Tools for ML/MXNet on Kubernetes.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

K8s Custom Resource and Operator For MXNet jobs

(Forks initally from tf-operator project now moved to


Custom Resources require Kubernetes 1.7


Distributed MXNet training jobs require managing multiple sets of MXNet replicas. There are three kinds of replicas sets which act as different roles in the job: Server (1-n instance) Worker (1-n instance) Scheduler(only one instance in a training cluster).

K8s makes it easy to configure and deploy each set of MXNet replicas. Various tools like helm and ksonnet can be used to simplify generating the configs for a MXNet job. Helm will be supported a few days later.

However, in addition to generating the configs we need some custom control logic because K8s built-in controllers (Jobs, ReplicaSets, StatefulSets, etc...) don't provide the semantics needed for managing MXNet jobs.

To solve this we define a K8S Custom Resource and Operator to manage a MXNet job on K8s.

MxJob provides a K8s resource representing a single, distributed, mxnet training job. The Spec and Status (defined in mx_job.go are customized for MXNet. The spec allows specifying the Docker image and arguments to use for each MXNet replica (i.e. scheduler, worker, and parameter server). The status provides relevant information such as the number of replicas in various states.

Using a CRD gives users the ability to create and manage mxnet Jobs just like builtin K8s resources. For example to create a job

kubectl create -f examples/mx_job_local.yaml

To list jobs

kubectl get mxjobs

NAME               KINDS


The code is closely modeled on Coreos's etcd-operator.

The MxJob Spec(defined in mx_job.go reuses the existing Kubernetes structure PodTemplateSpec to describe MXNet processes. We use PodTemplateSpec because we want to make it easy for users to configure the processes; for example setting resource requirements or adding volumes. We expect helm or ksonnet could be used to add syntactic sugar to create more convenient APIs for users not familiar with Kubernetes.

Leader election allows a K8s deployment resource to be used to upgrade the operator.

Installing the CRD and operator on your k8s cluster

  1. Clone the repository

  2. Deploy the operator

kubec create -f examples/mx_operator_deploy.yaml
  1. Create distributed mxjob demo.
kubectl create -f examples/mx_job_dist.yaml
kubectl get job
NAME                                DESIRED   SUCCESSFUL   AGE
example-dist-job-scheduler-1kkf-0   1         1            1h
example-dist-job-server-1kkf-0      1         1            1h
example-dist-job-server-1kkf-1      1         1            1h
example-dist-job-worker-1kkf-0      1         1            1h
example-dist-job-worker-1kkf-1      1         1            1h
example-dist-job-worker-1kkf-2      1         1            1h

kubectl logs -f example-dist-job-worker-1kkf-2-hz22k
INFO:root:start with arguments Namespace(batch_size=128, benchmark=0, data_nthreads=4,     data_train='data/cifar10_train.rec', data_val='data/cifar10_val.rec', disp_batches=20, dtype='float32', gpus=None, image_shape='3,28,28', kv_store='dist_device_sync', load_epoch=None, lr=0.05, lr_factor=0.1, lr_step_epochs='200,250', max_random_aspect_ratio=0, max_random_h=36, max_random_l=50, max_random_rotate_angle=0, max_random_s=50, max_random_scale=1, max_random_shear_ratio=0, min_random_scale=1, model_prefix='./wltest', mom=0.9, monitor=0, network='resnet', num_classes=10, num_epochs=10, num_examples=50000, num_layers=2, optimizer='sgd', pad_size=4, random_crop=1, random_mirror=1, rgb_mean='123.68,116.779,103.939', test_io=0, top_k=0, wd=0.0001)
[02:52:36] src/io/ ImageRecordIOParser2: data/cifar10_train.rec, use 1 threads for decoding..
[02:52:37] src/io/ ImageRecordIOParser2: data/cifar10_val.rec, use 1 threads for decoding..
INFO:root:Epoch[0] Batch [20]	Speed: 84.39 samples/sec	accuracy=0.159970
INFO:root:Epoch[0] Batch [40]	Speed: 85.53 samples/sec	accuracy=0.250391
INFO:root:Epoch[0] Batch [60]	Speed: 90.39 samples/sec	accuracy=0.284375
INFO:root:Epoch[0] Batch [80]	Speed: 89.56 samples/sec	accuracy=0.294531
INFO:root:Epoch[0] Batch [100]	Speed: 91.03 samples/sec	accuracy=0.337500
INFO:root:Epoch[0] Batch [120]	Speed: 85.62 samples/sec	accuracy=0.342187
INFO:root:Epoch[0] Train-accuracy=0.363281
INFO:root:Epoch[0] Time cost=190.507
INFO:root:Saved checkpoint to "./wltest-2-0001.params"

Project Status

This is very much a prototype.

Status information

The status information reported by the operator is hacky and not well thought out. In particular, we probably need to figure out what the proper phases and conditions to report are.

Failure/Termination Semantics

The semantics for aggregating status of individual replicas into overall MxJob status needs to be thought out.

Dead/Unnecessary code

There is a lot of code from earlier versions (including the ETCD operator) that still needs to be cleaned up.


There is minimal testing.


There are some unittests.

E2E tests

The helm package provides some basic E2E tests.

Building the Operator

Resolve dependencies (if you don't have glide install, check how to do it here)

glide install

Runing the Operator Locally

Running the operator locally (as opposed to deploying it on a K8s cluster) is convenient for debugging/development.

We can configure the operator to run locally using the configuration available in your kubeconfig to communicate with a K8s cluster.

Set your environment

export USE_KUBE_CONFIG=$(echo ~/.kube/config)
export MY_POD_NAMESPACE=default
export MY_POD_NAME=my-pod
* MY_POD_NAMESPACE is used because the CRD is namespace scoped and we use the namespace of the controller to
  set the corresponding namespace for the resource.

TODO(jlewi): Do we still need to set MY_POD_NAME? Why?

Go version

On ubuntu the default go package appears to be gccgo-go which has problems see issue golang-go package is also really old so install from golang tarballs instead.


You may need to remove the vendor directory of dependencies that also vendor dependencies as these may produce conflicts with the versions vendored by mlkube; e.g.

rm -rf  vendor/