Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage Podcasts Lin Sun and Neeraj Poddar on Istio, Wasm, and the Future of Service Mesh

Lin Sun and Neeraj Poddar on Istio, Wasm, and the Future of Service Mesh

In this podcast, Lin Sun, senior technical staff member and master inventor at IBM, and Neeraj Poddar, engineering lead and architect at Aspen Mesh, sat down with InfoQ co-host Daniel Bryant. Topics discussed included: the evolution of service mesh data planes and control planes, the new Istio 1.5 architectures, Istio WebAssembly extension support, and the future of service mesh technology.

Key Takeaways

  • A service mesh in one implementation approach to provide service discovery, traffic management, and cross-cutting communication concerns that engineers see when they adopt (micro)service-based. 
  • The data plane of most modern service mesh implementations run out-of-process as a proxy sidecar. This has evolved from library based implementations, such as Airbnb’s SmartStack or Netflix’s OSS libraries.
  • The recent release of Istio 1.5 saw the deployment packaging of the control plane move from a microservice-based approach to that of a monolithic implementation, named “istiod”.
  • Istio now also supports data plane extensions written in WebAssembly (Wasm). These extensions can modify requests and responses and perform out-of-band actions, such as authentication and authorization.
  • Standardisations like the Service Mesh Interface (SMI) can add a lot of value, but the user requirements, common use cases, and the core abstractions of the underlying technology must be well understood.
  • Multi-cluster and mesh expansion (out-of-cluster) support is continually improving in Istio and many other service mesh implementations.


Hello. Welcome to "The InfoQ Podcast." I'm Daniel Bryant, the News Manager at InfoQ, and Product Architect at Datawire. I recently had the pleasure of sitting down with Lin Sun, Senior Technical Staff Member, and Master Inventor at IBM, and Neeraj Poddar, Engineering Lead and Architect at Aspen Mesh. Recently, the InfoQ team and I published an ultimate guide on the topic of service meshes, which you can find at As a result of this, Neeraj and I had exchanged a few emails discussing the use cases for service mesh, and also exploring what the future holds in this space. When you reference some of the things that he'd learned from Lin and her very interesting work around Istio, the idea was born for us all to get together and discuss this topic in more depth.

Istio is a service mesh that was originally created by Google, IBM, and Lyft, but now has a very large community of people and organizations contributing to it. Both Lin and Neeraj are very active within this space. Although other service meshes do of course exist, I was keen to dive into the recent developments within Istio, particularly the 1.5 release which included several architectural changes, for example, moving to a monolithic control plane, and also, the addition of WebAssembly, or wasm support for creating plugins at the data plane.

Hello, Lin. Hello, Neeraj. Welcome to "The InfoQ Podcast."

Sun: Hi guys, super excited to be here.

Poddar: Hi, Daniel. It's nice to be here.


Bryant: Could you briefly introduce yourself please and share a recent career highlight or background.

Poddar: I'm the Co-founder and Chief Architect at Aspen Mesh. I've been working in the service mesh and Istio space for the past two-and-a-half, three years with Lin and a bunch of other good community folks. As far as the highlights go, I think one of the best parts of co-founding a company is seeing the company grow. I've seen it grow from 3 people to now 25-plus. It's amazing to see how much people can take on hard challenges and come out of the end of it with bright colors. I really enjoy that part. I'm still enjoying it as our team is growing. Working with the Istio community and making folks from our company integrate with that has been really satisfying. That has been a really good highlight for me.

Sun: Hi, my name is Lin Sun. I'm a Senior Technical Staff Member and Master Inventor at IBM. I've been on the Istio project for a long time now, almost three years now. Looking back at the highlight, last year, me and my co-author, Dan Berg, actually produced a book about Istio, "Istio Explained." It took a lot of nights and weekend effort. It's interesting to be able to put some documents together for our user from different types of scenarios, how a user can adopt Istio, how a user can look at the landscape of service mesh. That's a small highlight. We're really committed to this project.

Problem Spaces and Motivations for Service Mesh

Bryant: I know starting a company and writing a book are both amazing achievements, so kudos to you both there. Obviously, a company here, we've been talking about service mesh on and off for many years now, but I still want to admit to many folks though this is new for them. I wouldn't mind to get your take on what are the problem spaces or the motivations for service mesh?

