Argo Contributors Office Hours, 31 Mar 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Argo Contributors Office Hours Mar 31st 2022

Description

Argo Contributors Office Hours Mar 31st 2022

A

Hi everyone welcome to the contributor meeting, I'm going to be your host today, I'm leo. So, let's start by sharing my screen.

A

Okay, um so starting off with the three edge and discussion for next week, so this week we have john and remington.

A

Some of you uh found something interesting during during triage.

B

um Nothing super interesting, um john, do you have any.

C

No, I don't, I don't have anything really interesting, either yeah.

A

Okay, okay, nothing worth mentioning yeah. I know all right so moving on for the the next one. um So we have do we have volunteers for next week, so we we need a secondary and a.

D

Primary hey, oh, I think I can be primary. I wasn't here for a long time. It's it's a passion for fascia. Okay, yeah recognize your voice.

A

Okay, so pasha is going to be primary next week and we need a secondary any volunteers.

C

uh I would like to volunteer for that uh abhishek abhishek.

A

Okay, uh let me find you here are.

D

You in the list.

A

D

A

Foundation, thank you all right, so we got that done. um So we have one topic this week. I think it was added by javi.

A

uh Is that right? Yes, so javi um you wanna! Let me open this link with you.

E

Thank you. So uh this was an issue that we are seeing in one of our services, uh and uh the issue primarily is that uh in some environments, uh when the data is not available like when the traffic is not flowing through in the environments, especially the non-production environments, um we have seen that uh the rollout is getting uh uh failed. So we discussed last time that we can add some uh inconclusive state so that roller gets paused uh on further uh thinking. uh You know I discussed this with hari, so what we thought was.

E

Maybe it's a good idea to include a new flag if you scroll down to the bottom of the screen.

E

If you scroll down further, I've commented out that maybe, if you can include a flag like ignoring inclusive result for any particular environment, if you can add this, what what this will do is if the analysis lands up in a inconclusive state and if we set this, ignore conclusive results. So true, then it would not pause the rollout, but it will continue to uh go on with the next steps.

E

I've I've done some changes for it. If you want, I can do a quick demo. Is that okay.

A

You want to show the the the code.

E

A

Yeah, let me stop my sh screen sharing yeah you can share if you want.

E

So what we can do is, as part of let's say we can do this, both for background analysis or the analysis step essentially in the rollout analysis, if we can add this field, inclusive, ignoring conclusive result to true, then when it, when the rollout is run- and we can tell us.

E

So when the rollout is done, if the rollout lands up in an inconclusive state like this.

E

uh As this is set to true uh what we can, what we can do is we can kind of so in this case it is uh uh interclassical right. So now what I can do is I can deploy this.

E

In the template we do not necessarily have to make things, but uh the other thing we can do is there are two ways we can control this one is at metric level because even at metric level, we conclude the phase of the metric.

E

So if you want to restrict our change only to a specific metric, we can do so by adding this a slag in the metric so that that metric passes in the overall analysis process like in our case, we are seeing this only for one specific metric uh like the http traffic, whereas in other cases, if you want to put this flag at a global uh rollout level, we can introduce this flag at the rollout analysis level as well. Now, for now I'll just show this one.

D

E

Yeah, so uh in this case, if you see uh the messages, uh it is an inconclusive uh analysis run like this, and it is landing in a pos positive, whereas uh in in the previous case where uh is uh having, uh although the, although the state uh lands up in analysis is inconclusive, because I've said this to.

E

Because I've said this to true the inclusive results to true, it would further continue up and roll out the rest of it. So I just wanted to show this one. For some reason, I had a panic in getting these things going, but this is what uh I wanted to uh bring out the uh discussion. Like is this approach? uh Fine, uh if you go with uh bringing in this new flag where we can decide uh if we can ignore the inconclusive result in a particular environment.

C

Just to add to the point I think ravi actually wants to discuss or bring this to gc's point of to review with jc as well. So I don't see if you see here, maybe I can comment on that. I think more than uh inline analysis uh for background analysis. Actually it is useful because inline analysis there is a field called inconclusive failure limit where we can set like how many uh failures or exclusive values we can ignore. But for the background there is nothing means, even though it is there.

C

A background analysis is like infinite one. We can't really guess how many queries or how many metrics analysis will take. So this one will be useful for background analysis. What what I think yeah if any.

F

Question I think I didn't catch um so if the field is set to true and then your result is inconclusive, what happens next is is? Will we repeat the analysis.

E

So, by default, what is happening is if there is a the analysis is inconclusive. Rollout is paused right yeah, but is this new flag? If it's your flag, if we ignore inconclusive uh analysis, flag uh reset it to true, then it will not pause there. It will continue with the rest of the steps, but.

