Kubernetes Storage SIG Miscellaneous, 6 May 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes Data Protection WG - 20200506

Description

Meeting of Kubernetes Data Protection WG - 06 May 2020

Find out more about the Data Protection WG here: https://github.com/kubernetes/community/tree/master/wg-data-protection

Moderator: Xing Yang (VMware)

A

Okay, hello: everyone today is May sixth 2020. This is the data protection or group meeting. This meeting will be recorded and uploaded to YouTube, let's get started. Let me show my screen.

A

Today's agenda so today, we'll all be doing a design review of the warning group cap. There are several people asking about it, so let's take a look and discuss body today and then we will give an update on the container notify proposal which we discussed in previous meetings, and after that we can have some opening issues. If there are people have any issues to discuss.

A

So let me open this. So this is the working group proposal, so our first go over the.

A

Summary and motivation so this proposal we want to introduce a watering group API to manage multiple volumes together and a group snapshot API to take a snapshot of a warden group. So here we listed several use cases.

A

The reasoner is the cap, so this is the actually the application snapshot cup that introduces API to do application, snapshot and back up, but there are also other cases that are not covered by that cap. The first use case is that we can have a boring group that allows users to manage mobile learning together and those volumes could be ill onto the same application.

A

For example, they can use this to group all the volumes in the same stiphu set, and then the second use case is that for us, I think for some storage systems, maybe not a lot, but for some Street systems they always manage the volumes in the same group even for them to create one single volume. They have to create a cook first and then create creative onboarding in that group.

A

So this would be convenient and then the third use case is that we can also use this construct as a source to take a snapshot for all the volumes in the same one group. And then, if the system can provide consistent group snapshot support, then they can use this to do that, and the fourth use case is that this is, of course, is not something that we are continuing, our trust immediately by the destination. That's the potential future use case.

A

You can also use this constructor to do a group replication if a storage system supports it use case. Five is this is for the the volume placements use case, so we could use this one to say that we want to spread volumes across the street pools or stack them on the same soda pool, but this one since there are actually there is already a cap that is out so G pool.

A

So I think this is something that I'd like to figure out whether we can make a decision on this one, whether we should just focusing on this other cap won't in place, then they're, not using group concept for that. So this is a chibi determined, and the six use case is that this can also be used together with that application snapshot cap.

A

So this would be one resource managed by that application. Snapshot. Er D, second, seventh, is that if some applications don't use the workload API such as deficit, then they probably can just use this boring group and they can still manage those volumes together. So those are the use cases that I have here. There are any questions regarding the use cases.

A

A

Okay, so it goes so goes says to providing API team and motive audience to gathering a group and provide an API to take snapshot of a group of volumes and this the third ago. So that's we can also use this one to do. You want emplacement, but we want to decide whether we want to do that or not, and the fourth go is that I think this group API should be extensible so that if we want to support other features, then we should be able to use it.

B

B

I got a couple questions over here after he went through the use cases and goes okay. It.

A

B

Like some of the goals and some of the use cases, they don't necessarily they don't necessarily fit into each other's picture. For example, some of the storage systems does require volumes, get provisioned the time when you specify when you create a group, because for them to able to be able to support topology, they require, let's say consistency. A consistent, wouldn't group could be physically allocated in same black right yeah.

B

This doesn't sorry. Let me finish. Yeah.

A

B

Doesn't apply to existing Warrens in the system already in occupy. The second thing is that group snapshot doesn't really make sense for especially for some of the use cases either start there, for example, if you want to simple it, just manage the lifecycle of group volumes, a using one group to achieve. That seems to be a little bit too much from my perspective, because you know effectively, you cannot really do consistency. Group consistent snapshots were those randomly grouped together forums and a.

B

Secondly, even though underlying storage system that support group consistence, it's not shock, we ran into the risk where, if the worm are not allocated on the same rack or something that which the underlying storage system does require, however, we still expose this worm group API to the users. It makes this feature not usable.

A

Yeah I think there definitely a lot of things that we need to consider for so in the API section. They're actually like different ways for us to create a warren group, so yeah I think we can go through that and see you know.

A

Do we want to be able to start with something small and then is that something that we can make progress on yeah? Definitely there are a lot of concern. That's why this was actually we started discussing such a long time ago and then I think kind of getting stuck, because there are a lot of concerns. I guess.

