[00:00:00] Speaker A: Foreign.
[00:00:06] Speaker B: Welcome to the cloud pod where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure.
[00:00:14] Speaker C: We are your hosts, Justin, Jonathan, Ryan and Matthew.
[00:00:18] Speaker A: Episode 306 recorded for May 27, 2025 batch. Better have my sequel. Azure's maintenance makeover.
[00:00:26] Speaker C: Nailed it.
[00:00:26] Speaker A: Good evening, Ryan and Matt. How you doing?
[00:00:29] Speaker D: You got the rhythm right? Good job.
[00:00:30] Speaker A: I did. I was working on it in my brain, like how this rhythm has to be correct, which does not make sense.
Well, we've got another action packed show. We luckily did not get burned by more and more build or Google I O stuff. There was some interesting things that we did not cover last week, like the return of the glass holes. Always exciting.
The new Gemini powered glass holes.
You know which the demo looked pretty cool, but I have no confidence that it actually works as smooth as they showed it. Or the battery doesn't last more than 10 minutes, but pretty neat.
[00:01:03] Speaker C: Well, they're no longer alone, right? Doesn't Meta have some sort of integration with Ray Bans? Yeah, oh yeah.
[00:01:08] Speaker A: Meta has glasses and Ray Ban has glasses. But those are like really simple. They're just, you know, just photos really is what they do in those. Okay, this one here actually has Gemini integrated into the glasses. So as you're like the video they showed from IO, she walks around and then she comes out on stage to talk to the host at the time and she's like, oh, Gemini, what was the coffee company? The guy's. What guy's mug was from that she was talking to in the back. Wasn't even a question I was thinking she was going to ask, but literally had a coffee cup and it had a logo and so literally it told her, oh, that's blah, blah, blah, coffee and it's three miles away from here.
So yeah, it was a little creepy as well. Okay, now you're noticing things that no one else is noticing. You're remembering that at some context that's a little bit concerning. So yeah, again, the Glassworld thing might come back in a vengeance with some of these changes. So we'll see.
Well, a lot has happened this week in AI and the first up is Claude. Claude has turned four. Welcome, Claude. The latest model in the Cloud Opus 4 and Claude Sonnet 4 setting new standards for coding, advanced reasoning and AI agents. Cloud Opus 4 is the world's best coding model with sustained performance on complex long running tasks and agent workflows per anthropic. Opus 4 has 350 billion parameters, making it one of the largest publicly available language models. And demonstrates strong performance on academic benchmarks. Sonnet 4 is a smaller 10 billion parameter model optimized for dialogue, making it well suited for conversational AI applications.
Alongside the new models, they're also announcing a bunch of new features including Extended Thinking with tool use in beta. Both models can use tools like Web Search during extended thinking, allowing Claude to alternate between reasoning and tool use to improve your responses.
New model capabilities, including the ability to use tools in parallel, following instructions more precisely, and when given access to local files by developers, demonstrate significantly improved memory capabilities, extracting and saving key facts maintained, maintaining continuity and building tacit knowledge over time. Cloud code is now generally available. After receiving extensive positive feedback during our Research Preview, they're expanding how developers can collaborate with Claude, and Claude can now support background tasks via GitHub Actions and native integrations with VS Code and JetBrains, which will display edits directly in your files for seamless pair programming, as well as new API capabilities with four new capabilities on the API that enable developers to build more powerful AI agents, including code execution tool, MCP Connector files, API and the ability to cache prompts for up to one hour. And if you'd like Pokemon Clock now play Pokemon for you, which they demonstrate in the blog post. If you're curious to check out, There was several benchmarks mentioned in here, and my favorite one, LLM Leaderboard, was not updated yet with Claude, so I'll keep an eye on that to see when that shows up. But in the benchmarks they shared with us, they showed that for junt encoding Claude Opus 4 is 72.5% versus Cloud 4 sonnet at 72.7%, which is significantly 10% higher than Cloudsonnet 3.7 at 62.3%. I have no idea what those numbers mean, but they're bigger.
[00:03:57] Speaker C: I do, only because I've been in the midst of using this a lot and then going back between 3.7 and 4 largely due to like being rate limited.
There's a noticeable difference in in 4.0 it is better at delivering code working code the first time without having to go back through multiple iterations and it's a it's kind of neat. This is the first time I've ever actually been able to notice a difference, to be honest, it's that big of a difference where it's kind of neat. So it's like I don't think I remember seeing this big of a difference between 3.5 and 3.7, for instance. Like it was all kind of yay, it does the things.
[00:04:38] Speaker A: But it sounds like multilingual Q and A. It's only like 1% better. But others there was more significant increases. So high school math is now getting a 75.5%. So it's passing with a C. Before I was failing with a 54.8 and clouds on at 3. So it's getting better at high school math so my kids can cheat more.
[00:04:57] Speaker D: It's probably better than I am at geometry. Let's just go with that.
[00:05:00] Speaker C: Yeah, well, yeah, I can solve a.
[00:05:03] Speaker D: Proof better than me.
[00:05:04] Speaker A: Apparently OpenAI03 is still better than all of them because their Math score is 88.9%. I'm just saying. So if you need to cheat on your math test, high schoolers use OpenAI03.
But yeah, no, I overall glad I had told you guys previously that I paid started paying for Cloud Max. So I've been trying to use cloud code more. I still hate the fact that it's an NPM application because in fact my Macintosh with npm so that's a bummer. But number two, it's supposed to have this really simple integration where you go to Visual Studio code and then you launch cloud code and then it's supposed to say, oh, I see you're in Visual Studio and I'll install the plugin and then I'll integrate.
My experience with that so far is garbage. It does not detect my Visual Studio code.
