All unit tests for Apache Airflow are run using pytest.
The outline for this document in GitHub is available at top-right corner button (with 3-dots and 3 lines).
Follow the guidelines when writing unit tests:
- For standard unit tests that do not require integrations with external systems, make sure to simulate all communications.
- All Airflow tests are run with
pytest. Make sure to set your IDE/runners (see below) to usepytestby default. - For tests, use standard "asserts" of Python and
pytestdecorators/context managers for testing rather thanunittestones. See pytest docs for details. - Use a
pytest.mark.parametrizemarker for tests that have variations in parameters. See pytest docs for details. - Use with
pytest.warnto capture warnings rather thanrecwarnfixture. We are aiming for 0-warning in our tests, so we run Pytest with--disable-warningsbut instead we have custom warning capture system.
By default, in the new tests selected warnings are prohibited:
airflow.exceptions.AirflowProviderDeprecationWarningairflow.exceptions.RemovedInAirflow3Warningairflow.utils.context.AirflowContextDeprecationWarning
That mean if one of this warning appear during test run and do not captured the test will failed.
root@91e633d08aa8:/opt/airflow# pytest tests/models/test_dag.py::TestDag::test_clear_dag
...
FAILED tests/models/test_dag.py::TestDag::test_clear_dag[None-None] - airflow.exceptions.RemovedInAirflow3Warning: Calling `DAG.create_dagrun()` without an explicit data interval is deprecatedFor avoid this make sure:
- You do not use deprecated method, classes and arguments in your test cases;
- Your change do not affect other component, e.g. deprecate one part of Airflow Core or one of Community Supported Providers might be a reason for new deprecation warnings. In this case changes should be also made in all affected components in backward compatible way.
- You use
pytest.warn(see pytest doc context manager for catch warning during the test deprecated components. Yes we still need to test legacy/deprecated stuff until it complitly removed)
def test_deprecated_argument():
with pytest.warns(AirflowProviderDeprecationWarning, match="expected warning pattern"):
SomeDeprecatedClass(foo="bar", spam="egg")Some of the unit tests require special configuration set as the default. This is done automatically by
adding AIRFLOW__CORE__UNIT_TEST_MODE=True to the environment variables in Pytest auto-used
fixture. This in turn makes Airflow load test configuration from the file
airflow/config_templates/unit_tests.cfg. Test configuration from there replaces the original
defaults from airflow/config_templates/config.yml. If you want to add some test-only configuration,
as default for all tests you should add the value to this file.
You can also - of course - override the values in individual test by patching environment variables following
the usual AIRFLOW__SECTION__KEY pattern or conf_vars context manager.
Airflow tests in the CI environment are split into several test types. You can narrow down which
test types you want to use in various breeze testing sub-commands in three ways:
- via specifying the
--test-typewhen you run single test type inbreeze testing core-tests.breeze testing providers-testsbreeze testing integration-testscommands - via specifying space separating list of test types via
--parallel-test-typesor--excluded-parallel-test-typesoptions when you run tests in parallel (in several testing commands)
Those test types are defined:
Always- those are tests that should be always executed (always sub-folder)API- Tests for the Airflow API (api, api_connexion, api_internal, api_fastapi sub-folders)CLI- Tests for the Airflow CLI (cli folder)Core- for the core Airflow functionality (core, executors, jobs, models, ti_deps, utils sub-folders)Operators- tests for the operators (operators folder)WWW- Tests for the Airflow webserver (www folder)Providers- Tests for all Providers of Airflow (providers folder)Other- all other tests remaining after the above tests are selected
We have also tests that run "all" tests (so they do not look at the folder, but at the pytest markers
the tests are marked with to run with some filters applied.
All-Postgres- tests that require Postgres database. They are only run when backend is Postgres (backend("postgres")marker)All-MySQL- tests that require MySQL database. They are only run when backend is MySQL (backend("mysql")marker)All-Quarantined- tests that are flaky and need to be fixed (quarantinedmarker)All- all tests are run (this is the default)
We also have Integration tests that are running Integration tests with external software that is run
via --integration flag in breeze environment - via breeze testing integration-tests.
Integration- tests that require external integration images running in docker-compose
This is done for three reasons:
- in order to selectively run only subset of the test types for some PRs
- in order to allow efficient parallel test execution of the tests on Self-Hosted runners
For case 2. We can utilize memory and CPUs available on both CI and local development machines to run test in parallel, but we cannot use pytest xdist plugin for that - we need to split the tests into test types and run each test type with their own instance of database and separate container where the tests in each type are run with exclusive access to their database and each test within test type runs sequentially. By the nature of those tests - they rely on shared databases - and they update/reset/cleanup data in the databases while they are executing.
There are two kinds of unit tests in Airflow - DB and non-DB tests. This chapter describe the differences between those two types.
For the Non-DB tests, they are run once for each tested Python version with none database backend (which
causes any database access to fail. Those tests are run with pytest-xdist plugin in parallel which
means that we can efficiently utilised multi-processor machines (including self-hosted runners with
8 CPUS we have to run the tests with maximum parallelism).
It's usually straightforward to run those tests in local virtualenv because they do not require any
setup or running database. They also run much faster than DB tests. You can run them with pytest command
or with breeze that has all the dependencies needed to run all tests automatically installed. Of course
you can also select just specific test or folder or module for the Pytest to collect/run tests from there,
the example below shows how to run all tests, parallelizing them with pytest-xdist
(by specifying tests folder):
pytest tests --skip-db-tests -n autoThe --skip-db-tests flag will only run tests that are not marked as DB tests.
