Skip to content

Install and deploy

Preliminary requirements

Running Fractal Server assumes that

  1. It has access to a shared filesystem on which it can read and write.
  2. It has access to a database (currently supported: postgres).
  3. It has access to one of the supported computational backends.

These requirements are sufficient to use the local backend, while the following additional requirements are needed to use the SLURM backend:

  1. Fractal Server is installed on a SLURM client node, configured to submit and manage jobs.
  2. The user who runs Fractal Server has appropriate `sudo` privileges, e.g. to run `sbatch` for other users.
  3. The machine where Fractal Server runs exposes a port (possibly only visible from a private network) for communicating with the Fractal client.

How to install

⚠️ The minimum supported Python version for fractal-server is 3.10.

Fractal Server is hosted on the PyPI index, and it can be installed with pip via 

pip install fractal-server

For details on how to install Fractal Server in a development environment, see the Contribute page.

How to deploy

Basic procedure

The basic procedure for running Fractal Server consists in setting up some configuration variables, setting up the database, and then starting the server.

1. Set up configuration variables

For this command to work properly, a set of variables need to be specified, either as enviromnent variables or in a file like .fractal_server.env. An example of such file is 

JWT_SECRET_KEY=XXX
POSTGRES_DB=fractal-database-name
FRACTAL_TASKS_DIR=/some/path/to/the/task/environment/folder
FRACTAL_RUNNER_WORKING_BASE_DIR=some_folder_name
FRACTAL_RUNNER_BACKEND=slurm
FRACTAL_SLURM_CONFIG_FILE=/some/path/to/slurm_config.json

⚠️ JWT_SECRET_KEY=XXX must be replaced with a more secure string, that should not be disclosed. ⚠️

More details (including default values) are available in the Configuration page.

2. Set up the database

The command 

fractalctl set-db
initalizes the database, according to the configuration variables.

3. Start the server

In the environment where Fractal Server is installed, you can run it via The command 

fractalctl start
starts the server (with uvicorn) and binds it to the 8000 port on localhost. You can add more options (e.g. to specify a different port) as in 
usage: fractalctl start [-h] [--host HOST] [-p PORT] [--reload]

Start the server (with uvicorn)

options:
  -h, --help            show this help message and exit
  --host HOST           bind socket to this host (default: 127.0.0.1)
  -p PORT, --port PORT  bind socket to this port (default: 8000)
  --reload              enable auto-reload

Now the server is up, and depending on the intended use case you may have to create/edit some users - see the Users page.

Notice that you could also use more explicit startup commands, see below for an example based on Gunicorn.

Ports

You can get details on the open ports e.g. via ss -tulwn or ss -tulp. An entry like 

Netid                   State                     Recv-Q                    Send-Q                                         Local Address:Port                                               Peer Address:Port
tcp                     LISTEN                    0                         128                                                  0.0.0.0:8010                                                    0.0.0.0:*
shows that port 8010 is open for all the current virtual network (this means for instance that you can use the client from the login node of a SLURM cluster, from any computing node, or from any machine in the local network or with an active VPN).

Serving Fractal Server via Gunicorn

Fractal Server is served through uvicorn by default, when it is run via the fractalctl start command, but different servers can be used. When using gunicorn, you can use a command like 

gunicorn fractal_server.main:app --workers 2 --worker-class uvicorn.workers.UvicornWorker --bind 0.0.0.0:8010 --access-logfile logs/fractal-server.out --error-logfile logs/fractal-server.err

Custom Fractal worker

Fractal also provides a dedicated worker to handle the SIGABRT signal that gunicorn sends after a timeout (see this discussion). Since Uvicorn does not propagate signals coming from Gunicorn ref, we introduce a new worker that convert a SIGABRT signal into a SIGTERM one. This is a custom implementation, that can be optionally enabled by including the --worker-class option for the gunicorn command, as in 

gunicorn fractal_server.main:app --workers 2 --worker-class fractal_server.gunicorn_fractal.FractalWorker --bind 0.0.0.0:8010 --access-logfile logs/fractal-server.out --error-logfile logs/fractal-server.err

Postgres setup

See preliminary notes at
https://github.com/fractal-analytics-platform/fractal-server/issues/388#issuecomment-1366713291.

Fractal Server as a daemon/service

See preliminary notes at
https://github.com/fractal-analytics-platform/fractal-server/issues/388#issuecomment-1366719115.