I think it's because I'm using one from Homebrew and I think they're very happy path. This like, oh, this is somebody who downloaded Claude or sorry VS code and then install it in the directory. So I'm hoping that gets fixed. I see other people complaining about similar challenges and you know, there's all kinds of suggestions in the, in the GitHub forums how to fix this issue or manually install it. So I'm trying to work down the path, but I haven't quite figured that one out yet.
I'm hoping to get that one pretty soon though because I would like to be able to try the full experimentation on it.
[00:06:14] Speaker C: Yeah, I haven't been using the native integration. I've been using GitHub Copilot and you're just selecting the Cloud 4 model. And so it's sort of, you know, it's all layers of abstraction which I.
[00:06:26] Speaker A: Would do too if I could because I really like root code and Klein and those tools. But I had to pay for API tokens there where with my Mac subscription I get that code for basically for free.
So I like to use cloud codes Because I don't want to. I'm tired of charging $25 every couple days when I run out of API tokens to my credit card because that gets expensive.
And so I would like to use cloud code. So I have some, some desire to make that happen. So I'm hoping by the end of this week they'll figure out the code glitch because this was announced late last week after we recorded and I assume they're getting all kinds of feedback and bugs and like, hey, this doesn't work. Just, you know, honestly, Claude, just put it into the marketplace like every other Visual Studio code and I'll just go, I'll just go find it and I'll install it. Why do you have to be fancy? Yeah, like just give me the basics.
[00:07:14] Speaker C: Yeah, I wonder why. There's got to be some sort of weird technical reason they're trying. Either true or not true technical reason.
[00:07:21] Speaker A: Yeah, I think they're just trying to mess with.
[00:07:27] Speaker C: I do think they'll get it fixed though, because it's a very popular space right now. A lot of potential.
[00:07:31] Speaker A: Yeah. And a lot of people install things through Homebrew. So I'm not the only person who's got a weird path for their Visual studio install.
[00:07:39] Speaker D: See, the funny part is I don't think that's the weird path. That to me is the normal path.
[00:07:42] Speaker C: It is also because it's the best path.
[00:07:44] Speaker D: Like if it's not in Homebrew, it's not, it's not real.
[00:07:47] Speaker A: Yeah, it doesn't exist.
All right, well, anthropic releases Claude 4. So guess what, guys? Databricks has interesting new Claude 4 Opus and Silent models available to you directly in databricks.
[00:08:00] Speaker C: I look forward to this being announced in every cloud provider for the rest of the show.
[00:08:04] Speaker A: Yes, the new models power DataBricks Lakehouse for AI offering and customers can fine tune and deploy their customized versions of the models on databricks cloud platform. So get to training folks. You know, it's always the best part about a new LLM model. You get to upgrade your stuff every week and now have non deterministic answers. Always, always best.
[00:08:24] Speaker D: They were deterministic before, no.
[00:08:26] Speaker A: Nope. But at least like if you, you know, ask the prompt a certain way and you turn down the variability of your prompt, you had a pretty good guess that it would come back with the same answer. Now you change the whole model. That's off the table. So.
[00:08:37] Speaker D: Yeah.
[00:08:38] Speaker A: All right. OpenAI has added new tools and features to their responses API which allows developers to integrate OpenAI's LA models into their applications directly. Key new features include the web browsing tool that allows models to browse websites and extract information to answer the questions, a math tool for performing mathematical calculations and reasoning, code explanation tool that can explain code snippets in natural language, and an improved code interpreter for running code in a secure sandbox environment. These new capabilities open up powerful possibilities for developers to create more sophisticated and capable applications. Powered by the OpenAI language model. The web browsing tool in particular is a major step forward, allowing models to access and utilize information from the Internet to provide more comprehensive and up to date responses. Yeah, don't ask OpenAI what your weather is in your neighborhood because it would be wrong until this came out. These enhancements to the response API demonstrate Open API's continued innovation and leadership in the field of language AI. And there you go.
[00:09:31] Speaker C: Yeah, there's a neat. I haven't had a chance to use any of the sort of native tools, but now playing around a lot more with the Agentic and MCP servers sort of getting closer to developing in that direction.
But it's, I mean it's, it's sort of for me changing like it, it wasn't that particularly useful before. These patterns coming out right where you could update your AI responses with real data that's not part of the model. And I didn't even realize that was something I needed.
[00:10:01] Speaker D: Well, I felt like I needed it when there was like new services that came out and I wanted to like, hey, I want to write a script that hits the new PowerShell thing, but it doesn't know about it yet. That's where I feel like, like I hit the edges of AI and you know, early on in the LLMs. So now that they're kind of getting stuff in a little bit more real time, maybe not 100% or you know, from a work perspective, like, you know, it helps with that. From a personal perspective, I kind of use it as, you know, for lack of a better term, like college, in college intern, like hey, go research me the these 15 lawnmowers I was looking at the other day and tell me all the features of it and build me a matrix and kind of like let it do the initial analysis. So having it be able to search the web like for me from a personal perspective was you know, kind of key. So I very much noticed like I would change the model and the provider depending on what I was trying to do stuff.
[00:10:56] Speaker A: I can tell you all about this later today. I mean the Cloud Journey section, we're talking about my my journey in vive coding because I can tell you about this recency bias problem, an exact real world example that is kind of funny and was very annoying at the time.
So we'll get there in a little bit. All right, Moving on to cloud tools this week, Docker. I haven't seen anything come out of Docker in a long time, but they're back today to talk about Docker hardened images which are secured by default container images purpose built for modern production environments, dramatically reducing the attack surface up to 95% compared to general purpose based images.
