KubeEdge Community, 28 Oct 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: KubeEdge Community Meeting 2020-10-28, Wed

Description

https://github.com/kubeedge

A

All right, hello, everyone welcome to the cube age. Community meeting today is october 28th, it's europe-friendly time this week.

A

So uh so today it looks like we have new uh members joining the meeting, but would you like to give introduction yourself and after that we, after the agenda we can go through the 1.5 uh feature, development updates.

B

Okay, uh kevin: this is uh senthil. You want me to introduce myself now.

C

B

Okay, uh do you want me to do the presentation also now or uh later.

A

Yeah, I think we can first go through introduction to all from all the new members. Then we can the whole presentation.

B

Okay sounds good, kevin, so hello, everybody, so I am central, uh I'm currently working in ericsson, I'm working in ericsson for global ai accelerator here in india, so I work out of the chennai office in india and before joining ericsson, I was working with other companies, so I have worked with huawei in bangalore.

B

I have worked with nokia and I have also worked with I.t service companies like vipro and cognizant.

B

So my areas of specialization or my areas of interest is one thing is kubernetes and another one is edge computing and I do some open source work during my free time.

B

I write some blogs in medium.com and as well as in linkedin, so I am actually very keen on the cube edge project because it is related to its computing. So I am very interested in the cube edge project and I'm very glad to get connected with the community over slack. So that's my introduction. Thank you.

A

Thank you all right. um Let me check if anyone else would like to give.

A

A

All right- and let me ask in chinese again looks we have some member from chinese.

A

No all right, if not maybe we can start the presentation uh something. Would you like to share your screen.

B

Yes, yes, kevin. I am going to share my screen.

A

Okay, I released the screen, so you can share it.

C

B

So, are you able to see my screen.

A

B

B

So, thank you, everybody for joining this presentation, so I would like to talk about you about an open source project called as cube fledged.

B

So this is a project that I started in 2018 and I thought it would be very interesting for the cube edge community to understand what this project is all about, and I also feel that cubefledged will also be beneficial in cuba edge projects. So that's the reason I requested kevin to give me this opportunity to present this project.

B

Okay, uh again a little bit of uh me so, as I said uh earlier currently, I am with uh ericsson, global ai accelerator.

B

As a software engineer, and I specialize in distributed systems edge computing cloud native software development and very recently I have also picked up some interest in chaos, engineering and, of course, kubernetes and google cloud platform, um and I am a certified kubernetes administrator and also a google cloud professional cloud architect, and I maintain this open source project called as cubefleshed, and there are couple of more other contributors as well.

B

uh So that's about me, so I will not spend more time about me, so I will now go directly to the project.

B

Okay, so to give you all an overview about cubefletch, it is actually a kubernetes add-on, uh which I have implemented using kubernetes, crd and controller. So once a kubernetes cluster is running, we just install qfledged as an add-on, so that is how it gets uh installed or deployed.

B

So what is the purpose of cubeflexed, or why do we need q fledged? So let's say I have a kubernetes cluster right and this cluster will pull images from image registry right. So if we want, if we want these images to be made available for immediate use for deploying the pods, so when the pod starts up, cubelet will pull the image from the image registry.

B

So it will take some time for the images to get downloaded from the image registry onto the worker notes.

B

So, in order to reduce this time, uh cubefledged uh provides a feature for caching for caching the images directly on the worker nodes. So using cubefledged we can define what are the images that we want to be cased and we can also define to which worker nodes. uh These images should be cached uh and then cubefledged will pull those images onto the corresponding worker nodes. So the advantage of this is uh application parts. When they are created, they can start almost instantaneously because because the image will be already available in the node.

B

So if, if the image pull policy image, pull policy in the pod is set to.

B

Already present, so if, if it indicates that there is no need to pull the image, when the image is already present, then the pods will start instantaneously.

B

It is an open sourced project under apache 2.2.0 license uh and now coming on to the use cases. So what are the use cases uh in which we can use q flashed?

