Node.js Node-API Team, 12 Jul 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Node.js N-API Team meeting - 12 July 2018

Description

https://github.com/nodejs/TSC/issues/565

A

So welcome everybody to the napi tea meeting for July 12 2018 and we're just going to go through the current milestone that we're working on and talk about the issues that are in there. So our first issue is the blog to coincide with 6x + 8 X back board availability. um I have Runa, send me an email earlier today saying he's working on that and we're planning to get together, hopefully to sort of finish it off. I think it's next Monday or Tuesday.

A

So we're moving forward on that I haven't checked the PR for the 8x cember minor, yet or recently, but you know I think that's still moving forward. I know that Gabriel, you backported a few things for that. Already right, yep.

B

A

I think we're okay, for whenever that goes out, if I'm not mistaken,.

B

um Just uh just a question of policy: is it okay that if I initiate the back part of something that I think is important to have there because I one of my back ports is? Is a reaction to to miles asking me to back port it because it didn't land cleanly, but the performance one? He didn't ask yeah.

A

I know that that's that's fine I think, like normally more things would be backported, but there's sort of a big backlog and eight dot, X, okay, and so the approach he took this time was to sort of ask people what they thought was important. So if you think that that one's important and I think that you know, I I saw that I agree if it. If we can get better performance and you can get a backported, you know much work, it's definitely worth it right.

A

So yes, okay, good, so yeah um yeah and that's anybody else. If there's something you think we need to back port, you know feel free to tag it there. So I think. That's all there is to say on that one. The next one is to create the workshop for node summit, as Jim mentioned last week, he's not able to make it this week, but still working on it. He did send us an email saying that the third one is complete. I mean he has some time to comment and review on Monday.

A

So you know I think we should all have an action to take a look at the latest. One I I left him some comments in the issue, some some minor things: okay, but yeah. So if we, if anybody who hasn't looked at it yet if there's time that would be great I think other anything else, I think other than that. You know we pretty much have our deck, although hit I think we needed you to fill in your about. You slide yeah.

C

I did that last night: okay,.

A

C

I also added a slide, are ported the slide from our nationís old presentation about all of the nappy contributors. Okay, so a dislike.

A

For that, okay, so do we think we're pretty? Maybe we should just open that up and go one last run-through to see like? Is there anything else that left that needs to be done? So let me pull that up.

A

Okay, there's the link I will share it.

A

So here's the presentation as it stands. Okay, so we got our intro about about about about about.

A

So this slide I think we get rid of right.

C

Yeah I think that can go cuz.

A

This is, we've replaced, it I think with all of these mm-hmm. Okay.

A

So basically say hello, I guess we need to tag who was it who's? Gonna talk to this one I.

C

Can talk about this one okay,.

A

So is that is that you, yes, that's you.

B

Right there, yeah and I believe it was Jim for for no confi you if I remember.

A

See how I fit this? This is: okay, I'm gonna not put a picture there cuz it doesn't it kind of conflicts with all the other pictures. Yeah.

B

That's fine, but.

A

We'll remember that you're gonna do that one that kind of makes sense because you're the last one there, so you can just naturally lead into it. Mm-Hm, that's good!

A

So, in our workshop scheduled objections.

A

A

Are we missing something here, I'm going to Torres.

C

A

So we need to I.

C

Think did start writing one for a sink worker right. Well, that's what I was just like.

A

He's got the a sink worker. One is that which is that here.

C

It might be up one directory.

B

Maybe there are different kidding.

A

It's in the it's.

B

In the issue, yeah.

C

Special topics- a sink worker yeah.

B

A

Yeah, okay, so basically I'm going to.

D

A

That into there Oh somebody.

C

Already did yeah.

B

A

You're beating me to it great other.

C

Words, we need to also fix that a Inc worker to a sink worker- oh yeah,.

B

There we go why Oh wrong color- and this is a nice 8-bit, color, blue or ya- know sixteen color. That's four bit right. There.

B

A

Okay, so then I guess we'll leave the other. This one blank still here in terms of I, think he'll that hopefully will have it, but if not we'll just take it off the day out excellent right, so I think there yeah. We should discuss I, see that Nicola and hit s. You guys put together a list of modules which is good. Do we think that's that's enough modules, I!

A

C

Know if we've settled on which modules you want to use so, okay, I'm, discussion and feedback would be great right now, right.

A

So what's the I saw there is, you know, there's a number of proposals there are there concerns about some of them, or should we just we easily pull that list up? That's probably in an issue somewhere right or is it.

C

It was an email, sorry, give me one. Second, someone's knocking on my door. Sure list.

B

Of native add-ons for workshop I think that and.

A

Maybe if you just paste it in the chat, then we'll all have the current list talk boy sure.

A

So I was thinking attached because he might have it. Oh.

D

B

So Nicola Nicola had the header list and then he added some more yeah. So.

A

I am back yep, so we're just saying if you, if you have the list handy, if you could paste it in the chat, maybe yes give me one second, do you want the combined list yeah or whatever is the most useful for us to discuss and and move towards the list we want to use?