There's several startups that actually do this as well. DHI images are curated and maintained by Docker, continuously updated to ensure near zero known CVEs while supporting popular distros like Alpine and Debian for seamless integrations. They integrate with leading security and DevOps platforms like Microsoft NGINX, GitLab, Wiz and JFrog. Wait, I've got to go back. Did they say leading security platform Microsoft? That's wrong.
[00:11:52] Speaker D: You should talk to Amazon ciso about that.
[00:11:55] Speaker C: DevOps security and DevOps platforms.
[00:11:58] Speaker A: Microsoft DevOps is pretty popular I guess. Microsoft GitHub, yeah, they bought that tool.
These Docker hardened images are solving key challenges around software integrity, attack Surface sprawl and operational overhead from constant patching by providing a minimal focus based image customization is supported without compromising the hardened foundation, allowing teams to add certificates, packages, scripts and configs tailored to your environment.
They will automatically monitor and patch critical and high severity CVS within seven days, faster than the typical industry response times. Simplifying your maintenance world. This is nice. I'm mostly glad that Docker is releasing something that is not just bloat to their desktop client.
[00:12:38] Speaker C: Yeah, I agree.
They have really struggled to find their identity in tools like these are cool. I worry about the usability of these just because they are hardened and very stripped down and they don't come with tools like Curl or any kind of package management. And so I think a lot of people's build pipelines won't work with this necessarily, but hopefully that, you know, there's, it's tunable enough where you can use these images and pull in just what you need as part of your own application builds.
[00:13:08] Speaker D: Yeah, to me it's like a base layer image, you know, the golden image of that your security department hands you and says here you go, you know, it's a patch, it's here, you know, and solves that problem that you know, you have.
For lack of a better example in the old school days, all the Windows updates installed from patch Tuesday, if you start with that one every single week, you know the seven day for highs is pretty, is much higher than what I've seen in the industry. Normally I feel like that's Companies commit to 30 days for highs and 7 days for critical. So it's nice to see that they're, they've automated the process enough that they can commit to those without a problem.
[00:13:45] Speaker A: I think Fedramp Requires faster than 30 days.
No, you fed up high. I thought it was 14. I thought was the new guidance from the Biden administration.
[00:13:55] Speaker C: I don't, I don't think the response is due in 14 but I mean there's, there's all kinds of different layers depending on the severity and all kinds of other stuff too. So yeah, I could be a little wrong but I don't think there's, I think 30 days was the high severity one if I remember correctly.
[00:14:09] Speaker D: Yeah, normally it's seven at least from what I've seen. Seven for high 30 if you get below or if you hit 30. Most, you know, as a SaaS provider, most of your clients are good with that.
Obviously if it's you know, more critical or if it's a public facing server, obviously you unpatch that quicker.
[00:14:28] Speaker C: But I mean it's, you know, the reality is like you know the seven days which is, you know what they dictate for zero day type of responses like you're okay, sure put a number there. But the reality is you're completely beholden to the upstream providers and a company, all they can do is promise to deploy something as soon as they can.
[00:14:51] Speaker D: I mean most of the time it's within seven days of a patch being provided. If you like look at stuff, not when the vulnerability is found.
So in theory if you have a log 4J vulnerability and takes your vendor two weeks to get it in, it's you know, you have seven days from the time the patches obviously you should do whatever you could to remediate the issue but that's where the seven days is. It's from the patch availability.
[00:15:18] Speaker C: Got it.
[00:15:18] Speaker A: Okay. Well again glad to see it. Hope they continue to do something like this, but they're definitely competing with a couple startups. So I wonder if this is part of a business pivot like we can make more money if we made hardened images.
All right, let's move on to. Hashicorp is releasing Terraform McP server for AI integration.
HashiCorp released the open source Terraform MCV server to improve How AI models interact with infrastructure as code. Providing real time structured data from the Terraform registry, the server exposes module metadata, provider schemas and resource definitions in a machine readable format allowing AI systems generate more accurate context or terraform code suggestions. By leveraging the MCP protocol, the server enables AI models to retrieve up to date configuration details and align with the latest Terraform standards, reducing reliance on potentially outdated training data. The Terraform MCP server has been demonstrated with GitHub copilot integration allowing developers to access context aware recommendations directly from their ide.
The release is part of a broader trend in AI assisted tooling to unify developers workflows through interoperable interfaces as well as exposing you to all kinds of security risks. So please be careful with your MCP use cases.
[00:16:19] Speaker C: Yeah, I mean this is, I was trying to figure out how this would be used and so looking into it a little bit, it's not exactly what I thought which is it's not going to really allow interaction with my state files.
It's really good for sort of leveraging up to date information for like the providers and the Terraform public registries and that. So if you're, you know, if you're doing Terraform against code or doing Terraform against something that was released very recently, this was, you know, it's not going to be something that's known in the model. It's. You can leverage the MPC service and it will have the information about the new resource or what have you.
[00:16:57] Speaker A: So it's kind of neat.
[00:16:59] Speaker D: Yeah, I also read a little bit of how they were implementing it like with the Terraform server, with your corporate registry modules. So if you do have a platform engineering team that kind of has these modules predefined for you, it kind of will interact with those in that way which goes into what Brian's saying where like in real time we'll pull and say okay now you need these variables with your, with your VS code or whatever your IDE is. So kind of that registry piece of it I think to me is the key part that you can kind of leverage that for a lot of your, you know, different teams to use, you know, along with any of your other internal communication that you're fighting with people on.
[00:17:39] Speaker C: Well and yeah, just the, the ability to say, you know, in natural language what you want and, and to have it pull like those custom resources like you were, you know, from internal registry. It's pretty, pretty sweet.
