Magento Architectural Discussions, 27 Mar 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Magento Architectural Discussion, March 27th, 2019

Description

- guidelines for `final` keyword
- exceptions and module dependencies in Backward Compatibility guidelines
- guidelines for plugins, observers, dependencies between application layers, idempotence of Service Contracts

Meeting notes: https://github.com/magento/architecture/issues/135

A

A

Yeah about final keyword: there was a discussion in the community like channel I touched this link to this discussion and my question is we have a static check in magenta call right now: it's HP MD, but I will replace it with good sniffer. It doesn't matter, but it says that final keyboard is prohibited in magenta and there is nothing about that in the technical guidelines. So my question is: maybe we should add this or somehow explain to the developers how to use it? Why not to use it.

B

A

Is the current recommendation, so current recommendation is not.

C

Believe it implicitly management in terms of section we say it is not possible to define all my tetanic method, firefights and so on. So maybe there is no reason to make sure that someone else pacifically the obligatory Cobra got into this new change is very to the positive some design namespace, not just PHP, or something beside that I.

D

C

To use ESP ESP initially.

E

C

I might on PHP, with euphemism.

C

A snip, but it's pointed in not so good way, so I was forget something in space like design just to English. Some coding style checked from design.

A

The problem is right now, like name space, doesn't matter and it can be easily changed. My question is: we need some source of truth for a developer, so you can send Santa, link and say here is technical guidelines.

C

We completed today that you write accessible code and also we mentioned that final method. Encoders cannot be intercepted.

C

Some like Ruby the number for the reference in.

F

Okay, is you probably so we have a several dates and deadlines.

F

Right now we will let you know one rule I believe it's if we had at all this, but maybe we didn't mentioned it explicitly. This is that we should not change its data, except that object, which is subject now plugins as well as we should not change the state of object that we receive. In our events.

F

We need this rule to prevent significant changes. Its main effect not only other plugins in shame, but overall application liability also I put a traditional one regarding it, impotency for service contracts between Z's, okay, oh you know, that's we are going to support poor person. Kronos called our system are all our Web. Api can be called synchronously starting some time ago. So we have to be sure that all these actions can be processed correctly, so it must be. All of them must be impotent.

G

It would be good, like my question, was like what what should we do, this collection? What would if.

F

You can write, you can write a plug-in and cover incoming arguments. You can cover out current arguments, but you never should change it such because all other even expects to receive the same subject.

G

E

F

Can we have only one intercepted object which is subject I.

E

Understand what you're saying.

E

Foot soldier I.

F

Will put this combination.

E

Question should.

G

It should it be merged this item about data object. We would just have a generic items that object, states should not be changed in the plugin. Is.

F

A little different these data object. This is data transfer object to the expectant that Seyi it has to be. It has to be serialized at some moment through the network, so we don't want to put any additional application logic on on it, even wrappers, because you know assume you have this. Rather, we will have to replicate it on different points.

F

Our application in case we have as several instances which communicates between each others. We have to make sure that we have a same trigons and wrappers on both sides, so this is. This is different roles.

G

May be, then, we need to provide more examples in each of the items, because I would imagine to a lot of developers. It would not be clear why there are two items regarding plugins one for that object, so.

F

It's it's pretty straightforward role, don't put a plug-in for Gators.

F

And don't change the status option.

G

E

But it's different in sizes. You should.

E

F

Might have like an interpreter to, but.

F

You can change.

H

This example doesn't seem clear because we say layer must and depend on layers even walks it, and then we give some example which operates, presentation, layer and doesn't say anything about like which layer is dependent on which layer and a no to me to me doesn't look clear. It looks clear to everybody else.

F

From is clear: that's presentation, layer communicates to service control fire, not the other direction, or it's always through. That's. We can representation layer because business logic layer. We have access layer on disorder. If it's your service context as a dependency and your presentation layers akka blocks.

F

Definitely your your. Your service context has.

H

As a mistake, yes, so I believe examples are clear and also the statement itself looks like we should just say: polo player architecture provides reference to diagram, and the gut diagram should show the direction of.

F

Communication, you, you want more detailed examples for this one. The main purpose of this point to prevent cases when somebody initiate a state and expect to receive the state, for instance, initiator states in a block and expect to receive a result of the states in service contract.

H

Understand what you're trying to do here is like what you're trying to plan, but the phrase in itself doesn't seem clear to me, especially layers that invokes it like mustn't depend on whether the folks at I believe we should just say lower layers. My son depends on higher layers and.