A

C

This, oh, do I go to chat over here. Oh.

A

Did I do I have to promote sec? Sorry, you may be there. Sorry, let me I'm going to find once I go to sharing it's harder to find participants there. You are sir okay.

B

Yeah so there's doors.

C

Okay, I pasted both the lists that I came up with and Nico's list. Okay,.

A

So is that a right, okay, so there's the two so I, guess I I, don't have any additional input is, is like, as they're concerned, that some I know there is I saw some of the discussion where maybe some of them might be a bit hard to port or yeah.

C

I think Couchbase, for example, might be too long workshop.

E

Maybe we need to split the work between some attendees but.

E

It's it's very difficult if they didn't implement, didn't implement and it went on before so yeah I.

E

Technology list and I think that lib x, ml, Jas and coach, the coach base could be more difficult to support for.

E

Developer that never.

E

Implemented in native at dawn mmm right, so this is my my my concern for dress in are all good candidate to use in.

B

The workshop correct me: if I'm wrong, though we will also have some people who will be starting from scratch, though right, because they intend to to expose something to JavaScript, so so some people might not want to to tackle an existing module. Is that correct? That's.

A

Okay, I I think we may, in my mind the two key things. One is one have a big enough list and the okay from the the the providers for, if people who do want to do it, if somebody wants to start working on their own or has you know, wants.

B

To do something else, that's good too, so so then so then I guess worst case scenario is we have 24 people and they all want to port a different module right. So then, if we want to cover that, we need 24 modules, yeah.

E

A

What I was thinking so I, but I, don't know how easy it is to get a list of 24 good candidates. The other thing, too, is like the one you mentioned. That's hard I think we could have a. We could break out like a second here's, a here's, a hard like an advanced list like unless we think it's one like. If there's one I think there was one that was like hey this uses stuff that we don't expose in the API. It's like, okay!

A

Well, if that's the case, it definitely wouldn't be one that I want on the list, but if there's some that are like gonna be harder than others, but still doable. Maybe we have, like you, know: here's the regular set and here's like Oh for advanced users. If you think you already know what you're doing on on native modules, yeah.

C

You could probably split this up based on difficulty, so.

A

I think that I think that way would let us, you know, keep a few more and if there is somebody who's like you know, really experienced they you know, give them something. That's a little tougher. You.

C

Know do you have any thoughts on so there were at least some modules on the original list of like most popular modules or whatever that Jason had compiled, yeah, which don't seem to be maintained anymore, but there were certainly some candidates over they were. It looked like you know the fairly trivial module that could be easy to port I.

A

Would still pull those in like if they're still because they were there, because they were fairly heavily used right.

D

A

You know if we have somebody who has enough interest to pork them, maybe that's a gateway to having them, maintain it. Okay, I, don't know I, that's just my thought. Right like that.

C

Seems like a reasonable idea.

B

No well from from from what was pasted into the into the chat, even if we include Couchbase, we have, we have only like 13, so we still need nine more modules to cover the absolute worst case. I.

C

Think we can probably get I mean I, think we mean I, think we need more than nine, because things like ISO utf-8, validate I saw was an eco slash, but it looks like somebody pulled to that already last week: okay, well.

A

That's a good problem to have yeah I mean I, think you know the bigger the list. We can get I think we could cope like if we had half the number we needed. We could say. Well, you know, we've only got us, you know either find your own or maybe work on one that somebody else is porting as well as an exercise. Ideally.

B

A

We'd have at least 24 I have.

B

Another idea can we conceivably think of any C or C C++ libraries that we might want to port like from scratch, which could be interesting.

A

I I don't have any off the top. My head: it's not a bad idea, although in terms of adoption.

B

A

Your problem is getting people to move over from existing yeah.

B

Yeah for sure for sure yeah, but I mean just just for the for the sake of the workshop. So.

E

You you, you, are you proposing to create a new add-on from scratch, using C, C++ library, yeah.

B

There is one that that is interesting enough for the nodejs ecosystem and it doesn't. He isn't yet exposed I mean I, I, don't know it's a long shot much much longer than then then the existing add-ons, which simply needs to be ported, because they've already proven that they are interesting. Otherwise they wouldn't be around at all yeah.

B

You know just just because we don't have enough.

B

That's that's the only reason I'm mentioning this so.

A

Attach that other lists that we had what, if we, if we pulled the ones that weren't ported over from that, how many? How many do you think are there? Oh I.

C

Mean that list had like at least 100 modules on it right.

A

So if we just use that list from you know going down from the top I think my you know the oh you know: can we just pull that list say, take the top 30 and add it to our list here, I.

C

Just put a page to the list down to the chat, I think some of the top ones we've already ported yeah, but yes, certainly I, think you know some some are probably harder, like FFI I think might be one of the harder ones to port, but yeah I think we can walk our way down the list, see which ones haven't been ported yet and makes a case for them to be a candidate.

C

Okay, that's kind of what it that was kind of the exercise. I was going through yesterday, yeah.

A

If you so, can you continue that and sure.

