nd_service_registry is a Python module that provides a simple way to leverage Apache Zookeeper as a dynamic configuration and service registry.
The goal of the package is to provide a single foundational class that can be leveraged by any Python program for registering and monitoring services through Zookeeper.
The current use cases are:
- Register a server providing a service
- Retrieve a list of servers providing a particular service
- Execute callback methods whenever a service list changes
The main benefit of using this module is if you have several different tools in your network that all need to interact with Zookeeper in a common way. The most common functions are handled in this singular module allowing you to focus on your own app development more.
To install, run
python setup.py install
or
pip install nd_service_registry
Create a logger object:
>>> import logging >>> logger = logging.getLogger() >>> logger.setLevel(logging.DEBUG) >>> handler = logging.StreamHandler() >>> logger.addHandler(handler)
To create your initial connection object:
>>> from nd_service_registry import KazooServiceRegistry >>> nd = KazooServiceRegistry()
The KazooServiceRegistry object is a child of nd_service_registry that conforms to our ServiceRegistry specs, while leveraging Kazoo as the backend. The object handles all of your connection states - there is no need to start/stop or monitor the connection state at all.
To register the host as providing a particular service:
>>> nd.set_node('/services/ssh/server1:22', data={ 'foo': 'bar'})
Getting a list of servers at a path:
>>> nd.get('/services/ssh')
{'children': {u'server1:22': {u'foo': u'bar',
u'created': u'2012-12-15 00:45:03',
u'pid': 10733}},
'data': None,
'path': '/services/ssh',
'stat': ZnodeStat(czxid=6, mzxid=6, ctime=1355532303688,
mtime=1355532303688, version=0, cversion=1,
aversion=0, ephemeralOwner=0, dataLength=0,
numChildren=1, pzxid=7)}
One of Zookeepers great features is using it as a global lock manager. We provide two models for getting a lock. In one model, your lock is only active as long as your code is running:
>>> with nd.get_lock('/foo', simultaneous=1):
... <do some work>
...
>>>
Another example is explicitly locking a path for some period of time, then releasing it explicitly (eg, locking during one method, and waiting for an entirely different method to handle the unlock):
>>> nd.acquire_lock('/foo', simultaneous=1)
>>> <do your work... >
>>> nd.release_lock('/foo')
We wrote this code to be easy to use in Django. Here's a (very brief) version of the module we use in Django to simplify use of nd_service_registry.
<your django tree>/foo/service_registry_utils.py:
import nd_service_registry, time
from django.conf import settings
_service_registry = None
def get_service_registry():
global _service_registry
if not _service_registry:
server = "%s:%s" % (settings.SERVICEREGISTRY_PARAMS['SERVER'],
settings.SERVICEREGISTRY_PARAMS['PORT'])
_service_registry = nd_service_registry.KazooServiceRegistry(
server=server,
lazy=True,
readonly=True,
timeout=settings.SERVICEREGISTRY_PARAMS['TIMEOUT'],
cachefile=settings.SERVICEREGISTRY_PARAMS['CACHEFILE'])
return _service_registry
def get(path, callback=None, wait=None):
if not wait:
return get_service_registry().get(path, callback=callback)
begin = time.time()
while time.time() - begin <= wait:
data = get_service_registry().get(path)
if len(data['children']) > 0:
if callback:
get_service_registry().add_callback(path, callback=callback)
return data
time.sleep(0.1)
return get_service_registry().get(path, callback=callback)
Example use in your code:
>>> from nextdoor import service_registry_utils
>>> def do_something(data):
... print "New server data: %s" % data
...
