It makes more sense, also, we have a long term idea of generalizing
fixture support to items defined by other plugins, not just python, in
which case `--fixtures` would definitely not be python-plugin
specific.
and also fixes a regression in pytest 8.0.0 where `setup_method` crashes
if the class has static or class method tests.
It is allowed to have a test class with static/class methods which
request non-static/class method fixtures (including `setup_method`
xunit-fixture). I take it as a given that we need to support this
somewhat odd scenario (stdlib unittest also supports it).
This raises a question -- when a staticmethod test requests a bound
fixture, what is that fixture's `self`?
stdlib unittest says - a fresh instance for the test.
Previously, pytest said - some instance that is shared by all
static/class methods. This is definitely broken since it breaks test
isolation.
Change pytest to behave like stdlib unittest here.
In practice, this means stopping to rely on `self.obj.__self__` to get
to the instance from the test function's binding. This doesn't work
because staticmethods are not bound to anything.
Instead, keep the instance explicitly and use that.
BTW, I think this will allow us to change `Class`'s fixture collection
(`parsefactories`) to happen on the class itself instead of a class
instance, allowing us to avoid one class instantiation. But needs more
work.
Fixes#12065.
- Separate the requesting from the requested.
- Avoid the term "factory", I think most people don't distinguish
between "fixture" and "fixture function" (i.e. "factory") and would
find the term "factory" unfamiliar.
There are two non-optimal things in the current way scope checking is
done:
- It runs on `SubRequest`, but doesn't use the `SubRequest's scope,
which is confusing. Instead it takes `invoking_scope` and
`requested_scope`.
- Because `_check_scope` is only defined on `SubRequest` and not
`TopRequest`, `_compute_fixture_value` first creates the `SubRequest`
only then checks the scope (hence the need for the previous point).
Instead, also define `_check_scope` on `TopRequest` (always valid), and
remove `invoking_scope`, using `self._scope` instead.
Previously, the `obj` of a `TestCaseFunction` (the unittest plugin item
type) was the unbound method. This is unlike regular `Class` where the
`obj` is a bound method to a fresh instance.
This difference necessitated several special cases in in places outside
of the unittest plugin, such as `FixtureDef` and `FixtureRequest`, and
made things a bit harder to understand.
Instead, match how the python plugin does it, including collecting
fixtures from a fresh instance.
The downside is that now this instance for fixture-collection is kept
around in memory, but it's the same as `Class` so nothing new. Users
should only initialize stuff in `setUp`/`setUpClass` and similar
methods, and not in `__init__` which is generally off-limits in
`TestCase` subclasses.
I am not sure why there was a difference in the first place, though I
will say the previous unittest approach is probably the preferable one,
but first let's get consistency.
Previously, if more than one fixture finalizer raised, only the first
was reported, and the other errors were lost.
Use an exception group to report them all. This is similar to the change
we made in node teardowns (in `SetupState`).
pytest allows a fixture to request its own name (directly or
indirectly), in which case the fixture with the same name but one level
up is used.
To know which fixture should be used next, pytest keeps a mutable
item-global dict `_arg2index` which maintains this state. This is not
great:
- Mutable state like this is hard to understand and reason about.
- It is conceptually buggy; the indexing is global (e.g. if requesting
`fix1` and `fix2`, the indexing is shared between them), but actually
different branches of the subrequest tree should not affect each
other.
This is not an issue in practice because pytest keeps a cache of the
fixturedefs it resolved anyway (`_fixture_defs`), but if the cache is
removed it becomes evident.
Instead of the `_arg2index` state, count how many `argname`s deep we are
in the subrequest tree ("the fixture stack") and use that for the index.
This way, no global mutable state and the logic is very localized and
easier to understand.
This is slower, however fixture stacks should not be so deep that this
matters much, I hope.
According to my understanding, this code, which handles obtaining the
relevant fixturedefs when a dynamic `getfixturevalue` is used, has an
optimization where it only grabs fixturedefs that are visible to the
*parent* of the item, instead of the item itself, under the assumption
that a fixturedef can't be visible to a single item, only to a
collector.