C

I can go, I, think I stopped at like 30 or so, but I can keep going down that list just.

A

To basically get us to the point where, like okay, here's the ones that aren't obviously too hard I think like I said before is good, if we at least just fire an email off to the maintainer saying hey, this is what we plan to do. Let us know like I, don't think we may. We may not want to wait for you and okay, but if the email just says you know, let us know if you prefer, we didn't do that.

A

Otherwise, we're gonna go ahead and do it, okay, sure and then I think for the day, like I, don't think we should put them in the slides, necessarily I. Think just on the day, we should have like a you know, printed set of papers.

D

A

Know 24 papers that have the list that we can hand out to people to basically say you know, or even one which is got one piece of paper with the list that we have people circulate and strike out the one that they're gonna work on you.

C

Know we could do it like how rich does the kodan learn thing, which is these tiny slips slips of paper period, a package name on it? Yeah.

A

Either yeah yeah yeah, that's not bad I. Think even like a single piece of paper that we hand around you strike it out would work too, because it's not a huge number of people, but.

F

A

Wait yeah whatever so you volunteering to put together slips of paper, whatever whatever works, there I'm sure.

C

Yeah well I'm, well, looking in the beginning towards each get the list compiled then once we review that yeah put in the slips of paper is not too hard. Do that yeah.

A

We just need to have somebody. Do it? That's okay sounds good, so any any more. We need to talk about the list. I think that's I. Think by next we have week and while we have a week so hopefully by then yeah.

C

I can I can try to get this put together by tomorrow, because I think we probably need to reach out to the module owners sooner rather later. Yeah.

A

I think so, like giving yeah I think like sending it out at least Monday saying hey next week, we're gonna as an FYI. This is what's going on yep.

C

A

Us a give us a you, know, yell back, if you don't think, that's a good idea, yeah, which.

C

Is kind of why I was also leaning more towards modules that were actively maintained, because at least you knew there was a maintainer? Was you know around.

A

Yeah I mean that that's yeah, like I, think prioritizing those makes sense to me, but some of the others, if they're really still, you know if they're high up on the usage list, it's still something that you know us getting. Somebody to take over would be important. Yep I buy that okay, so I'm gonna go back to so we're talking about.

A

So I think anything else on the workshop, we should touch on.

A

I'd, encourage and I'm gonna try next week for us, since Jim isn't gonna, be at the workshop for us to go through the workshop. This, if you haven't already just to make sure we're familiar with it.

A

Okay document currents are adding ApS any PRI, so I.

C

Still need to ask somebody else to open up here. They didn't I'll open up here.

A

Okay, that's that's kind of what I thought happened: okay from the PR this week. Yes,.

C

A

Good escalates Anisha joined us. Yes,.

G

A

Where do we stand on that? One.

G

Sorry I checked to make sure I've heard a different way to check this button values an integer and that is done and I'm just trying about an efficient way of checking the date and really Jack's parts of it. But I should have a PRN soon. Okay,.

A

Sounds good and it sounds like the maintainer. He was basically saying: hey I want I, need those things or yeah.

D

A

So it sounds like if we send a PR he'll, be happy to merchants.

A

Okay, burning down the list of issues any particular issues. We should talk about her raise this week. I.

A

Was thinking that after the like I see Gabriel, you opened the one to sort of update the the bundled version of the API yep. After that we should probably put out a release. I was thinking.

B

Yeah I actually landed that today in cash, that's in there I I squashed, I squashed, doctor commits and and just mention the original commit IDs in the commit message sounds good. Is there anything else we think we should get in before we do a release? I I do have a small PR which, which I just realized this morning, that that I should have done in the previous PR, which was to to backport. This note note persistent, which we've been using for a while.

B

Now it's an old internal, because I had to change the the way the the callback bundle is done to use a class rather than a structure, because we needed destructor for it, but we don't really need a destructor if we had no persistent, because that was the only reason for the destructor, so I. So by copying note note persistent out of the no tree and into our copy of node internals I was able to put like reduce the Delta between between what the bundled version and the entry version. Okay.

D

B

There's this it's a minor touch-up on on the previous PR, so so I think we should get that in. If we're gonna release yeah.

A

There's no rush right, I was just thinking. It's been a while. It's been a couple months: yeah yeah yeah, your release.

B

Would be good and I mean it is a performance improvement. So all those folks who are not yet on a version of note that has this yep one to reinstall and reap the benefits.

A

Okay, so yeah we get so basically wait for that, if there's any other ones that people think we should get in before we land, let let us know otherwise we'll plan to try and do a release or like yeah, like.

B

In general, I mean, as as the impetus for for a shipping, a version of napi fades right yeah. Do we intend to back port all the new api's that we add cuz? You know, like you know, I, like I, mean I just landed thread-safe, not so long ago like do we really want it to bring that in or or is that going to be just like just like the sugar and no implementation so well.

A

It I think so. Iii think we want to continue to pull things in to know that on API itself, now that doesn't mean we need to back port everything to older versions right.