B

I guess one of my questions: it's really do we need to support those. This many goes over here right. Yes,.

A

That's why I want you bring this opera is so it's especially I think the one group group snapshot and what emplacement those are very two very different goals. Right, I think he usually, we started with a group snapshot and then we'll in placement. I think this is an something that it was added later so I do want to make a decision and then see what we want to I was.

C

Gonna say actually add even one more goal here, which.

D

C

Groups, yeah volumes that you can put into a consistency group and they are consistency across all the volumes in the group. Okay and I understand kind of the more use cases you coupled together, the more difficult it is to get this across the line, but we're talking about core kubernetes api s, and we want to kind of make them as generic and reusable as possible. So I would shoot up try to at least look at all the use cases and see. Would it fit? Does it make sense? Can we tackle this use case?

C

Okay, with trying to tackle more use cases rather than fewer okay,.

B

Sorry, sorry, would you mean by consistent group over here, he's saying the consistent snapshot or without what do you mean I mean.

C

Kind of the storage functionality that lets you say these group of volumes are going to be internally consistent crash, consistent, not just crash consistent for a single volume but across the entire group of volumes.

A

So force snapshot, at least we do have have it covered here and for replication it's so. We can also mention that one so I think.

C

It's a you know, feature that some storage systems have right.

A

So those newly it's either they support consistent group snapshot or consistent, good replication, so I some some stresses and they kind of support those two as two different features. Some can support both so here I have this consistent, consistent group snapshot here, replication I, think I think you mentioned here actually so here I did mention consistent, good replication here, so it is covered naturally by it's like a sub case of three and fork, and maybe I can add a supple it to make it more obvious.

C

If we can keep moving forward, I can comment offline on this yeah.

A

Okay, sure sure, okay, so I yeah I, don't think this part is important that you know if we, if we want introduce this, so we want to make this one generic and extensible okay and then now goes okay, so I just mention here that the we are not really, although we mentioned this use case, but this may be something in the future. We're not really focusing on talking about replication group here. So that's the non go.

A

Over to the proposal section so here we are proposing to introduce three new CR DS voting group on a good class and group snapshot to create a new voting group. So here we have two ways: one way is you create a group with existing volumes and either you provide a list of PVCs or just use a selector, and the second way is create a empty group. First then create an individual volume with that group ID and then basically add one one new volume to the group at a time and then for snapshot.

A

Group snapshot can be created from a source, rolling group and the restore so restore. There are also we could support it in two ways. Maybe we can start cinco and but then so there are two possible ways that well is actually the second one. We can just create one wall in front snapshot at a time in a creative. All you create a group with all of those volumes.

A

This is one way that so this way, I think it's a easier and then the second way, that is, that we gradually use the group snapshot as a source. So we can just take Cooke's wanna group from the group snapshot with with one api thing.

D

I have a question about that: okay, is this API meant to be purely at the kubernetes layer, or is any of this going to be at the CSI driver level? So.

A

Yeah, so this one wouldn't need to have a CSI traveler. They are guys. I haven't I, didn't really write those out because I want to see if we can agree on the higher level use cases first before type deeper, but we do need to have CSI interfaces as well. Yeah.

E

I think that would be like an implementation detail.

F

G

I'm not so sure that it's just an implementation detail I think that.

A

Part of design, of course, not just this, we do. We need to have that. That's part to make of this design complete I just thought. You know we have a lot to discuss, even just to figure out whether we want all of this or not yes, line.

G

Is the most important I totally agree with that, but yeah.

A

G

Worry that when you create too many options in the high-level design, the implementation can get so.

A

Yes, I think we could simplify things. I was thinking even like, for example, this I can think of like let's start with just this option, then you know, then we can decide why that you add this one day right. So this is one one thing and then for this great group of course, we also have like too few options. We can decide whether we want to start with something simple so but but then we also have this other place of what in place in the use case right.

A

So then, if this is already existing volumes and that doesn't make sense anymore, his wording is already placed right. So for that, then we will need this for volume placement. Then we need to say, create a new volume right otherwise for existing volume, then that that one really does not have a place here. So yes, we need to make some decisions and they may be narrow down to say well, should we start with something simple, you know so, but I think with you.