F

H

Why you know the main layer should depends on your presentation. It must never depend on presentation.

F

H

Persistence must not depend on.

F

The main layer yeah exactly exactly it's the same as as your your domain layer should not depends on low level knowledge, for instance, transaction state, or something like this, so this rule always will work in in both directions.

F

We are still going to build the plugins for our level entries like a resource model safe and by using this plugin. We still allows to call something like this for this contract to receive some data, in this case, a service contract which.

F

As a result was called as soon as a wall, overlayer could not should not depend something.

F

Even if this is lower level, not a presentation.

H

Standard approaches of where one is, you can talk only to one there below you and second, you can talk to anywhere below you I believe you want to have only one layer below you right, probably.

D

H

Specify it somewhere as well, because it's not clear from from from here that I mean directly the possible direction and number of layers you can skip is not specified here and it's not clear.

I

But I think I'm told is talking about different direction of dependency, so he says effective at least. If I understand correctly this rule, it says you must not depend on the layer above you, yes,.

H

And what I'm suggesting is to additionally say that you must not skip layers, you can depend on layer below you, but only talk to immediate neighbor. You cannot talk, you cannot skip layers when you talk.

E

It covers the case layer.

E

E

You probably can.

E

F

It's some fancy diagram for.

J

K

But how to know what is a lover if I call some layer from my begins, how I know it's over or not.

E

K

You definitely need their harm and good example. If myself tried to sent in my own will call some.

F

K

To sentient knowledge in this area, see you go to a presentation where to build. The email me is not clear, is hesitation layer, as that call it from the serozha is okay or not.

L

Here we have a concern regarding even not sleep inquires but contacted fires, not immediate. Oh how it sounds to me like your, whereas know for give me the role I would provide this.

E

Okay, so what your example.

K

F

A phone system of templates building.

E

I think if you.

F

Look like this, your application to properly design, because in your service contracts you have to manipulate, is a more abstract stuff like email and text, renderer and other. In this case, you will have another layer which will be a more lower layer, video when you will have a touch things like them: email, templates and whatever, but never direct here combines this stuff inside your service contract. Is this clear.

F

E

H

Yeah when I have question, are we talking about all kinds of messages.

D

Or we're talking.

H

About only like those which can like it should be, it important, post, post ocean update important right.

F

Oh, maybe we don't by.

H

Default I mean we can do it at IG important, but.

F

Yeah, we may be the post method for a nuclear service contract. Soon we.

F

We can one service contacts can be called several times with the same arguments.

F

So we have to guarantees it no matter. How many calls we will receive state of our application will be correct.

F

In my understanding, all of our new services must be put I.

H

Would say this requires separate discussion because it wasn't discussed with anybody.

F

We already implemented a synchronous, Web API. Is we previously because right now we have a this conflict.

H

They still have posters, nice to have put and the to operation.

I

Modes like simple.

E

F

But in this case, if you work exactly is the same as well.

E

So the truth, but if I know.

D

K

My question of how to send a mail in magenta.

E

F

Really depends their implementation. You know as soon as your implementation.

F

Is sent or not if.

K

Their possibility to make all might attend the content. You guys think about rest of our vacation. Do not.

D

K

I always cases very much.

I

What would be example where the method can be important.

K

Method that call external systems.

I

K

I

Is it not important if, if I have an operation with same arguments and I know that the operation I already performed, dependent and I know that a cold external system is not in that.

F

K

Definitely on next call it can skip I. Think.

I

For simplicity, we can say that our implementation can be important if an external caller is not in the important. It's it's a simplification, and we allow that allow allow external cause like in sending a second email to a customer. If the nature isn't.

E

I

The way we already have I think already have, with impotence in operations for.

E

F

We have plug-ins for this and you know that's mutable state of objects. This is main reason of incompatibility between different extensions. I know that this is I have to achieve, but in general I believe. All of us want to implement this rule. So we have a plugins for more advanced cases. Why should we recommend to use the events to change the object States because you can change and change anything everywhere?

F

As a result, your extension will not be compatible and your code will not be compatible with other system which expecting the system behaves in the regular way.

K

Five minutes.

F

Good question but I believe that this is not that we can do not. We can change right now, but at least we can recommend to do not change the state even right. Now.