B

So I have listed down three use cases, so one is real time applications so real time, data processing applications or real time monitoring applications process, time series data. So so these kind of applications will require quick, startup or even quick scaling. So in those cases, if the images are already available in the worker nodes, it will be useful and second use case is serverless functions whenever, whenever the functions are actually packaged as a container images and the functions need to react to events immediately.

B

So if the functions need to react or if the function needs to execute immediately, then also we can use cubefledged and third is, if you are running your applications on the edge.

B

Let's say you have designed a cluster which spreads across edge devices and let's say these edge devices are connected to the image registry in the cloud, and if there is some network connectivity issue in reaching to the image registry, then in those cases also cube fledge will be useful because because then the images in the local worker nodes will be used instead of trying to pull those images from the image registry.

B

So especially in the edge computing environment.

B

There is a need to tolerate network connectivity issues. So from that point of view, cubeflex will be useful.

B

And this project actually supports three different container runtimes docker container d and cribo, so we support all these uh three different container runtimes and we we also support all the three all these three uh platforms uh in linux platform. We support the amd64 architecture arm 32-bit as well as arm 64-bit.

B

So all these three platforms are supported and when it comes to installation methods uh again, we support uh three different installation methods for cubefledged, so either using yaml manifests, so the repo itself contains uh yaml manifests so which can be used to deploy cubefledged and the repo also has a helm chart and it also has a helm operator. So whichever method is convenient for the users, they can use the appropriate installation methods.

B

B

Now now, after the introduction of cubefletched here are the list of features that are currently available in cubefletch.

B

So the first feature is you can actually cache those cache images onto all the nodes in the kubernetes cluster. So that is one feature. So caching is basically we are pulling the images. We are pre-pulling the images uh onto the worker nodes, so it is possible to pull images onto all the worker nodes if that is what the requirement is, but if the requirement is to pull those images only to a subset of nodes, so if you want the images uh to be available only on certain nodes, so that is also possible.

B

So in the image cache manifest, you can specify a node selector. So if you specify a node selector, then queue fledged will pull those images only to the nodes that are resolved by the node selector.

B

So third feature is automatic. Caching for new nodes, so when the cluster is growing and if new nodes are actually added so cube, fledged will automatically detect the addition of this new node and it will automatically cache the image onto this new node. So there is no need to manually intervene for this new node.

B

Fourth feature is refresh, so if you want to refresh the cache okay so refresh the cache in the sense, if you want to actually pull the images again, if you have already, if the images are already available in the worker node and for some reason you want to re-pull the images so that is, that can also be configured, so you can actually configure cubefledged for automatic refresh of the cache.

B

uh So what happens is you will configure the refresh frequency- uh let's say 10 minutes or 15 minutes? So once this timer expires, it will automatically pull the images again onto the cache, and you can also manually trigger the refresh by way of annotating the image cache manifest. So if you apply a certain annotation, you can actually trigger the refresh and fifth feature. Is both public and private registries are supported.

B

So in case of private registries, you can specify the image pull secret.

B

You have to first create an secret of type docker and you put your credentials into that secret and then you reference that secret within the image cache within the image cache manifest and uh sixth feature is. uh You can also purge the image cache uh you can purge in the sense. If you want to delete all the images that are available in the cache you can do so and you can also, even if you want the entire cache itself to be removed, then you can also do that.

B

So so we have all these uh six features uh available in cubeflesh.

B

Okay, so thank you so much so any questions. Any questions. If you want you can ask.

A

Anyone have any question.

D

Hello, could you show us the crd files.

B

D

B

Yes I'll show it thank.

E

B

So this is the repository okay, so here under the deploy, I have the image cache.

B

Okay, I have it here, so I will probably zoom.

C

B

Okay, so this is the api, so there is a kind called as image cache. So currently it is in v1 alpha 1 version and if you look at the spec right, so if you look at the spec inside the cache spec, so this is where you actually specify.

