.Net Foundation Design Reviews, 12 Dec 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Reviews: GitHub Triage

Description

* 10:00 New APIs in SyndicationFeedFormatter
https://github.com/dotnet/corefx/issues/24668
* 10:30 Add a BCL tensor API to represent multi-dimensional data for machine learning
https://github.com/dotnet/corefx/issues/25779
* General review of issues labelled as api-ready-for-review
https://github.com/dotnet/corefx/issues?q=is%3Aopen+is%3Aissue+label%3Aapi-ready-for-review

A

All right we're another wise all right. So, let's start with this one here.

B

The try to add new properties in faithful matter or passing a time string. We don't know so right now. If the formatter see a day time string there, we don't know they only throw exceptions.

C

So problem is: we have two different syndication passes once for adam 1.0 x, for RSS 2.0 and those specs have reasonably clearly defined date-time format sutra supposed to be used, but the Internet's rate of wild west. On these things, lots people don't use those formats and use significantly different ones, sometimes and problem that customers had is with the original syndication feed or matter on full frame work. We just throw an exception when we hit something which doesn't pass, how it's supposed to pass and we can't possibly code up every single variant of every single date.

C

Time format, let mess all my hit so we're doing a best effort and then- and so we've added a few more formats to mrsa I believe. um But we've also been saying, and if that doesn't fit your bill, here's an API to and where you can plug in your own passer. It makes sense. So what.

D

Are those three strings that the parser takes and.

E

So there's version that the delegates down there, so we have two options: either we're gonna do the Fox or we gonna do the delegates as the bottom version. Oh okay,.

D

So it's date/time string, local.

C

Name, local name is local, it is, this is XML, so the first parameter is this is the literal string which appear.

F

C

Next one is, this: is the name of the XML element pound, which was around that string, and then the namespace is what was the name space, but in the wider limit around the string matters, you would apparently big.

D

C

And so custom might say because, like a tamaño is acceptable and I mean a extranet, especially like in that, when we're passing it, we want to be able to say this is name a namespace. Their customers can either say well, we'll just use default, one which is in there all they can do.

D

Is Adina those like Justin next immediate element, or sometimes people need to whole chain up to the root.

G

To be able to tell where they are ended up, like other attributes of people who care about stuff like that, um shouldn't be honest.

D

So so my second son I understand what the parameters are. Second one: why is it? Why did you decide to choose delegates, instead of just having a virtual method that you can override.

C

um Because this would allow someone to take the existing code and do a small change to be able to support it, whereas if you're gonna make it virtual so on has to derive the existing class and then connect. So it's amount of code, boilerplate codes or not to write. Just we have to do this, but it's so rare that you people.

D

I assume need to do it. Do we care about the few.

G

More lines of code, what happens if you need to override both a little different ways like if you have library a giving you a way to parse your eyes and library, be giving you a way to parse dates like do you now have to create a subclass type that can delegate to both of them independently.

D

Or we have two different individual methods, one for.

G

You know you are eyes one for daytime's all right, but I guess what I was saying is if the implementation is provided by two different people, you now have to write glue, whereas if this were a delegate, you could just hook up both methods independently. Well.

D

The chance of us you have ingested is so low that you have to write some glue anyway.

A

C

Your underlying.

A

C

A

Is it is a literary.

C

Xml reader or um yeah, isn't it we use x, maori demand, we're passing service, is not a serialized hoped it's actually reading the elements and full state tree.

D

So the only difference in my opinion so in my opinion, the differences between delegate based approach and the inheritances.

D

Inheritance is really Nair in my opinion, like it, you know it's. This is what excess like it's an abstract class and extensibility means I, operate some method and implement my custom logic, but if you don't have access to the code that instantiates the phone.

F

D

You kind of cannot use inheritance because then you need the I system to actually inject your own subtype right. Do you know you think common scenarios, people actually instantiate these for matters or it's? There is a g8 it for them.

C

They will instantiate it um so you're, seeing so your syndication feed format, we've got an RSS OH, syndication feed, formatter and then Adam won Oh syndication feed for matter. So, while this is appearing on the base, abstract I/o event, people don't ever generally because people don't drive this type since asically ever we it's certainly a base type for commonality. How do you instantiate one of your student.

H

You know one of the concrete ones you're just talking about um yeah or you get a factory.

C

Method to get it washed, it doesn't also currently use the syndication feed house which will generate a formatter for you. Okay, so you're not mooing it up directly.

D

So I honestly like to me, you know if the ritual method is kind of more naturally says it's an accessibility point.

D

Moreover, if you had the virtual method and then in the future, let's say you need more information to be able to parse successfully. Well, you can add some protected members that people can fish out from the formatter and kind of figure out where they are and what they have the moment. You add this delegate and you need additional data to be able to parse. You need to add a new delegate and, and then you have to decide which one you call and then yeah the other question.

H

I had is you said you added a few more for parsers and to these I assume the concrete types you're not going through this delegate. To do that, are you you're? You haven't yet another mechanism for adding parser up to the base type, so you maybe I misheard you, but I thought. Usually one of the things you said is you added a few.

C

More formats that you can parse, then I think now we have. This is the abstract type and we have two concrete types right and that's what people use like people don't generally derive this I.

H

Understand that part but I, didn't you also show you added a few couple couple more common parsers. So people don't have to use this extensibility point at all and.

C

We added support for a couple more mint state formats so like if it comes in an ISO 8601 formats instead of okay. So that's in the core implementation of the parsers you're not going.

H

Through this extension pointers when I'm getting it- yes, we do you to use, as example, is that how do you? How do you allow them to strings different date, parsers, together than um we don't.

B

Have different date, parts that we have one day per survey, that parts that came past different date format so.

H

They have to now parse all the different formats. Not just one type is it. Is it gonna be coming at more of one date time we're gonna evolve an item of your default.

B

One, that's your question basically right so this, so this syndication people matter has two child classes and those classes, but.

H

There's no equipment to call base stats in north state. If there's, if you want the default parsing as well, and.

C

But someone says this is a property someone wants to do that. They may capture existing one and change out later.

D

Yeah, so to change them, the parsers need to be able to tell whether they successfully parsed or not correct. How do they do this? How does this one says, I tried in a karatbars go up yeah if noggin, so this one can be on it. This is the last one yeah.

D

So that's another thing you could consider like this like yeah. If we you know staying in delegates, should it return then say or like nollie bool of you know, result and say, like I was a bar, so that's kind of changing. If you have more of these routines written.

C

By somebody else we did rap and we'd have to make a result and out which would mean have cheese named delegate. You can't have a funk shoe, which.

A

I mean I think, generally speaking, the problem of the fund is you know if you see signature like this, like you have no idea things.

C

Like Yale, which makers just try to with rest of the framework and a couple of places, learning new existed like.

H

C

Not a brush example.

C

Guideline was, these are how things that are future. Sir, is not a question with living uncle delegate and then.

A

One other thing you could do, which may actually make it already simpler- is that, instead of passing local name in the especi or I as literally two strings, you could pass in like an unknown next time, a qualified name or something that you get to lube either already. That already has them is because then it's kind of clear as the tie, because you see like string, comma XML, qualified name comer date, I'm all set, whether it's pretty clear that the string is the thing you're supposed to parse right versus.

A

If you have like three strings like I, don't know whether it's the first, the second or the third or like yeah.

H

Do we think it when you're parsing one of these features generally? Is there more than one type of date/time Porsche just needed.

C

Potentially, you can have a mix of atom and RSS elements and a single education document so.

H

People who use this then should always captured a woman the previous delegated, before they actually override it. Just generally, not something we people do when they on when we have sort of a pattern where there's a property they're, just not Luton I, don't know we get a common pattern where we wrap t, we grab the old one, and then we wrap it with the new one. That's kind of sounds like there's: gonna put a failure there for people.

H

D

Do it automatically well.

A

Or we have things like HTTP client, where you have the message Handler and then the magician doesn't attach.

F

Generally run Mei Mei.

D

D

So solved by just adding a little method that people can override it makes it super crazy. You can name parameters, you can access more data in the future. That was like that you know state of the parcel, while it's running or the formatter, the.

H

Only challenge for the face type is and seems like most of these poor manager created via factories. So how do you go and create wonders? Well, so that's.

C

Where it was yeah yeah, it relates one of those and he said.

I

That they use it, you know about.

C

You can I have a user calculator, but there is a this. Is that form on top is also a syndication feed class, where you hope it usage ICN. Can you somehow tell the feed class to producing?

C

No, no, you have to derive bat and then override the classes. That event is it common that people use? Does it feed factories.

I

H

But with respect, that's the most common way, that's what I was thinking, because so they don't want to choose whether or not to add a mirage. It probably chooses for your vision. The end point yeah.

C

So so, if you use there's a load method on syndication feet where you passed at the XML and it goes ahead and looks at the beginning, opening XML element and decide. What's that, where it's going to be a term or RSS and return to an access to a feed format or an atom feed formatter. Another.

A

Option we have is instead of having like those things virtual on the formatter itself. We could have a new thing called an arsenic, ation, feed, parser or something where those guys are virtual methods, and you can set that whole guy. And then this way you can just do the normal delegation right.

D

Okay, yeah and then you we could design a like. You know an easy way to change those people yeah, basically multiple of those in an array, and we just work with the chain. Yep.

I

C

One way of doing it so Shin do member when you use the load on syndication feed, whether its pre passes, where it's consumed as a whole document for returns it to you.

B

Sorry so you do syndication.