F

Would uh it be considered okay so basically, and the use case is to use it is background analysis, so it's kind of it's considered successful right in this case right yeah, it is culture, successful.

E

So what is happening- and the main problem we are seeing here- is that there are different in a given pipeline. uh Rollout is deployed in different environments and before reaching to protection right so in some environment, in some reproduction environments like qa or perth, when ah where there is no consistent traffic, uh this is landing in a failure state or uh if we add the additional conditions like result and uh the other result, actual condition.

E

We might get a positive state because it will become inconclusive but uh yeah, but the teams doesn't want that, uh the the requirement that they know that they don't expect traffic when when they deploy it in a certain period of time, but they wait and come and do that manual, roll forward right. So this is something came out of a practical problem. Do you know.

F

What somehow I was thinking that they would want to wait for for the traffic. Is it a use case where that, let's say they have a, you know, analysis and they wanted to be either afraid or successful, and if there is no traffic and the result is inconclusive, they just want to. You know: wait for the traffic. Is it the use case at all or.

C

Actually they don't, they don't want to actually so because for services sometimes they may not have a traffic or not all services will have traffic all the time. So, in those kind of cases they want to see three scenarios right, whether it is will be successful or that it can be failure, and if it is inconclusive state they don't want any manual interaction just to come in and promote because they feel okay say there is no traffic and then it's all inconclusive.

C

uh So, let's let me promote and restore that they want to. Actually if it is an inconclusive, there is no traffic that they deploy. So that's what they want to do and.

F

Maybe is it possible to just use success, condition expression in success, condition to accept no traffic? Basically, no in the success condition resulted feature um basic in success, condition we're supposed to access the result.

F

We have a result field which is json. This realized response from from the metric service, and I'm I'm assuming that if there were no traffic, you would get empty object in the result field, and I guess, is it possible to simply provide that expression that success condition expression would return true if result is empty or if it has, if it's not empty and it has appropriate value.

F

Maybe I can share my screen and show you what I mean uh it will be easier if you, okay, if you don't mind.

C

Yeah, I think that is correct. Actually you can do it. I think ravi has a corner case where he said like it might be.

E

C

Might it might hide some failures yeah, but.

F

Okay, that's my, but let me just quickly show just to make sure we talk about the same thing, and so this that's how success condition looks typically sorry like this and I'm assuming there may be wrong, but my assumption that if there is no traffic, this object is empty. It's basically nothing and this whole expression fails, consider it as false and that's why a rollout eventually became inconclusive.

F

Is it possible to do something like this uh like how? How is it going to be different in in this case this would return true and the role became successful and then, basically, I think you know everything continues. So I'm trying to understand how come we cannot do this.

E

We can potentially do some. uh We can apply some logic in the result field and achieve part of it, but it will eventually land at this point. It can become inconclusive or sometimes what is happening is when, when the data is not fetched right, let's say when the result value itself is null. Yes like there is no data.

E

What we are seeing is that rollout is uh actually failing.

F

I think I was hoping that we can just solve it by using this expression language, that's right! This will make it successful as I understand, because the result would be empty, so this whole thing would give you true and basically success condition would return. True.

E

Right so uh I.

A

Also agree with alex, I think this is more intuitive for users to to use instead of adding an additional attribute in the in the crd for handling this special case. um Having this, all the fine in the success condition to me sounds more natural, so if we have, we don't if we don't have this functionality today, such as alex is explaining maybe, instead of adding this additional attribute, we could implement this this knowledge, this logic. So it's another way of.

E

Right, I think the the challenge has been that when, when the data is not being retrieved, um I think we need to. We need to add that functionality. Then, if data is not at all being retrieved, uh the analysis run is getting failed. Then we need to handle that this case, where you know it should it should allow it to contin uh get some default uh result.

E

We have to the other way we kind of solved it temporarily. Is that when there is no data at all, we are defaulting it to zero and then applying this kind of conditions. So.

F

You think right now it doesn't even get here if there is no result. Yet. Okay, I understood now so yeah.

E

And I don't know how uh you know how easy it is for users to come up with this thing. Instead, if we have a flag, I thought it would be more intuitive, but you know such kind of conditions. I think we need to then update the documentation and let the analysis uh get that value as null or zero and then get into the analysis.

E

Let's not fail the analysis in that case, when there is no.

F

Yeah and the I propose that is because I know that success condition is based on that language, so it was. The language was chosen so that we don't have to introduce bunch of flags because it's supposed to be flexible. That's the the main reason- and I was assuming it's you we would get here already, because I saw tickets where people were complaining about errors here.

