Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

DevOps and Docker Talk: Cloud Native Interviews and Tooling

DevOps and Docker Talk: Cloud Native Interviews and Tooling artwork

A single API for multicloud with Control Plane

Sep 18, 202654 minAudio
  • Apple Podcasts
ShareXLinkedInFacebookWhatsAppEmail

Control Plane CEO Doron Grinstein joins me to discuss how they’ve created an AI-native cloud API that virtualizes multi-cloud and hybrid cloud to simplify management and ship faster. Watch the video of this episode. 🖥️ Watch demos from the live stream. 😇 My new GitHub Security workshop has launched! A free 2-hour workshop with hands-on labs to harden your repos and your workflows from common supply chain attacks. I'll cover how attackers are getting in, and then we'll lock down a sample repo so you know what needs to be done to protect your code. You'll leave with a deep understanding of risks and mitigations as well as a list of helpful tools to keep your repos safe, including my new "gasa" tool for scanning your repos and orgs. ★Show Links★ https://controlplane.com https://docs.controlplane.com/introduction get a free month to experiment with. email doron@controlplane.com Creators & Guests Cristi Cotovan - Editor Bret Fisher - Host Beth Fisher - Producer Doron Grinstein - Guest You can also support this podcast by subscribing to my YouTube channel and my monthly newsletter at bret.news! Grab the best coupons for my GitHub, Agents, Docker, and Kubernetes courses. Join my cloud native DevOps community on Discord. Grab some merch at Bret's Loot Box Homepage bretfisher.com (00:00) - A single API for multicloud with Control Plane (04:42) - Introduction: Meet Doron Greenstein (13:45) - What Is Control Plane? (17:40) - The Union of All Clouds (19:34) - Adoption and Onboarding (23:08) - Eliminating Infrastructure Toil (28:10) - AI Agents in DevOps (33:28) - Infrastructure as Code in the Agent Era (37:13) - Small Teams and Non-Developers (41:44) - Compliance and the Data Problem (44:59) - Scaling, Secrets, and Audit Trails (48:57) - Cost Efficiency and Getting Started ★ Support this podcast ★

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

Bret (2): were you sorta seeing where this all could take us in terms of agents helping us in infrastructure Doron: we wanted to reduce the cognitive load on engineers not having to stitch together Kubernetes, Istio, Grafana, Prometheus, Vault, and all the CNCF, solutions out there. And we created an API that gives you unified access to any cloud, any service, and we find that AI, if you just tell it, "Create me the infrastructure," it's gonna duplicate a lot of resources: and it's gonna be expensive, non-deterministic, and

it's gonna be non-secure, non-compliant. And where is the logging, the metrics, the tracing, the secrets management, the service discovery, the TLS, the DNS? And the list goes on and on. So what AI needs, just like an engineer needs, is a reduced cognitive load. Give it a deterministic API that says, "Hey, take my desired state and make it so. And by the way, make it so on any cloud." It makes AI a lot more effective. we find AI in a perfect, intersection because we can now make it do

tricks that is otherwise very, very expensive or even u- unattainable Bret (2): Welcome to another episode of DevOps & Docker Talk. I'm your host, Bret, and in this one, we go deep into the universal control plane for the cloud and the search for the singular management system, which we have been doing my entire 30-year career, been trying to find that perfect sweet spot of a tool that can manage all of my infrastructure from one API.

And what better person to have on to talk about this than the CEO and co-founder of Control Plane, the company building this unified API layer for controlling all your infrastructure. I'm having Doron Grinstein on the show, and we go deep into what they're building, And he has a very similar story to me, just trying to use the infrastructure that we had at the time and the tooling we had at the time, and I have managed infrastructure for so long that I have forgotten about all the tools that we've

tried to use over the years to unify everything, and we never quite get there. We get 60%, maybe if you're really lucky, 80%, but then there's always that other thing that doesn't quite fit. And it's exciting to see a company that's brave enough to take this on, because it's no small task. It's no small achievement to be able to say you have a universal API control plane that you can access from the browser or from Terraform or from Pulumi or from your agent harness, and it feels like that's what they're achieving.

And we walked through a bunch of demos. In fact, if you're interested in the demos, go check out the link in the show notes to go to the live stream, 'cause we demoed for about an hour after the conversation, and, we didn't put those into this podcast. So if you wanna check out all those demos, go check the link. But we talk about everything from how they onboard customers, what are their customers looking like in terms of their infrastructure, and how are they managing it once they get there. Surprise, not a lot's changed. They still use Terraform or OpenTofu. All the things that they were using before are compatible with their APIs.

But they also have MCP and CLI tools to manage this, and once you get into their interface, you realize that part of what they can do, because they're agnostic and they want to operationalize all of this across multi clouds, is they're setting up the Kubernetes distributions across the clouds for you. They're making the decisions around how the encrypted tunnels are gonna happen between the different providers, it feels kind of like what Docker felt at first, where they had this old saying at Docker of, "Batteries included, but swappable" and that is something that I feel like I could call control plane.

It feels like it's very easy out of the box, but they hide a little bit of that complexity with their opinionated defaults, which I like opinionated defaults, especially when I agree with those opinions. And we dive into some of their extra features that they're building on top of this universal control plane, like a, a sandboxing approach that is so popular right now that we have agent harnesses needing sandboxes, and we're now talking about the corporate vibe coders, which I call app creators, the people that are outside of IT that maybe need a safe space to experiment with

ideas that they can then hand over to either robots or humans to maybe develop further and get onto their production. Those people are coming, and us platform engineers and DevOps engineers, we're having to deal with that problem, and it feels like we need some solutions that help automate that middle layer out of the way with a higher level of abstraction, and it looks like that's what they built. So let's get into my conversation with Doron about Control Plane. welcome to the show. I'm glad to have you here. Doron: g-great to be here. looking forward to it Bret (2): Give us a little bit of your background, Tell me the origin story,

