Kubernetes SIG Apps, 5 Jun 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Apps 20170605

Description

Meeting minutes and agenda can be found at https://docs.google.com/document/d/1LZLBGW2wRDwAfdBNHJjFfk9CFoyZPcIYGWU7R1PQ3ng/edit

A

Welcome to the June 5th 2017, kubernetes, sig, apps and I'll get started by sharing the agenda we have for this week into chat, so in chat. You just got a link to the agenda that we'll be going over this week. We have some announcements, a demo and then some discussion, I'm, Matt, Farah and so I'll kind of be guiding it through things this week and so I'll jump right into the announcements. The first one we have is kind of a question or we're. Looking for folks who are interested in monocular.

A

For those of you who aren't familiar monocular is a a project that um kind of gives you like an app store for kubernetes. It has the home charts listed and then you can go, find them read through the details and even if you set it up in your own clusters, you can have things like an install button. I just dropped a link, yet thanks for at none for dropping the link to it in it's. Actually what powers cube, apps, calm, and let you go through that.

A

But if you're running your own cluster, you get some more capabilities, we're looking into things such as. What's the roadmap going to look like where it's going to go? Some of those features? Do you know just somebody want to contribute to some of it? Do they have suggestions they'd like to get into the backlog? So if you are interested, please speak up, email, add none and let's get involved.

A

This looks like an interesting project because it actually takes the technical command-line stuff and puts a nice interactive interface on it, which, for a lot of people, is something that's needed and useful.

A

That's the first one and then the second one we have is not. Is anyone able to demo their one-point-seven feature at cig p.m. meeting tomorrow morning, which is 8:30 to 9:30 Pacific Standard Time? They we have been allocated three slots.

B

Espanola, oh yeah, hello! Sorry! Actually this is Dennis I, don't actually I! Think I cut out for a minute, sir I think one of the announcers that I had written there was about doing some demos of the gaps with you tomorrow. So it's anyone that seven feature whether it's alpha or beta. If anyone's ready features, we can demo them raisin.

C

A

Me know: I, don't get you.

B

Added to the doctor over there as well and.

A

So people should reach out to you if they are our.

B

Current Alaska yeah.

A

D

An addition they can tell me now where they don't really know me.

B

E

C

B

Have we've been allocated? That's a gap, the sig p.m. meeting, so it's really a way that we can showcase what we've done for one desam to the other diem reps there just to make sure that other people are aware as well as if we need anything. We can also do that to make sure it's really I'm really. The point is for us to show the rest of the community over what we have building for one day, so there any features that are ready and.

A

B

I think I can add that to this image and tomorrow's the Emmys.

A

Fantastic, thank you so somebody's interested please reach out. Let them know! Thank you, and so the next thing we have on the agenda today is a demo of the app controller flows demo and some QA time. Stan. Are you on yeah.

C

A

C

A

Can hear you the floor is yours. Thank.

C

You very much so let me share my screen.

A

And I can see your screen excellent.

C

Well, hello, my name is Ken Laguna I am from Laurentiis I'd like to do a little presentation of features that recently led the Tanakh controller and which is called flows. I guess most of you are familiar with up control up to some extent. Actually, it's very easy to be familiar with. It seems at ease or should I say to us very simple: oh, it has resource definitions and dependencies between them, so you just define dependency graph out of them and let the controller deploy every sink in the order specified by the graph.

C

Here's how dependency graph might look like, and if you take a look at example, we come with a controller, they all have dependency graph that look more or less like this. The problem is that in real life it looks more like this one. One on this side and once you get to graph exists, becomes extremely hard to manage it to change this or simply understand, what's going on there, here's when folks come into play now. This graph may look like a mess of colorful boxes, but they actually do mean something.

C

The complexity comes from the fact that will be plus systems that made of many applications at the chaplain by itself contains of some kubernetes resources.

C

If we make dependency graph out of resources from each application, we get mess like this, but if we could redefine the graph in terms of applications and write, dedicated graph for each of them becomes much easier to understand and manage those do gestures close yet another graph vertex that you can use in dependency graph to present another graph, not only the tracks individual pieces of graph, but also break dependencies between the sources of different applications since flow graphs are all separate graph becomes possible to do changes to one system component without hitting there's nothing workflows like a black box with complete its resources so that you can change implementation.

