Apache Mesos Developer Meetings, 30 Aug 2018

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: GMT 2018-08-30 Developer Meeting

Description

Agenda: https://docs.google.com/document/d/153CUCj5LOJCFAVpdDZC7COJDwKh9RDjxaTA0S7lzwDA/edit

A

B

C

Then you mutant, if you're, trying to talk.

D

I don't know if anyone's gonna be joining but um I think it'd be great to walk through the stuff. He's been doing all right.

E

C

I also haven't looked at it yet so we.

E

Should we should probably reboot this meeting and maybe a working group? That's like got a clear mission or something like about like I, don't know, go to a developer or productivity or something you can because three now you know I, don't really know what kinds of things land on this agenda you can what's appropriate.

E

I, don't know who's interested in driving meeting them.

C

Wait right Oh, like a new working group meeting that yeah yeah.

E

This is like a remnant reworking your anything yeah.

F

B

Just thought you said: there's something to choose Gmail to RSVP, to Yogi's it's what yeah beta yeah.

E

B

Would just be honest, you know I don't want to go, please if someone said maybe even come by to our experience. Yeah.

B

The next time it gets good doing basil will be sign, are sorting like this. It's recording, okay,.

G

C

Here you guys able to hear us or is the background noise too loud.

D

Okay, guys sorry I just got a call, give me well. Why don't we start and then um I'll be back I'm over here, but I think we should start.

G

Alligator benchmarking results I've been doing.

G

Let me first share my screen.

G

G

All right, so there have been several performance improvement patches in landing through the alligator during the past, and so I decided to do a benchmarking across the previous several releases and to see where we stand. So it's a it's. A quick benchmarking. I just first want to use one particular test. The hierarchy alligator decline offers. So what it does it simulates a number of frameworks that have a fixed amount of bar once they have reached their targets. They start declining or subsequent authors.

G

So this has this Bureau is optimized various clan and wrong man. I, don't wanna call my five machine, so I didn't repeated a test, but I fixed the frequency so, as a result is very stable to set up as the data here. She is one once other agents with fifty frameworks and I only showed the first 250 allocations and and no code I've set in this step. What's the decline, China.

F

Communicator or yeah right within seconds.

H

G

Into max what is it next, Oh int max so again, so we're.

E

Actually know here which I go.

E

Image the dirty work: we don't suppress the key to finding when I keep seeing work filters right time. As my contingent change. That's when things everything could have finished everything good finishing the first side, so I get. This number turns out. There's no work to do which the meet your targets. They start defining all right.

G

Whistles, so the most interesting one, here's why the first time latest I have to admit I had I didn't go through the benchmark to code itself, but now I'd look at it seems, like the framework just keeps declining than.

H

G

Yeah, it's always declining yep. It's.

E

G

But we still need to make what happens as we build up more more filters and also we need to make allocations way other than the except I. Don't think it's a fundamentally different from a normal education cycle.

G

We still have to go through all the logical pathway.

G

Do we even make offers we make offers yeah? How do you know so? Well, they are log training out. Maybe let me share my.

G

Let me share my first screen.

G

D

I mean I just got back and now you're sharing stuff.

G

The princess logs like like this, so each allocation we make, we I think we always make offers after filtering on.

G

G

Going to get offered all the resources and we always the way that we treat office, we always define and set up the next few seconds, and we just keep doing that so I think each time its framework, what decline one agent forever so periodically this thing things we have a thousand agents. It take like a thousand cycles for each framework, not getting any offers at all I think something I, measured that the home I only measured the for first to 50 cycles,.

G

Basically, I think, although, like it has this weird behavior of keep declining offers it basically, it has the same logical path as a normal application cycle.

B

So it's pretty much that I don't study anything he does they're, not studying, and it has to them. Okay, I think it's confusing! Okay, so I you say it says it's simulating them having a fixed amount of work to do so. Basically, similarly, the Pringles that are always declining and all the efforts. So what so? This.

G

Is like benchmark in alligator with which it's done, it doesn't involve like master or task loans, so so I think the consular started with some resources already allocated. I think it's, it seems sense like we simulate frameworks that have mixed have fixed amount of work to do which had already been launched. So there are like maybe half of half of resources. How are they allocated, although there's no tile or anything.