Doron: I ran a company, sold a company, was chief architect at Dell, chief architect at, SAP, and then ended up as, chief architect, for VMware for cloud services. And, at VMware, I missed my interactions back at SAP with a guy named Dan Wilson, who was the best DevOps/platform engineer I ever met. And what I wanted at VMware was, Dan Wilson as a service. And, I really saw hundreds of, outside customers and internal teams

struggle with Kubernetes, Istio, Grafana, Prometheus, and having to reinvent the wheel again and again. And so had the idea for Control Plane, left VMware, started this company, called Dan Wilson. He was back at SAP and I said, "I want you to join me as my CTO and really deliver to the world, what you do as a service." And I thought it's gonna be a hard battle, but he jumped ship and he said, "Yeah, I'll join you." And, that was the origin story. And then we brought on a few other brilliant, engineers that have decades

of, practice on, on, cloud native architectures, and, we built this company Bret (2): this is pre-AI, right? So, th- we're gonna get into a whole bunch of details later on, agent-driven infrastructure, like MCP and CLI tooling, agent h- harness first, whatever we wanna call this thing, th- this new future we're in. but what did, what was it like to to be an infrastructure company and to see AI coming? Were, were you sorta like seeing where this all could take us in terms of agents helping us in infrastructure and stuff like that?

Doron: 100%. we, I think, got more lucky. We have more luck than brain, because it found us at a very interesting time. We, we wanted to reduce the cognitive load on engineers not having to stitch together Kubernetes, Istio, Grafana, Prometheus, Vault, and the things… all the CNCF, solutions out there. And we created an API that gives you unified access to any cloud, any service, and we find that AI, if you just tell it, "Create me the infrastructure," it's

gonna duplicate a lot of resources: a NAT gateway, a load balancer, brain, a Kubernetes set of nodes, and it's gonna be expensive, non-deterministic, and it's gonna be non-secure, non-compliant. And where is the logging, the metrics, the tracing, the secrets management, the service discovery, the TLS, the DNS? And the list goes on and on. So what AI needs, just like an engineer needs, is a reduced cognitive load. In the era of AI is less tokens, less context window.

Give it a deterministic API that says, "Hey, take my desired state and make it so. And by the way, make it so on any cloud." It is a boon for AI. It makes AI a lot more effective. we find AI in a perfect, intersection because we can now make it do tricks that is otherwise very, very expensive or even u- unattainable Bret (2): Yeah, I always struggled with, automation, pre-AI, in terms of, there's a point at which you're trying to automate things and automate triggers and intelligence with a deterministic language.

And I always would get to a point where I want this to be more automated, but the level of effort to automate that small thing was way more than it was worth in terms of the automation. And, and in DevOps forever, we've been talking about the toil, and I think that, I think the first time I heard that was from Google's SRE paper, was talking about, the toil of ops, and I always felt like there was a better way, a more complete way to automate things, but every time I would try to take it to the next level, it, it was like the management had no tolerance for me spending all of my time trying to automate these little things.

I'm like, "Yeah, but they're tiny paper cuts and they add up over time," and I was trying to, defend the, the case for more automation, but we would just never quite get there. And now I feel like the world is our oyster. I feel like we're, we're at a place where you're only limited by your imagination and your ability to, manage the harness and manage that agent inside it. Doron: we, we called it the cloud is your oyster. We truly allow you to have a singular API and leverage BigQuery on Google and Cosmos DB on Azure and S3 on Amazon and-- or any one of the services.

You can now write code and run it on your Raspberry Pi or Dell machine or your EC2 instances or EKS, AKS, GKE, whatever. And you can, not be confined by any factor. So think of all the clouds as They have merged, and now you can run the compute in a basement or on a cloud or on a secondary or neo cloud, but leverage any service of any cloud, and, and let AI do the work. Like, I need a Postgres, I need a Redis, I need a Kafka.

You can conversationally get those things up and running deterministically Bret (2): All right, so if we back up a second, If someone hasn't heard of Control Plane, how is this different than a cloud or different than a cloud management tool that we might all be using, like Terraform or, any of the dozen, different opportunities I have to manage my cloud with a CLI or some sort of service? Doron: Sure. So Terraform allows you to create resources. We provide the Terraform provider for control plane. We provide Pulumi, we provide other interfaces Control Plane is a virtual

cloud, allowing you to not have a cloud and be cloudless and still run on Amazon, Google, Azure, Oracle without a cloud account and pay by the millicore and megabyte of compute. So if you deploy the Bret app, you can run essentially serverless on any cloud or multiple clouds without having a cloud account. You can bring your own cloud account. If you're a SaaS and your customers tell you, "Hey, you must run on Azure," or, "You must run on Oracle," or, "Must run on Hetzner," you now can. Your code, without any change, can run anywhere, even on-prem.

We let you have observability built in We let you, as I mentioned, run by the millicore. We let you, communicate across clouds. If you have data in a VPC, let's say RDS, Relational Database Service, and you want to run the workload on-prem, yet communicate with RDS without poking holes in the firewall, you can. You can, run above the cloud. So today, when you run RDS as an example, you're paying a high premium to run a database. You're not paying for the compute. Control Plane lets you run any open source project very easily.

We've certified, lots and lots of, templates. You can run Kafka, Postgres, MySQL, Cassandra, Mongo above the cloud. So if you wanted to move your persistent store to Linode or Hetzner or Oracle or on-prem, you can. If you need multi-master replication, we support that for many database types. we provide, sandboxes, so you can deploy workloads or, or write code in a isolated environment. The difference about our, sandboxes is that they allow you to communicate

to your databases and other resources in private data centers or VPCs. And you can traverse the network boundary, easily connecting Amazon US East and Azure US West and Oracle in, the Middle East, et cetera. So you can think of all the clouds have, having merged and, with a hundred x simplification on how to consume them, but without any limitation. So if you need to use the native SDK for SQS, SNS,

Dynamo, or any other native service, you can we, we-- think of us as a virtual cloud, just like what virtualization did to servers. You didn't care if it's Dell or HP, we do to cloud. We make it so you don't care and the agent doesn't care if you're running on Azure or any other, cloud provider. In fact, you can deliver traffic close to where the customer is, and if there is a failure, let's say Amazon US East is down, you automatically get routed to GCP or to Amazon US West or somewhere else. So you deliver low latency and extreme high availability.

