opentracing-python-instrumentation

A collection of instrumentation tools to enable tracing with OpenTracing API.

Module

Make sure you are running recent enough versions of pip and setuptools, e.g. before installing your project requirements execute this:

pip install --upgrade "setuptools>=29" "pip>=9"

The module name is opentracing_instrumentation.

What's inside

Supported client frameworks

The following libraries are instrumented for tracing in this module:

• boto3 — AWS SDK for Python
• Celery — Distributed Task Queue
• urllib2
• requests
• SQLAlchemy
• MySQLdb
• psycopg2
• Tornado HTTP client
• redis

Limitations

For some operations, Boto3 uses ThreadPoolExecutor under the hood. So, in order to make it thread-safe, the instrumentation is implemented using span_in_stack_context() which forces you to use TornadoScopeManager.

Server instrumentation

For inbound requests a helper function before_request is provided for creating middleware for frameworks like Flask and uWSGI.

Manual instrumentation

Finally, a @traced_function decorator is provided for manual instrumentation.

In-process Context Propagation

As part of the OpenTracing 2.0 API, in-process Span propagation happens through the newly defined ScopeManager interface. However, the existing functionality has been kept to provide backwards compatibility and ease code migration:

span_in_context() implements context propagation using the current opentracing.tracer.scope_manager, expected to be a thread-local based ScopeManager, such as opentracing.scope_managers.ThreadLocalScopeManager.

span_in_stack_context() implements context propagation for Tornado applications using the current opentracing.tracer.scope_manager too, expected to be an instance of opentracing.scope_managers.tornado.TornadoScopeManager.

get_current_span() returns the currently active Span, if any.

Direct access to the request_context module as well as usage of RequestContext and RequestContextManager have been fully deprecated, as they do not integrate with the new OpenTracing 2.0 API. Using them along get_current_span() is guaranteed to work, but it is highly recommended to switch to the previously mentioned functions.

Usage

This library provides two types of instrumentation, explicit instrumentation for server endpoints, and implicit instrumentation for client call sites.

Server endpoints are instrumented by creating a middleware class that:

1. initializes the specific tracer implementation
2. wraps incoming request handlers into a method that reads the incoming tracing info from the request and creates a new tracing Span

Client call sites are instrumented implicitly by executing a set of available client_hooks that monkey-patch some API points in several common libraries like SQLAlchemy, urllib2, Tornado Async HTTP Client. The initialization of those hooks is usually also done from the middleware class's __init__ method.

There is a client-server example using this library with Flask instrumentation from opentracing-contrib: https://github.com/opentracing-contrib/python-flask/tree/master/example.

Here's an example of a middleware for Clay framework:

from opentracing_instrumentation import span_in_context
from opentracing_instrumentation.http_server import before_request
from opentracing_instrumentation.http_server import WSGIRequestWrapper
from opentracing_instrumentation.client_hooks import install_all_patches


class TracerMiddleware(object):

    def __init__(self, app, wsgi_app):
        self.wsgi_app = wsgi_app
        self.service_name = app.name

        CONFIG.app_name = self.service_name
        CONFIG.caller_name_headers.append('X-Uber-Source')
        CONFIG.callee_endpoint_headers.append('X-Uber-Endpoint')

        install_all_patches()
        self.wsgi_app = create_wsgi_middleware(wsgi_app)
        self.init_tracer()

    def __call__(self, environ, start_response):
        return self.wsgi_app(environ, start_response)

    def init_tracer(self):
        # code specific to your tracer implementation
        pass


def create_wsgi_middleware(other_wsgi, tracer=None):
    """
    Create a wrapper middleware for another WSGI response handler.
    If tracer is not passed in, 'opentracing.tracer' is used.
    """

    def wsgi_tracing_middleware(environ, start_response):
        # TODO find out if the route can be retrieved from somewhere

        request = WSGIRequestWrapper.from_wsgi_environ(environ)
        span = before_request(request=request, tracer=tracer)

        # Wrapper around the real start_response object to log
        # additional information to opentracing Span
        def start_response_wrapper(status, response_headers, exc_info=None):
            if exc_info is not None:
                span.set_tag('error', str(exc_info))
            span.finish()

            return start_response(status, response_headers)

        with span_in_context(span):
            return other_wsgi(environ, start_response_wrapper)

    return wsgi_tracing_middleware

And here's an example for middleware in Tornado-based app:

import opentracing
from opentracing.scope_managers.tornado import TornadoScopeManager
from opentracing_instrumentation import span_in_stack_context, http_server


opentracing.tracer = MyOpenTracingTracer(scope_manager=TornadoScopeManager())


class TracerMiddleware(object):

    def __init__(self):
        # perform initialization similar to above, including installing
        # the client_hooks

    @gen.coroutine
    def __call__(self, request, handler, next_mw):
        request_wrapper = http_server.TornadoRequestWrapper(request=request)
        span = http_server.before_request(request=request_wrapper)

        @gen.coroutine
        def next_middleware_with_span():
            yield next_mw()

        yield run_coroutine_with_span(span=span,
                                      func=next_middleware_with_span)

        span.finish()


def run_coroutine_with_span(span, func, *args, **kwargs):
    """Wrap the execution of a Tornado coroutine func in a tracing span.

    This makes the span available through the get_current_span() function.

    :param span: The tracing span to expose.
    :param func: Co-routine to execute in the scope of tracing span.
    :param args: Positional args to func, if any.
    :param kwargs: Keyword args to func, if any.
    """
    with span_in_stack_context(span):
        return func(*args, **kwargs)

Customization

For the requests library, in case you want to set custom tags to spans depending on content or some metadata of responses, you can set response_handler_hook. The hook must be a method with a signature (response, span), where response and span are positional arguments, so you can use different names for them if needed.

from opentracing_instrumentation.client_hooks.requests import patcher


def hook(response, span):
    if not response.ok:
        span.set_tag('error', 'true')


patcher.set_response_handler_hook(hook)

If you have issues with getting the parent span, it is possible to override default function that retrieves parent span.

from opentracing_instrumentation.client_hooks import install_all_patches,
     set_current_span_func

set_current_span_func(my_custom_extractor_func)
install_all_patches()

Development

PostgreSQL, RabbitMQ, Redis, and DynamoDB are required for certain tests.

docker-compose up -d

To prepare a development environment please execute the following commands.

virtualenv env
source env/bin/activate
make bootstrap
make test

You can use tox to run tests as well.

tox