Autonomous Driving

Jiushi Intelligence: Multi-Cloud Billion-File Storage for Autonomous Driving Powered by JuiceFS

Honghao Wang

— 3 min read
Jiushi Intelligence: Multi-Cloud Billion-File Storage for Autonomous Driving Powered by JuiceFS

Zelos of Jiushi Intelligence: Building Scalable Storage for Autonomous Driving

Jiushi Intelligence is a high-tech enterprise specializing in autonomous driving and unmanned delivery technologies. It has developed proprietary autonomous driving systems and AI chips capable of large-scale deployment.

With products rolled out in over 200 cities across China, Jiushi holds over 90% market share in complete vehicle sales in the L4 urban freight autonomous driving segment.

---

Business Growth vs. Storage Bottlenecks

Rapid expansion pushed Jiushi's data volume from terabytes to petabytes. The original Ceph-based storage solution began to struggle:

  • High costs and maintenance complexity
  • Performance bottlenecks for small files, metadata operations, high concurrency, and low-latency needs
  • Impact on simulation and training efficiency

Additional challenges included:

  • Frequent multi-cloud data transfers
  • Data fragmentation and high migration costs
  • Complex scheduling
  • Limited community support for certain tools

---

Solution: JuiceFS Unified Storage Infrastructure

Jiushi required a cloud-native, scalable, and low-maintenance storage architecture. After evaluating Alluxio, CephFS, and JuiceFS, they selected JuiceFS as their unified, multi-cloud storage layer.

All core business data — production, simulation, training, and inference — is now stored on JuiceFS, enabling:

  • Efficient massive data handling
  • Flexible multi-environment deployment
  • Unified infrastructure for varied autonomous driving workflows

---

1. Autonomous Driving Workflow & Storage Needs

Training Workflow Steps:

  • Data Collection & Upload — onboard calibration, data capture, and upload.
  • Algorithm Processing — model training/improvement, pass results to simulation.
  • Simulation Verification — if fail, iterate with algorithm team; if pass, move to test vehicle stage.
  • Test Vehicle Verification — if pass, deploy via OTA.

Key storage requirements:

  • High-performance I/O for massive small-file reads and low-latency access
  • Elastic scalability to multi-PB levels
  • Cross-cloud compatibility for unified access
  • Operational simplicity for long-term stability
  • Cost efficiency without sacrificing performance

---

2. Storage Options Compared

Alluxio

  • Complex deployment: Master + Worker clusters
  • Steep learning curve
  • Issues in community edition: hangs, I/O anomalies
image

Architecture of various file systems

CephFS

  • Metadata stored on MDS, data in RADOS
  • Complex operations: expansion & rebalancing are slow
  • Poor small-file write performance

JuiceFS

  • Flexible metadata layer: Redis, TiKV, MySQL, SQLite
  • Supports multiple backends: S3, OSS, MinIO
  • Simplified deployment with strong small-file performance
  • Fits multi-cloud architecture used by Jiushi
image

---

3. JuiceFS in Multi-Cloud Deployment

Architecture

  • Separate metadata & data layers
  • Metadata engines: Redis, MySQL, TiKV, SQLite
  • Data layer: mainstream object storage systems
image

JuiceFS Community Edition Architecture

Jiushi’s Deployment

  • Multiple clouds: China Unicom, China Telecom, Volcano, China Mobile, AWS
  • Self-built IDC: MinIO + Redis
  • Public cloud: OSS + Redis
  • Kubernetes integration via JuiceFS CSI driver and Helm Charts
  • Local NVMe SSD caching improves read speeds significantly

---

Practice 1 — Training Platform Integration

Platform layers:

  • Infrastructure — storage, compute, networking
  • Kubernetes layer — service orchestration
  • ML platform (Kubeflow) — GPU computing, multiple frameworks

Workflow:

  • Training tasks submitted via Notebook or VR
  • Resource scheduling via Training Operator
  • PVC pre-provisioned JuiceFS storage
  • Real-time monitoring of throughput: ~200 MB/s read-heavy patterns
  • Metadata engine tuning: TiKV outperforms Redis for read-intensive scenarios

---

Usage Stats:

  • 700 TB in one bucket
  • 600 million files, mostly small
  • JuiceFS handles high-concurrency small-file operations stably
  • Simulation bucket: PB-level large files stored in China Mobile Cloud Object Storage

---

Practice 2 — Multi-Cloud Data Sync

Setup:

  • Dedicated inter-cloud lines
  • Custom sync tool built around `juicefs sync` command
  • Sync same dataset to multiple clouds

Note:

Cross-cloud mounts are possible but network stability dependent — use cautiously.

image

JuiceFS multi-cloud data synchronization architecture at Jiushi Intelligence

---

4. Summary & Future Direction

JuiceFS Benefits at Jiushi Intelligence:

  • Handles production/simulation/training/inference workloads
  • High-concurrency small-file performance
  • Cross-cloud integrations with Kubernetes and Kubeflow
  • Reduced O&M complexity and costs
  • Stable at PB-scale data loads

Next steps:

  • Wider TiKV adoption for metadata
  • Improved cross-cloud sync mechanisms

---

Tip for AI & Content Teams:

Platforms like AiToEarn官网 offer an open-source AI content monetization framework for multi-platform publishing, analytics, and AI model ranking. When paired with scalable storage (like JuiceFS), such platforms can optimize workflows for data-heavy AI content pipelines in multi-cloud environments.

---

Would you like me to create a visual diagram summarizing Jiushi’s JuiceFS multi-cloud architecture for easier reading? That would make this case study even more accessible to technical teams.

Read more