A

So when know that NEP I would then have to be smart enough to say: okay well, I, guess: we've got it guarded by the the node versions right, so it would basically be you know if you want to compile, you would use the right node version and an API, no doubt on API has to like not not depend on the things that weren't available. If we haven't back ported them everywhere, right, yeah,.

B

Yeah actually I'm trying to think like well.

A

And some of those are still experimental right, so first they have to get out of experimental yeah.

B

True, I guess I guess the experimental now that we have a concept of an experimental and api. Then we may I think I guess we have a legitimate point by saying that we only support the stable ones right.

A

Exactly no doubt on an API, well Noda, an API can have wrappers for experimental things, but they are tagged. They need to have the same experimental if deaths, yeah, yeah right, so that that I think we can continue to do I'm in terms of something where and then you know, we've already got the model that if we release a like new API is they'll, go into a new node version and we'll need note, add-on, API to again respect the node version in terms of hiding things you shouldn't be using right and then separately, like I.

A

Think it's really on as necessary to back port things like I, don't think we should be trying to back port everything from you know. The most recent API is backwards. Yeah.

B

I suppose, since since since we are since we are like on a different schedule than nodejs, we could just you know, reactively back port and then and then you know, release a new version for whoever complained that this is missing. Or that is missing. You mean and know that on API or.

A

B

A

Wait a sec you're talking about into the bundled pieces if.

B

Necessary, yes, yeah.

A

Yeah yeah I would be I, think I'd be more reactive on that part. Yeah, okay, okay, like it's! It's because you know they're using they're using that they're already not getting advantage of the full they're, still not getting the full benefits and they have sort of the if they want to compile across versions, they can use the you know the node version has the guard yeah. That's kind of my thought. It's like unless people are saying, hey, I, that's causing me grief, I can't I'm not being able to migrate or whatever.

A

For though, for that thing just leave it as what it is it. You know if there's fixes sure, but if there's new things yeah, then we didn't make you think twice yeah, because we may not even port those back to you know: version six or version eight, either right, yeah.

B

A

And actually, in fact, I'd almost argue that we don't want to put anything in there that hasn't been back ported all the way, true right, because then you could get into a case where it's like well, okay, this works for me on sex. If I use the non an API version, but then it's not available for it anyway, yeah yeah, you.

B

A

B

A

Right, okay and then the last thing is documentation. Push on no doubt on API.

H

I have a couple that I'm working on I'll get a peter out. Hopefully, tomorrow, okay.

E

I I don't want a here for Frank Jean, okay,.

H

E

Tomorrow, III ended funko reference, so it's all on one PR, so take a look and for review and okay, okay.

A

So you're adding function reference to the same, pier or.

E

Yeah yeah, okay,.

A

So maybe you're saying we should wait until you add that to take a look or if.

E

You start or anyone someone can start to review the function, it's better because we are more.

A

Fast, okay, yeah I'll, try and take a look at I'm away tomorrow, but I'll try and take a look at it if I have done today, otherwise I'll make sure to get to it on Monday and he shall house the ones you're. Looking at coming.

F

Cuz I wanna get your infernal.

A

A light one, yes.

F

A

Just want to make sure we get this these in at least the first cut by next week's since we'll be at the summit the week after and it's a good death complete by then.

F

C

Okay, I need to jump off the call yep yeah. If you need anything, just email me I'll, be happy to help, ugh sounds good right.

A

Anything more to discuss on documentation.

A

No okay, I think that's everything. That's in our current milestone um are there other things that people wanted to talk about this week, yeah.

B

I have one regarding these, these non-validated versions of the of the api's. So it seems that it seems that for performance reasons we may want to have a version of of the of those api's which take javascript values and convert them to native values. Where, where we don't type check, we go straight to the conversion and if it fails and if it SEC faults, we take no responsibility, because it seems that that the the validation is taking a chunk out of the out of the performance. And so we've had one person complaining about this.

B

I mean they are working on a hash table. So I guess that is a pretty chatty application. But I have found that that if, if they switch from using node buffer to using v8 array buffer- and we we give- we provide versions of our API- is like, for example, like an app you get value, Ewing 32, and that we get array buffer info. Where we don't first check, is it a number or- and we don't first check? Is it an array buffer, but we simply, we simply cast it and go.

B

Then we can match the performance of what they're at what they have now, which is v8 findings straight and using node buffer. Of course, if they, if they kept using v8 and array buffer, it would be way faster than anything we can provide. But at least if we force them to use array buffer or if we advise them to use array buffer and give them these non-validated ones, we can match the performance of v8 plus node buffer, so so this this is.

B

This is a case study of a single ad on which, which seems to suggest that we should have these functions, and then we map how much is like when he is it like. A fifty percent.

A

Overhead ten percent what's uh well.

B

I I have the chart. Let's see where was that issue? Oh you yeah.

E

It's this table know that I think yeah.

B

Yeah I, have it I have it. It is now in the chat. Okay,.

