.Net Foundation Design Reviews, 9 Jul 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: GitHub Quick Reviews

Description

00:00:00 - Approved: System.Formats.Cbor: Add Reader and Writer methods for System.Half. https://github.com/dotnet/runtime/issues/38467
00:01:32 - Approved: Add PlatformID.Unknown value https://github.com/dotnet/runtime/issues/38984#issuecomment-656250578
00:05:57 - Approved: Request for Logging Internals to be public again https://github.com/dotnet/runtime/issues/35060#issuecomment-656257152
00:23:21 - Approved: Http header value Encoding selection https://github.com/dotnet/runtime/issues/38711#issuecomment-656265904
00:42:29 - Approved: Add static hash helper methods https://github.com/dotnet/runtime/issues/17590#issuecomment-656269790
00:50:49 - Rejected: Add EnumMemberAttribute constructor https://github.com/dotnet/runtime/issues/27655#issuecomment-656275959
01:03:59 - Approved: Inject existing object into MEF2 https://github.com/dotnet/runtime/issues/29400#issuecomment-656282604
01:18:40 - Approved: SequenceReader.TryReadTo(out ReadOnlySpan(T) sequence, ReadOnlySpan(T) delimiter) https://github.com/dotnet/runtime/issues/31355#issuecomment-656291313

A

Everyone, so we have a bunch of backlog to get through. So let's get started three eight four: six, seven reader and writer methods for half on C bore, approve yeah.

B

A

Only question I would have is, is, is half with this casing? What we're using everywhere I assume it is because we use single and double as the because of the type name. So.

C

That'd be my assumption.

D

A

Right I'm supposed to do something silly like.

D

A

One minute new record: alright, a bunch of stuff on logging that we could possibly reason about or we can wait and see if Miriam joins the call. I have a week and I'm driving.

A

Alright: three eight nine eight four had platform ID unknown.

A

Eric yeah I see you're on the call. Do you know about this? One.

E

We're having a bit of trouble hearing you yeah I,.

A

Heard your voice, but not your words,.

F

F

I mean I think it's just something for for wasm. What do we say for the platform ID.

A

Right what's so, this is the legacy API, because it's the enum right right.

C

In the proposal summary here is basically saying in order to avoid duplicating information, provide and explicit unknown value that we can return from the legacy API. Okay,.

A

So anything that we haven't already started restarting an answer will just get unknown.

A

E

I would probably call that other, rather than unknown,.

C

Maybe we should call it go look at system runtime and trough services, runtime information, yeah.

A

E

Scene of an implicated, and just just to confirm, like we never plan on adding anything else to this after unknown or other correct right.

A

Well, because it's it's very yeah, it's weird and it it ended up being effectively limiting, because the current values are mac. Nokia, s60, unix, win32, NT, win32, s1, 32 windows, which is different than 132, NT, wins and xbox. So it's it's a weird grab-bag already.

A

What got the zero value.

E

Winter beat us.

A

F

Mean should we be marking this thing as obsolete as well platform ID.

A

Probably, but all the obsolete, stuff's being sort of tracked differently, yeah.

C

I think this would be good for emos new targeted, obsolete proposal where we'd be able to basically say if you're using platform, ID you're, probably doing it wrong. You want this other thing. Instead, yeah and.

A

And really it wouldn't be on the enum, it would be on the the method that returns this because I don't know that I don't think we use it as input anywhere yeah I. Think.

A

Yeah I mean we could obsolete both so yeah, a unknown or other both seem reasonable if they had started with unknown as the zero value. That would certainly be better, but zero is already taken.

E

A

E

We keep the value as at seven as proposed here. Really. All that means to me is that we really shouldn't add anything after it. If we did plan on adding stuff after it in the future, that I would prefer that this has a value of like negative one or something I mean it's.

A

Not an order enum, so the number doesn't matter yeah.

E

It just makes it easier for people to check, especially since, if I have a net standard library, for instance, I won't actually be able to compile against the literal, unknown value, I'd have to say equals equals seven or less than zero or something.

E

A

You know equals.

E

Equals seven doesn't bad yeah, unknown.

A

Is how you do it yeah we well, it looks like we have a lot more unknowns and enums than others.

A

So I think good, as is, unless there's strong suggestions to use a different name.

B

Seems fine to me.

E

Or report content, because we don't want people using this say yes,.

C

Did we say we want unknown instead of something like other.

A

Yeah no one strongly objected, but I asked for strong objections.

A

There are a few others in existing names, but unknown is far more prevalent. Yeah.

C

Well, I was just thinking from from this scenario: it's not that it's unknown. In fact, we do know it, it's just you can't get it from this aprv yeah.

E

That was why I had suggested other, but it was just a weak suggestion. I didn't really care all that much.

C

And I don't have a strong objection, it's just given that it is more given that we do actually know what other might be a better name.

F

F

It looks like we use this in the operating system version string.

F

Would we expect that to say when it's other, we would expect it to say other in the string to match the Union value.

A

I mean is that off of that's off of OS version.

A

Because if we're reporting things like win32 NT, then like I, guess this Linux a Linux. What is Linux a UNIX right.

F

We always it always says UNIX right, even when you're on Mac, because mano said UNIX and then we match their behavior so max. Never, basically it anything. Unix flavored says yeah.

A

I think the OS version is probably not what we want like yeah.

E

We could we could also, since we were talking about off solution, we could also obsolete the operating system. Dog version string property. It wouldn't be outlandish for its rich on other here, because it is basically the stringify version of the enum value.

A

Yeah all right any objections to the current thing: modulo, whatever the string is gonna, do.

E

A

Hi Miriam spot there.

G

Yeah, so this one basically.

G

There's a bunch of overloads that we're missing and the person who created the issue requesting that a penis is basically asking for so longer messages. Those are the existing the old world. Let's take T one to T I'm, trying to see from my screen t I, think t43 I don't have enough resolution, but then they used to be public first of all and then not and then the question is so: we have two API, one is defined scope, the other one is defined.

G

One of them has eight overloads one two or one, two, three, four, five, six, seven, seven or eight overloads the other. One only has four look at the ask is to have the same amount for both of them and then.

G

And I believe I think used to be public and now they're internal I'm double check. So I was wondering if this is kind of like a slam. Dunk API.

G

A

Dude, if they were, if they were public and stopped being public assuming this assuming that was not done across versions, then there would be the question of do you want to bring them public again, as for just expanding that the generic matrix up to eight on both? That seems easy, yes to me, given the first question.

E

Generally, making API is consistent with each other is, as you had hinted, a slam dunk generally.

G

If this always was in our libraries, I would have been able to check the how the past ones were in API subnet right, but I can otherwise right. I want to confirm the fact that they used to be public. Is there any way other than api's I've got net, but I can check well.

A

Like if I look in AP eyes of dotnet I see defined scope, t1, t2, t3 and t1 t2 and t1, and not the top one, so I don't see, I, don't see the one that returns just func of lager and disposable, but I see the other ones.

F

It's actually below the defined scope. Thank.

A

You Eric I.

D

G

A

F

F

A

Don't know I would believe if you look at an ASCII chart, the less than might come sooner than the parenthesis.

F

Literally, just asking for defined scope to have up to six. Just like yes and.

A

Then adding seven and eight on both of them. Yes,.

G

And then I ask why eight I don't think they had enough I.

A

Think their answer was because that's what tuple does but tuple technically does eight in that it goes seven and then says, etc.

F

B

Fine, as requested uncle's.

E

H

E

Funk only used to go to eight I think, but there was a change in what taught not 4.5 I think to make it go up to 16, looks like four or.

A

E

A

So yeah I think so, given that these are already public and it doesn't seem like they went private or if they did, that, that was gonna be a breaking change. So, given that these are already public, certainly adding the tea for 36 is easy.

D

A

If you want to expand it up to t7 and t8, that seems fine, unless anyone can think of reasons why this is not good.

F

We go to eight because that was the old way we used to do it for other things, because.

A

It's where tuple says it's bored and had enough.

A

For tuple partially, this would have been object size, but it's a class-4 value tuple. It would be struck, sighs, yeah, I, don't really know why they went seven and then complicated other on to pool yeah eights, the historic. Why do you think you need more than this number.

F

But then an action in funk, we said: oh, we need sixteen yep I mean I'm, not I'm, fine, with not going off over 8 I wonder if we need 7 & 8.

A

Yeah I mean if they're I guess, if they're just asking, if no one really has a need for them yet, and certainly we can do the same thing that we do for tuple, which is when you hit the end. You make it. You make a thing to hold everything else.

A

Great so bring it up to parity is easy. If you don't think, there's a need for 7 & 8, then maybe we hold those back.

F

But if it's weird for me to go to take to 8 I mean if we're gonna go more I, don't know find out, go the pole. Why stop at 8? It's kind of the question like I think that the bringing up to parity is kind of a slam. Dunk like why. Why does one go to 3 & 1 go to 6.

F

But then the question is like: why go to 8 if you're gonna go more when.

E

Just out of curiosity I'm, looking at the implementation details right now, I don't see anything that would be prohibitive as an implementation detail from going beyond date. It's just it seems like it is touching a bit of stuff, though that's all.

A

Yeah, like I, imagine, there's a for funk. It's easy! We just define more and more of them for this I. Imagine there's a lot of annoying copypasta going on.

A

Because generics don't always compose nicely.

