Antrea Community, 24 Oct 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Antrea Community Meeting 10/24/2022

Description

Antrea Community Meeting, October 24th 2022

A

So good morning, good afternoon or good evening, thanks for joining uh this instance of the entire community meeting uh these uh today, this is the last meeting of October, which let's say the pre-allowing edition of uh the entire community meeting, which happens on Tuesday October, the 2050 for Europe and China and Monday October 24th for the United States and uh for today we don't have a much on the agenda, but we have a discussion proposed by holyang about having a more fine-graded updates for Andrea proxy and the reference issue which I'm sharing on the chat is a issue.

A

4291 say that I will let uh Hong Liang provided the details. Please go ahead.

B

Hello um on the share the screen: okay, could you see my screen now.

A

B

Okay, thanks: okay, just uh let's start for this topic, um how-to motivation. First I think it can avoid unnecessary updates for overflows rules or IDs Windows, not net force. Always another motivation is to use for for this. For this reason, I I could refactor the code of proxy and it can be more clear foreign.

B

B

And, of course, not all service has all these trivials, for example, we'll follow the balancer. Oh sorry, sorry for class IP. This is a dozen that has restarted how grass IPS or external traffic policy attributes.

B

Forces are updated, then responding flows, exercise and routes for the service should be, will be removed or faster and turn it and and then reinstall it. During this process there are some necessary updates. You can't do that, uh how a house two cases for the first cases. Maybe it's not a mix. Maybe it's not it's. It's not a very normal case, but I just I. Take it. For example, for example, updated attribute, plus the IP for notepad service in current design.

B

We know that the installed flows are these two flows are how simplified the flow accents, because it's too long and just simple, we can see the flows and IP sets, and we still this, but when we update the attributor cluster IP from one IP to another IP for this node password, maybe it's not an awareness is very normal, but in currently run um we will remove the outflows and episode entries for the this will will be remote will be only known and we we will.

B

We will reinstall the floors and actually sets every side entries that is as we as we can see, um and the only the only follows that should be updated is the is this one and because it starts the chance of class typing, not Adder attribute like Notepad.

B

As a result, it's unnecessary to update the flows and episodes and the episode entries colored in Orange for another case. Maybe it's not not very normal.

B

We just updated. Oh sorry, maybe this this is this case. Is it's normal? It's a normal case.

C

B

Just updated the node powder attribute for multiple service. We can see that we only need to update the IP set entries and the flows related relation to the node port for cluster IP for class IP flows as we don't change the class IP. For another part, we don't need to uninstall the flow and reinstall it.

B

Okay, that's the that's. The motivation um for the new design. I have listed all attributes for service and the describe describe into corresponding updates um for cluster IP. um Well, it is updated on class IP and the roads should be listed, which should be updated. Now.

B

um uh For this attributes, for this attributes, I think um only real, only related flaws rules or absence should be updated. Notch no need to no new tool, uninstall them all and then reinstall them. It's unnecessary.

B

uh Also, I also I also made I'm also made a procedural code. We can- and this is the this- is the install Services function uh you put in this new design we can have well, we have how clear code constructs we need to and we need to get if we we need together. If some this flows, exercise or root should be updated, then we needed we we, according to these flags, we up, we updated, we updated the flows, rules or episodes.

B

Okay, that's the that's! That's all any questions.

D

When I just want to mention that some attributes of a service are immutable, for example cluster IP, you can never change card typing. Once it's set.

D

Maybe not useful yeah.

B

um Yes, that's the that's a good question.

E

um Sorry, then sorry can cross the ID be set to none after it's sent.

B

uh Yeah until proxy we're doing the process, the class ipu is not um is is 100 science, it's user headline service without class ID.

E

Yes, I guess that, uh for example, is class, IP originally is num and then it's set to not now I'm, not sure this is allowed in humans. If you maybe we can validate.

B

uh You can't design, we also do precise, the non-class, IP I. Think.

E

A

In any way, it seems that it can, it can be updated or set to a user desired value.

A

According to the kubernetes CPI yeah.

