.Net Foundation Design Reviews, 7 May 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: AsnReader/Writer update

Description

00:00:00 - Approved: AsnReader/Writer updates https://github.com/dotnet/apireviews/pull/119
00:57:21 - Approved: Extend default value handling in JsonSerializer https://github.com/dotnet/runtime/issues/35649#issuecomment-625418116
01:07:08 - Approved: Add mechanism to handle circular references when serializing https://github.com/dotnet/runtime/issues/30820#issuecomment-625444164

A

B

Have three different machine sitting in front of me right now? This is a shil.

A

Yeah because moment is clearly better and you're more efficiently development. When you have more machines, I can do three things at.

C

B

What what ends up happening is I type on this nice keyboard right in front.

D

B

And it goes to one of the three machines at random and it's always like an egg hunt to try to figure out where I actually typed.

B

Also orange juice or mimosa.

A

I was about to say something inappropriate by there. It's mimosa.

A

All right, so then, let's go through it Jimmy do you want to do us, the honor for the next two hours.

E

Try and limit it to one okay, so this is the changes from rounds. One and two of when we were talking about this feature, the biggest one from round two will be where got rid of the ASN value reader and replaced it with a bunch of static, helper methods. So a lot of this should be easy and just making sure we're happy with the shape, but then that the new ASN decoder class will be a little bit bigger. The biggest change is we talked about at the end of last Thursday's meeting the the namespace.

E

We seem to have landed on system that formats dais and one so there's that and we get rid of the in the contents accelerator for the meta tag and.

E

Oh and they're in the asm1 tag, class I had a tribe decode without a decode, so I added the decode.

A

Here, where did I see the decode all.

E

E

Right so originally, when this library was written, it was replacing a shim layer. We had over windows where, if it failed, we got a cryptographic exception, so everything was written to throw a cryptographic exception for compat, but since this is being designed on its own now, there's.

E

Anywhere that would have previously thrown a cryptographic exception. I've introduced a new exception for it I called a s and content exception.

E

It looks like an exception, pretty sure it has serializable and the serializable constructor, but they didn't show up in the ref gin.

A

After that, there should have, because the constructor would.

C

A

Been protected right so I'll.

E

I'll note that I should look for the serialization constructor yeah.

A

I mean the attribute, probably doesn't, because we generally don't show it to be it's in the disk. Unless you opt in and then it's usually a mess, because we generate every single thing. You usually also don't want.

A

So the thing I don't understand, you said for comparison's, but this thing is new, so why would did we ever go through calligraphic exception? So.

E

All of the code that is driving this got written in to whoa I think this is how long it's taken to have time to go through API review.

A

This Damien's donate core to, although not would be good see so.

E

As part of the cross-platform work, we've had to do more and more manual parts and to get consistent answers between Windows, Mac and Linux, and so this this coming out is just a full product. It was an evolution of that. So when I wrote up the original proposal or the original code, it had cryptographic exception when I brought it up for the previous proposal. I had cryptographic exception, which made sense because we were burying it under cryptography, but didn't we be lifted it out of cryptography cuz? It didn't really belong there right. So.

A

Make sense all right so.

E

Unless anybody has any objections to the name, ASN content exception, we can move on to mm-hmm.

B

E

A

E

A

Is meaningless supposedly? Why? Because they don't think you were placed as invade your v-dub as in the code, or did you I did.

E

Okay, so at the end of Part two, we got two and here's the ref struct version of the reader. With the you got two.

E

You had the comment of enjoying my conflict of the thing I proposed and the rules that I wrote in the book. I mean.

F

A

E

Is the output of that of from replaced the ref struct with a bunch of just static utility methods?

E

So it's the same as the ref struct, except the two constructor parameters are now two formal parameters on every method, and it doesn't save state I like this, because it reminds me of win32.

B

And that's my comfort, blanket I think.

A

Nobody ever said this about Ben 32.

E

Right so so yeah a SN decoder is the name. I came up with for replacing the value reader because every method on it is decoding.

E

The state information goes away, which is the first block of removed members, and then now we got will start seeing the shape of everything else. So reading coded value used to take no parameters because it just used the constructor State. Now the data which is now named source and rule set are now formal parameters to read encoded value and every every member on this class will out.

E

The bytes read value, and so, in this case of reading coded value, it's returning a net standard to o compatible range aka. It tells you where, to start and how long to go, the range type is not in that standard. Oh, so I can't use it here.

A

So I'm not seeing the range is that for the donor or a.

E

Reading coded value, you have two outs that are scrolling off the right-hand side. Oh I,.

A

See somebody turning you mean our permit is working that makes with this did.

B

You say that the only reason you're not using ranges, because it wasn't that standard compatible, correct, so another tick in the column, for we really should consider making it that standard compatible, because we have other projects that are running into the exact same issue. Where we're contorting the API surface, to work around the fact that the type doesn't exist at level. Yeah.

A

I was about to say that, like I generally, don't like, when we do unnatural acts just for down level compatibility at that point, I'd rather use the type.

A

I mean it's not I mean we have done this before right. We have just what is it hash? Go whatever we did. We put a single type in a newb.

A

I think that makes more sense to me like if I mean I, don't know whether it mean it did you try range, are you? Would you be happy with it or is it just like add I thought I thought about it, but then I merely dismissed, I didn't design. It.

E

I'm pretty sure I tried it and I got a compiler. If we added it to a package from that standard to a compatibility, then I would have been missing that reference, but I, don't my.

A

E

A

I know I meant more like did you design it and did you like the designer? Did you just consider it notice that it's not an S in a tool, and it just immediately gave up well.

E

If you see so reading coded value, we it reads a source and a rule set and then now, if you drag and everything behaves correctly, so we can see the out parameters, it's outing content, offset content, length and invites read, and so if range was available, it would out range content range and into bytes red because the sorry bytes consumed that is funny--I, that we have to yeah.

B

So it sounds like in this case it actually would be a natural out type if, if it were, one of those situations was like an take it or leave it then whatever, but.

E

Yeah right and that's I think the only place that range would have shown up, and so it's an offset in the length instead of a compacted range. So it doesn't really change much. Just if range was in that standard, oh I would have used it here cur. How does the offset work? Is it not at the beginning of source, so in Assen, or rather bur encoded data?

E

The first byte will be this: the first byte in the tag, which is the return value which and that could have been multiple bytes and then the lengths telling you how long the content is could have been multiple bytes and then and then now you are at the first byte of the content. So content offset will generally be two for a small value and maybe up to eight or.

E

Complicated payload and then content offset plus content length is can be, is usually the same as bytes consumed, but bytes consumed can sometimes be 2 bytes longer.

F

Okay and then so the range would be in the source that you passed in. This would be the content start, so both of these intz or a range would be relative.

E

F

Okay got it there.

E

Yeah they're, both relative to source Sub Zero.

E

Yeah, so if you wanted so you, if you wanted to look at the content bytes, you would call read encoded value and then you, where you already had the span, you would slice that to content offset comma content length.

F

And then that's our range would be even better because you would just say source so that range yeah.

E

And so then the try value or the try version I put the return value as the first out.

E

Because that seemed to be the most logical flow of the tryouts, the tag the content offset, the content length and the bytes consumed.

E

E

For things so read, boolean being the simplest of all the things that you can read, so you give it a source and a rule set, and then it will out the bytes consumed, which is how much you need to move forward to get to the next value and it it like all. The other read methods that are for a data type asterisk, take a defaulted knowable asn.1 tag, so the default or null being used, the universal type that appropriate to this data type and any.

E

If you had anything from the application context, specific or private tag universes, then you would just pass those right and.

A

Then all the ones who had try copy you don't have to try company anymore at all.

E

So in the second round we replaced all the try copies to try reads so that we didn't introduce new verbs. So that's I mean I could point right there, but it's sort of a third green line on your screen. Right now, try read bit string source.

A

E

A

I can see that.

E

And so because that one is a reading and copy it, it outs, it's local data, the unused bit count which is special to bit string and then bytes consumed, and then it's written, because that is the order that we said to do it for operation status methods.

E

And yeah, so that's gonna, be the general shape of everything here is took, so the value reader had the same shape as the reader apply the transformation and then the the same stuff here of the constructor parameters become the first two normal parameters, except for the try copy where destination or I follow the rule. The destination always follows source, so its source destination will set.

E

Yes, that's true, and.

E

Yeah there was the suggestion to use bit array as or it as a supported type for aimed bit list values. So that's on here as well, so renamed bit list yada-yada-yada returning a bit array again. If we, if this round, we want to say no actually we hate that type, don't add it. Then deleting things is certainly easy.

E

And then yeah, so that's going to be the case of everything until.

E

Strings the weird one strings are not weird on on this one, though I'll be a little weirder when we get to changes in the reader all right, so anybody have anything about the shape of the decoder here. I.

A

Mean it seems like it's actually just applying the feedback we said so it seems fine to me.

E

So if someone did want the ref struct version of the reader, it's very easy to write. On top of this. It's just mechanical name, everything the same thing and just call slice on bytes consumed.