A

We should at least lease all of those options and then talk about them.

E

Have a question on restore restore okay, so what is the claim state on a restored volume? I mean has to be addressed. This is these are volumes which have data? They just can't be left there and a free pool for some provisioning to pick up.

A

It is the it it has a date yet because this is like when you create a create a warning for a snapshot right. So there's data in it. Yes, I.

E

Mean outstanding PVC could somehow match these volumes and snap. You know slap up these things which may not have been the intent. So what protects them from you know so.

A

You mean how do they know that, like they need to which snapshot to find.

E

Restore not a problem right, you construct a bunch of PVC at the end result of that. What there's no claim on those Peavey's of this part, which means an outstanding PVC good. You know under the covers just pick it up. Wow, oh.

A

Here we are I think here we are only talking about like a new create a new PVC, so I probably should make this more clear yeah. This is more like consistent with what do we have for the existing snapshot feature the exist existing, create warning from snapshot feature, so you need to create a new PVC Friday, that's what we hired currently. So that's what I'm having mine- oh yeah, okay, because otherwise I think I need something else that we need to design something I was just support in other cases, but yeah.

A

Okay, so for volume placement support, then we didn't see many results. There are already straight post exists and then, in the warning group class there is a allowed topology field for force to decide at the accessibility of the volumes to the pause and the notes. I just I mentioned the signs. It does not have a question just saying it's. What the that's the other cab the street before has the capacitive group does not have a capacity is really so ok.

A

So let me go over the definitions, volume group class, so he'll here right, so we have driver of your. We have the parameters. Then we have the sphere to specify whether we support whether this class is force. Group snapshot. Not this is a boolean and then this is allowed topologies so like, for example, this API I was thinking. You know, I, don't know if we really need it. So that's something we can also discuss. If you want to make it as simplified, maybe we don't really need to us.

A

We just need to know if it supports good snapshot not and then, if we want to use for the placement out- and we need to have something like this- maybe it can be just interview of fields instead of a new API object, so that I'm not sure.

A

You know the next one is what in Group one group, so it has a back and status. So in the spec we have group one group class name, and then we have a source here. So so this is a. This is optional right. If it's an now, that means that will create an empty warrant group.

A

Otherwise we will create that with some existing PVCs, so the source. So actually I. Should me just look at this right here. So this is a source field, so it can have a PvE PVC list or a selector. So that means we are select. We are just adding existing PVCs to this group. We are choosing one of this too.

A

Otherwise, if we are choosing this one, then that's for restore. Of course, this isn't. When you discuss it, we want you. Do we want to support this one right? What do we want to just L restore time? We can actually just create a one or in front snapshot at a time instead of introducing this, but this is something that we can discuss and then, if, if the sauce, if this source is now, then then that's an empty or a group.

A

Status field, so we will have the PVC list, so we know what volumes are added party B C's are added to this group, and then we have this indicating whether this one is created successfully or not, see how it was ready and then also having an arrow field, and the full group snapshot also have a stack and status for the group snapshots back. So here we have this source, so we maybe we don't really need.

A

Maybe we can just have a or a good name instead of the subject, though, this is if we need to support more than one it's kind of the source, but since we are not.

C

Real quick was that namespace, oh those.

A

Are names based did I say that I should I need Adams, those are name stays, so the class would be now namespace, but wouldn't group group snapshot this. Those should all be names, taste so add that and.

C

So, just to make sure I got the flow correct in my head. The volume group class object is not name spaced, yeah.

A

C

Who creates the volume group object.

A

C

A

Yeah, let me use it the.

C

User creates it with a volume group, class name filled out and the source which points to some existing PVCs and.

C

Do they also point to a volume group.

A

One good yeah so so warning cooks back has this voter group class names filled right, okay, yeah and then, and then the source could be if we are creating. If we are just creating that with existing volumes, then is going to be either a PVC list or based lecture that select the PVCs that we want now. This one is really for the third one is ready for restore. So if we don't want it- and you know, we can't start without this I think so. This is like yeah.

A

So the first you approach need add more comments. So the first you are for creating a new working group. This one is for creating is for libraries or.