You can also run breeze command to run all the tests (they will run in a separate container,
the selected python version and without access to any database). Adding --use-xdist flag will run all
tests in parallel using pytest-xdist plugin.
You can run parallel commands via breeze testing core-tests or breeze testing providers-tests
- by adding the parallel flags:
breeze testing core-tests --skip-db-tests --backend none --use-xdistYou can pass --parallel-test-type list of test types to execute or --exclude--parallel-test-types
to exclude them from the default set:.
breeze testing providers-tests --run-in-parallel --skip-db-tests --backend none --parallel-test-types "Providers[google] Providers[amazon]"Also you can enter interactive shell with breeze and run tests from there if you want to iterate
with the tests. Source files in breeze are mounted as volumes so you can modify them locally and
rerun in Breeze as you will (-n auto will parallelize tests using pytest-xdist plugin):
breeze shell --backend none --python 3.9
> pytest tests --skip-db-tests -n autoSome of the tests of Airflow require a database to connect to in order to run. Those tests store and read data from Airflow DB using Airflow's core code and it's crucial to run the tests against all real databases that Airflow supports in order to check if the SQLAlchemy queries are correct and if the database schema is correct.
Those tests should be marked with @pytest.mark.db decorator on one of the levels:
- test method can be marked with
@pytest.mark.dbdecorator - test class can be marked with
@pytest.mark.dbdecorator - test module can be marked with
pytestmark = pytest.mark.dbat the top level of the module
For the DB tests, they are run against the multiple databases Airflow support, multiple versions of those and multiple Python versions it supports. In order to save time for testing not all combinations are tested but enough various combinations are tested to detect potential problems.
By default, the DB tests will use sqlite and the "airflow.db" database created and populated in the
${AIRFLOW_HOME} folder. You do not need to do anything to get the database created and initialized,
but if you need to clean and restart the db, you can run tests with -with-db-init flag - then the
database will be re-initialized. You can also set AIRFLOW__DATABASE__SQL_ALCHEMY_CONN environment
variable to point to supported database (Postgres, MySQL, etc.) and the tests will use that database. You
might need to run airflow db reset to initialize the database in that case.
The "non-DB" tests are perfectly fine to run when you have database around but if you want to just run
DB tests (as happens in our CI for the Database runs) you can use --run-db-tests-only flag to filter
out non-DB tests (and obviously you can specify not only on the whole tests directory but on any
folders/files/tests selection, pytest supports).
pytest tests --run-db-tests-onlyYou can also run DB tests with breeze dockerized environment. You can choose backend to use with
--backend flag. The default is sqlite but you can also use others such as postgres or mysql.
You can also select backend version and Python version to use. You can specify the test-type to run -
breeze will list the test types you can run with --help and provide auto-complete for them. Example
below runs the Core tests with postgres backend and 3.9 Python version
You can also run the commands via breeze testing core-tests or breeze testing providers-tests
- by adding the parallel flags manually:
breeze testing core-tests --run-db-tests-only --backend postgres --run-in-parallelYou can pass --parallel-test-type list of test types to execute or --exclude--parallel-test-types
to exclude them from the default set:.
breeze testing providers-tests --run-in-parallel --run-db-tests-only --parallel-test-types "Providers[google] Providers[amazon]"Also - if you want to iterate with the tests you can enter interactive shell and run the tests iteratively -
either by package/module/test or by test type - whatever pytest supports.
breeze shell --backend postgres --python 3.9
> pytest tests --run-db-tests-onlyAs explained before, you cannot run DB tests in parallel using pytest-xdist plugin, but breeze has
support to split all the tests into test-types to run in separate containers and with separate databases
and you can run the tests using --run-in-parallel flag.
breeze testing core-tests --run-db-tests-only --backend postgres --python 3.9 --run-in-parallelYou can apply the marker on method/function/class level with @pytest.mark.db_test decorator or
at the module level with pytestmark = pytest.mark.db_test at the top level of the module.
It's up to the author to decide whether to mark the test, class, or module as "DB-test" - generally the less DB tests - the better and if we can clearly separate the parts that are DB from non-DB, we should, but also it's ok if few tests are marked as DB tests when they are not but they are part of the class or module that is "mostly-DB".
Sometimes, when your class can be clearly split to DB and non-DB parts, it's better to split the class into two separate classes and mark only the DB class as DB test.
Method level:
import pytest
@pytest.mark.db_test
def test_add_tagging(self, sentry, task_instance): ...Class level:
import pytest
@pytest.mark.db_test
class TestDatabricksHookAsyncAadTokenSpOutside: ...Module level (at the top of the module):
import pytest
from airflow.models.baseoperator import BaseOperator
from airflow.models.dag import DAG
from airflow.ti_deps.dep_context import DepContext
from airflow.ti_deps.deps.task_concurrency_dep import TaskConcurrencyDep
pytestmark = pytest.mark.db_testUsually when you add new tests you add tests "similar" to the ones that are already there. In most cases,
therefore you do not have to worry about the test type - it will be automatically selected for you by the
fact that the Test Class that you add the tests or the whole module will be marked with db_test marker.
You should strive to write "pure" non-db unit tests (i.e. DB tests) but sometimes it's just better to plug-in the existing framework of DagRuns, Dags, Connections and Variables to use the Database directly rather than having to mock the DB access for example. It's up to you to decide.
However, if you choose to write DB tests you have to make sure you add the db_test marker - either to
the test method, class (with decorator) or whole module (with pytestmark at the top level of the module).
In most cases when you add tests to existing modules or classes, you follow similar tests so you do not have to do anything, but in some cases you need to decide if your test should be marked as DB test or whether it should be changed to not use the database at all.