B

Do you have on here like helper methods that return the length of the incoming tag like the number of bytes that was required to encode the length of the tag.

E

So asn.1 tag: if you already have a built tag, it can tell you how it has a method calculate encoded size which tells you how long, how long in bytes that tag.

B

Would be sure and there's a legal and asm1 to have an over long tag, corrupt.

E

Right tags have to use their minimum encoding, no matter what in encoding rules that is okay,.

B

The context behind that question is I was trying to figure out like if, if somebody actually could build a full-featured reader, that also keeps track of like every single byte. That's in the that's in the incoming stream and what was in it. Okay, great yeah,.

E

The only thing that's not exposed is specifically encoding and decoding links.

E

The multibyte compacted format, slash lacks format for BER, but you could reason about. Oh this tag would have been this long, so I know that bytes two through five were the length like you could make a painter for it, but and you would understand the length because we told you the content length, so it can all be semantically done just some of it's a little ugly.

E

So all right! So if we're happy with that, then we can move down to a s and reader.

E

E

So I there was some feedback on the design of, in particular for read UTC date, where we have a parameter of what's your two-digit year max and for reading, a set of where we have the option of skip sort validation that it would be nice to encode those as state in the reader. So, in addition to the simple constructor that takes the encoding rules, there's now a s and reader options, which is the encoding rules plus those two things.

E

So when we get to those the members that have extra inputs, they can read those from or they'll get their defaults from the encoding options or the simple constructor that takes just the encoding rules will use the the things that would have been the default values in a default parameter.

A

The option is immutable.

E

A

Any reason why you don't expose them them that.

E

Because I deconstruct it and like.

A

The container rule said and sorry.

E

Yeah we have to scroll down again because I did this like a week or two ago: I, don't remember if I made it mutable so that you could use easy initializer patterns because then it only featured hasn't shipped yet or not, but I know that definitely the reader. This constructor shreds the reader options into values that it once I think I did make it rewrite so that you could use object. Initializer, okay,.

A

So that's a player, don't expose it right because then it would be mutable after a or.

E

I have to rebuild it on every read: yeah.

B

So if a quick question on the option, so it looks like the auctions- is a struct which means that it needs to have some kind of default value for the ASN encoding rules. What would that default value be zero? Is that legal.

E

Zero is birth, so.

G

E

B

G

B

Deus and reader options you'll get BRR, so does that mean that the ASN reader constructor should default a ascetic coding rules to burn as well for the constructor that takes a rule set, but no options.

E

Uhh well, I mean the struct you'd have like it's, not a default construct, you'd have to say, II fall to base and reader options. You could say defaults of rule set. If you want I, don't expect anyone to use default to vase and reader options. The the type does have I mean it'll always have a parameterless constructor, because it's struct, but you.

B

E

The default flow is, you would say, new and intellisense would tell you, there's our constructor and it takes an ASN encoding rules. Okay,.

B

It just seemed odd to me that I could that I could call the constructor overloaded, ASN reader, passing a new AS, andrea options, parameterless constructor, so I don't have to specify the rules, but if I call the constructor that explicitly takes your rules, I have to specify it like that. That just seemed like a disconnected me, but if we don't expect anybody to ever call the parameter. Let's constructor on this truck and yeah yeah.

E

And it's certainly not an expectation that I would have okay.

A

You need to have a rule set right, always right. That's the idea. Yeah.

E

Is it least trying to make it look like you have to always have specified? One default, of course, is always a valid sim default of T is always a valid expression. Default is sometimes a valid expression right. So then, again on this one, we see the general shape of there used to be over loads of thing and thing. Taking a tag is the first parameter now it's thing, or only one thing which takes a tag as the last parameter defaulted to default, which means use the appropriate tag.

E

So read bitless! Oh you like turns a bit array of where's UTC time. Okay, so read set of and read UTC time, which are a little bit lower.

E

Keep going yeah.

A

Try to actually at least pretend that I'm reading it.

A

Omit my screen froze again here we go.

E

Okay, so read set of and read UCT time UTC time, which are both in the center of the screen right now. So these both have overloads still the simple overload, which just has the nullable tag. It will use the value that was specified from a us and reader options and the if you don't, if you want a specific two-digit year max for this call to read UTC time, then you can specify it as a formal parameter.

E

Similarly, with read set of it'll use the options the default is false, meaning it will validate sort order, validation if the encoding rules required sort or sets to be sorted correctly and to digit your max is whatever the right answer is for x.509 certificates, since max is inclusive. I think it's twenty forty, nine.

E

So those are the only two kind of new things that were introduced on the reader.

E

And then we could see the ASN reader option struct the.

B

Next thing, and just to confirm like for things like read sequence, you explicitly did not have tri overloads like.

E

Right so the only tries that exists, our try copies so try read, character. String can fail because you don't give it a big enough destination.

E

B

E

If you have the wrong tag,.

B

It will throw pure well, you have things like try radium at 32, try radio at 64 which out to in some lungs yes, and they return false. If it's out of range, they.

E

Throw if it's not okay, okay caught it.

A

I think we discussed it last time right. Miss mention tags are considered bugs.

B

So, if one more quick question on that done, so if, if I have a value, say, 10 billion encoded in the incoming ASM structure and I call try to try, read into 32 it's going to return false if it has a well-formed value of ten billion inside of it right right does. Is the internal state of the reader mutated by that operation? Like can I call, try, read you in 64 and expect that to succeed, or did it skip if the try returns.

E

False it's the same as if you didn't call it. Okay got it same with the try, read character, string, try, read bit string and try read octet string, it's you. You need to.

B

Try again, and if a method throws an exception, then like most types, we just don't make any guarantees about the internal state restructure at that point, if.

E

It throws an exception, because the tag is wrong. Then the state is not modified if it throws and if it throws an exception. I guess if it's overthrows an ASN content exception, the state is not modified. If it throws a weird exception, then it's in a weird state, but in general slicing is the is I mean. Actually I can tell you from the representation. It always calls the decoder if the decoder succeeds it slices. So in this case, with the way it's implemented, it either is not modified or or any exception.

E

We've resulted in not modifying state. The.

B

The reason I asked is I feel it some customer questions over the past few months about like hey. Is it safe to try catch this operation and then call back into the reader or back into the writer and I responded with normally when an instance method on a type throws an exception. We say that that instance is dead at that point normally, but if we want to make other guarantees here, we should just document them. That's.

A

Fair really have a saying the instance is dead at that one. Yeah.

B

If something throws an exception like by definition, something exceptional happened, no by.

E

Definition, the method could not complete his task. Yes, but listen.

B

E

Not say exceptional behavior.

B

If a method could not complete its task, there was no guarantee that the method left the instance in its original state.

E

Right I mean again, it all depends on if it's throwing it designed yeah.

D

E

Designed exception or hey in the middle of this operation that I may have been doing. Multi slicing wound right, so yeah.

A

In fairness, document I think of exception, anything I used to think of multi frames, and then you probably at the point where you're handing it you typically don't have access to the instance. So it's probably like yeah I, just couldn't think of a case where I have a thought about that.

B

Because the questions that I was fielding were with regard to things like stream reader, where, like I read from the stream and that I failed, I can't unring scream again, like I, have, by definition, mutated the state of the underlying stream. In order to attempt to do my work right history, their.

E

E

Okay, so for the reader options effectively I designed it as a class and then made it a struct, which is why, as a kid, only ruleset, that's a required constructor parameter, but then it's the struct.

E

We can make it all or fully loosey-goosey like just get set with no formal constructor, but I do like that. It, it tells you like it, gives you the big hint of like this is a thing that you should have an opinion about.

A

So, do you expect this type to grow.

E

A

Because if you have more things to it, then at some point it's like yeah I, probably should have been a class.

E

Yeah I mean we can we can go either way.

E

I mean I can say that because skip sort order, verification is a boolean and the ASN encoding rules only use two bits and UTC time to digit. Your max is limited to the number 10,000. This is actually only one end.

E

It's a fancy hint.

E

So it has a ways to go before it would exceed the 24 by heuristics.

A

A

And the case I mean, if am I at that point, easier right.

E

So is that dropped the drop the formal constructor and just make everything get set? Yeah.

A

I mean like it: what's the whole problem with the you know, I mean he knew up as an option as he didn't pass in anything and then well. How do I said anything now? It's like it just seems some usability standpoint, there's no downside of having setters and it usually results with better usability.

A

I mean we would only make the rule susceptible, probably right, because the other two favors are inferred I. Suppose.

E

The other two values are set upon. Oh sorry,.

A

I'm yeah. Okay, then, in that case I make the whoops it's edible and call it a day. I guess.

E

Okay, yeah, basically, it's I was trying to express that ASN encoding rules is still a thing that you should have an opinion about essentially saying default to base and reader options is probably not what you want, even though it may be. What you want I mean.

A

The only downside with not making it a constructor parameters, basically the validation issue right because if you don't, if you don't pass one in right now now you have another rule set, and now you pass that to somebody else.

