How Bilibili Uses Large Language Models to Solve Big Data Platform Issues

Introduction

This article shares Bilibili's practice of building an intelligent assistant based on large language models.

Background

Bilibili is a video‑sharing platform with massive data volumes. Its big‑data platform supports many business lines, including AI and commerce.

Overall Architecture and Scale

The platform follows a “five‑layer integrated” plus “storage‑compute separation” design. The bottom layer is a distributed file system; the middle layer includes an intelligent scheduler and various compute engines such as Spark and Flink, as well as clients, real‑time streams (Kafka), OLAP engines (ClickHouse), and custom tools and CI/CD pipelines.

Every day the platform processes about 270,000 offline tasks, around 20,000 ad‑hoc queries, and roughly 7,000 critical real‑time jobs. The support team receives thousands of inquiries weekly, with each sub‑team handling about three person‑days of tickets, requiring dedicated staff to address task failures and slowdowns.

User Problems

For offline computation, users mainly ask two questions: why a task failed and why it became slow.

Why Tasks Fail

Kernel defects – upgrades without sufficient testing can cause large‑scale failures.

Dependency issues – component upgrades or bugs in shared resources can break dependent tasks.

Data quality problems – corrupted or malformed data leads to failures.

Other reasons such as memory errors.

Why Tasks Slow Down

Hardware aging – massive storage disks wear out, reducing read/write speed.

Resource scheduling pressure – high user volume and mixed deployment cause contention.

Data skew – uneven data distribution or problematic data slows processing.

Diagnosing these causes is labor‑intensive, prompting the exploration of intelligent methods to assist.

Need for Intelligent Assistance

Users typically submit concise, engineering‑focused questions, often just a problem description with a link or screenshot, lacking detailed context.