What Label Studio asks of the machine
Label Studio's requirements are about memory and storage; its documentation names no processor minimum. The requirements page sets 8 GB of RAM as the floor and 16 GB as the recommendation, asks for 50 GB of disk on a production instance, and expects port 8080 to be free. It runs on Linux, Windows or macOS and is tested against the latest Chrome, which is what your labellers should use.
| Question | What the documentation says |
|---|---|
| Memory | 8 GB minimum, 16 GB recommended |
| Disk | 50 GB recommended for production, before your images, audio or video |
| Task records | About 2.3 GB per million tasks in SQLite |
| Database | SQLite 3.35+ for small projects, PostgreSQL 14+ for real teams |
| When to switch | Over 100,000 tasks with 5 or more people working at once, or millions of tasks |
| Default port | 8080 |
The disk line is the one people underestimate. The 50 GB covers the application and its records; every image or clip you upload lands in the data folder on top of that. Count your dataset, double it for working room and backups staged before they leave the box, and size the drives from there.
Why a whole server suits annotation work
Labelled data is often the most private thing a machine learning team owns: medical scans, customer calls, documents with names in them. A dedicated server is one physical computer that runs your work and nobody else's, so the disks holding that data, and the memory it passes through, are not shared with another customer.
- No bandwidth bill for big datasets. The port is unmetered, so pushing a few hundred gigabytes of video up for labelling, and pulling exports down again, adds nothing to the invoice.
- Root access. You choose the operating system, the Docker version and where the data lives.
- A way back in. On rack servers, blades and the m510 and m710x cartridges, the out-of-band console gives you the screen and the power button even after a firewall mistake locks SSH out.
- Start again for free. A reinstall from the control panel takes about ten minutes, with no ticket and no charge.
- No contract. Billing is month to month, so a labelling project that ends in March does not leave you paying through December.
Which of our machines fits
Match the class to the size of the team, not the size of the dataset; storage is chosen separately on the order page.
- A pilot or a small group: the compact cartridge servers on the dedicated server page are built for one job. Pick one with 16 GB rather than the 8 GB floor, and NVMe if your media is large.
- A labelling team on PostgreSQL: the single-processor everyday servers give Label Studio, nginx and the database room to share one box without competing for memory.
- Many annotators and a busy database: the two-processor servers take more memory and more disks, and suit several projects running side by side.
- Pre-labelling with your own model: Label Studio's ML backend is a separate web server (it listens on 9090 by default). If your model wants a graphics card, put it on one of our GPU servers and keep Label Studio itself on an ordinary machine.
For a quick trial with a handful of tasks, a Cloud VPS will run the containers; it shares a host with other customers, so move to a whole server before real data arrives.
Up and running in an evening
These steps use the project's official Compose file, which runs Label Studio behind its own nginx with a PostgreSQL 17 container. The plan is to keep the web port on localhost until your own account exists and sign-up is closed, because a fresh instance lets anyone who reaches it register.
- Install Docker Engine and the Compose plugin using Docker's instructions for your distribution.
- Fetch the project:
sudo git clone https://github.com/HumanSignal/label-studio.git /opt/label-studiothencd /opt/label-studio(edit files there with sudo). Keeping it under/optavoids the locked-down home folders on Ubuntu 24.04. - The container runs as a non-root user with UID 1001, so it must own the data folder:
sudo mkdir -p /opt/label-studio/mydata && sudo chown -R 1001:1001 /opt/label-studio/mydata. - Edit
docker-compose.ymlso the nginx service publishes only on localhost for now (the fragment below). Docker's published ports skip ufw rules entirely, so this line, not the host firewall, decides who can reach the app. - Start the stack with
docker compose up -d(the older standalone tool isdocker-compose up -d; Label Studio needs Compose 1.25.0 or newer). - From your own computer, open a tunnel with
ssh -L 8080:127.0.0.1:8080 you@your-server, browse tohttp://localhost:8080and sign up your own account first. - Add the line
- LABEL_STUDIO_DISABLE_SIGNUP_WITHOUT_LINK=trueto the existing environment list of the app service (keep the database lines already there), then rundocker compose up -d --force-recreate. From now on new people join only through the invite link under Organization, then Add Members. Reset Link there cancels an old link. - For the team, drop
cert.pemandcert.keyintodeploy/nginx/certs, uncomment the twoNGINX_SSL_CERTlines, setLABEL_STUDIO_HOSTto your https address, restore the8081:8086mapping and recreate the stack again. Your colleagues then use port 8081 over HTTPS.