K

E

I

Mean I agree with this: it just very often I get asked. Why do we have? Why do we have rules that are not achievable or are not present in system today, I think historically, events were always advertised as a way where you can modify the state of your object. Maybe, like some people are asides and ways to modify objects, some didn't, but they are very often used for that in including our system.

I

You had you modify state and now.

D

I

I'm I completely agree with this rule it just. We have to all agree that we understand what what it means. Okay,.

E

E

Is there another way that you what.

F

I

Howie, what's better way to modify what was a better way to modify things and let me think about it, so I call a method right so usually when we, when we have this event, we have this events before something happens like, for example, before an entity is saved right, he's persistent, so we send an event. We trigger an event called before saved before right and we send that entity to that and people change something there. But today we have plugins in the plugin.

I

You can plug-in eyes the resource model, the actual persistence method, and you can modify argument. So you can have immutable, so you can provide new. You know, so you accept argument. You can create an object and have immutable behavior.

I

So having immutable. Behavior is better and using decoration approach and modifying arguments in you know decorator.

E

I

Sure I'm sure that there is a scenario where were there is only event. It's very convenient and.

I

I'm sure that there are cases yeah, but this is technical guidelines. This is.

J

F

Direction, oh, we yeah what we can touch you, but the.

K

Community that, if you the place where you cannot do not do not change data, just great who requests racism is another class which you'll be using inside extent.

K

Prices are just great bought, presents all good, we'll go to it.

F

This will be great.

F

Application involvement.

F

Change argument this rule plays in pair with rule, don't change the state of subject in interceptor. So yep you have enough instruments to do something you can substitute. You can substitute output.

F

See how we can achieve this, because each particular example may have a different solution and I I don't want to recommend a Mason with this poem because you know depends on a situation. Solution can be different from change in your application to build extension extension.

E

F

Good idea, blog post, this is good way: how to have to show a possible possible ways, helped us all.

F

F

Ability guidelines- okay, guys, be honest- updates in our compatibility guidelines. So right now, as adding a throw in a new exception, type inside of your court will be back button compatible change because.

F

You don't know who are the quality, calls your method and how's. This call already Hamlet for us. By solving your exception, you potentially may break a logic of other extension. Is this clear.

F

Also, including a new exception, or any type of exception,.

F

So each stage like this, we will be treated as a as a pie. Job change, I believe my my teacher.

K

F

Exception you, you throw useful Tom, you won't be catch properly as the result application due to broken applications in I. Leave you with a broken.

K

Maybe I'm wrong, but I believe help in technical guidance. This stage that your code should not if business logic should not depend on the exception by.

E

Business for your business logic means that when you.

F

Exactly but still you may rely on exception.

F

Exceptions from Isis: it's really invoked because this exception, you know you may have a decision to reduce this call or I, don't know, shut down resources. That's.

K

My concerns are for me, my church is nice that we not may be. Definitely you Pro break all except its major. If my change is definitely not major changes as messes minor changes, there.

F

Are no state like off working application, so this is. Why is this is major yep.

K

Definitely all changes can break things. Yeah.

F

But a dimensional argument do not do to save effector like a detainee exception. I.

K

My, for example, if someone create object, is construct, definitely can break very so.

E

We have a simple rule: major change is changed in API minor, changing is changing, so if you are extending something and when we change something which imports extensibility, we should increase minor version. If we change API, if somebody calls this method behavior changes. This is major. That's very simple. We cannot predict how many people be important based on major or minor.

E

It's very unclear rule.

E

So this this change, adding an exception, modifies exception, is part of the method. Api, that's why it's I.

G

Think it it makes sense, but I would also clarifies it. It would also be applicable to a non API classes, so they used by API classes, because if you use, if you have a collaborator in your implementation, and even though you don't mind if I API class and justify the.

G

Interaction in a.

G

E

G

I agree just may not be obvious by voices role. If we, if we are going to add that you cannot take new exceptions in the API class, then you probably should also say that it also implies to.

E

E

Very simple because you.

E

Must be either function cos function, B.

E

But you have a chain reaction. If you follow oil all been lying principles, you will. You will.

E

K

Yes, it definitely paid to change and open. If we decide it's major, if I can change all our next break, maybe.

J

K

Maybe different.

K

Because we don't have all our wires exception and interfaces by this exception right.

F

Now we are working to build a good static test to Cairo disease. Oh yeah, right now we have this problem and we are going to solve this.

F

In course, we sure I believe yeah.

K

For example, I, don't.

K

E

K

New rule you make.

K

K

You don't break my play. I required.