C

The fun flow without breaking colors cloak and around by users whose help up controversial I or they can be run from each other, which means that one flow can consume as ourselves and that we can can pull slows into bigger flows. So what can we do with the flows? We can define application from exam tomorrow. We can also define entities. Classification consists on, for example, it can be that's based classroom, not so that we could compose database application from log flows can even help those that are exempt.

C

Lifecycle management operations to people from the application, but nerds during the force must be reusable.

C

Most important piece of functionality in terms folks, your usable building work is ability to have parameters. Parameter value may be used inside versus differential with dollars. Ethics ashram on this slide. So now, instead of having separate graph for issues the applications note, we can have just one permit right. Slope resolve the data.

C

It is not struck that in two parameters, random values may come either from CRI or pass from another, so for casual flow is consumed by another flow.

C

Second principle: piece: functionality, application which parameterization will have a generic flow for class or not. But how do we build system that consists of verbal numbers, not replication, that I said when youth, when user runs slow, he can spit fire with account number of control over automatically. Do this number of reference issues in each name and last name in special parameter? Cope is name, it is variable. If you this, our nurses name, the niche flow replica will get its own water source experience. The other source will be shared.

C

The two more replicas now that that's possible to five six replicas, but also can be relative, so one can say, create me one more replica or delete two last replicas. This makes possible to remove upscaling and volcanic replication is just one flow from don't connect, it's even possible to specific actions that need to be performed before replicas get deleted, for example, can be unregistering out from the cluster.

C

Now I'd like to touch it is the example, which is a good demonstration of what can be done with close, the goal was to create application that is highly available and scalable and operations can be performed by a controller alone without additional operators. As a result, the application can be deployed together down and healed by running some salsa, also with his monitoring job, which periodically queries is because their health and if it falls and helping out at the magically, runs flow.

C

We should cover it now that it's this example of multi node application, which cannot be reduced to justification pod, for, for example, initial configuration requires of each node now, auras or not, and thus into not pasture to the pasture, is different, a different configuration from three node button, so you can adjust, take one speck and deployed several circles of it. Another problem is that there should be a service flourish of a node, so just need to replicate, set of resources.

C

Wrestling just be application, consists of three flows, one of the deployed initial cluster flow, the scale cluster app up and down and flow to recover and helping off. Let me very briefly walk you through them. The first one is TCP bootstrap which deploys initial cluster. On the left of this slide, you can see how the phone looks like and on the right would turn into after application. You see here the observes service for each of the pods and all sorts of the robot Street and Composition uses coop CTL utility to get list of services.

C

This is how we would know all the notes that are going to be in the cluster and then the majority job we started on the after opals are created the cluster. If any numbers not can be created, we just won't seem like a man.

C

The cursor flow, which can add or remove nodes to the cluster, it start with judge that registers limit in distant cluster get new cluster. Can she configuration are pushing to concede map? So when the bird starts, because all the information and all the register could be CD and the last flow is, it is very colorful which will remove some coconut from the cluster and create new node. Instead, their removal is performed by the job.

C

It's also used for down scaling, and then the slope called etcd scale to create new node in places once it was deleted. So here's november how one floor can consume another flow.

C

So I know if we have 10 left, because I wanted to show how it works. As you see it's not just presentation, they suggest you, let us start with Q&A and then I'm going to relight them. Okay, I can now. It was very fast.

C

So no questions and.

A

You won't have questions.

A

All right, winger, I, guess we're going to the live demo.

C

Right, Jen, just in case some questions later they're, going to share this presentation, sirs link to the commutation and also into pre recorded demo, especially for this PCB application.

C

All right, so do you see like this is a text here. If it's big enough.

D

He's big enough to me at least the Exile.

C

All right, so, what's the first thing I'm going to do is upload definitions to to cooper-, have very simple scripts and dusted, so just cause to do good. So now, I'm going to create an issue with this big configuration.

C

And this is how I run slow, so this is could be c1, speed, bootstrap flow name.