H

A

They're actually.

H

Do is there actually a storage system that supports restoring a whole group yeah.

A

So you so actually I see that loitering plantations. They actually just created one volume at a time. They just create a lot in France snapshot at a time.

G

The important thing would be if you're, restoring a whole group is you want to restore all of the individual volumes and then have the resulting volumes end up in a new group right.

A

Yeah I, so I need to double check and I. Think I, remember, I, look at the imputations in cinder seems to be they usually they just creating the video for waterings from snapshots and then put them into groups, but what I was trying to say they is that I, don't know if any serious matter you require I mean you can only give this whole group snapshot as a source. I think. Usually they just create one in Tampa and of course they put them in this group.

A

So I think that's so so this wise I think it's more for the convenience. It's probably not necessary. It's just community. We just I.

G

Was gonna say that there are storage systems where the construction of the group is such that you can't just add existing things into it unless yeah? Yes,.

A

G

New and so having a whole group restore would be the only way to enable creating a new group with existing data, because.

A

So, okay, so okay, let's, let's start to go back to this, think I have that somewhere, maybe in the beginning, so I understand the case that you're talking about I so that can actually be covered by yes, so create a created empty group first and then create each individual bottom. So those warring will be like a volume created from a snapshot. So you can just you can actually do a create, create volume from France snapshot and then add that to the scoop at the same time, that should work for those systems. Wray.

G

And for those systems you would have to. Similarly, when you're restoring into a group, no, the idea of the group you were storing into before you yeah.

A

G

A

I think this look flow with the work which I think this is not yeah. I think this one is actually not required. I haven't seen evidence that this is required. I'm just saying this is like a communist, but so for those systems they need. You have a idea right, so they can create an empty group first and then it can create.

A

What do you trust nap shot and then put that in that, so that yeah, so this would work, but this means that we also need to I think you need to make any husband they are to for the creative or in front of snapshot case there, because then we need you consider that the group ID that's right. You're.

G

Gonna have all kinds of restrictions about exactly what will and won't work at restore time depending on the implementation in the back end- and my worry about all of this is always that so many beckons have such brutal requirements that, like no API, will realistically so.

A

G

A

Understand I because I you know what done this before so I think this case would work for everyone. I haven't seen any that this does not work the case that does not work is you can for some store systems. You cannot add one individual. What in one existing, even volume to existing group that you cannot do, but if it's new, then that would work for everybody, so I.

G

Would be careful about saying everybody so.

A

I mean because I so since I leave the scene, sender I said I know what is for is Ives, so we can list clearly list us out here like what are the things that might be a problem, yeah, I, understand or definitely I think that was widely. We study this a lot time ago and kind of put it aside for a while. So there were definitely issues but yeah I think we just need to list all of those.

I

So for the group snapshot, yeah I think the idea of taking a group snapshot is good, but then that's going to output like a group snapshot ID right group.

A

Snapshot so you will have a oh yeah. You would have a group snapshot ID right at the end, so.

I

That's something that you did in Linder I mean because we don't we. We definitely don't have anything even close to that at vSphere and I'm.

A

I

A

You can use a name, so what we are talking about names here so I mean I, know I kind of like here should be actually we being Queen, Italy, really cool, we only say names so I think that's the.

I

Group ID is fine, but the snapshot.

A

Right because when we.

I

Understand so from the driver from underneath yeah we're gonna expect that we say group snapshot and then there is the expectation that they will come back and say. I took a group snapshot here is the snapshot ID of the group.

A

Yeah, so you should have that right, you're saying that you can't forget that well.

I

We don't have a group snapshot at the moment. Did.

A

Your group snapshot probably.

I

Snap, a whole bunch of things together: okay,.

A

I

Individual snapshot IDs, but not necessarily have an ID kind.

A

Of floor. Okay, because.

I

We're gonna need this when we do it store right, because we're going to lose all these metadata and the storage system has to be able to come back and say: okay, yeah, you want shot ID five and that consists of these other snapshots. I'm, not sure if anybody does that so standard for other people,.

G

What makes you think we're gonna lose all the.

I

G

My dataset are burned mm-hm, but but you surely you stored the metadata somewhere so that you know what to refer to when you're reconstructing everything you have to have something to start from.