B

A

So it can, you can um set cluster IP to your own value when you create the service, but then the API specification states that this field may not be changed through updates unless the type field is also being changed to external name or the type field is being changed from external value name. So let's say that uh you cannot. You should not be able to update the cluster IP.

E

Yeah in case, in the case that if someone really update the to the type to external name, all together, maybe we need to also delete all the rules related to the class ID.

A

That's right, yeah, I think so.

C

B

B

Maybe we can discuss later apply.

B

I I don't have I, don't have very.

B

A

But I I think that uh it it does not change. Your value does not change the value of the presentation or of the feature that you're going to implement. Because is it just that that specific use case where a cluster IP is updated? It's probably not going to be supported, but all the other use cases I believe, makes sense.

A

um Out of curiosity, do we have also to worry about the end point updates and do some changes, whether around updates or those are not. Those are not important because those are just part of the groups right.

B

uh Or for this for this one um we need to, we need to update an important group, sometimes, for example, for load balancer and when, when the external traffic policy and the internal traffic policy are the same, we just need one group and when they change the world, for example, we change the external traffic policy to collab to false to true, and anyway they are not the same.

B

Then we need another group, one group of four just the local endpoints another group for all importance understood this change will need to update any point uh when we don't need. Oh, let me see we don't need to in disk. uh uh In this case. Maybe we, we also need to add add a add. Some floats for remote endpoint uh um for the beginner assume that all external and the external traffic policy and the internal traffic policy are both uh local, then remote and the point flows are not and don't we don't.

B

We don't need to install remote and the point floats, but we exchange the one from local to Cluster the we need another group and this group under the flows through remote endpoint.

A

D

Could you show what the code looked like to express this table that do you have one flag for each uh for for each of for each column and one function for each row.

B

Oh yes, yes, a half! Oh, how? How.

B

um For every flight, I have a function for update before maybe account. Second, oh I.

D

Can send what does if need update any points, look like with.

B

The mint concerned.

B

No sorry got stocked.

D

I I mean I, just wonder whether this if need functions, have many duplicate code, uh or we could just you. You use one function to calculate o.

B

uh Shh, um how sort of a house a house out of this? But after many times we I I, didn't find a way to subscribe this and find the I didn't find such way, which I mean I need to restart.

F

B

Okay, uh like this.

B

But you you mean this: this functions, yeah.

C

Yeah, okay, it looks like a a it. It looks, clean, yeah, I, think.

B

That's the way how how a function for air reflex so.

B

Also I also at the also um they also have a process function for every flag like this, or maybe it's hard to reveal, because the change is so big.

D

Just one one minute suggestion: if each flag and each update function, uh uh if one flag and one update function, has no dependency on others, maybe you could just make one function to make to update the flows or lose on demand. Maybe you don't need one faster calculator, all these flags together and then and check them one by one. So if you, then, you could just have few steps. Each step take care of one flow attribute and you will also do the check itself.

D

So the caller could be.

C

D

Clear just one manual suggestion, but maybe let's see when we when reviewing the pr.

B

Ank: okay, thanks.

B

Okay, no any questions.

A

No, nothing on my side, just an observation that you know if we, since we are adding a lot of if statements for these updates, we will also need to reflect that in end-to-end test coverage, which I believe now you need to cover a few more scenarios.

B

uh That's a good Expressions I think I need to add. Some new into entrance has the underlined, but for Unitas how I think I have added enough unit has to produce for this for this refactor and because I adjust the I. Just came up with this idea when I read to when I wrote, Unitas for the for entry for the proxy packet.

A

Okay, good, so it's good that we already have coverage from a unit test perspective, good.

B

D

I, have one other, maybe unrelated question. Do we take care of updating any points? For example? uh Previously we don't support. uh Thank you. We don't support uh bundle, bundle at add and uh sorry.

D

A group, the member, how we call it in OS.

D

Oh yeah yeah, okay yeah, do it take care of buckets uh incremental update, for example, Peru is I think we just updated the whole group if any endpoints change right so could we do that incremental update and when a new endpoints is added to the service anyway, just to add one bucket using the new OpenFlow window file method.