If your test accesses the database but is not marked properly the Non-DB test in CI will fail with this message:
"Your test accessed the DB but `_AIRFLOW_SKIP_DB_TESTS` is set. Either make sure your test does not use database or mark your test with `@pytest.mark.db_test`.
If you want to see if your DB test is correctly classified, you can run the test or group
of tests with --skip-db-tests flag.
You can run the all (or subset of) test types if you want to make sure all of the problems are fixed
breeze testing core-tests --skip-db-tests tests/your_test.py
For the whole test suite you can run:
breeze testing core-tests --skip-db-tests
For selected test types (example - the tests will run for Providers/API/CLI code only:
breeze testing providers-tests --skip-db-tests --parallel-test-types "Providers[google] Providers[amazon]"
This is tricky and there is no single solution. Sometimes we can mock-out the methods that require DB access or objects that normally require database. Sometimes we can decide to test just sinle method of class rather than more complex set of steps. Generally speaking it's good to have as many "pure" unit tests that require no DB as possible comparing to DB tests. They are usually faster an more reliable as well.
There are some tricky test cases that require special handling. Here are some of them:
The parameterized tests require stable order of parameters if they are run via xdist - because the parameterized tests are distributed among multiple processes and handled separately. In some cases the parameterized tests have undefined / random order (or parameters are not hashable - for example set of enums). In such cases the xdist execution of the tests will fail and you will get an error mentioning "Known Limitations of xdist". You can see details about the limitation here
The error in this case will look similar to:
Different tests were collected between gw0 and gw7. The difference is:
The fix for that is to sort the parameters in parametrize. For example instead of this:
@pytest.mark.parametrize("status", ALL_STATES)
def test_method(): ...do that:
@pytest.mark.parametrize("status", sorted(ALL_STATES))
def test_method(): ...Similarly if your parameters are defined as result of utcnow() or other dynamic method - you should avoid that, or assign unique IDs for those parametrized tests. Instead of this:
@pytest.mark.parametrize(
"url, expected_dag_run_ids",
[
(
f"api/v1/dags/TEST_DAG_ID/dagRuns?end_date_gte="
f"{urllib.parse.quote((timezone.utcnow() + timedelta(days=1)).isoformat())}",
[],
),
(
f"api/v1/dags/TEST_DAG_ID/dagRuns?end_date_lte="
f"{urllib.parse.quote((timezone.utcnow() + timedelta(days=1)).isoformat())}",
["TEST_DAG_RUN_ID_1", "TEST_DAG_RUN_ID_2"],
),
],
)
def test_end_date_gte_lte(url, expected_dag_run_ids): ...Do this:
@pytest.mark.parametrize(
"url, expected_dag_run_ids",
[
pytest.param(
f"api/v1/dags/TEST_DAG_ID/dagRuns?end_date_gte="
f"{urllib.parse.quote((timezone.utcnow() + timedelta(days=1)).isoformat())}",
[],
id="end_date_gte",
),
pytest.param(
f"api/v1/dags/TEST_DAG_ID/dagRuns?end_date_lte="
f"{urllib.parse.quote((timezone.utcnow() + timedelta(days=1)).isoformat())}",
["TEST_DAG_RUN_ID_1", "TEST_DAG_RUN_ID_2"],
id="end_date_lte",
),
],
)
def test_end_date_gte_lte(url, expected_dag_run_ids): ...Sometimes, even if the whole module is marked as @pytest.mark.db_test, parsing the file and collecting
tests will fail when --skip-db-tests is used because some of the imports or objects created in the
module will read the database.
Usually what helps is to move such initialization code to inside the tests or pytest fixtures (and pass
objects needed by tests as fixtures rather than importing them from the module). Similarly you might
use DB - bound objects (like Connection) in your parametrize specification - this will also fail pytest
collection. Move creation of such objects to inside the tests:
Moving object creation from top-level to inside tests. This code will break collection of tests even if the test is marked as DB test:
TI = TaskInstance(
task=BashOperator(task_id="test", bash_command="true", dag=DAG(dag_id="id"), start_date=datetime.now()),
run_id="fake_run",
state=State.RUNNING,
)
class TestCallbackRequest:
@pytest.mark.parametrize(
"input,request_class",
[
(CallbackRequest(full_filepath="filepath", msg="task_failure"), CallbackRequest),
(
TaskCallbackRequest(
full_filepath="filepath",
simple_task_instance=SimpleTaskInstance.from_ti(ti=TI),
processor_subdir="/test_dir",
is_failure_callback=True,
),
TaskCallbackRequest,
),
(
DagCallbackRequest(
full_filepath="filepath",
dag_id="fake_dag",
run_id="fake_run",
processor_subdir="/test_dir",
is_failure_callback=False,
),
DagCallbackRequest,
),
(
SlaCallbackRequest(
full_filepath="filepath",
dag_id="fake_dag",
processor_subdir="/test_dir",
),
SlaCallbackRequest,
),
],
)
def test_from_json(self, input, request_class): ...Instead - this will not break collection. The TaskInstance is not initialized when the module is parsed, it will only be initialized when the test gets executed because we moved initialization of it from top level / parametrize to inside the test:
pytestmark = pytest.mark.db_test
class TestCallbackRequest:
@pytest.mark.parametrize(
"input,request_class",
[
(CallbackRequest(full_filepath="filepath", msg="task_failure"), CallbackRequest),
(
None, # to be generated when test is run
TaskCallbackRequest,
),
(
DagCallbackRequest(
full_filepath="filepath",
dag_id="fake_dag",
run_id="fake_run",
processor_subdir="/test_dir",
is_failure_callback=False,
),
DagCallbackRequest,
),
(
SlaCallbackRequest(
full_filepath="filepath",
dag_id="fake_dag",
processor_subdir="/test_dir",
),
SlaCallbackRequest,
),
],
)
def test_from_json(self, input, request_class):
if input is None:
ti = TaskInstance(
task=BashOperator(
task_id="test", bash_command="true", dag=DAG(dag_id="id"), start_date=datetime.now()
),
run_id="fake_run",
state=State.RUNNING,
)
input = TaskCallbackRequest(
full_filepath="filepath",
simple_task_instance=SimpleTaskInstance.from_ti(ti=ti),
processor_subdir="/test_dir",
is_failure_callback=True,
)Sometimes it is difficult to rewrite the tests, so you might add conditional handling and mock out some database-bound methods or objects to avoid hitting the database during test collection. The code below will hit the Database while parsing the tests, because this is what Variable.setdefault does when parametrize specification is being parsed - even if test is marked as DB test.
