Skip to content

Latest commit

 

History

History
76 lines (53 loc) · 3.21 KB

README.md

File metadata and controls

76 lines (53 loc) · 3.21 KB
Raw

High Availability Filesystem

A distributed filesystem library and fuse filesystem built on FoundationDB meant for storing metadata and doing file locking across a cluster of servers or containers.

You could think of HAFS like consul or etcd, but as a fuse filesystem - applications can use file locks and write + rename to do atomic updates across your entire cluster without any library support.

The HAFS filesystem also has the ability to store file data in s3 and other external data to create horizontally scalable directories that work well with write once, sequential read workloads.

Why does this exist

My project bupstash.io needed a reliable and fault tolerant way to serve repository metadata across many machines. Having been bitten by NFS in the past, I wanted to be confident in all the failure modes I would encounter when using distributed file locks.

Ideal use cases

In general this filesystem is good for configuration files and metadata that must be kept consistent across a whole cluster while also tolerating server failure - it has very low single threaded write throughput (unless you use the s3 backed file feature).

Features

  • A distributed and consistent filesystem suitable for metadata and configuration across a cluster.
  • Directories that can efficiently scale to huge numbers of files.
  • Optional s3 backed files that are efficient for sequential access of bulk data.
  • Support for distributed posix whole file locks and BSD locks with client eviction - if a lock is broken, that client can no longer read or write to the filesystem without remounting.

Current Limitations

  • Posix locks do not support partial file range locks (so non exclusive sqlite3 doesn't work yet).
  • Files backed by s3 or other object storage are not efficient for random access, only sequential access.
  • Because HAFS is built on top of foundationdb, it requires at least 4GB of ram per node, which can be a bit heavy depending on the use case and project budget (though it works great on an existing foundationDB deployment).
  • Currently HAFS uses fuse permission checks and is thus subject to TOCTU races if you allow multiple users to modify files. Unless you understand this limitation it is better to use a single user fuse mount.
  • Inodes may be accessed after unlinking like a typical filesystem, but in HAFS they expire after a configurable time limit, you cannot use an unlinked file indefinitely.

Caveats and Gotchas.

  • File locking is only effective at protecting updates to the filesystem, you should not coordinate updates to resources outside the filesystem without at least doing periodic file access and understanding the limitations of this approach.

Getting started

You will need:

  • A running foundationdb cluster.
  • The hafs binaries added to your PATH.

Create and mount a filesystem:

# Create the default 'fs' filesystem.
$ hafs-mkfs
$ hafs-fuse /mnt/hafs

From another terminal you can access hafs:

$ hafs-list-clients
...
$ ls /mnt/hafs
$ echo foo > /mnt/hafs/test.txt

From any number of other machines mount the same filesystem and have a fully consistent distributed filesystem - including file locks.

S3 backed files

Implemented but currently undocumented...