.Net Foundation Design Reviews, 28 Jan 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: Quick Reviews

Description

00:00:00 - Approved: Add StringContent ctor providing tighter control over charset https://github.com/dotnet/corefx/issues/7864#issuecomment-579398933
00:23:57 - Rejected: Is there a function to copy directories? https://github.com/dotnet/corefx/issues/15708#issuecomment-579405459

A

It's a to different concept. It just means that you as a color cannot maintain the collection it.

B

Doesn't mean Americans, it provides.

A

A signaling mechanism well and also signaling, just to one side of the exact.

C

Program here, models where you say I'm not mutating this. Therefore you don't have to make copies or make assumptions that I will mutate it or anything else. The.

D

Problem is, though, like I mean this is the this is the argument when time zones, it's on the team out in various areas like I I, generally agree with crystal, because you can't see what other systems look like. They don't have that so, for example, that it went on their path, locate, simpler stars as iostream library, locate winner tea streaming, API that they have read-only streams and and with white streams, and it's a mess because you, basically the problem, is not so much that you can express the contextual things with the callers.

D

The problem is the combinatorial results from that and like in video, only collections, it's relatively easy because you only only have two modes but the streams. For example, you have random access. You have read, you have ride ya, seek ability right like that, the holy compose all of those things and how you now pass things around in a same way and that's what he bus classes, iostream library, is relatively complicated. Considering that all you wouldn't want is just be the text file right and.

C

Well, firm, for most cases you say: if I'm not mutating this I take the read: only I, don't care about right support, I, don't care! If it exists, you only care that you can read and access elements. That's all you're going to do, and so you take that, and that means you can take both streams that can write in both and streams that can read because of the inheritance hierarchy. And if you care about mutability, yes and you take the spin it but.

D

It's a writable, the general only features you have. If you only have two features you can express it in hierarchy as soon as you have more features, how do you express a and B but not C? That's the problem of streams right, that's why that keep that you take its if it was just read and write as easy, but that's the the fundamental problem, I think for the idea that use interfaces to model features.

D

If you have combinatorial I, think for reading for read and write and with collections I could similar to spam and could be convinced that it will be workable model, but I think our challenges that we introduce that model being are 15 years more 18 years into the platform right, and so that's also why it is. If you would only implement with only interfaces today yeah it will be very hard for you doing to operate.

E

Of anybody else, anybody.

D

Else the other thing.

E

Is like our our span of tea is implicitly convertible to read. Only span of tea. Other languages, like Oh forbid that implicit conversion.

D

Can in our case that doesn't work in all cases- I, for example, extension methods. You have to do, kill him twice and like that, so they are definitely downsides and.

A

Also, it's a philosophy of the platform side, I.

D

Think it's not.

A

A confidence that see has this very extensive hierarchy and that black lines in shock started without because- and this was very like he put so much more weight on simplicity, then on you know, the kind of objectives that you were talking about being able to you know have some current is expressed in signatures in type system. Yeah.

C

I think that's also a big difference in the the general design that you wanted from 20 years ago, where the goal was mainly like GUI and web apps, where you want it simple, easy to write and understand, and today, where a lot of people are doing: micro, optimizations and they're wanting a more powerful language where you can express things, I assume.

A

I think we are still I would platform is about its simplicity and productive.

A

F

This read-only interface accomplished: does it enable people to write these micro, optimizations, so tection area.

D

F

Is this just a purity thing yeah to.

D

Me, like the argument of adding interfaces, more simplicity, that you can understand our design more because things are more uniform right now, it's kind of like odd and honestly, just a mistaken or, and that we don't have it and that honestly, these kind of oddities make it harder to explain to people. Does everyone's.

E

Calculus change, if we ignore the interface described here and say I read only set, is literally a contains T method that returns rule. Otherwise, it's an I read on a collection like you, you ignore is proper subset of is subset of overlap. Side equals the only API on here that snooze contains. Does that make everyone more comfortable with this proposal? No, actually we don't this yeah, because.

A

E

Very particular scenario is I: have a collection Oh, the only guarantee I need from this collection is o of one. Do you contain abelian? That seems like it should be.