E

B

Yeah I I put a charge, so basically we have, let's see napi yeah. So if you look at an API array buffer fast, that means that means we do not. We don't so, basically, where we're sending we're sending two buffers and to 64-bit integers from JavaScript into the native add-on. That's the that's! The measure.

A

These are the parameters right, so you, basically the parameters get passed in, but then you need to convert them using the call to yeah.

B

For the for the buffer, you need the pointer to the data and the length and for the n64 you need the in 64 yep, and so so so yeah so v8, v8 buffer is what they have now.

B

So using v8 bindings, not an API bindings and using a node buffer and and if they didn't switch to. If they didn't switch to an API, they could get much better performance by by using v8 and an array buffer. But.

A

Yeah I was just wondering: what's the absolute comparison like. Is it 100 percent 50 percent? What's the? What do you mean the absolute compare.

D

Well, like okay,.

A

If it takes 10 milliseconds with v8, it takes 20 milliseconds with any pi. Oh.

B

Well, it's let's see what we have now is well I'm, not sure he said at the beginning of this thread, which started in the other issue that that What did, he say, I think for a million calls it's 200 milliseconds just to just to do this conversion and and no useful work, and the hundred milliseconds of that is just to just to call napa get CB info is, is a hundred milliseconds of those 200 milliseconds, so so getting the parameters at all is is is very heavy but then, but then the validation would the validation would significantly reduce the other half yeah.

A

Trying to get a sense of his like is you know, for if it's a five percent improvement to sort of like well, not so interesting. If it's a hundred percent like it sounds like we're talking, 25 percent, possibly because if it's like huh fifty percent of the time is just calling in like making the call in in the first place right so and that's not gonna go away yeah. So if it was a hundred before that, we're optimizing, it best.

A

Fifty of that and if I look at these graphs unless they're, not logarithmic, they're, linear, right yeah, so the one so v8, which is the the green, is worse. What's the.

B

Today is v8 array buffer, no v8 buffer is what's being used today. That's that's. The released version is mine.

A

So that's like 10 10 million mm, and is that total calls, or is that to cut the cost of the conversion this is this is this is what this is, what our benchmark, outputs I'm, not sure well,.

B

A

So the the difference between that looks like about 25% between.

B

The blue, let's see an API buffer, is the orange. So so, if.

D

B

Just switch over straight with with no modifications, then that would be the orange and actually.

A

Pairs to blue, which is so it's like, 7 versus nine, or something right or 10,.

B

Well, okay! Well, let me see, one is what they have now, which is blue and then and then orange would be if they convert it straight and and this sort of greenish yellow would be if they switch to array buffer, but still we didn't make any API changes so that which is greenish yellow. That is the third last.

D

B

So we can achieve third-last without without any any changes. So.

A

That's better than V buffer, which is what okay.

B

It's not better than v8 buffer, but.

A

B

Better, it's better than I.

A

Guess I'm having trouble with the colors there's a yellow and the green okay, so it's over third last on there in the same order. Yes, so.

B

So I think the interesting ones are first versus third, last and first versus last.

D

B

Is what's released and third.

D

B

Is is what could be achieved if only the the add-on maintainer made changes from using buffer to using array buffer right, but otherwise no changes on our part right and then the last one is what we could achieve if we, if we provided a validation, free functions for for for both array buffer and for for n64 right. So.

A

It's yeah it's around the I'm, just spitballing the you know. One looks like it's around seven and one looks around nine right. Yeah. Oh you mean the you mean like the.

B

A

Y-Axis yeah yeah yeah yeah right, which, which says that okay so like that would say nine nine over seven, that's like 30%, so yeah, it's a this is substantial difference.

B

Yeah yeah I mean we could basically, we could basically cancel with with, with our combined efforts the the module. Maintainer is effort and our effort, we could essentially cancel the the impact of switching to an API for this particular module right. So my real question is is like.

A

Like I I, having two versions of the same function sounds a little ugly, yeah.

D

A

Is my first, my first few but but I wonder is like, were we to conservatives like what, if we just don't include the check and it's up to the the module owner to make sure they're doing the right thing like? If somebody was writing it today with men and v8, do they get that check for.

B

Free or not they do for buffer, because because uh the the nodejs implementation of buffer uses something called array buffer view, and it does the is array buffer view first, so so you, the node API, does it but but yeah I mean if they did v8, then it would be up to them to do to validate yeah, whereas.

D

B

Date on their behalf, so ya know.

A

B

A

That, in the one case, though, likes okay, the buffer verses, so like the n64 or what's the equivalent, if they were doing it on their own.

B

Well, yeah I mean if they used v8, then they can choose to simply cast to an n64 because they know they did it somewhere right.

B

A

If it came in as a parameter, it's still a handle like it comes in as a.

B

Javascript thing.

A

They still have to figure out they self to know.

B

What type it is! No! No! No, not necessarily because remember that that that every every native module is usually fronted by by a JavaScript module in the same package that the same maintainer brain controls. So so they can do the validation on the JavaScript side and make sure that the binding always gets like clean values. Oh I.