[00:17:50] Speaker A: All right, moving on to aws, the Aurora D SQL capability announced at RE invent is now generally available. This is a serverless distributed SQL database that offers unlimited scale, high availability and zero infrastructure management. It simplifies complex relational database challenges and the disaggregated architecture enables multi region strong consistency with low latency. Designed for 99.99% availability in a single region and 99.999 across multiple regions. I think I said five nines, I lost track of the nines.
It integrates with AWS services like AWS Backup for snapshots and restores AWS private link for private connectivity and CloudFormation for resource management and CloudTrail for logging. The MCP model Context Protocol Server improves developer productivity by allowing generative AI models to interact directly with the database using natural language via the Amazon Q Developer CLI and key use cases for DSQL could be microservices, event driven architecture, multi tenant SaaS, apps, data driven services like payment processing, gaming, social media require multi region scalability and resilience. The pricing starts at $0, which is only until you exceed the free tier with 100,000 DPUs and 1 gig of storage per month. Then based on distributed processing units and gigabytes per month, the specific rates were $8 per 1 million GPUs and then I think it was 32 cents per gigabyte per month. So if you need a highly distributed, highly scalable, available and manageable database solution, Aurora D SQL may be the right answer for you.
[00:19:13] Speaker D: The pricing of it's kind of going in line with like the Azure pricing and I feel like a lot of the other RDS type pricing where the compute is getting on like the low end is getting lower but your storage costs are getting higher, 33 cents a gigabyte, that starts to add up quickly. So it's just something to be careful of, especially if you have multiple read replicas in multiple regions.
Your 100 gigabyte server is, you know, know, times multiple adds up very quickly. So if you are doing this on the low end, you know, and just you know, in your own playground environment, just be careful of that.
[00:19:47] Speaker C: I mean it makes sense because you know I, I think that would be the normal cost drivers for a serverless solution and anything where you're doing a lot of high transaction rates and and like very, you know, performance conservative sort of things applications, you wouldn't probably use serverless technology. So you know like I kind of like this because most of my stuff is smaller and I'm not doing you know, billions of transactions a second kind of thing. And so the ability to scale down to zero and just pay for the data that's at storage versus, you know, having it live and accessible. Well, it is still live and accessible, but I'm not paying through the nose for that, which is cool.
[00:20:25] Speaker A: The interesting does not support MySQL it's just Postgres today. So I'm hoping for MySQL version in the future because I'd like to build. Use this for other purposes, but RDS is pretty expensive, so without having my SQL and you know, the ability to scale down to just, you know, serverless capabilities. It's not really an option yet, but maybe, maybe in the future we'll see.
[00:20:47] Speaker D: I feel like for years my SQL was first like with rds and then.
[00:20:50] Speaker A: It was for a lot of time. Yeah, yeah.
[00:20:52] Speaker D: And then I feel like it pivoted.
[00:20:54] Speaker C: To postgres, which is confusing because I thought postgres was like a new kid on the block, and that's why. And it's not. It's not new at all.
[00:21:00] Speaker D: No. Yeah. And that's why I'm like. It's interesting to me that this one GA'd with Postgres and not MySQL I.
[00:21:08] Speaker C: Mean, I think Postgres has more market share, at least right now, whereas I think it used to. I think it was MySQL in the early 2000s. I mean, that's my impression. I don't have any data to back that up. That's what it feels like to me.
Everything's Postgres these days.
[00:21:26] Speaker D: Yeah.
Or Microsoft SQL, if you really love that.
[00:21:29] Speaker C: If you choose poorly in life, which we all have.
[00:21:32] Speaker D: Yeah.
Hey, as long as you're not an oracle.
[00:21:37] Speaker C: Yeah, it's true. We could be worse. It could be worse.
[00:21:40] Speaker A: Could always be worse.
All right, let's move on here to our next show topic. Amazon ECS has increased the character limit for container exit reason messages from 255 to 1024 characters. So no longer will you waste most of the characters with exit error code 99 and then have the rest of the error message be cut off. So you don't know why you got error 99, which is basically container stopped. This will provide you that more detailed error message to help you debug your failed container more effectively. I'm surprised it took this long to get, but appreciate that it's finally here. You can access these extended error message via the Amazon Manager console as well as the Describe Tasks API via the CLI or directly from the API.
This feature is available in all AWS regions for ECS tasks running on Fargate platform 1.40 or EC2 container instances for ECS agent 1.92 and above. This is a great use case for debugging your container failures, which I deal.
[00:22:35] Speaker C: With all the time. Yeah, I think you could. If you were hosting your own compute layer for ecs, I think you could get to these before. But if you were on Fargate, you were completely screwed. You were never going to get that data.
[00:22:48] Speaker D: No. You get to the. The base logs on the system, it was just what shows in the console.
[00:22:53] Speaker C: Yeah.
So this is. I can't imagine that world. Right. So like when I was using acs, it was. I had access to that compute layer. So this has got to be game changing for people.
[00:23:06] Speaker D: I feel like Fargate 1.4 came out a long time ago. Is it still the latest version?
[00:23:14] Speaker A: I don't know if it's the latest version or not.
[00:23:16] Speaker D: I don't use Fargate 2020 came out.
[00:23:19] Speaker C: Wow, they've achieved nirvana. They don't need to implement.
What more do you need? Does it host your containers? Yes, it does.
[00:23:29] Speaker D: Apparently the latest version is still 1.4.
[00:23:31] Speaker C: I mean all the logic is that the ECS scheduler and stuff, that's kind of above the Fargate platform.
[00:23:38] Speaker D: Yeah. I'm just still surprised after five years there hasn't been a new revision of it.
Because I remember the first couple I felt like came out rather quickly, which I guess makes sense with the new platform. But you're right, they hit the Nervada moment.