from airflow.models.variable import Variable
pytestmark = pytest.mark.db_test
initial_db_init()
@pytest.mark.parametrize(
"env, expected",
[
pytest.param(
{"plain_key": "plain_value"},
"{'plain_key': 'plain_value'}",
id="env-plain-key-val",
),
pytest.param(
{"plain_key": Variable.setdefault("plain_var", "banana")},
"{'plain_key': 'banana'}",
id="env-plain-key-plain-var",
),
pytest.param(
{"plain_key": Variable.setdefault("secret_var", "monkey")},
"{'plain_key': '***'}",
id="env-plain-key-sensitive-var",
),
pytest.param(
{"plain_key": "{{ var.value.plain_var }}"},
"{'plain_key': '{{ var.value.plain_var }}'}",
id="env-plain-key-plain-tpld-var",
),
],
)
def test_rendered_task_detail_env_secret(patch_app, admin_client, request, env, expected): ...You can make the code conditional and mock out the Variable to avoid hitting the database.
from airflow.models.variable import Variable
pytestmark = pytest.mark.db_test
if os.environ.get("_AIRFLOW_SKIP_DB_TESTS") == "true":
# Handle collection of the test by non-db case
Variable = mock.MagicMock() # type: ignore[misc] # noqa: F811
else:
initial_db_init()
@pytest.mark.parametrize(
"env, expected",
[
pytest.param(
{"plain_key": "plain_value"},
"{'plain_key': 'plain_value'}",
id="env-plain-key-val",
),
pytest.param(
{"plain_key": Variable.setdefault("plain_var", "banana")},
"{'plain_key': 'banana'}",
id="env-plain-key-plain-var",
),
pytest.param(
{"plain_key": Variable.setdefault("secret_var", "monkey")},
"{'plain_key': '***'}",
id="env-plain-key-sensitive-var",
),
pytest.param(
{"plain_key": "{{ var.value.plain_var }}"},
"{'plain_key': '{{ var.value.plain_var }}'}",
id="env-plain-key-plain-tpld-var",
),
],
)
def test_rendered_task_detail_env_secret(patch_app, admin_client, request, env, expected): ...You can also use fixture to create object that needs database just like this.
from airflow.models import Connection
pytestmark = pytest.mark.db_test
@pytest.fixture()
def get_connection1():
return Connection()
@pytest.fixture()
def get_connection2():
return Connection(host="apache.org", extra={})
@pytest.mark.parametrize(
"conn",
[
"get_connection1",
"get_connection2",
],
)
def test_as_json_from_connection(self, conn: Connection):
conn = request.getfixturevalue(conn)
...To run unit tests from the PyCharm IDE, create the local virtualenv, select it as the default project's environment, then configure your test runner:
and run unit tests as follows:
NOTE: You can run the unit tests in the standalone local virtualenv (with no Breeze installed) if they do not have dependencies such as Postgres/MySQL/Hadoop/etc.
Ideally, all unit tests should be run using the standardized Breeze environment. While not as convenient as the one-click "play button" in PyCharm, the IDE can be configured to do this in two clicks.
- Add Breeze as an "External Tool":
- From the settings menu, navigate to Tools > External Tools
- Click the little plus symbol to open the "Create Tool" popup and fill it out:
- Add the tool to the context menu:
- From the settings menu, navigate to Appearance & Behavior > Menus & Toolbars > Project View Popup Menu
- Click on the list of entries where you would like it to be added. Right above or below "Project View Popup Menu Run Group" may be a good choice, you can drag and drop this list to rearrange the placement later as desired.
- Click the little plus at the top of the popup window
- Find your "External Tool" in the new "Choose Actions to Add" popup and click OK. If you followed the image above, it will be at External Tools > External Tools > Breeze
Note: That only adds the option to that one menu. If you would like to add it to the context menu when right-clicking on a tab at the top of the editor, for example, follow the steps above again and place it in the "Editor Tab Popup Menu"
- To run tests in Breeze, right click on the file or directory in the Project View and click Breeze.
To run unit tests from the Visual Studio Code:
- Using the
Extensionsview install Python extension, reload if required
- Using the
Testingview click onConfigure Python Testsand selectpytestframework
- Open
/.vscode/settings.jsonand add"python.testing.pytestArgs": ["tests"]to enable tests discovery
- Now you are able to run and debug tests from both the
Testingview and test files
To run unit, integration, and system tests from the Breeze and your virtualenv, you can use the pytest framework.