B

What are the images that you want to pull and on to what nodes that you want to pull so this cache spec, you can specify multiple list of images. So here I have one list and then here I have the second list.

B

So in the first list, I I have put these two images, and here I haven't selected any node selector. So here this list doesn't has any node selector, so these two images will be pulled to all the nodes, so whereas in the second list again here I have put two images, uh but here I have crea. I have specified a node selector.

B

So what happens? Is these two images will be pulled only to the nodes that have this label only only the nodes that have this node selector label. So likewise you can have a multiple list of images within this case. Spec and finally, this is where you put your image: pull secret for your private repository. So if it, if you have images mentioned from private registries, you can put the image pull secrets one by one here.

B

So this is how it looks like so once you sub you once you submit this yaml file using cube, ctl apply, then automatically q fledged will read this and it will automatically trigger pulling those images. That's how it works.

D

F

B

Okay, and uh if you have any further questions, also, uh please feel free to go through this repo, so I have added a readme file with all the necessary information.

B

So so you can simply go through this readme file. I have given all these steps also.

A

Thank you guys. I actually have a question, so uh I would like to know because, oh we know that cubelet they have the uh image gc mechanism right to save up the.

B

A

uh When, when it's under pressure, so uh I would like to know how uh cubafletch coordinate with this, I mean what, if the uh cache deleted by uh qled when.

B

B

Yeah yeah correct.

E

B

uh That's a very good question.

F

B

So what happens is.

F

B

C

B

If cubelet image gc uh deletes any image that is press present in the cache.

E

B

What happens is that's why we have a refresh option, so there is a refresh option, so it will automatically during the next refresh cycle. So here you have to specify the refresh frequency. So then, during the next refresh cycle, the image will be updated in the cache.

B

So even if image cubelet image gc deletes it due to disk pressure, you can you can make use of this refresh mechanism to have it again.

A

Okay, thank you and uh another question is more about the.

B

C

A

um We know that uh in some cases the node may be located uh in a very uh far away on the internet from the central cloud right, so the uh the connection between the cloud and edge might may be low quality. So I'd like to know, if there's any uh specific technology used here when fetching images or not.

B

Yeah yeah. That is also a very good question. uh So currently there is no such mechanism, implemented in cubefledged, for instance like a retry mechanism or or compressed compressing mechanism. So we don't have any such feature implemented in q fleshed. But if you think the feature like that will be useful, then definitely I can.

B

I can see to implement it, but currently, no, if, uh while pulling the image, if uh it will automatic, it will automatically do a back off and then try to pull it again and then once the final timer expires, uh then it will be reported as an error. So then it will retry during the next refresh cycle.

G

B

Image was not pulled during the first attempt, then it will again retry during the next refresh cycle, but in apart from that, we don't have anything special to cater to the network connectivity issues while pulling.

A

All right thank.

A

So maybe another option is that we don't download the images from the central registry right.

A

So is there any uh capability to load from somewhere like local storage or somewhere else, so maybe the uh the when the users they can prepare the uh images there.

B

A

They don't need to go through.

B

A

Things over internet.

B

Yeah yeah. That is also a very good question. uh I think, but we can implement those features, but currently we don't have that feature so currently it will.

B

It will try to pull the image only from the image registry, but if we want images from a different source, probably the from a disk that is attached to the worker node, then we can look at implementing it.

A

All right, yeah, yeah, okay,.

E

Anyone have any other questions.

E

E

A

All right, uh let me rephrase the question so the question uh was that uh his understanding is that there is a controller in the cloud creating some jobs to uh and the uh the jobs schedule. Actually, some pauses gradually to the nodes to to interact with the docker demon, to pull the image or is.

B

A

Working in this mechanism or other way.

B

Yes, it is working in this mechanism, so you can. You can actually refer to this section. So what happens is uh there is a job that is created for every node and every image, so this job will actually create a pod and then automatically uh cubelet automatically cubelet will pull the image and then we just stop it there uh we just execute and we just don't execute any command inside the container, so the container will will not start so we will just stop by pulling the images.