[00:23:55] Speaker A: All right. Moving on to Dynamodb Local is now accessible via the AWS Cloud Shell Downtown available in Cloud Shell, allowing developers to test DynamoDB applications directly in the AWS Manager console without incurring costs. It integrates with the existing DynamoDB API to enable local development and testing without impacting your production environment. Environment Developers can start the DynamoDB local in Cloud Shell, use the DynamoDB local alias without needing to download. To interact with the local DynamoDB instance in Cloud Shell, use the endpoint URL localhost8000 my favorite simplifies Dynamodb development workflow by providing the pre authenticated browser based shell environment accessible anywhere in the console. This also follows a thing we didn't mention last week was that now you can do DynamoDB local development as well on your laptop if you don't want to use the AWS shell. This is not an area that I am very familiar with the Cloud shell just because it came after I learned AWS and I'm an old dog who can't learn new.
Here we are.
[00:24:46] Speaker C: Yeah, I haven't used Cloud shell a whole lot and I imagine that, you know, like DynamoDB local I wish existed when I was doing more AWS work because this is it. That was always a pain point.
You could get by a little bit with like using SAM and other serverless frameworks. But this is definitely DynamoDB is a, you know, I like it but it is difficult and it does pretty require some maneuvering when you're developing an application utilizing it. So this is, this is handy. And then presumably I guess, I mean I've always used like cloud shells for like very simple CLA cloud tasks. I've never really thought about developing inside of cloud shell, but I guess people do it.
[00:25:30] Speaker D: Yeah, I think it's they're pivoting it to have the development futures probably because of the AWS Cloud 9 deprecation.
You know, they feel like maybe they're trying to get the features over to Cloud shell so they tell people, hey, move from here to here and kind of do it that way like so they have a way for people to continue leveraging the platform. But I'm like you where I'm like, I use it if I need to run like four commands or like some reason I can't do something locally or I'm fighting with logging into multiple accounts and assuming roles and a mad at life. I'm like, oh yeah, this really cool thing that's going to make my life easier still exists.
[00:26:10] Speaker A: I think the idea of it always makes sense until you're trying to do anything at scale or anybody else and then suddenly realize your database isn't as portable as you'd like it to be.
[00:26:21] Speaker D: Yeah, all right.
[00:26:22] Speaker A: It was announcing IPv6 support for EC2 public DNS names.
EC2 DNA DNS names now can automatically resolve to IPv6 addresses for EC2 instances and elastic network interfaces, allowing public access to IPv6 enabled instances over IPv6 previously ect DNS only resolved IPv4, which of course you had to pay for, which is annoying. Requiring use of specific IPv6 addresses or custom domains via Route 53 to access the IPv6 only instances. This enables easier access to IPv6 only and is available in all commercial and Gov cloud regions and configurable via VPC settings as just like your EFB4 settings are set.
[00:26:59] Speaker C: So.
[00:27:00] Speaker A: Thanks, you're long enough.
[00:27:03] Speaker C: This is probably. You know, I was just thinking as you're reading this, it's probably the only thing that's going to make us adopt IPv6 is by like you know, these cloud providers basically making the Internet, just making it default. Yeah, well, I mean just kind of doing it for you in a sense. Right. You haven't had the ability to do it at that load balancer level. So if you're hosting your application on the cloud, you can't really support IPv6 end to end.
And so this is, this feels like it's going to move a lot in that direction and I wonder if we'll see that really expand into like sort of the consumer cloud space versus the sort of telecom space where we see a lot more IPv6 and that hopefully never learn how to translate the subnet math.
[00:27:46] Speaker D: But yeah, no, I'm good. Yeah, I got enough problems where I get to like 20 ones and whatever I'm like, wait, where what?
[00:27:53] Speaker C: I can't. Okay. I've never been able to do it with IPv4 because I've always relied on calculators.
[00:27:57] Speaker D: So yeah, I mean, I think what you're really going to see is, you know, some fit ops person going to be like, let's just move off IPv4 to save $4 million a year on AWS IPv4 pricing, you know, but then.
[00:28:12] Speaker C: You'Ll find that one customer be like the driver, that one customer that's on some weird Internet service provider that cannot reach you. That's how you find it.
[00:28:21] Speaker D: I mean, I've had the opposite problem. I mean granted this was years ago, but when you had the dual stack albs where at one point I had a customer where I had to disable the IPv6 in order to make it so that it would work because the customer's ISP at home wasn't working for the OpenVPN and going routing through that whole way.
So we had to actually disable it in order to make it work.
[00:28:50] Speaker A: Interesting.
Well, if you've you run a lot of EKS clusters, you might have wanted to have them under a centralized dashboard and now you can. Thanks Amazon. EKS Dashboard provides a centralized view of Kubernetes clusters across AWS regions and accounts, making it easier to track inventory, assess compliance and plan your operational activities. It integrates natively into AWS console, limiting the need for third party tools and their associated complexity and cost.
The dashboard offers insights into clusters, managed node groups and EKs add ons with data on cluster distribution, version support, status forecasted costs and health metrics. And advanced filtering enables drilling down to specific data points to quickly identify clusters needing attention.
Setup for this is pretty straightforward using AWS organizations management and delegated administrator accounts enabling trust access in the EKS console. EKS Dashboard supports visibility Kubernetes clusters running on premises or on other clouds, though with more limited data compared to native EKs.
This feature will especially benefit organizations running Q&As at scale across multiple regions, accounts and environments who need unified visibility and control.
And this is at no charge.
[00:29:50] Speaker C: AKA you have a centralized team that you've shafted into hosting all the Kubernetes workloads and being the subject matter experts because there's no way that you can segregate that and decentralize it. And so at least we're making that those poor bastards life easier. So I like this except for the need for it. I don't like.