Custom pytest plugin runs airflow db init and airflow db reset the first
time you launch them. So, you can count on the database being initialized. Currently,
when you run tests not supported in the local virtualenv, they may either fail
or provide an error message.
There are many available options for selecting a specific test in pytest. Details can be found
in the official documentation, but here are a few basic examples:
pytest tests/core -k "TestCore and not check"This runs the TestCore class but skips tests of this class that include 'check' in their names.
For better performance (due to a test collection), run:
pytest tests/core/test_core.py -k "TestCore and not bash"This flag is useful when used to run a single test like this:
pytest tests/core/test_core.py -k "test_check_operators"This can also be done by specifying a full path to the test:
pytest tests/core/test_core.py::TestCore::test_dag_params_and_task_paramsTo run the whole test class, enter:
pytest tests/core/test_core.py::TestCoreYou can use all available pytest flags. For example, to increase a log level
for debugging purposes, enter:
pytest --log-cli-level=DEBUG tests/core/test_core.py::TestCoreYou can run tests interactively using regular pytest commands inside the Breeze shell. This has the advantage, that Breeze container has all the dependencies installed that are needed to run the tests and it will ask you to rebuild the image if it is needed and some new dependencies should be installed.
By using interactive shell and iterating over the tests, you can iterate and re-run tests one-by-one or group by group right after you modified them.
Entering the shell is as easy as:
breezeThis should drop you into the container.
You can also use other switches (like --backend for example) to configure the environment for your
tests (and for example to switch to different database backend - see --help for more details).
Once you enter the container, you might run regular pytest commands. For example:
pytest --log-cli-level=DEBUG tests/core/test_core.py::TestCoreIf you wish to only run tests and not to drop into the shell, apply the
tests command. You can add extra targets and pytest flags after the tests command. Note that
often you want to run the tests with a clean/reset db, so usually you want to add --db-reset flag
to breeze command. The Breeze image usually will have all the dependencies needed and it
will ask you to rebuild the image if it is needed and some new dependencies should be installed.
breeze testing providers-tests providers/tests/http/hooks/test_http.py tests/core/test_core.py --db-reset --log-cli-level=DEBUGYou can run the whole core test suite without adding the test target:
breeze core-testing tests --db-resetYou can run the whole providers test suite without adding the test target:
breeze providers-testing tests --db-resetYou can also specify individual tests or a group of tests:
breeze testing core-tests --db-reset tests/core/test_core.py::TestCoreYou can also limit the tests to execute to specific group of tests
breeze testing core-tests --test-type OtherIn case of Providers tests, you can run tests for all providers
breeze testing ptoviders-tests --test-type ProvidersYou can limit the set of providers you would like to run tests of
breeze testing providers-tests --test-type "Providers[airbyte,http]"You can also run all providers but exclude the providers you would like to skip
breeze testing providers-tests --test-type "Providers[-amazon,google]"Sometimes you need to inspect docker compose after tests command complete,
for example when test environment could not be properly set due to
failed health-checks. This can be achieved with --skip-docker-compose-down
flag:
breeze testing core-tests --skip--docker-compose-downIf you run breeze testing core-tests --run-in-parallel or
breeze testing providers-tests --run-in-parallel tests run in parallel
on your development machine - maxing out the number of parallel runs at the number of cores you
have available in your Docker engine.
In case you do not have enough memory available to your Docker (8 GB), the Integration. Provider
and Core test type are executed sequentially with cleaning the docker setup in-between. This
allows to print
This allows for massive speedup in full test execution. On 8 CPU machine with 16 cores and 64 GB memory and fast SSD disk, the whole suite of tests completes in about 5 minutes (!). Same suite of tests takes more than 30 minutes on the same machine when tests are run sequentially.
Note
On MacOS you might have less CPUs and less memory available to run the tests than you have in the host,
simply because your Docker engine runs in a Linux Virtual Machine under-the-hood. If you want to make
use of the parallelism and memory usage for the CI tests you might want to increase the resources available
to your docker engine. See the Resources chapter
in the Docker for Mac documentation on how to do it.
You can also limit the parallelism by specifying the maximum number of parallel jobs via MAX_PARALLEL_TEST_JOBS variable. If you set it to "1", all the test types will be run sequentially.
MAX_PARALLEL_TEST_JOBS="1" ./scripts/ci/testing/ci_run_airflow_testing.shNote
In case you would like to cleanup after execution of such tests you might have to cleanup some of the docker containers running in case you use ctrl-c to stop execution. You can easily do it by running this command (it will kill all docker containers running so do not use it if you want to keep some docker containers running):
docker kill $(docker ps -q)Tests that are using a specific backend are marked with a custom pytest marker pytest.mark.backend.
The marker has a single parameter - the name of a backend. It corresponds to the --backend switch of
the Breeze environment (one of mysql, sqlite, or postgres). Backend-specific tests only run when
the Breeze environment is running with the right backend. If you specify more than one backend
in the marker, the test runs for all specified backends.
Example of the postgres only test:
@pytest.mark.backend("postgres")
def test_copy_expert(self): ...Example of the postgres,mysql test (they are skipped with the sqlite backend):
@pytest.mark.backend("postgres", "mysql")
def test_celery_executor(self): ...You can use the custom --backend switch in pytest to only run tests specific for that backend.
Here is an example of running only postgres-specific backend tests:
pytest --backend postgresSome of the tests rung for a long time. Such tests are marked with @pytest.mark.long_running annotation.
Those tests are skipped by default. You can enable them with --include-long-running flag. You
can also decide to only run tests with -m long-running flags to run only those tests.
