How Bilibili Built an LLM‑Powered Assistant to Tackle Massive Data Tasks

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. Its big‑data platform supports many business lines, including AI and commerce.

The platform follows a “five‑layer integrated” architecture with storage‑compute separation: a distributed file system at the bottom, an intelligent scheduling layer, various compute engines such as Spark and Flink, client tools, real‑time data streams (Kafka), an OLAP engine (ClickHouse), plus custom tools and CI/CD pipelines.

Daily workload includes 270,000 offline tasks, about 20,000 ad‑hoc queries, and roughly 7,000 critical real‑time jobs. The support team receives thousands of inquiries weekly, each small team handling about three person‑days of queries, requiring dedicated staff to answer questions about task failures or slowdowns.

User Problems

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

Why tasks fail

Kernel defects, especially after kernel upgrades without sufficient testing.

Issues in dependent components; upgrades or bugs in shared resources cause failures.

Data quality problems.

Other reasons such as memory‑related issues.

Why tasks become slow

Hardware aging, e.g., disk wear leading to reduced read/write speed.

Resource scheduling pressure under massive user load and mixed deployment across departments.

Data skew or problematic data distribution.

Because diagnosing these failures and slowdowns is time‑consuming, Bilibili explores intelligent methods to assist in problem resolution.