[00:30:11] Speaker D: This was one of those things that I think I talked with AWS about like four years ago and I'm happy to see that it actually was finally released. It was you know, when they were discussing like future plans of EKs, you know and under some NDA that I had they were like, they were like they were looking into doing this and what would you think and how would you use it? And I was like sounds good. I would hate to be the person that has to use it but it sounds like a really good feature and it's always interesting to see when they finally ga something from your PFR from years ago and how long it takes. And I'm like oh wow, I forgot about this.
[00:30:51] Speaker C: I mean this has been a differentiator with GKE on Google for a while because their Kubernetes management is very seamless and they don't have the regional constructs built into their console and.
And so just by natively in Google you get these kind of centralized view where you can see all the clusters that you have. But it's. I mean still doesn't make managing Kubernetes easier.
[00:31:14] Speaker D: I do want to point out the quick tour of EKS dashboard makes me really just think of EKS in general clusters with health issues. 688 I'm like that sounds like Kubernetes. Yep.
[00:31:29] Speaker A: Clusters with upgraded 100% clusters to be.
[00:31:32] Speaker D: Auto upgraded in 90 days. 5118 Kubernetes is so noisy.
[00:31:38] Speaker C: Yeah doing a lot of logging and routing of logging these days.
So getting back in the dirty logging areas which you know I hear there's.
[00:31:48] Speaker D: A really good tool that you should look at.
[00:31:50] Speaker C: I will murder you. Shut your mouth.
[00:31:54] Speaker D: Good thing I'm on the opposite coast. But I'm pretty sure you will reach through the computer screen somehow.
[00:32:00] Speaker C: Yeah we don't talk about those days.
But yeah, so it's just funny because I'm just like looking through all these log buckets and be like, why is this one so much? Oh, kubernetes. Yeah. Okay, cool.
[00:32:14] Speaker D: Yeah, everything gets logged. I mean, in some ways it's good. Just find the needle in the haystack at one point. Yeah, good luck, sir.
[00:32:22] Speaker A: Or just use ECS because you're already there.
[00:32:26] Speaker D: I'm curious if this will eventually support. Like I don't know if it does because don't honestly care enough. But if it does support like EKs anywhere and then over time they make it be cloud agnostic, you know, and pull in your Google cluster and everything into one place. It'd be an interesting like future state.
[00:32:43] Speaker C: For this might just. I don't know, but it looks like, you know, like EKs anywhere the way that it's instrumented. Like it might just natively fit in there because it's just a EKS cluster and really all they're doing is the multi account orchestration layer at organizations.
So this might natively just support the EKs anywhere because those are just workloads deployed in an Amazon account, even though COMPUTE is running in your data center and so they're probably included.
[00:33:13] Speaker D: So the question does it work multi cloud or not? But future problems depends on probably a container running.
[00:33:20] Speaker C: If you're just hosting VMs in one cloud running the EKS agent. That's kind of the same, right?
[00:33:28] Speaker D: Yeah, though it's on. Good news. It's only available in US Yellow 1, in case you were worried.
[00:33:35] Speaker C: That's interesting.
[00:33:36] Speaker D: Which is interesting that they still chose Virginia to do it.
[00:33:39] Speaker A: Of course it is.
And good news, Cloud 4 is also available to you on Bedrock Riot.
[00:33:45] Speaker C: How did you.
I was concerned. Yeah, I'm glad.
[00:33:48] Speaker A: I know you were. I know. As move on to gcp, Vertex AI Studio is getting into its dark era.
Vertex ASU of course provides a unified platform to experiment with and customize 200/ advanced foundational models like Google Gemini and partners like Metalama or Anthropic Claude, which guess what, in a little bit we're going to talk about them getting Cloud 4 too.
Redesign focus on developer experience with faster prompting, easier ways to build and fresh ui, accelerating prototyping and pairing with generative AI models. Vertex AI Studio integrates end to end workflow from prompting to grounding, tuning, co generation, test deployments.
This also enhances prompt engineering and management functions and testing as multiple models. Basically the change for this is introducing a dark mode UI for Better visual comfort during long development sessions.
This gives you a much better vertex AI user experience because the white console of the Google console is blinding.
[00:34:39] Speaker C: It really is.
[00:34:41] Speaker A: So I appreciate this one. Thank you Google Dark mode for life.
[00:34:45] Speaker C: Yeah, I'm there for it. You know, I also feel a little dirty because it just feels like, you know, as developer community, it's just pretty princesses like I need dark mode or I can't see. But it's also. I love it so much.
[00:34:59] Speaker A: All right, that didn't get you excited. I have a new Gemma model for you. Gemma3n is a powerful efficient mobile first AI model optimized run directly on phones, tablets and laptops. It seems to be the trend these days. Enables real time multimodal AI experiences with advanced on device capabilities. Model leverages a new shared architecture co developed with mobile hardware leaders Qualcomm, Mediatek and Samsung. This positions it well versus other mobile AI offerings from potentially Apple, which has failing terribly at Apple Intelligence. Gemin3n uses an innovative technique called per layer embeddings or PLE to significantly reduce RAM usage, allowing larger models run on mobile with 2 to 3 gigs memory footprints. Integrates closely with Google's broader AI ecosystem, powering the next generation of on device features like Gemini Nano and Google Apps. And developers can preview core capabilities that come with Android and Chrome. Key use cases include real time speech, transcription and translation, voice interactions and multimodal understanding of combining audio, image, video and text inputs, all processed privately on the device.
Is available to you now in preview with progressive ability to follow later. And no pricing yet, but I assume it'll be free for a period of time or forever, who knows?
[00:36:02] Speaker C: Yeah, I mean with as. As pricing with generative AI goes, it's. You never know what you're going to get.