B

E

Should be a different interface if you actually read through the description of this issue, that one method is actually what people are asking for, like people aren't asking for like overlaps, that equals stuff, like that they're asking for contains yeah.

C

But we also can't expose in I, contains interface and have it threaded through the system properly. We can no.

E

C

Can't make I collection inherit from it.

E

Stuff right collection doesn't have to have my collection doesn't guarantee alone contains. We would have to implement a new interface I of one contains and then have various collection types set really and dictionary derived or implemented just.

G

Took China just quickly, even some of our sets can table it right.

B

Nothing about assets and it sort of one, the specific type hash that Senatorial one sure, but a set just means you can't have the same thing twice. That's all sorted.

G

Services login to yeah.

B

So there is no such thing as an interface. It says: I can provide controls in a1, okay,.

E

H

Have a boolean.

E

Contain deficient contains the aficionados because they can always.

A

E

A

Home and obviously yeah yeah yeah I think the strongest argument for adding it is basically to round up like make a consistent set of Freedonia and read our data.

B

By the filling in missing, anybody who wants contain should just use, and presumably we already have I said special case, and if we don't, we can.

E

Okay, so said, that's so that's our answer about so we say if you want contains which again, if you read this issue is really what people are asking for then use link because, like no one in the issue is actually asking for I read only set, as it's listed here like seriously. No one was asking for it.

B

Well again, I don't even see anybody talking about the key connections that they got. You propose good yeah.

D

It says hey once you think it would be nothing.

B

E

Like the fact that this is even on the screen, I think is derailing. The conversation I, don't think what's listed on the screen is actually useful. To the conversation, I mean.

D

I would agree to view that to me is profile subset in the other ones are not that useful I think I've probably used overlaps at some point and that's it. But if we were to add an IV, no only said thing it would even we can never add to it. Those make more so.

I

Why why are we considering that, since no one's asking for word.

H

A

E

Just consistently, yes, yeah, that's not the issue that was opened.

E

Here, like we're, I mean we're talking we're talking about adding an interface for consistency, but it's not actually solving by a scenario. The customer.

D

And I think I think if the only thing people care about these entertains I would argue that it's already solved yeah. That's then we should justify his.

B

Pieces and since I read only list with that and the pairing between sets and list is defined. It would be great to reestablish it.

E

B

Is let's put on the whole, the missing part, all right.

E

I mean we've got twenty minutes on this already, it's like better! That's really what I'm trying to get out reckon so I've spent 20 minutes talking about something, that's not actually addressing the issue that was but.

D

I think it is an issue that it may still feel being you probably should address right, like I mean I, don't think we're saying that, like the only anybody said like you should not have them, it's just super noisy and is there another way to do the customer scenario? I think the question is more like well, there is a good issue. We don't have that interface, it looks odd. Do we care and I think that's that's kind of the conversation to me like I would not be opposed. I probably would do that these guys here.

D

I will probably not do this one and then call it a day. So.

B

Yeah I think we would have read Emily said: there's I read Emily said we would make hash that sort of said. It's implement that interface, possibly the read-only set helper type that can wrap another I set. Look like they don't if we don't think it's the big ugly one.

C

Reason don't read.

B

Only set wrapping.

C

And I said, or as for the struct that wraps, we don't look to have read only list or dictionary wrappers. So maybe that's a separate discussion. We.

B

Have the complicated read on the collection wrappers that can only wrap a reto or they can only laugh grab a list of deep which names faces that in read-only collection.

C

I

C

That's why I didn't find it.

D

Yeah I know we have it's like so in it, okay, let's maybe take.

A

D

Off, maybe I actually did not see this one here. So if you take these two guys off and we just talked about adding the interface and implementing it, does anybody feel strongly that we shouldn't do that? Would they take are innumerable, is parameter or her it all myself, I think their model of Avadh hash that and sorted said taken it again. You know yes,.

B

The theory that you could maybe then we could, in a future version, possibly five say I, said, extends I read only set and then either the language supports the explicit interface looking for moving the member to the base type or well.

