[2.3] [WIP] Synchronized services...

[fabpot]

Q	A
Bug fix?	no
New feature?	no
BC breaks?	no
Deprecations?	no
Tests pass?	yes
Fixed tickets	#5300, #6756
License	MIT
Doc PR	symfony/symfony-docs#2343

Todo:

• update documentation
• find a better name than contagious (synchronized)?

refs #6932, refs #5012

This PR is a proof of concept that tries to find a solution for some problems we have with scopes and services depending on scoped services (mostly the request service in Symfony).

Basically, whenever you want to inject the Request into a service, you have two possibilities:

• put your own service into the request scope (a new service will be created whenever a sub-request is run, and the service is not available outside the request scope);
• set the request service reference as non-strict (your service is always available but the request you have depends on when the service is created the first time).

This PR addresses this issue by allowing to use the second option but you service still always has the right Request service (see below for a longer explanation on how it works).

There is another issue that this PR fixes: edge cases and weird behaviors. There are several bug reports about some weird behaviors, and most of the time, this is related to the sub-requests. That's because the Request is injected into several Symfony objects without being updated correctly when leaving the request scope. Let me explain that: when a listener for instance needs the Request object, it can listen to the kernel.request event and store the request somewhere. So, whenever you enter a sub-request, the listener will get the new one. But when the sub-request ends, the listener has no way to know that it needs to reset the request to the master one. In practice, that's not really an issue, but let me show you an example of this issue in practice:

• You have a controller that is called with the English locale;
• The controller (probably via a template) renders a sub-request that uses the French locale;
• After the rendering, and from the controller, you try to generate a URL. Which locale the router will use? Yes, the French locale, which is wrong.

To fix these issues, this PR introduces a new notion in the DIC: synchronized services. When a service is marked as synchronized, all method calls involving this service will be called each time this service is set. When in a scope, methods are also called to restore the previous version of the service when the scope leaves.

If you have a look at the router or the locale listener, you will see that there is now a setRequest method that will called whenever the request service changes (because the Container::set() method is called or because the service is changed by a scope change).

[lsmith77]

will review the PR in more detail when I get home but I think proxy objects might be a better solution. we could support automatically "unloading" of predict loaded proxies referencing request scoped services when we enter/leave the given sub request.

[fabpot]

This PR is transparent as far as the services are concerned, which is one of the primary characteristics of the service container, the objects must not know that they are managed by the service container. How would proxy objects work with listeners?

[lsmith77]

As the DIC is the one that creates the proxies, it would it could track them and therefore ensure that the state is managed.

So when I access the request proxy for the first time in a service it would fetch the request service from the DIC instance inside the proxy and so the proxy would become initialized. Then when I leave the scope/enter a new scope the DIC can automatically revert the back to the uninitialized state, ensuring that the service would be initialized again based on the now current state of the DIC.

[alvarezmario]

@fabpot By the way, the example you put about the locales, is currently an issue in the 2.2 branch, but not in the 2.1.
I switch back to 2.1 just because of this.

[fabpot]

On Twitter, someone suggested volatile instead of contagious?

[fabpot]

@lsmith77: Have a look at the PR, my solution is much simpler than that, no moving parts.

[lsmith77]

Well I dont think my proposal is anymore complicated, especially of we add support for proxied services anyway.

My concern is mostly about "propagating" or "recalling" the setters, because in theory the service could have assigned a request scoped service to another dependency and it may not be practical to have to write logic in the setter to then propagate this as well.

Now in my proposal instead of recalling the setters, you simply iterate over the proxies and reset their state. this will then also update any dependency that are not managed by the DIC.

from a users perspective its more reliable and i think from the DIC point of view its not any more complex either. again assuming we want to add proxy support anyway.

[sstok]

Is using array_key_exists() really needed here? I know isset() is (was??) faster then array_key_exists().
And I'm pretty sure using null service names is not allowed.

About the naming +1 for volatile.

[lsmith77]

volatile indeed sounds better .. does Spring not have similar needs?

[jorgelbg]

I think that the volatile name does not make justice, yes the service will be volatile changing with each request, but it will be fixed to the request itself, so it will not be available outside the scope of the request which will kind of confusing?.

[fabpot]

On Twitter still, someone suggests "synchronized"

[merk]

Looks excellent.

May I suggest Propagate, given the name is already in use in the code?

