A distributed file system. For a high-level description of TernFS, see the TernFS blog post on the XTX Markets Tech Blog. This document provides a more bare-bones overview and an introduction to the codebase.

Goals

The target use case for TernFS is the kind of machine learning we do at XTX: reading and writing large immutable files. By "immutable" we mean files that do not need modifying after they are first created. By "large" we mean that most of the storage space will be taken up by files bigger than a few MBs.

We don't expect new directories to be created often, and files (or directories) to be moved between directories often. In terms of numbers, we expect the upper bound for TernFS to roughly be the upper bound for the data we're planning for a single data center:

10EB of logical file storage (i.e. if you sum all file sizes = 10EB)
1 trillion files -- average ~10MB file size
100 billion directories -- average ~10 files per directory
1 million clients

We want to drive the filesystem with commodity hardware and Ethernet networking.

We want the system to be robust in various ways:

Witnessing half-written files should be impossible -- a file is fully written by the writer or not readable by other clients
Power loss or similar failure of storage or metadata nodes should not result in a corrupted filesystem (be it metadata or filesystem corruption)
Corrupted reads due to hard drives bitrot should be exceedingly unlikely
Data loss should be exceedingly unlikely, unless we suffer a datacenter-wide catastrophic event (fire, flooding, datacenter-wide vibration, etc.)
The filesystem should keep working through maintenance or failure of metadata or storage nodes

We also want to be able to restore deleted files or directories, using a configurable "permanent deletion" policy.

Finally, we want to have the option to replicate TernFS to multiple regions, to be able to scale up compute across multiple data centres, and to remove any single data centre as a point of failure.

Versioning and releases

TernFS is actively used in production, but the project is still evolving quickly. All releases are currently in the 0.X.Y range and should be considered pre-1.0, with the following meaning:

A change in X (e.g. 0.4.5 → 0.5.0) may remove or change internal APIs. Upgrading across minor versions should not be skipped; read the changelog.
A change in Y (e.g. 0.5.3 → 0.5.4) adds new functionality or small configuration changes, and may contain bug fixes.

Components

 A ──► B means "A sends requests to B" 
                                       
                                       
 ┌────────────────┐                    
 │ Metadata Shard ◄─────────┐          
 └─┬────▲─────────┘         │          
   │    │                   │          
   │    │                   │          
   │ ┌──┴──┐                │          
   │ │ CDC ◄──────────┐     │          
   │ └──┬──┘          │     │          
   │    │             │ ┌───┴────┐     
   │    │             └─┤        │     
 ┌─▼────▼────┐          │ Client │     
 │ Registry  ◄──────────┤        │     
 └──────▲────┘          └─┬──────┘     
        │                 │            
        │                 │            
 ┌──────┴────────┐        │            
 │ Block Service ◄────────┘            
 └───────────────┘

servers
- registry
  - 1 logical instance
  - ternregistry, C++ binary
  - TCP bincode req/resp
  - UDP replication
  - stores metadata about a specific TernFS deployment
    - shard/cdc addresses
    - block services addresses and storage statistics
  - state persisted through RocksDB with 5-node distributed consensus through LogsDB
- filesystem data
  - metadata
    - shard
      - 256 logical instances
      - ternshard, C++ binary
      - stores all metadata for the filesystem
        
        file attributes (size, mtime, atime)
        
        directory attributes (mtime)
        
        directories listings (includes file/directory names)
        