B

So that is how it works. We create jobs and then the cubelet will talk to the container runtime, so we support docker, container d and cryo. So all these three runtimes are supported, and that is how we pull those images. Your his understanding is correct.

A

So you mean you're, actually asking cubelet to to interact with the runtime to download the image right.

B

Yes, correct: okay,.

A

Any other questions.

A

A

All right, thanks again and thank you for the great presentation.

B

Okay, thank you so much kevin. Thank you so much everybody uh for your time. uh If you want any uh new features to be implemented in cubeflesh, please feel free to go and raise the issue here then definitely, I will take a look and try to implement it, and second thing is: uh if you feel uh this will be useful in cube edge, uh then again, you can also try to use it uh in cubase.

B

Okay. So thank you. Everybody. Thank you for your time, thanks kevin bye. Thank you.

A

A

Let's move on to the next item: the next one is a gateway proposal by fan fees.

A

G

Still there you could process the process.

A

So let me let me rephrase the english first.

A

The discussion is about the gateway proposal, so we actually uh have some discussion already in previous months, so uh the background was that currently youtubers didn't provide support for uh ingress or another part. So that's that makes the automation uh that makes there's no automation for uh these services, exposing for the external uh request when the part is under.

A

So actually the background is trying to resolve that so initially, initially uh the uh contributors think that we can directly implement ingress, but the uh problem is that ingress is actually a cluster-wide api, so it can not work very well when the nodes are located in different age network.

A

So then they move to the idea using the sdo gateway api and the easter gateway. Api is a little bit different to the ingress api. It's actually not cluster-wide, so so the initial proposal here uh previously was to uh actually just uh implement the steel gateway in qh.

A

A

A

A

A

G

A

A

All right, so uh let me just rephrase in english, um so I just uh uh covered the background before and the uh proposal last. The time is actually uh covering about how to how to introduce edge gateway capability. So so actually this one.

A

Depends on different network environment, so we know that ingress is a cluster by the api. So in some cases the user may still have the network that all the nodes and the masters are able to communicate with each other.

A

So in that case, people may still want to use ingress because that's very uh vanilla functionality and they already have some third-party adoption on that, and the second case is that so educated way is trying to provide actually uh ext, make it possible for the external request to to reach out the sum of the parts located in the in the subnet specific subnet.

A

So in the in the network, in the network environment that agent locates in different subnet, it's very useful, and the third thing is that actually for the age mesh under discussion is actually covering four layers.

A

So the underlying is the uh actually the container networking. uh Currently, if there's no any ci plugin stored, h, edge gateway can provide the part to product communication through host network. So basically the the network of nodes and secondly, uh h made each mesh currently have a equivalent capability of.

A

Equivalent capability of load, internal load, balancing just uh just the same to cluster ip, but it it it's able to work on the uh hosting network. If, if the there's no cni plugin stored, the third layer is node level. Dns resolution and the fourth layer is advanced.

A

The traffic routing traffic management, just like service mesh. So so the discussion is that uh each gateway should each mesh should be able to provide different capability sets.

A

Due to different networking environment, so with this background, as you get away this one this capability, we think it should be decoupled to the other functionalities. So maybe we can just introduce as a new uh crd or new api provided by cubh and in the in the cases that the underlying network is all reachable between nodes.

A

Maybe we can also find a way to integrate with istio. So that's then, the user can get full power of steel.

A

Alright, that's all about this topic, any questions or we can move on to the developer, updates.

E

E

A

E

A

E

E

E

A

A

Okay, so this uh the discussion is uh still about the edge gateway, so just want to uh identify the different case between uh using a service mesh like istio.

A

If we, so, the question is actually that, uh if we support least watch from age for the third party applications, do we still need the age gateway.

A

So actually the lister watch on age is providing support for the applications talk to kubernetes api, but the data plan, the part to product communication across different subnets, still need the uh still need additional functionality support it's not yet covered by this. The watch out age. So in that case, we, if the cross subnets communication not yet supported, we cannot integrate with service mesh because the the end-to-end communication may still be broken right. So in that case, we may still need to consider whether we should use educated way.

