89979ca1eb
This PR was merged into the 3.3-dev branch.
Discussion
----------
Not allowing autoconfigure, instanceofConditionals or defaults for ChildDefinition
| Q | A
| ------------- | ---
| Branch? | master
| Bug fix? | yes (removing risky behavior)
| New feature? | no
| BC breaks? | no
| Deprecations? | no
| Tests pass? | yes
| Fixed tickets | see #22530
| License | MIT
| Doc PR | n/a
This PR *prohibits* using `autoconfigure`, `_instanceof` and `_defaults` for ChildDefinition.
Additionally, I added many "integration" test cases: we need to test and prove all edge cases. These are in the `integration/` directory: the `main.yml` file is parsed and compared to `expected.yml`. Both are in YAML to ease comparing the before/after. We need to check these out and make sure they're right and we're not missing anything else.
This PR removes MANY of the "wtf" cases, but there are still 4 that I know of... and of course they all deal with parent-child stuff :).
A) [MAJOR] [autoconfigure_parent_child_tags](https://github.com/symfony/symfony/pull/22563/files#diff-fd6cf15470c5abd40156e4e7dc4e7f6d) `instanceof` tags from autoconfigure are NEVER applied to the child (you can't set `autoconfigure` directly on a Child, but you still can set it on a parent and inherit it... sneaky). We could throw an Exception I suppose to prevent this `autoconfigure` from cascading from parent to child... but it's tricky due to `instanceof`.
B( [MAJOR] [instanceof_parent_child](https://github.com/symfony/symfony/pull/22563/files#diff-14666e9a25322d44b3c2c583b6814dc2) `instanceof` tags that are applied to the parent, are not applied to the child. Again, you can't set `instanceof` directly on a Child, but you *can* set it on a parent, and have that cascade to the child. Like before, we could maybe throw an exception to prevent this.
C) [MINOR] ([autoconfigure_child_not_applied](https://github.com/symfony/symfony/pull/22563/files#diff-3372a1dcaf3af30d14a7d0a6c8bfa988)) automatic `instanceof` will not be applied to the child when the parent class has a different (non-instanceof-ed) class. If we could throw an exception for (A), then it would cover this too.
D) `_tags` from defaults are never used (unless you have inherit_tags) - fixed in #22530
A, B & C are effectively caused by there being a "sneaky" way to re-enable `autoconfigure` and `instanceof` for ChildDefinition... which opens up wtf cases.
## Wait, why not support `_defaults`, `autoconfigure` and `_instanceof` for child definitions?
1 big reason: reduction of wtf moments where we arbitrarily decide override logic. PLUS, since `_defaults`, `instanceof` and `autoconfigure` *are* applied to parent definitions, in practice (other than tags), this makes no difference: the configuration will still pass from parent down to child.
Also, using parent-child definitions is already an edge case, and this *simply* prevents *just* those services from using the new features.
## Longer reasons why
The reason behind this is that parent-child definitions are a different mechanism for "inheritance"
than `_instanceof` and `_defaults`... creating some edge cases when trying to figure out which settings "win". For example:
```yml
# file1.yml
services:
_defaults:
public: false
ChildService:
parent: parent_service
# file2.yml
services:
_defaults:
public: true
ParentService: ~
```
Is `ChildDefinition` `public: true` (so the parent
overrides the child, even though it only came from _defaults) or `public: false` (where
the child wins... even though it was only set from its _defaults)?
Or, if ParentService is explicitly set to `public: true`, should that override the `public: false` of ChildService (which it got from its `_defaults`)? On one hand, ParentService is being explicitly
set. On the other hand, ChildService is explicitly in a file settings `_defaults` `public: false`
There's no correct answer.
There are also problems with `_instanceof`. The importance goes:
> defaults < instanceof < service definition
But how do parent-child relationships fit into that? If a child has public: false
from an _instanceof, but the parent explicitly sets public: true, which wins? Should
we assume the parent definition wins because it's explicitly set? Or would the
_instanceof win, because that's being explicitly applied to the child definition's
class by an _instanceof that lives in the same file as that class (whereas the parent
definition may live in a different file).
Because of this, @nicolas-grekas and I (we also talked a bit to Fabien) decided that
the complexity was growing too much. The solution is to not allow any of these
new feature to be used by ChildDefinition objects. In other words, when you want some
sort of "inheritance" for your service, you should *either* giving your service a
parent *or* using defaults and instanceof. And instead of silently not applying
defaults and instanceof to child definitions, I think it's better to scream that it's
not supported.
Commits
-------
|
||
|---|---|---|
| .. | ||
| Bridge | ||
| Bundle | ||
| Component | ||