A

But I guess an option is to general expectation is that the party like the options type themselves, are usually not in a position to validate anything, because usually you can set things independently and then the entry point is usually the thing that you actually pass the options to and then that that part is usually the one that validates that the combinations of options you passed makes sense right. I mean.

E

This constructor does validate, as do the the two sets. I mean sorry, the boolean set doesn't validate that you gave it the standard dotnet true, but the UTC time two digit, your max will throw if you give it a short.

A

E

Never out of range yeah.

A

For atomic weight is that make sense right about I mean usually it's like what options they use these combinations of things, and then certain combinations are nonsensical or whatever, and usually the type that exposes them is not the party that validates that the net sends I, don't think it will be bad to remove the required parameter either like it seems within the general flow of how options work, I mean.

E

And so the other another potential way of doing this is to remove the ASN encoding rules and instead of having the.

E

Constructor that takes an ASN encoding rules and then the alternative being a constructor that takes the options. They could be a constructor that takes encoding rules and either overload or default taking an options. So.

A

That you still.

E

That that would make it where you are more searching for options would have been better.

E

Yeah yeah, you missed it. Well, it's the beginning of the type. That's it that you're in the middle of the dif of.

E

I went too far. You searched freya-san encoding options, ASN reader options.

A

Options: yeah no I mean like where the rules themselves yeah. Oh okay, it's never mind. Okay, oh it's just you know, I see, yeah.

E

Right, so to keep the encoding rules as the decision someone has to make to the reader. Instead of having that rules built into the reader options, I could take them out of the reader options and then just make this or change the shape of the constructor overload to take a rule set comma and options and.

B

This is all in under the guise of like we really want the caller to care about the rules that correct it.

E

Feels important to me I can't extremely articulate why, but you.

A

Can't because them it might the default.

E

Yeah I mean the default you would get for. The reader like burr is the most lacks, so it it would be the most forgiving option, but that's not necessarily what you want.

A

It to mean the most logical way would be you leave the rules on on options and you remove the constructor that takes the rules themselves. So the only way you can pass stuff in is by options and options are optional, so you can just say: there's the default option. If you don't like them, you have to construct them and set them. You want the way you want. It.

E

A

E

Goes again against by my gut feeling of that, the rules are important, like I've kind of I want them to always be a formal parameter in my in my in my gut.

E

And so I left the ASN encoding rules constructor here, because most people won't care about the other two options.

E

E

E

And I think that there's only really two people in this part of the conversation, so we can probably work that out. Offline yeah.

A

I mean to me that I mean that's a player that feel right, I'm, just I'm, just judging it by how other readers work, but I guess most of the readers are not multi-format right, they're, usually single format, and then it's like well you're. Reading XML. You already know that. So you don't you don't have to specify that much, but yeah.

E

Kinda living, oh and.

B

I'm, actually just looking at what we did for Jason serializer right now, but I think they're, probably different enough. Where it doesn't, the pattern doesn't really match so.

E

Okay but I think I, my gut inclination to avoid the the default or that you defaulted into. A set of rules is make the second constructor take an encoding rules and a reader options and get rid of the encoding rules from the reader options. Right.

B

That makes sense to me then.

A

You would make the the options be forthright.

E

Either default or overload I, don't remember how we feel about constructors having defaults I.

A

Mean it's me: it's the same as any other method. Okay,.

B

Just out of curiosity area options being a struct instead of a class out is also intentional, because it was only one.

E

Inch big, okay,.

E

My my nod to low allocation sure.

A

I need to get a new computer also.

E

It meant that you didn't have to earn as instruct it has the bonus of no one should have the feeling that they get to change it after the fact and have that matter, because it was a strap. So it was passed by value. But that was a side point.

B

Yeah I'm from a performance perspective, I, don't really care I, just making sure that, but it was an essential design decision. So great.

E

Okay, so I think to the writer I guess the names of the two, the two things on the reader options were very verbose of skip, set, sort, order, verification and UTC time two-digit year max to express what rule they serve in the reader. As the in the overloads that let you override the value you gave them the option they use a they don't they're, not as verbose, because there's the context of the method name.

E

Then otherwise were at the writer.

E

E

All right, so there was a comment and a previous review of that it. Sometimes it's useful to drain one writer into another, so I added a copy too, because that's the name that we said that we wanted.

E

And essentially it's the same formula: all the overloads collapsed into one with a with a default and then.

E

ah There is a taking right, go ahead, Yuma so.

A

The rider does not actually hold on to the thing that rights to right.

A

You always pass in the the array where you're basically right to correct the.

E

Writer maintains its own buffer and then, when you, when it's done, you can try and code it into your buffer or there's encode, which is return me. A copy of the buffer.

A

I see but because.

F

The copy hauler works.

A

Because I mean I'm not seeing any data that you're passing to the constructor, so wondering how you actually copy anything.

E

So copy to would be it's the same as calling encode or.

E

Right encoded value of so destination that right encoded value of this dot, encode, except it avoids the intermediate erase.

A

I mean so, if today, when I call copy to what is what is actually what happens, it.

E

It takes it, writes the data that it would have returned in a new array to encode into the new writer, as if you called writing coded value. Oh okay,.

A

So, okay, so then the writer Basin has its own private data that basically magically grows, and you know writing to it and then copy to base. It just provides what you have said you have given it so far to the other writer right.

E

And we could, and instead of copy to, if it, if now you're, looking at it you're like wow. This feels really weird: it could be removed or an overload to write encoded value which would change the source and destination, and so it would be destination, dot. Writing coded value. This met up, you know, yeah and.

A

What happens when I call copy to and then followed by encode is encoding me giving me an empty array, or this is giving me the exact same data that you just wrote to copy to it.

E

Will give you the same data? The data is preserved and is still grow until you call reset, at which point it just goes back to the MTU internal buffer, that make sense.

A

E

C

A

C

Of the internal buffer, rather than no cob, do.

E

Not return spans that are not provided by the color.

C

So this is not intended to be used in any performance. Critical areas do not return spans. That.

E

Are not provided by the color? That's not always true Jerry, that isn't that that's the short rule written down, and it is a very firm belief that I have so I can't write. An API violates.

H

Plenty of places where you've got some static, cache data or something and it's perfectly reasonable to return a span, a fair.

E

Enough, but in this case it's gorillas raised, so if it returned the span, it could immediately be invalidated and it was originally using the array pool, which is the worst kind of span to return. That's fine.

H

I'm, just saying like that high-level rule is to.

E

Vary from summarizing.

H

E

I guess don't return a span unless it's yeah there's a different than less because we made it include the static array, return property thing. Yes,.

B

I think I'm was a query earlier, who mentioned the performance concerns? Yes, because there's a reset method on this writer, you could actually just have your own private pool of these or you could have a like a thread, static, writer and just reset it whenever you want to real to him that should that should work, just fine, so you'd still be able to use these high-level api's in a performance, critical scenario: yes, I mean.

C

B

Still have to copy at the end, but I'm copy for fast, though, a lot of huge concern are.

C

We expecting the buffer to be relatively small so that we wouldn't care about copying.

E

I mean it, it depends on what you're doing. Certainly, if you build a sign CMS with an embedded payload and what you've signed is a two gig file, then this is as a two gig buffer. If but, if you're, building a certificate that comes out to like three K, so the expected value is 4, K or less, but there is no upper bound and.

B

Then copy of 4 K is not going to be measurable in any real-world application. How.

A

Can you do it, but yeah I mean.

B

If you're, if you're generating asn.1 payloads like that, frequently inside your application, like what on earth are you doing I mean.

A

That's yeah the biggest that goes back to correspond right like. If you expect this to be in a hot loop, then, if the typical buffer size is 4 K, then that's probably not going to be acceptable right. If you expect this to happen like a thousand times, were you know in a given process lifetime, then it's probably going to be fine right, but even.

B

Even in a hot loop like a 4k man, coffee is really not gonna hurt. You, like processors, are really really really good at men, coffee, yeah and.

E

I mean essentially it's if it returned to span you're, probably copying that into your destination buffer anyway, so you would be doing the copy. So, instead, you just try and code into the destination buffer. You were gonna, you were gonna, do the copy into yeah.

A

I mean I'm generally agreeing with your point of like. If you have a global array like handing out spans, seems like a recipe to you know, shoot yourself very badly, so in that sense it is similar to lists of T very basically decided against it.

A

But like yeah I think the I think the conclusion is: if he ever really discover that this type needs to be used in the context of a super hard loop, then I, guess the shape of the type is not really amenable for doing that right. He would have to create a different type that does that I.

E

Mean we could easily extend it to this one, it's the caller would have to provide the buffer as a constructor parameter and then now it's theirs and there's.

A

Any reason we couldn't do the same thing in above the reader. Anybody have a bunch of statics that somebody can write something on top of.

E

So when we get to so most of the primitives that works, fine, the nested content types.

