Layer5 Meshery Development, 27 Aug 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Meshery Development Meeting - Aug. 27th 2019

Description

Working on a PR for @giltene's wrk2.

A

Good good well grisha's, halfway on so let's get going anyway sure.

B

So I can share my screen. Okay, basically with playing this w RK 2 for the Total Request count, I was getting the number, but it wasn't really make sense so I'm just trying to see. Where is the point? So the addition right now is similar to percentile I'm, just adding stats total count which is supposed to take. The total count returned.

B

Here and what else yeah? That's the change here and in the script changed supposed to be in two places, so it's gonna push for the Lord. That's very gonna, be the connection and similar to methods and the next thing it should push as total count, and when we come to law we should be able basically to see this value.

B

So I'm, basically trying to understand better the code and make sure complained of a value right. So.

C

A

Was just a recount from last time that we spoke? It was that we were getting I guess it was two things that we were getting through for I/o. We were getting for each individual HTTP request. We were getting I, guess a latency time, and then we also had a Total Request count that was put out.

B

Yeah, so for the fourth year, that's the statistics. We have right now we were getting, but it is supposed to check also in his side, because the total count I am talking. Is that one? Let's say there is certify or equation over time. It should show you this start and time. You see like, let's say from zero seconds to zero zero for millisecond and the next start at that and the next basically. But it's like increasing one percentile at a time.

A

B

Let's say yeah one percentile to percenter, you remember for tag. Willie are key to what I was showing that that's exactly as a percentile high. It's increasing I am assuming the same way. The faulty is increasing like every time. Zero point. Zero is zero point, ten twenty like right and when it goes to the like further, you see it's even going more being more precise.

B

So that's so! Basically right now we get VSWR k2 and I'm also assuming that that's logic 4002 from double our key okay, considering that was quite famous and they were first in the market and basically here just the count percent should be fine to put in Jason just we may need to add start and end time, she's missing right now, yeah, that's I, guess also grish supposed to confirm for for Tia, but I think that's illogical. It should be the same way you're working okay, but we I don't think in the chart.

B

We we will showing you.

A

Here I'll bring an example, I.

B

Can stop sharing that you can show.

B

How are you doing Lee yeah.

A

Good yeah real good Girish, if you're on now you wanna chime in on the bit of.

A

For a oh and it's either.

C

Yeah I'm sorry I haven't had a chance to actually check that sure I will see if I can actually do that. Probably now, but what is actually a question? Chuckle yeah.

B

The question was in regarding in 440th that early did you manage to see my screen when I was sharing that the one you gave I think.

C

Yeah I think I saw yes, so.

B

Basically, the only thing like right now missing should be the start and end the person frequency changing should be similar to whe, to I guess how it's going from zero to hundred.

B

To zero to 1, plus sorry yeah percentile, wise.

C

Okay, I think I kind of understand. I just came in second yeah. Sorry, you guys can continue so.

B

I put in the chart: yeah, that's agree.

A

It's not hard to hear you in your videos not up either here.

B

If you're going to read me the table showing the percentile how it starts from zero to come to one point: yeah, that's if you go down there here, that's a table! Yeah!

B

It's how it's changing.

A

Okay, so yeah do we have the output of WR k2 and for IO side by side that we can compare the two yeah.

B

That's the one for the party- oh, let's the last time with Krisher I'm gonna. Just put here. That's fort area and that's you see is right now. Wl cube is a JSON format. Okay, that's yeah.

C

That's that's exactly the example which I was using for reference last time, yeah.

B

Yeah I just hate that okay.

A

And then one of the things that we're trying to confirm is that way that we acknowledged last time that we're missing is what the the is for each individual. Is it each individual percentile we're? Is it every like Millison millisecond.

A

We're putting out for iOS putting on statistics every millisecond and so we're trying to match that format.

B

It's just the question is the frequency of percentile in wlky's, you like start from zero, zeros and 0.10. That's what in 40 doesn't show that explicitly. It shows only the value of the percentile.

A

And yeah, and so for the 40s seems to be it's as it's incrementally spitting out statistics. It's doing it.

A

It's doing it on start time, increments, start and stop time increments. So for every millisecond that passes I, guess it it set out, and so then, in the format that wk2 is using is that it spits out statistics every tenth or as a percentile every it's.