A

All right, I think.

A

This covers what we talked about and.

F

He does say in the beginning he said I'd like to have a logger message to find method that accepts more than five type marks.

A

That would be six she's more than fine. We have the parity number I'm.

E

In in general, like I'm, I, wasn't honestly that keen on seeing that we have a thunk that goes up to sixteen like that. uh That honestly seems like we're missing a fundamental language or runtime feature, rather than something that we would rather do then make you know sixteen different versions of a given type along with all of the plumbing required to run them, but I. You know in the face of that feature, not existing is this the best we can do, maybe that I I, it seems like it seems like cruft right.

A

Well, I mean really hitting up the limit of where functions are realistically defined, right, like things that are longer than this typically have outs and can't use funk anyway.

C

Iii think it's worth noting that well there's two parts, one most other languages handle this with very attic templates, for example, so you can actually have arbitrary number of parameters to the template or generic, but outside of that, once you start getting above. Like four arguments, you start getting into calling convention weirdness where you're actually like basically killing perf and maybe rather than exposing something that's more than let's say three arguments which is generally where we do the cutoff to switch to params anyways.

C

We could define the one that takes, for example, T four.

C

We could recommend people that take something more than let's say T, four, as they should pass in a value tube will here, which is infinitely extensible will be passed more efficiently and solves the issue of I've got a lot more parameters. I need to pass.

A

Yeah that mean that would be and that'll be true for all things. Returning long funks.

A

Or number of argument phones.

C

Assume the reason that they did 16 years because funk itself goes up to 16.

A

Well, they only went to hate nope.

C

Our current implementation of funk goes up to 16. Now, I, don't yeah.

A

Yeah no I mean in this proposal they only value.

A

But four or five and six are make defined, scope, look like define, and that makes total sense to me if we've already done the one. Let's do the other and there may be perf that you don't want to do it, but symmetry is better than purity.

F

Right now in my head, so what's symmetry with phone connection mean let's go up to 16. If we're gonna go more well.

A

Self symmetry, so the local symmetry like. So that's what well I think this is an easy yes and then the comments of if more.

D

A

Then go for it in on both defined and defined scope and if, if not, then maybe stop at six. But if there's a good scenario for eight, then.

F

So we wrote up check up to eight, but if there was a good scenario for nine I.

A

I would say that this is a check up to sixteen. They can't actually go to sixteen. They can over go to fourteen, because they have one of the stars.

F

What stopped me so.

A

C

Yeah I, don't think we actually would have lager support above like for wood. We, because that would just switch over to trans, wouldn't it.

F

Apparently, no one knows yeah.

E

I, don't know yeah, but until someone's actually asking for that, though, it's probably not something. It's a concern for future us I, guess yeah.

A

I mean so the I mean, based on the comment of you, know, picked a Taz. That's where tuple and value tube will stop I, as we've already said, I would stop it six and six. If you need, if you need more than I, don't think you need to come back to this meeting.

A

Unless Steve wants to jump in with a really quick, why we don't want to add more generic overloads.

I

Sorry I have no context yeah.

B

I

A

You could have just seen a lot of tees on the screen and what no? No thanks.

D

D

E

Aren't we defining a custom delegate type that takes them all those ends.

A

Because guidelines say don't use in okay.

D

A

All right wait: weird: the there was another blocking one I, don't know how I lost it.

A

And I feel that there was a different one.

D

Apparently, don't I prove the right thing.

A

Yep, weird cuz, there was something about a attribute: I thought. Oh no, okay, alright losing my mind.

A

Blocking was three days ago, but Cory just marked it ready for review yeah way to jump to the top Cory yeah.

B

Basically, we have a couple needs for people to be able to define encoding in HTTP headers, the HTTP spec. Actually doesn't it doesn't formally support in coatings, it's kind of a contract between the client and server. They just have to know that if they send on ASCII characters that they will be interpreted correctly, so this is a little bit of a dangerous feature in terms of like people misusing it, but essentially for reverse proxies.

B

We need to be able to pass headers in unaltered between asp and HTTP client and we've got a few other requests just for talking to very specific servers.

B

So we added these properties on sockets handler and multi-part content that allows people to essentially plug in a callback that says for this header name, and this request message used this encoding.

A

A

The guidance we have for when to not use funk is when you think that there's not enough information and the types to tell people what is going on. I know that.

D

Looking at this I don't register.

J

A

It's like I, don't know that I would know what string is or what the string is, but possibly the only way that I would like I can be convinced either way that funk is right or that a custom delegate is needed just so that you can have so that you can name the string parameter and let someone know what they're gonna get back.

B

Certainly, have a delegate I'm a little less worried about this. We kind of don't want people to use this API like this is super advanced scenario, so it not being super easy to consume. Just by reading the API service, maybe isn't terrible I.

A

I

I'm fine with the Fung I, do have a question on the request, message and response message. So, first I'm not it feels like we're trying to predict future scenarios that require their inclusion I'm, not sure that they actually are warranted. I know Corey. You said you had some scenarios where you thought maybe potentially theoretically they could be used, but the response one in particular we're going to be handing out a response message: that's not baked. Yet that seems dangerous.

B

Yeah we could change that to the request message. It might actually be better in that case, because I think there are a few places in our implementation, where we like expect 100. For instance, we can end up rereading headers over multiple responses for a single request so that maybe we change the response message to or press message, but even.

I

With a request message like there are, depending on the order of the headers, we think it end up seeing differences in the request message right like if we get a location header, we will update the request message: URI with that location, so, depending on the order, the request message will change per invocation of the delegate yep.

I

That also seems weird I mean, obviously they could they. They could close over something that might give them access to the request mentions that was being used, but we're explicitly handing it to them. I, don't know it feels yeah.

A

So the the scenario here is: please write these bytes as I, give them to you right, more or less yeah yeah, but that the API takes a string, because that's what.

I

A

The API is thinking.

I

The API is taking the name that the string header, the encoding, is used for the value. Sorry.

A

I mean the existing like add header and and stuff.

I

B

A

I

Are all strings right.

A

Because I'm just wondering if there's like, if varying and codings is needed, or if the only real use is I, would like to give you the bytes directly for this header value. Well,.

I

In some case, that's for the request direction, but in the response direction the server is sending bytes and they're handed out to the user as a string.

A

But this kestrel need that direction as well. This.

I

Has nothing to do with kestrel.

B

Well, the reverse proxy will need to read the bite from a response and forward them to the client making the request. So yes, this is some.

I

Some server sent something it may or may not be kestrel and the you. The value of the header, needs to be translated from bytes to a string.

A

Okay, so this is your reading, a response, and you want it to you- want to give it an encoding value that essentially just passes everything and you're going to give it to that same encoding value which is going to just put it back. So you tell it. This is utf-16 le and and it just parties on without validating that it's sane, but whatever right, yep.

A

Otherwise, you would end up with potential. Your p-- means that you get a replacement character inserted into the response, whereas.

A

Something else would have passed the bytes direct, because it seems like we're. I mean I, understand that the yeah I mean encoding is the right thing, for we had a string, we're turning it into bytes. Now we have some bytes and we're turning it into a string. It just feels like it's a very powerful and flexible thing that has a very scoped need and and Steve's concerns of every time that you see the request. It may be subtly changed based on how other headers got interpreted.

B

I

Pretty clear, I don't have any problem with the delegates. I was just questioning the second argument to them. Yeah.

B

It's really the main use that I can see for having the request message is if you want to change how you're interpreting headers, based on like what domain you're accessing, for instance, if you have a client, you know it's common for a client to be used as a singleton to access multiple domains. Maybe you just want a special case, one of them at the same time, I could easily see the recommendation being you have this special domain, just give it its own HTTP client.

B

D

Working without having.

B

The request message: at the same time, once we define this delegate, we can't well, we can, but it's not ideal to add a second delegate down the line. If we find out that something in the request message might have been useful.

B

The message also has like that property bag on it that we talked about it earlier in the week. So it's there.

E

There's also an extensible.

B

Way for users to pass in state if they need that.

E

By the way, I have seen Mathare scenarios where you want to inspect the hosts before making a determination as to the encoding. I have actually seen that in practice, I've seen applications that, when they communicate across the internet, use puny code when they encode the host header. But if they're communicating across the intranet of they use, utf-8 I, don't know why. But it's what I saw in practice.

A

Of course, then, there's that whole you.

D

A

Does this host mean what domain is it in there? Sorry, what scope is it in problem, but that's their problem? That's.

I

Interesting I guess you could come up with an encoding that went to puny code and then encoded that into bytes absolutely yeah I. Had we don't have one of those but I guess you could write one yeah.

A

I mean you would just call the whatever thing that has a name. That's obvious once you know what it is. It's nothing.

I

A

I

A

I

A

Id and mapping dot, whatever the method on that is and and then run that through the classic ASCII and then run that through encoding guy, ask.

I

Uri, didn't you know whatever yeah I.

E

Saw it in practice, yeah well,.

I

This whole, this whole thing exists for not good ideas, but see them practice right indeed,.

A

Okay, so all right so was that both request, header and response header selectors take the request message as their.

I

Feel like you're, going to do that my biggest concern was ending out the partially filled response message.

I

So if we change it to request message, we feel like that's useful for the host or, for you know the URL or whatever, and accept the fact that it may have changed between you know since the last invocation.

I

I guess: that's reasonable! Okay,.