C

Free download mm-hmm and then it returns you a class driver's education, beautiful matter, yeah you when you get a class back, has it already consumed no annals I've assumed it yes right. Okay, so could have a property. So we need a property event to switch it out, because we can do it on a constructor on the thing also be touching. Lots of API is like the load API on syndication feed.

H

Ya, the Lord, a PID, looks like it actually creates and Adam Peter, and then it calls can read it passing a maximum reader and then, if it can't read it goes and tries, it creates an RSS two 1s certificate by.

D

The way I don't feel strongly about any of the recommendations, but like just to let you know like what are the traders in this space I think the delegate based extensibility API is kind of not as easy to involve going forward. Like you have this delegate, this is the signature. This is the property and if you want something more in the future, you need.

I

Over any rapper.

D

T's and then you have a problem which one is called in, which order and like can't.

H

Overload importantly,.

C

Have a syndication feed or produce one. So if you derive it, then you can't you can't use at T after and fight your own syndication page well,.

D

So that's what you are saying to work around this problem. You can cut the basically a type called parser and.

D

You basically implement the parcel abstraction and then you add it to the property. So that helps you a lot. It doesn't solve all the problems, but it does hurt a bit because you, you know it's easier to change them. You can change them. Actually, you can add any ice to the parser. Like you can add more. Let's say: let's say in the future: this I'd shoot. It would be so nice to know the name of some element higher up, not just those two three strings.

D

Well, you can add the fields to this parser thing and set it right before you call it. Each.

I

C

Input your wavy parser works. You'd have to do a massive redesign. We have to get and function. It just worries about its current locality and the external document that.

D

Was just an example and I'm saying you can you can basically add properties and fields and whatnot to the parser? You can add new methods to the parser versus once you expose this delegate. That's your matter. You.

H

Can add that in there there's nothing wrong with you actually considering this an event for this as well, especially the parchment? If your default parsing didn't worked in the environment same force to date, then you get to handle or can have the extra pieces of information.

C

You need adjusting is bad idea, trying the default pacifist and only if it doesn't work cooling, because you could have a failure mode wherever Parsa is incorrectly passing but believes it's correct. It succeeded. How would.

H

You handle that in a delicate case, because you're still going to run want to call the base one I'm guessing and you can decide if too cold if base one first or if you first execute your logic right. So yeah.

E

So well the inside, but you can decide in your code in your own delegate, I decide like first I'm gonna, try parse and if it fails, I'm gonna call the my base or whatever was there before. Oh so.

J

You called the delegate first, we just come delegate, and so it becomes.

C

Mostly, like I said, you see it's.

J

Up to you as a customer's an extender to do the caning right.

D

Yeah I thought we were before discussing it, and somebody says that first, the building code runs and then their delegate at the end. Now, if it's vahana.

E

D

Okay, it's always going to addition. It's the whole thing.

E

I'm not worried about much about the extensibility of the delegates and I think it's per day. You know advanced scenario anyway, but still has good points that you know it's more. You know future extensibility and, like you know, providing more information or adding one more parser and things that we look our selves quite a lot with this butter and.

A

One other option we have is instead of having visually normal property.

K

On it, we could.

A

Just make it a collection, majors can just call add, and the default one is in there and.

C

It's the one it's bill, allow you to reorder appreciation so well, it's a collection where you can order in any way. You want didn't need you that without its delegates are multi delegates by default, you can yeah, but those are really hard to control the order, because there's no BSC, basically.

A

To array them and then somehow get a gotta get back that is they're ready, like it seems a bit user word.

K

You have to chain on your cell phone yeah, the Coalition for Angus.

I

L

Mean if it were going on that path, apologists hold it together into one abstraction and then have an array of that abstraction in.

G

That handles all parsing things you have keep in mind with the complexity that you just mentioned is only if you want to inject yourself in the middle of a delicate chain, adding yourself to the beginning or the end of online account yeah.

G

How do you add yourself, delegate dot, combining put your delegate first and then the existing delegate is a second frame, because you have to had.

H

So, switching gears a little curious about the first sentence, which is you're playing the a district or only not standing or that stuff murmurs there. That reason, oh.

B

We don't happen to tear this property Festival hopeful. Now we yeah.

E

But you are open to eventually port it to to desktop if there's a vehicle and everything right. Yes, there's nothing preventing us now.

A

Generally speaking, like the one thing we don't want to do, is you don't want to diverge all right? Like I mean if you will diverge in the sense that you know, we cannot version everything simultaneously, but we don't want to basically a fork in the road where a framer goes down. One path and core goes on another path and you have the same teacher in just different API shades, so that would be bad.

A

So basically that means whatever we design. We have to be able to say: oh, you could port it back to that. Stop. If we don't do something that we know wouldn't work right, yeah, yeah,.

A

J

Those are the knows.

A

That I have them miss anything.

J

C

Agree on absolutely ground, first one, it's just not! It was just trying to go with convention, not.

E

Everything is such a guidelines yet right, you know, yeah.

C

E

Knitting just the corner cases. Yes, so.

C

That as part of what this is about, I, like the X mollified name idea, should think about I couldn't find just couldn't remember the name of class which combines them I. Think that's in there, it may I may have a moment. You could just look at the video with the video consider like whatever that thing is the.

A

C

Gives you separate um there's two properties which is modality.

F

And the name and.

C

Any API is on XML reader, take a namespace in the local name. If I went looking.

A

But whereas indication fee for me to live on core assembly wise.

B

Systems model does because.

L

Already depends on everything else, both right.

A

So that would mean we should be able to be able to take dependency on that thing and comes from XML.

C

Our last hope you, like our mystery XML reader world, not be X document world well,.

A

I think XML qualified name is the is the old XML Dom API there's also an X and an X document one, but that thing is called X name and that you can convert from screen directly to annex name excellently slightly nicer than X of a qualified name, but I thought the next time qualified entity. Part of the other video AP, is, if it's not yet just take. Whatever feels natural.

A

If you have any way to express your concerns with the extensibility I think I just mentioned so.

D

Another virtual.

A

Method yeah, as I said, I mean.

D

So I will definitely not do that generic France I will either do the delegates, the you know, specific delegates or the base type you just inherit from a base type and you add a bunch of those to an array or a list, and then you call them repeatedly. They have met up, try ours. So if you cannot Paris you just return false, then you, you call them in a loop.

D

At the end, you can throw the same exception every single time, for example, one additional problem that the delegate solution has since at the end of the delegate, its user code Rani, they can be throwing random exceptions and then higher higher up code will be getting random exceptions depending who who wrote the code.

I

C

D

C

M, ah it's a passive throws. We record that and then, when it comes to actually accessing the property where it's been set, but we we've sort of a pointer is easy.

C

Because another thing that problem would be beautiful matters being, if you have an exception start, please just fail. Ops and, like you, have one date, which is bad. You can't process your old documents yeah so or do you think about hazing scindia? um Yes, so one thing about having named delegates with a boolean response and a nap, um let's leave I said it's easy to add yourself to the end or to the beginning of a chain we're a delegate.

L

A

Along that, you guys have to actually call a guy, then also in that way right. Otherwise, you would just call the first one and then nothing happens right. You can just create an internal help, her we just hold it higher and then they said the contract would be you take the first. It easy yeah.

A

Yeah, then that would work what.

G

Would would be the type again.

C

All delegates, a multi-car surrogate or not so someone can either replace it or if I want to do something, I will do combined income, sir, so combining you just say, plus.

E

C

To my new delegate and it just good I.

A

Just consider that I'm not sure that people get that idea very quickly, like they're, usually dead, which is 2 plus equals on whatever the thing is, and if you make it any vendor, you can only do plus equals if you make it a property that is of the type delegate that we just set. The thing replace the thing: that's in there yeah and then so. The question is that can hold you communicate to your customers that they probably want to chain it, because I probably want to use your default parsers. All these.

C

Product changes are going to do we're going to need documentation for each properties, so you put it in documentation with sample code showing how to do it, and no mr. Bob staff we've used at unit tests as best practice sample codes to provide at our customer. So if there's any question, we just point a mare and say now: this I should do it.

E

Though you will do prevent questions, I will do insert it before the questions arise. Right.

L

C

A not a documentation.

C

C

Really don't think we would have scenario well, someone want to insert in the middle because they would get a feed format with just one on it to begin with, and so pick up their own things they want to put in to it just would be a binary plus equals you can order them perfectly. How? If you want, without doing anything, tricky.

A

We just write the thing they don't agree with your autumn would say: is that people don't tend to read documentation for they use an API. They try to make sense of it by themselves and if I had to guess they probably have Bach summer in their RSS see today they write their processing code that addresses the one bug they run the thing again it works, and then three months later they realize oh I, regret something else, because I can't parse the other formats.

A

But it's not necessarily obvious when you replace the build in one that you just didn't chain it correctly right, yeah receiving a virtual method, it's kind of natural that oh I didn't call the base. I mean it's kind of making the normal way of doing things right. So this is the Galligan list like most like that they are multicast to begin with,.

C

So what do you think about if they, if there's only one set on there and it's their own one, we have an implicit run outlast if, as Affairs fails, I.

A

Think that would work. The only problem of that is that, if you do this way, you basically give up the control and the customer site to replace prepend rep and which may be okay.

C

Why it's only if they don't run ours, but we would hold. You know that um so we're going through them and we've got a static. This is our default one. So we just comparing. Is this delegate this? Yes, then we never we've run out you for one.

A

But when a men is like, you can have code logic along light, so like if it's not your one, when your cache one first and then one customer one, but it means the customer can't control with it. Once.