E

You can't you don't know what the I'll just assume, all bytes or all tags are one byte. You don't know what the second byte is until you know what the last byte is. That's.

A

Because right, yeah well.

E

It's made for very fast readers. This this is in coaching set is, is made for on the assumption that you read something more often than you write. It I see, yeah I, guess at the end, so everything's length, prepended and so for nested, payload types, you don't know what the length is until you're done.

C

E

Could have a make an ibuffer.

C

Writer, as a doctor, you.

E

So if we ever get the need, maybe but again I like I I, don't I I.

B

Feel like we're designing for a scenario that doesn't exist right, yeah.

A

It's fear like I'm, just trying to assess it because for the reader we basically said: okay, we have a static helper because we don't want to give people a rev struck beater, but if they ever need one that can easily build one right here, we are saying: oh no, we don't need that, and that seems just like the surface. At least you know inconsistent right. If we, if we believe we always have way more readers and writers and writers are not performance, critical, then I think that's fine and.

E

We have there ways that we could deal with it in the future of take a ibuffer writer, take a buffer, add static, encode methods for all of the primitive types and then you're on your own, or we figure something else out for the nested types but yeah there. There is growth here that doesn't require returning the span.

C

And that it's it's just a little weird, it's kind of a low-level API, but it's doing its own allocation and copying. It seems like a mismatch to me, but if it's not a critical path that we can add it later, if it's needed yeah.

E

It's more a really hard to get correct, it's just above the lowest of low level, so.

A

It also different somebody made my mom buy a low level right, I mean memory, stream is relatively low level and it does the exact same thing. I did basically happily allocates for you I think. As long as people are clear that that's what happens, but it kind of as I said earlier, like it seems, implied because you you don't give it anything right so till it has to sent the data somewhere right. It shouldn't be an entirely mystery to the caller that this thing allocates.

E

Okay, so if you scroll a little bit we'll see the first thing, that's returning a scoop so push octet string.

I

E

So all the push methods in this version all returned scope. Scope is a nested type inside the writer, it's a struct that is I disposable and has no other public members. And so what scope does is when you? So if you say using push octet string and when it when that using exits, it calls the pop-pop-pop octet string for you and then it's safely at impotent and other stuff.

E

So it's just letting your logical scopes and your lexical scopes be the same thing. Any.

A

Reason that you have a pub octet string, then.

E

I left it in case. For some reason, somebody wants to write the linear code and not the using statement.

A

A

Only concern is that this accidentally results in double pops right. That seems the.

E

If you manually call pop and then dispose the scope, it's a no op if you dispose the scope and then call pop it'll, throw a stack imbalanced exception. Well valid operation exception for you called pop. When it that's not the current context.

E

You I mean, unless you happen to have done, push off texturing toys, but.

E

Yeah so I mean I can I can hide the pop I? Can you know editor browsable hide the pop, but it I mean.

A

For symmetry I give you that it looks nicer to have both. It seems odd to only have one, but then again it seems yep a little bit odd to have both at the same time when people like am I expected to I mean it might not be clear to me, but that they push dispose and pop right, because that seems exactly to me that that disposing school pops right. That's the other thing, I guess it's! Maybe it's just the naming but yeah.

E

Yeah and maybe it's we want to leave it as void, push and come up with another verb for the scope, but I guess.

A

I didn't generally I, don't have a problem with the pattern where we use I disposable as a as a glue for pairing up methods.

A

My only concern is that people need to know that that is what's happening right and so that don't do things twice. I mean granted. It's only not a big deal, because if you do it twice, you immediately find out that something is not working right. So it's not like there's a sign unexpected back in your program that you will find very late right.

E

So leave it, as is I, won't.

A

Have a better idea of anybody else's also an opinion. The space.

B

Okay, I: don't think it's really hurting anything being there to be honest,.

E

Yeah like if, if somebody yeah I mean, as you said, emo somebody calls it so if they call it inside the using so if they usually using and they call it inside the pop or the disposed, is resilient to that. If they do it in the wrong order, then they'll almost immediately get an exception. As they were author, like they wrote the code, they tested the code. It blew up in their face and certainly the API Doc's that I've written on it says returns a thing that calls pop for you when it's disposed yeah.

B

And and I think if we look at the adjacent serializer or the jason writer as a pattern to follow here, I believe they did something similar where the there are multiple ways to kind of close an object, scope but, like jeremy said, you'll find it very early in testing. This isn't something that's going to bite you in production, unless you literally never tested for code.

A

Yeah that doesn't seem like the code that you write once and then just push and walk away right, yeah.

E

So Cheryl you wrote it and then sent it to a reader somewhere. That then told you they couldn't read the thing that you wrote.

E

And then yous going to the speck again and find the bike that you were missing, not that that's ever happened to me so all of the Pusha balls, which are octet string, sequence and set of return, the same opaque scope, value, sorry, the same, opaque scope type. Certainly it has a different value and.

E

It should be the last major actual change, but there's scope in its glory.

A

You I'm curious how the heck did you make it idempotent.

E

It it remembers enough state to be able to verify if it's already done, but.

A

It read only struck. How do you remember that you already did something.

E

Because it as a nested type, it has access to the writer state and so the reference type.

A

E

On to is the writer and the int that it's holding on to is: where was I in the state? Does it look like I'm? Still, there generally.

B

Speaking if struct I disposables need to be resilient against double dispose that state needs to be in whatever they're wrapping, not in this truck itself, because when you make a copy of the struct, the state, its duplicated I, see.

E

Hard to be it impotant if everything is a struct, but in this case I have a reference type that can hold the state for me, but rather I'm a reference type that I'm comparing the state against.

E

And then yeah tag class got bigger because it could.

A

Generate justification but yeah, oh, my god, you're literally in one of our segments. It's a I was not prepared for that.

A

So why did the values all change? Oh, they didn't you just just right to an end. I see: okay, yeah.

E

The feedback was, it doesn't, need a bite, so just be the standard int they got bigger because it could the.

B

The tag class on the tag number are basically Ord together and emitted into the stream. Is that how that works?.

E

If your tag number is less than 31, yes, modulo, the primitive slash constructed bit.

E

So they, if your tag numbers less than 31 than it or is the two things together and ORS together or is in the constructed bid if needed. If, if you hit 31, then life becomes complicated. Okay,.

B

But it sounds like it sounds like on the wire in general. This was meant to fit into a single octet in general. Yes,.

E

I the rid think until the it was the 2015 version of the spec or the one before that nothing had a number bigger than 30, and there are definitely well known operating systems that may or may not always assume that it fits in one bite.

B

E

Was that don't know anything fingers.

B

It's more like pointy necessity. Yes,.

A

Then let me just quickly check one thing: I think.

A

I, don't know what I did but I. For some reason, teams is now at 60% of utilization. There's something really fundamentally wrong.

B

Oh, so you use somehow reduced it from 90. Congratulations.

A

My machine is at 100%, but the rest is going down for OBS, so I know what's happening here, but every time I tie back and Italy I'm gradually there we seen a character by hand. Oh my God. Thank you.

E

So emo, since this is in designs and what we reviewed was a comment on designs, should I update the design to match what we did and then close. There came at the PR. Yes,.

A

Yeah, so I was about to say that so like I. Consider at this point we're done if they hasn't be, the one I was just considered it approved. The way we usually track designs is that we approve a particular snapshot of a design right and, if later on, changes need to be made as a separate PR and it's a separate approval right, that's I think the easiest way to deal with that and yeah I guess. Ideally, this would have not been a comment, but just a PR abut didn't guess.

A

In this case this would have been awful. Oh yeah right. It is what it is.

E

Yeah I mean that was that was the. When do we merge a design that where you asked and I said what do you want me to do so?

E

Okay, so I'll update the the content, that's there and then we'll deal with that.

A

All right so I think they owe me on the call, by the way.

A

I guess I'm perfect. Do you want to talk to this? One.

I

A

Sorry to put you on the spot, but I honestly, don't expect this to finish. I expected us to take two hours.

A

I

So last week we discussed this so this this issues to discuss have been ignored, default value, semantics for Jason serializer. Currently we have ignore no value but no way to apply this to value types. So if you scroll all the way to the last comments, there was some feedback last week regarding whether we should apply to this realization, so the gist is Steve and I spoke to James and cake shortly after the meeting on Friday and the feedback was, should we do any surprises. This realisation doesn't make sense, but we basically said hey.

I

We know sort of applied the semantics for Newton soft when we first designed ignore no values um so wanted to be sure that it actually needed to be applied to this realization. So, after speaking with him, couldn't really think of a good reason why we should which reflects some of the concerns from API review last week, and the sentiment is that most people probably do not notice that it's been applied or required for it to be applied. So um that's it.

I

I worked up dated the proposal to make it applicable it's a serialization and and fix up the concern with the ignore condition when all and went to how to specify that it's only applied toward to serialization.

I

So the first thing here is, we just ignore condition or it exists today we have never win, never win all and always so proposal here is to rename when know to, when writing default values, which is very clear that so, if the condition is set to this petite and would only ignore when writing novel is meaning that, if knowledge applied on the serialization would apply, it I mean if knowledge in period. um So if so yeah that's the first thing. So if we scored down to the next question, is no?