B

Okay, so here is how the animation cycle latency changes, small freezes are other right. You.

G

Can't say that yeah that is correct, but I'm also saying that DC is also similar to what happened when you do normal. A location yeah done that there would be a threat to what happens unless you're super, essentially without super scampering. So it's basically simulates as posture without some question. Gotcha.

C

And there's no turn the tasks run forever. No.

G

There's no task actually per se like it, so the way that we simulate have a fixed amount of work to do is just make half of asian resources as allocated. So in that in that iraq, you.

I

Allocate a benchmark there's! No! I right so I guess by that I mean we don't recovery. Small is wrong. Fire.

G

Okay, so here are the results so on two rafts, but they are the same data. So this is the y-axis is, is allocation, cycle latency and then why are two x axis is the allocation cycle. So I measure 50 allocating cycles and we can tell for release for for, for different builds, 1, 5, 1 6 x, one seven and the last one is the current master plus a copy-on-write optimization.

G

So so as we as we make go through the allocation cycles, more more futures that build up so so the latency naturally increases, but we can see our optimizations are getting getting. Working done are getting work done, so the latencies are like better across releases and what happens here is you go? That's 15.

E

Like is this gonna skyrocket, operating yeah.

G

J

What what's the special.

E

E

Yeah wide look looks like like exponential, like it's pretty worried. What happens if you.

E

J

I don't understand why we.

E

Have pick a number that correspond with like it starts to go yeah, it's like the critical point.

C

Happens right at the end of this test.

G

Yeah it shouldn't, you, don't seem to be, and it's I.

B

Guess we can try writing to this with motivation, see house where you have well, it sees like you at some way to which some number of creators, and then it's probably it's- probably becomes exponentially build up like problems.

E

E

In this cluster you could optimize it I think, but normally you can actually get it in every permutation or resources filters on a box. You could have flipped it like one CD and two and then read that was like.

B

Interesting is having the access the number of Fraser since the allocation cycle and Steve.

I

Yeah mom's into in the number so.

G

Here here is here is a seeker log, so it doesn't see. The numbers seem that the number of almost zero.

E

It starts anymore, it's not right because everything has been meaning is, but it takes us a very long time to do that.

E

You know it's not very, it's not a very long time, it's like to max the previous term, which makes me think that, like, if you have a number, it takes X to do something I think it's already a stabilizing yeah yeah, already like when you go from 48 49 there I would assume that in 48, like.

E

In creating your like an expected randomness of like if you're picking one person expected.

E

Distribution is like she averaged at all. It should be in the middle point. Some of them were on the end of the list, slimmer at the front of the list and on average like, if you averaged all those times like he did really good leave sort of the funders didn't list yeah. You know like it averaged that say you can't ever at the middle and then the next round, it's essentially everything's at the end, go to the end. We get think any so becomes.

C

Interesting yeah, but nonetheless, from my round 26 27 around 50 Terron yeah at around 50. We go out by like a factor.

B

It's a state, stable afterward, because the number of filters in increased- if there's nothing else, go for we're gonna double it snow for super may then it was a motorized air filters. It's not there I, don't think performance yeah, but others we got its tiny. We can create. With this configuration we can test more than 50,000 filters topography. Well, you would enjoy more more occasions or something yeah we're just how alex gets.

I

More, it's really cool that we're getting what 50 percent improvement over 105 now yeah. Yes, so.

G

That's the number here this is normalized an agency to run pipe and yeah, so we are making progress across releases for this measure. For this thanks mark yeah, I think.

E

In practice, this benchmark, there's no reservations or any medical whispers. Yeah right resource aren't that expensive to copy compared to clusters for people for reservations like labels, so we're.

I

I

I

Yeah I guess I thought that, like SDK would just tag one unique ID for resource or something like that. Her reservation I mean I, think it's working. Okay,.

G

Yeah so I guess you're saying you practice were there more than this benefits Amy, it's just also like it. When Cora is involved, we were also there like much more percentage improvement because we're doing like more computation, but this is.