C

Before after they're not saying if, if our one is missing from the multicast delegate right and we run the customer, we run whatever customers got, I am and if it doesn't succeed in passing, we always try ours. It was.

A

A full row- people, wonderful yeah,.

C

Use it as a fullback if they haven't.

A

Got it already in their ordering I guess what I'm trying to say is like I understand that problem saying is that it just means that the customer doesn't have control with the yours once before or after all, right, oh yeah, because they chose to like you know over.

C

I sang and they really want to.

A

Run if you want to.

C

Announce first, they.

A

Just do it and.

C

Then you know your rent and then you don't have to do it again. Yes,.

A

C

Is if I want to have as run first to make and do an append of theirs after house and if they want base to run first, we can do a prepend or just a straight-out, replace printer. It works because yeah once we're done with everything we've specified and if it still has a path, there's no harm in saying much one hours just to net.

C

On your end, we're trying.

E

To kind of compensate.

E

But if you have the virtuals does it's where we have troubles actually with multiple read? Are multiple people? Try and do you know, add custom parsers evelle right just they have to know about each other, they go ad and they.

D

Are called in order so.

E

The virtuals were 200, you you, here's my parser yeah and one more there's a list, but each.

D

Of them will call the base level is no calling a base. This first first parser in the list is their parser, and then you can rearrange the list. However, you.

E

Would have like.

D

Rule of rapper of do list reppin a list. Basically now there would be a property called parsers and it returns the list of parser or URI parser and you can keep adding stuff to it.

E

Okay, the argument with the virtuals, like you know, having a clause that you can inherit in one of the key values holes that you can call your base right as well, which is the default parser.

D

Well, if we go with so initially, we were talking about just coming virtual method on the education or matter yes in the party in the collection of Arthur's thing. There's no reason for you to call by as it has it's a pure abstract class from the base. Behavior is just the first item in the list, so.

E

They're, basically making rich delegates, you know: vid applause, kinda,.

C

But then Christian is like it's just redoing emoticons delegate functionality, but making the API self is slightly more obvious.

A

Yeah yeah that, but that's exactly what it does I mean to detect the fact that delegates are multicast like I, wouldn't, say people get that necessarily.

C

Like that wants point, emoticons delicates.

A

Well, you could argue that the whole design is not the best one which I would agree with, but like that, that's what we have I felt like the question is given an API surface like can people look at it and know what they're expected to do all right and the question is you know if I see a property that is a getter and setter like I?

A

Usually don't assume that a you know mess with the state, and you know insert myself in the middle necessarily right, because I usually think of Delegates the things you said not as things you. You know inject yourself somewhere into it right, like usually for events, it's kind of clear, because you have no choice, there's no replacement, but you just you just do plus equals or minus equals. But for settable delegates it's like what you just replace the whole thing. A little.

D

Bit to me, as I said, let me try to explain one of it like the to me that abstract parcel. The main advantage is that you can add more information if you like in the future, rediscovering that to parse a date, and we need some additional piece of information. You cannot do it with that delegate those three strings. Each person is all you can pass it with an abstract class. You can add additional virtual member that has more parameters, notes and.

M

D

Dynamically check whether the customer gave you the you know the new advanced parser and you can pass more information to them. Okay, so.

C

Well or you have a context, object which you can version over time, it's also Razzles can say, is combined to stick without it. The one who promises be an interface which has some properties on to get this from it, and then that can be version to add more stuff. Do yeah I mean I just want to avoid so.

C

Yep and a large problem people have is having too much boilerplate code to be able to use any extensibility points, but, like its biggest that's the biggest issue that you see with the and I want to avoid going down that path. And so you don't want to make it like 15 fights through I call. It part yeah, yeah,.

A

C

Your partner should just be like you know, it's generally going to be like two lines of code. You know it's going to be a date/time tripod, pass exact with some sort of string and saying you have to create a class and which draws from someone and override this property and the rest method, just too heavy weight for something which is really a one line at one. Second, but I, guess the.

A

Way I would describe. It is like well I think when you're after is the same thing we're after, which is you want simplicity right, yes,.

F

It is necessarily result.

A

Of how many api's you have, because you know see API is a good example of something we usually have one method that takes 15 arguments, and you know this is very tricky calling convention where, if this is this- and this is that this happens or this so like you have this simple.

A

Water is super complicated, right and I'm. Feeling like this is going down a very similar path where you say well in order to chain there's this very specific way to do it, which is not necessarily very obvious right versus another design, may have let's say twice. The number of API is because it's two methods or it's you know you know one metadata type, but it may lead itself to like completely obvious Co. Did you have to write because it just fits in with your mental model? It's right, I mean I, don't think we have the.

A

You know perfect answer right now, but I think the idea is that that that's lazy, something you should think about and say, like you know, you know, start inviting some temper code against that and like try to figure out for yourself. Like, oh, is this the code I would write the first time and the answer is most likely. No, but we have their own API. It's right, I'm, like that's. It's a thing kind of like the the bar from from the simplicity standpoint.

A

um So but I totally get the point of, like you know, the more data you push in, like the more objects you have to allocate like parsing, is also more expensive. More all that so the.

C

Attili is a trade-off somewhere, but I've got no problem with us having more classes interfaces as customers, so she'd only have to write the amount of code needed to get a job done and in this case no, but it's a one-liner to write a parser. So what about having a? No because you're right back, we multicast delegates are non-intuitive right. How do you use.

E

C

E

Is a community recommendations? So what about a combination of property date parser and a default syndication late, parser with virtual methods, so I think the virtual methods be killed because there are cases when you cannot override right. Why not events, then you have both Doggett's and can override if need it. Why.

K

Not what sigh, why not events, then she can't really override if you're already before parsers already so I think.

A

We events is, you can only attend like there's no pretense, we.

E

Can inherit from default state parser and inject the class into property yeah.

A

That's what I said originally that it seems to me like the the closest thing for something that gives you both the.

E

Public event and sorry public event handled it's a parser and predicted virtual on dates, time, parse and like how you can inject a custom XML urs over into XML Reader.

A

In passing it into the the instance of the syndication seems vomiter no core understands if it's firm, yeah.

C

So I think requiring it to be a virtual, something which one overrides is a no-go, because syndication feed in sunshades fees and so you've got no control over what types being associated um I think there's some strong merit to having a abstract class or an interface which has table keys on it, which people can actually well. So you could do just a.

D

Simple start forget about the you know, interface or abstract class. Just have a simple strand to date. Has those three fields: struct Spanish, oh you're, advocating avatars, but what basically yeah exactly- and that is the reason we have developed. What do you mean starts own version.

C

If you are passing it to us, if you get a new version, it being passed to an existing filed library van is my problems with. If you don't inherit right, you know yeah well, it would you can't inherit worse trucks and so versioning. You can't create a new version which inherits from that you can. You can.

G

Just add a new field to our struct at the side view it's not like win32, where it breaks like the jitter makes everything work yeah. We don't have sizes.

D

The size of structs is not a explain. You contract I'm, very it's not part of the wider contract. We're.

I

Not going to prom move ready to run stuff and gent.

D

But forgiven version of the framework, it's like.

A

D

Change if you change code, you read, you have problems who sent it so like unless you don't change anything in the library, yes, that you don't otherwise your engine and then joining this one, and you have a difference.

E

Was to determine its version resilient, so it's gonna, you know, make sure that takes the right version, size of the classes and structs it. Okay,.

D

E

Had we had fields so starts all the time: okay, I never break customers, so maybe for next time like if you guys can look at the option of the you know of the class or struct or something and let's look at the toast proof, two proposals how they look, how ugly you know, how do you search and look inside that? Do we want.

J

To like do another session, or we just agree like what a leg of my my preference at this point.

C

Repose Slezak would be keep the idea of a name delegate but turn it into a list, because multicast delegates are non-intuitive and take out those parameters, all the things being passed in and have I've instruct or an interface being passed where then they can go and retrieve her from it, which gives a versioning capability and doesn't Mme delegates definition is stable, even if you had stuff so I like the struct idea. I'm, not.

D

Sure about delegates and a list, because it's a weird combination like we either do delegates or we do a list of like you, know, parsers. We don't because it would be a third party like we basically said hey, we don't like how delegates saying, let's invent a new way of changing delegates, let's have a list of those and then we call them like yeah, it's Yeti like it's a new thing in that net. We don't do it. We either kind of accept the negatives of Delegates or we do something else like Alice parsers, when.

A

You could have the single type called like. You know, parser right, that has two methods on it right, like the parts day time. The parts other thing and you set this as a property on the syndication feed right people would just derive from that guy add their own stuff on it, and so you get the chaining by you override the parser. You call base at the time you want I'll still come by, it doesn't damage multiple providers. Right is the next.

E

Scenario, oh, it does.

A

Because, basically, you derive from the from this on the type of you. If you want to chain in five other things, it works like any other abstraction where you, where you basically taking the other guys, have them as field, and you call them, and you want to call them right like it's exactly as any other reppin abstractions. It.

E

Still compact, let's.

A

E

But in prepping obstructions I have to know everyone know you.

A

Can just say, oh sorry, perfect in inheriting that's what I meant? Yes, but there's only one you hear it from right and, like the one gives you the default default right and you can control whether you want to call the default or not by calling base or don't call base right. If you want to compose with five other things. Yes, you write a generic type which really that's.

E

Weird that each of them they'll probably call the default right. So that's what kind of breaks out if you decided I, don't want to call my default, but Ivor I'm wrapping someone who decides to call the default like you're, fine.