I

So how am I gonna get my so my group snapshot ID from the driver level, so.

A

We were actually able to even do do this for like lbm cinder, so I was thinking that, should you see this so.

I

Like we have like VM snapshots, an ID from the underlying system, okay,.

A

So not really, but this is not necessarily a VM sense. All right. No.

I

I know so so would this be like a synthesized ID that we strictly keep at the kubernetes layer and we don't expect the driver to support this I.

A

Actually need to think about that because I so.

A

In in cinder we, when we implement this one, we did implement this one for a suti system that does not have like consistent group snapshot right. So if they don't really have this concept mmm this to a system, but we allow them to kind of group things together. So it's mostly the information would be stored. The cinder layer would have something it's just a and yeah I think that, should you should work, because that incriminated is in kubernetes database right.

A

So you would have this group snapshot name and then you also have other individual snapshot so.

E

A

Yeah, so here so let me I think this is probably I need you go back and check. I think this is probably what we did. There's you have so so for the warring groups. That would have a group snapshot name so.

I

Like the restore procedure would be, we need to recreate the individual snapshot. Crt use.

A

I

A

Individual snapshots right so you take so when.

I

You take snapshot.

A

You take individual snapshot right and then you put them you basically for each snapshot. You will have this group snapshot name, so you know that they are together. This is how you say: okay, all of the snapshot belong to this group. Actually.

C

Brings up an interesting point, I'm thinking based off of what they said. Why does restore even need to be coupled at a group level when you take a snapshot that hey I want these snapshots to be crash consistent after the snapshots have been created? Why does the restore have to be root? Yeah.

A

I think that's not necessarily right, so I I put that as a optional. Basically, this one just.

G

A

G

They don't have to be group to restore time, but if, if you want to end up with a a new group of things that you can again take a consistent snapshot of, then you might be required to tell the system that at the time you do the restore otherwise the result of the restore won't be something you can turn back into a group. So.

A

You could do that just by accreting a new empty group first, and then you just to create a one loading front snapshot at a time and make sure that Y is added to that group. So that should work I.

G

A

This was not required so.

G

At a high level, that's true, but the implementation details for that will get really really messy.

D

I guess I'm I'm I'm also like I'm agreeing with Ben here that, like how do I, okay, I, took my crash, consistent group snapshot and then I want to get that back. Would I then specify a each individual snapshot like that seems okay,.

A

So you know you wouldn't know at this time right. So you have your snapshot. Have a group snapshot names? You know what a snapshot it you need. Then you, basically just you create a voting group first, then you create a volume from a snapshot. Add that to the group. So you do this one at a time until you got to all of them all of them, whistled all of them, but.

C

The assumption there is that the storage system supports adding the volume is one by one right. Yeah.

A

I, don't see anyone I mean she's, but yeah. At least we assumed her that the those two types of a store system like either block or fire they all just.

D

Thank you, yeah I'm, just thinking from a user perspective like.

A

D

Gonna I'm gonna hide all that behind just specifying a group snapshot and doing that for them.

A

D

Want to make it easier.

A

For them, you want them to just like yeah.

D

A

Case then, that that's that's, you know, that's the reason to have this lot right, so this is for convenience, but I see that joy ever when they implement this one. They still are. They just afraid why the time anyway,.

A

G

So so I expect I guess I expect there will be implementations of this, where the the group snapshot will be implemented by basically co-locating all the data for all the volumes on one thing and taking one giant snapshot of it in such a way that it can't be restored piece by piece. But of course the driver can pretend to restore things piece by piece by you know when the first restore is requested, restore the whole group and then, as other restores requests that you could just provide pointers to.

A

It should be. This should work because at least I have seen this tremendous library a lot of drivers, this increments one. So that's yeah,.

G

I'm, just I'm, trying to think through how we got interface is gonna. Look if if what the back end is actually doing is taking one giant snapshot of everything and the only way it can restore it is to restore everything if.

A

They actually have so they do have those in DVD or snapshot IDs, actually race. It's not like yeah.

G

If what you actually have is just one snapshot of everything, all those other snapshot IDs would have to be synthesized and faked, which there's nothing wrong with doing that, because you can make it work, but it gets messy when you do the restore it's. Basically, the only.