A

And look like in the chat Levi proposed a delegate head, a header, encoding, selector of to context.

A

Instead of using funk, yeah.

E

Where to context is the request message or the HTTP content like I? Don't I, don't know if it's a good idea, but it allows you to have a naming.

I

E

I

You mean you're, suggesting that delegate would then be used for kestrel as well for like they would change their signature. If.

E

They wish, by the way, I had a quick question. A related question in the header encoding selector scenario are: is the HTTP content itself also unbaked like? Is there any risk to accessing that's properties.

I

The HTTP request content should be baked as it was supplied by the caller before it was passed into the send. We also should not have started actually forwarding it. Yet, if we're sending headers I guess, with the exception of, if something like a redirect caused us to do it again, we would have already tried to send the content once.

B

Yeah, the the multi-part content property is interesting because the HTTP content there, basically is just this, so that could could be it's basically a way to overload a method. When you can't overload one.

B

I'm not super sure what you could do with that HTTP content. You have it, but.

E

Zuma bleah that selector is only selected for the outbound for the request off for the response correct, because you can't generate a multi-part content until you've already decoded the headers. So you have one kind of a chicken-and-egg problem right like you can't set a property on an object. You don't yet have well.

B

No, the multi-part content can contain additional that can contain multiple requests or responses.

B

So that that selector would be used for the inner requests or responses, not the the main requests headers, if that makes sense, yeah.

A

Yeah, so for the custom delegate or not I guess I'll leave it as a note that it was suggested and it's weather the weather. It seems valuable to be able to name the parameter. If you believe that this is sort of a we have a scenario that needs it, but we kind of hope, no one other than Europe or theoretical. Your competitor would use it and that's.

I

I believe I I misheard. What you said, I thought you said to be request, message or HTTP context, not content. That's why I was asking about Kestrel no.

E

Sorry, maybe I yeah I guess the other. The other question here is: uh is this now forcing you to materialize header names or strings, and we are already doing that as an implementation detail or as a public contract as.

B

The contract well I mean in theory, we could follow.

E

The original bytebuffer.

B

And kind of late initialize, the header dictionaries, if we wanted to, but in reference I, don't know when we would do that for a request.

I

Well, so for requests messages like they're already strings, that's just the way it is unless we were to add a new API that allows you to pass them as bytes, in which case we would have to materialize them to pass it a string.

I

Similarly, for a response, we can, you know we only have to technically. We only have to materialize them as Corey was saying, if someone actually either accesses the header or it's one, we need internally to do our job and if we were to add an additional API that allowed access to the raw headers, we could do that in a way that didn't require that, but then again, if we did that, you wouldn't require an encoding either.

B

The only other note I had is this kind.

E

Of is a dangerous API.

B

To use because if you think this is just a quick way to add Unicode support to HTTP client you're gonna have a bad time when the server tries to interpret your utf-8 as Windows 1252 or something like that, maybe worth adding a dangerous prefix to these.

I

And I gave a safety bang coaster that I personally, don't think we should. There are plenty of other ways to use the API as I need to be client to do dangerous things. I mean.

A

Assuming it's not currently validating that you don't insert a bunch of null characters in your string like you, it's already possible to probably send malformed or confusing data and yeah.

B

I mean in in that case, though it's kind of you have to maybe not intentionally, but it it seems much rarer to do that than to just say, oh this, this is how I add Japanese support to HTTP client. That seems like a much more innocent error that that anyone could make right.

A

But if the problem is they're asserting an encoding and then they they call a server and the server comes back with malformed requests. Like that's, you learnt, you gets max very quickly, yeah.

B

We're just depending on the server to validate in that case, which they should be.

A

So yeah I don't think it's right. Well, it may mean that you're speaking gibberish to the other side, you can do the same thing with TCP clients, so I don't or socket or whatever yeah I. Don't it doesn't hit my personal.

A

Maybe we need scare words here bar.

E

Yeah I agree with Jeremy is the suspect. If, if there's a mistake here, you're going to discover it very quickly, it's not it's not going to surprise you in production and that's the kind of thing that we tend to be scared by.

B

Okay sounds good.

E

All right, it's good to bring up, though I'm glad you did ring it up. Thank you.

A

Any thing wrong in my notes.

D

A

I may have done this. The wrong order. I'll see a female stool picks it up.

A

All right, I guess this one's me so cryptographic digests we're out of blocking things by the way the current priority or the current filter. Is things marked as five ready for review oldest to newest, so cryptographic, digests and dotnet are polymorphic for some fairly reasonable reasons, but frequently you just have some data and you want to hash it in a one-shot, but you have to create a disposable object, ideally in a using frequently not and then use it, and people have been asking for a long time of why isn't there a static somewhere?

A

That will just do what I want, and so this is that so well. The only thing that I wrote in here is sha-256. That was just to avoid a bunch of copypasta across all of the simple digest types which are md5, sha-1, sha-256, sha, 384 and sha-512.

A

So the instance method is called compute hash. We can't use that because that's already the name of the instance method and across four years the best we came up with was get rid of the word compute.

B

I like it is there a static hash size on these, so that you could initialize the destination bytes easily, though.

E

You're expected to know when you call them ya.

A

Mean it's sha-256 that tells you in the name. It's whatever 256 divided by 8 is.

A

E

Which I think we decided trying to do math.

A

Sleeping is hard: okay,.

A

D

Looks good to me cool.

E

That was easy. There was a question in chat. Why? Why is this on the wire up? Why are these static members on the types themselves, instead of having a static method on hash algorithm? That also takes a name as a parameter.

E

So you would pass in hash algorithm named dot, sha-256 dot 512, whatever. As an argument.

C

You know coming from math and math F I, don't like static helper classes,.

A

Well, we are in so yeah we already have. We do already have hash algorithm name and we already have hash algorithm. The only weirdness that would be on hash algorithm is that that would know so. Okay, so I, don't like it on on hash algorithm, because there are types that derive from it, and people really like calling static methods on derive types.

A

So someone will be like they called I mean we can use new to mask it at the lower levels but like if you you could then do h max off if 256 dot, hash, which looks like you're doing H max sha-256, but you're not you're, doing whatever the hash algorithm parameter is.

E

It also means that we, we would only be limited to the types that we know about and if you'd have, if you wanted to define your own my custom hash type, you wouldn't be able to use the same pattern we did. You would have to literally define your own static helpers as we're doing here. So.

A

E

Is probably the pattern we should just have people do yeah.

A

Yeah, it's a reasonable, very reasonable question, but I think that the static or the derived type static invocation would lead to a lot of potential confusion and that the other. This is a pattern that we can just follow and that helper method is easy enough to write.

E

It does this issue also concern itself with HVAC or just with non key toshing I did.

A

It with just non keyed hashing, because once you have an H Mak, you have the key and now just create the instance.

E

Okay, final question done: yeah I assume. If you pass a null byte array into the hash method, it throws an argument, no exception.

E

A

E

What if you pass a null read-only span of byte as opposed to an empty read-only span of byte? That should work so.

F

E

We did have an issue a few months ago, where somebody was surprised that uh that they changed the call sites on some of the crypto API is to go from byte array to read-only Savannah byte and they started seeing incorrect behaviors because where they should have received an exception telling them you're doing something wrong. Instead, we computed cryptographic cryptographic computations over empty inputs which broke the security of their application right.

I

But we didn't, we decide we're we're not gonna make a distinction between read-only spans created from empty versus no I. Think that's a we just shouldn't be doing that right.

A

So the yes, the the one that takes a byte will throw in a any of you, called the managed one and gave it nonsense. The span one while I have written a few places that need to make a distinction, because the underlying spec says there's the distinction between Nolan empty.

A

The the general rule we have is the span from Nolan's span from empty are the same span and.

E

Yeah yeah I agree with that rule and I think we should follow that rule but I'm. What I'm, trying to figure out I guess is. Is there a pit of failure here, potentially where a developer has? It is calling the API there incorrectly passing us null not expecting us to treat it as a zero length buffer like is there? Is there something we can do to this API to make it more apparent that, like hey developer, we really anticipated you to passing data here if.

A

They have a byte array, then they'll still get the argument or they still get the null because, though, if they have a byte array, they'll call the byte array overload and they'll get the exception if they've already turned it into a span that was on them. Okay, sure.

B

I

B

Like we're being consistent with how we're using spend elsewhere and the B scale, yeah.

E

So I guess one one issue, then, is: if I have a byte array and I call the hash overload that takes both a source and a destination that is implicitly converted to span, which means that that could mask Knolls as a problem.

A

Yeah but you're less I mean it's more advanced you're, probably yeah I mean there is a. There is a border distinction between the the byte returning 1 and the destination writing 1.

A

Okay, we can I mean we can add other overloads to the destination. Writing one I, just think that we classically haven't, except for we don't waistband of char, with string.

E

Okay, that's why I just wanted to throw it out there. It sounds like consensuses, it's not a realistic problem. Okay,.

A

All right so as the.

D

A

The expanded definition.

A

So as API review, I will mark it as approved then as feature owner punt to future. Yes, Kevin you're still allowed to do it right now. If you want to.

A

That was, it was important to have the API review discussion in case someone wanted to get it in five, we're not committing to the feature in five for anyone who did differently in triage yesterday, all right, inna member attribute constructor.

A

We have a value property and they want to make a value, constructor parameter.