C

And I want to have cluster of the Dollar Tree string off, so I specify Electric account, which can be any number.

C

You the graph and we can see like deployment, was very fast, so we should see what was deployed.

C

There are three inch output, there is monitoring choke service for each of them out and one service certification. So, for example, we can see it is the work by just.

A

C

Good CDL, exact.

C

C3 know for running and now I can use another. It flow do what kill it so, instead, the bootstrap assay CB scale and, for example, plus two plus two replicas.

C

Notice, it creates rheticus one by one because and you cannot create a work with someone like wanted to build it one by one.

C

So now it's five bot free initial to first word.

C

Those five foot, five notes now and I- can also, for example, make it for so instead of plus two si minus 1.

C

So root for now running and concedes them by just one feel like a mint. Oh I can delete a neutral book only the notes or delete lots of cabinets yourself, and it will automatically recover get master flow which normally not go to Emily. But it's called from from monitoring joke I.

C

Can probably show you llamo business or close-range language, so you do all right. So.

C

This is, it is the example, so everything we have here is like normal resources, word service, and so on, don't be in your shoes, whatever it is: slow flows, so very simple persons which define basic purposes so.

C

In alpha main space kind flow, then export, it means that it can be run from CLI, otherwise can only be consumed by a masterful construction and destruction, our labels selector for dependency label. So this is how we know which source belongs to which flow by featuring dependencies basis. Labels construction is work on the created when I do something and destruction is what happens when I a remove racket. So, when I delete the petition note, if you get the dependence smart with this label, replica circle good, my namespace for it, because the shared between different flows.

C

Here we have three flows and which sucked off with great initiative Croatia skills. It's go up and down, but again remove not created by Gustav. Show this scale flow so like the sheriff, Lucas and parameter. Just creation of what parameters can be accepted by the flow. So here we have two parameters and both of them can default window, which means they're, not mandatory.

C

So I can we see like and past cluster name but Felicity, but they can pass different name. I was a fearful and if I want to have several clusters in one and space because I don't complete that need to give them different names and loudness problems are always how often money trying to checks for unhealthy notes. If this is bootstrap, so.

C

Scale flow we're similar just different labels, the only one parameter and it's sequential, which means replicas, are deployed one by one click in it at more than one, not at the time.

C

Color flow, so not message magical here and you have the pencils look like, so this is dependence between grass mowed. Your problem, familiar with what was added here, is a big pass arguments between use between nodes in the graph. So, for example, then here we use joke, whispering trust, part and dream, and we pass his parameter from inputs of flow and in order for dependence to match against flow. Now most of dependencies have labels which match what was in the flow definition.

C

So it's really simple: I wish there was some joy forces problem used to have one, but it's not too bit outdated, but has big advantage over operators that just connects resources. You don't need to be any programming just set that set of resources and dependency between them. That's it but whisper motivation and replication. There are now a lot of things that can be done with sub controller, which were possible to do before, and this example just needs really good because configuration we should generate like it's very dynamic.

C

If you try to do it manually, you will need to do a lot of things. If you want to add another note, you need to change existing resources and that's often you should take a look at it to be like the citizen. It just be example on which, based from from Korres guys.

C

It's very very simple, especially on scale you need to have it to be of now. 11, not you definitely just want to just manually.

C

Right so totally there are some questions of the one hip down through.

C

Your belly, it.

A

Noticed so one of the comments from chat is, it looks like you specify the edges twice: a parent nominates, its children and each child nominates its parents is that right or its parent now.

C

The dependency have its parent and child. So if you draw rough dependencies arrow between two two nodes and be generous parent and child, like it's furnished ability specifies independently.

A

Okay, somebody else asked: could you comment on the variables that are passed between the parent and children, for example? Where do they come from.

C

So usually they came from slow argument, so you should take a look at slow, oh here's stuff. So it's has two parameters, cluster name and let us probe into the deadlift here and then, when you define dependences.

C

So, for example,.

C

If you made a special variable, which is replica names provided by reserved controller itself, but other variables they can be, then they come from so inputs are the variables.

A

That are provided by app controller documented somewhere. Yes,.