        file to blocks mapping
        
        block service to file mapping
      - UDP bincode req/resp
      - state persisted through RocksDB with 5-node distributed consensus through LogsDB
      - communicates with registry to fetch block services, register itself, insert statistics
  - CDC
    - 1 logical instance
    - terncdc, C++ binary
    - coordinates actions which span multiple directories
      - create directory
      - remove directory
      - move file or directory between from one directory to the other
      - "Cross Directory Coordinator"
    - UDP bincode req/resp
    - very little state required
      - information about which transactions are currently being executed and which are queued (currently transactions are executed serially)
      - directory -> parent directory mapping to perform "no loops" checks
    - state persisted through RocksDB with 5-node distributed consensus through LogsDB
    - communicates with the shards to perform the cross-directory actions
    - communicates with registry to register itself, fetch shards, insert statistics
- block service
  - up to 1 million logical instances
  - 1 logical instance = 1 disk
  - 1 physical instance handles all the disks in the server it's running on
  - ternblocks, Go binary (for now, might become C++ in the future if performance requires it)
  - stores "blocks", which are blobs of data which contain file contents
  - each file is split in many blocks stored on many block services (so that if up to 4 block services fail we can always recover files)
  - single instances are not redundant, the redundancy is handled by spreading files over many instances so that we can recover their contents
  - TCP bincode req/resp
  - extremely dumb, the only state is the blobs themselves
  - its entire job is efficiently streaming blobs of data from disks into TCP connections
  - communicates with registry to register itself and to update information about free space, number of blocks, etc.
clients, these all talk to all of the servers
- web
  - 1 logical instance
  - ternweb, go binary
  - TCP http server
  - stateless
  - serves web UI
- cli
  - terncli, Go binary
  - toolkit to perform various tasks, most notably
    - migrating contents of dead block services (terncli migrate)
    - moving around blocks so that files are stored efficiently (terncli defrag, currently WIP, see #50)
- kmod
  - ternfs.ko, C Linux kernel module
  - kernel module implementing mount -t eggsfs ...
  - the most fun and pleasant part of the codebase
- FUSE
  - ternfuse, Go FUSE implementation of a TernFS client
  - slower than the kmod although still performant, requires a BPF program to correctly detect file closes (see -close-tracker-object flag)
- S3
  - terns3, Go implementation of the S3 API
  - minimal example intended as a start point for a more serious implementation
- xtx/ternfs/client
  - A Go library to implement TernFS clients
  - Used by every client except the kmod
daemons, these also talk to all of the servers, and all live in terngc
- GC
  - permanently deletes expired snapshots (i.e. deleted but not yet purged data)
  - cleans up all blocks for permanently deleted files
- scrubber
  - goes around detecting and repairing bitrot
- migrator
  - evacuates failed disks
additional tools
- ternrun, a tool to quickly spin up a TernFS instance
- terntests, runs some integration tests

Building

% ./build.sh alpine

Will build all the artifacts apart from the Kernel module. The output binaries will be in build/alpine. Things will be built in an Alpine Linux container, so that everything will be fully statically linked.

There's also ./build.sh ubuntu which will do the same but in a Ubuntu container, and ./build.sh release which will build outside docker, which means that you'll have to install some dependencies in the host machine. Both of these build options will have glibc as the only dynamically linked dependency.

Testing

./ci.py --build --functional --integration --short --docker

Will run the integration tests as CI would (inside the Ubuntu docker image). You can also run the tests outside docker by removing the --docker flag, but you might have to install some dependencies of the build process. These tests take roughly 30 minutes on our build server.

To work with the qemu kmod tests you'll first need to download the base Ubuntu image we use for testing:

% wget -P kmod 'https://cloud-images.ubuntu.com/noble/current/noble-server-cloudimg-amd64.img'

Then you can run the CI tests in kmod like so:

% ./ci.py --kmod --short --prepare-image=kmod/focal-server-cloudimg-amd64.img --leader-only

The tests redirect dmesg output to kmod/dmesg, event tracing output to kmod/trace, and the full test log to kmod/test-out.

You can also ssh into the qemu which is running the tests with

% ssh -p 2223 -i kmod/image-key fmazzol@localhost

Note that the kmod tests are very long (~1hr). Usually when developing the kernel module it's best to use ./kmod/restartsession.sh to be dropped into qemu, and then run specific tests using terntests.

However when merging code modifying TernFS internals it's very important for the kmod tests to pass as well as the normal integration tests. This is due to the fact that th

Ternfs

Install / Use

README

Goals

Versioning and releases

Components

Building

Testing