jhub-SwarmSpawner enables JupyterHub to spawn jupyter notebooks across Docker Swarm cluster
More info about Docker Services here.
Python version 3.3 or above is required.
pip install jhub-swarmspawner
git clone https://github.com/rasmunk/SwarmSpawner
cd SwarmSpawner
python setup.py install
You can find an example jupyter_config.py inside examples.
Docker Engine in Swarm mode and the related services work in a different way compared to Docker containers.
Tell JupyterHub to use SwarmSpawner by adding the following lines to your jupyterhub_config.py:
c.JupyterHub.spawner_class = 'jhub.SwarmSpawner'
c.JupyterHub.hub_ip = '0.0.0.0'
# This should be the name of the jupyterhub service
c.SwarmSpawner.jupyterhub_service_name = 'NameOfTheService'
What is jupyterhub_service_name
?
Inside a Docker engine in Swarm mode the services use a name instead of a ip to communicate with each other. 'jupyterhub_service_name' is the name of ther service for the JupyterHub.
It's important to put the JupyterHub service (also the proxy) and the services that are running jupyter notebook inside the same network, otherwise they couldn't reach each other. SwarmSpawner use the service's name instead of the service's ip, as a consequence JupyterHub and servers should share the same overlay network (network across nodes).
#list of networks
c.SwarmSpawner.networks = ["mynetwork"]
You can define container_spec, *
- and networks inside jupyterhub_config.py.
The command
and args
definitions depends on the image that you are using.
I.e the command must be possible to execute in the selected image
The '/usr/local/bin/start-singleuser.sh' is provided by the jupyter
base-notebook
The start-singleuser.sh args
assumes that the launched image is extended from a version of this.
c.SwarmSpawner.container_spec = {
# The command to run inside the service
'args' : ['/usr/local/bin/start-singleuser.sh']
}
Note: in a container spec, args
sets the equivalent of CMD in the Dockerfile, command
sets the equivalent of ENTRYPOINT.
The notebook server command should not be the ENTRYPOINT, so generally use args
, not command
, to specify how to launch the notebook server.
See this issue for more info.
The spawner supports Docker Swarm service placement configurations to be imposed on the spawned services. This includes the option to specify constraints and preferences These can be imposed as a placement policy to all services being spawned. E.g.
c.SwarmSpawner.placement = {
'constraints': ['node.hostname==worker1'],
'preferences': ['spread=node.labels.datacenter']
}
To define which images are available to the users, a list of images must be declared The individual dictionaries also makes it possible to define whether the image should mount any volumes when it is spawned
# Available docker images the user can spawn
c.SwarmSpawner.images = [
{'image': 'jupyter/base-notebook:30f16d52126f',
'name': 'Minimal python notebook'},
{'image': 'jupyter/base-notebook:latest',
'name': 'Image with automatic mount, supports Py2/3 and R,',
'mounts': mounts}
]
It is also possible to specify individual placement policies for each image. E.g.
# Available docker images the user can spawn
c.SwarmSpawner.images = [
{'image': 'jupyter/base-notebook:30f16d52126f',
'name': 'Minimal python notebook',
'placement': {'constraints': ['node.hostname==worker1']}},
]
Beyond placement policy, it is also possible to specify a 'whitelist' of users who have permission to start a specific image via the 'access' key. Such that only mentioned usernames are able to spawn that particular image.
# Available docker images the user can spawn
c.SwarmSpawner.images = [
{'image': 'jupyter/base-notebook:30f16d52126f',
'name': 'Minimal python notebook',
'access': ['admin']},
]
To make the user able to select between multiple available images, the following must be set. If this is not the case, the user will simply spawn an instance of the default image. i.e. images[0]
# Before the user can select which image to spawn,
# user_options has to be enabled
c.SwarmSpawner.use_user_options = True
This enables an image select form in the users /hub/home url path when a notebook hasen't been spawned already.
With 'type':'bind'
you mount a local directory of the host inside the container.
Remember that source should exist in the node where you are creating the service.
notebook_dir = os.environ.get('NOTEBOOK_DIR') or '/home/jovyan/work'
c.SwarmSpawner.notebook_dir = notebook_dir
mounts = [{'type' : 'bind',
'source' : 'MountPointOnTheHost',
'target' : 'MountPointInsideTheContainer',}]
With 'type':'volume'
you mount a Docker Volume inside the container.