Sun: We were seeing a trend of people moving from monolithic to microservices. As part of that trend, as people go into microservices and cloud native, we've observed that there's a common set of problems among these microservices that many people start trying to solve. For example, how am I going to connect to my microservices? How am I going to do retries? How am I going to observe my microservices? How am I going to secure the communication of my microservices? How am I going to enforce policies on my microservices? We've seen a set of problems, and people start to solve them for each individual language like Netflix OSS, so I would call it like a first generation of service mesh, especially for a particular programming language, then we've noticed the trend, people are not going to always write in one language. They're going to start with multiple different languages. I think that's when the service mesh notion was more formally born around using a sidecar proxy to be able to take care of all that for microservices. Then have a control plane to allow a user to intelligently program their sidecar proxies.

Poddar: What Lin said makes a lot of sense. I think for me, another thing to add on that will be, as companies have been moving or traditional enterprises are trying to compete with innovators in this space, they are realizing that the real competitive advantage is their developers and the ability to add applications at a far greater pace than their competitors who are trying to disrupt this space. In that sense, as people are moving towards microservices, they realize that a lot of the work in microservices from developers gets tangled with the work with operators. That's what a service mesh gives you, in addition to what Lin said. It gives you a way of basically separating those layers where each of those personas can do their own work, and still be successful, and give the best outcome for the company. Operators try to enforce organizational policies. They were able to do that in monoliths through creating load balancers and network devices like BIG-IPs and whatnot. You want the same policies, but you don't want to push that burden on developers. That's what service mesh gives you, that decoupling between the operators and the developers, in my mind.

The Control Plane and Data Plane

Bryant: Another problem that I see folks getting their head around is this notion between a control plane and a data plane. Lin, you already mentioned about the sidecar proxies there. Neeraj, you mentioned some of these topics as well. Would one of you mind, breaking down how you see the control plane and the data plane, please?

Poddar: Whenever you have application data flowing through in your system, you want that application data to actually flow through the minimum thing so that it's performant, but you still want to apply those policies. In the Istio land, wherever your application data is flowing through is what constitutes data plane. This is mostly the gateways and the sidecar proxies, which are collectively making up the mesh. As an operator and a user of the mesh, you have to configure things. You have to configure things in the language of the platform that you're comfortable with. That's where the control plane comes in. Control plane takes in the user provided configuration, whether it's a platform configuration of Kubernetes or Istio specific configuration, and provides an abstraction layer, and converts it to a data plane specific configuration. It gives you the right separation, at the same point it makes the data plane more pluggable. In future we have the flexibility of switching out the proxy if we want. I really like the way Istio was started with this flexibility in mind, and this nice separation fits very well for people who are used to this paradigm in traditional networking.

Sun: I totally agree with that, Neeraj. One addition I would add is from a user perspective, it has been providing tremendous value to our user to have the abstraction layer provided by the control plane. Because one time I actually did an exercise myself, so I deployed a simple Bookinfo sample of Istio. I took a peek of that proxy configuration of one of the microservices, there are thousands of lines of the Envoy configuration. I'm not even doing anything fancy. It's like round robin routing, and maybe a header-based routing. That's it. What Istio really provides is through Istio resources to having a user programming with the abstraction layer for virtual services and destination rule, and be able to interact with control plane. Then have that configuration translated to a more complicated Envoy, or whatever proxy configuration it is. That's real value to our user without needing to understand all the detail of the proxy configuration.

Istio 1.5

Bryant: I love Envoy. I obviously work with Envoy a lot. I don't think anyone would deny the configuration language, the YAML, the JSON, it's designed to be machine written. It's very verbose. I totally get your picture around the control planes. You've already mentioned that both of you are big Istio advocates. I would say other service meshes do exist just for neutrality purposes. Istio has been coming along in leaps and bounds. I keep an eye on the project. I notice over the last year in particular, it looked really nice, 1.5, the release, really good stuff. Say, for folks that have heard of Istio maybe not paying too much attention over the last six months. Could you share some of the highlights in the project, please?