A

Was yeah I was trying to say you know, for the is for the conversion to like you know, get nappy get value in 64 yeah. They have the same. You know if they're using not they're, not using nappy, they're using then or or or v8, if they, if they don't I'm, just trying to assess like if they don't add the check themselves and they try and convert will they crash? Yes right.

B

A

Be no different well, except that we've done the check that don't we have to check yeah.

B

Yeah yeah yeah yeah. If we, if we, if we were to remove the check, what would be behave exactly as the engine behaves right and I'm kind of wondering.

A

If that makes sense yeah, then then the next questions is. Would that be breaking? But yes, like you know if we said that wasn't breaking my first thought would be. Maybe it's that we should just remove the checks yeah.

B

A

Because you know if it's that big, an overhead- and you know, okay, we're adding safety that you didn't get with the other api's. But you know mostly.

B

We're not giving you.

A

The option of being performant right, that's a big trade-off, yeah, so I'm that would be my wanderers. Can we get rid of? Can we get rid of the cheque like this? Getting rid of the cheque, make any sense just as opposed to adding a new method.

B

Yes? Yes, yes, yes, that's. That's! That's an important question and.

A

Then, in terms of being breaking it's like okay, well, I, guess today you could, you could make the call check the result and if it told you the result, was wrong type or something yeah yeah. You get nappy invalid argument.

B

A okay but yeah it would, it would be different. I mean the the things potentially things with crash where they didn't crash before yeah.

D

B

So you know you it could be considered a regression at that point. Yeah.

F

B

So that would be. That would be bad I guess.

D

B

Because then, because then we would have to treat it as a regression and then add the check back in yeah.

A

How many of these methods do we think we'd need to duplicate.

B

Well, I, don't know pretty much snappy get value star right and there's quite a lot of those. Aren't there well I'm cutting pipe grepping. Yes,.

B

So we got double in 32 you in 32, in 64, bool string, string, string and external.

B

So- and this is this- doesn't take into account nappy.

B

This need single quotes, nappy get and then I don't know if this works dot star info, so we get nappy, get CB info, that's that doesn't do validation buffer, even if we don't do validation. Actually, we don't do validation in app. You get buffer info because we know that no does it on our behalf. All.

E

B

Nappy yeah so array, buffer, typed array and dat of you. We would need we need to remove validation. So that's a total of one two, three four five I'm, not sure.

B

If four-string you you'd want to remove validation like mm-hmm I, suppose you might I'm not sure what what the general use case is for Strings like if you know, if there's like lots of bulk string transfers going on from js2 native in some module, but okay, let's in the worst case, we have 1 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, so 11 more api's, all of which do exactly what these do except eject. That's.

A

Not so crazy, I wonder if my thought I'm wondering is like nap. You get value in 64, understood, unsafe, yeah, exactly I, fast or unsafe, or or whatever oh yeah. I was thinking the unsafe part because it gives you the you know.

B

A

Exactly you can use this there's a warning. The question.

B

Be then to come.

A

Safe thing with buffer right, so right could those just could the unsafe versions just call the like, do the check and then call the safe, the unsafe.

B

Version, oh! No, because because the the the whole reason why we create them is, is that performance is an issue, and so we don't want the stack construction destruction at that point. I, don't think of another function. Call yeah I was thinking for the for the non safe for the safe ones yeah, but but if the safe one, if the safe one made and made a function, call to the unsafe one, then then that that's a function called. That's that's setting up a function, call yeah.

A

I'm just wondering if that's actually significant I guess is it that is the is a single function. Call like we're at that level of of you know the D is check being equivalent to that or is the is check, something it's because we're going back across.

B

A

B

True yeah I'm, not sure, I'm, not sure how heavy that the check is on the on the v8 side, I'm, just thinking more on, like the maintenance side of things right like well, my idea what my idea was to just to just the in line in line right, it's and and just just have a common function. Yes, but it's in line so yeah.

A

Okay, yeah, that's hot yeah that.

B

A

B

Yeah I I, totally I I am I'm a big fan of do not repeat yourself, yeah.

H

B

I would definitely do it with with an inline function at that point, and then that would be like the meat and potatoes of each one, and we just call it. It does seem like.

A

You know if there's a 30% difference as long as I mean I assume the in this case, like they're, not doing much else in their call right, they're, just getting the parameters and going back or something like yeah.

B

Yeah I mean it's, it's I mean I, I, don't exactly know what what it is, but but the fact that this person came out and pointed out that this particular thing is a bottleneck for them. I assume that this particular thing is a bottleneck for them, and since it is a hash-table right, the the hash table doesn't really do much it just you know it's mostly just set get set, get a if from from what they're saying at least so.

D

B

So, that's why that's why the this is this chatty use case bothers them. So much is what I gather from the discussion right.

A

Yeah cuz, we kind of I mean we kind of new chatty things. There was gonna, be more but yeah it's so it's just that balance like is. Is it save enough that now more people will use it then would otherwise right and versus the extra maintenance? Now 11 functions isn't a huge number, so yeah yeah.

