Red Hat OpenShift Operator Framework SIG | OpenShift Commons, 23 Nov 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Operator Framework SIG November 2018 Full SIG Mtg

Description

Moderator: Sebastian Pahl (Red Hat)
Jaeger Operator Introduction and Update
OpenTracing CNCF Project
Operator Framework SIG November 2018

A

All right so hi I'm Sebastian I'm in charge of the operator framework from the engineering side here at Red, Hat and Diane, is absent today and left me in charge of moderating this. So this is going to be quite freeform. There was an agenda that where some people wanted to present there are their operators. I, don't know if you're all here but I would say. If that's the case, if one of you had an operator to present like, for example, the Jaeger operator, then I be happy to hear about it all right.

A

So first we're good. Let's talk about, let's hear from the llego operator. Let's that's sounds like a fascinating topic. I.

B

Hope that you can see my screen here, all right, so the Jaeger crater, so my name is Sean. Kelly I am a star ng as well, and work on the.

B

They should be started at Cooper to trace their couple thousand microservice, but right now it's country being used and production by Bridger. And so, if you don't know yet, if you gave yourself.

C

Good, so do interrupt me if I get muted, again yeah.

A

I was checking on my side. Go ahead. I think it seems to be going fine. Now, all.

C

Right cool, so um so white wedding is any part, makes its way of the point.

B

Here is actually.

C

Do have templates ready for human IDs for open shapes and we've been telling people just get our template change it to suit your needs, and that's why I mean it should be easy. The thing is, people do not want to take our templates and change it, even though we are quite comfortable with kubernetes and all those primitives that the turbine eighties talks about like deployments and pollen and not before not that used to it and they are looking for an easier way to deploy Jaeger in comparative analysis and so on.

C

So the community came up with a home charge and when we start and the home start actually makes it easier for people to deploy either there, you can just add some configuration switches and I get stalled now, and the thing is so we heard when we started thinking about building an operator. We thought well, that's actually a good way to start. So we just take our chart. We use the helm, add operator kit and you're done with it right, but it's not that easy in our church.

C

It's not really problem with the chart itself or with the hell map operator kid. The problem is the way that air is deployed can be quite complex, and that is because the area itself is used following a microservices architecture, meaning we three or four components that are that can be executed individually right. So we have the agent, you have the collector, we have the query and they all have different audiences. So people running the collector probably want to you know for people with a very high concurrency scenario.

C

They probably want the collector to run on its own point and the query in another part: very, you know that it's just for either she opened the UI and see and look at traces. So if the query crashes, we don't want that to influence the collector.

C

But then there are people who are quite happy to run everything on the same quad right because you know- or you know, with the same binary and that's not something that is very easy to accomplish with the home tracks, it's doable. But it's not that easy and it becomes quite messy to be honest. And then we have different agency strategies, so the agent is the component that runs very close to the application and it receives tracing information from the application and then sends data to a remote collector.

C

Typically, the agent runs on the same machine and the application at local host, but will co-host with something that is not well defined? Relatives right, I mean it could be a denote, the cumulate knows, or it can be the pod. So local host can be two different things and the right answer is like that. What is local host is actually both right. I mean. If we are in a multi-tenant scenario, then we probably want the agents on on on the pod level, so local host.

C

There is a point where, if you are on a single tenant mode- and we don't care that much about security, we can say a few megabytes by the pointer agent as a demo fetch and to complicate things a bit more. We have different storage backends, so we support in memory. We support elasticsearch cassandra, Kafka and probably more to come. So, as you can imagine, the the helm charge can be quite complex and the easiest way to support to deployment scenarios right now.

C

If you have to helm carts- and that's not- you know the ideal scenario, so we then decided to start the operator from scratch. You know using the operator SDK, and that was quite a good decision.

C

You know when we look at the current viewpoint, because we can hide all those complexities into our operator and give people one llamó file with either very semantics. Software people actually can understand in our terms, and it can be very simple. So it's the simple scenario possible, so people just create a CR like this with gkv with the name for operator displays simples and the operation. We just assume some defaults like a first read, it's been used for deployment like this. One is the e memory storage and we use it all in one deployment strategy.

C

So we have one part and everything went from that part and liquid. You know the ingress pages you create routers if you're running, knocking shaved and so on, of course. Well, that's quite abstracted from the user. A more typical scenario would be something like this, so we specify the name. We specify that strategy is productions. We have different pods for different opponents and the storage being used here is electric search that who server can be found on on this address here. Username and password is also defined there.

C

Now, once we got into features parity with the supernal state plates and openshift templates, we could start doing cooler stuff, you know so you know things like interesting.

C

A openshift world proxy is before running in open ships right so Sophia's we did not have in our segments and it's something that we can detect by at runtime, and people can actually disable that by specifying flag here in the fear they can just say well why I do not want a side harder or oh hi, but by default we ship Jaeger with the UI secured in when running in negotiation, and then one other thing that we can do with anger with the arrow operator.

C

Is we allow people to specify a notation, they're saying I want the sidecar to be ingested into my deployments later. As a you know, eager agents of this iceberg and our operator will control the events related to deployments and for the deployments breeze is a notation, we just add a new container to the pod yeah. So that's what we have today so that their current stage in terms of roadmap, our next big task is, you know, moving from zero, zero, seven, two zero one zero.

C

There is quite a change, but I think we will manage and then expand the site for injection. So right now we support only deployment but only deployment, and we want to support also other pod types like spatial sides and then demo sets as well and if you change them or operations like stuff, which is the actual purpose of having operator right. So one of the scenarios we have in mind is seamlessly updating the underlying the other versions. So with there is a newer version batch minor or even major.