[00:36:11] Speaker D: Even once they give you the pricing, you still don't know.
[00:36:13] Speaker C: Yeah, I mean it's neat to see that they're developing models to be directly right on the edge. Like I, you know, you know, it's kind of. It feels like early web days were, you know, chasing down performance gains wherever you get it.
And this feels very much in line with that. So it's neat.
[00:36:29] Speaker D: I mean that makes sense. You know, these things run a massive compute and eventually if you can get on your phone offline, still be asking general questions, you know, or on low bandwidth, it's going to help to have that computer on the edge. And then, you know, imagine millions of edge computing devices like laptops, you know, at a company. Right. If you can run kind of like the Old school folding at home or any of those on all your company laptops.
On the, you know, AI PCs that you have, you can start to really do a lot with them.
[00:37:03] Speaker A: All right, moving on to Google's announcing major updates to its intelligent agent platform, providing more robust development tools Intuitive manager and seamless agent to agent communication. The Agent Development Kit ADK adds new capabilities to create sophisticated agents with greater stability and adaptability.
The basically A2A is the leading industry adoption for platforms, introducing new capabilities for building, deploying and securing them. And the updates provide a comprehensive flexible platform for building intelligent agent solutions, unlocking new possibilities across the agent industry.
Those improvements include the Python ADK is now v1 for stability for production ready agents and the Java ADK is now at 0.1 extending agent kept layers to the Java ecosystem which is not too bad.
[00:37:46] Speaker C: Yeah, it is kind of neat. Like I've been playing around with this a little bit thinking about, you know, agents being sort of like independent, like sort of components, like logic codes, what like, you know, almost thinking of it like as a Python class or a microservice. And so once you start sort of developing in that model, like having a boatload of agents gets difficult to manage. So this is a way that you can sort of manage them all in code. And so this is kind of nice. You can version it out and have deployment models that are a lot more reliable and a lot more, you know, maintainable in the long run.
So this is kind of cool. I look forward to more improvements in the space.
[00:38:26] Speaker A: Yeah, me too.
You know, the agent to agent stuff is really cool. The biggest thing that you get though is security.
That's the biggest risk we've seen so.
[00:38:34] Speaker D: Far in quite a few of these authenticated connections. How you manage it, it's gonna be the wild west.
[00:38:42] Speaker A: Well, and data leakage the models is a big problem with the MCP stuff. So there's a lot of, lot of dangers of MCP you should definitely be a little cautious about.
[00:38:50] Speaker C: And there's not a lot of the wild west. Yeah, there's not a lot embedded in the platforms regard Rails are very far and few between right now on all these platforms.
[00:39:00] Speaker A: Yeah. So I will see a lot of that improving. Oh yeah here soon.
[00:39:04] Speaker D: Don't worry, we'll end up with like you know, HTTP and HTTPs. We'll end up with MCP and mcps, you know, over time.
[00:39:14] Speaker A: I mentioned it earlier, I already said this one but I'll say it again. Anthropis Cloud Opus 4 and Cloud Sonnet 4 on Vertex AI.
[00:39:20] Speaker C: You're welcome.
[00:39:21] Speaker D: My life is now complete. Thank you all.
[00:39:23] Speaker A: You're welcome.
Google's advancing sovereignty, choice and security in the cloud with their updates to its sovereign cloud solutions.
Three new offerings are available in their sovereign fleet, including Google Cloud Data Boundary which allows deploying sovereign data boundaries to control data storage, processing location and manage encryption keys externally. The Google Cloud dedicated, designed to meet local sovereignty requirements through partnerships like Thales S3NS in France and Google Cloud Air Gapped. A fully standalone solution not requiring external network connectivity. Tailored for intelligence and defense sectors. These solutions leverage Google's massive global infrastructure of 42/regions and 202 educations and key partnerships across the those regions. This enables customers to choose solutions aligning with business needs, regulations and risk profiles. Not a one size fits all approach.
Again, this combines your local control with access to Google's leading security like AI powered defenses, confidential computing and post quantum crypto. So if you are in the data sovereignty place and you need to address your growing data residency concerns without sacrificing your cloud benefits, Google's got your back this week.
[00:40:22] Speaker C: Yeah, it's kind of nice the way that Google does this versus AWS. Right. Like AWS has the GovCloud and it's almost like a separate product now whole separate authentication, whereas these are built in.
[00:40:33] Speaker D: Is a separate product.
[00:40:33] Speaker C: It is, yeah. And so but you've not dealt with.
[00:40:36] Speaker D: Govcloud where you have to like set up another account and each govcloud account is tied to your corporate.
[00:40:42] Speaker C: It sounds terrible, but I have done it in Google and I didn't have to do any of that. Sounds great.
[00:40:48] Speaker D: Yeah. So every GovCloud account has associated commercial account with it and you create it with the GovCloud Create API which pretty much fires and you get no information back. And then like 30 to 45 minutes later it magically creates you a decode.
[00:41:04] Speaker A: Nice.
[00:41:05] Speaker D: It's. It, it's a disaster.
[00:41:07] Speaker A: Yeah, it sounds magical.
[00:41:09] Speaker D: Yeah. It's not magical.
[00:41:11] Speaker A: Well, and I think I heard about GovCloud is that you have to have a commercial cloud to even create the govcloud, which kind of defeats the whole purpose. So you always have to have this commercial entity hanging off of your federal entity.
Right?
[00:41:22] Speaker D: Well, that's what it is because all the billing is done through the commercial entity. So there's a one to one ratio of commercial to GovCloud and it's not like your, your account names I think match. But then like everything is different. Like you're. Because the arms are all different. Because it's like aws dash Gov, you know, rn, aws. And it's like gov, it's like all your terraform. Especially if you're doing any like, anything, you have to like use different data lookups. It's a load of fun.
