[Messenger][RFC] Improve Messenger to support other app consuming/recieving the message

[tristanbes]

Description

Hello,

I'm a new user of Messenger; Thank you to all contributors of this promising component; Yet, it gave me a hard time to solve some basic problematics.

Example

Let's say you have 2 different Symfony applications.

The application A (big, monolithic) is responsible of publishing the message to the queue using messenger. Until here, it's ok, everything is covered by the doc.

Then the queue (Amazon SQS/SNS) triggers a new Lambda execution, passing the message to the function (which is a SF Application B, responsible of generating a PDF); This behaviour is specific to Amazon, but it can be generalized of "messages can be pulled or pushed".

The problematic

A. On all examples on the doc, you assume that the consumer is the application that put the message in the queue itself;

But what if it's not the case ? What if it is another Symfony (smaller) project, responsible of generating the PDF only ? What if it's another language ? (JS ?).

B. The Messenger component sends the messages with some namespace inside;

So if your message class lives in App\AwesomeBundle\Message\PdfMessage in application A, and App\Message\PdfMessage in application B, then application B won't be able to decode the Message;

C. What if the message is not pulled by the messenger component itself, but somhow is recieved (pull vs pushed) ?

---

I achieved this scenario, and here's what helped me:

• Happyr/message-serializer
• http://developer.happyr.com/symfony-messenger-on-aws-lambda
• http://developer.happyr.com/sns-retry
• @Nyholm (thanks again ;D)

Do you think some of these problematic are generic enough to be covered by the component code or by the documentation ?

Thank you :)

[dunglas]

Isn't the Symfony Serializer integration enough for these cases? When using Serializer, the message is encoded in JSON, and can easily be decode by any app able to deal with JSON.

[weaverryan]

We’re covering this in the coming days in SymfonyCasts:

A) symfony Serializer for turning messages into JSON if the worker will be another app. That’s one line of config.

B) custom Messenger transport Serializer is your app is consuming messages that originate from another app. That’s one class and one line of config (and if you want, you can use the Serializer component from within the this custom transport Serializer).

Could the problem be just a lack of documentation to point this out?

Thanks for the convo!

[ragboyjr]

So, I actually was able to successfully get messenger working with magento 1.x to a symfony application over redis. Both applications send and receive messages from the other.

What we ended up doing was just create a third repository for storing our shared message classes between the two, and then installed the repository as a git submodule and used composer path type repositories to include into each of the applications.

I think using shared classes is safer than maintaining two separate versions of the class, but, you still have to be careful that any message class changes are done in a backwards compatible way to remove the tight coupling of those apps being deployed together.

[Matth--]

We have an application that is synced up with a .net process. We use a custom Messenger Serializer (described by @weaverryan) for inbound and outbound messages.

[rjwebdev]

I am struggling with this issue myself currently at work where I have to publish messages onto a queue where another agency will pick up the messages.

I tried to create a small package that contains interfaces which would be implemented by the classes that will be published so both of us would end up installing this package. In my messenger config, I bound the interfaces to the different queues but this did not work since the class names were added to the envelope and not the interface names.

[pounard]

@rjwebdev

but this did not work since the class names were added to the envelope and not the interface names

That's one of my main concern with messenger (and symfony/serializer component as well): you cannot give a map of names or aliases that would map to PHP class. In a project we wrote this: https://gist.github.com/pounard/25d2a8d4502a6bb61e660957196a5773 (which is then filled with an alias to PHP class array map, from a custom bundle configuration).

Then you specify your messages with business oriented names and not technical ones. One missing bit in my example code is that we also have a custom messenger's serializer that maps the names the other way around to send them instead of PHP class names (PHP classes got converted to business name when published, then converted back into PHP class names when consumed).

[pounard]

@ragboyjr

I think using shared classes is safer than maintaining two separate versions of the class, but, you still have to be careful that any message class changes are done in a backwards compatible way to remove the tight coupling of those apps being deployed together.

If you have more than two applications (currently we have only one but we are targeting a client environment where we'll have at least 5, each with their own maintainers and release plan, plus one end-user centric dashboard that will arbitrarily consume events to populate a client dashboard).

In this future, having other technos than PHP is considered, and having stalling application for 1 or 2 year is also a possibility: we have to specify our messages as a high-level business-oriented specification unbound to the technical aspects. Which means, backward compatibility when possible, having the most recent upgrades able to talk using more than one version of the protocol, etc...

In that regard, sharing code with every application is not possible.

[pounard]

@tristanbes

Do you think some of these problematic are generic enough to be covered by the component code or by the documentation ?

I think it should, I opened #33373 #33368 (messenger) and #33399 #33397 (serializer) for discussion some time ago. You might also want to participate in #32049

[ragboyjr]

@pounard for sure, you gotta do what works best for your team and roadmap. A custom serializer might be the best way to go if you end up sharing messages across various languages. If you can keep in php, a shard repository that can be updated at its own cadence in the individual repos might be less work.

[pounard]

@ragboyjr

That's the usual answer I get :) I do not always agree though. Regarding incoming content type I still think that the messenger component should handle this natively. For the rest, every bit is subject to subjectivity and one's need. But all remain interesting topics to discuss.

To go further in this direction, there's still a probability that we'll switch from the messenger to a more bare-metal AMQP implementation instead one day, if the messenger doesn't easily answer our needs; that's why all our business code sends its messages via a custom interface, and messenger, in our project, is completely hidden behind it. It's replaceable.

[weaverryan]

Are you suggesting that the message type would be set on the content type and then a system would allow you to map each type to a class in your code (that the Serializer works then use fire deserialization)? Or do I misunderstand?

For us, our custom Serializer is a basic switch case statement: we determine what kind is message is coming in (e.g. by reading some info from a header or just looking at the field), then use that in a switch case to create and return different message objects for each.