B

Percent all is seconds. Is it because hmm because it's running short a second, it can't be the seconds.

C

It is a person time yeah.

A

It is actually literally from the zeroth percent, so from the first request to the very final request: yeah, it gets very it's sort of an a sort of logarithmic, percentile, that's being put out because you don't get very granular for the first percentage, every tens place and it goes by five up until 75 yep and then it starts to get much more granular.

B

But the value it's giving like you see like 12. Is it the percentile value? Is it the same in 4000 BC? Then then, if the frequency is finally, we can assume the person showing the values that value.

A

Gray shaniyah: what's the key summarize the problem statement and what we're trying to solve.

C

I mean violently, like you know, we are trying to get similar stats from wrq too. So, as you mentioned, you can see like you know, in Florida yo, like you know, we have the timestamps or the start time and dying. Of course, like not the start. Time of the test is actually at the top, so from then on, like we can actually compute the time when each and every request was sent and time what now I took for it to get back the response essentially and the percentile, like you know, value or the percent value.

A

Specific because we're not for arrow, isn't calculating or exactly delivering statistics on each and every request sent, rather because you here it was a one on a request. That's.

C

True, that's true, it is actually the time. Yes, it is time. Is that's correct, so in that time window, which is like you said, every every one, millisecond I think it's actually computing like an how many requests actually made it so so that that's kind of like a how for IO is actually doing its math I.

C

Think what we should try to do is like you know, maybe not exactly get the same behavior, but actually something close so that we can maybe did use the data like an asked for what IO is actually getting so that changes all over the place. So.

A

The what which numbers are we used of the JSON that's put out here and which values are we using as we're graphing.

A

We're taking the total cumulative we're taking the per histogram count for the number of requests that fall within that latency.

A

We also are looking at the cumulative the blue line in terms of total total percentage request, sent.

A

So I'm still kind of looking for so so I think you said the problem statement, the two don't spit out the same statistics exactly they both have the same data. Trying to give you the same format. What format are we trying to get it in is that it seems like the common.

A

The common measurement would be percentile.

C

You know what let me do this I can actually get the data that's used for rendering the graph so I mean the the response coming out before IO is actually not the exact analogy for I'm doing the chart. It is computed so from the data that 4i overturns its computer in the UI to actually render the chart. So there are two different things now, I think what.

B

C

Easier for WR k is actually to just give the data that is needed for the charts, rather than not worry about Rio data, maybe because yes for the chart, we only seem to care about. You know one or two things like you know, which is the percentiles, and you know the accumulator time that's taken. You know, rather than you know and I think for WR k to, like you know, probably has the data, like you know, I, like Sacco, actually shared his screen, so maybe we'll doc.

C

We can actually maybe draw a very similar diagram, but you know, if chart without having to add an extra layer of computation. Maybe I mean that's I. Think the easiest thing to do, because I don't think it's going to be easy to actually get the you know. The times like is like for IO has been getting the data yeah.

B

That's where I see the key difference is fourth year is based on start and time, giving you the percentile value, but as a doubler kit is more frequency of percentile, giving the devaglia.

A

Is this graph, while we can that data to look be in the format needed for the first graph? Is that gonna be the same format needed here in.

C

A case, what's the second we're up here, you can see mostly what we are using is actually the percentiles, so there was a mini median as actually p50 average and then a seventy-five. Ninety. Ninety nine. Ninety nine point, nine, a max so I think we should be able to maybe get something along these lines. And now the interesting thing with this rap is it's not like a continuous graph. It is actually like mostly points, so you know for the whole graph. It's just like.

C

You know, for example, for this result said, like you know the first result set. You know all it's taking into account: it's actually the values you can see. Like you know, it's a all is taking to account is like, though you know, the total of the overall QP s, the total max. You know, duration, the P 99.9 Eurasian, P, 99 Eurasian, ninety yeah, but you can see that, like you know, this chart here is actually just taking. It account individual values. I think you know we can actually get this.

C

Like especially be you know the piece you know the percentile value screening double your k2 is actually spinning out, so you know so we can think it choose from that.

A

Already have these points because they're so common records they're already pre-calculated, so this one isn't an issue yeah.

B

A

Like it's the other one, that's.