B

Are there any other properties on this type that are workers well well,.

A

The generally in an attribute, if something's, a settable property or if it's in the constructor, it's a get only property. And if it's a settable property, it's not in the constructor.

A

But this would be because this is obviously back forming something. But.

E

Is there a subject matter expert for this issue.

A

Says serialization I think that's you! Oh.

E

No I'm, just the guy, removing it I think you're, the Arianna I hope not I.

A

Feel like SRS is a specialized subcategory of system run time.

A

Paolo emo hello, oh.

E

It's datacontract serialization, okay,.

A

All right so emo, maybe you'll, have some good opinions here. The proposal on the screen is, we have the existing attribute in a member attribute. It takes value as a get set property and they want to add a constructor overload to accelerate setting the value.

A

But I feel like we generally don't add those accelerators once the property has a setter I.

K

Mean it is a weird design, though, all right like I, don't know why they did the public in the first place, but I.

A

Agree so we can say that we're fixing it to what they should have done now. I.

K

Mean the question really is like what other properties does operate? That does this thing here is.

E

Values, value cedra was ever called.

A

Which must be good only yeah.

E

Because, presumably setting it to null actually is interesting compared with it having its default value of null. So.

K

What does it mean if you apply the attribute of out passing anything to the constructor and not setting value? No.

A

K

Doesn't mean there's no value now or does it just mean take the underlying value I?

K

That seems odd, and this.

D

E

Applies to what.

K

Data contracts, realization or.

E

Yeah so I'm I'm, looking through the full framework sources, and the only place that is value set explicitly has ever checked is ever read, is to ensure that you did not set the value property to null or an empty string which presumably, the values that are could have chucked, which.

A

Would be the same as using the default constructor.

E

No, so they they say if you set the one check that they perform is. If you have set the value property, you did not set it to nowhere empty, but.

F

If you have never.

E

Set the value property, they have a separate code path for that.

A

Where it's, what figuring out value based on what you applied it to.

A

D

The question is, what does.

A

The enum member with you know, open square bracket, you know member close square bracket, they have a default constructor. It does something.

A

The we classically, if I recall, correctly, don't overload attributes which is a redundant way of saying we don't back spill the settable properties into constructor parameters.

K

So everybody maybe don't overload them and we certainly have other constructors all.

A

Right I feel like we generally try not to but I mean.

K

That is true because we tried to keep them simple, but, like I mean there are plenty examples of a I've done that, even though I can't think of one right now obviously had this argument with Jana, I think and I pointed out, there's some very common one that do that.

A

Fair enough I mean like aside from the fact that it's an attribute, and that means that now we would have a constructor parameter that corresponds to a set of all property.

K

Yeah I have no idea where that went with the property, because if you just look at the docs pages, they always set the property, which seems.

E

Yeah it again looking through the source code, it looks like applying the attribute and not do and not setting the value. Property does have a small side effect, but it's not something that it looks like most people would ever run into. It looks like it looks like setting the property I'm, sorry setting the attribute and not setting the value property, simply marks the thing as I am allowed to be serialized, even if, for whatever reason, WCF decides that this wasn't originally serializable Incirlik.

K

Looking at the dock page for the details, it seems like not passing in any value is legal and it means it's just Co lized as the string of the inner member. But then what happens? If you don't put in a member on the inner member, does it blows up at one time and says you can't sue, doesn't really know.

E

They use a they use, the name of the member it. Basically, it's the same as not setting a value property at all, but there is an additional check, such that's. What is this there? There is an additional check that says if, if the field is not marked as serializable, but it has a value that is annotated with this attribute serialize, it any way.

A

Which lets you have like three members of an enum be serializable with the others, failing I, guess, I guess.

E

K

Like it seemed like they liked the route documentation has an example where one field does not have the attribute, and it says not, this realisable enumeration and like it seems like if you do that it should blow up. Oh, it should block this realization at least or something I don't know like. It seems highly confusing to me in general. I would assume if I mark the entire enormous data contract and all wages in that enormous realises over the fall, and then any member is only used to customize what it is utilized under.

K

That would have made sense to me design wise, but like I, wasn't there when they decided I mean in principle like looking at the docs. It seems there is a distinction between setting and not setting, and so we need to preserve that, but overloading the constructor seems reasonable. The only downside of doing that after the fact is, if people inspect the attribute using the static analysis, thingy right, whether don't actually instantiate the attribute but I just look at the attribute construction that caller has to handle.

K

Now the constructor passed a string and value set with to the string. Well, let's now basically two ways you can configure the attribute and unless you instantiate the attribute you you actually have to handle that in code right. If the item is just instantiated, then it's fine, because the constructor will set it and then, if they ask for the value it will have the right radio does that make sense or all.

A

Right so the the complication, so the reason yeah, because I found the guidelines. Finally I think my cursor was on the page passed them so I had to search the whole book is: do not provide constructor parameters to initialize properties, corresponding to optional arguments. Aka, don't have a setter and a constructor parameter, and then the other one is avoid overloading custom attribute constructors. So yeah.

K

And I think that we.

A

Have a do not for this scenario, which is probably the metadata reader, gets very hard yeah.

K

That's why I'm saying this is custom, attribute data thing in reflection, why the base it just tells you what the signature looked like and yeah that one has to have handle in all the two cases: I'm, not sure that applies to data context. Realization, I think there just need to instantiate the attribute so Rapala just work with atlas, but that's why it would be careful with doing it after there to bhadori shipped.

A

So it sounds like this: one is a no I mean.

K

E

Also, keep in mind that if we did, if we did decide, we want to approve this in the future, we should get the area owners to comment the area owners. So far, it doesn't look like if you even looked at the issue yeah, unless it is me and Joe, which means no one. No, it's a it's Hong in Steve, okay, so.

K

Do we actually care about that? A concerted addition? Is it something we actively invest in Norris, there's, also something that we don't recommend people using. We.

E

We actually have unit tests for this exact scenario.

K

For what for what scenario, the fact that there is no constructor that takes the string or the.

E

Fact that you construct it and never set the value property.

E

In fact, system text Jason also honors a property. It looks like she's.

K

Okay yeah, unless the area I want to feels like this is a cool feature: I personally, don't care one way or the other I.

K

Mean on the surface, it looks reasonable right, it seems redundant to say value, equals yep.

A

But apparently they thought it was important to have to make a distinction between the default one and having specified values. So I would.

D

Have done it as.

A

A default and there I would have done it with overloading without a settable property, I think but yeah.

E

So let me also, let me let me clarify point I just made system text. Jason actually does not honor this. The unit test I'm looking at is for Newton soft Jason.

K

K

The cool thing mechanical keyboards that can hear when it's happening.

A

All right and it's not blocking so just close I, don't say no on yeah.

K

You just close it I mean be nice in the comment, but doesn't that that seems to be the process. Yeah.

A

K

Label for that right, I guess.

A

Yeah normal does anyone have any objections to my words here.

E

You have the best words, the biggest words.

K

That goes into the show notes are today.

K

A

Yeah we're actually we're officially reviewing a milestone: 5 ready for review, nope non-blocking we're sort of caught up evil, I, don't know where the proposal isn't. This I guess up here.

K

F1 has a method: I have to have a similar functionality. It's just an API.

K

Yeah I think I know what the center here is, but I don't I mean this is one of those areas like the area owner market ready for review, who marked ready preview Eric.

A

K

In principle, I think the it doesn't knock look bad. It seems in line with what is the route I begin, its container configuration areas yeah.

A

Yeah configuration container.

K

It has yeah, it has all these things called with blah blah blah.

K

What's funny is that it has this thing called with part.

A

Do you not have meth in the database.

K

No I do container.

E

Configuration yeah.

K

It's container configuration it's system, composition done hosting I'm, looking at it right now and actually and okay, maybe I, don't know.

K

It's a bit weird that, like naming wise I, would have expected something like whiff part that seems more consistent.

C

D

K

Yeah I guess: if Nick commented on it and he proposed it, I think that's probably the right terminology, then.

A

K

Basically, the so the one thing with math is that basically the way this works is they call it exports? Because, let's say you have a time called. You know my fancy screenwriter and usually you usually don't want to export it under that type name and usually export it. As an ID text, writer or you know, a text whatever the base type is that you usually write from, and so the idea would be when you save of instance, you would pass in the instance, and you would also specify what is the.

K

What is the the the actual thing you want to export it under for would say you know with instance, text writer, then you would say new my fancy text right all right and then the other overloads here. The text rings are just because the way math works under the government should note that mefi two works under this cover the same way, but it may be one all exports effectively just strings and so usually they're just a fully-qualified type name.

K

So I am, I put basic over the proposal that nick has, which was the simple one right.

K

He basically had the live export of tea, tea, exported instance, contract name and air as optional, and then he just has the equivalent of that. That is dynamically typed right so that you don't have to do generics. So that seems reasonable. To me, I mean.

A

The only and maybe this is inherent to why you would want to do it. The only thing that really bothers me on this one is.

A

It's set up for generic inference, but that's exactly what you don't want if you're calling it.

J

A

If you're calling it you're wanting to change the wire type of.

A

The instance but I don't there's not really a good way to.

K

A

I'll gently say with instance new example. Then it's gonna see realize it. As example, I mean certainly that I guess the method could throw if, if thing that get type is equal to D export or a type of T export, but.