We provide a five-nine guarantee, translating to five minutes, fifteen seconds per year of cumulative downtime. So the reason customers come to us is for unbreakable compute, for, extreme ease of use, for, cost optimization. Many of our customers save about fifty percent compared to running natively on Amazon or Google because They can pay for what they use and not pay for a cow when you need a glass of milk. 'Cause when you buy a Kubernetes cluster, you pay for the brain nodes,

the worker nodes, you pay a hundred percent of their utilization, even though they may consume twenty percent. And on control plane, you pay twenty percent if you're consuming twenty percent. Bret (2): so you built this sort of unified API layer over multiple clouds and, It sounds like you're adding almost like your own custom services on top of that like you mentioned the, the sandboxing, and obviously someone can, design their own little, EC2 instance or whatever, or they can, they can try to design their own sandboxing solution. But it sounds like you're, you're adding these as, specific services on top of that unified API, This is actually a long question.

I'm spitballing here. But, you could implement something in one of those sandboxes or one of your other, ideas on top of any of the clouds, whether or not that cloud has a particular offering labeled that, right? 'Cause when I think of unified cloud, I think of, reducing it down to its core elements and then, only getting to use those core elements. But it sounds like you're enriching on top of that API to give us higher level services. Doron: yes, people assume because we unify the cloud, we give you, lowest common denominator. But in fact, we're the opposite.

We give you the union of all clouds, because now You can write the code with a native SDK for Google, for Cloud Spanner, for Amazon, for, SQS, for, Azure, for cloud, for, let's say, Entra ID or Cosmos DB or whatever machine learning service that is on Azure, and we allow you to consume those services without credentials. Why is it important in the era of AI? You do not want to leak to the LLM, the credentials. So neither in design time, meaning coding, nor at runtime are keys exposed, and

in fact, you don't even have to rotate the keys because you get ephemeral, zero trust, least privileged tokens. and we patented the technology called Universal Cloud Identity that basically gives you the identity, the IAM of all the clouds in a much more, streamlined manner, and the code is devoid of credentials completely. Yet the code can only do the exact, operations on the exact resources you granted it access to. Bret (2): what are you seeing with adoption?

as a, as someone who's… I've managed infrastructure for over 30 years at this point, and, w- we've been chasing this dream in ops land of, the unified control plane, right? we've had for decades now, ever since really the cloud or before, we've had all these attempts at unifying. You know, VMware wanted to do it. Microsoft tried to do it. we had just all these different pieces, and the best we could do was, like, stick to one OS and then use their, I think they, system management tools or whatever. And then I feel like ever since then I've been chasing this dream, whether it was Terraform or Chef and Puppet before that, and then we got Pulumi, and

we've had all these different attempts at, one control plane essentially that controls not necessarily my workloads, but just the infrastructure behind it, and it's always felt like a compromise. It always felt like if I'm gonna do Terraform, that's great, but it's gonna get much harder as I go multi-cloud, and I'm gonna be spending my days in TOML hell just trying to make sure I don't break things with a, with an update. So I'm curious of, like, some customer stories around, onboarding. How do people get to control plane from where they are,

and what does that look like? Doron: Yeah. So we have, over, eighteen hundred, now over two thousand customers. Some of them are very large, some of them are tiny, doing vibe, coding type projects. And most of the larger customers, the-- none of them are greenfield. They're all coming from Amazon, from Google, from Heroku, from, uh, Azure, and they all have a sprawling, degree of Helm charts and, Terraform or Pulumi or Crossplane, artifacts. And we have, mechanical translation in the CLI allowing you to take a Kubernetes

YAML and enrich it to the Control Plane YAML because Kubernetes doesn't have the concept of the global virtual cloud, a cloud made up of N clusters. So in Control Plane, I can take three clusters or three hundred and turn them into one global virtual cloud or actually subdivide them into many global virtual clouds and deploy an application to the GVC, which is like a VPC, but spread across two regions or a hundred in one cloud or many So our

customers typically have a couple days up to a couple weeks, depending on how large the customer is, of migration, which is, again, that mechanical translation of their YAML or Helm chart or other artifacts to Control Plane. But we integrate very nicely into their CI/CD pipelines. We integrate with Flux or Argo CD. so it's very natural. The learning curve is very shallow. Once you learn Control Plane, you feel, and what I hear from customers, we run circles around our peers.

And one customer, I think the best compliment I ever got was after using Control Plane, moving back to not using Control Plane is like going from iPhone seventeen to Nokia from nineteen ninety-seven. That's-- And, and now, that was very rewarding for me to hear, and I would-- I'm eager to demonstrate. and my objective is after this, call, y- are you gonna go back to the old ways or are you gonna adopt Control Plane to deploy your code? I think it's the latter, because, it's-- we, we truly give you,

Dan Wilson as a service, the one I, I mentioned, I started with. He's our CTO, and he really, rather than you having to spend or the AI having to spend millions of tokens reinventing the wheel, focus on what matters, which is your rocket shipping system or your, whatever the thing that you're doing and not be, knee-deep in the toil. and there is a lot of toil. It's one thing to have toil, but why should every company reinvent the same wheel? Logging, metrics, tracing, secrets, m- mutual TLS between services because

you want a service mesh, audit trail, scaling, scaling strategies, failover Should every company reinvent that wheel? Does that make sense? Should every company have an Exchange server engineer babysitting email? That was the case. It isn't anymore because it doesn't make sense when you can, shift left or whatever, delegate that, mundane responsibility to somebody who is excelling at it. And think about it. If every company reinvents the same items I just mentioned, it represents technical debt because now you need to maintain, fix the CVEs, upgrade Kubernetes

every three months, upgrade Istio, upgrade Prometheus, upgrade Grafana. When do you have time to do the actual work when you're babysitting an infrastructure that your peers are doing the same thing? We say stop with the babysitting of infrastructure, start delivering value to customers, Bret (2): We talked about this last time when we, we had a r- a call a couple weeks ago. The, the assumption that I have about services like this, like if I'm looking at your homepage and I've not talked to you and I don't know anything about the company, is that, to me, there's always a compromise, I feel like.