Sun: For Istio 1.5 particularly, I think one of the major highlights is really to make sure a user can easily deploy and operate Istio a lot easily, because prior to 1.5, a user would have to worry about multiple components of control plane. That means they would have to worry about checking out the logs for multiple control plane components, make sure each of them are upgraded correctly, make sure each of them have the right scaling policy, make sure they are healthy and up running. With 1.5, there's a big change in our control plane components that we are actually merging all these control plane components back into one single control plane components called Istio daemon. That really simplifies operators and administrator to be able to install, and deploy, and upgrade to be able to manage that one single thing. We're also seeing on the multi-cluster land, which is where I spend primary of my time in the project now, that the remote cluster, it's actually much easier to config to talk to the primary cluster per se, the other cluster, when you have multiple clusters with Istio, because of that one single Istio daemon. It's tremendously simplified the configuration for that.

Poddar: An interesting point here is in Istio, and obviously Lin knows this, is we try to come up with themes for every release or themes for the year. The theme for the past few releases has been operational simplicity, and improved user experience. What Lin said totally falls in that camp. I just want to give you an anecdotal story around, if someone is hearing this for the first time, they might be thinking it's counterintuitive. People want us to make microservices and then use Istio to simplify the problem. Then why are you going from microservices to monolith. I totally get that. That's why I want to put some statements out there, which makes sense to people who are trying to deal with this dichotomy in a way.

I think, in my opinion, what's happening here is we went in the route of microservices assuming there are lots of different personas, which will be managing different pieces of Istio. That's why the microservices paradigm made sense. You can have Citadel, which is issuing certs managed by security ops, and Pilot managed by somebody else, and Gateway managed by somebody else. We found in reality is that most of the time there is one operator who is doing all the work and adding burden on them to do upgrades and look at logs, like Lin said, for 20 different things doesn't make sense. Even though from the developer's point of view, we still have that logical separation in the code. From development in the community point of view, from the working groups, we are still microservices. In the deployment model, we are monolith.

Reacting is a Positive Sign of a Healthy Community

Bryant: I saw a very nice blog by Christian Posta talking about this stuff, actually. I think one of the things we can all agree as engineers is it's better to look at something and go, if we can change that. Do it now rather than punt it a few years, and then wish you'd done that. I'm not saying it's a mistake, but as in admitting as things change, reacting is actually a very positive sign of a healthy community I think.

Sun: That's very true. Yes.

Poddar: Yes. I think we both just said the same thing. It's just so nice to recognize that what your users are suffering for and what their needs are, and then going back and looking at some of the additions we have made. That's definitely a sign of growth and maturity.

WebAssembly Support in Istio 1.5

Bryant: It's a bit of a hot topic, a bit of a buzzword, but another thing that caught my eye in 1.5 is the WebAssembly support, the wasm support. I wouldn't mind to get a few thoughts on what you both think of that. The potential, maybe the drawbacks with this approach with wasm as well.

Sun: I think Wasm Assembly support is really interesting. For Mixer, you guys all know when Mixer was a single component. It was on the path of data plane. Neeraj talked about data plane as part of the customer traffic, as a user travels through their microservices. Sometimes they have to go to the Mixer component to check if the car is even allowed. That has been an issue for many of our users. We actually don't like that actual hop to Mixer. What if Mixer goes down? What if Mixer is deployed in a different region, or different zone? There was a lot of concern on that. I'm very excited to see the WebAssembly support in the community. I hope to see more documentation on this. Honestly, I haven't got my hands on a lot on this myself other than the sample, the solo team put out. A really simple sample to replace the headers. I see it as a tremendous value. I really think of it as providing Docker images but proxy images to allow extensions into that proxy, which in our case is for the Envoy proxy. I do expect to see a lot of more movement of moving Mixer plugins into the WebAssembly runtimes and be able to distribute that runtime maybe into a particular hub, like WebAssembly hub. Then other people can easily grab it and consume it in their proxy. I think that's a really tremendous value for the community.

Poddar: That's the immediate tactical advantage, which is super important for the users. A little bit on my background. I have worked in proxies and load balancers for the past three, four companies. Data plane programmability is a big thing for customers. Aspen Mesh is part of F5 Networks. I was part of a startup called LineRate Systems, which was acquired by F5 Networks. The reason we were acquired was we had Node.JS programmability. BIG-IP is sticky for everyone because a lot of our customers come back and say, "We had Black Friday sale, and I was able to write an iRules LX and save my site." I think that's the power that WebAssembly gives you. It enables users to dynamically program with a much enhanced user experience. I'm really excited for where this lands. I think this can be a game changer, not just for Istio, but the entire proxy landscape.


