As AI workloads diversify, many organizations struggle with data scattered across multiple systems, inconsistent access methods, and complex cross-cloud data movement. Haluo, a major Chinese mobility platform, faced these challenges while supporting model training, data processing, and agent-based applications simultaneously. Their solution was to introduce JuiceFS, a distributed file system, as a unified data foundation. This approach connects Notebook environments, training clusters, evaluation pipelines, and agent runtimes through a single file-based interface. The key insight is that treating 'everything as a file' simplifies data management and enables seamless data flow across different stages of the AI lifecycle. For engineering leaders, this case offers a practical blueprint for consolidating AI data infrastructure, reducing operational overhead, and improving collaboration between data scientists and platform engineers. The pattern is particularly relevant for teams planning to scale agent-based systems that require flexible, shared access to training data and model artifacts.
Haluo adopted JuiceFS to unify data access across model training, data processing, and agent workloads, solving fragmentation and cross-cloud complexity. The case highlights the trend of treating files as the universal interface for AI infrastructure.