Remove this optimization for the following reasons:
- It doesn't save much (one loop iteration in `matchfactories`)
- It slightly complicates the complex fixtures code
- If some plugin wants to make a fixture visible only to a single item,
why not let it?
- In the static case (`getfixtureclosure`), this optimization is not
done (despite the confusing name `parentnode`, it is *not* the parent
node). This is inconsistent.
In `CallSpec2.setmulti` the `params` and `_arg2scope` dicts are always
set together.
Further, the `get_parametrized_fixture_keys` accesses `_arg2scope` for
all argnames without a check, which I think rules out that the code
protects against plugin shenanigans.
After removing the suppress, adjust the comment and code to make more
sense.
Dicts these days preserve order, so the sort is no longer needed to
achieve determinism.
As shown by the `test_dynamic_parametrized_ordering` test, this can
change the ordering of items, but only in equivalent ways (same number
of setups/teardowns per scope), it will just respect the user's given
ordering better (hence `vxlan` items now ordered before `vlan` items
compared to the previous ordering).
* Improve error message when using @pytest.fixture twice
While obvious in hindsight, this error message confused me. I thought my fixture
function was used in a test function twice, since the wording is ambiguous.
Also, the error does not tell me *which* function is the culprit.
Finally, this adds a test, which wasn't done in
cfd16d0dac where this was originally implemented.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
ruff is faster and handle everything we had prior.
isort configuration done based on the indication from
https://github.com/astral-sh/ruff/issues/4670, previousely based on
reorder-python-import (#11896)
flake8-docstrings was a wrapper around pydocstyle (now archived) that
explicitly asks to use ruff in https://github.com/PyCQA/pydocstyle/pull/658.
flake8-typing-import is useful mainly for project that support python 3.7
and the one useful check will be implemented in https://github.com/astral-sh/ruff/issues/2302
We need to keep blacken-doc because ruff does not handle detection
of python code inside .md and .rst. The direct link to the repo is
now used to avoid a redirection.
Manual fixes:
- Lines that became too long
- % formatting that was not done automatically
- type: ignore that were moved around
- noqa of hard to fix issues (UP031 generally)
- fmt: off and fmt: on that is not really identical
between black and ruff
- autofix re-order in pre-commit from faster to slower
Co-authored-by: Ran Benita <ran@unusedvar.com>
Change our mypy configuration to disallow untyped defs by default, which ensures *new* files added to the code base are fully typed.
To avoid having to type-annotate everything now, add `# mypy: allow-untyped-defs` to files which are not fully type annotated yet.
As we fully type annotate those modules, we can then just remove that directive from the top.
This is a useful addition to the existing `listchain`. While `listchain`
returns top-to-bottom, `iterparents` is bottom-to-top and doesn't require
an internal full iteration + `reverse`.
Refs #11662.
--- Problem
Each fixture definition has a "visibility", the `FixtureDef.baseid`
attribute. This is nodeid-like string. When a certain `node` requests a
certain fixture name, we match node's nodeid against the fixture
definitions with this name.
The matching currently happens on the *textual* representation of the
nodeid - we split `node.nodeid` to its "parent nodeids" and then check
if the fixture's `baseid` is in there.
While this has worked so far, we really should try to avoid textual
manipulation of nodeids as much as possible. It has also caused problem
in an odd case of a `Package` in the root directory: the `Package` gets
nodeid `.`, while a `Module` in it gets nodeid `test_module.py`. And
textually, `.` is not a parent of `test_module.py`.
--- Solution
Avoid this entirely by just checking the node hierarchy itself. This is
made possible by the fact that we now have proper `Directory` nodes
(`Dir` or `Package`) for the entire hierarchy.
Also do the same for `_getautousenames` which is a similar deal.
The `iterparentnodeids` function is no longer used and is removed.
Add a function on the `FixtureManager` to register a fixture with
pytest. Currently this can only be done through `parsefactories`.
My aim is to eventually make something like this available to plugins,
as it's a pretty common need.
I could find 2 plugins that would be broken by this (pytest-play and
pytest-wdl), but they will be better served by just copying
`_fillfixtures` instead of use the private function.