Armin Ronacher's Thoughts and Writings

Nameko for Microservices

written on Wednesday, April 8, 2015

In December some of the tech guys at onefinestay invited me over to London to do some general improvements on their nameko library. This collaboration came together because nameko was pretty similar to how I generally like to build certain infrastructure and I had some experience with very similar systems.

So now that some of those improvements hit the release version of nameko I figured it might be a good idea to give some feedback on why I like this sort of architecture.

Freeing your Mind

Right now if you want to build a web service in Python there are many tools you can pick from, but most of them live in a very specific part of your stack. The most common tool is a web framework and that will typically provide you with whatever glue is necessary to connect your own code to an incoming HTTP request that comes from your client.

However that's not all you need in an application. For instance very often you have periodic tasks that you need to execute and in that case, your framework is often not just not helping, it's also in your way. For instance because you might have built your code with the assumption that it has access to the HTTP request object. If you now want to run it from a cronjob that request object is unavailable.

In addition to crons there is often also the wish to execute something as the result of the request of a client, but without blocking that request. For instance imagine there is an admin panel in which you can trigger some very expensive data conversion task. What you actually want is for the current request to finish but the conversion task to keep on working in the background until your data set is converted.

There are obviously many existing solutions for that. Celery comes to mind. However they are typically very separated from the rest of the stack.

Having a system which treats all of this processes the same frees up your mind. This is what makes microservices interesting. Away with having HTTP request handlers that have no direct relationship with message queue worker tasks or cronjobs. Instead you can have a coherent system where any component can talk through well defined points with other parts of the system.

This is especially useful in Python where traditionally our support for parallel execution has been between very bad to abysmal.

Enter Nameko

Nameko is an implementation of this idea. It's very similar in architecture to how we structure code at Fireteam. It's based on distributing work between processes through AMQP. It's not just AMQP though. Nameko abstracts away from that and allows you to write your own transports, while staying true to the AMQP patterns.

Nameko does a handful of things and you can build very complex systems with it. The idea is that you build individual services which can emit events to which other services can subscribe to or they can directly invoke each other via RPC. All communication between the services happens through AMQP. You don't need to manually deal with any connectivity of those.

In addition to message exchange, services also use a lifecycle management to find useful resources through dependency injection. That sounds like a mouthful but is actually very simple. Because services are classes, you can add special attributes to them which will be resolved at runtime. The lifetime of the value resolved can be customized. For instance it becomes possible to attach a property to the class which can provide access to a database connection. The lifetime of that database connection can be automatically managed.

So how does that look in practice? Something like this:

from nameko.rpc import rpc

class HelloWorldService(object):
    name = 'helloworld'

    @rpc
    def hello(self, name):
        return "Hello, {}!".format(name)

This defines a basic service that provides one method that can be invoked via RPC. Either another service can do that, or any other process that runs nameko can also invoke that, for as long as they connect to the same AMQP server. To experiment with this service, Nameko provides a shell helper that launches an interactive Python shell with an n object that provides RPC access:

>>> n.rpc.helloworld.hello(name='John')
u'Hello, John!'

If the AMQP server is running, rpc.helloworld.hello contacts the helloworld service and resolves the hello method on it. Upon calling this method a message will be dispatched via the AMQP broker and be picked up by a nameko process. The shell will then block and wait for the result to come back.

A more useful example is what happens when services want to collaborate on some activity. For instance it's quite common that one service wants to respond to the changes another service performs to update it's own state. This can be achieved through events:

from nameko.events import EventDispatcher, event_handler
from nameko.rpc import rpc

class ServiceA(object):
    name = 'servicea'
    dispatch = EventDispatcher()

    @rpc
    def emit_an_event(self):
        self.dispatch('my_event_type', 'payload')


class ServiceB(object):
    name = 'serviceb'

    @event_handler('servicea', 'my_event_type')
    def handle_an_event(self, payload):
        print 'service b received', payload

The default behavior is that one service instance of each service type will pick up the event. However nameko can also route an event to every single instance of every single service. This is useful for in-process cache invalidation for instance.

The Web

Nameko is not just good for internal communication however. It uses Werkzeug to provide a bridge to the outside world. This allows you to accept an HTTP request and to ingest a task into your service world:

import json
from nameko.web.handlers import http
from werkzeug.wrappers import Response

class HttpServiceService(object):
    name = 'helloworld'

    @http('GET', '/get/<int:value>')
    def get_method(self, request, value):
        return Response(json.dumps({'value': value}),
                        mimetype='application/json')

The endpoint function can itself invoke other parts of the system via RPC or other methods.

This functionality generally also extends into the websocket world, even though that part is currently quite experimental. It for instance is possible to listen to events and forward them into websocket connections.

Dependency Injection

One of the really neat design concepts in Nameko is the use of dependency injection to find resources. A good example is the SQLAlchemy bridge which attaches a SQLAlchemy database session to a service through dependency injection. The descriptor itself will hook into the lifecycle management to automatically manage the database resources:

from nameko_sqlalchemy import Session

import sqlalchemy as sa
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class User(Base):
    __tablename__ = 'users'
    id = sa.Column(sa.Integer, primary_key=True)
    username = sa.Column(sa.String)


class MyService(object):
    name = 'myservice'
    session = Session(Base)

    @rpc
    def get_username(self, user_id):
        user = self.session.query(User).get(user_id)
        if user is not None:
            return user.username

The implementation of the Session dependency provider itself is ridiculously simple. The whole functionality could be implemented like this:

from weakref import WeakKeyDictionary

from nameko.extensions import DependencyProvider
from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker


class Session(DependencyProvider):

    def __init__(self, declarative_base):
        self.declarative_base = declarative_base
        self.sessions = WeakKeyDictionary()

    def get_dependency(self, worker_ctx):
        db_uri = self.container.config['DATABASE_URL']
        engine = create_engine(db_uri)
        session_cls = sessionmaker(bind=engine)
        self.sessions[worker_ctx] = session = session_cls()
        return session

    def worker_teardown(self, worker_ctx):
        sess = self.sessions.pop(worker_ctx, None)
        if sess is not None:
            sess.close()

The actual implementation is only a tiny bit more complicated, and that is basically just a bit of extra code to support different database URLs for different services and declarative bases. Overall the concept is the same however. When the dependency is needed, a connection to the database is established and when the worker shuts down, the session is closed.

Concurrency and Parallelism

What makes nameko interesting is that scales out really well through the use of AMQP and eventlet. First of all, when nameko starts a service container it uses eventlet to patch up the Python interpreter to support green concurrency. This allows a service container to become quite concurrent to do multiple things at once. This is very useful when a service waits on another service as threads in Python are a very disappointing story. As this however largely eliminates the possibility of true parallelism it becomes necessary to start multiple instances of services to scale up. Thanks to the use of AMQP however, this becomes a very transparent process. For as long as services do not need to store local state, it becomes very trivial to run as many of those service containers as necessary.

My Take On It

Nameko as it stands has all the right principles for building a platform out of small services and it's probably the best Open Source solution for this problem in the Python world so far.

It's a bit disappointing that Python's async story is so diverging between different Python versions and frameworks, but eventlet and gevent are by far the cleanest and most practical implementations, so for most intents and purposes the eventlet base in nameko is probably the best you can currently get for async IO. Fear not though, Nameko 2.0 now also runs on Python3.

If you haven't tried this sort of service setup yet, you might want to give Nameko a try.

This entry was tagged python and soa