C

Because it just needs a normal chain right, you based on something. So if you've been driving, something which isn't the default, so you got so you got a base class and you've derived, but to have a new one events on derive that they call base. They.

E

Have to know about each other at build time right. You know I'm more talking about, like you have like two libraries. Each of them inherited good, like you know something, and they will do both compotes.

C

Also is still contracted me or some in which the implementation will only be like one line of code right. How much code does someone have to write to be output to execute out one line code, and this solutions is true, although, like you know, this is why we.

A

Really like it speed up my heart to judge for sample code I I'm somewhat suspicious that you know if you look at even two inches like like this like this will be not much thing. It is actually a one line of code like it needs to be like at least some if checks like it probably is like on the order of let's say five to ten lines of code.

C

You know I, don't watch your other Porsche.

A

C

You should totally kept.

A

Your catch an exception.

C

To something in like in to capture the existing pasta, I could do no capture exist, Bastian to tempura grab a one-line code. Then you have a closure to find resetting it to equals, and then you do your one like now. Yeah.

A

Both the exception block with the delegation to the other one right like I, mean it's pretty clear that we can use some sample code here now. Yeah I'm willing to get like it's a like. It's probably not 100 lines of code, but it's probably not one line of code like it will be multiple lines at which point the question is going from five to eight yeah seems like not such a big deal going from one to five. Okay, that's.

D

A bigger jump back from five to eight right, I think if the parsers in real life were just one-liners, then I I would say yeah. Your argument is strong, but I'm. Also with you mind a bit skeptical did the harvester for a day. Time would be a one-liner if.

C

You go by what the spec says. It is a one-liner answer, only not one-liner, because people deviate from the spec exactly.

D

But if you, if defeat, actually has a spec compliant data, and then you don't you to do anything your parts, yes spec man, okay! Well, then, this accessibility point is for like wacky cases and then it's is it ever gonna be one-liner the wacky case no.

C

D

Sir itself, right it.

C

Could be like song would have and I so basics, one variant where we don't put team if it's based in I've seen that in some and that's just one liner should just be a format string for a lot of them. No most of them are going to be, they are still a standard, but we're just a different standard and a difference, and it is going to be a one liner because it should fit into our date time passing format, strings, I,.

D

See so you've done you only ready for to you. First, a subset of.

C

They died, there are many different daytime formats, as Jason ones is ISO. 8601 is up three different RS versions, and we do a bunch know we do the ones which are defined of a standard and a few couple of extra variants of it, but we can't come up with every single one. There is.

D

So which ones that are built into the VCO? You don't attempt to parse.

C

We usually built into the vclt Michel.

F

C

Can do HH : mmm now, like you, can, is flow.

F

For my line for about super yeah, you know, but.

C

Business and its science and.

A

C

It's presumably.

A

A

Informants during light, which which that and the operation itself is a one-liner yes, but like I, said, the pollen is not that I had to Paulo mr. policy around. You know. Okay, this is my pricing code. When do I dispatch to the built in one we tried before to a trial after that gives it an if check with a pool with an outdoor. Is it, like? You know a try-catch and like that's always on the order of like you know, when.

C

We switch it to agree, we should switch it to having a boolean response code and not use a try/catch. Then you shouldn't. If it got a default one, they shouldn't have a need, a try/catch because we're not going to throw it would just be. You know if try passed exact, otherwise, I call now call be no:1, um but if it's a multicast delegate, then you don't even need to do that. Just to return your boolean true or false, and do your plus E or dot combined all.

A

Right so that may be reasonable. I mean the.

C

Question is what.

A

Happens when you have to like handle these guys forever, then it gets a bit more interesting but yeah. So.

D

I would like to avoid sending them back and like going back and forth so they have to come here again I. Could we just like serve them like some parameters and let them tease the rest? That's.

A

Kind of what I try to do here, like two bases, say: here's the kind of things that you brought up as like trade-offs, right and three I think the the number one thing that I think you know we should have as like sample code, the API or I didn't see. Okay, do we like that outcome? You.

E

Can do it offline if we can no hey? Ideally, we should I would like to see that you know what.

C

Are we comparing should be a way to just update the thread and we have a pull request. We do have power quest with uh actual tests and as well Orvis be just so. I won golf like right. The.

G

Dog just yeah yeah. We.

C

Can play to it to me I'm saying because um after them.

D

C

So we're saying no to be idea of a list of Delegates. How do we feel about multicast delegates, I? Think it's it's.

D

Fine, they have negatives, but you know we have them so and I also think that's like the non-trivial chaining with solder that it is a negative, but most people will just you know like combine in whatever first.

C

People probably use or.

I

Just a place upon.

C

Which men in know, if capturing someone in create a closure, will be in a one question probably chain like that. Rather than that and delegate combined or plus equal.

D

Might bill so I think we should just like either go with standard, multicast, delegate or event extract, or this abstraction in the list of those and I. Think it's up to you just right. You know, update the IPIN samples of every officer that actually does something interesting, I.

L

D

Yeah we, if you, if you can update this today, we can look right to the end of.

D

D

Or through online or through the ether.

A

Alright and I think we're done with this one sweet you're free to stick around or any when you want to do. If you.

K

Eric in years, yep.

M

I can hear you.

A

That should be enough for first pass. Did you want to give some old video? Oh.

M

Yeah sure so, in order to support some machine learning and AI scenarios and I met kind of, though the first, the first.

M

Container of data mostly used for numbers basically, and so a tensor multiply, multi-dimensional data is like you know, one dimension is a vector or a list kind of thing. Two dimensions is a matrix. Three dimensions is a cube. Tensors is the term for the abstraction. Over top of you know, n dimensions of this data, and so the idea is, is like all these machine learning libraries kind of have their own types, for this is what I think is a matrix.

M

This is what I think is a matrix and in order to like Interop between, say things like math, dotnet and c and d k, people have to copy the data from one format to the other, so the.

F

M

To make a b c l type that that people can use, that's like an exchange type that everybody kind of agrees on, and so we don't have to do this copying around. It's also tuned for interrupt with natives scenarios as well. So if you have a native library that takes matrix data, you can you can interrupt tensor data with it without doing the copying.

M

Similarly, if you have a native library that returns native memory, that has the the matrix data in it, you can make a tensor that wraps over top of that native memory. Without doing the copying a.

A

Huge part because I think the other part that it's worth mentioning is that these tensors tend to be large, like we don't talk about arrays that have like maybe a few thousand element to talk about things that have a few thousand elements per dimension times like 25 dimensions, or something like that right, and so these things tend to be. You know not. You know a few hundred bytes, it didn't write.

M

Like a decent example to think about is is an image right and if you have a hundred by a hundred pixel image, so you know 100 by 100, then every you know red-green-blue is is a third dimension in there right. So it's a hundred times 100 times three and now, if you want to do like video machine learning on videos right, you have one of these images every every so often right.

M

So if you, if you have a video, that's you know a minute long and you take a picture every every two seconds or something you know it. Things multiply really quickly in this world.

E

Just to me, like specific interpretation on top of Memorial Dean,.

M

It's similar to memory of tea, but it's got its got. Higher-Level, interpret specific interpretation of Memorial tea like so another. Another thing that I failed to mention in the in the preamble is we have dense representations of tensors so which is kind of what we've been talking about right. We have an array and every element in the tensor takes up a memory slot, but then we also have sparse representation of tensors.

M

So again, if you go back to the image- and it's like you know, 75% of this image is all white, but then I only have drawing on 25% of it. You can have a sparse representation that allows you to not have all the quote-unquote zero values represented in memory. You only represent what actually has values.

M

So that's another kind of dimension here, the.

E

Four models that you know is there some industry for me, yep.

M

Did you hear the question you're back yep, but there is. There is a handful of industry format, sparse formats we've chosen to implement kind of to start. One is like a super industry format called orchids, press compressed, sparse column and then the other one is is a dictionary based, sparse.

E

We have option to extent in future tumult to other formats. Is that totally yep.

F

That is exactly correct, so I have a question if.

D

D

um Reverse tried, the default, I assume is false, but it says that CNT K is using true.

D

Shouldn't we default to what CLT, Oh ones, sorry.

M

I heard I have a question.

D

So I was asking about the reverse: tried parameter to Destin, sir.

D

That's true: should we reverse the default so.

M

What reverse stride really means is, is column-major like, but the reason we don't call column-major is because the word column kind of implies to two dimensions right, so we're reverse stride is, is a good thing to talk about like whether this is a good name or not? Well, we can defer that for now, but the default of it is C C style. Arrays are row-major right, so dotnet arrays are real, major and so like. If it feels.

F

Natural that in.net, the.

M

Default to B is the same as the array, because, if you look through some of the api's there's like take an array and two tensor it, and so, if if the default would be that you wanted reverse, try true call a major. We couldn't just take an array and then two tensor it either. You would have to you'd already.

F

Have to have that array.

M

And call a major format which you probably don't have or we'd, have to copy the data out and do the formatting yeah I guess that.

A

Makes sense the meta feedback would be. We should obviously choose or defaults in such a way that common operations do not require crazy gymnastics right either for Interop of the majority of the libraries or you know, for normal data structures like a race right, but there's clearly some tension because there's not a single format to rule them all, so we. So. This is something that we need to probably like. Look at and I compare and like look at how many operations or copies or translations they would meet in common cases right.

M

So I mean kind of just reiterate that net you know C style arrays are our real major Python and dra defaults to row major.

M

D

Makes sense it ends it.

D

Dotnet erase not C n TK and makes it.

L

All right good, the only thing that he said really good so.