I

What does this look like globally? So whether we use this, you know globally as well see we have a default signal condition. Initially, a name is a default ignore condition. So this determines when to ignore default values, realization initialization, so we have the same same options and default to be just never meaning that would always he relies on T cell as default values.

I

So if once one wanted to per property, ignore a value on a sterilization, if it's default, then do it set it to when rights in default values. So this is nice because we're able to add when reading default values twenty-four values in the future, if we ever needed an insect receipt, so the final thing, what the proposal was to make it no default values obsolete so this yeah. So this initially was specific to no values. Now we have a more generalized thing that applies to transition and sterilization.

I

So perhaps we don't need anymore, especially since we're concerned about inter plane to this realization as well. um The I do plan for me personally would be to make these so put obsoletes for this release and then in.net six remove it since, like the library is fairly new and then you know, would have absolute eight, so they have one for each cycle and then wouldn't you know, advertise this option and just have people start toward the new global option yeah. So that's!

I

This is just a plain feedback from last week um to see. If, if any questions we.

E

I, don't know that you'll ever be able to actually remove it, but obsoleting. It isn't as much of a problem.

I

A

I think we said this last time. I think I'm also somewhat skeptical, that removal will ever be viable.

A

But um I think that design makes sense to me. Does anybody have any concerns with them? The only kind of where I was hoping we would land yeah.

E

I think the overall shape is reasonable. I wonder if the wind riding default values should not have the S on the end is there? It applies to this property and this property's default value or as the default, but it's for every particular property, its default value, but it's not. There is no notion of plural on a per property for default.

I

It's just weird: when it's applied globally to see single a fester, that's nothing.

E

Yeah but I think that the word or the fact that it says it's the default, ignore condition when writing default value. I, think that that, in my head, that feels grammatically accurate to get rid of es but having the.

B

S matches the S at the end of always.

E

Touche been always yes, always all the way right it down. I mean and yeah.

A

A mica level is low because I think either one would work, but I don't know. Do we have I'm.

E

Just thinking of it as sitting on an attribute sitting on a property of when writing default values, because the plural doesn't feel to me like it applies to this property. So it feels like it's a mode question instead of the value question, I think so my when I read it it. It feels like a state to me instead of a value comparison.

I

It could be interpreted to mean for every instance of this property across the object default values. For all those wastes is the one thing I guess, so maybe that doesn't make any sense so.

B

From chat, someone suggested uh what, if you just leave off the value suffix entirely.

A

Of inviting default.

A

I think I, like it.

I

Yeah I think that makes sense yeah.

A

Yeah actually like it, it's also shorter.

E

It feels a little incomplete to me, but not enough to I'll, say anymore about it.

B

So, thank you for it. For that suggestion, and.

E

E

Tedious nitpick thing in your description of the default, ignore condition for Jason serializer options. You say: it'll be an invalid operation exception. If you pass never sorry always that should probably be an argument. Exception.

E

Because you're you're not complaining about the state of the object, you're complaining about the value of the of the argument, so it should be an argument exception that makes sense to me so.

A

Where do you see that and.

E

In the XML doc for the.

A

Open yeah, yeah.

E

Because, presumably you're also throwing an argument exception if it's a if it's a value, that's not defined, so it would be, if used to gonna, throw the argument out of range if it's always or a value. That's not defined.

D

I

E

Otherwise, I think it looks good to me.

A

A

A

A

A

A

All right so Dan.

A

A

A

All right so I think there's one more.

A

Jason related right this one here.

A

Let's mark this blocking so that I'm going away.

J

J

Can you click on the second link that talks about the references dollar proposal, but.

A

This one here not.

J

The second one, this.

A

J

And the coming day.

A

Thanks teams, what is happening here? Why is it.

J

Another one they do, what the pretty one.

A

Yeah I, don't know why the other one popped over. No, it seems almost like it scrolled back. You.

J

Can scroll at the bottom.

A

One more time, Oh touching the keyword all right so just work. Alright,.

J

So, on top of the reference handling, but in a proposal I wanted to add a reference resolver, which is essentially something that we already have internally, that is just about is the backup references where we have all the where we keep all the objects. All the reference type objects that we have on here. We found on serialization or visualization, so.

J

So when they're, what I want to do is to expose reference reference resolve it as an abstract class similar to web nuta. So has, and a few reasons for this is that someone might want to define how they store their references. For example, someone wants to route the references and, and they can achieve that by use creating the resolver out of their civilization. The state say they can use, send it concurrently to the city later.

J

Another option will be to once want the new compare models that it is worth how to work on it gets exposed and that chimera mode exposes the state within the serialization.

J

They can access the reference resolver and interact with it with the with the methods proposed and also this. This proposal also addresses the previous video that we had about changing reference handling to reference handler used to make this not look like an enum and more like a class to sound like a class martin indyk.

J

So maybe you want to scroll a little bit down to the top to the examples proposal.

J

Basically, I'm just okay I'm, adding the reference resolver class with the port, with formatted studies to interact with with and in order to enable that I'm I'm also adding a constructor for the reference standard class that takes a takes a funk and in that phone you can use so to get together. You get the references to nuts root. You can use delegates the restaurants, our constructors, with a funk and in order to get it, get that reference to read you just create create an instance.

J

Respect instance would use an instance out of the educational setting just spell that frequently into the funk- and it's pretty much is and also keeping my acute myself when you call the reference counter this new reference under constructors, you will be the CID Elijah Williams pairs that you are perceptive references right.

J

um Is scroll down a little bit more just for examples.

A

Or this one or do you have samples? Nobody said yeah.

J

For example, here someone might want to use instead of int, which is where we have someone like one just to use good for data use. They use three three days which they use cast the string on the obligation into the GUI, then just keep it on manage on a dictionary and to meet the ID. They just create any good and return it on the get reference method, but.

H

Could you scroll back up for a second to the definition.

H

A little higher.

H

H

Reference handler has these two defaults yeah, which presumably have a built-in resolver I. Guess it's like reconstructing a reference handler with some funk that returned a new resolver. Is that right? Yes,.

A

Okay, so default doesn't preserve any references, presumably right, so you can probably override the.

F

A

In return now, and then preserve would do it via into that fear. Yes, interesting.

J

Well also, given my said before, does not create a result right because it is not going to share anything right well, so.

A

Can you tell me turn now from this funk and then basically, the behavior of default.

J

Well, we, you are not our Reverend Sun Ray shield, you are not implementing I mean you are not inert in four different colors you are. You shall have to unite from the front of shoulder now.

A

I know, but let me do it basically, if could I create a reference handler and passing in the funk and that funk returns now is that legal, because that would be logically.

J

A

Of default right, or is the idea that you have to do.

J

I know it's not legal yeah. It's is that we can throw like invalid operation. On that case, I see.

H

And presumably, the way, the way this is defined, the there's some internal rather than public method that actually invokes the funk yeah.

B

Okay, does the funk itself need to be provided in a kind of state, but the serializer like serializer options, something along those lines.

J

Okay, maybe that so.

B

It sounds like for each each time I call serialized. The serializer will invoke this funk to get a potentially fresh instance of the resolver correct. Yeah no state is passed into this call site. Does a resolver need any information whatsoever from the serializer in order to create a new instance of itself.

J

No, no it doesn't, it doesn't need any information for create anything I.

A

Guess the general idea is: if you need any state, you have to create the closure for yourself right. No.

B

But it's so creating the closure means that I'm capturing State at the time that I'm registering the resolver. What I was wondering is: is there any information specific to a given serializer instantiation? That needs to be floated in here, and it sounds like an insurer's. No well.

J

There was a consideration is to know to pass a pool just to know when you, if you call this from serialize or these here lies just you know what what diction, as you want to you want to initialize, but that I discussed that with Assam and things like that's like not really necessary and just a little bit more confusing. That necessary right also is.

I

We the options are locked after the fossilization, so this shouldn't change. If you had the options the first time, then you probably don't need it again for subsequent realizations I, don't think yep.

J

A

Other words, you could capture that yourself if you needed them during this realization.

D

So so one of the things that we also discussed offline was.

D

If you create the reference resolver.

D

Particular options that could cause you know essentially memory leak because the options are probably used throughout. You know the executable or you know, AB domain, so that may or may not be the intent to share the same references across everything right, but we don't have today a nice way to scope it just for a particular entry point. That's called you normally call it. You know: Jason serializer does serialize or deserialize there's no, not today any object to kind of hook on to that would save for the lifetime of only this serializer deserialize.

D

I want to preserve references, you're missing, that's missing. Today. um It was mentioned mm a little earlier that yeah I was working on a design that you we had that state thing, and you can maybe hook up to that, but just to note that if somebody decides to create you know their own ah static or something tied to an options class, they could get into trouble. So does.

H

That mean that at every reference, resolver needs to be thread safe.