And so when I think about the complexity of the cloud APIs, just for AWS, the, "This Week in AWS" podcast with Corey Quinn, I've been, I've been on his show, he's been on mine, but I, I always find his articles about, yet another AWS service that we don't understand what it is, and it's redundant. And, and so he talks about the hundreds and hundreds of services they have, and it's, th- there's always a customer for the things, but it may not be you or me. And I always find it hilariously complex that I am so behind on how many of those products I don't know about and I've never used, and it just,

Am I even an engineer if I don't understand half the things that AWS deploys, right? Sometimes I question it. But, I'm curious about, how your API relates to their product APIs, and how you guys have been able to innovate around covering all the bases to make sure that we're not left with, this reduced abstraction that's only the same core fundamentals across every cloud, but you're allowed to, you're able to e- enrich it a little bit. Tell me a little bit about that. How's that work? Doron: Yeah, that's a great question. So I mentioned we're not the lowest common denominator, but the union, and what does that mean?

It means if Amazon came up with a Bret service a microsecond from now, we would support it because we're not trying to, abstract the API, the native API to the Bret service. We integrate at the identity and access management layer. And so we would mechanically be able to pull permissions from their IAM IAM system relating to the new Bret service. The coder, AI or human, would use the native SDK if there is one to the Bret service or the C-- or the API.

We just inject the token necessary for the Bret service to do its work, and we're able to do it whether you-- the workload is running on-prem, on Hetzner, on DigitalOcean, , on Google, on Amazon, on Azure. So because we integrate at the IAM, the identity and access management layer Bret (2): Yeah Doron: We are supportive of any new service that Google, Amazon or anybody else provides instantly. There is no work that we need to do. And the likelihood of Amazon or anybody else changing their identity and access

management APIs that we exploit are zero, because if they change their APIs, the cl-- the world would break, right? Because everybody, millions of people have, utilized their services. So if they were to change, they would give you multi-year head start. but, because again, we integrate at the IAM level, when you consume, let's say, RDS, but you're on, Azure, the, the, the workload is running on Azure or on-prem, we obtain the correct token from their STS, their security token service, and

know how to inject it at runtime from the SDK to RDS, in our case, And, so in effect, the workload is liberated and no longer captive to the subject cloud. You can now move amongst the clouds. People spent millions of dollars moving from Amazon to Google or Google to Oracle, but on Control Plane, the cloud is your oyster. You can move from Amazon to Google, from Google to on-prem, from on-prem to whatever with a click of a button, and you can leverage all of them.

So we flip the script where the clouds want to-- their objective is to lock you in. We flip it on its head and, turn them in-into just commoditized utilities. So when you consume electricity, your TV doesn't care if it's Edison or somebody else. It cares about the interface. With Control Plane, you now have a virtual cloud operating system that utilizes the power, the whole power of Amazon, the whole power of Google, et cet- et cetera, but without locking you in.

Bret (2): Yeah. I wanna change the conversation a little bit to be more specific on some of the AI, uh, stuff that we're all managing right now. So, this whole year for me has been really focused in understanding other DevOps teams and, platform engineering teams, like understanding the way that they're using agent harnesses to do their job. I think, if we f- if we rewound a year ago, maybe even three or four years ago when Google Copilot was kinda like the first IDE to throw AI completion in, and even in those early days, I think I demoed it when it was in beta when we were on this channel.

I remember one summer we were, I think it was 2022 or something. It was before ChatGPT came out, and we were, like, looking at the, the early GPT model that was embedded in, in GitHub Copilot, and the, VS Code UI. And we were, we were trying to see how far we could get with YAML and TOML by tab completing, right? Like, we were trying to figure out how much time can it save me as an, as a ops engineer, but we were really just talking about file completion, completion in a YAML statement. So, it was saving me keystrokes, and it maybe, was saving me potentially some human error that would hopefully be caught by a linter,

but it wasn't a huge gain, right? And so you fast-forward a little bit, and then we're all, we're doing that with ChatGPT. People are copying and pasting. And then last year when we got Claude Code for the first time and we under- we started to understand what an agent loop is, it, it started to shift and I started to notice that my, my production friends were all dabbling with letting their, the first thing they would do is let their agent have, a read access token to some part of their infrastructure so that it could maybe run the kubectl commands for them or run the AWS CLI commands for them. And I was doing that same thing for, like, GitHub CLIs for, for CI.

CI. Like, I cared about GitHub Actions and, my build status and stuff like that, so I was, I was managing that middle portion, and I was using agents to, keep track of it. And now it's like these are all inherent features that are really built in. They, they've got the background workers, the sub-agents, and things that are making this really possible. But I'm curious what you're seeing with your customers. how, how are they driving their infrastructure with AI? Like, I guess maybe if we can get to some, into some nitty-gritty details around, are they focused on only using their agent harness? Are they using, MCP to, to Control Plane?

Are they using, Control Plane CLIs? Like, how are they actually doing this work, and what are some of the stories you're hearing? Doron: Yeah, our customers, pretty much all use our MCP server, and we always pride ourselves in providing all the accessibility. Of course, at the lower end is, is the API, but everything you can do in the API, you can do in the CLI, Terraform, Pulumi, Crossplane, UI, and MCP. To us, these are symmetric interfaces.

Of course, with MCP, you can now bring the intelligence to let it discover what can be done, and now you can have conversational infrastructure. You now have the equivalent of a enterprise architect/platform engineer/DevOps/SRE bundled into one that you can have a conversation with. you can tell me the app you want to build and the dependencies that you want to have, and the MCP will go and negotiate and build those things that we want to build for us. And then if we need to troubleshoot or need to look at observability, the

MCP is able to look at the traces, the logs, the metrics, a- and add the intelligence to give you actionable items. So you feel like you have the best SREs and platform engineers at Google with you twenty-four seven, and you don't pay them. So you can vibe code a very complex system made up of 20 microservices all intertwined with databases and queues and, and caches and say, "Okay, give it to me codified." And our codified version is always backwards compatible.