C

A

Thing about when we go to go to format the thing about your suggestion: grecian it makes intuitive sense. Sort of cause for potential concern is to the extent that we ever changed these graphs. Maybe that would still be the same data and that would be fine I. Don't I, don't think that we've completely digested these and really assess that these are the best use the best way to communicate to users that this is what's going on.

A

Maybe they are maybe it's just it's a complex thing and it just takes a while for people to digest. I I think that if this one's not entirely unintuitive I think this blue line here is you just if you don't read anything, you just see the books like a cumulative thing, going from zero to to 100 that if you looked at a histogram before it's, not on obvious that that's what's happening, that you you're, probably always find your P 50 is you've got the most count in there.

A

I think where things start to get somewhat confusing for me is just is it feels like it's somewhat arbitrary 50, P 75 P those lines don't match up with the bucket sizes that these counts are being put into I. Think the counselor every 50. You know it's a bucket size of 50, um which I.

A

Think, logically, that makes sense that well clip that you're always going to find the P 50 somewhere somewhere in the tallest bucket, whether that was toward the end that front or the back or.

A

I think that that statement is true and when you look at the P 75 I guess it depends on the scale that's used. You may find that the P 75 falls in the same bucket or shortly thereafter you I guess you start to wonder: I, guess I such a question. Is this histogram valuable? This is.

B

A

Percentiles that are the the and so okay, so thinking about that out loud, just as we go to make some decisions around how the data is formatted, I think the other thing to think about is that this multi, the having multiple scales on the same axis to me, isn't naturally intuitive that okay, so I've got a high-flow up top here. I've also got a high latency response up top here. Does that mean that there's a correlation.

B

A

B

Be maybe two-dimensional for is beautiful. Eighty I don't like if making two graphs separated well.

A

Yeah, that's the point of this. One was what so here, I think, if, if you just did to you, get what you were describing, what.

B

Like this yeah, but you know I, think yeah. Is it because the response time and yeah that's what.

A

This one is missing is the percentiles that we were highlighting as potentially being the most important what this compared, and it's really the comparative graphs that are they're, not the entire point. You get a lot of interesting information out of this.

A

What what these graphs should hopefully facilitate over time is a separation of or helping people understand, where certain configuration changes affect their performance, where different versions of their software affect performance, where their infrastructure affects the performance and if there was distributed tracing and within here, how much of that was on the within the app versus elsewhere?

A

That's kind of a separate, separate question, but.

B

Then we have right now in the left, the percent and then right count it's different. Yet that's.

B

Is it in the graph right now like for me a bit confusing count the person in the left and in the right count? It's is it always the correlated like percentile increase and then the count increase right all.

A

Right and kind of a funny thing about this percent is that it's going to 110, which, if you know necessarily make sense, neither does the negative 10 can't make sense. Maybe it just draws a prettier graph, but then yeah over here. The count also this this- the duration is 1.4 seconds. So, okay, so only 18 requests percent. Okay. So this is a courage. I go to a bit of a longer duration, one 30 seconds to a Guri Garissa like we're we're, mostly just we're actually kind of deviated aggressing from what we're trying to accomplish here.

A

So I'm looking to you to have a sort of a direction on some of the on the statistics that taco is augmenting to confirm whether or not this is the right output or whether or not those are the right statistics that we need to get I.

A

Think this charting discussion is a slightly separate. I only brought it up in context of identifying if we weren't gonna bring the same statistics, but slightly different I was just bringing it up in context of what we're just flying. I.

C

Think we already had a very similar discussion. Some time back, I mean we started off with work for IO has been probably, but is that the right thing in the long run yeah? We don't need to have to give it some serious thought, hello, the amount of data and the type of data that's being returned. You know it's also important. Like you look at all the little generators like you know all they provide. You is a either some kind of a latency information and know kind of the accumulator count. That's pretty much all they provide.

C

So what for I/o has done is like you know, with that data they'll try to enrich it with some time. Metrics like you know, and and try to chart it out in a sensible way, I mean by sensible, I, don't know to who again, but it looks like I'm in fourth I was being used in several projects, so it looks like I mean it's, it's ok. People are ok, seeing those charts, I, don't think anybody has actually taken data from forward I/o and has tried to do a different chart. I, don't think so.