M

I put two two scenarios here. The first usage scenario is is basically the some code that we presented at connect. If you, if you watch any of the connect keynotes there was, though the amnesty, you know you draw a digit and then this thing tells you what digit you drew. So this is the code that takes a bitmap, basically the drawing on the screen and then turns it into tensor data and so kind of the.

M

The first lines is getting the width and the height of what what you want the image to be, and then you resize that image. So that's kind of the first three lines aren't really the important stuff.

M

The first line, that's important is the new dense tensor of floats that takes in the dimensions you want, and then this optional boolean, that's whether you want reversed ride or not, basically a row row major or a column, major and then you know, start looping over width and height and getting the pixels for each color, converting them into a grayscale and then putting them into the tensor and then returning the tensor, and so that comment about the turning.

M

The black background into the white digit- that's just because the image comes in with black writing on a white background, but them that the learning model is. You know a black background with the white writing. So that's just also kind of more details. That's not important to, though, to the API.

M

The second scenario is something that I played around with there's: there's old libraries like law pack or linear algebra pack, though it's an old library that was written originally written in like Fortran and then eventually kind of modernized in the 90s, and so here it's taking two to dense, tensors, a and B and then solving for x in the equation. That's in the summary so a times X equals B and solve for X.

M

So it's basically B divided by a, but you don't really divide matrixes matrices, so so anyway, what this is showing is being able to take a sensor and then Interop it with with existing native code.

M

um You know some errors checking at the top that says you know this must be a matrix. It must be square. The number of rows and B has to match the number of rows in a and then because, because lapack will mutate the values in place, you actually do need to make a copy in this scenario, because the idea being the caller doesn't want a and B to change, and so in this exact scenario, but other.

F

Native libraries.

M

Might have, you know, might not might return native lat made of memory, so you don't need to make the copies but anyways. In this case we do many copies for a and B past the pointers for those for those tensors into this law pack, II I, don't know how to say, but it you know, DGS.

G

Look forward to see the reference point two so if ever ever I or for dense tensors, if you, if they're represented, if they're, backed in memory by a multi-dimensional array or similar, then dangerous, get minimal. Reference will probably point to that array itself, but for sparse tensors well, but then return right.

M

Like notice in this case, like the API, is actually tensor, that's right. First for sparse, you would have to have another method right and the native method would have to take a quote: unquote, compatible format of of that sparse. So so concretely, if, if the native, if you want to scroll down, we can look at if you want to go to compress, sparse tensor. What's underneath tensor I think.

A

The general gist is effectively I think we expose the member layout if we don't hide the memory layout from you. So like this API here buffer is specific to against tensor. Like you know, the the other ones have other representations and they give you access whatever their representation ends up being. So when you look at the device like basically, as you saw earlier here, I think there's a class hierarchy involved right, so Dennis tensor derives from tensor right and so that you know you ain't. You basically have to know what you're doing.

A

If you want to. You know efficiently hide.

G

That you know, then.

A

I get that I guess.

G

What my question is coming down to is: do we have API so that looks about right, so cuz, you know some.

G

Reference so the the api's looked the same if we're all on two different types, but it actually does different things under the covers. Like gives you different representations in I think to populate the.

D

Tensor data, so if you populate the tensor, it doesn't matter which kind of tensorial right, correct, I'm, talking for interruption area moments, you doing inter I, think there's no escape like the native code, expects some format and so some other representation.

A

Or then we don't have a single buffer. They have like three different buffers because they were, you know, I, don't know the coefficient inside the the coordinates that I've filled and yeah yeah. Something like that right. um What is a bit weird here is that, like Y, twice clone, why do I have to cast the clone without because.

G

I know double returns, an object yeah but like. If you want method, you don't have to say this by the bullet, because that the abstract tensor clone returns.

M

Tensor of T, so I mean we could new. We could make new clone methods on dense tensor. You know compressed, sparse tensor, think.

A

Makes more sense, because that seems like a bit weird that I have to like I will then want to get back the same thing. Clearly right like that.

M

Those methods so.

E

What a high-level question like in the beginning, you basically say that this is about like making the vendors for their libraries to do that, but it all feels more, like you know, having better Intel and being able to write dope netcode interrupting better with those libraries.

M

Yes, sort of well so, which one is it's kind of its kind of all. It's kind of.

M

Sigelei right now, Lucy.

F

M

Any machine learning types or libraries, but it it's compared to other languages, dotnet, isn't used in the machine learning space.

M

Oh, the proposal is in order to start getting done it more involved in the machine learning space having a type that represents multi-dimensional data. Better multiple dimensional data will hopefully will spark more AI.

E

In machine learning, you would expect that some of those libraries might support their parts or maybe entirely into donut, is the third important I mean.

M

What one example, one example is tension. Tensor flow is a very popular machine, learning library out of Google right now. Right Miguel has this: has this repo that has tensorflow sharp, which is C sharp bindings for tensorflow? And and so it's it's not a rewrite completely of tensorflow. It's it's bindings over top of of tensorflow.

A

Enduro yeah, but I think interrupt is like if you interrupt it's more like what the customer can do by themselves. Bindings is basically pre can't interrupt code written for you, okay, first, so so.

M

That's that's one.

M

Similarly, there's there's many libraries in Python to do to do machine learning, scikit-learn kind of one of them, one of the more popular ones like there is no Interop story from done it to Python, or at least C Python.

M

From what I can tell like I asked some people, everybody told me it doesn't work, and so, in those cases where somebody who would say say they were like playing, maybe they're, investigating and researching in Python, and they decided okay I want to productize this coat, but Python isn't fast enough for me now I want to port this Internet. It's like that is a scenario as well of okay.

M

I have this this stuff in Python that I want to I want to run in my productized app, but I can't call Python for my productize data, so I'm now porting this python code. So I.

A

Mean in general, like I, think like methods like this one here we don't expect most customers to riot like what we expect, customers to basically uses api's like.

E

This are the questions who.

A

E

The high level code and under the verticals in to the low level libraries which are.

A

In different languages,.

E

A

Binome will be providers, we don't provide anything like this and we don't provide a pre conversion to call into a lot back for its outright. So what we provide right now is just you know, the exchange type plus the you know being optimized for doing stuff like this. There is a general discussion right now in the machine learning, space and indepted that you know nicolas and other people that driving is like. How far are we going? We've done that should we have like, should we, as a team, for example, provide a la peg wrapper?

A

Should we write our own? Should we should we should we just provide, like you know, so we just work with other people like you know that do jason nets of machine learning so that they can write those libraries, but the end goal is clearly in the sense that you want customers to be able, to just add a new gear packet reference and make just basically hack away with you know, simple API is that do that and I think productivity.

E

Of c-sharp or never think well, keeping the performance and being able to you know, write.

A

And I think like this is basic, the very first step. So I don't know we don't have anything like you know this today, but eventually they will still have that because you know, if you guys, like this yeah they're, not fun, to call yeah.

M

Exactly and that's why I kind of have both of these scenarios in this proposal is that top scenario is really easy to understand and I'd expect lots of people if they were doing machine learning and Donette to be writing code like that right, nowa, potent SAR fill it up with data, give it something. But yes, like the second scenario, is still important, but not as many people are going to be writing this. We still need to be able to do this. We still need to be able to do scenario two, but.

D

You know like once we had a language pinning support to spans, it will be, it will get simpler and safer. It basically will get to a point where it's well. It depends what's the signature of the native library, if it's nice, it's going to be nice. If it's not nice and well, that's how the native library works and, of course, wrapper improves.

A

A lot but I think the general problem is I mean like what we are going for right now. You know, you know you know Asha and us in particular elective is. We want to democratize the whole machine learning space wait a minute, but in so like the libraries are nice from you know. If you know what the memory layout is, it all makes sense. But if you look at message like this, like like first of all, you have to have like the five-year degree in math, to understand what the this thing even does and then.

A

Secondly, the exits. You know that the deposit conventions are not so straightforward, so I think what this it would. You know what these two lines of code look like, and the grand scheme of things is almost irrelevant right. All your all. Your problems are understanding the space I.

D

Don't know I agree, but what I was saying is what you said would improve upon the native library, whereas I thought we are kind of discussing is the tensor code is the tensor. Api is all good enough and I'm saying like once we had the pinning support I, don't think how we can make it any better and it better yeah for this particular sample. The rest is kind of not super self explanatory.

A

So I think like arrogant, come here if you're, if I'm wrong but like I, think Polly. We should just look at the at the base type and then, as the derived words right, yeah.

M

There's a really I'm, just gonna, say because we're gonna have to get to this. Some point: there's a very important piece at the end called range: oh yeah, that's like like nobody is pushing on this and I absolutely have I. In order to do tensor correctly, we need a range. We need a range type somewhere, because one of the methods in tensor is is slicing tensors and.

I

M

The idea of slicing is is give me a subset of this data right and in order to have a subset of an n-dimensional data like you need a range for each dimension, so, anyway, point being is I kind of want to push on on range as well.

M

F

L

Because because.

M

I think it's important and it's gonna take a while to get people to come to a conclusion on range. You.

A

Think the thing with range is I think a set had a great sample. Go like I. Think range on on on span is, you can say it's very minor, syntactic sugar, but it saves you like I, don't know six seven characters to the grand scheme of things it doesn't do much. But if you look at a tensor where you have like six dimensions, for example, I give very common operation that you have is that you have when you're, for example, train models.

A

You have you maybe have a tensor like a CSV file on disk or something that effectively has the inputs and the outputs like mixed together. But then what you want to do is you want to basically take the tensor that you have and get a new tensor. That only has basically, you know the last column. Now that nothing is.

