pytest-sentry 0.1.5

Tracking flaky tests as errors

Let’s say you have a testsuite with some flaky tests that randomly break your CI build due to network issues, race conditions or other stuff that you don’t want to fix immediately. The known workaround is to retry those tests automatically, for example using pytest-rerunfailures.

One concern against plugins like this is that they just hide the bugs in your testsuite or even other code. After all your CI build is green and your code probably works most of the time.

pytest-sentry tries to make that choice a bit easier by tracking flaky test failures in a place separate from your build status. Sentry is already a good choice for keeping tabs on all kinds of errors, important or not, in production, so let’s try to use it in testsuites too.

The prerequisite is that you already make use of pytest and pytest-rerunfailures in CI. Now install pytest-sentry and set the PYTEST_SENTRY_DSN environment variable to the DSN of a new Sentry project.

Now every test failure that is “fixed” by retrying the test is reported to Sentry, but still does not break CI. Tests that consistently fail will not be reported.

Tracking the performance of your testsuite

By default pytest-sentry will send Performance data to Sentry:

• Fixture setup is reported as “transaction” to Sentry, such that you can answer questions like “what is my slowest test fixture” and “what is my most used test fixture”.

• Calls to the test function itself are reported as separate transaction such that you can find large, slow tests as well.

To measure performance data, install pytest-sentry and set PYTEST_SENTRY_DSN, like with errors.

Transactions can have noticeable runtime overhead over just reporting errors. To disable, use a marker:

import pytest
import pytest_sentry

pytestmarker = pytest.mark.sentry_client({"traces_sample_rate": 0.0})

Advanced Options

pytest-sentry supports marking your tests to use a different DSN, client or hub per-test. You can use this to provide custom options to the Client object from the Sentry SDK for Python:

import random
import pytest

from sentry_sdk import Hub
from pytest_sentry import Client

@pytest.mark.sentry_client(None)
def test_no_sentry():
    # Even though flaky, this test never gets reported to sentry
    assert random.random() > 0.5

@pytest.mark.sentry_client("MY NEW DSN")
def test_custom_dsn():
    # Use a different DSN to report errors for this one
    assert random.random() > 0.5

# Other invocations:

@pytest.mark.sentry_client(Client("CUSTOM DSN"))
@pytest.mark.sentry_client(lambda: Client("CUSTOM DSN"))
@pytest.mark.sentry_client(Hub(Client("CUSTOM DSN")))
@pytest.mark.sentry_client({"dsn": ..., "debug": True})

The Client class exposed by pytest-sentry only has different default integrations. It disables some of the error-capturing integrations to avoid sending random expected errors into your project.

Accessing the used Sentry client

You will notice that the global functions such as sentry_sdk.capture_message will not actually send events into the same DSN you configured this plugin with. That’s because pytest-sentry goes to extreme lenghts to keep its own SDK setup separate from the SDK setup of the tested code.

pytest-sentry exposes the sentry_test_hub fixture whose return value is the Hub being used to send events to Sentry. Use with sentry_test_hub: to temporarily switch context. You can use this to set custom tags like so:

def test_foo(sentry_test_hub):
    with sentry_test_hub:
        sentry_sdk.set_tag("pull_request", os.environ['EXAMPLE_CI_PULL_REQUEST'])

Why all the hassle with the context manager? Just imagine if your tested application would start to log some (expected) errors on its own. You would immediately exceed your quota!

Always reporting test failures

You can always report all test failures to Sentry by setting the environment variable PYTEST_SENTRY_ALWAYS_REPORT=1.

This can be enabled for builds on the master or release branch, to catch certain kinds of tests that are flaky across builds, but consistently fail or pass within one testrun.

License

Licensed under MIT, see LICENSE.