H

D

Yeah, it depends how it's saving these. um Let's.

H

Go down to the good example below well,.

A

I, don't think it's right say right. The balance is specific for this particular D civilization, right yeah, because the.

D

A

Because the funk is invoked ratio that you see last ever the center.

B

Is shared not the.

A

Not the actual resolvers yeah.

B

It sounds like the provider has to be thread safe, but not the resolver itself, yeah.

H

And the provider is generally just new, whatever yeah.

J

Right yeah: well, there's no one wants to read them, but.

A

I mean I had the same at a similar feeling like to me. Raw funks are nice. If you really absolutely positively sure you never need anything more right as soon as you don't want to pass in anything else, you're kind of screwed with that design is why we'd have to I mean you're, not entirely screwed. You could have an overload to Reference Center right that takes funk off something else right, but.

A

Yeah, maybe it's good enough, I mean if you feel, if you feel it comfortably the only person in state it's fine, otherwise an abstract based type might be better well.

B

As as you said, adding an overload that takes the funk of state to handler would work right. We just deprecated this one and say call the new one. If you need to like it's, not the worst thing in the world. That's.

H

Not just just to confirm so every time the serializer does a serialization and deserialization. It invokes this funk. It gets a reference resolver, it uses it for that serialization deserialization instance, and that only and then it drops it on the floor.

H

Yes, sir true triples, I hope.

D

H

D

False I think I mean I me I'm David, you can reply, but we don't. We don't clear it at the end of a serialization yeah. Would you save that past the operation.

J

No, we don't say we pass the operation bill yeah, so the only reason to is to leave after civilization will be to have it out of the civilization all right. So, instead of just delegating the new that I get into the New York to the constructor, you just assume the instance.

B

So there is, there is a potential pit of failure here then, where right now we expect people to pass within the funk itself like open, closed paren arrow operator new by resolver. But if someone says bar resolver equals new, my resolver and then they passed open, paren, close paren arrow, that instance that I instantiated as part of a closure. They now have a memory link inside their application and they won't notice that memory leak until their app at some point in the future falls on the floor time or.

H

They hide it static instance and their delegate is just open. Paren, close paren arrow s, underscore blah yeah.

J

I get it I, get a bad to the proposal. I, don't know you want to scroll down a little bit more and I think what we can do, for that is to maybe create create.

J

Another, like a rule, is to check that we are all the people is, is actually opening to not doing a new operation on the phone.

J

Would that be good enough, I mean.

A

Maybe I'm not quite understanding it's like, but but if even if you were to return a static right, the funk is still scoped to this instance of options. But if I drop these options on the floor and I create I, create a new jason, see other options with different settings. Then I also don't have a memory leak, but I only have memiliki of the options themselves now also are effectively aesthetic correct.

G

E

You see realize you make one set of options and you and you serialize multiple times. This collection is just growing and growing and growing and growing and growing.

H

E

And then, eventually, the dictionary can't allocate anymore in tombs. Yes,.

H

Sunscreen, the resolver is shared across all instances of all options that are created with deserialize people. Sorry, those of us can eat it one time so this D sterilized people method is using the same underscore resolver every time it creates right for every operation, that's performed, so all all operations from now until eternity that call deserialize people are using this same underscore. Resolver instance.

I

H

A

I

Grow Oh fancy. Yes,.

E

Because they have the dictionary.

A

And that, but you know, grow indefinitely, yes and it's.

E

Going to be called concurrently, because I mean assume that this was a bcl method, it's static, so it's thread-safe, and now that means the resolver instance head to that instance, which normally wouldn't have had to be thread. Safe, does have to be thread safe in order to preserve the thread. Safety of the static method right.

H

I mean so I, so is there? Are we trying to enable the scenario where someone who uses this I'm wondering if we shouldn't yet have a generic factory with a new constraint and I? Just you know, activator create instance, the the type.

J

Yeah, there's very strong ones that is actually asking for arena patina the.

A

Same instance, the problem with that approach is now this guy cannot hold on to any state all right so now, if you need to pass in any state to resolve it, you're kind of screwed in them well,.

B

Not necessarily because I the thing is we we want to make it easy for someone to do the common case and possible for someone to do the uncommon case, but not we don't want people to run into a situation where we force them to use a complex API and they accidentally get it wrong. So I could see a world where both of these exists both the easy one that Steve suggested and the complex one that David just talked about.

E

So, for example, you would say reference handler: equal Jason reference handler dot, creative T, and that is a activator, create instance. T.

B

Basically yeah, and that would be the one that we direct the common case to, and then we tell people like if you really need to customize this, for whatever reason funk it I.

E

Wonder how that's gonna show up in Auto captions.

A

I

A

Maybe I mean I'm not sure how big of a bit of player that is because it seems like why would you cache the resolver yourself that seems kind of.

B

A

B

People accidentally do it like I I, know that people in the stream can't see this, but I posted in two chats two different variants of how to register a resolver. They both look innocuous. One of them will lead to an whom, within your application,.

E

Eventually, I mean, if you only.

B

E

Once it's fine, but but.

B

I mean if you have a web service, presumably you're, calling this over and over and over I. Think.

D

The problem, and- and this is kind of looted to before- is that the cert Eliezer doesn't have a concept of a state object. It has this options, thing right which isn't not intended to hold state. You know, dare to hold your options.

D

So if you have a funk hanging off of that that this holds state, you know they're kind of coupled and today there is no state object created when you call serialize or that lets you pass in a state object. But if that was added, I think that that would be. You know the place that you would want to tie your resolver to because are your handler to, because and it'd be up to the caller to say: hey I'm, passing in this special state that I want to preserve between multiple calls to the serializer.

D

That's very explicit! It's just that we don't have today. So that's why it is what it is right now.

A

And pose I mean my gut feel is that the funk is good enough, but I can see the argument. I mean like move like the second example. If I had, rather you basically closed over local burger won, this whole thing gets hosted somewhere teams also in the general case, fine, because it usually means it's tied to the invocation of that method. Right I mean unless that method is due to the entirety of your program where effectively the hosted instance is stacking around for a long time. It also seems relatively fine I mean.

B

Let me give a counter example: there I am writing an asp net application where I create the options once within my program, startup routine I, pass it to asp net and then it just reuses that options over and over and over again, if I've now additionally closed over a singleton, resolver inappropriately I now have potential thread, safety issues and room issues within my application, which my testing might not catch, see.

A

Again, it's a fear because he would register this usually and basically, the startup method, which yeah that's only one in the whole application, which is not good yeah by.

E

The way mo you're currently streaming your chat window instead of the Talking Heads.

A

D

Wanna, do you want to talk about the Newton, softest penis, a sure.

J

So actually asp.net step into that issue with Newton self, because they was, they were catching the Hastings utilizer itself, as an instance prefers, and by catching the whole distance realized today else they are also catching the resolver. So every time to do any request, they will use, keep, they will just keep adding to the resolver over and over, and they will. They will eventually reach out route and bound memory. Count right, part of homework.

E

I mean I think the phone like the funk is reasonable for somebody who needs to do a complicated thing, but I think Steve has the right idea that we should make an accelerator somewhere that just takes a tea, so reference resolver, Factory of tea or reference resolver, Factory dot, create of tea or something that just returns the funk for you and makes it so that like because it essentially.

A

So Kobe said.

J

D

And analyzed at.

E

That complains, if you don't use the word new in this funk well.

A

That means that we.

E

Need a better way of doing it, it.

A

Seems like the easiest way to do it is you just have a create method on reference and ler? It takes the tea and that T is a reference resolver and then returns event handler yeah.

B

A

B

To be a create method, as well rather than a constructor parameter that takes us on cut sorry I, just poke over you.

J

Okay, so Daniel yeah, so you will not be able to do there.

J

The the template that we were mentioning about reusing the same same instance on differentiation. Oh right, well,.

E

You could like we can leave the the funk constructor here, but then you could also say reference hand reference handler having a static, creative T that just builds the funk or nothing goes to new T or.

H

Have both be a factor create method, one of which takes a T and one of which takes a funk up. T yeah.

A

Although, arguably when you have both, they are equal, equally convenient right. So I like the argument, still holds that by people cereal overload, maybe they like I'm, not sure that they did. The generic one looks simpler. Basically, I'm.

H

Not sure anyone's ever said that functions taking functions were easy. Well,.

A

I think the problem is it.

H

A

Because, if you, if you look at you know what Levi said earlier right, if in the context of asp net- and you start up method, fung's are fairly popular or fairly common right so like. If that's and that's basically, when you accidentally close over a local, then you really kill yourself right and I think that's where my concern is because in ASP nets in the in the configuration thing like everything home like a lot of things, take phones, but it's a fairly common pattern there.

A

So that's why I think people might accidentally just you know when they see them both as overlords. It will just take that one. Maybe.

H

But then you kind of have to make two mistakes in a row. First, you have to gravitate towards the more complicated overload and then you have to do the wrong thing. With that overload.