>>> service_registry_utils.get('/services/staging/uswest2/memcache',
... callback=do_something)
New server data: { 'path': '/services/staging/uswest2/memcache',
'stat': ZnodeStat(czxid=8589934751, mzxid=8589934751,
ctime=1354785240728, mtime=1354785240728,
version=0, cversion=45, aversion=0,
ephemeralOwner=0, dataLength=0, numChildren=1,
pzxid=30064903926),
'data': None,
'children': { u'ec2-123-123-123-123.us-west-2.compute.amazonaws.com:11211':
{u'created': u'2013-01-08 16:51:12', u'pid': 3246, }
}
}
- Warning: LC_ALL and LANG settings
Due to an unknown bug, if Django cannot find your LC_ALL LOCALE settings (which often default to 'C'), nd_service_registry or kazoo crash and burn during the init phase. Its unknown why at this point, but we've found that its best to unset LC_ALL and set LANG=en_US:UTF-8 (or some other valid setting) before you start up your Django app.
If you use Celery, set these options in /etc/default/celeryd.
If you use uWSGI, set them in your uWSGI config file.
Running the Django shell:
# unset LC_ALL; LANG=en_US:UTF-8 python manage.py shell
The ServiceRegistry object tries everything that it can to make sure that the backend Zookeeper connection is always up and running.
- Fork Behavior
In the event that your code has created an ServiceRegistry object but then gone and forked the process (celery, as an example), we do our best to detect this and re-create the connection, watchers and registrations.
When we detect a fork (more on that below), we re-create our Zookeeper connection, and then re-create all Watcher and Registration objects as well.
- Fork Detection
Detecting the fork is extremely tricky... we can only really detect it when you call the module for new data. This means that if you have created a Watcher or Registration object, those objects will not be aware of the fork (and thus the loss of their connection to Zookeeper) until you make another call to them.
Because of this, I strongly recommend that if you can detect the fork from within your application (Django signals perhaps?), you should immediately call the rebuild() method on your ServiceRegistry object.:
>>> from nd_service_registry import KazooServiceRegistry >>> k = KazooServiceRegistry() >>> do_fork() >>> k.rebuild()
The goal of this module is to be as self-contained as possible and require as little code in your app as possible. To that end, we almost never raise an Exception once the module is loaded up and connected.
We do raise a few exceptions, and each one is documented here. Whenever we can though, we instead just return False as a way of indicating that we were unable to perform your command now ... but that we will take care of it later. Whenever we do this, we throw a WARNING log message as well.
- nd_service_registry.exceptions.NoConnection
Thrown if you attempt any operation that requires immediate access to the backend Zookeeper service. Either a set() operation, or a get() operation on a path for the first time.
Also thrown during initial connection to Zookeeper, if lazy=False.
(It should be noted, a get() will actually return the cached results even if Zookeeper is down. This allows the service to fail temporarily in the background but your app is still able to get the last known results.)
- nd_service_registry.exceptions.ReadOnly
- If readonly=True, any operation that would result in a write will throw this exception. Most notably, a set() operation will fail with this exception if readonly=True.
Detailed implementation details and instructions are in the individual library files.
Running them
$ python setup.py test running test test_creates_simple_json_outputformat (nd_service_registry.bin.ndsr.get_tests.GetTests) ... ok ... ... test_default_data_produces_expected_dict (nd_service_registry.funcs_tests.FuncsTests) ... ok test_encode_creates_dict_from_single_string (nd_service_registry.funcs_tests.FuncsTests) ... ok Ran 15 tests in 0.108s OK
Integration tests require you have a Vagrant VM up and running with Zookeeper available. This VM can be easily setup
$ vagrant plugin install vagrant-berkshelf Installing the 'vagrant-berkshelf' plugin. This can take a few minutes... Installed the plugin 'vagrant-berkshelf (1.3.4)'! $ vagrant up ... $ python setup.py integration ... $ running integration ... Make sure that the enter/exit functionality works in non-blocking ... ok Test that a blocking Lock works ... ok test_decode_converts_json_to_dict (nd_service_registry.funcs_tests.FuncsTests) ... ok ---------------------------------------------------------------------- Ran 19 tests in 3.360s OK