elmiko's notes

Shout out to Erik Erlandson for introducing me to Baker.

Recently I have been working on some issues in Sahara which require me to observe the creation and destruction of multiple clusters. I also frequently need to restart DevStack and then get my stack configured for the tests I want to run. In the past I have used the Horizon interface and the ReST API to perform operations with Sahara. This time around I wanted to investigate the Python saharaclient interface to do the same.

One of the things I really enjoy about the Python interface to OpenStack is using the REPL to query and manipulate a stack. I like the immediacy of being able to interact with the stack and run small pieces of code if need be. The objects also act as a quick way to filter the information coming back from the various services.

Much of this could be achieved with a few bash scripts and templated json files through the ReST interface, or with the Sahara command line tool, and I don’t consider the Python interface to be a superior method of interactions, but I do find it to be a useful tool for my needs and an interesting window into OpenStack. I’m also a child of Logo , which makes REPL type software feel very familiar to me.

####Setup

I am running all these commands against an install of DevStack that I have slightly customized by changing the passwords and some of the services. I am running the Neutron network service and I have disabled Sahara because I want to run a custom version for debugging. All of this code should run on a default installation of Sahara, so it should not be necessary to run a custom install.

####Code

What follows are snippets of code that I assembled into a small file to run from the command line for automating cluster configuration. There are a few other steps I optionally perform with Keystone and Swift, but I’ll save those for a more general post. I’ll go over each of the snippets with a small explanation and then present the entire script at the end.

First I setup some variables to use with the clients I’ll need to interact with the stack. I’m using the v2 Keystone interface here more for convenience as I’ve found some inconsistencies when using v3. I need to keep track of the different project names so that I can create an admin scoped token with the Neutron client and the non-admin for my Sahara work.

ADMIN_USERNAME='admin'
ADMIN_PASSWORD='openstack'
ADMIN_PROJECT_NAME='admin'
AUTH_URL='http://localhost:5000/v2.0'
NETWORK_URL='http://localhost:9696'
SAHARA_PROJECT_NAME='demo'

This function creates the Sahara client, all the OpenStack Python interfaces use this methodology of client objects. The clients provide a wrapper around the ReST calls that are made to the underlying services. Most of the client interfaces I’ve worked with also provide manager objects to work with the underlying collections that they wrap. In the case of a Sahara client object this might look like client.clusters, which returns an interface to the /clusters endpoint.

def sahara_client():
    return sahara.Client(auth_url=AUTH_URL,
                         username=ADMIN_USERNAME,
                         api_key=ADMIN_PASSWORD,
                         project_name=SAHARA_PROJECT_NAME)

In order to create cluster templates I need to know the management and floating pool network identifiers I’d like to use with my cluster instances. To do this I’m going to instantiate a Neutron client and query the networks, this is where I need the admin project scoping.

def get_public_private_nets():
    '''Returns (public id, private id).'''
    c = neutron.Client(api_version='2.0',
                       auth_url=AUTH_URL,
                       endpoint_url=NETWORK_URL,
                       username=ADMIN_USERNAME,
                       password=ADMIN_PASSWORD,
                       tenant_name=ADMIN_PROJECT_NAME)
    retval = {}
    for net in c.list_networks().get('networks', []):
        if net.get('name') in ['public', 'private']:
            retval[net.get('name')] = net.get('id')
    return (retval.get('public'), retval.get('private'))

Next I put together a function to create a template for a cluster configured with the vanilla Hadoop 2.4.1 plugin. This is a very basic cluster but my testing is quite simple in this case. I am using the method of embedding my node group templates into the cluster template. Alternatively, I could have created node group templates separately and then referenced their identifiers in the cluster template. For my tesing purposes though, I won’t need to reuse the node groups so I’ll keep the template creation to a minimum.