E

Good at least have some grass.

B

Yeah, go there no more.

B

A

So you wouldn't be growing. Oh.

E

Okay, but in the lodge that kept going more times right.

F

There, oh right yeah, we missed the pledge over.

E

Here, yeah, yes, but it's interesting talking to one user recently, they seem to have actually noticed. There were performance improvements to be on without us really having to communicate it too. Much so must have actually.

H

Yep: okay, yeah! That's.

G

That's right yeah, so this is just our luminarias kitchen benchmark. Oh yeah, we're working on some like more coffee house. That was also including and stay tuned.

D

What's exciting, it's exciting to see the stuff get an improved huh yeah.

E

Yeah we're not done either I know the.

D

Best part yeah, that's the most exciting part.

D

Yeah, no! It's it's really good! Okay! Well, thanks for presenting that stuffing boy.

C

D

Have a feeling that there was a different protocol when Judith was here for getting people to join these meetings, because, let's join.

B

That, while she was speaking, everyone and 80% and then fighting billion, so it was more active, yeah. Okay,.

D

I'm gonna I'm gonna chat with Matt Turks, but about having someone step in and help help do that. Well.

E

Maybe like maybe this meeting should become like a working group. That's focused on something because, like honestly, this presentation that magnet me, we really doing bring them performance. One yeah make.

D

Sense, yeah I mean I, mean I, don't think we need this meeting per se either. I think it was just an opportunity for anyone to be able to come and chat about stuff, and so maybe the right thing is to just keep pushing hard on the working groups right now, rather than trying to focus on a catch at all. Yeah.

E

Probably the gap Bri and always like I, don't know one where we talk about programming, c++, amazing stuff, like the build and abstractions we have librarians the abstractions do that yeah yeah.

K

E

Probably some room I.

K

Think, like one catch all that we have is a community working group in some sense, where we talk about lots of stuff from freaking any what we are working with the mostly focus on stuff around community interaction contributions. But that's another catch all come on in group where we could. Actually that's why we cannot talk about other working groups that we could spin up, and you could start talking about it so have that.

B

Okay, it's announcement, I guess this year, P for mrs. Khan was extended, if you say, and if you don't want to try to New York and you might not have to like poor it's a bad idea will remove conference you it's mostly al anbiya and when you know someone we sent tonight. So if you haven't submitted anything yet to thank and the format of the conference will be different than the previous mrs. stones, it is will be a user conference, so will be more and forward to media by the community and secure.

B

Is that appears more to you? Then I'm gonna be right for you.

E

D

I'm back sorry, no no.

C

Someone had mentioned that Canton Ruth was maybe this. You guys done that you might want working group leaders to do like summaries at some point. Is that something you don't, though, yeah.

B

We don't like that and I think you know what's going to talk to people, and so but one of the ideas is, do you have a presentation for me great, you see what what's new tops minute talk or.

E

Furniture or just I think what we wanted do.

B

You have that evokes another day that you put sample it's being done. What's in a robot and then talk like a in-person meeting or something yeah we haven't fully defined. The foreman would be like a even a just working group.

E

Working groups, young or just knit one anyway, and it ain't gonna- be good. Yeah I mean, if you just say what anyway yeah Ruby.

B

Okay, you could such mean Ella. What's new, you know working Brooklyn's, don't spin another thing on the outside, just think they more salacious a few are just missing something else. They never explained them.

K

So September 4th that's right after the long weekend. Okay, so it's.

B

A whole Tuesday then, will be happening today after the weekend.

K

They're not taking next Friday adding to the long weekend. Okay, oh.

B

Yeah location change will be good for your nose. Yeah.

K

I guess I had to change my tickets now.

K

Then I can a back channel to another hotel and everything and book my tickets and now can I blow blow back on the face.

K

That's wise thing to say: save some money for the company, but it looks like oh.

D

No, no, this is still gonna save money for the company. Oh yeah,.

J

J

H

She say pen recording.

D

That's not that's not why the decision was made, though, but it is cheaper. Okay, all right anything else. We want to cover.

C

B