Installing MLReef on offline server

The best way to run MLReef on your own on-premises infrastructure is the MLReef Nautilus package. Nautilus is a single docker image containing everything necessary to create machine learning projects and run ML workloads.

Nautilus contains:

  • MLReef Management Service
  • Postgres
  • Gitlab for hosting Git repositories
  • Gitlab Runners for running Machine Learning workloads
  • API Gateway

Installation

Two steps need to be done in order to run MLReef Nautilus locally on a server which has no internet access.

  1. In the first step, run bin/build-export-nautilus-offline to pull and tar all required images at one place.

Then copy the tar files to the offline server at a location of your choice.

  1. In the second step, run bin/build-run-nautilus-offline with tar files location as an argument.

This will start up a the local instance of MLReef with persistent docker volumes named mlreef-opt, mlreef-etc, and mlreefdb-opt containing all user data on offline server.

The installation on an online server:

git clone git@gitlab.com:mlreef/mlreef.git
bin/build-export-nautilus-offline

Copy the tar files from mlreef-images-tar to the offline server. Copy bin/build-run-nautilus-offline script to offline server.

On the offline host:

bin/build-run-nautilus-offline -d $THE_PATH_OF_TAR_FILES -s $PIP_SERVER(optional)

Example:

bin/build-run-nautilus-offline -d mlreef-images-tar
bin/build-run-nautilus-offline -d mlreef-images-tar -s localhost:10010/simple
bin/build-run-nautilus-offline -d mlreef-images-tar -s https://python.example.com/

This will start up a local instance of mlreef with persistent docker volumes named mlreef-opt, mlreef-etc, and mlreefdb-opt containing all user data. The container comes up with a default runner running on same docker network on localhost.

In order to run MLReef Nautilus locally with local volume binding, you will have to replace docker volumes in docker run command of bin/build-run-nautilus-offline script with persistent data volumes.

Example:

  --volume /root/mlreef-gitlab-opt:/var/opt/gitlab   \
  --volume /root/mlreef-gitlab-etc:/etc/gitlab       \
  --volume /root/mlreefdb-opt:/var/opt/mlreef        \

Notes for pip server in offline mode:

  • If the pip server is running on the same offline host with MLReef, localhost needs to be used for pip server URL, e.g : localhost:10010/simple .

  • If the pip server is running on some other host in intra network, the DNS host entry needs to be configured for docker.

    Example for Ubuntu:

  1. Get the DNS Server IP:

$ nmcli dev show | grep 'IP4.DNS'
IP4.DNS[1]:                             192.168.0.1
  1. Edit 'dns' in /etc/docker/daemon.json (create this file if already not there). Multiple DNS server IPs can be added separated by comma.
{
    "dns": ["192.168.0.1"]
}
  1. Restart docker
$ sudo service docker restart

Now, the PIP server host should be accessible from mlreef service as well.

Installing a pypi server

Install pypiserver (https://pypi.org/project/pypiserver) with this command:

pip install pypiserver                
mkdir ~/packages                      # put offline python packages into this directory.

Copy some packages into your ~/packages folder and then get your pypiserver up and running:

python3 -m pip download -d ~/packages -r <requirements file>

Start the server with this command, you can choose a different port number:

pypi-server -p 10010 ~/packages &      # Will listen to all IPs.

From the client computer, type this to test if the pypip server is working:

pip install --extra-index-url http://localhost:10010/simple/ <package-name>