Some of our tests are quarantined. This means that this test will be run in isolation and that it will be re-run several times. Also when quarantined tests fail, the whole test suite will not fail. The quarantined tests are usually flaky tests that need some attention and fix.
Those tests are marked with @pytest.mark.quarantined annotation.
Those tests are skipped by default. You can enable them with --include-quarantined flag. You
can also decide to only run tests with -m quarantined flag to run only those tests.
Our CI runs provider tests for providers with previous compatible airflow releases. This allows to check if the providers still work when installed for older airflow versions.
The back-compatibility tests based on the configuration specified in the
PROVIDERS_COMPATIBILITY_TESTS_MATRIX constant in the ./dev/breeze/src/airflow_breeze/global_constants.py
file - where we specify:
- Python version
- Airflow version
- which providers should be removed for the tests (exclusions)
- whether to run tests for this Airflow/Python version
Those tests can be used to test compatibility of the providers with past (and future!) releases of airflow. For example it could be used to run latest provider versions with released or main Airflow 3 if they are developed independently.
The tests use the current source version of tests folder and current providers - so care should be
taken that the tests implemented for providers in the sources allow to run it against previous versions
of Airflow and against Airflow installed from PyPI package rather than from the sources.
Running tests can be easily done locally by running appropriate breeze command. In CI the command
is slightly different as it is run using providers build using wheel packages, but it is faster
to run it locally and easier to iterate if you need to fix provider using provider sources mounted
directly to the container.
- Make sure to build latest Breeze ci image
breeze ci-image build --python 3.9- Enter breeze environment by selecting the appropriate airflow version and choosing
providers-and-testsoption for--mount-sourcesflag.
breeze shell --use-airflow-version 2.9.1 --mount-sources providers-and-tests- You can then run tests as usual:
pytest providers/tests/<provider>/test.py- Iterate with the tests and providers. Both providers and tests are mounted from local sources so
changes you do locally in both - tests and provider sources are immediately reflected inside the
breeze container and you can re-run the tests inside
breezecontainer without restarting the container (which makes it faster to iterate).
Note
Since providers are installed from sources rather than from packages, plugins from providers are not
recognised by ProvidersManager for airflow < 2.10 and tests that expect plugins to work might not work.
In such case you should follow the CI way of running the tests (see below).
When you implement tests for providers, you should make sure that they are compatible with older Airflow versions.
Note that some of the tests, if written without taking care about the compatibility, might not work with older versions of Airflow - this is because of refactorings, renames, and tests relying on internals of Airflow that are not part of the public API. We deal with it in one of the following ways:
- If the whole provider is supposed to only work for later airflow version, we remove the whole provider by excluding it from compatibility test configuration (see below)
- Some compatibility shims are defined in
tests_common.test_utils/compat.py- and they can be used to make the tests compatible - for example importingParseImportErrorafter the exception has been renamed fromImportErrorand it would fail in Airflow 2.9, but we have a fallback import incompat.pythat falls back to old import automatically, so all tests testing / expectingParseImportErrorshould import it from thetests.tests_utils.compatmodule. There are few other compatibility shims defined there and you can add more if needed in a similar way. - If only some tests are not compatible and use features that are available only in newer airflow version,
we can mark those tests with appropriate
AIRFLOW_V_2_X_PLUSboolean constant defined inversion_compat.pyFor example:
from tests_common.test_utils.version_compat import AIRFLOW_V_2_10_PLUS
@pytest.mark.skipif(not AIRFLOW_V_2_10_PLUS, reason="The tests should be skipped for Airflow < 2.10")
def some_test_that_only_works_for_airflow_2_10_plus():
pass- Sometimes, the tests should only be run when airflow is installed from the sources in main.
In this case you can add conditional
skipifmarkerforRUNNING_TESTS_AGAINST_AIRFLOW_PACKAGESto the test. For example:
from tests_common import RUNNING_TESTS_AGAINST_AIRFLOW_PACKAGES
@pytest.mark.skipif(
RUNNING_TESTS_AGAINST_AIRFLOW_PACKAGES, reason="Plugin initialization is done early in case of packages"
)
def test_plugin():
passSometimes Pytest collection fails to work, when certain imports used by the tests either do not exist or fail with RuntimeError about compatibility ("minimum Airflow version is required") or because they raise AirflowOptionalProviderFeatureException. In such case you should wrap the imports in
ignore_provider_compatibility_errorcontext manager adding the__file__module name as parameter. This will stop failing pytest collection and automatically skip the whole module from tests.For example:
with ignore_provider_compatibility_error("2.8.0", __file__):
from airflow.providers.common.io.xcom.backend import XComObjectStorageBackend- In some cases in order to enable collection of pytest on older airflow version you might need to convert top-level import into a local import, so that Pytest parser does not fail on collection.
In CI those tests are run in a slightly more complex way because we want to run them against the build provider packages, rather than mounted from sources.
In case of canary runs we add --clean-airflow-installation flag that removes all packages before
installing older airflow version, and then installs development dependencies
from latest airflow - in order to avoid case where a provider depends on a new dependency added in latest
version of Airflow. This clean removal and re-installation takes quite some time though and in order to
speed up the tests in regular PRs we only do that in the canary runs.
The exact way CI tests are run can be reproduced locally building providers from selected tag/commit and using them to install and run tests against the selected airflow version.
Herr id how to reproduce it.