So I'm sure Kubernetes deprecate stuff every three months. You have some deprecations and new APIs. It's a fairly unstable substrate. So if you wanted to take YAML from 10 years ago and just run it, you'll run into issues. So we made an early choice. Bret (2): At a minimum warnings, possibly errors, yeah Doron: Right. And we made a decision. We, we noticed that six years ago and made the very, I think, good decision of always having backwards compatibility. So if you take YAML from six years ago and apply it today, you'll get the exact same

artifacts on top of any compute substrate, whether it's on-prem, Azure, etcetera. Bret (2): we-- Are-- When you s- when you say that, are we talking about your customers can continue to use Terraform or Pulumi? I'm assuming that that's how they're, they're interacting with you, right? Like, they're, they're talking to your API from these particular management tools, right? I'm ass- I'm assuming, um, OpenTofu is s- is supported. What else, what are the other ways? 'Cause, o- one of the things I've started to see with agents is if, if you're a singular cloud, it might just be better if your, if your

agent is the complete interface to h- whatever the description language is. I've seen some examples of people actually switching away from Terraform back to the native, cloud formation or whatever of their cloud, because they look at it as there's this agent abstraction between me and the YAML or the JSON. I don't have to be the expert in every single line of that anymore, so I actually care more about, some of the inherent functionality of a different, of, of the language tooling itself rather than, we always looked at Terraform as maybe a little bit more to, easier to understand than cloud formation.

A little bit, a little less comprehensive, a little, little less verbose. But if you've got an agent in the way, that matters less. Are you seeing anything like that? Are, are you seeing people, switch, products now that they're using agents, as their interface? Doron: Somewhat, yes. I've seen people that, that used certain… I don't want to name names 'cause I don't want to dish on any one platform, but there are many equivalents. Like when-- whether you use Crossplane, Pulumi, CloudFormation, Terraform, and a few others, they're functionally equivalent. They're like French, English, Spanish.

You can express the same ideas in these languages. They do things in slightly different manners, like Pulumi is more, programmatic than, than, declarative, and Terraform is more declarative. But at the end of the day, they express a desired state, and there is a translation layer that takes the desired state, like with DAG on, on the case of, Terraform, and applies in the right order, calls the APIs that are affecting the, the actual resources in the cloud. It's just a translation, right? So like you said, if I don't have to write to that language,

why can't I talk in English and let it make the calls directly? But here is the thing. Terraform, as an example, or OpenTofu, invested heavily in doing the DAG and doing the validation and doing stuff that the agent itself will have to reinvent what they had done to apply things in the right order and give you the rigor that they have already baked into the platform. It is better, like in the Excel example I've ga-- I've given, to let the agent leverage a tool like Terraform than to learn, the underlying, API.

It's, it's possible. here it's, it's, it's not an exact science. Here is subject to, to, judgment. It can call the, the direct, it can call the API, it can call the CLI of the AWS, let's say. It can call, a CDK, There are many, again, equivalent, logically equivalent, options here. So Bret (2): So Doron: our Bret (2): you, I Doron: w- w- we, we just give you the, the one API, and it's your choice. You tell the AI, " Use the API, use Terraform, use YAML, use JSON, JSON, use, CRD." We give the choice to the customer or the AI, and the AI would

choose the, the least token-consuming manner that it can operate in Bret (2): So when someone's migrating to your platform, if they're using CloudFormation, d- I'm assuming you don't support CloudFormation because you're sitting in front of the AWS. Like, you're, you're Doron: Correct. Bret (2): different Doron: we, we don't support cloud formation for control plane. We have customers that use cloud formation to create their S3 buckets and their RDS and their Dynamo, then, then they tell control plane, run a workload that connects to the S3 bucket to RDS to Dynamo.

They're not, in conflict with each other. They're just not overlapping. There is no reason for us to introduce a cloud formation concept 'cause it's a proprietary, lingo of AWS. So we are operating at the cloud agnostic, layer. We fly above the cloud. We don't wanna fly in, in that cloud. We fly above it Bret (2): Do you see, do you think that Terraform is still the primary way your customers are managing their infrastructure? Do you see that as, the majority? Doron: I would call it 60%. Bret (2): Okay Doron: Just because of inertia, Not because it has a, a, a

specific advantage, like I said. Whether it's Terraform, Pulumi, or Crossplane, or something else, All they are, are syntactic sugar on top of the API with some DAG, that, that decides what to create first and what to destroy first, and maybe save the state and so on. So to me, those are fairly commoditized, languages. So for, for the audience to understand, we don't compete with Terraform, not even a little bit. It's like a car doesn't compete with, a sandwich. It's two different animals completely.

Bret (2): yeah. so we talked about customer base, and, like, the different size of customers, right? And it's hard for, any one cloud to always, a-adopt all customers, right? We, we all think of, the digital oceans and, the Vercels of the world, that they're all like a very niche cloud, that they've got a very specific purpose, and then you have the three hyperscalers that are just everything for everyone, but it's also impossible for anyone to get on board really easily, right? Do you, do you see, are small teams of, with n-no developers or very few operators or anything like that, do they-- Is Control Plane being used by

those kinds of small little companies? 'Cause I, I have a scenario where I'm thinking, I have, companies that are, they, th-they might have one or two IT staff, but, they don't have internal developers. So usually when it, it's hard to figure out, as a consultant, your heart, it's hard to figure out, what do I do for them? Because I either have to put them on a hyperscaler, and then now they can't do anything without me, right? I become a bottleneck, which I, I, me as a consultant, I don't ever wanna do that. I wanna be instantly replaceable re-usually. Or I have to put them on some sort of niche cloud that, if it's too niche, it

may not be here in five years, or it might be somehow limited i-in, in what it can do, and then I'm li- like a Lovable for AI or whatever, like these, these little fly-by-night dashboard companies, right? Do you… W-what's it look like for the really small teams on, on Control Plane? Doron: The good news, you will always have a substrate to run upon 'cause it can be, like I said, your on-prem, it can be Azure, Google, Amazon, Oracle, et cetera. And we have customers that have, non, traditional developers, certainly no dev… We have DevOps engineers that use Control Plane, platform engineers,

but we have people that are just vibe coding and not know anything about… They don't know the if statement, and they still deploy successfully on Control Plane. case in point, my younger son, wanted to build an app. He never wrote a code, any piece of code in his life. He was able to write a full game, which was pretty impressive, and run it on Control Plane, and he didn't even ask me how to do it. So he deployed the whole game with a back end, and, I was shocked. You know, I, I helped develop the product, and I was shocked at, at, at what he was able to do, and he's never written a line of code