C

It's only fills them with a synonym guess, one or two more like flow name, so so here how they used see that it's name is part of so flame will not control the process or social. Just puts will read the comment here for each of replicas and because charity giving anything you get, but man for each other. A business cup controller has lists of places where I can replace variable names. For example, it doesn't do this in bar street. That's why I put all my variables into environment.

C

You use environment, variable superscript, those there's only few predefined variables, which are documented and narrow, startles a/c, underscore everything else's comes from flow equipment.

A

Okay, thank you. Does anybody have any other questions.

A

Going once twice thank you for a great demo and walking it through not just the code and showing it to us, but how it works. Thank you.

A

And so the next thing we have on the agenda.

A

Is a discussion on our discussion topic is the snapshot proposal and how end-users could use this feature for their stateful applications? Do we have anybody on to talk about that.

A

Do we have anyone who can leave that you.

A

It sounds like the person who wanted to have the discussion. Topic may not be here today. Does anybody else capable of discussing this particular topic.

D

I can give you some general information about it, I'm going to depth in the implementation. The basic idea is that you would be creating, via the cloud provider, storage snapshots of PD or EBS, that not too persistent volumes. It's actually a really intricate proposal in terms of how we work, you'd also be able to rebind. So if you created a staple set with the persistent volume claim, you'd be able to map the persistent volume claim to a previously taken snapshot and bootstrap your staple application off of basically off of a previously restored snapshot and I.

D

Think the primary use case that it probably will help people with in terms of a data availability plan, is not so much disaster recovery because, with EBS and PD, you know with cloud storage solutions.

D

It's not like back in the day when, primarily you were building stuff on top of a cheap array of commodity storage, the the chance of disk failure is much lower, but it's somebody fat-fingers and drops all your tables if you've been taking incremental snapshots on a regular basis, you could do a point in time restore for your storage application back to to these snapshotted volumes. um It looks pretty cool, but it's not at a state where I've had a chance to play with it at all. So now.

A

That sounds really useful because the ability to oops we screw something up to restore from that quickly. That's.

D

Important it is, but honestly, from my perspective and most of the stateful applications that I've used in production environments, they provide some level of application level snapshot and feature like, for instance, with Cassandra. You can take incremental snapshots and upload those as the hard links to the sort of trim tables up to a cloud storage provider. You can do the same thing with elasticsearch. It provides default, snapshotting features, um actually those will go directly to GC p or both Annie as 3 compliant storage ray so and a lot of times.

D

If you use the application level you're going to be in a better state and if you're just snapshot in the block device. The other thing that I don't really understand how they're going to solve very well is you're, going to have to lift rights and eff sinks off in order to get a consistent block, storage snapshot. So I, don't know how you would do the point. I'm snapshot on the running application unless you lift all the right pressure off of it.

D

A lot of the application level snapshots that are provided as functionality for storage applications come with a bacon solution to allow you to continue to take rights and get an eventually consistent slap shot. So I don't know if this solves that problem.

A

Okay, so we've got some questions. We might want to reschedule this one. Then somebody asked the question is, as you mentioned, Eadie's is that empty directories is that Eadie's or eby's uh EBS.

D

Embedded block storage in PA, persistent disk yeah.

A

Okay, that's what I thought all right, then I'll actually see about getting this one rescheduled for a different time, since we have 27 minutes remaining. Did somebody have something else they wanted to talk? Will consider this open floor time open discussion time? Does anybody have anything else, they'd like to bring up and discuss.

E

Going once just that, hey guys got really just that core OSS was really great thanks to everybody who made it awesome and also to the shards hack night. That was really just fantastic. It's not really the discussion. It's more like a comment. There was nothing else happening thanks.

A

For pointing it out, yes, and then it seems Matt butchered race with you, okay, good and I'm sure he loved the idea of a charts hack night with a couple home relationship yeah somebody's asking. If there's a link to the snapshot proposal, does anybody have that that they can share? We can add it to the agenda. I might be, I can find it in attics in the agenda after the fact that would be great. Thank you all right.

A

Well, if we don't have anything else, then we will give everybody 25 minutes back of their time and we'll reschedule this one thanks. Everyone for coming and have a wonderful rest of your day.

C