C

Can't we make I set from and I read only set today, since it's being added at the same time as the interface is added, there's no risk of over low resolution breaking.

D

No ice it exists today, no.

C

No, but you can you can make that the problem with making I collection in from a diary, read-only collection, for example, is because people who have methods that take both now break notes. If I've already said that.

B

I've written a type, that's I, set I, get a compile break, but I have area not explicitly implemented any of the members. I get a compile break for not explicitly implementing the members from I read over yeah I.

A

Already don't I said: oh man, rice. If we shouldn't exclude a point where it doesn't have new members, is it still a breaking change? Yes,.

D

Because there's different identities well.

B

Then I had literally no members defined on it. It was a marker interface, then that would violate our rule of don't that marker interfaces, but it wouldn't break anybody, know.

D

Know if it's basically a subset of what already yeah.

A

Because they're.

D

Different identities, I said contains and I would only see, contains at two different methods: interface, implementations to compile it says, you're missing the other one yeah we'd.

C

Have to do it using games, yeah, yeah.

D

And that's the discussion said earlier, where it's unclear may be concluded that thing that it's a breaking change, but somebody recently challenged said that this should work for that scenario. So we have to look into this.

B

We moved it down, we could look at dem, I set to say, move the implementation like go, ask it, but if somebody was already explicit I said they can't like. We can't have both directions them both ways right, cuz,.

C

I thought the problem with dim was only for interfaces which already exist in people are already implementing yeah.

D

I think the issue be pointed out was the dim approach was that it can result in the runtime error of ambiguous matches where the compiler doesn't know, which one is the correct for interpretation right.

B

If we had each one of them point to the other one, but yeah.

F

B

Soon, as you say, you want to yet I said it's like well, you need to go deal with this now and.

D

That is to pull them off, like you know the interface hierarchies that we have, because some of them are shared diamonds by construction right, where you have, we implement an idea, nary and an I collection. At the same time, they both inherit from the same things. Now you have to default into physical mutations for the same thing.

B

I

All right, yeah.

B

I can agree with the original thesis from the issue and with Christophe. The top one adds consistency to our I collection type and IV. No introduction head makers.

E

So we do only what's highlighted on the screen them yeah.

B

We could also do what the other thing, the next blob of green, which is provide the the class read-only set, which wraps a set and prevents access.

H

To the mutating add, but that's the well so if.

A

We decided that we are adding this interface. The next level of the discussion I would want to have is do we want all these members and there isn't, for it is if the main reason why people want it is contains, and they don't want to use link, and they think that they can tell you know better protein say stuff like this implementing this interface. That just has basically contains- and you know, maybe like a smaller subset of these methods and it's super simple, implementing subset operations.

A

It's kind of not super easy, I.

C

A

Don't think I'd agree just.

C

From the perspective of not many, people actually write their own implementation over a raw array. They typically wrap an existing collection, and so you would just call mutable set dot, is subset of.

A

And all these Meadows proper, subset subsets, we have them in our I said.

C

Yeah there are there. This is basically every member of I set that it.

A

C

A

So effectively.

C

The same thing we did for read on the collection list.

A

C

B

Your movies ad union, with intersect with, except with in symmetric, except with it eludes the rest yeah so is subset of, is superset of the proper superset of the proper subset of overlaps and set people which is asking if the set equals it's not set.

B

That is not upper equals. Is the bird.

B

Which I guess is superset of 87.

B

Yes, I mean all of these. Things could easily you can have default implementations in the center. Faced with you know, extension methods, because you can write them all in terms of the taste.

B

It wouldn't set equals, wouldn't be very efficient. It's going to walk everything.

D

So people gently on what was that I'm, not I, guess.

B

Accounts the same, and it contains everything that, yes, it isn't.

C

So so the interface and expanding half set and sort of ascent implemented credit.

E

The interface contains everything. Inside later there.

B

So then I was asked. Is there a reason we want to say no to the universe? I'd rather talk.

C

My only opinion there is the both read-only dictionary and read-only collection are both an object model and so for consistency. It should likely be there as well sure, like I'm, fine, with saying the namespace is wrong. You want to add.