If the volume doesn't exist it will be created.
mounts = [{'type' : 'volume',
'source' : 'NameOfTheVolume',
'target' : 'MountPointInsideTheContainer',}]
For both types, volume and bind, you can specify a {name}
inside the source:
mounts = [{'type' : 'volume',
'source' : 'jupyterhub-user-{name}',
'target' : 'MountPointInsideTheContainer',}]
username will be the hashed version of the username.
This kind of volume will be removed with the service.
mounts = [{'type' : 'volume',
'source': '',
'target' : 'MountPointInsideTheContainer',}]
It is also possible to mount a volume that is an sshfs mount to another host
supports either passing {id_rsa}
or {password}
that should be used to authenticate,
in addition the typical sshfs flags are supported, defaults to port 22
from jhub.mount import SSHFSMounter
mounts = [SSHFSMounter({
'type': 'volume',
'driver_config': {
'name': 'ucphhpc/sshfs:latest',
'options' : {'sshcmd': '{sshcmd}', 'id_rsa': '{id_rsa}',
'big_writes': '', 'allow_other': '',
'reconnect': '', 'port': '2222', 'autoremove': 'True'},
}
'source': 'sshvolume-user-{name}',
'target': '/home/jovyan/work'})]
To enact that a volume should be removed when the service is being terminated, there
are two options available, either use a anonymous
volume as shown above, which will
remove the volume when the owning sevice is removed. Otherwise you can control whether volumes
should be removed or not with the service with the autoremove
label flag. e.g.
mounts = [{'type' : 'volume',
'source' : 'jupyterhub-user-{name}',
'target' : 'MountPointInsideTheContainer',
'label': {'autoremove': 'True'}}]
Or
mounts = [{'type' : 'volume',
'source' : 'jupyterhub-user-{name}',
'target' : 'MountPointInsideTheContainer',
'label': {'autoremove': 'False'}}]
With the default being 'False'.
You can also specify some resource for each service
c.SwarmSpawner.resource_spec = {
'cpu_limit' : int(1 * 1e9), # (int) – CPU limit in units of 10^9 CPU shares.
'mem_limit' : int(512 * 1e6), # (int) – Memory limit in Bytes.
'cpu_reservation' : int(1 * 1e9), # (int) – CPU reservation in units of 10^9 CPU shares.
'mem_reservation' : int(512 * 1e6), # (int) – Memory reservation in Bytes
}
There is the possibility to set parameters using user_options
# To use user_options in service creation
c.SwarmSpawner.use_user_options = False
To control the creation of the services you have 2 ways, using jupyterhub_config.py or user_options.
Remember that at the end you are just using the Docker Engine API.
user_options, if used, will overwrite jupyter_config.py for services.
If you set c.SwarmSpawner.use_user_option = True
the spawner will use the dict passed through the form or as json body when using the Hub Api.
The spawner expect a dict with these keys:
user_options = {
'container_spec' : {
# (string or list) command to run in the image.
'args' : ['/usr/local/bin/start-singleuser.sh'],
# name of the image
'Image' : '',
'mounts' : mounts,
'
' : {
# (int) – CPU limit in units of 10^9 CPU shares.
'cpu_limit': int(1 * 1e9),
# (int) – Memory limit in Bytes.
'mem_limit': int(512 * 1e6),
# (int) – CPU reservation in units of 10^9 CPU shares.
'cpu_reservation': int(1 * 1e9),
# (int) – Memory reservation in bytes
'mem_reservation': int(512 * 1e6),
},
# dict of constraints
'placement' : {'constraints': []},
# list of networks
'network' : [],
# name of service
'name' : ''
}
}
When JupyterHub spawns a new Jupyter notebook server the name of the service will be {service_prefix}-{service_owner}-{service_suffix}
You can change the service_prefix in this way:
Prefix of the service in Docker
c.SwarmSpawner.service_prefix = "jupyterhub"
service_owner
is the hexdigest() of the hashed user.name
.
In case of named servers (more than one server for user) service_suffix
is the name of the server, otherwise is always 1.
Docker Engine in Swarm mode downloads images automatically from the repository. Either the image is available on the remote repository or locally, if not you will get an error.
Because before starting the service you have to complete the download of the image is better to have a longer timeout (default is 30 secs)
c.SwarmSpawner.start_timeout = 60 * 5
You can use all the docker images inside the Jupyter docker-stacks.
DockerSpawner CassinyioSpawner
All code is licensed under the terms of the revised BSD license.