C

I mean unfortunately, I- don't have the best answer here, but we.

A

Get we don't have to figure out that the charts as much as just the data yep yeah? Just what do we need to change in to PRK to.

C

You know what look I will say: let's you know, okay, so let me do this. Let me see if I can actually get yet. Eva is the chart data that was used for rendering and let's see if that will actually help us just to you don't kind of come to a closure on this one yeah.

B

Then I can see the.

C

B

That can I need to take it. We.

C

Might have to make some changes in the UI to handle the VR key to stuff separately, then for I/o, because now, in the case with for I/o, there is some processing that's happening, which probably will not be needed and if WR get to is going to be providing it at that particular format. Yeah.

A

So that can that could be okay, because I mean we could signal what the data says, but, and it doesn't make sense to kind of reverse engineer and reverse back from what we're trying to display.

A

Having said that, I don't think that we've proven yet that we can't get the same style data out of W rkq as for I/o, so that there would be no changes in the UI.

A

Which I think since we're gonna be making changes? It's probably ideal that we're making them to be in the same in the same format as for a oh. We can so that there is changes all in one place and WR k2, not on you. I.

C

um I mean like I'm looking at like analog data, we got cut from poor audio, that's not easy or I mean this is exactly what I wanted to accomplish like nowhere, yeah, hey, let's not touch anything in the UI I mean, let's just make changed so that the data from the vrk is in the same format as for IO, but looks like it's not going to be possible. That's exactly like know why I said: okay, you know what you know: it's not going to be possible.

C

I, don't think, like you know, for IO, uie or the chart, but you know we are drawing in the UI is actually using all the time data that for IO is providing to the content, because in the front end code, essentially what it's doing it's it's again doing some summarization coming up with a new chart, new data set, which is actually given to the chart for rendering the WR kid to data, seems very closer to that, rather than the raw for diarrhea.

C

That's exactly like what it is so, which is the reason why I'm like yeah I, think that it just be easier to do that.

C

Yeah, because if you look at the chart, okay I don't know who is charged Aaron. You know I'm going to share this. If you can actually share this with I mean this is pretty much like I know what you're looking at.

C

C

Yeah, so if you're looking at you're looking at this I hope, you can see my for IO chart yeah. So if you, if you look at it Sokka, you have to tell me this now. All this guy is using for charting there's. Actually, the cumulative count which I think WK 2 is providing or I mean, as increases I, remember, seeing the count increased, zakharik, yeah.

B

C

So this chart is pretty much like another very straightforward. Looking with the WR key to date and.

B

The first Sun tile is from zero to hundred and how you have this.

C

Is zero 102? You can see right there, so that's 100 percent or ninety nine point three and yeah: that's hundred percent. Can you see the screen yeah.

B

C

Did you just follow my mouse pointer? You can see. So that's cumulative ninety-nine point, seven, one! Seven! That's ninety! Nine point. Seven! Three! Three! That's ninety nine point, seven three. Three, probably that's a bug, but that's final person. You see that this is so. This chart here is actually a cumulative of the counts, which means it's it's kind of what you want to build up off. The total count as a time progresses. That's essentially what it is, but at the bottom they have the time. So we have the accumulation or accumulation of the count.

C

But do we know the time is actually the question. So what this is saying is accumulatively. Okay. This is how it went it's kind of like what this is showing but against. There is this response time in seconds so which means for for a percentile value of like a percentile value somewhere here, like laughs, roughly 50%, so that's like roughly around the median the time taken for that request was somewhere between like 40, milliseconds I would say it's to the point. So do we have that data from the vrk so.

B

Basically, as I understood, we have like, if you go to WT the table, so basically, here you have the count. The total count, the percentile from zero to hundred and the response time, assuming the value of percentile should be yeah. That's the response time if you opened the link in the zoom, the doubler tea table, so we basically here, we have the percentile count and response time and he ended here in the song link just.

C

Give me a second.

B

Oh, is that E is that.

C

The King world link you actually yeah, alright, okay,.

C

B

Please go to the table. Yeah. That's value, percentile total count. So is it these three values we need like percentile. Think of these from zero to hundred like one is means hundred yeah.

C

Yeah actually, potentially I think the value probably is in milliseconds or something yes, so here in the summarization table, that makes clear so yeah.

