Ceph CDS Hammer, 28 Oct 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: CDS Hammer (Day 1) - RGW: Bucket Index Scalability

Description

http://goo.gl/U4b70r

28 October 2014
Ceph Developer Summit: Hammer
Day 1
RGW: Bucket Index Scalability
Guang Yang

A

So the upside for the next one, the bucket scalability or bucket index scalability I guess- is that it's a holdover from the gh summit, so there's actually a pretty detailed pad. If you look at the blueprint, as well as a decent, detailed description, well, maybe we could just hammer through some of his points and look at the status and then we can take it to the list after after the summit. When he's able to talk about it,.

B

um See all things pedal and I had our the open.

C

I'll put in the racy channel.

C

Well, it's freaking start with just a status, so I mean he has a. He has a pull request open for this yeah.

B

C

That, like you,.

B

Yeah we've been through various cycle for free few, uh most of it is done, um and most of it is ready. um The thing the main thing that it already is yeah, but tying it up with the multi region, multi-zone stuff, because the the problem is, we have some logs with it resides within the packet injection. We need to make sure that everything works correctly and there are maybe just choosing different kind of markers to tune for the food, the multi region for further, when we have shouted bucket index.

B

That and that's basically what we do internally or hope we have markers that point. It just Matt, multiple bucket entry, so that we know it when we list back at what point we are at.

B

So yeah, so so the the feature is is going along pretty nicely. We try to grind. What's working on it, we were trying to make it everything obstructed nicely so that we don't try the same thing over and over and over, um because we it was changing the current all packet access Taffy's in synchronous, asynchronous greatest operations, most of it and now everything needed to change, because when you do operations on concurrently in multiple back it charged.

A

Objects, yeah yeah.

B

You know you wanna, you don't want to do it's synchronism on everyone, he need to have a sliding window, and so it's kind of bright and its complexity. But but ok I think that's that working ok now.

C

So it looks like an ideal etherpad and we talked about adding two different list commands, so you can do the regular traditional list. That'll still return, the same, ordered results and would sort of read from all shards in parallel and aggregate the results of the client side and then also adding a new list. Unordered command. That would sort of.

B

C

Shards individually, yeah.

B

We that wasn't implemented yet. Oh my.

C

Side of it again: ok, yeah.

B

um It should be trivial to add, though, but.

C

Ok, some people don't care I, guess all right, but that position I, guess the position identifier needed to be a tuple instead of a just a teeny yeah after the.

B

S3 API at least provides you with with an entry that says: what's the next marker is so you can use that I'm, not sure that most clients will know we'd even look at that entry, so it might be.

B

It might be a problem for these clients, but but it's going to be a different command, so yeah.

C

Yeah I won't matter yeah, yes, okay, um I, remember there was one question that I think I think maybe you were on vacation when they did this last time there was one question about the atomicity of the multi-part upload.

B

How he said well, better.

A

B

Don't think so because you need in the multi-part upload, you need to all the entries to be on the same bucket truck so you're gonna hash, hash value that the target entry.

C

I see okay, so it all end up in the same place. No one matter, yeah.

B

Then I'll guys know.

C

B

Meet because unless we want to change it drastically the way it works, no matter I, don't see why not get in the same shot anyway, yeah yeah.

C

It sounds like a better bet.

C

C

Has Geno Guang I looked at all doing the blind buckets I know they had a hack to do this earlier, but that's.

B

Not that I'm, aware of I'm and I didn't know that he he was involved with that specifically, uh but here I am yeah yeah um yeah. The plan buckets is kind of like the Silver Bullet of performance, job right performance, at least yeah.

B

The problem with blind buckets is the kind of dropping all a lot you're not going to have multi-regional his own you're, not going to have object. Versioning.

C

C

Guess you could have an alternate version of object person that only is of the llh.

B

Well, you can have an s3 like a swift like object, versioning, maybe yeah.

C

Yeah: okay- let's not worry about that then yeah! Okay! This is this going to conflict horribly with the the object versioning branch that you have going probably.

B

Well, in terms of the object, avenging branches, I had uh I kind of did some a few major cleanups, um so my changes go out or no other.

C

Oliver three yeah all.

B

Over the place, and so they're gonna be lots of conflicts, but I don't think that anything specific yeah yeah um big, because the we didn't really touch the same areas other than in tactic changes here and there and I'm not not afraid to merge. I well, I think before have tried to get better haha I I, don't think. There's gonna be a major issue issue: okay,.

C

Okay and that we're currently just waiting for him to come back with the is he gonna work on the race gateway agent, bits or.

B

C

Are we gonna help.

B

Him um you're gonna, okay, all.

C

Right, hopefully, some of the new folks we have turning next week Oh week after next. Some of them will somebody will be working on the agent again yep that'll help yeah.

B

That'd be nice, it would be nasty of someone um helping there in that area. There's some pending code that I still need to review, though oh and we need to have some discussion about what the next steps gonna be. But um in testing sonam require some time testing and and then we we need to think whether that's only phase one and phase one is basically you take her back it and say uh this bucket is gonna, be no has at least many shots.

B

What you do it before the bucket had any tear in it um or give to it. You know right, but what to do with buckets that already exists, and you want to show them.

B

There there can be different approaches. One would be like quiesce all the rights to to that bucket and convert it to 20 shouted one which should be I. Guess should work somehow, like probably not the most elegant solution, but much should be pretty easy to wear up, except for creating rights which is not something that we we really do be good but yeah. Let's say.

C

The right always does a like a prepare type operation on the index. Alright, the first step right so that that code would fail and block or whatever, until it.

B

Yeah, we can do that for how long, though yeah.

C

We can put ya how big it got before you the first right where's that it is there's the bucket object, that's in the metadata pool that specifies how it charted all.

B

Right, yeah yeah with the bucket object. We add.

C

A monitor to it, it's the same one that has the ACL on it. That's in the cash on the gateways, either.

B

The one to ten cacl I'm, not on ripped no there we have two different one is a bucket entry point and one is it actual bucket instance, yeah I, say I think that bucket instance, the other one is: has the ASEAN.

C

Okay, but it but it if I'll get cached by the gateways, so that they'll remember with the sharding yeah or how about this? Ok.

B

Yeah, it's cached and another.

C

B

Right pull the ended up a message.

C

That goes around.

B

Every metadate data change there's an not if I message the closer one out.

B

Unless you turned off the case, the.

A

B

But uh nah nowadays we cash it. We have a double double layer, cash, double gate. We actually cash, the packet info, um her and accredit yeah and decoded busting bucket info well to save the encoding decoding that happens used to happen every own operation.

C

B

um So yeah, so that would basically do it. So I don't know if there is any need to do something more come more complicated than.

B

Like detect, whether a bucket has lots of objects inside and whatever you need to do.

C

Other little boy, yeah I, think that would be that'd be the end goal. This would.

B

Be face please three, maybe sorry.

C

Stace to the online Rashard in and then be automatic.

C

Look I'm by the ways to go.

C

C

Well, I, don't think we have a whole lot more to talk about here with that. Well, where we should take a take, a break Patrick, oh yeah,.

A

I think that sounds a good plan.