JuiceFS keeps file metadata in a database and file contents in an object store. This container runs its S3 gateway, so everything on your server that speaks S3 can talk to it. The official binary, unmodified, supervised by s6-overlay. The file system is created on first boot, so there is no console step between installing the template and using it.

A one-knight job: I build it, keep it running, work through the issues and add what people ask for, until nothing is missing. It is free, with no accounts, no telemetry, no ads and no paid tier. No asterisk anywhere. Nothing readable ever leaves your own walls. Forged on evenings and weekends, with heart and stubbornness.

If it has earned a place on your server or computer, toss a coin to your knight: it helps cover the costs and keeps the project alive. It also makes this knight's heart beat a little faster. Three ways below, whichever suits you.

Table of Contents

1. Why this container exists
2. What is this?
3. Quick start on Unraid
4. Connecting a client
5. Configuration
6. Choosing where the metadata lives
7. Backup
8. How AI is used here
9. Support this project

1. Why this container exists

JuiceFS publishes two ways to run in Docker, and neither one fits a Community Applications template.

The first is a Docker volume plugin. A plugin is not a container, so Unraid cannot install it from a template at all. It also carries a documented flaw for the simplest setup: with SQLite the database file ends up inside the plugin's own container, and upstream notes that it stops working once the service restarts.

The second is the juicedata/mount image, which expects you to write the whole command line yourself, including the metadata URL, the storage backend and the credentials. That works, but a template built on it would be a docker run line in a text field, and the file system would still have to be created by hand in a console before anything could use it.

This image closes both gaps. Every setting is an ordinary environment variable, so every setting is a field in the template, and the file system is created on first boot.

2. What is this?

The container runs juicefs gateway, which serves an S3 API on port 9000.

The gateway is the right mode for a server like this. It needs no FUSE, no --privileged and no shared mount propagation, unlike the mount mode. A plain port is all it takes.

Underneath, JuiceFS splits a file into two parts. The metadata, meaning names, directories, permissions and where the pieces are, goes into a database. The contents go into an object store as chunks. Out of the box this container puts the database in a SQLite file under /config and the chunks in /data, so a fresh install needs no second container and no external service.

One consequence is worth knowing before you start: because files are stored as chunks, the contents of /data are not browsable. You will see JuiceFS's own directory layout there, not your file names. Everything goes in and comes out through the S3 gateway. If you want an S3 API in front of a share you can still read normally with a file browser, VersityGW is the better fit.

The JuiceFS binary comes from the upstream release and is checked against the published SHA256 during the build, so a re-tagged release fails the build instead of shipping quietly.

3. Quick start on Unraid

Install the template from Community Applications, set a secret key if you want to choose your own, and start it. Nothing else is required.

On the first start the container creates the file system and logs what it did. If you left the secret key empty, one is generated and written to /config/.s3_root_password, and the log points you at it.

On every later start it finds the existing file system and leaves it alone. That check comes before anything else, because formatting an existing volume a second time would orphan every object already in the store.

4. Connecting a client

Point any S3 client at the container:

Endpoint:   http://<server>:9000
Access key: juicefs          (or whatever you set as S3_ROOT_USER)
Secret key: the value you set, or the one from /config/.s3_root_password
Region:     us-east-1        (any value works, JuiceFS does not check it)

With the AWS CLI:

aws --endpoint-url http://<server>:9000 s3 mb s3://backups
aws --endpoint-url http://<server>:9000 s3 cp ./file.txt s3://backups/
aws --endpoint-url http://<server>:9000 s3 ls s3://backups/

You can create as many buckets as you like, because the container runs the gateway in its multi-bucket mode. Each bucket is a top-level directory in the file system. That matters more than it sounds: in the plain mode the whole file system is a single bucket named after the volume, and creating one fails with NoSuchBucket, which is the first thing most backup clients try to do. Set MULTI_BUCKETS to false if you want the single-bucket behaviour instead.

5. Configuration

Some settings are read on every start, and some only once, when the file system is created. That distinction matters: changing a create-only setting later has no effect, and the log will not tell you so.

Read on every start:

Read only when the file system is created:

To change any of the second group on an existing install, you have to create a new file system, which means new metadata and starting over with the objects.

STORAGE and BUCKET accept every backend JuiceFS supports, so the same container can put its chunks on another S3 server, on Backblaze B2 or on a MinIO instance instead of on the local disk.

6. Choosing where the metadata lives

META_URL is one field, and it decides which database holds the metadata.

The default is SQLite, a single file under /config. For one server that is the sensible choice: no second container, no network in between, and a backup of /config captures it. What it cannot do is serve several machines writing at once.

For that case, put a Redis or Postgres URL in the same field:

redis://:password@192.168.20.10:6379/1
postgres://user:password@192.168.20.10:5432/juicefs?sslmode=disable

Redis is what the JuiceFS project recommends for speed. Be aware that it keeps data in memory, so its persistence settings decide whether a power cut costs you the file system rather than just a cache.

Whichever you pick, the database is not optional and it is not a cache. The chunks in the object store are unreadable without it. Section 7 follows from that.

7. Backup

Back up the metadata database and the object store together, and from the same point in time if you can.

With the default settings that means /config and /data. If the metadata is in Redis or Postgres, back that database up with its own tools and keep the schedule close to the object store's.

A backup of only the object store is not a backup. The chunks are there, but nothing says which chunks made up which file.

8. How AI is used here

One knight builds this, and AI is one of the tools I work with, the same way I work with an editor or a compiler. It helps me write code and documentation and it checks my work, and that saves me a good many evenings. It does not make the decisions, though. I read and understand everything before it ships, and if something here breaks, that is on me and not on the tool.

You do not have to take my word for it. The code is open and every release note is written by hand. The issue tracker shows how problems actually get handled, including the ones I got wrong the first time. If you find something that is not right, open an issue and I will look at it.

9. Support this project

Questions, bugs, ideas? GitHub issues →.

A one-knight job: I build it, keep it running, work through the issues and add what people ask for, until nothing is missing. It is free, with no accounts, no telemetry, no ads and no paid tier. No asterisk anywhere. Nothing readable ever leaves your own walls. Forged on evenings and weekends, with heart and stubbornness.

If it has earned a place on your server or computer, toss a coin to your knight: it helps cover the costs and keeps the project alive. It also makes this knight's heart beat a little faster. Three ways below, whichever suits you.