B

Do we want to add the class while we're doing this? What.

D

The column that I have of the time proposed was the question is: what implementation is this based on? Is it sorted? Is it high, but.

B

It says it so the constructor takes an eye set, it's just going to forward.

B

I

Asking for this who's.

B

Asking for the read-only wrapper yeah.

F

B

Like says it once then, I read only said and I don't want them casting a way that we don't mean it's like I'm gonna give them my container. Instead of it, there wasn't.

F

B

F

Use cases is clear: it's just it is anyone asking for this, or was this added to the issue Gurley for completeness.

E

There was a separate issue which is not linked to this, for someone was asking for read-only sudden-like type. They were looking for read-only half-sunken, particularly they.

B

E

B

That this was added for parity and which is the other reason we're adding the other case and.

C

For parody, it needs to be a class, not a struct.

I

C

But it would be really nice if we had read-only collection, which was actually collection, not list and dictionary, and read-only list and read-only set, which were strux. Therefore simple, wrappers that were optimized with no heap allocation or.

E

Just have a read-only keyword that removes the ability to calm you to believe the eyes on the interface.

C

E

We'd only own interfaces.

C

First, yes,.

E

You're part of the LD outright No, so.

D

What do we would think about implementing I only said also on the immutable ones, probably right on.

B

The immutable any other types that are simple the thing I said we should also add yeah yeah.

D

Immutable sort, it said an immutable hash.

D

D

Generated so bad on this laptop all right.

D

One item going.

B

And there alone I.

D

Think this one might be fast or slow.

E

Click-Click the first line at the top.

D

I'm on the screen, this way, yeah.

E

And that it links directly to the API proposal.

E

We already have a class called spring dungeon. Yes, it's part of the HTTP namespace I mean.

D

This seems consistent, I mean I, just wrote this back for the JSON HTTP extension methods and taking taking the content, and the media type is pretty common pattern than that. Yeah.

E

So there are existing overloads right now that allow you to specify the media type in the encoding as separate parameters, but control over the actual content type header value. None of those existing constructors will give you that full control, whereas if you pass in the actual header here, you now have full control, so I don't think we get head everything to take this year.

E

So there are existing overloads to take encoding and string for content out before media type and the way that those header the way that those constructors work is they concatenate them together and generate the header string for you. But that means that you don't have control over the header string. At that point, you don't have control over what actually gets emitted. That's a content, type response and I see, whereas if you pass in the actual content type header, which is.

D

E

That second parameter is hey, guess what you now have full control over. Yes,.

D

Ooh, that's what I meant so the other ones that I saw there were two overloads or tweet one. Without that, just the video one roof takes, the content, type is or the media type is a string and one that takes media type header later. That seems to be a pattern that the other API is there. Yes, in this case, it seems like just an omission. This.

B

One is it takes a string and a text encoding and then a string of text encoding input text label of what the media type is. You.

A

Know, as opposed to.

B

Taking a video, yeah and.

D

Presumably you don't.

B

Need the very reason we don't also need the encoding you.

E

Can get the encoding from the media to another million? It's all! Okay, because you put the encoding object on the beginning, yeah, it's it's a char set as a string, which means that the string content here would need to parse it out. And yes, we know and if.

G

You look at if you look at the existing constructor for string content, the one that takes an encoding and a string media type. It just uses those two things to create the media type header value. So this is just skipping all that and just taking the media type header and.

E

Steve also gave a compelling argument for like well. It could also save an allocation of who plays those so I'm, just just making sure.

B

That we're not losing a piece of flexibility, yeah I was trying to follow what the encoding no.

I

It's a good question. We did ask.

B

It what name are you and then hope that the media type edit value' knows how to look the encoding, backup.

E

Yeah it can't it would, it would require parsing, but that's generally pretty easy. I mean it's almost always going to be a utf-8.

E

I

B

The defaults next point.

B

I guess that's the mime type yeah, not the encoding time.

D

Is there a function to copy directories.

B

D

Apparently the answer is no, but that's.

E

D

But it um I could've.

E

Sworn have you had, one I mean: are we going to build robocopy into the framework we couldn't I.