[pounard]

@weaverryan

Are you suggesting that the message type would be set on the content type and then a system would allow you to map each type to a class in your code

content-type and message type are both very different things, I think they can be handled very differently.

Message type may or may not be exposed by the bus implementation, for example in AMQP you can retrieve messages with the 'type' property (as of now, the messenger stores it alongside the message header, in the form of the PHP class name). Converting from an arbitrary type to a PHP class could be a feature of component/serializer, or could be handled by the messenger serializer (which I think makes more sense).

Regarding content-type (i.e. format in component/serializer semantic) should be handled by the messenger serializer class.

For us, our custom Serializer is a basic switch case statement: we determine what kind is message is coming in (e.g. by reading some info from a header or just looking at the field), then use that in a switch case to create and return different message objects for each.

I think that a basic version of that based upon configuration could be in messenger core, all it needs is a map to be injected, using the dic or annotations (either one or both). You then wouldn't even need to implement your own.

[kafoso]

Symfony Messenger's headers are very tightly coupled with not only Symfony Messenger, but also with PHP as a whole through its utilization of PHP namespacees.

This is quite the hindrance for us who wish to utilize other programming languages that publish and consume messages from the same RabbitMQ server.

The "type" header always points to a PHP class (full namespace). But what if the message is an integration event, consumed by an entirely different code base with a different namespace? It is a very greedy coupling.

Stamps are a great feature. I like how they're handled. However, again the coupling to Symfony and PHP is tight, as headers like "X-Message-Stamp-Foo\Bar\Baz" will exist, where "Foo\Bar\Baz" is - again - a PHP class reference.

You are completely locked into this naming convention if you use Symfony\Component\Messenger\Transport\Serialization\Serializer and must then write your very own serializer (implementing Symfony\Component\Messenger\Transport\Serialization\SerializerInterface). This is of course doable, but I feel Symfony Messenger should support overriding of header keys.

This could be fixed for stamps by introducing a new interface, e.g. "NamedHeaderStampInterface":

<?php
namespace Symfony\Component\Messenger\Stamp;

use Symfony\Component\Messenger\Stamp\StampInterface;

interface NamedHeaderStampInterface extends StampInterface
{
    public function getHeaderName(): string;
}

Stamps that implement this interface will use the string provided by getHeaderName(), instead of "X-Message-Stamp-". Yes, there is a very real risk of name conflicts, but that should be something developers sort out themselves.

Mapping the header names to their respective classes is another thing which must be handled, such that when consuming, things may get glued back together.

A thought on how to handle the "type" header differently

Let developers define a core namespace for the messages being sent to an exchange. This namespace is then subtracted from the full class name.

Example:

<?php
use My\Application\AMQP\Event\Foo\Bar\Baz\UserWasCreatedEvent;

$namespace = 'My\Application\AMQP\Event';
$userWasCreatedEvent = new UserWasCreatedEvent;

if (0 !== stripos(get_class($userWasCreatedEvent), $namespace)) {
    throw new \Exception(sprintf(
        "Class \\%s must be located under the namespace \\%s, but it is not",
        get_class($userWasCreatedEvent),
        $namespace
    ));
}

$type = substr(get_class($userWasCreatedEvent), strlen($namespace));

echo $type;

Output:

Foo\Bar\Baz\UserWasCreatedEvent

To stitch it all together again, the consumer will simply have to prepend the namespace. We have now gotten rid of the ultra tight coupling in PHP and to the application from which a message originated. However, the remainder of the namespace will still have to be exactly the same.

The above is also very easily converted to namespacing in other programming languages like C#, Java and Node.js. E.g. simply replace the backslashes with dots.

---

Anyone else thinking this would be a great change?

I may even be done without backward compatibility breakage by simple writing an alternative Symfony\Component\Messenger\Transport\Serialization\Serializer, e.g. Symfony\Component\Messenger\Transport\Serialization\CrossApplicationSerializer, which then handles - you guessed it - the cross application aspects.

[pounard]

I just think the messenger component should have a type to class map, per default returning the class name, but configurable, either using some YAML, some static method call, or using annotations (I dislike annotations I would not use them anyway). The messenger's serializer is way too coupled to messenger own logic making it very hard to replace or decorate. We done that for a year, and ended up ditching the messenger component completely, using a 100% custom bus using pgsql now.

[carsonbot]

Thank you for this suggestion.
There has not been a lot of activity here for a while. Would you still like to see this feature?

[tristanbes]

Carsonbot (🤖 ) I didn't work on this problematic since 1 / 1.5 year. Has this issue improved since ?

[carsonbot]

Thank you for this suggestion.
There has not been a lot of activity here for a while. Would you still like to see this feature?

[adioe3]

I just think the messenger component should have a type to class map, per default returning the class name, but configurable, either using some YAML, some static method call, or using annotations (I dislike annotations I would not use them anyway). The messenger's serializer is way too coupled to messenger own logic making it very hard to replace or decorate. We done that for a year, and ended up ditching the messenger component completely, using a 100% custom bus using pgsql now.

This is exactly what we've implemented in a separate bundle in my company and it works well. I am in favor of this approach for two reasons:

1. free-form event names are great (user.farted.today versus App\Events\User\Fart\Today)
2. you can keep backwards-compatibility, i.e. this would work:

routing:
  super.custom.event: transport
  Old\School\Event: transport

events:
  super.custom.event: Some\Php\Class

Now I can integrate non-Symfony applications, yay!

[carsonbot]

Thank you for this suggestion.
There has not been a lot of activity here for a while. Would you still like to see this feature?

[carsonbot]

Hello? This issue is about to be closed if nobody replies.

[kafoso]

I would still like to see at least opt-in support for a loose coupling.