K

It's not always the bug right so, for example, like I just recently bought an HP NetApp and build a DI system. It's the same pattern right, you say add what does it add in forgot? What the name is? Basically, you basically add a edit, add a service right and then you can type it under a different thing and that's the same pattern. You just have to specify the type name in that case. Well,.

A

But in that, but in that case, generic inference is fine of if you're, if you were instantiating the type that you want, then it works and if you didn't, then things explode mightily and if this will explode mightily, then fine and probably find anyways. It's just it's weird that the exact time that you want to use this method. You don't want generic inference to work.

K

It's true, although like yeah I mean it's hard to say right, like I mean if you actually have a service like in my case, that is not based on an abstraction that we just the service, because all that ones is just the global thing. I want to import I, actually don't have a base type. So in this case, Jerrica difference actually was helpful, but I would agree in general. You want to have it explicit type, they're, the.

A

Only time I can think of that you would want the type that it would be like I can think it would be OK for the type to match if what you're doing is taking a t export from somewhere and you're, saying always write this down as the base class and if someone act, if the base class wasn't Stan Schiphol, then then you could end up with I. Didn't give you a specialized form, just serialize it, as is so.

A

Maybe it makes sense, but it's you know it's an API that the only success the the ideal time you want to use it based on what understand? Is you don't want? Our generic inference is wrong. On the other hand, this does stop you from passing an nonsense combination of a type object and an instance. So right so it gives you some compiler sanity on your call. So yeah.

K

If you think I mean it seems like I mean, if you really cared I mean there would be a fine example of an analyzer about that. You would say, you probably didn't mean to use generic and friends here.

A

A

Yeah, because pretty much anytime, that you didn't specify the type name with this method would be probably a bug. So, okay, so you think the names look messy.

K

Yes, so I would I would just say: I would I would extract the exemption methods as Nick wrote them not what's in the proposal.

K

So let me just comment on this: I'd go with.

A

So the with export of tea, instead of with instance, yeah.

K

Base, it will be just instance, method on the container configuration they would return a new and actually don't know whether they returned this or a new instance, but basically the same signatures right, except that it's not extension, methods, I, think the terminology export makes more sense.

K

E

It seems to be consistent with other terminology already used by mouth just from reading the issue, description and.

K

I mean Nick boomer, wrote the whole thing right so like it, yeah I would just trust his gut on that. Okay.

A

But he had suggested it as extensions, and you think it should just be instance, methods well.

K

I think he worded extensions because he.

A

Wrote it in a demo project? Okay, yes, exactly.

K

So that was just more like so.

A

His was an FYI okay, I mean you could have just pushed the approve button there email, it's not like. It knows. Well,.

K

I know I just wanted I mean it's just my low-level way of sharing my screen right, fair.

E

Enough I I would probably change T to t export because it matches the API that's on with part which.

A

E

T part, instead of T and I'd, probably flip. These two arguments.

K

Sorry I can't see what are you doing right now? I agree.

E

Yeah type and object should be flipped because you want the type to be the first thing. So.

A

That's both the generic and non-generic yeah, then on generic, when the essentially the generic one, then it looks like you're calling the non generic one, but you have compiler safety instead of just whatever runtime behavior you get when it doesn't make sense.

K

Yeah, exactly that makes sense, because the other bomb wouldn't compile right. If, yes, then T possible.

K

A

So while on the one hand, it means that the first argument isn't always the exported instance, it's when you read it left to right. It's with export a type name than an instance, then an optional contract name.

A

So we do have two optional parameters here. Does that mean that we want the simplified overloads that don't have optional parameters.

K

Good question: what doesn't matter do in general? Do they actually I think Murphy to was also gently? Overloaded, yes, I, don't see to optionals, hey.

A

I was just gonna say they probably predated us accepting optional parameters generally and I.

K

Mean that was done before we had optional guidance right.

A

K

Why most of the I mean his extension methods? He didn't care right. He just wanted to only type it once sure.

A

And just the just the guidance that we wrote in the book is once you have two or more, you should consider adding the simpler ones so that there's the view that doesn't show a bunch of square braces and intellisense.

K

Yeah I would probably just do one overload that just doesn't have any optional and then one that basically has yeah.

A

Yeah, that's the guidance.

K

Yeah, that seems, yeah I, guess yeah, it's fine hold on. Let's move window, I love the window. Sorry.

E

Judge just to confirm you're, saying, have an overload that contains neither of the optionals and have an overload that contains both correct. Yes,.

D

K

Then there's a bit nonsensical that you know effectively. If you only pass the first two they're logically, both applicable right, but the compiler would prefer the the actual overload. But AB AC allows you to pass in the metadata without passing in the contract name.

K

Which is better than having three littles.

A

Right and changing all the t's to T exports.

D

K

I did that already named them, call it I, don't piss the Lord again, there's I.

A

Missed one, if I'd been smart, I would have just copied export pasted that, instead of needing to highlight the T but right so that's I, think you and I have the same thing except mine, says: T export now.

K

Yeah be switch should swap the type names right if he didn't do that already. I did.

A

Oh yeah, so I I fixed, so it's type comma instance.

A

And then I guess we're doing.

A

Presumably the nullable annotations form of this is these are obviously nullable and the exported instance is not because exporting null is probably not good and the type is required. I.

K

Would agree with that I think it's not very annotated. If so, then yeah that's marked them.

H

K

Yeah exporting now it seems that I don't think it's legal.

A

All right, any other things, people see that's wrong here,.

A

Which will look exactly like emos thing and just in another bubble, yeah.

K

D

One more API approved.

K

Are you sweating by oldest to newest yeah, interesting, cuz, I think we did we already decide what to do for extend process.

B

K

Oh, you look at the fire I see! That's why yeah.

A

Things that are still things that are marked as five seem more than things that are floating so.

E

My pink David uh we'll see if he this year, oh.

A

Right and then on that previous one Oh add the note. This is that we also consider it valuable for a analyzer that finds when generic inference was used.

K

Okay, I am against this API. We should not do it.

E

You're trying to beat trying to beat you know: David Fowler's, I,.

K

I tried to start the conversation at a at an interesting angle. I guess I feel.

B

Like didn't, we review this API and reject it, and wait for David provide more. Let.

K

Me check yep. Yes,.

B

Feedback that causes it to be reopened, hey.

E

B

K

So that muslin is catch me at last time, so, like originally I I think originally I said we should do it and then people talked me out of it, and the argument was that this API will allocate on your behalf, which is surprising on the sequence, with our API and you're saying: there's other API is already they do that they.

H

All do that yeah yeah I, don't agree, but this API is super useful and the other ones do the same thing already.

K

Yeah to me, the and that's kind of what I said as well: I like to me it's kind of like what would the person do otherwise right and the argument was well loose use the other API, as do the stack, I love yourself, and if you don't like the stack, I looked and you know, go.

J

K

Or pull yourself whatever, but like dead I can see this being boilerplate right, inconvenient.

H

As hell, yeah and yeah I mean basically all the code has to do it at the call site. It's kind of the hard part right stock, how at some size bigger than that there right from the pool, and then you know three.

K

H

K

Do you have an extension method today that you basically wrote yourself that doesn't yep.

H

Well, what this does? No, not because you can't yeah yeah right.

A

I mean you could yes, in a span of temp space, that the caller used as a stack Alec that gets used different fits and yeah yeah.

H

And you'd always have to allocate you can do the it first and see the size was small enough. Who didn't do that right.

A

Makes your I mean.

H

A

Overflows more predictable, exactly right, I mean predictable, is better than non predictable.

E

K

Yeah I mean the question is like: if we do this like, if you only have an extension message, that's the same thing. The risk is low. My only concern is we give you this API and then everybody you update all your called sites and then two months later we get this. You know bang email that says. Oh, my god, we just request the cast with performance on a hot path that that's you know that would then come back to like we told you so kind of thing right.

K

Well, yeah I mean I, agree generally speaking that if an allocation happens really enough, then it probably doesn't matter it.

J

H

K

It's probably really cuz. He buffer solution are right.

H

And I think it I guess it depends on how long your tokens are, but mere may or may not matter, because they, because the alternative is use, place first and then you call Trevor to. But you still, you still need to do the whole, like slice, create, create a subsequence treat here and then call like try get span or whatever. This is all kind of painful right. So.

K

What is your general of like approach thing is something like you search for a semicolon, and then you say: okay now pass this stuff off to the header parser, and it's just the span that represents the head until the semicolon. Yeah.

H

Yeah, so I have been working a bunch of parsers to try to find the most optimal way to use secret Reader and the fastest thing to do is to gasp I'm keeping it keeping. The secrets around is good for not allocating, but it's really chunky I'm hard to know to use right. So you kind of want to find the block of memory you want quickly and then turn it to a span and their purse yeah.

A

H

A

I mean, but you know just speaking the other side of it, so we make sure that we're.

D

A

The decision, if you're using the sequences, it seems like you're generally trying to avoid allocations and be willing to write more code to avoid doing allocations and while a one, T delimiter is probably or means that you'll always get a span allocated I mean I. Guess that's not true. If it was found three pages later, you're gonna allocate a copy of three pages, and but something like the PIM footer, wherein certificate plus ten hyphens is I'm gonna, just say like 18 characters.

A

Long, that's even itself likely just to now be complicatedly on a on a straddle and so you're.