B

Mean I feel that every time I've needed this I've written it so I don't think we'd have a.

B

Simple for loop or I, learned in like 2008 didn't have I, haven't, looked to see it with infinite sense. The fact that somebody made an API proposed on the market ready for review, suggested oh yeah,.

D

I think the problem that we have and I'm not sure what other things we edit since then, but in the past we didn't want to add io API. Is that have policy right? So the question is: if you have to swallow sorry this for loop, where you walk over the things, what happens if one of the copy operations fail, do you undo the you should go and you throw an exception? Do you like? What do you want the baby to.

E

Be yeah so Jeremy did comment on that and he said like behavior aside, the API shape is probably sufficient, assuming that we actually settle on a behavior and just talk another.

B

Because, presumably because somebody come in so we have directory and so that move too, but not copy too and well, you can say well it's because move is atomic. So there's no policy yeah if it's a different device move, is actually a serial copy public which could afford in the middle yeah. So we have if, if we have move- and it doesn't check that it's not staying on the same device, yeah.

E

D

I already having a thing with policy but I think what the u.s. does is the copy fields. It deletes the new thing and it just keeps the old thing like doesn't believe the old thing. So you don't lose any day that versus I, don't.

F

Like directory permissions, sorry, what were you saying for directory permissions? Could make this potentially a complex thing to implement, so it could be lagging. Is there's a question of if this would do that? But if it would, then it might be good to have a convenience method here that users wouldn't have to incorrectly implement it themselves. I think.

D

That's the gentleman in favor of the API, so whatever policy problems, we have it's probably better for us to implement them.

D

I

F

Of pains me to implement yet another synchronous files just submitted.

E

Yeah, what we, what we could do them is if we want to say that behavior should be part of the API. We can pump this back with plate. Please describe very succinctly or very uh not so simple, but please, please describe what yeah precisely what the behavior from this method is yeah, given all these out cases, because someone actually did comment on the education, so yeah, 2017, yeah and.

A

What about the comment about synchronous versus asynchronous I,.

F

Mean there's no secret asynchronous API is some of these in Windows, so well also on directory.

D

There's nothing asynchronous today: everything that's yeah.

C

I think if we were to add async filed api's we'd have to that be a different set of API design and probably in all at one thing. Yeah.

E

Yeah I would leave the synchronous, whatever assistance.

I

A

E

Synchronous, then, maybe so that the state conservation token and then what do you do when the token is corrupt? You abort the operation. You delete what you did between files, you pair.

B

With cancellation.

E

B

And as soon as you get an access denied, whatever policy we had, we have do.

C

We take cancellation tokens for synchronous, API is today, yes,.

B

We have recommendations to fix it there. You know some of them do like blocking collections. You know the.

G

Only places the only synchronous operate operation stay they take a cancellation token are those that are waiting on a synchronization, primitive.

E

This would not be waiting the same permit, but it is a long-running operation potentially yeah. Would that not qualify I.

G

Mean if that's gonna be our new stance, there are lots. Api's would be going out. That's like.

B

I would say that this one meets it, because if you were making a you know GUI program, you would put a cancel button on this operation. Therefore, yeah I mean I, think yeah.

C

I think if it were like a GUI, you wouldn't use this synchronous. Api you'd write an explicit async API on a background thread and with reporting.

A

F

I think I wrote a proper.

A

Console too, you probably want to cancel as well. Well.

C

Same thing, you're reporting information via some kind of UI, just because it's not buttons and wind forms and it's in a console doesn't mean it's not a UI.

C

Yeah III think it's probably simpler, just to say, like synchronous, API, czar, blocking and asynchronous are non blocking and can be cancelled. And then, if we see a need for being able to block and stuff, we should consider adding a cig. Api is for stuff, like this.

A

I don't know it feels a bit like a hug, but I agreed that we already have. We are already kind of. We have a lot of debt because we have a lot of operations that don't do it, but I am NOT I I'm, not thrilled about the method that taken that can take half an hour, not having ability to cancel it and I'm, not in not having ability to record progress, 13 and so on and so on. Maybe we should properly design it with progress, reporting and cancellation, because.

