Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Not allowing autoconfigure, instanceofConditionals or defaults for ChildDefinition #22563

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged

Conversation

weaverryan
Copy link
Member

@weaverryan weaverryan commented Apr 28, 2017

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 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 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) 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:

# 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.

@weaverryan weaverryan force-pushed the exceptions_parent_instanceof_autoconfigure branch 3 times, most recently from 335b11b to 9939921 Compare April 28, 2017 02:18
@nicolas-grekas nicolas-grekas self-requested a review April 28, 2017 10:30
@weaverryan weaverryan force-pushed the exceptions_parent_instanceof_autoconfigure branch from 4944a47 to 710d9a8 Compare April 28, 2017 15:50
@nicolas-grekas nicolas-grekas added this to the 3.3 milestone Apr 28, 2017
@weaverryan
Copy link
Member Author

This should be finished now.

It highlights a clear philosophy: basically, when you set a parent key, you are opting out of the autoconfigure, instanceof functionality. And we try to warn you with exceptions where possible.

Your parent class can still take advantage of these... and most of those settings will cascade onto the child (notably, not tags, consistent with child/parent functionality). So, your parent's Definition is configured... and then that's applied to the child.

Also in this PR (but a bit related): we prevent calls and arguments from being applied to the autoconfigure instanceof. This is because you can't override / opt out of these. And by throwing an exception, we still leave it open to allow these for 3.4, without breaking BC.

Again, please check the test cases in the integration/ directory!

Cheers!

Copy link
Member

@nicolas-grekas nicolas-grekas left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍
100% agree with Ryan's comments, this is required to make things consistent and remove the need for multidimensional merges that have no generic solutions.

@simensen
Copy link
Contributor

In talking about this with @weaverryan today, I asked under what circumstances might people want to use parent in which they couldn't solve the problem with _defaults, _instanceof or autoconfigure? Neither of us could think of a concrete example.

Personally, I've only used parent a few times. All (I think?) of those cases were to cut down on manually wiring the same thing across several similar objects. If _defaults and _instanceof/autoconfigure had been available I think I would have been able to avoid parent altogether.

Said another way, it feels like _defaults, _instanceof and autoconfigure eliminate the usual need for parent altogether. So it seems like disallowing their use together ( parent plus the new features ) for children seems fair. At the same time, we could consider marking parent for deprecation to get people to use the new features.

…ildDefinition

Also, not allowing arguments or method calls for autoconfigure. This is a safety
mechanism, since we don't have merging logic. It will allow us to add this in the
future if we want to.

The reason is that parent-child definitions are a different mechanism for "inheritance"
than instanceofConditionas and defaults... creating some edge cases when trying to
figure out which settings "win". For example:

Suppose a child and parent definitions are defined in different YAML files. The
child receives public: false from its _defaults, and the parent receives public: true
from its _defaults. Should the final child definition be 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 the parent
is explicitly set to public: true, should that override the public: false of the
child (which it got from its _defaults)? On one hand, the parent is being explicitly
set. On the other hand, the child is explicitly in a file settings _defaults public
to false. There's no correct answer.

There are also problems with instanceof. The importance goes:
  defaults < instanceof < service definition

But how does 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.
@weaverryan weaverryan force-pushed the exceptions_parent_instanceof_autoconfigure branch from 710d9a8 to a943b96 Compare April 28, 2017 21:09
@theofidry
Copy link
Contributor

👍 for that PR.

You could still need a parent to extend a definition of a library, so I don't think parent should be deprecated.

@fabpot
Copy link
Member

fabpot commented Apr 28, 2017

@simensen we had this discussion during DrupalCon with @weaverryan and @nicolas-grekas and we came to the same conclusion. We should deprecate parent in 3.4 and remove it in 4.0. I like when new features allows to remove some other ones.

@fabpot
Copy link
Member

fabpot commented Apr 29, 2017

Thank you @weaverryan.

@fabpot fabpot merged commit a943b96 into symfony:master Apr 29, 2017
fabpot added a commit that referenced this pull request Apr 29, 2017
…defaults for ChildDefinition (weaverryan)

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
-------

a943b96 Not allowing autoconfigure, instanceofConditionals or defaults for ChildDefinition
@weaverryan weaverryan deleted the exceptions_parent_instanceof_autoconfigure branch May 1, 2017 12:56
@fabpot fabpot mentioned this pull request May 1, 2017
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

7 participants