B

Oh, no, no! No I! We don't support that because in for the new open floor we just um we just. We can just support the very, very large service, very, very relaxed. So it's with many different points for the job. You mean the foreign.

D

Update I mean the library does support uh incremental, update right the package. You.

B

Mean the Toyota update yeah. uh We don't support that.

D

B

D

Library, deposit right.

B

Yes, libraries particle, but we don't use. We don't use this now. I.

C

D

It is probably more common scenario for serious update. Normally people doesn't change the service class IPO or the service type or the node port and and your dad attributes, and you you listed button because of the nature of service and and points the endpoints change a lot and very often right. So.

C

D

It be better to support incremental and importance update without.

B

D

B

That's that's very good.

B

I think like I can I see I. Think I can add this to the tradition to the to the design.

B

Yes, that this is a good idea.

D

I'm not sure how it could benefit the implementation, but then, if I imagine that if we have a very large service, I I, just at a scale cluster, it already has. For example, hundreds of endpoints and one new endpoint is adding to the service. If we could calculate the data and the only append one more packet to the group, I think it should receive some CPU cycles and memories to construct the whole endpoints group.

B

Okay- yes, yes, reasonable about that!

B

Okay, um I'll! Think about how to implement Implement what you said. Oh thanks. Thanks.

G

um Hold on I have one question about the Open Floor changes, because you know that entry has a catch, a memory cache for the installed open, blind Trace so either if we found that so one attribute is changed and although the expectation is only one before lunch is uh needs updates, but the proxy may may send two open branches updates to the Agents of the floor layer, but only one OpenFlow changes. It happen happening on OES, so I, don't think the the.

F

G

This change will take much much worthy values for a trip, because that's for the the connections between entry agents to OS actually only well open, employee related managed messages are sent so I, don't I, don't know if the effort has has the values as our expectations.

B

Yes, we we just don't change the uh we just changed the very to without changing the closer for single service. Maybe it's not where we was yeah.

G

I I don't know if for um app tables or episodes or the routing configurations, uh we have some memory cache local people entries, but just focusing on open, Flow entries, I think the the work is only happening on the memory agent memory, but it doesn't take any effect on the OS. That's what my concern.

B

Okay, I mean okay. Well, we can. We can discuss it further offline for this. For this.

A

Sounds good, then, are there more questions or comments for online.

A

I'll wait a few more seconds to see if there are other questions.

A

All right, it seems that, therefore, it's all for this presentation and a whole angle are you planning to include uh to propose this change for entry at 1.10.

B

um Before we have, you have um how how things appear and maybe before I can I could get enough, reveal comments, but no maintenance approach that I think it can be immersed in 1000.

A

A

Cool, so thanks a lot all young for this presentation and uh I think we can move to the uh any other topic that we haven't did on the agenda. I mean we don't have any other topic on the agenda, but if there is a any other topic that you would like to propose for discussion, please go ahead.

A

Okay, that uh there is a there are no other topics, uh just a quick announcement from me. You know that uh last week we released the Andrea 1.9 and tia 0.3. You can check uh you can check the release, notes for new features and changes that have been made in this release and um for the next for the next meeting. I think that you have to be aware that the timing might be a slightly different so um because Europe is changing this weekend to uh moving off DST.

A

This means that now uh the since the meeting is based on a British time, a meeting in China should shift by one hour from 12 P.M to 1, pm and I believe that, as for the US, since there is only one week of difference between the time shift in the U.S as well, the next meeting should be at 9 00 PM as usual for the US for China. It will be 1pm until Europe and USA would switch back to DST next March.

A

So it will be 1 pm for the next five months and to say that I believe that is all from me and uh does anyone have anything else to report to comment any announcement.

A

A

Perfect I believe, therefore, that's all for today, thanks everyone for attending and uh let's meet again in two weeks time. Thanks a lot and I wish everyone a great day or a great evening or good night.

C

Thank you cheers.

A

Bye-Bye bye.