D

A

Up doing a lot of allocations and a thing that you've already written a bunch of code to not allocate, but if that's what all the other convenience methods are doing, then yeah. That's what they're doing yeah.

H

And that one is interesting, so the pen for the pen thing we were scanning for the end sequence. In that case, you could easily avoid the allocation and I get this appraised.

H

I think this is more for smaller talkers, where you don't expect to actually end up having giant blob between the beginning and end up where you're searching so header header segments or are less likely to have that same issue, but I get the the the API generally can put you in a into a place where you you may and allocating a big thing, because you thought it was convenient to write.

A

Because if the behavior on everything that was these try read like try read to was it's false if it's not found or the first occurrence is or if it's not found in the first page right, if I will give you the span or it's not found in a single like that's where it gets all.

D

A

Starts becoming ambiguous, but really it's I'm telling you false, because someone needs to allocate something and or it's completely not found.

D

A

But yeah, if that's, if that's what they're doing it's, what they're doing the feedback before was it feels like this is just setting you up for you wrote a lot of code to avoid allocations, and maybe it's allocating every time he called it. Yeah.

H

I mean I am I was torn initially, but the alternative is this thing: either uses the pool and then your your fraught with let'em issues, because then, who calls returned and when do you call it? Do you always call it or do you call it if you span that that was returned is brought by a pool to race, it's all messy. So this seemed like the best of all worlds.

E

So, just just to clarify, like you, can write this as an extension method today, correct yeah.

D

E

There isn't try Ritu, that returns the sequence correct. Are you using that extension method today in your own internal stuff, yeah.

H

Look no somewhere else. Yes,.

E

So you anticipate that if we add this API, you would consume it pretty much immediately. Yeah.

H

H

Discussion both two things: I have a writing code to find the best way to use this parser, and then someone someone had tried to convert their own code using arrays to use listicles reader and they were complaining about a lack of APA like this I think. The proposal is for a get span from read-only sequence itself, and then we were playing with it, and then it turned out that this was the better API to have.

H

I still want the advance to that that I think I think that one got punted Bart was called set position. It was kind of awkward.

A

All right, so it sounds like, even though the API has a nonzero pit of failure.

A

People want it so.

E

It already exists, so you were saying you know: I do.

K

Really want to add the blue like this, though, like that seems the bool.

E

Yeah I mean all of the other try Ritu API Scott is just a limiter equals true. Okay,.

K

I have that okay, it's.

E

Good symmetry yeah I.

H

Mean if, if in the future, we can think of a way to pass an alligator or give it to the sequence reader. That would be an interesting compromise. We're in these methods, you give the reader an alligator or I.

J

Don't know a pool and then a locator is that it doesn't really help you with. Like stack, a look cuz. You can't even.

H

Oh I know I know to be honest: I.

K

Mean having an interaction force, take a look, would be nice, but I, don't know how it would do that. Like that, that's it's! The general problem that we have of these API is. Is that you'd like to have them policy free, so you can, but that makes them slightly less usable I.

E

Would actually push back against an accelerator for that, because it would make it extraordinarily common for people to write this inside of a loop and then stack overflow.

K

Well, it doesn't have to be trivial right. It just means like I know: it's not doable right. You basically have to bake that policy into every single API, but you buddy have to provide an API such that somebody can do stack a log on the call side and that that indirection makes the signatures fairly and we'll be sure.

E

And it it sounds like it can be done with a little bit of contortion, because it is possible to write an extension method. That is the API that's being proposed here. Yeah.

B

E

If you have that extension method, then you can build one that works with stock outlook under the covers yeah.

K

I guess you can I mean the problem is making that generic would be hard right.

K

E

A strongman you have try Ritu, that takes a destination span, which is your scratch buffer, basically, and try. Ritu also returns a span, which is here's where I actually put the data which could point to the scratch buffer, or it could just point to what was already in the read-only sequence and be copy free. That was.

H

One of the original designs we had for peak I think the arc, try feat, repassing, a buffer and then we're talking about. Maybe it should be lazy, so should be a fun within it doesn't work for stuck garlic.

E

It doesn't need to be a thunk like you can pass in a stack out.looked buffer as a scratch buffer. It's just.

H

G

E

To write to it, the.

B

May you always have to allocate it that's.

E

That that's okay, because you do it once at the beginning of the method, and it's generally pretty short like say: maybe 512 bytes, like the overhead of that's actually not too bad. You.

H

H

Make it faster.

E

Sure, if you want to.

K

Yeah I mean like yeah. You probably have to do that, but I mean other than that. It's just a pointer, decrement right, yeah.

F

K

That's really what that size boils down. To so I mean unless you have sec issues, I, don't think. Ladies Tagalog buys you much I, don't.

H

Know we we pay a lot for it for zeroing the for ziering the sweeter memory. It's actually super expensive in Castrol yeah.

K

That I'm saying I mean the energy thing I buy, but that's even the lazy case. You probably want to avoid that cost right. Yeah.

A

So the signature on this is all busted, because outs go after all other parameters, but it matches.

E

H

Mean you're all there when this happened right, the first one I did not review sequence. Reader.

K

You probably did but like there were too many durations I, don't remember that yeah.

E

It was, it was.

H

Sorry never mind yeah I think we have so far in this. It solves a very specific problem. Yeah.

A

I think I've seen pieces of it before but like the original version appeared I, don't know what I was on vacation or something.

K

Yeah this I, you know I, think Pamela and Kristoff, designed this by yelling at each other and I. Think that's see the outcome that we have here, but yeah I yeah. It's probably not the easiest API we have, but that's there's also why I don't care so much about accelerators here, because convenience all up here is not right. So.

A

Yeah, given that there's already these other two try read two's that are routing the span and allocating, and this one's just adding the one that takes say that the delimiter is not a single T, but a span of T all right seems: okay, there's a footnote that says that we also.

A

Declined this API for being a complicatedly, allocate this one. Let me think second, one doesn't allocate the first one does yeah how we rejected the read, only span version for the same reason. So, presumably we approved this one as sequence. Only.

H

This one was for a retraining exactly that name is better than my name, but I would want it. For the same reason, I was the other one. Yeah is like gonna allocate it randomly I like it, especially it kind small late too,.

H

And I do think in the future. Maybe what Levi said actually consider like different API that maybe do weird like taking a scratch. Piston return.

H

Stop that I mean that that that is that one is more fine I'm worried about people passing in like a pool, a pool de Reyes, I guess.

A

I'm reading into the pool to Ray, then using your data.

H

And returning it, it's.

A

Fine, you know the scope of when you're you're reading it yeah.

E

You're you're a leaf function as a leaf function. We don't really care about life. Tying our shoes, I.

H

Guess I get that for rent and return or pretty cheap. Isn't that bad.

A

Because if or something like header processing, if you're doing a header and then moving on and a header and moving on, you would just use the same scratch buffer over and over and over again you've you've read it you've done what you needed to do. You've either sent it to an output stream or done your logic and now you're. Moving on to the next one.

H

You're trying to not allocate at all right so so you're reading you weren't trying to copy you're just trying to get the span of what one one is well.

E

Mason header in the first place, yeah.

H

E

Funny thing about headers is the the fastest way to interpret them is actually going to be not to use sequence reader at all, but rather to say, like just read the first 4k of the incoming request put it into a single contiguous buffer and then parson yep.

H

See we don't copy, we we only copy if we need to, and since we have 4k before us anyway, so casual reads: 4k right and you have 4k chunks. So you read 4k at a time and you parse whatever you can't expand and then only when you get on the boundary. Do you go to this slow, slow path? So it's faster until you hit the edge and then you go slow and in.

E

These cases, you wouldn't be using sequence reader at all, you would just be using utf-8 parse or what I really we.

H

Use serial to get the extents of like the line, for example, and then we parse this bad from there quints.

E

Reader can't operate over spam sequence. Reader requires an array. Then it.

H

E

H

E

Sorry I thought you were saying you you, you wrapped a sequence. Reader.

H

E

Existing span: okay, yeah.

H

H

Think the goal of this type is to go from chunky to more fine-grained, to zoom it and then parts.

E

A

So the try read to that is just adding like because the existing try read to s that our outing spans already have the allocating problem. This one potentially makes it worse because the delimiter itself can now be straddling a page boundary, but.

H

So I think I think in the future. If we do choose to solve it, maybe the scratch buffer. Is there right, API design? For that one sure.

E

I mean the implementation of this won't be too bad. The implementation of this can be written as an extension method in like four lines of code. So whatever call the sequence one and call the sequence, one is the sequence of single span return, its unread span, otherwise call to array and return up, yep.

A

Of course, then you're using the the sequence through the return, which is I, think what it's trying to avoid, but still that's implemented. That's plumbing details now yep, okay. So since this issue called out revisiting this, one I still think that new API shouldn't make the mistakes of the past and that we should continue like that was we already had the overloads like we're, adding an overload to a thing that already has a problem. I I really think that we shouldn't be adding more pizza. Failure to this type I had.

H

A scenario here so late, this one was because I was it for it specifically. This one was about reading like was it like prefix based, let's like left, base prefix systems Trey recount like this came from Trey.

B

It's right here.

H

Yeah today's Apple is a is a prefix length. Payload and the code you want to write is try read length from the bite from the byte array, so you read like a big-endian integer, and then you want to read that many bytes from the sequence by the span right.

A

So in your proposal, you only used it as a sequence: yeah Jesus.