A

It only is the first two problems you want to project pieces out essentially, and so, if you look at code that does that for like three or four dimensions, you're saying you range, you know one comma, two or something you also have to represent. That you want to say, take everything from the minimum to a maximum value of sorry, comma beta to the other. One there's like so many combinations that you talk about a line that has like 100 characters which in Python because they have syntax, is like five characters and you look at it.

A

It really makes sense. So this is where the thing really, you know is little more minor, syntactic sugar. It really makes a difference for me and.

M

Even how would you do it? So if you, if you want to go to the slice method, real quick.

K

Search for range.

L

M

You need you like. A range is, let's just say, a start and an end right.

D

M

Of int like well, how do you make this API? How do you say? I want a list of a start and a pair of start and ends. It's like well either. You gotta have a tuple of you, know, events or you need some type right, because you need to have a list of these types. So it's like slice on span because span is only one dimension. You only need two inch but since if, since we're now n dimensions, it's like well now you need n two inch so like. How do you have to represent that?

M

We all agree: okay, I'm, not seeing nothing.

M

You do power, so the proposal just to kind of close out on this run on range. The proposal that I'm putting forward is system that range in system that memory nougat package.

D

So so let me give people some segments because I've been working with both Eric and we talked to Jared and some other people kind of some of the trade-offs or things that we discussed regarding range. So one question is: is range of ins or longs so, for example, a range of ins does not work for things like pipelines, it's not clear whether it works for I.

D

What if we make it long if you're at a click and one for pipelines, it's just working with range range is so large, then they don't fit in int this kind of the wrong thing to do so you can have a signature in it. Lets you advance if I plant by you know four gigabytes, but it's gonna be like I know not the right thing to do.

D

But having said that, there are some scenarios where, like in you, ran out of spicy needs for like woody segmented. Pallets second thing is: somebody commented from the community that they wanted to make range of T, so ain't I, you can put any type range of T. Has the problem that suddenly, the syntax is kind of you know.

D

Prepare the tea's and the types need to be compatible. I've got some.

G

Other negatives: well, if you have an operator plus plus and comparison, you can theoretically make it work, maybe maybe just keep calling it in a loop until it.

D

More complicated that the nice thing about range just based on needs, it's so simple. We know that it's going to work and you know there are like its. You can trash it to what we were not used to prove.

H

Why don't you do not yeah.

D

Well, so another one, a it's related! It's gonna answer a question, so so the C sharp compiler wants with syntax. Basically, if you type open square bracket number.

F

There's something like this here, your agent actors, if you click that link, if you click the link in the details,.

D

They will create, you know a value of this type range type, so I can just cavity the signature.

D

This time then use the Centers for this to work with tuples. It would probably be a bit strange. We would have now two different syntaxes for tofu, like the parentheses, syntax and now.

H

We already have slinky support for it, but if they want to have specific support for changes addition, so we.

M

Also talked about how to handle lungs, if you scroll down there, the idea is: there's multiple.

D

Yeah, so this is some of the background. I I do I, did kind of conclude that I think we should start with simple in price range, and if we have a need for Long's, we just add longer range, and then you know, pipeline soup means longer range and then language we is very general. You just apply some attribute to your type. That has two fields, and you know the label just sets those two fields to whatever you you're putting in the square brackets.

H

M

The idea is, if your type wants to support range with Long's, you could put both you'd have methods that take both okay.

D

And also it's a bit more efficient if it's a small start with ins and Long's and and so on. That's all.

D

So now now the location of the type Y system, memory.

M

So a couple couple reasons: well a couple reasons that it's not in the same package as tensor is because one is I, don't see, tensor going in box anytime soon, like I, think it it's a it. It should chip in its own library kind of for a while, probably indefinitely so system memory is system. Memories.

M

Brand new is another reason that I picked that one and so adding adding types into it now versus some other library, some other package, it was it and then three tensor already has a dependency on system memory, and so sensor wants to be net standard one one and and run you know on that neck or 202, one desktop you WP xamarin, like basically anywhere right.

D

I was definitely not proposing that we could range in tensile, but I was thinking. Maybe we just put it in again system run time.

M

But then it's in system run time how what tensor work on done in four six, four six one.

D

Correct so I understand that you make it part of who eventually system memory will have to depend.

A

On that one technical problem that we have is that the is going in box into one. The question is like what I don't want to have is like a version into one where everything in members inbox button range, and so the question is because then you have like they know it gets even welcome to get know. You have, you know partially just.

E

H

A

Would have to go in budget to one.

B

Exactly so, the question is: do we have enough.

A

Want me to pull it off now, because it's a language feature, that's going in there I think I would say if we, the language feature, isn't necessarily.

M

Like it's in proposal stage right now, there's no hard. It's going in anytime, no.

A

That's true, but I mean like spent. My experience with language features is that there is then you know they tend to take a long time to mature and they do tend to change things fairly drastically along the period right. So we ship a version of range based on the current language proposal and then the guys change their mind on what they can project to, and then you have to reduce a new type that is different from the one image.

D

But with Jared we showed him this dive and we asked him what he thinks about it and he was optimist, but we should probably get something in writing from them or kind of delicate bit or before. Yes, why.

M

Did his gonna help.

D

M

Like honestly, the worst case scenario in my mind, is I ship arranged in the tensor library and then the c-sharp API doesn't work with it, because it's not like the it's, it's not the you know real BCL range type. It's just.

D

We put a wrench in system memory and then it doesn't work for the language. Now. Language introduces a new range and now system memory has awards in the range that it can. It cannot use or shouldn't be using.

M

Memories, at least core enough that whatever they do in in the language, is okay. We have to make it work with system memory. Well,.

E

That's kind of forcing their hand right. You know, honestly.

M

That's what I'm doing for everybody right now is I'm forcing the hand because, like nobody else, needs this as much as tensor needs it. So we don't need. The language span, doesn't need it right. It's like tensor needs it, but everybody kind of wants it. So.

A

But but you, but that kind of speaks to the other thing right. I mean if you put range in tensor like system, not tens of a deal or something right. That whole thing is in Clearview for a while we can make whatever changes you want to make to. Those API is eventually we get arranged. If you really really liked it works for tensor really well, and then you know the language guys can add it to the language, and that may just mean that it will take us another year until this thing is stable.

A

But then we have the best possible time and then the kind of the discussion again, you know where the sting ships, because I agree with you that range at that point shouldn't ship intends anymore. It should actually go in box and then we can find a home for that thing.

D

But then so do you think that every single you know we cannot add ATI system memory after we inbox it? So that's I'm, saying.

D

Basically means that in next release we have to create yet another, because this is the memory became. Like you know, the the uber library.

D

We're jumping around it basically became system that OOP for that net core 12.1 a lot.

E

Of these around Memorial II, you know around span right well,.

D

The bells are, the signal are dependent. Well, it depends exactly what you mean by around, because there are like utf-8 parsers and for matters and yes, they use span. Just like many IPS is raised. So you know, like all the api's internet framework, around arranged.

E

Man ISM, you think- and this is the funnel of the span right. It's not that we have random things that are entirely.

A

Let's not talk too much about that. I think the takeaway for the right thing is like 10 shows absolutely no range. We need to make sure that the design that we have works for c-sharp as well as for tensors and then the question is: can they get our ducks in a row? So we can ship the whole in to one, because it will be the easiest for me quick standpoint if they can ship it into one we after you create yet another assembly somewhere we can put whatever the range sides of being really. You.

K

A

D

Right I will just try to get the language people to look at it, but it is great because you will support answers out of pen and for firework. Yes,.

A

J

Best place forever right so.

A

You can ship not a bad words of range to go ahead. Favor, it means potentials will be substantially less. It cannot be actually in the same deal as stand. Ups. Well, not long term I could be I mean that could be a home for it and you, but again liked it I.

A

The reason I don't like it, because the basic turn stands us into this partially, because now we eventually get the type format it down somewhere, in which case you have all the problems are possible, so it would really really try hard to keep tensors pure, like similar to immutable. I would not try to put types in there.

E

A

E

It will be terribly, you know how to think about it. Is you know, tensors, you know, and you have wrench because chin is like. Is it sitting on top of the platform for some while and when it's marching, like so you're, basically saying hey, this is the pure goodness, and you know it has dependencies. One says: oh, that will eventually come into the path for this.

A

Thing thing for range: the control set is: can we get thumping one of the other types and nothing bad happens? If we can then yeah, we have to find it another sound before that and that that would try to avoid that, if possible, No.

D

So can we just send the type to the language people and say we want to ship it in 2.1? Can you sign off on the device sign off on the design eric like? Do you think they were literally.

M

I literally I copied this from that proposal and I just removed I numerable from range because I didn't need it. Yeah.

A

But is that that shape right now sufficient? Because the mostess other thing that we talked about like howdy.

A

What I'm trying to say was: is this in this shape? You currently have sufficient, because I remember from the discussion with set, we needed a representation, for we don't have a lower bound and we don't have an upper bound yeah.

M

And I think the idea there was like a sentinel value like negative one for start means you don't have a lower bound and negative one for end.

M

A

D

A

To think about, because that over every comparison.

D

A

D

Presumes that range can only be positive numbers, yes, which is not necessarily true. Oh I, see yeah, that's you!

D

A

Yeah, so that's the.

M

Design me, if you think this might be I might be ignorant here, but why? How could you start negative well.

D

It's just like you can have offset in a stream negative. My.

K

Bounds actually, this is.

D

A

Range here we're.

