Run on Kubernetes

Note

Support for kubernetes is experimental. APIs may change in future releases.

Mars can run in clusters managed by Kubernetes. You can use mars.deploy.kubernetes to set up a Mars cluster.

Basic steps

Mars uses image repository marsproject/mars by default. Each released version of Mars has its image since v0.3.0. For instance, the image for version 0.3.0 is marsproject/mars:v0.3.0. If you need to build an image from source, you may run the command below:

bin/kube-image-tool.sh build

A docker image with Mars tagged with the current version will be built.

Then you need to make sure if you have correct client configurations for Kubernetes by running

If it reports an error, please consult documentations for kubernetes or your cluster maintainer for more information.

As Mars uses Python to operate on Kubernetes, you also need to install Kubernetes client for Python locally. It can be installed with pip or conda:

# install with pip
pip install kubernetes
# install with conda
conda install -c conda-forge python-kubernetes

After all these steps we can create a Mars cluster with one scheduler, one worker and one web service with kubernetes and run some jobs on it:

from kubernetes import config
from mars.deploy.kubernetes import new_cluster
import mars.tensor as mt

cluster = new_cluster(config.new_client_from_config())

# new cluster will start a session and set it as default one
# execute will then run in the local cluster
a = mt.random.rand(10, 10)
a.dot(a.T).execute()

# after all jobs executed, you can turn off the cluster
cluster.stop()

When you want to use this cluster elsewhere, you can obtain namespace and endpoint from the custer object and create another KubernetesClusterClient:

# obtain information from current cluster
namespace, endpoint = cluster.namespace, cluster.endpoint

# create a new cluster client
from kubernetes import config
from mars.deploy.kubernetes import KubernetesClusterClient

cluster = KubernetesClusterClient(
    config.new_client_from_config(), namespace, endpoint)

Customizing cluster

new_cluster function provides several keyword arguments for users to define the cluster. You may use the argument image to specify the image for all Mars pods, and the argument timeout to specify timeout of cluster creation. Arguments for scaling up and out of the cluster are also available.

Arguments for schedulers:

Argument Description
scheduler_num Number of schedulers in the cluster, 1 by default
scheduler_cpu Number of CPUs for every scheduler
scheduler_mem Memory size for schedulers in the cluster, in bytes or size units like 1g

Arguments for workers:

Argument Description
worker_num Number of workers in the cluster, 1 by default
worker_cpu Number of CPUs for every worker
worker_mem Memory size for workers in the cluster, in bytes or size units like 1g
worker_spill_paths List of spill paths for worker pods on hosts
worker_cache_mem Size or ratio of shared memory for every worker. Details about memory management of Mars workers can be found in memory tuning section.
min_worker_num Minimal number of ready workers for new_cluster to return, worker_num by default

Arguments for web services:

Argument Description
web_num Number of web services in the cluster, 1 by default
web_cpu Number of CPUs for every web service
web_mem Memory size for web services in the cluster, in bytes or size units like 1g

For instance, if you want to create a Mars cluster with 1 scheduler, 1 web service and 100 workers, each worker has 4 cores and 16GB memory, and stop waiting when 95 workers are ready, we can use the code below:

from kubernetes import config
from mars.deploy.kubernetes import new_cluster

api_client = config.new_client_from_config()
cluster = new_cluster(api_client, scheduler_num=1, web_num=1, worker_num=100,
                      worker_cpu=4, worker_mem='16g', min_worker_num=95)

Implementation details

When new_cluster is called, it will create an independent namespace for all objects including roles, role bindings, pods and services. When the user destroys the service, the whole namespace will be destroyed.

Schedulers, workers and web services are created with replication controllers. Services discover schedulers by directly accessing Kubernetes API via the default service account. Pod addresses and their readiness are read by workers and web services to decide whether to start. Meanwhile the client read statuses of all pods and check if all schedulers, web services and at least min_worker_num workers are ready.

The readiness of Mars services are decided by readiness probes whose result can be obtained via Pod statuses. For schedulers and workers, when the service starts, a ReadinessActor will be created in the service and the probe can detect it. For web services, the web port is detected.

As the default service account does not have privilege to read pods in Kubernetes API, we create roles with capability to read and watch pods using RBAC API, and then bind them to default service accounts within the namespace before creating replication controllers. This enables Mars containers to detect the status of other containers.

Mars uses Kubernetes services to expose its service. Currently only NodePort mode is supported, and Mars looks for the host hosting the pod of a web service as its endpoint. LoadBalancer mode is not supported yet.