H

What do they do? Yeah, but yeah I mean yeah again.

A

Like in this exact.

H

Try this exact.

A

Scenario right, like you're using it as a rate, read string and you're like oh I, can easily convert a read. Only span of char or we'd only span a bite into a string by going through encoding, so I'm gonna use, try, read whatever the past in length was out to a span. It's like okay! Well, it turns out this was Warren piece. So it's taking your sequence of 900 pages of of 10k memory and okay. Let's go build one giant contiguous buffer to pass it in. That's probably not.

H

A

You wanted so the.

H

Thing is sequence: reader is inconvenient to use and sorry read. Only a sequence is inconvenient to use and sequence. Peter was the solution for that. So the question becomes like when you can narrow down your search to find the set of bytes you want. What do you do next so in in this case? Right, let's say: I have this buffer if I want to parse it further I know how to create a new sequence Peter over that stub buffer, all right, so it's C.

H

So there is this edge where you kind of fall off a cliff and spans are easy because spans are contiguous and everything supports them. When you have a sequence: you're trying tintin Maryland you're, trying to narrow the search to get a span or a subsequence and then do more parsing right.

A

Well, I mean this the easy you know easy answer: yeah.

H

A

You get, the sequence, is payload dot. Give me the span if you're a single span, because I know it's there somewhere like it's. Probably you know a couple lines of code. You get the single span if it's a single span or two array, which is exactly what the overloads going to do. But you're going to see the calls to to array in your code, yeah.

H

But that could so that we have that code in Castroville right that code is everywhere. Everyone has that get to to buffer or not I'll, be even thought about adding a method to sequence. To read. Only sequence I was like to array me to erase sort of possibly to array late. That might be a better for this stuff because they would give you a spam back, but it would give you a span that maybe it's a copier, maybe isn't a copy, and it has the same problem. But no it's on the core type right.

A

H

You would end.

A

Up with you know, two contiguous memory with an optional, allocator or but.

H

Yeah now we're.

A

H

Passing alligators.

A

H

Like what is the thing you're returning when you call try read and you get you get some sub payload by you're, gonna, further parse, what you do from there, because I I could create a second a second secret reader and do more parsing and the purpose garbage. Just because this thing is heavy.

A

H

A

Right but you can just say if you know reader, try read and then you just immediately do payload to array and yeah you're, making copies and small things. But that's if you want the easier looking code instead of the better performance like calling the one that outs of span is worse performance. Well,.

H

That's that's the key right so like for the most part, if you use pipelines and use this, it won't allocate unless you hit a boundary which push by default is, is how bigger the the third period I think it's four K at least my default.

H

E

That is the scenario reading small tokens from the payload, because that's what's in the example here like. If, if the scenario is to read small tokens like you know, 32-bit integers from the payload, then we don't need the try read that out to span. But the try read that, as a sequence would be useful.

A

I, don't I think the point.

E

Is if it's small.

A

You're, probably not going to get an allocation and in the rare case that the four.

B

E

A

A boundary you're like whatever just do it the problem.

E

Is I guess what I'm doing is I'm thinking of like what would the implementation of try raid big-endian look like like the implementation would say, take the unread span. Try calling binary, primitives, try, read, know.

D

E

Succeeds return it it. It was slice first slice, four bytes. No, you don't need to do that. Binary primitives doesn't require that I know.

H

E

You would slice first, no, you wouldn't why not, because if you get the unread, the unread first spam from whatever is left in the sequence, binary primitives will attempt to read four bytes from it. If it returns false, then you say: try read that returns the out sequence. If that returns false fail, see.

H

If that returns.

E

True and binary primitives.

H

This isn't that this is basically trying to a recent let first and then it tries to read that many bytes come in sequence, right.

E

I'm only talking about the imagination.

H

E

Try reaping onion.

H

I say: okay, okay,.

E

The implementation of that I wouldn't use the first overload, but I would use a second, oh.

E

A

It would internally be yeah info yeah. Well, really, it would just probably take sequence dot. Give me. The current span is its length, bigger than four yay past that too yeah. So it's never gonna call. Try read count anyway. It's just gonna. Look at the next at the current position: zero yeah.

E

It would call try, read count only and they slope out yeah.

A

Yeah my so my personal feeling here is that this, the one that out sauce pan should be deferred until the allocator problem is solved and that that should be true for all new all net new method groups on sequence, reader.

A

B

Kind of agree I mean.

H

B

I think someone did a did the math and kind of calculated the probability of allocating based on buffer size, and it was amazing, very, very small, I- think it was like in in the single digit percentages and for most buffer sizes right.

A

But if the thing that you're doing here with parse is reading a bestir like, while the average value of that length is going to be 60, the the big theta of that is pages and pages that you're now concatenated together into a single temporary array and a thing that looks like it's not allocating.

B

Yeah, that's so.

A

If you added the span one that I would have it do something like throw as soon as you pass in the bigger than a thousand and I.

H

Mean you could I mean if you want to throw.

H

A

H

Tryna, try read yeah.

A

You call tried, read count or sorry: I tried to read exact of more than okay, it's like out to a span and it's like nope. That's too big I refused.

H

To do it, I mean really the performance, he thought it something huge, but.

A

Like whatever the number like I would I would.

H

Want it to be throwing.

A

To tell me you're doing this wrong like yeah, and then we would debate on that, and so that's why I think we should wait until allocator is solved of what do you want me to do when I need to when I can't give it to you from existing contiguous memory by.

H

The way like that either right, but nothing nothing solve this just so. We know like at some point: you're gonna have an issue with with a single contiguous buffer and even the pool. Even the array pool masses are two Meg's right for buffer right. So at some point, you're gonna need chunks and the question the only question becomes like what is our model for doing this kind of computation? Do you want to get it? Yes?

H

Do we recommend getting a second stickers reader from this sequence, I'm doing for the parsing, or do you like I, don't know, I mean.

A

So in that model, let's assume that it takes a sequence, reader allocator and it has try, allocate of a size and you passed in one that refuses to go above 2 Meg, because that's what the array pool effectively limits itself to and then it would be well allocate returned false, so try reading exact returns false. So now you call the sequence 1, and now you decide what you want to do with it. You can call to array and let it allocate contiguously.

A

But if you're gonna do that, you would have just made your new out your allocator, be new array and yeah.

H

I guess what I'm saying is: we need to add more guys that take only sequence if we don't do things that that take span. If we don't want to expose things that return span here, then we have to expose ApS. That's it secret reader, I.

B

Don't know for me.

H

Right so the alternative becomes. We should explore spans or tell everyone to do to area which is bad, which is bad for performance, or we expose more things that understand the sequence and then do loops under the covers.

H

In very specific cases, then I.

A

Just like so my my feeling on it is I think you acknowledge that it would be nice if there was an alligator model where someone could control what happens when it crosses boundaries. Yes,.

H

A

And the moment we add one of those we would add an analyzer that says you called the one that didn't specify an alligator you're doing it wrong switch to the new one, because even.

I

If you want okay.

A

I

Know consider, but it certainly would like. There are plenty of situations where someone's using this thing and they don't care about a tiny allocation once in a blue moon and they're, not gonna bother right, fair.

H

I

H

It feels like the trade-off here to me is the BCL has a ton of spiny. If you guys for very small for reasonably small count. This works well, alright and allocates rarely, and we don't really understand the guidance for what happens when the cut, and so, if you end up having to call a thing that picks bands and you call to array you're in the same boat that you would have been if you called Tribune exact passing in the spine in the first place. Alright.

H

So it feels to me like I, don't know how to how to rationalize those two things well,.

A

Cuz, the other one would be try read exact that takes a span to fill and then now you've solved both prop like. Yes, it's always kinda.

D

A

Never sliced, but it's it's the. Where did where does the memory come from if I have to straddle pages? Oh this one I can I call that late. Then.

H

I'd call that try copy two or something like that are try, try.

A

H

A

Copy exact yeah, but it's I mean so we've yeah. Really this is you know, without being anyone who's ever used. The type I see.

D

A

What it is that it says it's trying to solve and then what the API is that are proposed and what they have to be doing and and either like I could understand that if they return false, if it's not an, if I can't give you a slice like try, read exact outing. The span you're like give me four K and it's like I, can't because it's not in the next page. There.

D

Is not I, don't.

A

Have that much left in this page, you need to do something more complicated now, and so I can understand that if it does that, but the ones that are just silently allocating feels like you do all this work to switch your stuff to sequences and now it's not helping as much definitely not helping as much as you think it is and that it may have invalidated all of your work right.

H

But but the trade-off there I think is you use this thing to find narrow narrower buffers, so you can avoid how often allocates like that is a trade-off, because the code complexity of trying to parts of sequence by itself, but is just harder. The opinion did depending on how the parser, obviously, but but we basically have have chosen to not add this type everywhere in the VCL unspun. Is that type?

H

So the question is: what is the fastest wait and to narrow down this type to get a span for some subset of buffers, a bytes of T's, I? Guess.

B

Is the problem we're trying to work around here that we want to be able to pass in the temporary working space of some sort that we need, like a tribal return for know that no this? This didn't exist in the initial thing or no, your temporary space doesn't have enough room either.

H

You're trying to avoid copies in the very common case when, when you're also so, for example, like try, read, could be like I, don't know some, let prefixed a message right, and maybe your system has a setting that says my mass buffer size is 4 k, r2 k, who knows and I have 4 K buffers I know for the most part it if I allocate it's gonna, be some some rare case right and Michael could even throw before calling try.

