def setup_persistent_compute_target(workspace, cluster_name, vm_size,
                                    max_nodes):
    """
    Set up a persistent compute target on AzureML.
    A persistent compute target runs noticeably faster than a
    regular compute target for subsequent runs.  The benefit
    is that AzureML manages turning the compute on/off as needed for
    each job so the user does not need to do this.

    Args:
        workspace    (str): Centralized location on Azure to work with
                         all the
                                artifacts used by AzureML service
        cluster_name (str): the Azure cluster for this run. It can
                            already exist or it will be created.
        vm_size      (str): Azure VM size, like STANDARD_D3_V2
        max_nodes    (int): Number of VMs, max_nodes=4 will
                            autoscale up to 4 VMs
    Returns:
        cpu_cluster : cluster reference
    """
    # setting vmsize and num nodes creates a persistent AzureML
    # compute resource

    logger.debug("setup: cluster_name {}".format(cluster_name))
    # https://docs.microsoft.com/en-us/azure/machine-learning/service/how-to-set-up-training-targets

    try:
        cpu_cluster = ComputeTarget(workspace=workspace, name=cluster_name)
        logger.debug('setup: Found existing cluster, use it.')
    except ComputeTargetException:
        logger.debug('setup: create cluster')
        compute_config = AmlCompute.provisioning_configuration(
                       vm_size=vm_size,
                       max_nodes=max_nodes)
        cpu_cluster = ComputeTarget.create(workspace,
                                           cluster_name,
                                           compute_config)
    cpu_cluster.wait_for_completion(show_output=True)
    return cpu_cluster