Instrumenting custom codeedit

Creating Additional Spans in a Transactionedit

Elastic APM instruments a variety of libraries out of the box, but sometimes you need to know how long a specific function took or how often it gets called.

Assuming you’re using one of our supported frameworks, you can apply the @elasticapm.capture_span() decorator to achieve exactly that. If you’re not using a supported framework, see Creating New Transactions.

elasticapm.capture_span can be used either as a decorator or as a context manager. The following example uses it both ways:

import elasticapm

def coffee_maker(strength):

    with elasticapm.capture_span('near-to-machine'):
        for i in range(strength):



Similarly, you can use elasticapm.async_capture_span for instrumenting async workloads:

import elasticapm

async def coffee_maker(strength):
    await fetch_water()

    async with elasticapm.async_capture_span('near-to-machine'):
        await insert_filter()
        async for i in range(strength):
            await pour_coffee()



asyncio support is only available in Python 3.7+.

See the API docs for more information on capture_span.

Creating New Transactionsedit

It’s important to note that elasticapm.capture_span only works if there is an existing transaction. If you’re not using one of our supported frameworks, you need to create a Client object and begin and end the transactions yourself. You can even utilize the agent’s automatic instrumentation!

To collect the spans generated by the supported libraries, you need to invoke elasticapm.instrument() (just once, at the initialization stage of your application) and create at least one transaction. It is up to you to determine what you consider a transaction within your application — it can be the whole execution of the script or a part of it.

The example below will consider the whole execution as a single transaction with two HTTP request spans in it. The config for elasticapm.Client can be passed in programmatically, and it will also utilize any config environment variables available to it automatically.

import requests
import time
import elasticapm

def main():
    sess = requests.Session()
    for url in [ '', '' ]:
        resp = sess.get(url)

if __name__ == '__main__':
    client = elasticapm.Client(service_name="foo", server_url="")
    elasticapm.instrument()  # Only call this once, as early as possible.
    client.end_transaction(name=__name__, result="success")

Note that you don’t need to do anything to send the data — the Client object will handle that before the script exits. Additionally, the Client object should be treated as a singleton — you should only create one instance and store/pass around that instance for all transaction handling.

Distributed Tracingedit

When instrumenting custom code across multiple services, you should propagate the TraceParent where possible. This allows Elastic APM to bundle the various transactions into a single distributed trace. The Python Agent will automatically add TraceParent information to the headers of outgoing HTTP requests, which can then be used on the receiving end to add that TraceParent information to new manually-created transactions.

Additionally, the Python Agent provides utilities for propagating the TraceParent in string format.

import elasticapm

client = elasticapm.Client(service_name="foo", server_url="")

# Retrieve the current TraceParent as a string, requires active transaction
traceparent_string = elasticapm.get_trace_parent_header()

# Create a TraceParent object from a string and use it for a new transaction
parent = elasticapm.trace_parent_from_string(traceparent_string)
client.begin_transaction(transaction_type="script", trace_parent=parent)
# Do some work
client.end_transaction(name=__name__, result="success")

# Create a TraceParent object from a dictionary of headers, provided
# automatically by the sending service if it is using an Elastic APM Agent.
parent = elasticapm.trace_parent_from_headers(headers_dict)
client.begin_transaction(transaction_type="script", trace_parent=parent)
# Do some work
client.end_transaction(name=__name__, result="success")