A

Yeah I'm saying I'm, not sure it's a mistake, though, right because that's what I'm saying in asp net it's so common, then it's just you know. That's all you, you know if you have to write ten method call to at an end of them are I mean.

E

We could also have so get rid of the public constructor once we're having the static creates, but the reference handler dot. Creative T, which takes takes the T as a T with a constraint of new and has no arguments and reference Handler dot create manually, which takes a hunk reference resolver. So you had to write the word manually before you get to the one that takes the lambda yeah or.

B

Yeah or just make reference Handler abstract, have reference handler of T on the concrete shield, and if you want to implement your own handler, like you, subclass it and override the create hem or create resolver method like no function volved anywhere there.

A

See Chris say this last part again like you: basically, it would introduce the type the basic is the factory they were. You saying so.

B

What reference handler becomes abstract? You create a concrete sealed implementation, reference handler of T constrained, the T to both reference, resolver and new, and if I really want to customize how these things are created, I sub class reference Handler and override the create resolver method. Oh yes,.

A

Allanon because then maybe.

B

Like it I don't know.

A

Yeah, no because then it's like, if you really want to soup, you can but know you talked about ten lines of code rather than two right, so that seems reasonable.

A

Yeah I think I'd like that and then basically referencing the dog. Creative t is basically the thing that everybody's using because that's super convenient and it does the right thing well,.

B

You don't even you 380, you can just new reference handler of T, like that's the easiest way to create one of these things right. New reference.

E

Handler of good reference, resolver closed angle bracket, open paren, close paren, yeah.

B

That's that reads more naturally, for our developer audience I mean I'm, not I'm, not interested in shutting down the scenario right, I'm just interested in this, for the people don't accidentally do the wrong thing. Think.

A

About it now, I think that makes sense to me. So let me just sketch it and then we can talk about whether we like it or not.

J

So you're suggesting to make okay, so you make abstract reference Handler and until someone wants to use this custom preserves, they need to inherit small reference handler right and but I mean we consider I can see your mouth merging or doing abstract there's another class and get rid of reference. Resolver I mean like just moving the methods to reference country, but then, if you add something else to reference all this, it wouldn't make sense.

E

Reference handler still stays the factory it instead.

A

E

A funk, it gains a public abstract reference hand, a reference, resolver, create resolver method, and so.

B

E

You want to do. I have custom options that need to go into my resolver or anything weird. Then you would extend reference handler so that in your create resolver, you can do whatever complicated things you needed to. But if you just want the easy thing you would say your handler was reference handler of gooood resolver. Okay,.

A

And this one would have this guy here. So almost.

J

A

J

Wait but then again, let's assume that in the future we discussed before the we, some people might want to a big nerd option. So if you want to I, don't know, for example, if you want to have reference cancer with ignore, and then you have these two reference, that's all what about create! Well, I! Guess you can return, no I mean you can just overwrite it to return all in just enabling, nor in case we are ignoring the feature. Does anything what.

B

They can or be your static singleton.

J

Well yeah, but will then people be confused? What's that, maybe they they can also customize signal behavior.

B

You wouldn't be able to customize the ignore behavior if it's a static singleton, because it gives you back a reference handler. Okay,.

J

E

Tami awake what but ignore what would you be ignoring.

J

Oh well ignore these two separate options and instead of doing pressure, it ignores the references right.

D

But witness what's that beam today, you get an exception right. If you, if we see that stuff.

J

Have ignored today we don't expose. You know that, yes, on getting better I mean.

D

If you have reference IDs in your Jason today, you deserialize it with default. We.

J

Throw right now we send that to the Jason session. Data will flow, because what that that's.

D

What I thought ignore was for so: okay, okay,.

J

B

On line 14, you need to add a reference resolver constraint to to.

A

B

E

B

Reference resolver constraint to.

E

T, a new Kaaba reference, resolver referenced resolver.

A

E

I, don't remember if there's a canonical order, yeah.

A

E

This right and.

H

Then, on line 11 create resolver would also probably be protected right.

A

B

A

B

Oh, why wouldn't it be public? Why would it be hello because, if you're, if I'm creating my own, if I'm customizing, the serializer I, might need to call this in order to get a reference handler, what I, not or in order to get to the solver right, neutralizer sure, if I'm writing my own serializer or if, in the future, we ever intend on adding extensibility hubs, materializer, okay,.

E

Like if it took parameters, then I'd say it should use the plate method pattern, so it would be protected. Abstract create resolver core but says it doesn't take any. If we don't expect that we're going to be adding overloads, then.

E

Public seems fine, like cuz it'll need to be called by the serializer and yeah. We could make it into.

A

E

Internal yes, fam or SM, so we could do that so that we could work around it and certainly, if someone extended the type they could and make a sibling method that made it public but I, don't see a reason that we would need to I, don't think protected, helps I, guess, yeah.

A

I'm positive but I mean, if you want to test it right, you can easily construct any unit test. If you really wanted to I mean I, don't know like it doesn't buy us anything. I agree that this you make people jump to hoops now hey what.

H

About down in reference resolver itself, the string reference ID. That means we're creating a unique string for every object yeah. We already do.

K

That yep we do that on the on the default revenue server and that's not an issue performance.

J

No not really I mean my knee. Doesn't drink, write.

D

It out anyway, Jason you.

H

Right, but how are you is there a separate.

H

Realizing off the wire and I have a thousand references to the same instance. Am I creating a thousand strings in order to call this a thousand times.

J

Yeah I mean you are costing an into a string thousand times, I guess, because that's the ID that we have on the day for reference server.

B

We could, as an optimization, just like cache the strings 1 through 20 or something and make both the default references over and the serializer aware of this. If they're encoded in.

E

Their canonical utf-8, not escaped representation well.

B

Do I have to be honest, Caped anywhere well,.

H

I guess I'm wondering if we're creating a string from something. What are we creating the string from from my name? Why isn't districts indexed by it.

J

A

Let me do a channel, isn't isn't the idea that the resolver is the party that decides what that string is right? Basically, you you're calling get reference, and then I say this is the string representation right in my good example, it's the two string version of the couid, the default one would use two two string of an end and whatever other encoding I want to have for my instances. That's basically what I am responsible for right. The resolver is the one deciding what the string is.

B

I think there's a mismatch here, because good reference is called during serialization. Add reference is called during D serialization right.

A

Because you're, basically on both sides right, yes,.

B

So when you're reading stuff from the wire, the incoming dollar sign ID is a string which may be an int if that's how it was written out, but it could also be the work through right.

J

H

Okay, so it's not actually integers. It's arbitrary data on the wire yeah.

C

A

That the resolver put in yeah.

E

Whatever, whatever get reference decided to write, it could be using. You know, base 62 encoding to minimize space on the wire yeah.

B

But to to Steve's point, though, would it so it might make sense to take a span or a memory here, but the problem with that is presumably the resolver is going to have a dictionary backing its data and trends and and memory really don't work well with the dictionary well.

E

I mean the thing for the highest per that you would want would be the of the read-only span of byte, but then somebody would have to run it through an escape to to work out the Jason escaping if they, because you could write I mean something could come in from the wire. We're literally every character uses the backslash, hew syntax and that's legal yeah. So it needs to handle that so I don't think we can really do better than the string and still have any modicum of usability.

A

I mean you could pass in, basically the whatever the representation of the e of the Jason YouTube ad reader is right and then just say you, you are responsible for taking this thing out and whatever representation you see fit right.

E

But that would be that you have to you'd have to run it through the JSON string, D escaper.

B

E

A

Are too many Pizza.

B

Failure there yeah.

A

I mean I, don't know, like my my gut feel, is if you care about a high performance, you probably shouldn't use the API. We canvas.

E

Yeah I mean that that was gonna say that too, if, if you've got, if you doing reference handling- and you have a graph that has the same object in it a thousand times your serializing wrong.

D

Like if we use an nad frame or two string string could be you know what your primary key is right.

A

And I mean to Stevens Point, alright, I think the problem is you effectively have to create the string for every single instance right which, if the underlying value isn't in that's not particularly efficient right and for.

B

The little wicked is.

A

Almost always an INT, we.

B

Could address out, though, if we needed to address that more generally, we could do that within the reader itself, by having the reader keep like a small in turn, cache of strings that it's seen while.

A

B

This object graph, so if we were to solve that at the reader level, this would just get that benefit for free. We wouldn't have to special case us.

E

Or nothing and it references either or usually going to be returning if it's from entity framework and it's the primary keys and it's probably pre calculated, so it's not actually allocating anything to return. If it's doing a dictionary, try get value, then, if you, what you stored was the key or the string form, then then you've already computed that so there's a lot of. Even if you hit this a thousand times, it's not necessarily that you're calling in that to string a thousand times you can like. That's on. You know.

A

I think get reference would be you would it would almost have to go out of your way to not make it unique, for instance, right because the you.

F

A

This thing in the dictionary you compute the thing and you return the one right the next time. You just look it up the negative.

E

You've sorted is running.

A

By definition, right, if.