K

A

Of the bounds of a relative position, yeah.

E

If I take range of something no.

A

Take now thinking of a text buffer right here, the text buffer yeah. You have a currently selected phrase in.

E

The buffer sure and if I'm operating on you know some support, suddenly that thing operating on be able to step.

A

They 10,000 selected worth starting at 5,000. Now you want to do some computation that says: oh yeah, the other thing is minus 200 characters in front of your thing: Luke shifted basic, yes, which is actually super common for any sort of you know like. What's what's the Delta between my current carrot and my target carrot? Well, I give a good translators.

D

For parsers, it's super useful yep for backtracking and stuff, like this I mean okay,.

E

Let's take it offline, but I would like to know more about the use case. It's.

G

A little he was wrong.

E

It you know small parts, kind of steps out of bounds, so just.

A

Like I think, the bottom line is like every time you have offsets in length. What do you want to avoid? Is you wanna? You want to look at a code and the code should be like very simple, like X is lower than Y, then those emergency to come up as much as possible and the more special cases you address a all could be negative, like that, more more, like special conditions and I.

A

Think over history has proven that if you have negative offset, you tend to have fewer special cases and stuff just plays out, usually as soon as you do. This thing, where you say.

L

A

Means is it's, it's bet it's a bad value like you know, bad things tend to happen so.

M

You bring up a good point is that it's not start in length its start and end. That's.

A

Also true yeah, we need to also agree on what we want the semantics to be here. I spoke.

M

To Ashley offline- and she said it's- they they're pretty hard on. It's not gonna, be start in length. I.

A

Mean it will be truly try the property called length right, which gives you effectively what the Delta is right. I mean.

M

A

M

A

Pretty much all range types I've ever seen head start and end length because they keep depending on what you do is the month more convenient than the other like well.

M

A

M

Was, and also is like, is it inclusive or is it exclusive.

D

Well, that's also.

I

D

Yeah but then the question is what is the city language syntax map map to so it would map to start and end see, but.

A

This is the difference between different languages like I, think in Python they have dot dot and then they have colon and 1 means. 1 1 means the other I forget, which one is it I think the colon 1 is end and the top dog is langurs, no I. Think I I don't know.

D

A

D

Like we need to separate late, because we have so many 80s or I- think that the framework that operate on length, porting code, to something with ranges without having ability to basically say start and length, will be just really bad well, it also depends.

A

On what what that is, I mean I think the biggest problem is using constructors. They only see two ends. You don't know what they are like. The way they've done it in BS. Is that they're tech strange thing he basically takes start in length and then you also have a static method called from bounds.

A

You can say tech span from balance and you say start nth and like quite frankly as long as we are consistent with what we have and that we review I think if you'd see two is humid today like because we have no other API syndrome.

J

Of the take end positions but I getting for language, syntax I mean the language syntax. It should be start end up saying at $3, 7, nothing sure.

A

They get the mop right, but, like I made, the question I think first, I think for syntax. I could be convinced. That end is good enough, because I mean you look at the their primary example. You, with the for loop turn right yeah like that. Like.

D

It will be really weird that this is an N right yeah. Is it ok, let's not optimized, for like the stupid example, but it's nonzero, but I think it's.

M

It's not really a stupid example. Right, like for loops from 0 to 10, is like very classic.

A

And is it incorrect, it's no I think it's the same as follow, but it's exclusive. It goes from 0 to 9 yeah.

M

I think I think for 0 based languages or zero based indexes like it has to be exclusive. So there's gonna be interesting discs, so.

E

Something tells me they're, not good, all two point, one there's gonna too many questions, wretched stuff, so Michael you.

A

Could pull it in I mean it's just I think so probably just said: go for it. Let's pick it. Let's prepare the disaster. What what.

E

You know we said a fencer will be.

E

What do we want to ship it like? This? Is the final thingy uh-huh.

A

First of all, let me let Eric speak again. I would say, let's not design the thing yet because to me, that's almost a different thing: I mean I would like I would say: okay, duly noted, we need to figure out what what range has to do and how we ship it I. Think yeah cover that part. All the questions that we have about the general API but I think what carol was saying is a.

D

Sure, let's go for it, try to make it work for two point, one in case it does whatever something like in case. We cannot make it work. What do we do and I think the only reasonable answer not right now is we keep it in in the tensors and the implication is we will have another package.

J

D

We make that we.

J

Keep the trenches as pretty to mean it means that we can break.

D

No, no! No, but at some we need to when, when language syntax is added, we will want range in the platform. Yes, yes, so we will break the tensor out of the preview, make it ring. No, no! No! No forget about the tensor. That's right! We literally don't care what happens with answer. What do we do with the range at that point when we want the language yeah, exactly just to keep chilling as a new and being yes and usually.

E

D

E

Make it, inter.

D

System we cannot well yes, so, but now we have discussing the contingency plan. If it doesn't work for a you know, system memory 2.1. We basically are signing up conditional ooh. That is it not in the future. To be fair, though, like.

A

We have the problem with like I mean anything so I think it would be convenient for us to add it to 2.1 if they can't add it to 2.1 I, don't think anything bad happens, because this is not the last name which we chose every ship right like there will be other things it had the exact same process well,.

D

But I taught me so I thought we said from now on all the kind of boobs that we know that we cannot have to imply. We have been meeting a review and we kind of basically like we don't do it very three person. Yes, okay, so let's not meet me.

A

The number one girl what you shouldn't get into one, because the cost aside, if they can't make it into one I'll honestly say to me- ranges like almost in the same range as tuples right. It's like a feature that we really really really want to get out of band because you really really want to support for femoral get is, and we can't probably based on our experience with couples sketch what the cost would be. Now the question is: do we actually care or do we say? Oh you know too bad.

A

We didn't make it four to one so we'll be only in core and too bad. If you're running on senator framework, you don't get it. That could be a output of.

D

That but I would say: if that's the decision, we should be making it in this meeting. We should talk to like either and other question.

M

More, that's it for tensor for tensor, it's not okay. To just say this only works on dotnet core to one, which basically means that's.

E

M

Non-Starter for tensor in has to work, but.

E

Here's my question was like relate like rv4, or do we plan to ship tensors as like RT an equality or something like final non preview during the tuple on time frame? Yes,.

M

That was the plan.

F

M

This ships, this ships are stable version with donek or two point. One I.

E

Say that I didn't get that we.

M

Can change that if these ranges reasons see.

N

Why does tends I have to be stable.

A

Into one because one or two one is not that long, so.

M

Okay kind of the the reasoning is: there's some problems with uptake nobody's gonna, take it as a dependency, if it's in preview and so getting you know, feedback and users on it. It's hard.

F

Right now that.

M

You know especially right now when it's in corvex lab, because it's like well, we could just throw this away. So why would anybody want to treat- and you know outside of just playing with it? Why would they actually want to take a dependency on it? Use it on the you.

A

Know uptake is the primary thing, but I would disagree with you on the fact that previews the thing that holds people back because I mean like we had those beautiful collections where we shipped essentially a preview version for I, want to say something like nine months and we had a decent number of downloads like in the order of like I forgot, the number but like 20,000 30,000 dollars or something like.

A

We got a ton of feedback and we did grab the API service as a result of that I think the biggest problem usually is, as you said, like it's a quarter flex, it's a core of xlab. It's a to set up because it's on it's on it's on it's um I get usually many of our things. Don't just work on full frame with they're pulling like 50 other packages to work on full frame work.

A

So these are all the technical problems that we usually end up having, but I think we could solve those issues, regardless of whether you know tensors is stable or preview. So.

M

C and TK is is one of the major people. We want to be able to use this right. They ship stable packages every couple months, maybe.

M

They they shipped 206 months ago and then at two one four months ago and then a two to three months ago, and then there are two three twenty days ago and a two three one, seven days ago, if we, let's just take the plan of tensor, is nine months in preview. It's like they're, not going to take it for their next. My.

A

Point is not so much the drink for much. Don't forget about the timeframe like the the the biggest problem usually ends up eating. Is you have a decently sized API servers that we haven't really even track the API service? Yet what I'm, after is having enough traction on the API servers that we know the API is good.

A

That's that's really my concern I'm saying given that to one ship like only next year, like what's our chance to make enough inroads with partners in that timeframe, to convince ourselves that the API shape that we have is good enough. Well, it was questioning.

D

That that that line of I didn't get, maybe that's also a confusion about dates. I, don't think he was saying that it was same shape with the point one. It will just ship in the timeframe, so two point one can shape and then tensor can shape stable. Two months later,.

D

So it doesn't matter as long as range is stable. You know that's what is it about range being stable into one or is it apart, like.

A

A

E

After available for two one, like I've I've much, both tensor has dependency on ranch, which means when stanza goes. You know stable ranges to grow, stable, I. Think I. Think range is one thing, because the API is so.

A

Small I could be convinced that he pulled us off this one is pretty large and it has a non-trivial interaction with the rest of the ecosystem and.

D

So Ellen can you clarify? Is it? Is the plan to ship this stable in the 2.1 timeframe or at the time when 2.1 ships I mean.

M

The current working plan was when 2.1 ships tensor is stable as well, so.

D

At the same time, you don't you do it, you, you think it would be unreasonable to have like a month or two months. Delay I mean.

M

I, don't see what to.

A

Me it's more like, as I said, like it's validation time I mean like I highly doubt that we will be able to get a significant number of downloads and uptake to let it in the design.

M

Agree, especially, it's living in core effects lab yeah.

