticket handling, through fairness‑focused saturation limits and virtual‑queue mechanisms, to planned dynamic tuning and expertise‑aware routing for future scalability.
The Bilibili live‑streaming team’s S14 tournament support showcases how systematic business‑scenario analysis, precise resource forecasting, accelerated fault‑drill and stress‑test workflows, and optimized tooling can deliver stable, low‑cost performance for massive, high‑concurrency events like the 2024 League of Legends World Championship.
The rebuilt tag system introduces a three‑layer architecture, standard pipelines, Iceberg‑backed storage and custom ClickHouse sharding, a DSL for crowd selection, and a stateless online service, achieving 99.9% success, sub‑5 ms latency, and supporting thousands of tags across dozens of business scenarios while planning real‑time processing and automated lifecycle management.
Bilibili’s AI‑driven game‑recognition system extracts real‑time LoL events through OCR, hero detection and hot‑spot tagging, generating high‑energy timestamps and interactive overlays that let viewers jump to key moments and view detailed statistics, enhancing spectator engagement and analytical capabilities across major esports tournaments.