B

F

The use case was exactly as you described: people were trying to test in pre-production environments, they were no traffic and they were getting results like the empty object here, and that was the expression was failing so, but maybe there is, there are two cases. Maybe there is another case where the query returns, non 200 response and.

F

Yeah yeah. If, if this is the case to turbulence, I would yeah, I would recommend to consider another flag that will.

F

Okay, to make a recommendation, I I wanted to learn more because I'm pretty sure we already have some vlogs that uh it's like, basically not every metric return, json, and I think we already have some kind of configuration to explain to roll out.

C

F

How to treat that result, and maybe if we have something for jason, I would see if now I'm thinking about the flag, that's a that would say, you know, accept known 200k and let success condition handle that.

G

Can ignore errors flag or something like that.

G

And I yeah and it.

F

To be different yeah, I think it's not so easy for me to recommend. I need to uh think more, but somehow I when, when you explained, uh ignore uh inconclusive flag, I it wasn't clear to me what is it going to to to do and looks like right now ignoring conclusive?

F

Basically, it would ignore errors it. It would work like this flag. It would. It would kind of assume too many things. It would think that if you set this flag to true that means uh known 200 response from uh metric provider means it means inconclusive, and then it would ignore it basically- and I think it's it's not true for everyone like some people, it will be confusing like for, for some developers.

E

uh My use case where I was trying to set uh ignore inconclusive issue is essentially in such cases where we have success, condition and failed conditions and stuff like that or when we get zero values. In some specific cases uh like in the example that we were discussing, and um let's say if I change this from blue to green, and if I try to deploy this one.

E

Is it started picking it up.

E

In this case, ideally uh earlier, as you have seen, it would uh ideally fail. In the uh I mean it will get into the rollout analysis and it will pause the rollout because passing in this flag, where I'm saying ignore intention to true it, would uh you know.

E

E

You know make it uh go through the rest of the steps like as we are seeing here. It will not uh pause the rollout at any particular step. Essentially, it is making that to qualify as uh successful.

E

So yeah, so this is this one we are declaring it here, but essentially this flag will translate into an analysis and spec. uh Let me show the analysis.

E

So yeah, essentially the the metric is inconclusive in the analysis run, but because we are setting this flag, as incompletive run, ignore inconclusive to true. We are qualifying this to be successful so that it can proceed um for the rest of the steps and.

F

Release repressed reload assumed it's inconclusive because I have a feeling I I would really try to see. If we can, I mean because rollout assumed it is inconclusive.

F

The only condition, the only you know, reason why it did so. I think it's because it tried to run both success and failure. Query good falls from both of them and basically and then then it's fair to say, rollout is inclusive. If it didn't even try, I think it should fail yeah. So I would, I would try to see if we can just use this additional check in success. Query basically not result or result is something if and then yeah.

F

I I'm really surprised that you're allowed thought it is in conclusive state right now and I would expect it to to run successfully. Maybe can we just try? Can we see if we can uh I'll.

E

Try again and get back down.

F

Okay, yeah and then the the less set settings will happen. Roll out, I think the the better. In my opinion, uh this is, it looks like an edge case and it could be, and that's why the success condition is, is a lim. It's like it's not declarative, uh set of flags. Success. Condition is a language so that it can be flexible yeah. If, if we can work around this each case using this expression, I think it will just make things simpler.

F

Let's try please, let's check.

G

All right dude: do you still like the idea of the error case when say the uh metric servers down or something like that having that flag or do you think that's not needed as well.

F

Yeah in case of uh metric server down, I think we should fail in this case right.

G

And pause like it would do yeah, currently, okay, okay,.

F

I'm speculated I mean I'm not that user but yeah it.

G

F

Would want it to fail like I would want to make sure at least I I was able to query metric server. It gave me nothing.

G

Yep, that's right. I agree.

A

All right, that's right, so I guess let's continue in the in the issue. The discussion right. I guess it's, the most appropriate.

F

Since we already spoke about it, I'm going to just you know, reply in the issue and repeat what I just said just for the people. So you know by.

A

Cool thanks thanks alex thanks harvey um all right. I get. I guess, that's wrap that wraps up the the the topics we had in the document. Does anybody else have um last-minute topic you wanted to to discuss this meeting.

A

Yeah go ahead, go ahead.

E

Alex I need some help and reviewing the notification engine pr's that I've been training for a couple of months now, if you can get some time to review them, I think that will be helpful.

F

And I know that russia also uh experienced notification engine engineer partially. If you can help me, I will- and I will uh try to yes, okay I'll just try to prioritize it as soon as possible and review the tears.

F

Thanks very much.

A

All right anything else, anyone.

A

Okay, if not, I'm gonna give you back those minutes and.

B

See you next week, thank you, yeah. Thank you leo. Thank you. Thanks leo.