despite me wanting him to, to learn. But he was able to do it. So we have many people with no coding experience that are able to be very successful. Bret (2): Yeah. Do you think that's a-- With AI, I mean, I, I know my answer to this question, but, it's like a leading question. Do you think that with these agents getting better and better and th- there's the… Now we got Cowork and, and Claude Code built into people's Claude and ChatGPT subscriptions for people that aren't technical, right? You don't have to use the TUI anymore. And so I feel like we're, we're heading to a, pretty quickly to a place where non-IT staff are getting more and more power to be able to potentially do things like

vibe code their own app and deploy it on the internet, and If they don't have the IT staff to help them manage it, they can, they can at least get by with some, some simpler stuff. do you see that happening on Control Plane too? It sounds like it is, since your, since Doron: 100%. We are, we are releasing a, more consumer-oriented… I'll come out and say it. We are, the, shipping very soon. We're in the tail end of, Vers- not Vercel, but, Lovable and Replit, Base44 killer. And why we're able to do what they do, but they're not able to do what we do, is the coding agents.

when you look at Replit, most of the effort has been put into, Guardrails and skills that are just, wrappers, assisting agents for, for the LLM. and, and those are now in the LLM itself most of the, of the time. The heavy lifting is allowing you to run on any cloud, letting you scale using a hundred different scaling strategies, letting you fail over automatically, leveraging hundreds of diff- the operational stuff. So we've been hard at work building this operating system for six and a

half years, and it's not something you can vibe code, absolutely not. And so we built this operating system, and for us to layer the, consumer-facing app builder is a lot easier than for an, an app builder to create an operating system for running, anywhere. Replit is I, I played with it a little bit. It's a, it's a good tool. I don't know that it will let you scale to Netflix grade. I, I think what happens is, and I wrote an article that I'm publishing, but many

people, when they start scaling or need security or need compliance, they get off of that vibe coding platform, where Control Plane takes you all the way to Netflix grade cloud native maturity. You get the compliance, the security, the scalability, the, unbreakable compute. Bret (2): I would just assume that a lot of those platforms are not like, SOC 2, PCI, you know, HIPAA, Doron: HIPAA, right Bret (2): all the things that like a dec- any decent sized company is gonna care about, and there's probably someone in marketing vibe coding something on Lovable

that's completely not up to standards. For me, my experience so far has been one of the biggest challenges in like the vibe codings products out there, is they don't really answer the data problem, right? Like, 'cause usually m- everyone that I've talked to so far, like creating a dashboard is cute, and these are, these are important things for teams. but th- th- there needs to be data on that dashboard. There needs to be data in that app, and it's usually existing cust- company data locked up somewhere, usually not behind like a perfectly designed open portal with a proper API across every surface with, with automatic, token delivery

for people that request it to ins- Like, there's just this whole infrastructure of cloud native that a lot of companies they haven't matured to yet. And that's to me, like building these vibe coding platforms for companies, which I think, yeah, again, like we could probably do a whole separate podcast on, but like that world is inherently necessary that you have a very modern mature infrastructure before you can do that. ' Cause a lot of it is about, automation and security and encryption and, l- like you said, auditing and compliance. That's, that's where I live, and that's where the vibe coded app, while it may

not be stable or highly redundant, it has to… It, it… Those things are almost like optional, maybe someday things. But what's not usually optional is for this thing to be useful, I need to access that SQL database, and how do I get that, right? How do I ensure that that data doesn't leak? How do I make sure that this is behind the firewall? Like, there's all those levels of complexity that I don't see the Replitz and the Lovables of the world ever truly tackling. So I think you're right that this is probably… These, these solutions are probably gonna come more from infrastructure companies like yourself rather than a, a startup that's focusing on AI as a, as a sort of, magic button.

So, I'm, I'm interested to get into that. So I know you brought some demos, and we could talk about this product forever, but I, I feel like we've, we've covered some of the questions that I wanted to, to know about. l- l- what are we gonna, what are we gonna look at today? Doron: So I want to show you, I, I want to do a tour around the platform, and I'll give a sneak peek at some of the power you get right out of the box, and traverse the network boundaries, allow you to run an app anywhere, even on-prem, that consumes any service of any cloud, show you the observability, show you

the audit trail, show you the Terraform, capabilities and, and stuff like that. So just give you a general tour so people know what exists. And, most people don't know that they have at their fingertips so much power, and they keep struggling with Amazon. And, w-where you can get a hundred x the productivity by leveraging, a, a platform like Control Plane. and not only that, save a bundle because you're not running full EC2 instances, you're running the millicores. So you can literally run your Bret app for six-- five and a half cents a

day, a day meaning twenty-four hours, literally less than six cents a day. and yet people are spending money on NAT gateways, load balancers and, Kubernetes nodes and pay hundreds of dollars per month when they can be paying, cents on the dollar. And that's what I want to show. Boom questions? Bret (2): no, this is, this is a lot. That's, I'm trying to think of my workflows and, like, how my workflows for creating infrastructure and managing infrastructure need to adapt to this model, 'cause it's, it's flexible enough. you have click ops, right? but it seems so easy to hop back and forth between the click ops, 'cause it feels

like the click ops is backed because you have that versioning and the audit trail. you don't, it doesn't, it looks like you've removed the risk of the, sort of the rogue click that now is no longer compatible or, or can- isn't tracked in a way that you can update your Terraform. this is the classic problem of, we're git- we're trying to do GitOps, and we're putting all of our infrastructure in, in Git, and we're storing changes there, and we've maybe got some automation for Terraform plans and applies, or we've got, we've got some stuff going on there, and that's the only way we're supposed to do it. And then we get some person that shows up and in an emergency or because they

didn't know or whatever, they go into the AWS console and make a change, and now I'm bifurcated, and I have no path to success because AWS isn't gonna really easily export, it's not gonna to export Terraform for me. And so I have to figure out what they changed before, I risk blowing their stuff away with my future Terraform apply. And so we, we've always had this challenge in infrastructure of, sometimes we do need the click ops. Sometimes we ne- do need to be able to go into a GUI really quick, troub- especially when you're, like, troubleshooting or just trying to recover from outage.

