OPPO's Application Distribution: Leveraging Big Data, AI, and Intelligent Computing for Cost and Efficiency

OPPO shares its practice of applying algorithms and big data in the application distribution domain, focusing on cost and efficiency improvements.

1. Application distribution business scenario – OPPO’s OS has over 600 million active users worldwide, with 260 million monthly active users in China and 180 billion monthly downloads, providing massive data for analysis. The distribution service covers all user actions, not limited to the app store, and faces challenges such as low‑frequency recommendation accuracy and high variability in game traffic.

2. Data system construction – The data warehouse exceeds 40 PB, with more than 5 trillion records, 4 000+ offline tasks and 200+ real‑time tasks. OPPO adopts a layered architecture (access, foundation, model, application) and addresses three main challenges: development compliance, metadata management, and concise ETL design. Solutions include 150+ data‑governance rules, comprehensive metadata (events, fields, dictionaries, HDFS, tasks, tables), and an innovative partitioned table for efficient ETL.

3. Intelligent compute system – To meet growing AI and model demands, OPPO built a dynamic‑limit‑flow architecture. Versions V1.0, V2.0 and V3.0 progressively introduced dynamic throttling, VIP user identification using ATL truncation and DSSM twin‑tower models, and value‑aware resource allocation, achieving 15‑20 % traffic support increase and revenue growth.

4. Unified modeling across all scenarios – OPPO unifies more than 20 traffic channels and 200+ scenarios by standardizing data, sharing global features, and merging samples. The unified data foundation grew from 2 TB to 30 TB and from 100 million to over 10 billion features, enabling multi‑task modeling with shared and personalized networks.

5. Future roadmap – The plan emphasizes edge‑cloud collaboration, large‑model deployment on devices, and richer user‑profile construction to improve model training, inference, and product experience.