[fabpot]

@merk: isSynchronized() sounds good to me but isPropagatable() sounds weird, no?

[merk]

You're right. It sounds weird.

The correct way of saying it is isPropagable -- https://en.wiktionary.org/wiki/propagable

Another suggestion: isInfectious() 😏

[marcospassos]

Synchronized sounds good! 👍

[marcospassos]

I think that @lsmith77 solution may solve the issue #6985.

[lsmith77]

how about "scopeSafe"?

[drm]

I suggested synchronized on twitter, because for some reason I couldn't comment here with my phone...

The reason I think it is a good name is because it implies that the container does some work for you.

(edit:) And volatile has an "unstable" ring to it, but that's of course a matter of taste.

[jorgelbg]

What about "Scoped Services" for the new concept?

On 2/8/13, Gerard van Helden notifications@github.com wrote:

I suggested synchronized on twitter, because for some reason I couldn't
comment here with my phone...

The reason I think it is a good name is because it implies that the
container does some work for you.

---

Reply to this email directly or view it on GitHub:
#7007 (comment)

[fabpot]

I've just updated the code to use synchronized instead of contagious.

[igorw]

propagators? s/e/o

[fabpot]

I've renamed it to synchronizers as this is more accurate now.

[stof]

@nomack84 if you want the issue in 2.1, consider the case of using the Intl locale after the subrequest, and you face it too (see my report in #6932)

[Crell]

Mixing the assignment into the method_exists() call like that makes the code quite hard to read. Can we break it out to a separate $method = line and then the if() statement? That would improve readability considerably.

[Crell]

Overall I think this looks like a solid approach, and would resolve the issues we ran into in Drupal. It should be fairly easy for us (and anyone else using the DI component) to adapt to it. Nice work, Fabien!

"synchronized" sounds like a fine name to me.

I think this works better than proxy objects per @lsmith's suggestion, since it avoids new object instantiation entirely. It looks like existing constructor-based dependencies would still reinitialize, though, which gives developers a choice of whether or not it makes sense to reinitialize their services.

[ghost]

@igorw wrote a very interesting article about the problems with syncronized services and scopes here: https://igor.io/2013/03/31/stateless-services.html I think it's worth reading as he presents some interesting points and demonstrates the problems very well.

[sstok]

A little off topic, but if we abolish scopes completely. How would we handle scope="prototype" then?

Really like to idea of @beberlei !

And it should be possible to introduce this without BC breaks no?

[beberlei]

@sstok my idea would only be about removing scope="request" (in the long run, i.e. 3.0)

[lsmith77]

i would like to propose that we hold an IRC meeting to discuss this .. ideally tomorrow evening since we do not have much time until the end of the dev phase for 2.3

[ghost]

I will attend if it happens.

[beberlei]

I am afraid i am not into the topic deep enough to be able to responsibly say something about this. In the beta phase we maintained the rule that only people with insight into the code are allowed to vote on changes. I am not sure that except @fabpot anyone (maybe @stof ) can reasonable say that he can make an informed decision about this topic.

My impression was that synchronized services is only necessary, because the request is wrongly a service, although its a data/value object. I proposed an idea about how to possibly solve this, and slowly deprecate request as a service.

This was done without thinking about consequences like:

1. How to maintain BC
2. How to adapt the current listeners to it
3. How a good solution would look like and if my proposed solution is really the best one.

"request" as a service is a very fundamental concept in the MVC stack of Symfony2, it may well be that this pushes us to necessary solutions that we could have avoided much earlier, but not anymore in the 2.* lifecycle.

I know Fabien tried to work with the request stack prior before merging this PR, so his decision to merge this took all the three points into account, which i cannot say from my experiments.

I don't think it makes sense to discuss this topic further without having a concrete alternative on the table in form of a pull request. Since Fabien tried several different approaches here, I am not sure that any option is not something he has considered already.

So without his definitive input on the topic with regard to the possibility of an alternative solutions, it doesn't make sense to discuss it.

[webmozart]

I agree with @igorw's blog post that this PR is problematic. More than being a blurry concept, "synchronized" has a very different connotation for any developer that used Java before (in Java, "synchronized" marks exclusively accessible code/services in multi-threaded environments).

I also agree with @beberlei that the request should not be a service. It is no "service" by definition, as it does not process any business logic. It's just a data container.

To me, a "requestContext" service with a getCurrentRequest() method seems like a simple and totally suitable solution.