B

It's it's it's um it's it's! It's kind of a kind of a tough call, but I suppose we need to. We need to look at the big picture and see how much, how many more of these performance sensitive use cases we're gonna run into. You know yeah, because, for example, one idea that I floated was you know.

B

If you want to reduce the number of parameters, why don't you set up an array buffer and then and then just use, use these data views to to set things inside the array buffer on the JavaScript side and then pass a single array buffer and cast it as a native structure, yeah.

D

B

The problem with the problem, with doing it that way for this module was that that there are two array buffers involved and their data cannot be contiguous because of you know the structure of the module. So you know the. What do you call that the application doesn't doesn't allow those two buffers to be contiguous? So that's why that that that idea didn't work and it it also has to be zero copy.

A

B

Right, I guess.

A

That that idea, though, if you extended it to passing two things instead of yes, yes,.

B

Yeah, like you.

A

Can always take it one front one further right if there's two things that can't be contiguous well,.

B

Then just past two, that's yeah yeah, but if there are four things being passed right now and and since the buffers seem to be sort of application, specific I would say that the best you could do is go down from four to three or actually from five from five well from five to three, because no actually it has to be from five to four, because the external cannot be bundled since, on the JavaScript side, you cannot get the value of the pointer and and and stick it in a structure.

B

So so the external has to be separate. The buffers have to be separate. The only things you could bundle potentially, and even those with great difficulty are the two in 64, because there there are actually no api's for for writing in 64 is into into into an array buffer, yeah.

A

B

That is that, if.

A

That's what it is you're kind of standing on your head too, so yeah.

B

Exactly yeah I mean if it were, if it were normal data like in thirty twos and and and boolean z' and strings and stuff, then I'm not sure about strings, but at any rate integers you could easily used an array buffer versus like like an array. If somebody came and said you know, accessing a race is slow. I would immediately tell them using array buffer to transfer this.

D

B

That that would be that, but in this case it's not so so like it's not sort of coagulate to bull, do you have to put it that way? So so then, so then you gotta ask yourself is: is this? Is this validation you know? Should we provide an option for not validating right, yep?

B

Yes, so I I don't know. Maybe I should open an issue or I mean we already have an issue for this. So so I guess I can ask the question and I don't know just shoot shoot an email to or just shoot a mention to collaborators and see see what what distills from that yeah.

A

I think that makes sense yeah the thing I think like in addition to the graph cuz like it there's a lot of data there. It would be good to say here's the proposal, here's what you get today by default and like call out the percentage difference, yeah.

B

Okay, okay, like I, can I can I can remove most of the the most of the ranges from that graph and just keep the keep they're irrelevant once cuz I mean this is kind of an experimental graph with playing around with different things. Okay, one.

A

Ten hundred, that's the.

B

That's how many calls you've made? No, no, no, the number! Oh yeah, you mean the x-axis. Yes, yes,.

D

B

Mean the difference, is the differences become accentuated the more times you run it right. So that's what that little straight right.

A

Yeah, so it's more like on each individual one. What's the percentage difference I mean each individual iteration. Well, like you know, if I make one call yeah and it's you know, it's 30% slower yeah,.

H

A

Be like that would be the clear okay. How much is the overhead? You know? How much are we if you, if you use this this method instead of the other one, you know I'd almost be interested like for each call: what's the Delta, maybe it should be the same like, but if it's like 30% faster and then like here's for a typical one, where we have four parameters and we have to get it, you know we have to get those four parameters.

A

You know the overhead is half as much before we get started kind of thing, as it's obviously related to how many.

B

Parameters you have as well right. Yes, yes, in fact, we we have. We already have a benchmark for that for for various combinations of parameters right and III. When I wrote, this benchmark, I basically used that and just added another case where we have this particular combination. So yeah we were already we're already looking at things like I, don't know like sending a hundred numbers across and and stuff like that, right.

A

Yeah I, you know cuz I'd be interested in just the raw numbers like okay, here's, the here's, the 11 new things we have to add. You know one on a single like if you're just gonna have a single parameter of this type. It saves you axe right, which is oh I,.

B

See: okay, okay, so so write the version of the benchmark that that tests, the tests, the the safe versus unsafe, yeah.

A

Oh yeah I like trying to show the compare it like. What's the benefit, if we provide these these eleven unsafe functions, yeah, okay, they don't necessarily have to write the benchmark, but if you ran it on both, though like, if you just gather the data and then present it as the here's, the Delta benefit you get sort of in the the one.

A

You know that the degenerate case, where you pass in a single parameter, don't do any work and return yeah and then I guess the more typical one of like you know, couple parameters and, and then I guess it's. The other thing is like well. If you actually do some work, you know, does it actually matter and the answer might be? Yes,.

H

Likely no yeah.

A

Yeah, so it's it's clear like we're really this these things would be to serve these really chatty things. Were you doing nothing but hey if the benefits big enough? Maybe it's just worth it right, yeah, yeah, yeah, so I think I. Just think, like you know putting it like that, it would be a very concise here's. What we might change, here's, what the benefit might be, and here's specifically the case is where it matters right and then you know, yeah getting I mean I. Think we just want to we'll need to just meet.