D

G

Store but you're pretending that you have a bunch of smaller things that you're restoring one by one do.

C

You have the ability on your side to restore from a group volume group. Id I mean.

G

It it depends how we write this. We have lots of different backends lots of different software implementations. Some of them would actually be much more granular and have lots of snapshots. You can just consistently take them other ones.

G

The way we would do it would be to just basically have one big thing that we would internally partition, pretend it's lots of small things and then take a snapshot of that whole thing at once, and then, when you do a restore you get a restore of the whole thing, whether even if you only wanted to individual pieces and we'd have to do the bookkeeping on our end.

G

To remember that we restored a bunch of stuff you didn't ask for, and then, if you do ask for later, we would say: oh I already have that here.

C

It is the hardest challenge. Is that we're gonna have to support a bunch of different storage systems? Some storage systems may allow you to create an empty volume group and then add in volumes at a time. Others may only allow you to add empty volumes, and there may be some storage systems that don't allow you to create an empty volume group. They say you must specify you know five volumes at creation, time together and that's it.

C

So we have to think through those use cases and then on the snapshot on the on the snapshot restore side. We have the same problem of you may have. Storage systems like Dave was mentioning where they don't really have a concept of restoring by group ID. They just create a bunch of individual volumes and they would have to figure out how to synthesize a group ID or the ones that you're mentioning been where they don't have individual snapshot IDs.

C

They only have a group ID, and so we need to decide whether we're going to support all of these use cases. If so, what is that API going to look like or, if not, which subset are we going to support and why yeah.

G

And and the latter is it okay, we can say we can intentionally say we're not supporting this case, but we should be upfront about that.

A

Okay, yeah sure it.

C

Would be great to get real-world kind of input from some of the storage vendors on the call and maybe offline if we could to say hey who can support, which mode that way we have real data to operate off of.

A

Yeah so please, you know, comment on the cab. Were they shouting me yeah? Let us know.

B

Another thing interesting: the work here is that do we expect the wooden group snapshot to manage the lifecycle of wooden snapshots created by it.

A

It by deletes it should, yes, it does leave it.

B

You know, since the one snapshot objects created by the ones that wouldn't group snapshot it's visible to users right, it is possible, it is there's a risk that users may directly use those snapshots. Yeah.

A

So yeah yeah, that's a good point, so we could. So those are the things that we probably choose. The certain thing we pollinates you like say: how do we prevent those? If you know they, we don't want them to delete one snapshot from a group snapshot and you know, then you could cost the whole course I'm sure to be invalid right. So we need to think about how to disable that not sure it's a we need a web cook. The whole thing mm-hmm so yeah. So things like that at me. We.

B

I mean if, if direction is that women creative woman group snapshot where some newly Agassi as I call I, would say, including one snapshot as in in the media, where he doesn't make sense, because, because eventually, the won't group snatcher I should be managing all this and exposing those two and user seems to be a lipid too much. If is, on the other hand, we dedicate individual snapshot creation, then yeah. It makes sense. We.

A

Just natural well be created yeah, they will be created. Why.

B

It has to be created if you're going through the CSI call directly.

A

That's how actually your well, so if you look at the this case that we're actually creating one at a time is not like either. So, if you look at this we're actually either we are adding, because this symbol, PVCs or we are I- owe ya doing a snapshot not talking about the wall in the group right it took about on you. The snapshot groups. Nash up are not here.

D

A

Part: okay, yes, so I think that also is tied to we need to decide. How are we going to do restore if we can only restore by the time, then we do need to have those right. Otherwise, maybe we can look at you know if we only need just for this one type of restore just you know we have just restore from this group snapshot. We don't really need to worry about the individual pieces, then maybe we don't. We can hide those, but also I, think for maybe for other drivers. They couldn't really support that.

A

So that's the thing. If we want on this one, do you work for everybody? Then we need to expose those sure.

B

A

Yes, I think that same thing we need to decide. I know those yeah, that's actually pretty yeah yeah. Those are the that that are either weird right. So you really suppose you use an either screwed snapshot, but then you also have those you need to snapshots that are showing as resources and how are you going to protect them from being deleted by accident right? So those are the things that we knew needed is we need to decide how to what I do.