It’s worth noting that for the most part the cluster template dictionary is equivalent to the JSON representation with the exception of the neutron_management_network key. The key for the management network identifier is net_id in the saharaclient. I’m not sure exactly why this exists, from looking at the source I imagine it is a backward compatibility with versions of Sahara that defaulted to Nova networking.

def cluster_template_vanilla24(name='vanilla24'):
    float_pool, mgmt_net = get_public_private_nets()
    return {
        'name': name,
        'plugin_name': 'vanilla',
        'hadoop_version': '2.4.1',
        'net_id': mgmt_net,
        'cluster_configs': {},
        'node_groups': [
            {
                'count': 1,
                'name': 'v24-master',
                'flavor_id': '2',
                'node_processes': ['namenode',
                                   'oozie',
                                   'resourcemanager',
                                   'historyserver'],
                'floating_ip_pool': float_pool
            },
            {
                'count': 3,
                'name': 'v24-worker',
                'flavor_id': '2',
                'node_processes': ['datanode',
                                   'nodemanager'],
                'floating_ip_pool': float_pool
            }
        ]
    }

Lastly, I put it all together with a wrapper function to create the cluster template on the Sahara server and return the resulting object. This function shows an example of how to use one of Sahara’s manager objects. The call to cluster_templates.create is functionally equivalent to making a POST to the /cluster-templates endpoint on the Sahara server. To further manipulate the cluster templates on the server I could also call .list to perform a GET on the endpoint, or call .get(<id>) to do the equivalent of a GET on /cluster-templates/<id>.

def create_cluster_vanilla24():
    s = sahara_client()
    t = cluster_template_vanilla24()
    return s.cluster_templates.create(**t)

At this point what I usually do is import the file in the Python REPL and use the functions directly. Something along the lines of:

>>> import config_sahara as c_s
>>> c_s.create_cluster_vanilla24()

Recently though, I have been introduced to the Baker package which facilitates making command line scripts. Instead of needing to create a rich interface with argparse all I need to do is expose a function or two. This is where Baker really helps outs. Instead of invoking the Python REPL I might just run the script from the command line. When combined with multiple cluster templates this becomes a quick way to get things setup. For example:

$ ./configs_sahara.py vanilla24
Cluster Template {u'neutron_management_network': u'23d26aa0-6a86-4ee7-874f-e92d4157a0a1', u'description': None, u'cluster_configs': {}, u'created_at': u'2014-09-26 03:27:41', u'default_image_id': None, u'updated_at': None, u'plugin_name': u'vanilla', u'anti_affinity': [], u'tenant_id': u'e4d4f5b56ed647e3bc4cef323432d61d', u'node_groups': [{u'count': 3, u'name': u'v24-worker', u'volume_mount_prefix': u'/volumes/disk', u'created_at': u'2014-09-26 03:27:41', u'updated_at': None, u'floating_ip_pool': u'2aa5e578-59b5-42dc-a826-1f9d0f1a0bf1', u'image_id': None, u'volumes_size': 0, u'node_processes': [u'datanode', u'nodemanager'], u'node_group_template_id': None, u'volumes_per_node': 0, u'node_configs': {}, u'auto_security_group': False, u'security_groups': None, u'flavor_id': u'2'}, {u'count': 1, u'name': u'v24-master', u'volume_mount_prefix': u'/volumes/disk', u'created_at': u'2014-09-26 03:27:41', u'updated_at': None, u'floating_ip_pool': u'2aa5e578-59b5-42dc-a826-1f9d0f1a0bf1', u'image_id': None, u'volumes_size': 0, u'node_processes': [u'namenode', u'oozie', u'resourcemanager', u'historyserver'], u'node_group_template_id': None, u'volumes_per_node': 0, u'node_configs': {}, u'auto_security_group': False, u'security_groups': None, u'flavor_id': u'2'}], u'hadoop_version': u'2.4.1', u'id': u'5193f5a9-6c85-4697-8a28-106aa00c185b', u'name': u'hdp2'}

Messy, but I think with a little help from the json package it could be made more readable. Similar output will be generated in the REPL if using the commands directly. It’s worth noting that once you put Baker into the script it won’t import cleanly anymore into the REPL. Next time I visit this topic I’ll present a little larger script showing a more indepth stack configuration for use with one of Sahara’s new features for Juno.

This is the script that I generated to use with creating vanilla Hadoop 2.4.1 and HDP 2.0.6 clusters. Thanks for reading and happy hacking!

Configuring Sahara cluster templates with the Python client