You don't wanna stare at a bunch of l- code in TOML forever. So you're, you're usually in the, in the dis- in this GUI app trying to figure things out, but you also know that y- it's, you can't touch it. It has to be read-only because you need to make your changes somewhere else, and you're so nervous. So it feels like you've kinda thought about that a little bit. Doron: Yeah, and everything, like I said, everything you can get as, YAML or, or JSON. But here is our app running, and so we give you a more consumer-y interface to this thing. but the bottom line, it's really backed by the control plane API, and the control

plane API lets you run anywhere, and with, with determinism, and, and it's the same. You c-- Again, you can run it on your Raspberry Pi or Dell machine. And the cloud wormhole, which I haven't demonstrated, allows you to very, very easily connect to an RDS instance in a VPC or connect from a, the seven different locations here to my own Redis on my machine. And we have a service catalog where you can run above the cloud. If you want to run CockroachDB or Cassandra or really any open source

project, we certify these s-services. Like let's say I want to run Hermes Agent or Langfuse or Keycloak or, you name it, Postgres, Mongo You can run them at the cost of the infrastructure instead of paying a premium to RDS. You pay double what the compute is, but Control Plane lets you do backups, lets you do leaders election, lets you do multi-master replication with different add-ons. You can run Ollama, you can run whatever you want to run. and you can conversationally say, "Hey, get me an N8n, run me a Hermes agent,

run me, know, OpenBaO or run me whatever, Postgres very, very easily." you, you can do all of, all of that stuff. And again, you can run anywhere. Bret (2): Yeah Doron: w-we're not going to do an exhaustive demonstration. one of the things that, that people care about is security, and you can set very granular policies of who is allowed to do exactly what. You can say, this agent is allowed only to push image but not delete an image. This Lisa can only, read a certain value but not write. Secrets are only available to X and Y.

So everything is very, very granular. and th-there is a lot, a lot more. You can set custom domains. we have customers with twenty thousand different domains, and they run perfectly on Control Plane. It gives you the confidence to operate above the clouds without being bogged down by the complexity of any particular cloud. And the thing I wanted to show, because we have this thing called Capacity AI That is by default on. If you look at the spend, you're spending twenty-five millicores 'cause the system

learned over time that Bret video game only needs twenty-five one-thousandth of a CPU and thirty-two megs of RAM. If you go to the control plane pricing and plug in-- And you can bring your own compute, or you can use control plane compute, or you can do both. So twenty-five millicores and thirty-two megs of RAM cost you a dollar seventy-eight per month, period. You don't pay for load balancers, you don't pay for nodes, you don't pay for anything. That's all you pay. We charge you by the second. let's say a full hour would cost you, No, sorry, a full day would

cost you less than six cents. So if you ran the Bret app for a full twenty-four hours in a location, it would cost you six cents. That's why I'm saying sometimes running it in more locations is even cheaper than running it in a single location, 'cause if you were to stand up a EKS cluster, you would pay hundreds of dollars in a month to heat up the planet Earth or to make Jeff Bezos richer. We let you pay for what you eat, not for heating up the planet Earth. So extremely cost-effective Bret (2): In that, in the, the workflows you were showing

as you would, as you were doing this stuff, I w- guess one of my questions is Deployment configurations and design can often be, hard for teams, especially when they start to adopt Kubernetes, right? Because then they have to pick Flux or Argo, and they have to decide on a particular way to deploy that, and then they have to configure it all correctly, and then when they have b- a bunch of different apps, yeah, so my premise was, like one of the, one of the challenges that I, I train people on and is a lar- a d- decent amount of my consulting over the last six years has been once teams adopt Kubernetes, there becomes this deployment complexity because it's

not enough to just have Kubernetes. Then, then you need, your code's stored in a place, GitHub, GitLab, whatever, and then you gotta connect all that together. And, in the old days before containers, a- and even when we first, first started using containers, it was very much "I'm gonna write a CI job that runs kubectl apply," or, "I'm gonna run a Docker run, and, and it's gonna connect to the server from my CI." And then we grew up, and we realized we maybe need a more deterministic methodology, and then we started adopting Argo and Flux. And, and so for the longest time, I've been f- I've been advocating

and teaching those technologies. We call it GitOps, but it doesn't necessarily have to be labeled that. and I'm curious how you, how your platform works with that. Does that m- It, it… You've got a lot of, glue there that feels like it makes a lot of these rough edges streamlined and it's almost like you're applying opinionated defaults, which I love because I'm a, I'm a very much a, an opinionated defaults kind of guy. when it comes to like the code in the repos, and is it as easy as someone on your platform pointing to repos, and then it figures out how to get it into Kubernetes?

do they need to even worry about Argo at this point? 'Cause you've been saying things like deploy this app, and it's doing it. I don't know how it's doing it. I know it's, I know it's running on Kubernetes. But what's that, that deployment look like for Control Plane? Is it automating it? Is it, is it doing Argo for you? Doron: So it c-- you can use Argo. So we have, a CRD, a control plane CRD, so you can have a cluster and, use Argo or Flux very naturally. the way it works, so you said, "I'm not sure how it works," so let me explain. you get the YAML.

Let me share real quick. and again, YAML is just one of many. So if… let's take this guy for-- as an example. I can export, this YAML. So you can see the kind is workload. It's a new kind that doesn't exist on Kubernetes, and it has the different, like what's the memory, what's the scaling strategy, what GVC does it go on? And you j-just say CPLN apply and give it, minus F and give it the YAML, and it deploys it. Like I said, we have other ways to deploy, but at the end of the day, we have a data service.