A

A

Okay, so, and with so for group snapshot that we have status, that we have this ready to use flag and also we have a list of bonus snapshots, so in volume snapshots that we have the group snapshot, name and then the here so I have this slice, because if we want to support both, you want to do is this just for both the useless. You take a good snapshot and placement. Then then this could be in you know, different groups, that's why this is a string. This is a like array means that just one name.

A

So so this is the llamó example. So if we create a so this is the group class. So it's a group snapshot. So this is supports group snapshot. So we have the statute set to true, and then this is the warning group. You lost you this class, so this basic says that okay want to add all of those volumes that matches the label in to this warning group and that's that plans to this one, a group class one in a good snapshot. We just specify the source.

A

Is this warning group and and a PVC belongs to this rolling group? We, so we all have this group name here and then, if we also do that for group for the policemen's, then in the in this Placement class, that would have this topology is and then in the group we specify the placement group class and then the PVC, then you know it's possible for it to belongs to both the motor groups for group snapshot and also placement. Then in our case it belongs to two different groups.

A

So here we supports a one group that supports group snapshot and also supports a placement group, so yeah, so this will get very complicated. We really want to support all of this.

A

Yeah, so that's, that's it I haven't add all the details on in the see what we are going to do. It was, you know, CSS back and how the how this controller would handle this and the Laurel details, of course, is not not here. Yeah just your joins. You get some high-level feedback. First.

A

And so obviously, I need to update this and there's a lot of concerns. Let me talk about today, so any any other comments on this cap.

C

I'm excited about this. This cap I think it's a gonna, be a fun and evolution of the current API. Do you think it's going to be challenging because there's so.

D

Much different.

C

Types of storage systems, so what I did was add a item to tomorrow's SIG's storage called to talk to vendors about this? Okay, we can say: hey, you know, we've got this kept in order for this to work for everyone. We need to kind of know what works for you and we can start getting their input. Yeah.

A

Sure yeah that sounds good know about this part. Do we still want to consider this for placement group? We have a placement.

C

Group is important, I know that was handled by the by the other cap right.

H

C

Of independent from that yeah.

H

I would say for now because the whole volume placement discussion is still ongoing. Let's leave it out of here. The puzzle to you know kind of.

D

Implies to scope.

H

And avoid confusion, and you know in the future we can look at potentially adding in but okay yeah for.

C

Now I think like.

H

C

Think, quite before elfin I have a story for if we plan to do placement via volume groups, we should have a past, at least towards that. It doesn't necessarily have to be part of alpha that we should scope yeah.

H

But I think right now like right now, this design is already tackling so much yeah.

A

H

Makes it really it's really hard to follow.

C

Us to kind of paint ourselves into a corner by focusing on one or two use cases that.

D

C

Then coming back and realizing it doesn't I.

H

Would so a suggestion I have is instead of like, instead of sprinkling, all of the use cases together like separate them into each section and be like this is what the API looks like for this use case. This is what the API looks like for this use case. Okay, so you want to set that away.

A

Like the complexity.

H

Instead of placement separate.

A

Those two basically yeah.

H

Because, right now, it's like the the design is like switching back and forth between all of the different use cases to explain each of the fields. So it's like kind of hard to get get no good.

D

H

To end picture of a particularly use case- mm-hmm, okay, and that way it also gives us the ability to be like you know, for the first phase, we're gonna tackle this this one or two use cases for the second phase. Here's like ideas we have for how to address the other use cases but like we don't have to you, know, really get iron out. The API and you know, get the details for these use cases. Yet mm-hmm.

A

Okay yeah, he says.

J

So I have a question: is the provisioning user, the user that's doing the provisioning is the plan as they would be aware, you had the two bullets of whether you add the volumes to the storage group at the beginning or whether you.

D

J

Tag the volumes dynamically with the storage group do they need to be aware of which drivers support, which method or is the expectation that all drivers will support both methods, which we've we've heard some kind of fit with one or the other, so I'm at Dell, and we have like five different systems and they do things slightly differently. My concern is: what is the user experience when he tries to provision against yes.

A

Yes, I think that's something that we need to discuss right. So, if your drivers, some drivers, can only support one way, the other driver or can support here in a way. Is that what you're trying to say or can they do things the same way, I think all your drivers support the same kind of weight? You create a what a group well.