Nginx has data plane programmability through filters for quite a bit now. Envoy had C++ filters, but they were really hard. Wasm on the other hand gives you a much easier user experience and developer experience. As a company and as a vendor, we have dual advantages. As a vendor, now I can make extensions for our customers without making a lot of custom changes within Envoy. Similarly, our customers can make some specific changes for what they need, which we might not be able to provide. It's really nice in terms of performance and security, because the security paradigm here is really nice. In terms of VM isolation, the risk is low. You can't crash your Envoy even if you have a semi-faulty filter. I wouldn't say it is totally foolproof just yet. From the performance, it's very native to the machine code. Overall, I feel this is a big thing. I see a lot of good trends emerging from here in the coming months and years.

Sun: The language is a huge thing. A lot of people are having trouble with EnvoyFilter today because it's only in C++, or Lua. What if I don't know those two languages? I'm stuck. I don't want to pick up a new language just to write my own filter.

Poddar: Exactly. Look at Nginx, for example OpenResty. There are so many companies which are just born out of OpenResty. As a co-founder, I just see opportunities here.

Discouraging Engineers from Putting Too Much Into a Plugin

Bryant: I totally agree with everything said here, but it's my job to play devil's advocate today. My background is very much in the Java space, and I did a lot of work on Netflix's Zuul back in the day. The beauty of Zuul as an API gateway, you could dynamically inject groovy scripts at runtime to change functionality. It definitely, as you mentioned, Neeraj, you could do denial of service rejections by dropping in a script that would reject certain requests, and so forth. The flip side was, I made quite a bit of money as a consultant, fixing lots of these gateways where people have put business logic in the gateway, where they'd highly coupled their gateway and the services. I'm a little bit concerned, we might see the same thing with wasm, and Istio, and Envoy. I don't know if either of you have got any thoughts on how we might discourage engineers from putting too much into a plugin?

Poddar: That separation is not very obvious and clear, when you just start out. I think the community and all the developers need a bit of maturity to actually understand what should be in the application code and what should be as a dynamic plugin, or a semi hack that you put in an infrastructure code. You want your infrastructure to be stable. These should be well thought out plugins. This shouldn't be, "I woke up at 12:00 in the night and I wrote something so that I don't have to change my code." If I put on my operation's hat, or a security ops person's, I think this is as critical to any other infrastructure. You can't just drop plugins. Whether we will get there, we will. It will take some time, I think. I'm totally with you that there is a concern. I think people like Lin and I, we should start educating now and get ahead of the curve a little bit.

Sun: In fact, we're having similar challenges in the community here. I believe, Neeraj, we had a discussion maybe a week or two ago, some of these Envoy filters in the community, should we actually standardize and provide proper API'ing of our control plane, so a user doesn't have to build their customized EnvoyFilter.

Poddar: Exactly. It's hard to differentiate. When there are easy bailout things that you can do, or easy cop-out things, that's what users tend to choose. We just need to provide some guidance and guardrails, and then, hopefully the best thing emerges.

Bryant: I should say, I've definitely made the mistake as well. I was sounding a bit judgmental. As you all know, I've definitely built my own filters, which are not a good idea as well. We've all done it.

Poddar: If you look at iRules repository for BIG-IP, you will see some crazy things. They're scary.

Tips On How People Might Learn, Share, and Build Up Knowledge

Bryant: The thing is, often if you can do it, someone will do it. I've been there. I've done it early on in my career. It's all about knowledge. That's an interesting question in general. I'm guessing a big part of what you all do in the community with Istio is sharing this knowledge. Have you got any tips on how folks might learn or how they might contribute perhaps to sharing this and building up this body of knowledge?

