This is a client-side library that can be used to instrument Python apps for distributed trace collection, and to send those traces to Jaeger. See the OpenTracing Python API for additional detail.
Please see CONTRIBUTING.md.
pip install jaeger-clientimport logging
import time
from jaeger_client import Config
if __name__ == "__main__":
log_level = logging.DEBUG
logging.getLogger('').handlers = []
logging.basicConfig(format='%(asctime)s %(message)s', level=log_level)
config = Config(
config={ # usually read from some yaml config
'sampler': {
'type': 'const',
'param': 1,
},
'logging': True,
},
service_name='your-app-name',
validate=True,
)
# this call also sets opentracing.tracer
tracer = config.initialize_tracer()
with tracer.start_span('TestSpan') as span:
span.log_kv({'event': 'test message', 'life': 42})
with tracer.start_span('ChildSpan', child_of=span) as child_span:
span.log_kv({'event': 'down below'})
time.sleep(2) # yield to IOLoop to flush the spans - https://github.com/jaegertracing/jaeger-client-python/issues/50
tracer.close() # flush any buffered spansNOTE: If you're using the Jaeger all-in-one Docker image (or similar) and want to run Jaeger in a separate container from your app, use the code below to define the host and port that the Jaeger agent is running on. Note that this is not recommended, as Jaeger sends spans over UDP and UDP does not guarantee delivery. (See this thread for more details.)
config = Config(
config={ # usually read from some yaml config
'sampler': {
'type': 'const',
'param': 1,
},
'local_agent': {
'reporting_host': 'your-reporting-host',
'reporting_port': 'your-reporting-port',
},
'logging': True,
},
service_name='your-app-name',
validate=True,
)The opentracing-contrib project has a few modules that provide explicit instrumentation support for popular frameworks like Django and Flask.
At Uber we are mostly using the opentracing_instrumentation module that provides:
- explicit instrumentation for HTTP servers, and
- implicit (monkey-patched) instrumentation for several popular libraries like
urllib2,redis,requests, some SQL clients, etc.
Note: do not initialize the tracer during import, it may cause a deadlock (see issues #31, #60). Instead define a function that returns a tracer (see example below) and call that function explicitly after all the imports are done.
Also note that using gevent.monkey in asyncio-based applications (python 3+) may need to pass current event loop explicitly (see issue #256):
from tornado import ioloop
from jaeger_client import Config
config = Config(config={}, service_name='your-app-name', validate=True)
config.initialize_tracer(io_loop=ioloop.IOLoop.current())The recommended way to initialize the tracer for production use:
from jaeger_client import Config
def init_jaeger_tracer(service_name='your-app-name'):
config = Config(config={}, service_name=service_name, validate=True)
return config.initialize_tracer()Note that the call initialize_tracer() also sets the opentracing.tracer global variable.
If you need to create additional tracers (e.g., to create spans on the client side for remote services that are not instrumented), use the new_tracer() method.
This module brings a Prometheus integration to the internal Jaeger metrics.
The way to initialize the tracer with Prometheus metrics:
from jaeger_client.metrics.prometheus import PrometheusMetricsFactory
config = Config(
config={},
service_name='your-app-name',
validate=True,
metrics_factory=PrometheusMetricsFactory(namespace='your-app-name')
)
tracer = config.initialize_tracer()For development, some parameters can be passed via config dictionary, as in the Getting Started example above. For more details please see the Config class.
When using this library in applications that fork child processes to handle individual requests,
such as with WSGI / PEP 3333, care must be taken when initializing the tracer.
When Jaeger tracer is initialized, it may start a new background thread. If the process later forks,
it might cause issues or hang the application (due to exclusive lock on the interpreter).
Therefore, it is recommended that the tracer is not initialized until after the child processes
are forked. Depending on the WSGI framework you might be able to use @postfork decorator
to delay tracer initialization (see also issues #31, #60).
The OpenTracing API defines a sampling.priority standard tag that
can be used to affect the sampling of a span and its children:
from opentracing.ext import tags as ext_tags
span.set_tag(ext_tags.SAMPLING_PRIORITY, 1)Jaeger Tracer also understands a special HTTP Header jaeger-debug-id,
which can be set in the incoming request, e.g.
curl -H "jaeger-debug-id: some-correlation-id" http://myhost.comWhen Jaeger sees this header in the request that otherwise has no tracing context, it ensures that the new trace started for this request will be sampled in the "debug" mode (meaning it should survive all downsampling that might happen in the collection pipeline), and the root span will have a tag as if this statement was executed:
span.set_tag('jaeger-debug-id', 'some-correlation-id')This allows using Jaeger UI to find the trace by this tag.
To use this library directly with other Zipkin libraries & backend,
you can provide the configuration property propagation: 'b3' and the
X-B3-* HTTP headers will be supported.
The B3 codec assumes it will receive lowercase HTTP headers, as this seems to be the standard in the popular frameworks like Flask and Django. Please make sure your framework does the same.