B

C

Worst time you know, or or the worst time taken by a request was twelve one, five milliseconds and that's exactly what yeah.

B

C

So so so yeah, so we have the day of, or one chart right here so which means lucky. No, so at least that part is.

B

Taken care of it.

C

B

Now that's what I was showing in this see it wasn't with law that I can I did get some data, but it wasn't correct anyway. I am assuming I can fix it and peel it, for these will have these three values in JSON. So.

A

Already here, yeah.

B

But not I will push it to the JSON from last clip. That's missing one. That's I am working.

B

We don't need to change any logic in computation. Let me say it this way, for this rebellious.

A

C

Just don't write this thing.

C

B

C

Now, with the total count here, since we do see that maybe maybe we can actually draw the histogram by computing- a difference in the in the town between the two values like so, for example, between a 10 percent and a 20 percent percentile, we can count the number of requests.

B

Difference yeah.

C

Difference between these two I need to understand the 4io stuff, better and I'm. Trying to understand, like you know what is a window, because in some cases you can see this is longer Oh.

B

Let's show he's ticking on comms, so histogram calmed what's histogram calm, so.

C

This is saying roughly, if I understand this correctly, so this is like.

C

So roughly around the 30% I'll sort of 30 millisecond response time mm-hmm there were 541 request yeah, so in in the bucket from 25 to 35 milliseconds, they were 541 requests. So in the in the bucket from 14 milliseconds to 16 milliseconds, there were two to 3/8, but but.

B

Then it's not the response time yeah! That's it.

C

Is based on what the response time, though, now coming back to yours. So this is what I'm saying so for a second okay I'll say forget about the percentile column and just focus on the response, time value and the total count. So maybe we can do a similar math like, for example, we can say okay between 4, millisecond and 5 minutes are 6 in the bucket from 4 to 6 milliseconds. We can count like the total number of requests. By doing a Mac look mean the values maybe but between like between this.

C

So so this value is- or this is a number of requests at 4, milliseconds I mean like in a minus 1, so it's gonna be three nine five zero at four milliseconds and roughly. If we take this, that's going to be 15 8:03, so this probably subtracted from this might actually give us the number of the number of responses in that bucket of time.

A

C

So so between 4 & 6, then you know it I think it's just one of your direct subtraction, like you know so this minus this will probably just give us the total count in this time window and the same, like you know, from this value to the next we can. We can, you know, count the number of number of requests in that bucket and so that's kind of what the histogram is doing.

C

I just need to understand. You know how. How is this trick your mind? You know because it seems to be around I'm here. It's like to second, but here it is like from 30 to 35, and then here look for example at the bottom is from 45 to 60, so yeah I'm, just trying to understand like how it's coming up with this.

A

When you look at from 60, it goes from over to there from 50 to 60. That's 10, 60, 70 and 60 go through the 60 and then is there a 65.

C

Am I hovering over the right button? You.

A

Are and then, if you go to the neck, go to the next one to the right 60 so.

C

In the bucket from 60 and 70 requests.

A

Yes, there's 14, and so it's just a bad graph or a bad. They shouldn't have all those vertices, there's 14 from 70 to 80, or rather actually should go down to 80 right there yeah. Actually, the they've got a bad, a very confusing and bad graph, because they've got a 202 and so did they're not displaying the value they shouldn't.

A

They should only be displaying those values once not twice, because if you didn't know how to interpret it, you might assume that there were two between 80 and 90, which there are go back to the left. One earn a down if you would.

A

Yeah interesting I didn't think there, just in this case they're hitting the same they're hitting their sang from 60 to 70, whether you're at the 60 or you're at the 70. It's going to give me zero zero at the at the low end. You know, and then, when you go to start from 70 to 80 it's 14 and 14, it would be a less confusing thing if they just had it at once, and they would have it on the right hand, side so you'd, say yeah after that time period.

A

It passed this as the count in some respects. It's accurate that they put it on both the left hand right hand side, because, because they're saying there was a fourteen of them could have been like 79 milliseconds, but they're just normalizing and saying that it was between 70 and 80, and so anyway, of the so there's one thing this meeting is going to close and so I'm gonna. Send you guys a different lengths that we can carry.