Sun: I do find out, we're making this easier for our users. Istio, at least the old perception from a lot of people was tremendous pieces for people to learn. What we have been doing as work in progress in the community is make sure our documents are easier to consume, make sure people can adopt it based on their need, and they don't have to learn everything. We've been trying to educate people that you don't have to learn an Istio-specific API. If you just want to get certain features like telemetry, observability of your microservices, you only need to learn minimum Istio API, if you just want to secure your microservices. We had a lot of users having a hard time with Istio API, mainly because of the networking API, which is actually very mature now. There were people having a lot of hard time to understand gateway API, virtual services, and destination rule, and ServiceEntry. Also, I think, most importantly, how to have these APIs work together. How do I combine these resources to work together to do what they wanted? That was a little bit more a learning curve for people.

In my book, I try to educate people, when you don't have multiple versions, you don't really need to look at the network API. We provide default retries for you. We have a little bit of default configuration to plugging along. I would say, stop based on your user cases, and focus, go there. Don't try to learn everything of Istio all together, because it could overwhelm you because of the rich feature the community provides.

Improving Istio, and also Taking Advantage of It

Poddar: If you try to do everything in Istio, it will take a very long time. The other side of this learning, I would like to say is there is one about users who are trying to use Istio for their advantage. Then there are people who are trying to invest their time and improve Istio by being a developer or being a community member. I think we need to do better on both fronts. Our community should be more receptive of diverse people coming in and actually making contributions. By that, I mean we need to be making sure our working groups are more open, the meetings are more open. We have to make sure that it is in a time where it is appropriate for people from Asia and other regions, including UK and Europe to join. We are doing that. I think we are learning. It's really good to see more and more contributions, both as a developer and then also users coming back and saying, "I wanted to do this, but I actually couldn't do it." Then ask questions and go to GitHub and fix it. I think we're seeing the entire spectrum as the community is maturing more. I think, both the user experience and the developer experience will be better.

How Istio Will Interoperate With the Service Mesh Interface

Bryant: I wanted to dive into a couple of deeper topics now. The first one is around interoperability. I've found some really interesting conversations of late around the service mesh interface, the SMI spec. I'd love to get both your thoughts on where you think the value is of that. How Istio is going to interoperate with SMI, for example.

Poddar: Here's my thinking in general about APIs and abstractions. Any developer when given a chance will try to create another abstraction, thinking it simplifies life. In some cases, it makes sense. In some cases, it does not. For the SMI, I think the benefit here is the operator can change the service meshes underneath. At the face value, it looks like there is a reasonable advantage of it. I have not seen much of customers in my day-to-day interactions, who are looking to do that on every year basis. Like, "I'll go to SMI, today I will use Istio. Tomorrow, I will use Linkerd. Day after tomorrow I'll use Consul." I get the broader principles and why you would want to go towards SMI and maybe create an abstract layer. We have to see if one of these technologies is already emerging as the de facto standard. This is akin to Kubernetes versus Mesos versus Docker, and you know who won at the end. I'm not trying to project anything. I'm not trying to say they're doing something which is not awesome. It's just that I think as a user, you should be making sure that you choose it for the right reasons. Are you really in need for changing your underneath infrastructure so often? That's my take.

Sun: I think I'm in a similar boat as Neeraj. I think it is a little bit too early in the evolution of service mesh solutions and the user adoption today, to conclude that we actually need a standard API specification to help our community. Think about how service mesh landscape today we have so many service mesh vendors, and each has their own implementation. Some have more features than others. If you end up using SMI, when you're troubleshooting problems, you still have to go into the underlying vendors to troubleshoot the problem. At the end of the day, you may end up needing to learn every single vendor that you use underneath of your SMI. That's my struggle with it right now.

Could SMI Provide a Consistent Workflow Across The Board?

Bryant: I think it's an interesting point about standardization and the lifecycle of a product. Definitely, I did a lot of stuff in Java and JCP. We were only standardizing the boring stuff, the stuff that we all knew, the stuff that we were building on top. I also empathize with the SMI folks who I understand their goals. I think one other thing I've heard around SMI is a bit like using HashiCorp's Terraform. The underlying cloud infrastructure is all different, but you use a consistent workflow with Terraform.

I wonder, do you think perhaps maybe the SMI could be in that space in that it provides a consistent way of working? I can imagine and say pick a number, 70%, 80% of users in service mesh land, really just want to do routing, observability, mTLS, the basic stuff. Do you think the SMI could provide that consistent workflow across maybe your estate has got a bit of Consul here, Linkerd here, Kuma here, Istio here? Any big organization is going to have a few things in the mix. Do you think SMI could provide value there?