A

This group needs to decide so getting some extra feedback from people would be good to help that okay, okay, any other things, people want to talk about this time.

E

In the chart there easily yeah, yeah, okay, so.

I

Something not very big, so we had the diagnostic working group meeting yesterday we were talking about diagnostic tools and their tears of support. So one question that comes up is if node project is saying that this particular profiling to is Tier one.

I

How does that intersect with the the ear, the napi or the add-on API, because many of those api's are not available through NEPA at this point, so III wanted to basically bring that up here and see what people's thoughts are, whether whether the the level of support on the tier of support the new project claims for a diagnostic, you should have intersection with with the add-on API or not.

A

My my initial thought would be it's a strong indicator that we should have something in a formal API, and you know any.

D

A

Is the preferred place I think for if we want to make it like platform independent right? Yes, so.

I

A

Think III yeah, my thought is just yes. I see some alignment there so.

I

I think in the diagnostic working group we were discussing it purely from from a diagnostic utility point of view and I, don't think we were discussing I, don't think we did enough consideration of how it impacts other vm's and napi I. Think maybe that should be part of the discussion above.

I

What should ultimately would mean that we in part of this working group will have to take some of those easy eyes and promote them to an API right. So if there's an intersection between tears, oh so small and meet something, something is a tier 1 thing for know to support I. Think anyway, I should have yeah.

A

I think that's a good, a good sort of a good track of thought that we should follow. Does anybody else got thoughts on that I'd.

B

Like to see some of the some of the api's we're talking about that, we might want to expose via an API. Is there? Is there a link somewhere.

I

So in the diagnosing working group actually there's a document that Michael was showing and yeah.

D

I wasn't talking.

I

About epi, they were more talking about Kathy of tools so, for example, the CPU profiler or the heap profiler right. So so those have EPS to go along with them, but the discussion there was more around the capability rather than the API is.

A

Yeah I hadn't answered the question. Gabriel I had a gist which was a draft that I'd shared with a few people, but my next action is to create a PR so that we could give you that link okay, but it won't list. The API is it'll, be it's like, like Ali said, it'll basically be like you know: profilers debuggers, it's it's the sort of higher level tools right, but I think it can be used as one of the things that you would pry.

A

You know used to prioritize what you think should be part of an API.

B

A

B

Just trying to figure out like who would who would write applications against these api's, and why would they and I mean? Would they consider it important that they be a bi-stable or are those folks writing applications against these api's such that they are already fairly committed to know Jess and therefore committed to maintaining whatever they write to the point where they do not necessarily need this API stability? So you know what what is our audience for these api is basically I. Think.

A

It's more obvious when you look at the tools but they're.

B

Yeah yeah that.

A

They're tools that are working with node and so they're consuming either node or v8 API seduced a bunch of the work sometimes uh-huh.

I

So right now it it is pretty much weird API is the most release. True, okay,.

A

So any PI we would provide a way to state. You know basically a way to standardize what node will provide. Okay as.

B

Opposed to them just.

A

Using you know, what's available today, yeah yeah yeah.

B

A

Yeah so yeah I think Ali, that's that's a good thought of, like you know, I think we we've already had the thought of an API could be broader. You know we focused on the module needs to start with, but it's sort of the place to put address the same issues in a broader scale and that's probably a good way to think about prioritization on that front.

A

So did you wanna like was there you just want to raise that or as an idea or is there something more concrete? You wanted us to start thinking or looking at yeah.

I

So at this point, I just wanted to get people's thoughts and opinions on this. One I think both working groups probably need issues on this topic. I think the intersection goes both ways right. So what we as part of this working group, what API, we should probably add for nted to any API. Probably this would be good input but, on the other hand, for diagnostic working group I think whether or not something is Tier one. We should probably think about.

I

Availability of API is Right a memory of any PII as part of the criteria or at least discuss it as part of the working group right.

A

Yeah, like you, basically one of the criteria might be to be in tier one. You have to be using an API, api's or.

I

Any bi yeah it yeah. If we are not providing neutral API is for a given capability, then it is, it doesn't make sense to say this is a tier one thing for node to support right.

A

Yeah, like you made that, may it may be that that's part of the yep, like you have to.

C

Have a good test case.

A

But you also have to be using these api's because if they aren't there for you to use, obviously it's not on our top tier.

B

Yep yeah this sounds good, but it does mean that that we have to. We have to have the API is ready to go yeah.

I

It doesn't mean that we, actually you have to do the work yeah.

A

Yeah absolutely but I think that's a longer-term thing like today, we're still fairly focused on getting adoption for what we have. There they'll certainly be a point where we then say: okay, if we've achieved that what else should we start adding right, and this would be a good way to drive some of that I think so.

I

Unfortunately, I have to drop off now, but I think I'll convert this Dewey Hughes on both doctor working groups sounds.

A

I

A

Yep anything else, people want to talk about.

H

Okay, guess we'll.

A

Talk to you all next week, then all right go.

E

A

Thanks for sticking around and no worries, bye.