nginx:
ports:
- "127.0.0.1:8080:8085"
# - "8081:8086" # back on once certificates are in place
Leave the db service without a ports entry. The Compose file sets PostgreSQL to trust connections without a password, which is only acceptable because nothing outside the Compose network can reach it. If login fails with a CSRF error behind HTTPS, set CSRF_TRUSTED_ORIGINS to the same https address, with no trailing slash.
Where the files live, and keeping them private
Label Studio keeps projects, tasks and annotations in the database and puts uploads in the data folder, mydata in the Compose setup. You can also point a project at Amazon S3, Google Cloud Storage, Azure Blob Storage, Redis or a folder on the server. With bucket storage the media is not copied: it stays in the bucket and reaches the browser through pre-signed URLs or a storage proxy when someone opens a task.
For data that must never leave your own hardware, keep the media on the server. Mount a folder into the container, set LABEL_STUDIO_LOCAL_FILES_SERVING_ENABLED=true and LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT to that mount, and add it as local storage in the project. The same UID 1001 ownership rule applies to that folder.
Backups, upgrades and keeping watch
A single server is a single copy, so plan your backups before the first project starts. Two things need saving together: the PostgreSQL database and the mydata folder.
- Database: dump it with PostgreSQL's own tool, for example
docker compose exec -T db pg_dump -U postgres postgres > ~/labelstudio.sql, and copy the file off the server. - Uploads: copy
mydataoff the server on the same schedule, so the media and the records that point at it match. - Annotations: exporting a finished project as JSON or CSV (and, for image detection and segmentation projects, COCO, YOLO or Pascal VOC) gives you the labels in a form training scripts can read, independent of Label Studio.
- Upgrades: the Compose file follows the
latestimage. Take both backups first, then rundocker compose pullanddocker compose up -d. Pinning a release tag instead oflatestmeansdocker compose pullfetches only the version you chose, so an upgrade happens when you decide on it. - Watching it:
docker compose psshows whether all three containers are up, anddocker compose logs appis where errors surface.
Restore onto a scratch machine once, before you need it. A dump you have never loaded is a hope, not a backup. A month-to-month server makes that rehearsal cheap; see server migration for how we approach moving a running service between machines.
When this is the wrong home for it
You manage the server; we keep the hardware and the network running. If nobody on your team will apply updates, watch disk space and test restores, a hosted labelling product is the safer choice. An ML backend on a second machine is a separate web service: fence it to Label Studio's address and require authentication like any other service. Keep the app behind HTTPS and invite-only access.
Questions people ask
- Can I stay on SQLite instead of PostgreSQL?
- For a small project, yes. The documentation says SQLite can cope with tens of thousands of tasks if you avoid complex filters and multi-user pipelines, but loading can slow to more than ten seconds per task when imports run during labelling. For more than 100,000 tasks with five or more people working at once, for millions of tasks, or for many concurrent users, use PostgreSQL, which the official Compose file already does.
- Will my images be copied into Label Studio if they sit in S3?
- No. With cloud storage the media stays in your bucket and is loaded through pre-signed URLs or the storage proxy when a task opens; only the task data goes into Label Studio's database. If you want nothing outside your server, use uploads or local storage instead.
- How do I stop strangers creating accounts?
- Set
LABEL_STUDIO_DISABLE_SIGNUP_WITHOUT_LINK=trueon the app container. An invite link on its own does not close the sign-up page; the variable does. After that, people join through the organization's invite link, and resetting the link cancels the old one. - Does Label Studio need a GPU?
- Label Studio itself does not; its requirements list memory and disk only. A GPU matters only if you connect an ML backend running a model for pre-annotation, and that backend is a separate server you can place on a machine with a card.
- How soon can I start labelling on a new server?
- A ready server is handed over within 30 minutes of the payment being approved; the payment and fraud checks before approval are the part that varies. The Docker setup above then takes well under an evening.