K

F

Same situation, the same situation: you have your extension, which never calls which never calls at product save massive and always calls on the gate, and we change a save signature. Okay, you're you're still will will work, but all other extreme should do not. Oh. This is not example, I believe the changes of a API class signature is significant. One.

F

Because your example.

E

Exceptions but.

F

Another getting composure, I put I, put additional section into this document, which is also dependency and I, believe that adding of composing dependency for an existing model is breaking change because after client.

F

Installed version exists, a huge probability that some photos will be disabled. All you know that's how many blog posts and tweets we have regarding, as effect, how to disable some agenda extensions or encode. As a result, I believe that you are you caught in party releases should not depends on depend on exist in religion models, because we should we cannot grant use it. This model is already enabled, and after a bunch of great client view, a client instance will be in workable state.

F

So this is still safe to add any new model and depends on this new model, but in case you want to add dependency for an existing model into your new model or for an existing model. This will be a breakage, for instance, if you identified that's okay, your wish list, as has to depend on MSRP model you. You cannot introduce this dependency in a compatible way, because you cannot guarantee that MSRP model is enabled it's a particular instance. As a result, and as a result of updates, you will put a break a client instance.

F

Because you will deploy this new model, given you code, so this new model cannot be disabled, I.

F

Can explain this to technical writers and think how we can rephrase this, but definitely we solves for several times such such cases, when we are put in some some dependency into existing models modulus for for correct for existing modulus. As a result, it is impossible to upgrade the client, because.

F

Existing as excuse me released previously released models.

F

So, for instance, if you're Rich List model does not depend on MSRP model, you cannot say that your wishlist model will will depends on MSRP model in patch series, because you don't know is this: MSRP model exists at the quality instance or not okay, to introduce as it changes for minor minor versions, I believe, but not for much.

K

Yes, it means that you cannot, for example, in 23.1, introduce new models. You.

F

Can introduce a new models, but you cannot introduce a dependency on the previous.

K

F

K

I mean that you cannot register new models. That depends on this.

F

Yep, you can release a new models which depends on the existing one models, but you cannot release all the J released models with a new dependency to another release models, but.

K

I still have the same problem if I have disabled.

J

K

Only about installation system problem just with resolve and dependency, because if you current model, can depend on exact variation of.

D

Models and you install.

K

It model can depends on a different version. It will be a conflict of models. I understand.

F

Your question ad: this is just a little bit Azure scenario because forever for new models, you just checked: that's your version is not satisfied and you cannot install it, and this is okay. You do not introduce any behavior with existing model. This is what's at a worse situation, because you already have it.

F

Maybe you are even in.

F

F

H

D

Just coming back to the retry example, if I would have looked at a recipe, I don't know with Magento service contracts, mesh well with rest semantics, the.

C

D

I see it might consider these all kind of our carbonyl concepts.

D

When a client makes a cult list of doesn't matter, is internal or external.

D

The only one corner case where you won't really notice over did the work. That is when the next a network socket goes up right. You go to reset in all other cases unit. Please get an HTTP response back. So if you got a full guarantee from the server that, oh five to say, 500, you know that silicone are completed or an issue or whatever line can safely send an on I'd important request. Again, that's a guarantee to the so assured implement.

D

So our featherless server internally can have retry lotteries. It's completely, oblivious to the client, so we'll try its best when it gets our requests.

D

Regardless of server implementation, client has to look at the response code to tell was that night important, essentially the server transfers that problem to the client? You know whether the client should retry or not, because so will indicate very clearly that I feel unless you see a 200. Ok from my service, I've guarantee I'm gonna take care. It doesn't matter whether we do asynchronously in the.

D

Yeah I know so I guess from my perspective, in the service, giving for an on item request to our response. So where is guaranteeing that? Okay, it was done annually. But if it's an error response of some sort right, it is expected for the trying to retry. That is whatever.

E

D

When you use database gives you the.

E

D

Few things Klein did not find should not be generating IDs. It is a service responsibility, because, if you look at the rest, is everything about documents right you use, clientís completely alluvial line is saying: okay, I'm, making a non idempotent request here and bring of course, I want to create a new document. It on men could be product or whatever it is coming in a service respond to say yes, I caught document, and here is the ally of that new document.

D

Now for all subsequent calls, you have to present this ID to you know, make changes to it, delete it. So the pure resumix client is not.

B

D

Client's own documentation, whatever it is whatever below that a sense, so I can have its own ID mechanism for reference.