Many users and other packages depend on different parts of Qiskit. We must make sure that whenever we make changes to the code, we give users ample time to adjust without breaking code that they have already written.
Most importantly: do not change any interface that is public-facing unless we absolutely have to. Adding things is ok, taking things away is annoying for users but can be handled reasonably with plenty notice, but changing behavior generally means users cannot write code that will work with two subsequent versions of Qiskit, which is not acceptable.
Beware that users will often be using functions, classes and methods that we,
the Qiskit developers, may consider internal or not widely used. Do not make
assumptions that "this is buried, so nobody will be using it"; if it is public,
it is subject to the policy. The only exceptions here are functions and modules
that are explicitly internal, i.e. those whose names begin with a leading
The guiding principles are:
we must not remove or change code without active warnings for least three months or two complete version cycles;
there must always be a way to achieve valid goals that does not issue any warnings;
never assume that a function that isn't explicitly internal isn't in use;
all deprecations, changes and removals are considered API changes, and can only occur in minor releases not patch releases, per the stable branch policy.
Removing a feature¶
When removing a feature (for example a class, function or function parameter), we will follow this procedure:
The alternative path must be in place for one minor version before any warnings are issued. For example, if we want to replace the function
bar(), we must make at least one release with both functions before issuing any warnings within
foo(). You may issue
PendingDeprecationWarnings from the old paths immediately.
Reason: we need to give people time to swap over without breaking their code as soon as they upgrade.
After the alternative path has been in place for at least one minor version, issue the deprecation warnings. Add a release note with a
deprecationssection listing all deprecated paths, their alternatives, and the reason for deprecation. Update the tests to test the warnings.
Reason: removals must be highly visible for at least one version, to minimize the surprise to users when they actually go.
Set a removal date for the old feature, and remove it (and the warnings) when reached. This must be at least three months after the version with the warnings was first released, and cannot be the minor version immediately after the warnings. Add an
upgraderelease note that lists all the removals. For example, if the alternative path was provided in
0.19.0and the warnings were added in
0.20.0, the earliest version for removal is
0.22.0, even if
0.21.0was released more than three months after
These are minimum requirements. For removal of significant or core features, give users at least an extra minor version if not longer.
Reason: there needs to be time for users to see these messages, and to give them time to adjust. Not all users will update their version of Qiskit immediately, and some may skip minor versions.
When a feature is marked as deprecated it is slated for removal, but users should still be able to rely on it to work correctly. We consider a feature marked "deprecated" as frozen; we commit to maintaining it with critical bug fixes until it is removed, but we won't merge new functionality to it.
Changing behavior without a removal is particularly difficult to manage, because we need to have both options available for two versions, and be able to issue warnings. For example, changing the type of the return value from a function will almost invariably involve making an API break, which is frustrating for users and makes it difficult for them to use Qiskit.
The best solution here is often to make a new function, and then use the procedures for removal above.
If you absolutely must change the behavior of existing code (other than fixing
bugs), you will need to use your best judgment to apply the guiding principles
at the top of this document. The most appropriate warning for behavioral
changes is usually
FutureWarning. Some possibilities for how to effect a
If you are changing the default behavior of a function, consider adding a keyword argument to select between old and new behaviors. When it comes time, you can issue a
FutureWarningif the keyword argument is not given (e.g. if it is
None), saying that the new value will soon become the default. You will need to go through the normal deprecation period for removing this keyword argument after you have made the behavior change. This will take at least six months to go through both cycles.
If you need to change the return type of a function, consider adding a new function that returns the new type, and then follow the procedures for deprecating the old function.
If you need to accept a new input that you cannot distinguish from an existing possibility because of its type, consider letting it be passed by a different keyword argument, or add a second function that only accepts the new form.
Issuing deprecation warnings¶
The proper way to raise a deprecation warning is to use the
from the warnings module in the Python standard library, using the category
DeprecationWarning. For example:
import warnings def deprecated_function(): warnings.warn( "The function qiskit.deprecated_function() is deprecated since " "Qiskit Terra 0.20.0, and will be removed 3 months or more later. " "Instead, you should use qiskit.other_function().", category=DeprecationWarning, stacklevel=2, ) # ... the rest of the function ...
Make sure you include the version of the package that introduced the deprecation warning (so maintainers can easily see when it is valid to remove it), and what the alternative path is.
Take note of the
stacklevel argument. This controls which function is
accused of being deprecated. Setting
stacklevel=1 (the default) means the
warning will blame the
warn function itself, while
correctly blame the containing function. It is unusual to set this to anything
2, but can be useful if you use a helper function to issue the
same warning in multiple places.
Testing deprecated functionality¶
Whenever you add deprecation warnings, you will need to update tests involving the functionality. The test suite should fail otherwise, because of the new warnings. We must continue to test deprecated functionality throughout the deprecation period, to ensure that it still works.
To update the tests, you need to wrap each call of deprecated behavior in its
own assertion block. For subclasses of
unittest.TestCase (which all Qiskit
test cases are), this is done by:
class MyTestSuite(QiskitTestCase): def test_deprecated_function(self): with self.assertWarns(DeprecationWarning): output = deprecated_function() # ... do some things with output ... self.assertEqual(output, expected)
Documenting deprecations and breaking changes¶
It is important to warn the user when your breaking changes are coming. This can be done in the docstring for the function, method, or class that is being deprecated, by adding a deprecated note :
def deprecated_function(): """ Short description of the deprecated function. .. deprecated:: 0.20.0 The function qiskit.deprecated_function() is deprecated since Qiskit Terra 0.20.0, and will be removed 3 months or more later. Instead, you should use qiskit.other_function(). <rest of the docstring> """ # ... the rest of the function ...
In particularly situation where a deprecation or change might be a major disruptor for users, a migration guide might be needed. Once the migration guide is written and published, deprecation messages and documentation can link to it.