A

Another problem is that we know that edge gateway.

A

The idea is to limit the backhand, the endpoint inside the inside the same subnet, that the access point is reachable, but the difficult is that we know that in kubernetes the designs that the part are defined and by deployment and the schedule by scheduler and the service is actually just a kind of uh internal load, balancing relation binding to the other points. So they are.

A

A

Restrict from uh restrict the product location from a service api, so that's the difficulty so in cases users didn't enforce the uh enforce. The restriction to make sure the uh the part, part and location and defining the deployment are the same idea with the service or ingress definition or gateway definition.

A

Then the entrance functionality will be broken so so here we need to balance the trade-off of designing this entrance. Behavior.

A

Okay, so um let's move on to the one to five feature updates. So, according to the calendar, I think we have only let me check so I think we have only two weeks until we make this release right.

A

A

Check all right, so, let's go through the uh feature list efcc to the age and that one is already implemented the uh k. Yet I mean get data any updates on this. We're still waiting for review.

A

A

A

Still under design.

A

Improving a releasing process.

A

I think this one is actually pending waiting for review and also the debian package support.

A

This is also under review.

A

Okay and support list watch any updates on this.

A

That's the time it was said. We need some review comments as.

E

A

A

Okay, uh okay, admin uh reset, but this one is implemented so website upgrade uh this one. Actually uh we almost finished the upgrade, but the uh the actually the styling looks not uh not to improve enough. We may still need some standing improvement, so maybe we will just uh have it on the uh development branch and wait for more improvement. Then we can replace the main set.

A

So here we already support different versions. We just need to update the configuration and the home page. We I think we may need need to add more uh section like introduction to the uh overall architecture.

A

And the recent update is that we have a blog list, page design, but the the detailed styling still need improvement and the other events.

A

Okay, so the community, we have the uh join community page and also the partner page and the uh case study page, but it's it looks very still in the early stage for the styling thing, so functionality functionality-wise uh it's enough, so we we just need more improvement on this.

A

A

Let me add more details.

A

Okay, data pipeline- this one is pending uh motorbikes, mapper, go implementation, so this one we have a pull request. This.

A

A

A

A

A

A

A

Okay, so this one we can get it in and the uh each mesh optimization this one is still under.

A

Review and uh the uh pay admin support uh uninstall cloud core as a deployment. Any updates on this look still under review and the optimize output log all right. So maybe we can just uh review it and get this in maybe edge mesh. We need more discussion linked. This is completed and tests. This is a long term thing.

A

uh So that's the one fix redownload service file, any updates on this or just the waiting for review.

E

You are building a tester agent.

A

Foreign, okay,.

A

um Add the option, past package path and updates on this checking.

A

A

Me: okay! Next one.

A

So you're gonna take a support cloud code running as systemd.

A

Service- oh all, right. This one is pending.

A

What about the next one use device api both on cloud and age and any updates on this.

E

A

F

A

A

A

A

F

A

A

F

A

E

A

E

E

E

A

um uh Configure domain url for cloud core, this one is simply implemented and the upgrade to kubernetes 119.

A

Yes, this one is already, uh I think, already got some review, just a need. A final review and approval.

A

Okay, the original e3 test, this one is still.

A

Under test okay, oh yeah, we have, we already have the result.

A

All right, so uh we have the list. We just need to check. If any uh any issue, we need to fix okay uh gateway to a page, so this one is under.

B

Design, probably.

A

Not getting able to carry 1.5.

A

All right, let's worry about it, other.

A

All right so for these uh features and the works already done, we still need to double check. If we have the uh finished the document work. If not, please update the document, especially the user user guide and user.

A

Menu all right, um that's all about the meetings things today. I think.

A

All right, thank you all for joining the cubeh community meeting, so the next time will be november, the 4th passively friendly time next week. Thank you all again for joining see.

A