Poddar: I think once we know what those 70%, 80% use cases are, maybe. I think as those use cases emerge, and as the space consolidates, like you said, the boring stuff can be standardized. Or, is one API already the standard, because getting users to debug two APIs is definitely worse than one.

Sun: Really, goes back to your point, how boring it can be? If it's consistently going to work for every single platform out there, there may be a value out there. On the other hand, we also have to look at how is Kubernetes involved in this space? In the last KubeCon in the U.S., I believe people are already talking about the next version of the Ingress API. I'm sure you guys probably have seen it. It also has Route API, it has Virtual Gateway API. It's going to build the standard of the Traffic API if that's getting adopted by Kubernetes. If Kubernetes, everybody already agrees is the de facto standard for container orchestration system, so that may be the API.

Poddar: I totally agree. On top of that, I'm still not convinced if this is the right API to abstract. Most of the enterprise customers that I talk to, they struggle with, I want my application to work like this. Don't tell me networking APIs. That's my other struggle with service mesh spec. It's still trying to abstract networking in more networking terms. It's not talking to the language that people want to use, which is, tell me my application behavior, and then do whatever you have to do.

Bryant: I think it's probably a whole podcast on that, and developer experience, and a bunch of things there. A bunch of other topics I'd love to cover as well. Just picking another one, which I think, again, folks are asking a lot about at InfoQ, is this notion of multi-cluster support. I'll put into that topic as well. It might be an abuse of powers here. The Linkerd folks referred to as mesh expansion too. You got multiple Kubernetes clusters, but you might then have some VMs, or some old mainframe stuff running out of band on your COBOL, or whatever. We're hearing a lot about COBOL at the moment. How do you see the Istio and the service meshes dealing with that thing, where there's not just one Kubernetes cluster? There may be many, and there'd be other things in the mix too.

How Istio and Service Meshes Deal with Multi-cluster Support

Sun: We already have a pretty rich support, I would say, around multi-clusters. We support Inglewood official documentation around how you can set up homogeneous multi-clusters of Istio. How you can have Istio maybe running in one cluster for the xDS serving, but on the other cluster you also have a lightweight Istio running but not doing xDS serving, just to manage the certificate and sidecar injection. In the future, we're actually going to involve a little bit more in this space. I'm currently working on developing a little bit more configuration to have a central Istiod notion that people would just maybe have the remote cluster fall back totally to the primary clusters in Istiod. That's coming.

Multi-cluster Pattern In

We also have replicated control plane, a multi-cluster pattern in, where people can config heterogeneous clusters where the cluster doesn't have to be the same as far as the services deployed within the cluster. They do have a shared single Root of Trust, but the services don't have to be the same. You can config locality load balancer among these clusters, so to be able to fully leverage different functions provided by the mesh. We also have VM expansion support where you could potentially bring workload VMs to participate into the mesh. You can have part of your services are running VM and part of your services are running in Kube, and you could potentially load balance the amount that you use. That's also a hot space involving We already have some guidance for our users today on those.

Poddar: We just added a new API called WorkloadEntry, which gives you a uniform way to use the Kubernetes primitives and say, I want to add these VMs. I have migrated two of these VMs as pods. Now I want to load balance across VMs and pods, and VMs of the same services uniformly. That's a really interesting API. I totally agree with all the options that Lin said. I think broadly, the use cases are trying to do locality aware load balancing, or trying to do a failover. I think we are there. We can do a much better job of making sure our documentation and the user experience is much better compared to where it is now.

Sun: One thing I would add is troubleshooting. Sometimes if you ever run through a multi-cluster scenario, you will notice that if it fails, it's a little bit harder to troubleshoot today. That's an area we definitely want to look into, to provide more guidance and automation to our user.

What the Future Holds For Istio and Service Mesh

Bryant: These distributed systems. Yes, who knew they're really hard, aren't they? I hear you. That's great stuff. That leads nicely into wrapping up now. What does the future hold for Istio and the service mesh space?