J

I'm trying to understand what the expectations are. Oh.

A

Okay, I well I think we need to decide like what, which was the way to go right. So so I think this this one most drivers should be able to support this, but this one I think is very useful.

A

I also think that you know, of course, then decides that you're, the the PVCs that is selected has to be correct. Right, if you say it's a consistent group, but if they are not on the same tour, maybe your system cannot support because it's a good place. So maybe that's a problem so so I think this one, the first one would be harder it could you know they? Could it get me wrong depending on what your support? But if you just want to group them together, not really saying I'm guaranteeing consistency, then that's part.

A

Okay, you just group everything together that that is in the same application, so I mean think it depends. There's.

G

Also, the concern that, no matter what we design like the people who consume this API, were the applications that consume this API may prefer one or the other mechanism to the inclusion of the other one and that's gonna drive what people actually implement right. If we design this to give you two options, but in practice everyone uses just one of them, then vendors are going to ignore the other one, because yeah so.

A

Yeah, so probably that's something that we would also like to have some input from vendors right. What do they think is more useful? Maybe we can start one too that you implement with rather than supporting both right. So maybe that's something that we also like to well.

G

I been ders, you mean like the application vendors who would write the same issues right, not the storage, vendors I think.

A

We need input on both right, so such a vendor just to.

K

Give one input they're going to be applicable from a store builder side as well just take a case off the replication group which we might think or to use that as the use case for this, because in the Amonte we have some construct, like the application group, where we use the single construct as a volume group and pivot group and the application just creates this volume group based on the applications, which means to be consistent from application point of view.

K

And then it just binds all those PVC in a group and take a consistent in a crash. Consistent snapshot of all the PVC which are part of that group and then takes the copy on the other side. So the same can be applicable for the back of a swirl right. So it all depends on that. What all application we want to bring into that consists, the group and how those are related and what the requirement in terms of honoring a crab mentality right.

A

So if your support application consistency, then yeah, then we- and there is another proposal that we are working on, which is two choirs application. So yeah, then the application can do something like that. I think we were talking about earlier. It's like a story system, consistency so.

J

A

J

Was we would always take consistent snapshots if they were a group, but if we're going to allow certain storage systems to to not to take snapshots of groups but they're not going to be consistent, there probably.

D

J

To be a way to expose to the user whether the group is completely consistent or not.

K

One we are talking about application consistent and where you might have distributed application using this group or the respect, is the crash consistent, so I think really. We are talking here in terms of the crash consistent snapshot right, not any application level.

A

Yeah so right yeah, so here we're actually talking about because for publication level consistent and then we can achieve that using the this execution. Core content notified that.

K

A

See something like that: yeah I think we are already at the hour, so we can come back and talk about this more later. I.

K

Had one last question before share with us, so we are also adding a support to add any previously in their group dynamically right depending on the leader, because just to start with, you might have created that volume group, but later on, you added more volumes through the particular application. So, yes.

A

You're talking about adding a existing, are you talking about adding what so after group is created, adding in why existing warren yeah, yeah.

J

So I think that's the.

A

One that we need to be a little bit more careful, because some sources cannot do that but-but-but. If we are supporting this. You know this first case that we actually kind of a saying that that is. Okay, you have existing.

D

A

So I mean we need to at least describe what's the problem with some assertion. They have to know a group ID ahead of time. Monday.

K

A

K

Example, I can give I suppose you have one application for that. You are using three-volume the street or they start doing, and you have some replication policy which just creates a group with all the PVCs being part of that group, and you take a snapshot consistent set of that group, and then you replicate to the DSi now.

D

K

Some point of time, I had another two volumes to the same application, but consistency group might crash. Consistency might again depends on the volumes which are part of that application. Right, so I want to again add those new easy to the say me acting volume group and then take that snapshot. So that is a valid you screen I.

A

Company last and yeah I will just say that that's it I think there are some stores and cannot support that. So, but that's something we definitely do discuss and see what whether we want to support.

K

B

A

Yeah, adding existing one at you, yeah, definitely yeah, okay, so we'll stop sharing out I think we have some other copy blog. We will continue in the next meeting thanks everyone.

B