A

So I think I mean I agree with you. The minimum thing is, it has to be on you get. It has to be an official release, otherwise we'd never get to anybody using it and that one I think we should do earlier. Rather than later, and given that it's preview, we can't bring any API as you want. You can break it apart, move to rename whatever you want to do, but I think great worth of things to validate.

D

This whole thing holds water by the way. I, don't believe it has anything to do with core effects lab it's more local versus michael, correct, true, but we can't, I think, became produce and bizarre effects lab. We cannot work.

A

Produce sign bits out of core effects lab. Please.

G

A

Yeah I mean from a customer standpoint: they don't care about it, but it's.

G

Really about you know.

J

J

Does that make sense powerful, do you do we? Should we inspected lee something that often, but how, in practice, definition.

A

J

A

J

A

Wants to do it, we just do it, but there is no schedule for that.

M

Like we just did drawing.

M

A couple versions are those right, so.

E

Like once a month, we once per two months roughly in practice, yeah, but we can pull it off and it's not high overhead for the preview in.

A

From uniboob we had some releases, they were like a week apart, because somebody found a particular bug and you just fixed it and uploaded a new version like it's not like. We can ship as 3po many hours in quickly as we need it's.

E

About investment on our site and yes, we have come at it so.

M

So I'm not I'm, not a hundred percent committed to like tensor, has to ship stable at the same exact time. Dinette core 2.1 ships are TM, so it's kind of like it. You have to have some sort of stake right to say, but.

D

Also, it doesn't matter like so for the purpose of the discussion of range it almost it doesn't matter. Basically, you want to ship before the subsequent parachute knocked out that core. Therefore, in that network, two point one that is the range needs to be there. I mean I. Think I need this requirement, I'm, not saying that it's doable but I'm saying if we follow the requirements, that's the implication.

D

Right range is a general purpose type. It's not urgent.

J

D

Point one point, one more something pinch of range, one, one additional DLL that will be in place, so we can extract with you know it's it's just. It just adds more complexity, and here else.

H

A

The theme is that people are trying, it has to be because it means we we already have to support a version of memory that is type forward, is the platform and.

E

Friendlier there in its own deal like good point.

D

All right who's gonna work with the language team very quickly do.

A

The range thing so we could do I can I can help Eric like bring up the discussion again with mats and see whether the kick we're on the on the on the calendar. I, don't know what that backward looks like right now. I have attended their meetings in a while, but I mean the nice thing. Is we don't need an end design from that might be just basically to agree on you know. Basically, you know the the rough corners of the design.

D

Point well so imagine, for example, that initially because the distract without.

D

And then the language in the sides that the interface is required. We cannot enter interface because we cannot interface to packages that are being that. That's.

A

What I mean by that? The rough design points and I mean like like things like what interface has to implement, what the shape is the language binds to that I mean those are the I.

D

Wasn't polish that needs to be there API that they sign up to support otherwise we're screwed? They don't need to have.

M

A sign-off from them that, when they do this language feature this.

M

A sign, a sign off that this type, looking exactly like this, will be supported in the language feature. Thank you.

A

Limited in the in the language of laminating is organelle to those discussions. It has to do with things like. How does this tie typical resolution like? How does this next time with the other syntaxes over here like? Is it? Is it unambiguously possible, in all cases?

A

Oh, how do we line up this feature and we want to have code that was composite, but they never know to compile it consumed by law over you know those things play and those things they don't need to login that can they have like you know all the one way the world they want.

A

I mean, quite frankly, I, don't care whether they're support language semantics in you know, whatever the next release is seven five or seven, eight or eight like what whatever timeline they put it on, but they cannot change the way, the language by it to the API and that's the thing yeah okay, which seems I mean, given that the proposed is pretty detailed, already like it seems, not unreasonable.

A

We have enough design possibility to write.

E

That we remind you your your comment on value topple. If we you might, we should not do anything like that ever again to.

A

Be very clear, like I'm, not pushing for this being in I'm, not saying making this loop is easy or hard. My statement is mostly realization of the ball. Pivot, which is languages evolved, the platform evolves, and there will always be cases where somebody will make a business decision and say I want this feature or this slice of the future. No pun intended to be available. All of that and I mean our system has to handle this. Somehow it's just you know how expensive all those things so.

D

A

With you, except I, think we.

D

I talked with Peter and we basically said span, let's grandfather it, because we basically didn't realize that we have this massive problem. By the time we already signed up all customers and like kestrel and pipelines in America, it would be just a mess to completely change it and give with these problems just just go for it.

D

From now on, we don't make in this meeting in the API review meeting these kind of decisions it needs up, maybe with I agree with you that we will probably have to do it, but we agreed on a separate process, for this is to be parents. Visit and more people need to be in the loop than just us before we kind of practices and I.

A

Think yeah like to me it's you know, I think what this room is good for is things like all we you know finding out in the dependencies and say it would be really good if you could ship it into one because then we don't have to solve the following of you. Can assume things are there if it can't make me four to one then well, you know is whatever it is like it's well I think you.

D

Know if we cannot make it for two point, while at least more people in our organization will understand like that, the trade offs, because I think so far, there's been a bit of a kind of wishful thinking like let's not do it, and now, for example, once we decide not to do it for a range there will be. The implications would be very, very obvious yes and like not that it's you know we would be in that nice price, but at least the implications would be obvious.

D

So, like more people will get feel like educated on oh well,.

E

More people different force to understand it up.

D

Side up, yes, I.

E

Will say these.

D

Are education exact, more people in the month, but.

A

D

Good because, right now you know because right now we are in the mud and other people are not, and they basically imagine this a you know: fiction area, beautiful wharves and everything just works. Correct, yes, yeah.

N

Let me tell them: you know that this API and they want the cake.

N

Discussion minimums.

A

Well, that's a discussion, I think right now that is going on on that topic. Yes, so um how about this? So we didn't really cover any. He did a good job one hour. Eric like this is something you need to keep it on. This features have another one: we using push this off for next year and focus on the range discussion, um I.

M

I would really like to get a like a preview version of this out with panic or two one preview. One.

J

F

E

Should we should finish up this year, I nearly. Is there any problem actually shipping to preview this way without the API review? Are we gonna feel, like you know, so,.

A

We generally have never shipped previous without any apparent loss reserves.

M

Right I saw a commit or a PR from Eric st. John in the core FX lab. That was like take review feedback from from the API review. Like has this already. Has there already been around of was.

A

Around with Eric crystal and myself and I think that was one people in the room and it was like a very early look at this whole thing, I think, but since then I think this thing is also quite a bit I wouldn't I would say before which day we should probably do another realm. So then the question was just technically: when does it have to happen like I'm?

A

Finally, setting up a two hour slot on Thursday like getting me home right now, zero, like you guys, are fine, then it's then we set up it so for Thursday in.

N

M

Format: okay, do.

N

You see what is.

M

The format of what I presented- okay, Mike or do you want things changed? Yes,.

A

I want things change like you. What would you will you present here is the reference assembly, because you have all the bodies and you have all the attributes and everything is fully qualified.

A

If you open the thing in API review, where you can also just copy and paste that representation, but ideally just sent to the apex or send me, the module and I can create the apex out of it myself.

A

Did he ever decide.

M

You fit tll yeah.

A

M

Okay, yeah I was I was unclear on how how you wanted it like formatted formatted, instructure gross. Well again, you.

A

Look at this thing, for example, right, usually what you want is great, that's really good! You know this would just want to see me :. You don't want to see America like we just want to have a short, because it's almost like an OD human compiler I see I want to read the thing like okay I mean.

N

M

I can run that as well.

M

Okay, I used I used what we gen the ref assemblies for usually.

A

When I look at like ten api's having it in the get up, issue is usually fine because I just write the bullet points after, if we review, like you know here where you have like hundreds of api's and apex, is somewhat nice and because I can just tag the api's of the comment and then it's somewhat more context but like yeah, it's not like the high order bit here like yeah. Just save me, the module like I can just open it present it and add comments as we go all right.

A

Then let's have a meeting for Thursday. We should have another conversation for favorite on branch.

A

And with Matt's on what the design points are that we can close on that alright.

K

A

Week, sorry, everyone next week, that's a great question: I will be in next V. Now, what about you guys if you.

D

Guys are move like there's no point for me and just not available. C'mon I will be working next week, but I would prefer to work from home work from home as well. I would kind of like.

E

D

You know I don't want to work from home, probably but I want to work from home several days. So.

A

D

A

What we only have we just do it quick like seven.

E

A

You ready I can make it a tone. Yeah I, don't write like three next week or you can just do the one or you another.

E

Option is to do it offline literally, like you know, one two, three, four five things which are simple.

E

Which I would like to get for the holidays, because I might jump on them by living a free things in networking, so yeah alright did.

M

You want me to sync up with the language people.

A

And technically, on that review, meeting set as well, but I did a horrible thing. The last three weeks of attending did Jeanette Center yeah. If you just see see that's fine, okay,.

M

A

Then the other thing is talk.

M

Do you wanna be involved as well.

A

Cc me and any other questions like like: do you have done any thinking on the lower bound upper bound kind of thing, because I think that would also be useful if you have some code samples for all that I.

A

Think that's the thing that I don't think they cover it all right now, in their respect,.

M

Okay, so you want like a like scenarios: why that's it useful yeah.

A

Like I think, what I would like to see is like the sample code, that Seth had where you basically slice out a whole dimension essentially or you skip the first two or something like that, but they don't take everything after that. Yep.

M

Okay, I can easily grab scenarios for that yeah.

A

Alright, then thank you and then you guys online and then see you next week.