H

Read so I know that like if, if, if what the person passed me is over over that length, I can just throw an exception. I feel like. In that case you don't mind the occasional allocation and the code is simple. You call try, read you get a spam back and then you call a get string. The alternative is, they have to call a try, read you get a sequence and then do value is single span, dot, first, bad, otherwise, to array everywhere and.

A

What about try read exact count, comma span of T scratch, though let's please come up with a better name for that comma out, read only span of T value, which is fine, we can even say if the. If, for some reason you passed in a scratch space, that's less than count either that's an argument exception or then it'll allocate a new thing and probably go for an argument exception, because.

H

It breaks the flow right cuz. Imagine you're typing a bunch of like coding sequence, not to be clear, spawn somewhere else to get past in here, I mean.

A

H

A

It's valuable one like so just brainstorming, yeah.

H

A

We added in span out read-only span yeah, and then you wanted to say, but most people are gonna, be okay with the I I promise, I'm passing small numbers, and if you need to allocate a 16 byte array, I, don't care like.

D

A

Adding the other one I just it feels like. We need to give someone the way: that's not potentially creating big numbers uh as a surprise. I.

D

B

Seng is you you're constantly having to guard against these large numbers and things I, don't know if saving someone on this one specific API is going to be significantly helpful to anyone trying to write a robust parser, no.

H

But this at least this pattern, if we, if we want to see that instead of trying to figure out the allocator, we add overlook that take a scratch space. That is one alternative right, yeah.

A

This is really it I mean it's not so much about this method, it's about figuring out. How do we, how do we lessen the pit of failure and right and make a pattern going forward that that is sustainable right.

H

So yeah I like having an alternate. It's like the current, was allocated in umbria cases. If you want to avoid it in all cases, then you have to own the memory before you call the method, which is a little bit unfortunate, but there's no way to fix that I.

H

Think and I can either be a stack up memory or it can be a ripple or whatever, but you own the memory and the the return value is either a result of it either that memory or it's from is from the listicles reader itself, so you may copy buddies yeah. That seems like it would like a reasonable alternative stood at over Lowe's for all the try reads that have this scratch base, that takes bands and now we've kind of kind of indicate to you that the one that doesn't have this well allocates and.

B

That would be you are, you are saying like if it goes over this scratch space, we're returning false I, think.

A

If count exceeds.

B

A

Dot length you throw of you, obviously don't know what you're doing.

H

I guess it was throw yeah.

B

We can't have false meaning, not enough buffer space, and the thing doesn't exist in the original sequence. So.

H

Question does that make it hard to use.

E

You would only want to. You would only want to use that particular overload if you were reading a known small issue amount of data like say that I'm reading acquit from the stream, like I I, know for a fact, like stack allocated, sixteen bytes for instances no I'm going yeah.

A

So if it was, it was like the bestir of the thing that could potentially have been war in peace, but is probably a sentence.

A

You would have read the count if the count is less than some threshold I'm willing to deal with here then try read exact into possibly using some scratch space, otherwise, okay, this is really ugly. I really do want. The sequence now like this is this is too big. So.

H

Interesting so spicy questions that I have for you guys like. Is there any precedent for using different data types that, but first should maybe be int, may be a short to indicate that this is really for smaller buffer sizes? Is that is that a thing.

E

Really because int and long or the two datatype are the two numeric types that are easiest to use in the framework. Once you start dealing with short light s, bite like you, have to start casting everywhere and casting can mask problems because.

H

It's normally unchecked, so to me, try count feels like if you were to pass in a number that was over I. Don't know pick a big big number I likes something that seemed it would look like. It was really her performance. I could see throwing, but it's kind of Anna. We don't do it anywhere else, so it feels odd to do it. In this one, IP I, your purchase house gets worse yeah.

E

I, wouldn't the the whole scratch thing like I, wouldn't recommend that most people do that to be honest, because it it is, you know a a complexity that most people don't need to deal with, I'm, just thinking about cases where people need people really really really need to avoid allocations for whatever reason they're inside of a crazy hot code bath. At that point, maybe having these ancillary api's are useful. I mean.

A

Reading a page of tweets where you know that they're all bounded right, you'd start with your scratch space, that's big enough, and then you would just while processing them all read them off using the here. This is the temporary space. If you need to allocate something if I happen to straddle the page boundary right like so.

H

I think it can't throw that I think it has to like I, think this car space has to be a hint or maybe a best-effort, and if it, if you exceed the scratch base, it then does two array, so you can still like it. Maybe.

A

It's just a you controlled both arguments at that point, yeah.

H

I'm good exactly.

A

You could have seen wait. This is bigger than my scratch. Space called the one that doesn't take it or something, but it doesn't mean that the overloads a little weird, because one of them gets infinite scratch space and the other one gets finite scratch space, but yeah so I mean like but yeah the the something taking scratch space feels at least like enough of an out or nothing.

D

A

There's the subtle extra complexity suggested to you that, like you, should at least stop and think. Maybe this isn't what I want always I like.

H

It yeah we we had this in the discussion when we first it and then we kind of said yeah kind of complex cuz. This is the whole stock, hot dance that you can't do here. You can't hide it. Basically, the caller has to do colic and then, if it's too small dude the whole like right from a pool still think so you could only really use this if, if it didn't throw- and if you knew it was within some small smallish range first, a colic so I like this idea.

H

mmm So should we have an issue that just says so forget this one method, but for all of the methods no, and it does have to be in this release, if you don't get to it, but can we agree that we think this pattern could be used across the type I.

A

Think it's a pattern that could be easily used across sequence. Reader I, don't know that there would be a whole lot of other places. We would use it.

D

A

A lot of places are producing spans, but it's certainly a pattern. I've used in helper methods of I stack Alec in the collar pass a you need to do it. You probably need to do a complicated thing in the middle. Here's, your working space, yeah and then just reuse, the same working space over and over and over again.

A

Yeah, like I I, like that it suggests to someone there's more complexity going on than you may be, giving consideration to yeah.

B

H

Question like so I want this I want this API. What happens if the scratch space is too small, forget count if we did if we did like try, read spine taking in other other limiter, and you pass in scratch base. Do you think it would throw or do you think it would return? Faults may be false because they couldn't throw the scratch base.

H

A

H

That is right, yeah.

A

The answer from.

H

A

A try method if it returns false.

A

You should have a clear course of action and if the clear course of action from the one that span is to call the one that outs, a sequence, that's a clear course of action so false, like the we simplify it to say it's single purpose, but Christophe expanded it one says single recovery and that the what you would you wouldn't want to do more complicated things like you shouldn't then need to have effectively re-implemented the method yourself to figure out what what false meant and what you need to do about it.

A

So it's either one simple thing or we think that 99.9% of people are all going to write. This exact same thing in the in the false case: Oh go call the sequence one or or whatever, so the try, reading try read two saying false was I, couldn't find it or the first one I found I couldn't copy into the scratch space or wasn't already in a contiguous buffer like you would have the the sequence. 1 will will tell you if, like you, can you can move on there? It's like try read without allocating something false.

A

Oh try read, maybe with allocating something or try read, give me the sequence because I don't understand. Life is hard like.

H

Alright, let me propose one more thing, then, infinite time here right.

A

Because there are negative 18 minutes left so yeah.

H

If we, if we had the sequence overload and we had a way to get a span from a sequence, if we put all the onus on the read, only sequence, to get a span from that. Well, what would you call the method to return either?

H

Not a copy or a copy? But you have it be. The same kind of pattern is this because maybe then, then we could just say: okay get get a sequence and then, when you want to spy, you do sequence, stop blah blah blah dot method, and that could be an out for like this whole pattern.

H

A

Mean like you could have sequence dot, because the sequence will already tell you its length right. That's.

H

A

One not a Oh pages right now, so it can out a sequence and then you could have you know, sequence, you know dot either make contiguous or try make contiguous or whatever, and then you can give it scratch space for right to here. If you need to otherwise give me just the span slice that, like that, does the oh I'm I'm, just a single I'm, a sequence: wrapping a single span, I'll just give you the span, and otherwise you give me a place that I can copy to so.

H

You have to array which is the typical copy, but that's the that's.

A

H

Always allocate yeah, we could do eyes array or not.

A

Yeah well because I mean a thing that I do think is valuable, is if you end up with the like the the five T sequence and you're like I. Just I really want this to be in a contiguous thing, and I already have some scratch space here, so right, so, okay. So so it's sort of like two contiguous and then you could have a like just allocate the array.

A

If you have two or be the span, if you can and then you could have a two contiguous with scratch space that could easily throw of dude the scratch space, you gave me a smaller than count you. You should have known that.

H

You will find the get spam proposal in here.

K

To get spam from what you mean.

H

There's a bug somewhere that has I guess bad, a method on reading sequence that may be more at this file by way and find it yeah.

J

K

Yeah, it's like, unfortunately, not a good name to look for. Oh.

H

It's close: what's in five, oh yeah.

A

All right, unfortunately, youtube I have to run and.

D

A

The host of the YouTube stream, so it'll drop when I leave, but.

A

H

A

And I'm not doing anything with the try weed exact because I think we lost leave.

H

Absolutely yeah.

A

So good discussion see everybody later. That's.

K