E

He almost ossipon progress reporting like I'm working right now with the api's in.

D

Fairness, we have, we have a PR thunder back to that end, exact same cam, delete with recursion, get all files that gives you back in the rain. There can also run for a long time. You can't cancel them.

C

Yeah and depending on your file system, you, it could take anywhere between you know, hours to seconds, even for the same amount of files, I totally.

A

I'm saying I understand that we are deep in the hole you know here, but we that one, if we were designing today from scratch, I would be a bit more careful about what we do is yeah cancellations and progress reporting there.

B

Were like the fact that this is recursive like I would expect this, to name it copy recursive, just to avoid the ambiguity or visit copy the files or is a copy. The tree or pilot copy tree. Don't.

A

Have a pattern where you possibly and then they say deep or just shallow I think.

C

We call a recursive, not deep or shallow. Okay, actually I think it's an enum and its top direct free versus oh yeah. There is.

D

That's enumeration for donita's is reducible. Okay,.

A

And what does it called recursive.

D

B

Because the top directory only the search option, if that name, don't really specifically.

B

D

You're, finding floods, you wanted to be.

D

So, what's the conclusion then so Jeremy said shape is okay. Behavior need some more thought, but like I'll, be okay with the dude with the API, then we'll give you if.

E

You wanted things like status reporting, then this would have to be obtained right.

D

But the magaman would be none of the api support that today they do the same thing. None.

E

Of you yeah, but.

D

I believe it was.

I

D

Where or like get all files.

I

That gives you back an.

B

Event reporting supposedly, if they answer a minute something they get all files, gives you back the file info right. So it like it's not gonna, take an hour. The delete, recursively like I, can't think of a file system. Word delete, would take hours, I mean like I guess. If you had a period of at your storage, that makes me mad.

D

A UNC path with the phone apps that I deleted that took three hours to complete so.

B

Versus copying files were especially if it's a problem device boundary it's like now, you have to care about all the data and not just all the tree structure yeah. So this is significantly significantly longer. There's the question of. Is it Apple? Preserving, probably not because it you can say the CP char? Is it like there's just a whole bunch of stuff, but I think it's an easy thing that we can say to the Shady's fine, but I, don't think that it I, don't know that it makes sense. I also.

C

Don't think it's terribly hard to write for each violin directory, dot, enumerate directories, directory dot, delete or whatever yeah.

D

Yeah I mean like the code that the person wrote like they say the code right I think it's not like. Oh my god, I cannot possibly write this by myself. The only problem with it is, if you have a file.

B

System that supports more than later 30,000 deep directory structures. You might want to stack whatever why.

I

Would you blow the stack in that case, it's a cursive implementation, that's what you say! Yeah! That's.

B

Because they use were here, oh.

C

Because he yeah, because you wouldn't need to use recursion, you could just do filed up copy yeah. Well, you need to remember.

B

Which direct to take? Oh.

C

Yeah fair, you.

B

Need the you, if you cared about blowing up the stack, you would change the list here, the directory in the merchant. If you also we're going to help your directory structures not like.

E

B

As remounted itself in the pathway, yeah I mean.

E

At that point, you're just gonna go forever. Anyway, it's will your honor this space yeah. So.

B

I, don't like have inclined to say new API review. Fine. These are reviewed. No. There.

E

There are too many unanswered questions for feature. I agree, yeah,.

D

F

Know it seems like the implementation will be potentially very platform-specific as well. If we do get into permissions.

D

A

I, don't have a strong opinion. I would have a slight preference with things like this being a just a separate library. Then it can have a bit more policy. It can, basically, you know, implement xcopy and then being done properly with progress, reporting and cancellations and kind of a having.

F

Like a holistic library feature, our Microsoft Visual, Basic library does have this method.

B

It's public, inaudible API. We don't need none of them. I think.

D

It would be one it's not cross-platform and I think it depends on the window, shell, so I, don't think you can call it, but.

B

Yeah so yeah hi I agree with Christophe. It would be better.

A

D

Right this one will go offline. Let me do all other stuff, all right things, people online and unfortunately,.