Poddar: This is something that I've been thinking about for quite a bit, not just as a community member, but also for Aspen Mesh, since our futures are quite tangled. I think there are some tactical future enhancements that I think are going to happen and emerge. Then there are more strategic ones. One tactical thing I would say, multi-cluster support, and multi-cluster being a reality is pretty obvious to us that that's going to happen. I see more and more users coming and saying, we have these many clusters, they are multi-cluster. Tell me how I can do this versus that, whereas, last year, only Lin and I were talking about it. It's more real.

Second thing, which is more real, is I think the scale. I constantly see people coming from big companies like Alibaba and few others who are saying we have hundreds or even thousands of nodes on which Istio is deployed. Tell me how I can actually make Pilot work there. Tell me how I can make the control plane scale. That's the thing that, again, as an abstract exercise we have done, but now it's real. That's the immediate things that are going to change towards.

The more strategic piece for the future for me is mostly around three things, one, which you've already touched about WebAssembly. I think WebAssembly will emerge and change a lot of data plane programmability, and might emerge as a proxy standard. I really look forward to what customers and developers alike come up with the crazy things you can do, keeping in mind that that might be troublesome or create a new market for consultants. Who knows?

Second thing, which is very interesting, which I've seen, is the application of service mesh in different business verticals. This is around service mesh Kubernetes. They started as a need for enterprises. What we are seeing is the same needs are relevant for telcos and 5G providers. 5G vendors are going to take this space. They're going to make this a more interesting problem and more interesting solution, similar to how the telco vendors modified how load balancers used to be talked about. I think this is very interesting for me.

Similarly, I have seen some emergence in IoT sectors where IoT folks are saying we need to create an IoT platform. That platform does not look very different from an enterprise platform, just that we are talking about different protocols. I think the long-term vision here should be, how can we make service mesh, or particularly Istio apply to other verticals, to other protocols without trying to reinvent the wheel? I see a fair amount of momentum there. I think, as it advances, I think we will go in directions that we haven't thought about. Two directions, which are really interesting is how can we make this work in constrained environments? Service mesh right now is deployed as sidecar proxies, which take up quite a few resources. Think about if you have to run a service mesh in a far edge 5G tower. I'm really looking forward to that evolution. Similarly, performance, if you're trying to get this work for telcos. You have to be really high performant. Iptables might not work for you. You might have to go to some really advanced technologies. That's exciting. Really, I think that will push us to new innovation.

Sun: I totally agree with what you said around multi-clusters, performance, and also WebAssembly provide new ways for a user to be able to extend Istio. From our perspective at IBM, we spend a lot of time multi-cluster. Make sure different interesting patterns around multi-cluster can work seamlessly for our user. We want to promote the central Istiod notion where Istiod doesn't have to be running on the remote plane. That's something I'm driving in the community. We're also working on the community multi-revision. I think that's super interesting. You guys probably know Red Hat OpenShift already support multi-tenancy. I'm super excited to see the environment workgroup did a lot of work on multi-revision support in the community that allow a user to be able to run multiple versions, or even same version of Istiod on the same cluster, to allow a user to seamlessly upgrade between versions and also for multi-tenancy purposes. That's a tremendously useful feature to our users to be able to do a lot more testing before they move to the new version, and also, isolate their projects by different meshes within the single cluster. I feel with innovation in single cluster space, we're also innovating in multi-cluster. There's also a lot of talk on mesh expansion and federated mesh. How are we going to connect two meshes together? Those are super interesting areas. I'm expecting the community to continue to be involved, and hopefully, play a little bit ahead in the service mesh space.


Bryant: That's a lovely insight into interesting roadmap items, I think folks can look forward to and perhaps get involved with as well. That's splendid. The final couple of things, though, if folks want to follow you online, what's the best way to do that, Twitter, LinkedIn?

Poddar: Both for me. It's nrjpoddar.

Sun: They can reach me on twitter @linsun_unc. I'm reasonably active on Twitter. I got a notification when Neeraj liked something.

Bryant: There you go. Thanks both of you for your time there. I really enjoyed chatting to you.

Sun: Thanks so much for having us.

Poddar: Thank you. It was really great to be here.

More about our podcasts

You can keep up-to-date with the podcasts via our RSS Feed, and they are available via SoundCloud, Apple Podcasts, Spotify, Overcast and the Google Podcast. From this page you also have access to our recorded show notes. They all have clickable links that will take you directly to that part of the audio.

Previous podcasts

Rate this Article