E

You started this dictionary of object to int and then returned the to string yourself then, but yeah, that's the you. You made a smaller dictionary at the expense of a more calm, more expensive, get reference and.

B

Keep in mind zero through nine dot, two string or cash anyway, so that might end up helping matters on the camera front side of things.

A

Yeah, if you only have 10 references, though it seemed stupid,.

J

By the way, sorry I I checked it before referring this over that well and yeah. We do have the string array computed, and so we just look them up and return them on the good reference. So there is no, as a thousand two strings cause right.

J

And also another, a lot of difference from youth from Newton self is that reference rhetoric, the referendum, solar, is an interface and we are having an abstract class. So we will attack to make this an interface.

E

Abstract class, good interface, bad.

B

Abstract classes generally are preferred over interfaces. We can also add methods to them in a more realistic fashion.

B

Because I think the only way we could add methods to interfaces is to use default interface, implementations, which now means you can no longer target that standard to a okay.

E

J

E

Scared of them.

E

A

Okay, so I think.

A

But is what we see on the screen, then what we kind of like.

E

Yeah I mean I, think line. 16 would technically just be able to say new T, because T had the new constraint, so you wouldn't need create instance, but otherwise, yes,.

A

Fair enough I mean not that I care for the implementation, but yeah.

C

E

And I assume that the reason that the that it's one reference resolver and not one for serialize I'm one for DC realizes just because it makes the most sense to an implementer to have the same sort of dictionary and look up semantics. Yeah.

J

A

Yeah, it seems odd to split them.

J

Well, internally, there two dictionaries right but yeah.

A

Because the Corp after using the resolver form, is different. But.

E

Right but the yeah just because on the one hand, it is a little clear, the relationship if it's split into two types but I think everybody's implementation. It would everybody would have the. Why do I have to write two types: I'm writing essentially the same code in in read and write. So it.

D

Scenarios to make sure that you can kind of round-trip these ideas is that true David, you.

J

D

You know today we create a new integer, every single you know run through and you and depending on your object, graph you'll get different integers, for you know the same object or, and by using this feature with a common cache on that you can control, you can preserve the IDs across multiple calls. Yes,.

J

Yeah, so well, essentially with the funk now, with the with the factory becomes, you can preserve the idea.

E

I mean you could also just have your your reference. Resolver type use a static dictionary and that's what you chose to do is your implementation, so.

B

Why would you want to reuse IDs across all instances.

G

J

One pointed that out in the example that we were discussing before that someone might want to elevate the.

J

The memory the memory memory says that they need for when they do multiple civilizations, they thought they they don't send the whole object. You know in a whole payload station or multiple one. So in order to know what the reference were on the on the multiple caster, they did, they choose to preserve the references.

B

Sure so then they would have to do serialized in the same order. Well.

E

But wouldn't your I guess as part of this who's responsible for understanding uniqueness. Like do you call when, when serializing do you call get reference on every object in the image graph, yeah.

J

J

Yeah I get the reference and these reference before that it's called on TLS and add reference and resolve reference. I call on this year, like so.

B

Then there so that means that you're actually creating an ID for every single object in the graph, even if it's never referenced correct.

B

J

Well for access for drugs, except for value types, yeah.

B

But I mean this: if, if I'm serializing something using this feature, presumably I have reference types, not value types. So this means potentially that going back to two I think Steve's argument about performance earlier should getting the ID of a particular object be deferred until it's actually necessary. So.

D

This is a one pass, algorithm I mean that would I think implied two passes.

J

It would it's a one part and you could not give it, because you need to write dollars an ID when you see the the object and then when you we are going to reference it. You need to verify ref and you need to do. You need to have the same the same ID that you were before.

B

E

Of two calls, because so looking at this, if it's, if what you're doing from your side is, you call is referenced and it always unconditionally called get reference assume that somebody's using a dictionary or you're dipping into the dictionary twice yeah.

G

E

It makes sense to be string, get reference, object, value.

G

Comma out, bull first use yeah I step into that and put.

J

Since that's different from Newton's us- and maybe someone wants to wants to know if something is reference before taking action, I split split the cows into two methods, so that's more user-friendly, regular I guess so, maybe if otherwise, if someone calls get reference and they might end up, are we starting the? Are you look for the object, which is probably something that might or might not want always.

E

Well, but I may have get reference gives the string and the have I seen this before like because really what you've written here is is dictionary contains key, followed by reading the dictionary it's a little different, but that is like on anti-pattern 101.

J

I have parentheses in the proposal where I, what I exactly did what you were mentioning so, instead of having it reference and as an each reference is they get ringgit reference object. Value will already exist, but that said that, now you, you cannot look up for the reference without adding to the dictionary.

E

But if you're always going to call get reference wouldn't like is there? Is there a scenario there like you're? Not your oh you're gonna call get reference whether they return true or false.

E

Yeah so get reference doesn't know if it, if is referenced, return true or false, unless they're, depending on the absolute deterministic ordering of the methods that you're going to call in saving state between them, which seems like it's even better to have it just in one call.

J

Well, yeah I, don't have a concrete sample for that, so I mean uh and I was proposing. The first I mean I was proposing the first. Also so I don't have a strong opinion about changing to only get reference to having to have only get reference that are getting reference. A nice reference right.

J

And yes, and and and they and occult sites that I saw the on github, they usually use do what they describe, which is called. It's called, is reference and then college ID reference used to do the same, which is to do the same light that we will do normally with one single method.

E

Now I think, unless we can come up with a scenario where it's important to have them in two methods. I think I mean again if you're using this performance is obviously not your highest goal, but that custom equality comparisons can be expensive and so double dipping. The dictionary is not necessarily a good idea. I see you so I would I would recommend making it get reference making it reference return both things and it feels like the most natural representation of that is adding an out pool on or to what you currently have. Okay,.

E

Cuz try get reference feels weird, no I wouldn't like that one, because that would ya get referenced. That outs is this: the first time I've seen that object seems better, because now the the boolean has a name which tells you what it what it means so.

B

What you could do there, then, is instead of adding a bool, you could have an in pool which is insert or just may whether it exists, and if get reference returns null, it means it didn't exist.

E

But who would ever call it with I'm probing.

E

Because the infrastructure is always gonna call it with I need an ID for this, but you have to tell me if I'm writing, down dollar sign, ID or dollar sign Ref.

E

B

E

Object or referencing this object, but it always needs a name so.

B

You're saying that there's no need for a stand-alone query: API, the only the only API you really need is at it. Unless it already exists and tell me what was already there yeah, okay, so add, or.

E

Get existing it's a target existing that needs to give you both the ID and whether it added or got existing so as well. The.

G

E

To know are they in? Do they need to write down the value still with dollar, sign, rep or a dollar sign ID, or are they writing instead of the value dollar sign? Ref? Okay,.

B

Yeah I've got it that.

E

J

Keep in mind that this is actually hold Newton, who looks so maybe it will make it more major migration currently, if, when they try to pass me in their code to our code, wouldn't.

B

That be in general. How much are we trying to mimic that API.

E

Well, I believe our goals, where we will be inspired by it, but we will do the right thing for the platform instead,.

B

E

Copy they yeah yeah.

B

Because I mean if you, if you tip the scale, all the way to the other extreme, it's oh, if you're satisfied with it, just keep using it right.

B

J

Okay, so there was one.

J

A

So what does it mean for the API, then.

E

Remove is referenced and added out bull.

A

E

A

I owed the bull, we're out.

E

Bull existing on get reference.

B

B

E

Like that, and now we can debate the word existing okay.

E

I mean oh yeah, existing already exists, we could invert the meeting and say first use or.

E

Something that conveys that this is saying the purpose of this boolean is whether or not this object has been passed into this resolver previously. Maybe.

A

You can just say: oh, it already exists.

A

A

Might be a slightly.

A

Apologies, my cleaning lady, is apparently dismantling part of the house in.

E

This a thorough job.

A

She just declares bankruptcy today.

A

All right anything else or is that it.

D

I mean just on that note, then I mean dad reference to. Is it ever? Is there any scenario that you'd want it like a triad to see if you've already added it.

A

I mean the sealer. This should handle that no well.

D

I mean I think it would normally just replace it. I suppose. Have you guys.

E

That's presumably just whatever the resolver is going to do if two things are defined with the same ID yeah.

D

Yeah I'm just wondering yeah. If people are calling that sudden they're not the implement or they would find out if they would care about just actually, we tell when we.

D

So internally we throw when that happens, yeah exactly okay, okay, but that means we call get reference. First, then,.

J

No ad references for this utilization, okay,.

D

So, okay, so the dictionary itself, we if we add to what we just throw so yeah using that version of the index or whatever that's there also.

J

We call triad and we throw yeah, we call triad and if it's not valid, which oh, that teacher, which out that this is a duplicate, ID.

E

Yes, all right, I think it's probably good. Then.

A

Until it's not.

A

A

All right and then we're 12 minutes over, but at these three items. So that's a pretty good. Thank.

D

A

Right guys see you later bye-bye.

A