C

We we want to give the users the possibility of updating that automatically, because we're quite good at maintaining backwards compatibility. So it should not be consciously at least batch level operation of the problem and linked to that. We also want to do data migration, so we do provide already for Cassandra. For instance, we do provide simple files or CQL files that migrate data from all over the newer version, so we can just apply that by integrators and in the process of developing the operator.

C

We also got some questions, so those are the open questions and we don't have good answers yet. So the first one is, you know: tight versus freeform configuration. So at one side we have a human and operator that once you pass an option to the binary and in between the binary and and the person, we have the operator the Haggar operators. So should the operator just relay the options from from the user to the binary?

C

And if the option is, you know a valid one, then you can crash later and you know, if you, the part, will be in a question of State or should we validate everything in the crater, meaning that the operator knows about all single options from the binary and that kind of flies that we have to have a one-to-one version of the operator and the underlying Jaeger version. So it is something that we are discussing and we don't have a good answer.

C

Yet, if you did discuss about that related to your operator, let us know and tell us what it decided. The second big topic is handling. The penalties have biggest tour, the biggest topic that we have right now and it's in the context of elasticsearch. So when deploying either in for a production strategy, we need a backing storage. You know not a persistent, lacking search, so not new memory, so we need the current, the Folies elasticsearch. So how do we make sure that there is an elastic search going Starks yogurt instance?

C

It's a girl, so a an instance to something to exist there when the operator runs. But if there is, if you are calling in saying they want to use elastic search, what should we do? Should the operator aware of that to the perhaps provision elastic search in the username? How to do that and then, finally, is how to detect where we are running. So, if you running in open ships, there should be a way, an easy way to detect that not wanna open before community I mean in plain kubernetes or gke, and so on.

C

Of course, it would be beneficial for us to know where they are running so right now we we have a command line interface flag. Where people will specify the platform, then we just do progress, Pacific logic following that flag, but if we can detect this by default, we wouldn't require these exercises like it's.

C

Basically, so it is when I add, if you don't have access to those questions, I don't know how much time we have, but if we don't have time to feel free to think me, I'm on either on the irritation room feel free also to open questions, issues on the github repo for the eager operator and I'm on freenode on the chianti channel as well, so which is which oh yeah, but with it.

A

Thank you very much. Anybody have any questions.

A

With a few for the first one to go back to your question that your your own question about platform-specific logic, have you thought about at least we have thought about it a bit, but we don't have a generic way of doing it. Yet have you thought about probing for api's and trying to adapt yourself depending on what API is you find in the cluster? That's.

C

A very nice suggestion, no I have not tried that so she's not pressing right now. Florence I did open it thread on on a mailing list on this topic, but when I, open they're, just in strategy implemented flag the explicit flag and move on, but here's a topic that I am coming back, Twitter does I think it provides better usability to the domain user. Yeah, probably okay, yeah I mean it makes a lot of sense right. So he's called a rock if there's a route bit over to default.

A

Yeah, it's a problem, that's recurring with a lot of other operators and that's why I like a common way of solving these things.

C

That could perhaps exist at the operator level at the SDK level.

C

You know absolutely.

A

And we're looking for good ways of implementing it. So any suggestions, of course welcome for the SDK or as other tools like both are. It would be good.

A

What about your your handling of dependencies, like you mentioned OLM juju. So would your next step be that you would need a very a a Goods elasticsearch operator to depend on like yeah.

C

Awesome so a lot of my problems, I think so there are actually two types of dependencies right so when, when creating those rights, I had other types of dependencies, but I realize the different seeing other ether components is more important as a general topic, so the other type of dependency is when we have batch charts that have to be finished before a part in stock right.

C

So now some quick case, if somebody's using Cassandra as the backing storage, then we want to run a create a schema and batch job that we create all the tables indexes and so on so forth. So this is one type of dependency and we're currently solving that by using the weight all facility that there's a provides I think it could be managed easier. You know, I think it's super bowl's right now for us, but it's the document now back to the Ephrem level. Dependencies yeah so having a producing quality elasticsearch operator would be awesome.

C

One other thing to think about is I. Think I hinted that before so we don't need elasticsearch. When the operator starts, we need a elasticsearch when certain CRS come in and when that happens, then we we need a magic search to be ready to be used right, so not only installed, but it started ready and available and I think that's one thing that is not done yet I think it just starts the provisioning and then handles that control back to us and when we start it's not ready yet right lecture is not ready.

C

Yet so we, when you start in a crash stage, and we stayed there for a couple. Research got.

A

It yeah that's good feedback and I would suggest working with the OLM team and that's maybe there's two specific problems, one of having other quality operators- and that's that's a that's for all of us, the community in general to to be slowly building them so that we can start composing a bit more versus building and then, but for the OLM. One yeah I would definitely suggest reaching out and seeing if your pattern is something that is interesting to others,.

C

Those were mainly my passions and I mean we have a lot of features to implement in you know, for the Jaeger operator, there are quite a lot of scenarios like monitoring catering.

C

But in I mean the electric percolator is something that we can now so we can also help with. So it is so important for us that we, we would eventually, you know, join their community as well, and that our features also prevented their.

A

Great, alright I don't have any more questions, I, don't know. If have it, we have any more topics or questions, but I'll give everybody a few seconds to react to that. One.

A

Nope all right well, we've had a very great conversation. Thank you and presentation about Jaeger I'm personally excited about this, since it's it's a cool project. So thank you very much for your time. Everyone and once Diane, is back she'll help me get this recording onto YouTube thanks. Everyone.