We have an-- we have the control plane API, and you make a call to the plane itself is, is running on Amazon, Google, and Azure, and the-- our data is distributed. It calls a component we call the which is a control plane, Go-based application. It runs on the edge, we call them edge clusters. So you can have one cluster, or in our case, we have one, two, three, four, five, six, seven clusters. These are different clusters. They can have hundreds of nodes. we provide clusters that are, cloudless, meaning we provide clusters, or you can

bring your own, you can bring your own cluster like I did in the Ron office. I created the Kubernetes brain in the cloud, but I could have run OpenShift or EKS or whatever, in the local environment The-- what you need is a Kubernetes cluster with the actuator and a few other components like the, the log, collector, the metrics collector, and so on. But we control planify a cluster, and then it becomes part of that global virtual cloud. We use NATs under the covers to have our, control plane, if you will, communicate the desired state to all the clusters, and they all take the desired

state and actuate it in the cluster. So under the cover, you have, stateful sets or, deployments or, pods and all the Kubernetes goodness that is happening, but it has been abstracted. And if you want to see the Kubernetes resources and, and it's a bring your own Kubernetes like I have here, you can get the kubeconfig and go edit or use the dashboard and go edit. But we… You noticed we didn't have to issue kubectl apply commands and all that because there is an actuator component in this cluster

that went and did the actuation of the desired state Makes sense. And also, if we are down, the, the cluster operates autonomously and serves traffic. Maybe you can't upgrade the version of the workload, but A, we're very highly, redundant and highly available. so we're very, very unlikely to go down. But if we were to go down, we are the control plane, not the data plane. The data plane is where the data, flows, and you mentioned that earlier. so hopefully that gives you kind Bret (2): the goal, right? Is like the control plane never, never takes down the data plane

when it's, when it's unavailable. Yeah. Doron: Exactly. A, we're extremely… Yeah, and we test everything, like, like I can't even tell you. From day one, we created what we call the Observer Project, and it does twenty-four seven millions of tests simulating node failure, multi-node failure, cluster failure, AZ failure, region failure, whole cloud failure. And we do, fuzzing. We do random failure, introduction of failure modes with a core objective of making sure you never go down. So one of the things you get with Control Plane is mainframe-grade

availability over commodity infrastructure Bret (2): All right Yeah, the fractional CPU, stuff is, is pretty flexible and pretty awesome. When you're running small stuff and you're a small company and every single workload you add has, this minimum cost to it, and Even when you try to run, small instances in an EC2 cluster, you're limited on how small you can go, and then you're inherently bottlenecked there if you ever get any decent traffic because it, it doesn't know how to… You, you have to build a lot of intelligence in in order for it to auto-scale up and correctly change the node size, if you need more memory

because, a classic example is, a legacy app with a PDF export and someone decides to PDF export a report that isn't batched properly, so then it consumes all the memory on the server for one report, and then the, and then your tiny little app crashes the whole thing all because you wanted to print a report. So there's always, there's always, sizing challenges because you just want things, especially when, when it's, a small business kind of situation where, they, you tell them, $500 a month for something that they consider not worth all that is a hard sell when you're like, "Well, that's what it takes to be minimum redundant

across regions in this particular cloud." And the idea that I can say, "Well, you're, you're only using this app internally, so we don't even need a single CPU. We need a half of one. We need a third of one," Doron: Yeah, we need a fif-fiftieth of one or, one, one-twentieth of one. Yeah, and that's… And you can now give people redundant availability for three dollars, five dollars a month. you can run an app for five dollars a month that is-- has the availability of a real enterprise. Bret (2): how do people get started? They go to Control Plane.com and sign up for an account?

Doron: Yeah. Yeah, it's very easy. so they go to Control Plane, they click sign up, they follow the prompts. we ask for a credit card, and I find some people get intimidated. we're happy to give you, especially to your listeners, shoot me an email directly, Doron, D-O-R-O-N, @ Control Plane.com. I'll give you, basically a, a free month to experiment with. Of course, don't do crypto mining and don't do crazy stuff, but happy to let you experiment on the platform. Even if we were to charge you, you can run so cost effectively that unless you

reach five dollars in spend, which is a lot on Control Plane, we don't charge you. So if you're under five dollars a month, we don't even charge you 'cause it's not worth it to send to Stripe. So many people can run real projects on Control Plane for less than five bucks. There is no platform fee, there's no per seat fee, there's just usage fee, and it's very, very cost effective. Bret (2): That's what I used to use Heroku for, so now m- now maybe you're my new Heroku Doron: Think of us as a non-toy Heroku. Bret (2): Yeah Doron: the non-toy, replace your, Heroku with, the, the, the cloud du jour. But, we give you all the power of all the clouds, but, with the

ease of use of Heroku or better, Bret (2): thank you so much for being here, Doron. Doron: thank you. Really enjoyed, the conversation. You asked great questions. super happy we connected Bret (2): And we'll see you guys next time. Ciao.

Open the transcript file

Hosts and Guests

Hosts

  • Bret Fisher Host Feed-stated

Guests

  • Beth Fisher Producer Feed-stated
  • Cristi Cotovan Editor Feed-stated
  • Doron Grinstein Guest Feed-stated

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. CI/CD via agent tools: Skills, CLIs, MCP, and more with Semaphore

    Aug 13, 20261 hr 17 minAudio

    What does an AI-native CI look like? The Semaphore team joins me to talk about their focus on making your agent harness the gateway to testing, fixing, and deploying your code. Video podcast version here:...

    • Apple Podcasts

    Credit the people on this episode

  2. K8s Maxxing with AI-Native Platform Engineering Stack with OpenChoreo

    Jun 13, 202655 minAudio

    OpenChoreo is a “batteries included”, AI-native Kubernetes platform stack for Platform Engineers that combines GitOps, Observability, AI Agents, and Workflows into a custom K8s distribution “super pack” that is managed...

    • Apple Podcasts

    Credit the people on this episode

  3. Docker AI, what’s new with MCP, Agents, Sandboxes, and more

    Apr 7, 20261 hr 16 minAudio

    Michael Irwin of Docker joins me to run through Gordon AI improvements, Docker Hardened Images and what's now free, Docker Sandboxes for running agents in proper isolation, Model Runner updates including MLX support on...

    • Apple Podcasts

    Credit the people on this episode

  4. Backup S3, Google Drive, iCloud, Notion with Plakar

    Mar 20, 20261 hr 7 minAudio

    Bret is joined by the founders of Plakar - Julien Mangeard and Gilles Chehade - to nerd out over backup engineering. The kind where you're building your own file formats and cryptographic layers, not just wiring up cron...

    • Apple Podcasts

    Credit the people on this episode

  5. Your Images are Out of Date (probably) - The Silent Rebuilds problem

    Mar 4, 202637 minAudio

    Container base images (like Official Docker Hub images) are often updated without new tag versions. I call this Silent Rebuilds. There's no way to know this happens without image digest-checking automation like...

    • Apple Podcasts

    Credit the people on this episode