[fabpot]

It might seem obvious today that the Request is not a service, but back in the days, I fought hard to avoid the introduction of the Response as a service as well. So, we only have half of a problem today ;)

The Request object is indeed a very special "service", or put another way, this is the only data object that is/should be managed by the service container. Why? to have the ability to inject the Request into other services.

But just to be able to get the Request from the DIC, we had to introduce a lot of different notions: the request scope, synchronized services, strict arguments, synthetic services, and perhaps more. Even if all those concepts might be of interest by themselves, the truth is that they are not that often useful when managing "regular" services (and Silex/Pimple is probably the best demonstration of that).

But as of this PR, I think we now have a working model for managing the Request and its life-cycle in the context of the DIC. I say "working model" because an average developer can now depend on the Request without knowing anything about the internals.

That being said, I'm the first to recognize that things are complex behind the scene just for this one object (even if this is probably the most important one). I also agree that having so many features just to be able to easily work with the Request is a strong smell that something is probably wrong.

So, what about @beberlei proposal? Here are my first thoughts:

• It seems that it would fix a lot of our problems quite nicely and easily, but we would loose one thing that we have as of 2.3: the ability to inject the Request like any other service (except the possibility to inject it in a constructor which seems like a very small drawback).
• I tried this approach when I was working on this PR and I hit a wall quite fast. But IIRC, it was because I didn't want to inject a request registry instead of the real Request.
• Backward compatibility is a big concern here. Not that introducing a request registry would break BC, but we would probably have to keep everything we already have (including the synchronized services) for the foreseeable future anyway. So the question is if it makes sense to maintain two ways of doing things.
• Related to the previous point, this new way will need to be documented along side the current way. We will need to make it easy to grasp and not confusing for people used to the old way. That should probably be not too difficult though.

As many of you probably know now, I'm not that good at thinking/talking about concepts without real/working code. So, I will play with @beberlei code in the next couple of days and get back here with my findings and thoughts (and I highly encourage others to do the same to get familiar with it). Then, we will list the pros and the cons of both approaches and that will allow us to discuss what to do for 2.3 and/or 3.0.

Enough for now, let's go back to the code now!

[webmozart]

What do you think about reverting the PR until a final decision is made?

[igorw]

A note about BC in relation to @beberlei's proposal: It might be possible to create a ProxyRequest which extends the Request, overrides all public methods to delegate to an internally stored request registry. There is some cost in terms of complexity/performance, but it could be done.

[fabpot]

@bschussek Whatever the decision we take, we won't revert this PR as it fixes bugs with our current approach. And as we would have to keep the current way for BC reasons, this bug fix should stay in anyway. That's one of the reasons I did not consider using a registry/proxy in the first place when I was working on a solution for the problems we had.

@igorw We discussed that already (it was @lsmith77 proposal IIRC) and that's out of question: the request is used quite frequently, so we won't introduce a proxy for such an important object in Symfony.

Another drawback for @beberlei proposal I've just thought about:

• As of 2.3, a service depending on the request (and doing it the right way) does not need to depend on the event dispatcher anymore, which obviously won't be the case with the registry. And that's an important property: a service depending on the request does not need to care about the machinery we have; it should be reusable outside of the context of Symfony.

[lsmith77]

When passing the request via action method signatures, we could pass in the plain, unproxied, request instance. There is no reason to ever put a controller into the request scope, same applies to listeners, that I guess also are guaranteed to get the proper request instance via the event object?

[fabpot]

@lsmith77 You are right for controllers, but for listeners, the problem is a bit different: they get the right request when the scope is entered, but before this PR, they were no aware that a request scope finished. And this is the only problem that this PR tried to address.

[lsmith77]

Just FYI .. we already have one place where we use a registry which is the ManagerRegistry for doctrine.

[henrikbjorn]

@lsmith77 which is also a pain point when using the bridges outside of the full stack framework as everything in components relies on a manager and the only implementation of it i have seen is the one in symfony.

[stof]

@lsmith77 the goal of the ManagerRegistry is to allow multiple managers in the app (and retrieved lazily), not to stack managers on top of each other. So we don't have the issue of poping the old one.

[samuelvogel]

What actually is the status of this issue? The fix discussed here is marked as merged, but #6932 is still open/broken. It's a little hard to comprehend this long discussion without background knowledge!