- Make sure to build latest Breeze ci image
breeze ci-image build --python 3.9- Build providers from latest sources:
rm dist/*
breeze release-management prepare-provider-packages --include-not-ready-providers \
--version-suffix-for-pypi dev0 --package-format wheel- Prepare provider constraints
breeze release-management generate-constraints --airflow-constraints-mode constraints-source-providers --answer yes- Remove providers that are not compatible with Airflow version installed by default. You can look up
the incompatible providers in the
PROVIDERS_COMPATIBILITY_TESTS_MATRIXconstant in the./dev/breeze/src/airflow_breeze/global_constants.pyfile. - Enter breeze environment, installing selected airflow version and the provider packages prepared from main
breeze shell --use-packages-from-dist --package-format wheel --use-airflow-version 2.9.1 \
--install-airflow-with-constraints --providers-skip-constraints --mount-sources testsIn case you want to reproduce canary run, you need to add --clean-airflow-installation flag:
breeze shell --use-packages-from-dist --package-format wheel --use-airflow-version 2.9.1 \
--install-airflow-with-constraints --providers-skip-constraints --mount-sources tests --clean-airflow-installation- You can then run tests as usual:
pytest providers/tests/<provider>/test.py- Iterate with the tests
The tests are run using:
- airflow installed from PyPI
- tests coming from the current airflow sources (they are mounted inside the breeze image)
- provider packages built from the current airflow sources and placed in dist
This means that you can modify and run tests and re-run them because sources are mounted from the host, but if you want to modify provider code you need to exit breeze, rebuild the provider package and restart breeze using the command above.
Rebuilding single provider package can be done using this command:
breeze release-management prepare-provider-packages \
--version-suffix-for-pypi dev0 --package-format wheel <provider>We have special tests that run with the lowest direct resolution of dependencies for Airflow and providers. This is run in order to check whether we are not using a feature that is not available in an older version of some dependencies.
You can test minimum dependencies that are installed by Airflow by running (for example to run "Core" tests):
breeze testing core-tests --force-lowest-dependencies --test-type "Core"You can also iterate on the tests and versions of the dependencies by entering breeze shell and running the tests from there:
breeze shell --force-lowest-dependencies --test-type "Core"The way it works - when you run the breeze with --force-lowest-dependencies flag, breeze will use
attempt (with the help of uv) to downgrade the dependencies to the lowest version that is compatible
with the dependencies specified in airflow dependencies. You will see it in the output of the breeze
command as a sequence of downgrades like this:
- aiohttp==3.9.5
+ aiohttp==3.9.2
- anyio==4.4.0
+ anyio==3.7.1Similarly we can test if the provider tests are working for lowest dependencies of specific provider.
Those tests can be easily run locally with breeze (replace PROVIDER_ID with id of the provider):
breeze testing providers-tests --force-lowest-dependencies --test-type "Providers[PROVIDER_ID]"If you find that the tests are failing for some dependencies, make sure to add minimum version for the dependency in the provider.yaml file of the appropriate provider and re-run it.
You can also iterate on the tests and versions of the dependencies by entering breeze shell and running the tests from there:
breeze shell --force-lowest-dependencies --test-type "Providers[PROVIDER_ID]"Similarly as in case of "Core" tests, the dependencies will be downgraded to the lowest version that is
compatible with the dependencies specified in the provider dependencies and you will see the list of
downgrades in the output of the breeze command. Note that this will be combined downgrades of both
Airflow and selected provider dependencies, so the list will be longer than in case of "Core" tests
and longer than just dependencies of the provider. For example for a google provider, part of the
downgraded dependencies will contain both Airflow and Google Provider dependencies:
- flask-login==0.6.3
+ flask-login==0.6.2
- flask-session==0.5.0
+ flask-session==0.4.0
- flask-wtf==1.2.1
+ flask-wtf==1.1.0
- fsspec==2023.12.2
+ fsspec==2023.10.0
- gcloud-aio-bigquery==7.1.0
+ gcloud-aio-bigquery==6.1.2
- gcloud-aio-storage==9.2.0When your tests pass in regular test, but fail in "lowest-direct" dependency resolution tests, you need
to figure out the lower-bindings missing in hatch_build.py (for Airflow core dependencies) or
in the corresponding provider's provider.yaml file. This is usually a very easy thing that takes a little
bit of time to figure out especially if you just added new feature from a library that you use, just check in
the release notes what is the minimum version of the library that you can use and set it as the
>=VERSION in the hatch_build.py or provider.yaml file. For hatch_build.py changes you do not
need to do anything else, for provider.yaml file you need to regenerate generated dependencies
by running pre-commit run in the provider directory after adding the file to git or just letting the
pre-commit to do it's job if you already has pre-commit installed via pre-commit install - then just
committing the change will regenerate the dependencies automatically.
After that, re-run the breeze shell --force-lowest-dependencies command and see if the tests pass.
breeze shell --force-lowest-dependencies --test-type "Providers[PROVIDER_ID]"Sometimes it might get a bit tricky to know what is the minimum version of the library you should be using but in this case you can easily find it by looking at the error and list of downgraded packages and guessing which one is the one that is causing the problem. You can then look at the release notes of the library and find the minimum version but also you can revert to technique known as bisecting which allows you to quickly figure out the right version without knowing the root cause of the problem.
Assume you suspect library "foo" that was downgraded from 1.0.0 to 0.1.0 is causing the problem. Bisecting technique looks like follows:
- enter breeze with
--force-lowest-dependenciesflag (thefoolibrary is downgraded to 0.1.0). Your test should fail. - make sure that just upgrading the
foolibrary to 1.0.0 -> re-run failing test (withpytest <test>) and see that it passes. - downgrade the
foolibrary to 0.1.0 -> re-run failing test (withpytest <test>) and see that it fails. - look at the list of versions available for the library between 0.1.0 and 1.0.0 (for example via
https://pypi.org/project/foo/#history link - where
foois your library. - find a middle version between the 1.0.0 and 0.1.0 and upgrade the library to this version - see if the test passes or fails - if it passes, continue with finding the middle version between the current version and lower version, if it fails, continue with finding the middle version between the current version and higher version.
