SEALFS
Describes a decentralized, cloud-native distributed file system with separate data and metadata storage, user-mode file system, and RDMA support.
What this file does
Describes a decentralized, cloud-native distributed file system with separate data and metadata storage, user-mode file system, and RDMA support.
When to use it
- Building a distributed file system for cloud environments
- Designing a system that separates data and metadata onto different disks
- Implementing a user-mode file system with FUSE or syscall hijacking
- Exploring RDMA-based network storage for performance
Assumes this stack
SEALFS
The storage system of sealos, aiming to be a high-performance, highly reliable and auto-scalable distributed file system which fits the cloud native environment.
System Architecture
The architecture of sealfs is decentralized, and there is no single metadata node. sealfs hopes to improve the read and write performance as much as possible and solve the problems of storing large amounts of small files.
Main Components
Sealfs consists of the following three components:
Server
Server component is responsible for storing files and metadata. sealfs separates data and metadata into different disks, since metadata is undoubtedly the hot file on distributed file-system. This way, users can choose better hardware to store metadata.
Client
Client component implements the file-system in user mode. It intercepts file requests, stores, and addresses them through hash algorithms.
Manager
Manager component is responsible for coordinating the cluster.
The System Architecture can be shown as follow:

User Mode All The Way
With specific hardware, sealos hopes to support user-mode completely, from file request hijacking on the client side, to the network, and to the storage, for maximum performance improvement.
More designs can be referred to:
Design Document
RoadMap
Currently, we are committed to improving the performance thoroughly. For other design aspects, such as high reliability and high availability, the priority would be lower.
- first version Function:
-
Client:
- fuse file system interface
- System call hijacking(file system of user mode)
- location algorithm
- batch process
-
Sever:
- bypass file system
- file Storage
- disk manager
- catalogue manager
- Metadata persistent memory storage
- file index
- file lock
- Persistent data structure
-
Manger:
- heart manager
-
Network:
- RDMA
- socket network
-
Test
- IO500
- function test
-
Compile
rust version 1.68
make build
Quick Start
Start Manager
# edit manager.yaml
vi examples/manager.yaml
# start manager with manager.yaml
SEALFS_CONFIG_PATH=./examples ./target/debug/manager &
Start Servers on a Node
./target/debug/server --manager-address <manager_ip>:<manager_port> --server-address <server_ip>:<server_port> --database-path <local_database_dir> --storage-path <local_storage_dir> --log-level warn &
Start Client on a Node
./target/debug/client --log-level warn daemon
Create & Mount Disk
./target/debug/client --log-level warn create test1 100000
./target/debug/client --log-level warn mount ~/fs test1
LICENSE
What's inside
Architecture overview, 3 component descriptions, roadmap with 20+ checklist items, compile and quick-start commands, license
Change this for your project
- Replace
labring/sealfswith your own repository name in the README links - Replace
examples/manager.yamlwith your own config file path - Replace
./target/debug/with your build output directory
Where it goes
Keep it in your repository where the agent or team that needs it will read it.
Worth borrowing
- Separating metadata and data onto different disks to optimize for hot metadata
- Using a decentralized architecture with no single metadata node
- Prioritizing performance over reliability in early roadmap
Related Documents
Zig 0.16.0 Context Document for LLMs
Documents breaking changes from Zig 0.13/0.14 to 0.16.0, covering syntax, I/O, build system, and standard library updates.
绘制对比图片 KITTI
Documents the topic remapping and data flow needed to connect Fast-LIO2 frontend with SC-PGO backend for KITTI evaluation.
Beads Workflow Context
Guides when to use bd graph-based issue tracker versus TodoWrite for session vs multi-session work, with commands and workflows.
EPIC: usda-data-crush-stats
Maps a multi-year USDA grape crush data pipeline from raw downloads through bronze, silver, and gold layers to final CSV fact and dimension tables.