- continue that way until you find the version that is the lowest version that passes the test.
- set this version in the
hatch_build.pyorprovider.yamlfile, regenerate the generated dependencies file and re-start breeze with--force-lowest-dependenciesflag and see that the library has been downgraded to the version you set and the test passes.
By default masking secrets in test disabled because it might have side effects into the other tests which intends to check logging/stdout/stderr values
If you need to test masking secrets in test cases
you have to apply pytest.mark.enable_redact to the specific test case, class or module.
@pytest.mark.enable_redact
def test_masking(capsys):
mask_secret("eggs")
RedactedIO().write("spam eggs and potatoes")
assert "spam *** and potatoes" in capsys.readouterr().outYou can apply the marker pytest.mark.platform(name) to the specific test case, class or module
for prevent to run on unsupported platform.
linux: Run test only on linux platformbreeze: Run test only inside of Breeze container, it might be useful in case of run some potential dangerous things in tests or if it expects to use common Breeze things.
By default, all warnings captured during the test runs are saved into the tests/warnings.txt.
If required, you could change the path by providing --warning-output-path as pytest CLI arguments
or by setting the environment variable CAPTURE_WARNINGS_OUTPUT.
root@3f98e75b1ebe:/opt/airflow# pytest tests/core/ --warning-output-path=/foo/bar/spam.egg
...
========================= Warning summary. Total: 28, Unique: 12 ==========================
airflow: total 11, unique 1
runtest: total 11, unique 1
other: total 7, unique 1
runtest: total 7, unique 1
tests: total 10, unique 10
runtest: total 10, unique 10
Warnings saved into /foo/bar/spam.egg file.
================================= short test summary info =================================You might also disable capture warnings by providing --disable-capture-warnings as pytest CLI arguments
or by setting global warnings filter
to ignore, e.g. set PYTHONWARNINGS environment variable to ignore.
pytest tests/core/ --disable-capture-warningsBy default, all environment variables related to Airflow (starting by AIRFLOW__) are all cleared before running tests
to avoid potential side effect. However, in some scenarios you might want to disable this mechanism and keep the
environment variables you defined to configure your Airflow environment. For example, you might want to run tests
against a specific database configured through the environment variable AIRFLOW__DATABASE__SQL_ALCHEMY_CONN.
Or running tests using a specific executor to run tasks configured through AIRFLOW__CORE__EXECUTOR.
To keep using environment variables you defined in your environment, you need to provide --keep-env-variables as
pytest CLI argument.
pytest tests/core/ --keep-env-variablesThis parameter is also available in Breeze.
breeze testing core-tests --keep-env-variablesBy default, the database is cleared from all items before running tests. This is to avoid potential conflicts with existing resources in the database when running tests using the database. However, in some scenarios you might want to disable this mechanism and keep the database as is. For example, you might want to run tests in parallel against the same database. In that case, you need to disable the database cleanup, otherwise the tests are going to conflict with each other (one test will delete the resources that another one is creating).
To disable the database cleanup, you need to provide --no-db-cleanup as pytest CLI argument.
pytest tests/core/ --no-db-cleanupThis parameter is also available in Breeze.
breeze testing core-tests --no-db-cleanup tests/coreAirflow's CI process automatically uploads the code coverage report to codecov.io. For the most recent coverage report of the main branch, visit: https://codecov.io/gh/apache/airflow.
If you wish to obtain coverage reports for specific areas of the codebase on your local machine, follow these steps:
- Initiate a breeze shell.
- Execute one of the commands below based on the desired coverage area:
- Core:
python scripts/cov/core_coverage.py - REST API:
python scripts/cov/restapi_coverage.py - CLI:
python scripts/cov/cli_coverage.py - Webserver:
python scripts/cov/www_coverage.py
- After execution, the coverage report will be available at: http://localhost:28000/dev/coverage/index.html.
Note
In order to see the coverage report, you must start webserver first in breeze environment via the
airflow webserver. Once you enter breeze, you can start tmux (terminal multiplexer) and
split the terminal (by pressing ctrl-B " for example) to continue testing and run the webserver
in one terminal and run tests in the second one (you can switch between the terminals with ctrl-B <arrow>).
Each coverage command provides a list of modules that aren't fully covered. If you wish to enhance coverage for a particular module:
- Work on the module to improve its coverage.
- Once coverage reaches 100%, you can safely remove the module from the list of modules that are not fully covered. This list is inside each command's source code.
You can run tests with SQL statements tracking. To do this, use the --trace-sql option and pass the
columns to be displayed as an argument. Each query will be displayed on a separate line.
Supported values:
num- displays the query number;time- displays the query execution time;trace- displays the simplified (one-line) stack trace;sql- displays the SQL statements;parameters- display SQL statement parameters.
If you only provide num, then only the final number of queries will be displayed.
By default, pytest does not display output for successful tests, if you still want to see them, you must
pass the --capture=no option.
If you run the following command:
pytest --trace-sql=num,sql,parameters --capture=no \
tests/jobs/test_scheduler_job.py -k test_process_dags_queries_count_05On the screen you will see database queries for the given test.
SQL query tracking does not work properly if your test runs subprocesses. Only queries from the main process are tracked.
For other kinds of tests look at Testing document