[00:41:50] Speaker A: I'm not sensing the fun in your voice. I'm sensing, oh, I could go on.
[00:41:54] Speaker D: For a long time about this. Random APIs don't work because they haven't been approved to work there, but they don't tell you that till you open a support together.
[00:42:00] Speaker A: I can tell you that happens on Google too.
[00:42:02] Speaker C: It absolutely does. Their enforcement, I mean, any kind of like enforcement at a platform layer is always going to leave the, you know, sort of the feedback layer for the end user a little bit left in the dark because it's sort of like, why did that fail? And then, you know, it's got this cryptic error message. And then, you know, you talk to your cloud team, they're like, oh yeah, that's because it's, you know, defined in the data boundary or defined in the organizational policy that is only allowing certain things.
Yeah.
[00:42:29] Speaker D: Okay, so Microsoft just gives you internal server error. Then when you ask Copilot, which is now embedded in the console, obviously how to fix it, it just says go as Microsoft support. So it's a load of fun.
I work on the fun clouds, GovCloud, Microsoft Cloud. Next up is Oracle Cloud. I think maybe in my life.
[00:42:48] Speaker C: Oh, man. Oh, don't do this to yourself.
[00:42:51] Speaker D: Like, no, no, I'm not planning on it. I think that might be a hard limit. No good.
[00:42:56] Speaker A: Yeah, I mean, Azure might be a new hard limit for you too after this experience.
[00:43:01] Speaker D: It's up there.
I understand the sharp edges and I walk into them every day.
Yeah, by opening my email.
[00:43:10] Speaker A: Well, if you have unstructured data that you want to add to a BigQuery table, you can now convert AI generated unstructured data to that BigQuery table with this new BigQuery feature that converts unstructured data, images and text into structured tables. Using advanced AI models like Gemini 2.5 Pro and Flash, it builds upon ML generate text to streamline the process of extracting insights and making unstructured data compatible with existing data analysis workflows. And while AWS and Azure offer some AI services for unstructured data, the tight integration between BigQuery and Vertex AI and the ability to directly generate structured tables sets GCP far apart. And if I knew how to work any of that work, I would be excited about this.
[00:43:44] Speaker C: I suppose so. I don't know what The AWS mature sort of offerings that they're referencing there. But like the, the ability to sort of take a bucket of unstructured data and then have this. It's a. It's effectively data labeling AI data labeling of your images and your, your unstructured data and then populating that metadata into bigquery tables which is pretty rad, right? Like the this, you know, there's only a few years ago where you had armies of people basically doing the same thing by looking at a like opening of the text image, writing down some identifying factors about it and then you know, putting that into an application that saved that into a database. So this is pretty neat to have that all run on the background and just enter it, you know, then you have a whole bunch of things that you can do a query search and you could pull up that data much easier and kind of neat. I like it.
[00:44:40] Speaker D: Yes.
[00:44:42] Speaker C: Give me all the pictures with cats in them. That's all we, that's all we need. Okay. Yeah.
[00:44:48] Speaker A: Google is now offering you Mistrals AI Leshat Enterprise. An AI assistant for enterprise search, custom agents, document libraries and more all available via the Google Cloud Marketplace allows you to build custom AI agents without code MIS OCR 25.05 is a powerful optical character recognition model for document understanding. Is now available as part of the managed service and compared to other cloud AI platforms. Google Cloud offers an open and flexible ecosystem to build custom AI solutions by integrating pre trained models like Mistral's Leshat Enterprise. Art leverages Google Cloud secure Scalable intra integrates the services like BigQuery and Cloud SQL. And Mistral's OCR is one of the 200 plus foundational models in the Vertex AI model garden. Key use cases for Google for this one is research analysis generating insights from data code development, content creation with lechat, digitizing scientific papers, historical documents, customer service docs, all with ocr. Industries that benefit include finance, marketing, research institutions, customer service, engineering, legal and more. And the pricing on this is usage based by document page count.
[00:45:47] Speaker D: The concept of the paperless corporate environment is still not here and this proves it that we still have to OCR stuff.
[00:45:55] Speaker A: Right?
[00:45:56] Speaker D: We should have done a show title with that.
[00:46:00] Speaker C: Yeah, I mean you still have to, you know, OCR even if it's not paperless, it's just PDFs. It's that unstructured data. And the challenge that comes, you know, with that it is sort of.
I feel like I don't understand what the products are in here because I get it to Google Marketplace. Offering, but I'm not familiar with lechat Enterprise to know if that's a standalone product or not.
But yeah, I mean solutions like this, you know, I think we'll see more and more. I think there's already a bunch out there. I think this is comparative to like Google's Asian space that they announced at Google Next and you know, like having having the ability to sort of take all your corporate information and make it usable is, you know, a pangea that everyone sort of still needs. Even with solutions like this, it's they're difficult to implement and and hard to get. So hopefully it's making improvements in that direction because sweet Jesus, if I have to search the wiki one more time with its native search integration, I've got to throw my computer through a wall.
[00:47:00] Speaker D: You just want a new computer.
[00:47:02] Speaker C: 2 I really want a new computer.
[00:47:04] Speaker A: You should just use elasticsearch.
[00:47:08] Speaker C: I can drive to you Justin.
[00:47:14] Speaker A: True.
[00:47:15] Speaker D: This episode you get right on a watch list somewhere.
Sure, we don't know why you can't board the plane, but it says you're on the no fly list.
[00:47:29] Speaker B: There are a lot of cloud cost management tools out there, but only Archera provides cloud commitment insurance. It sounds fancy, but it's really simple. Archera gives you the cost savings of a one or three year AWS savings plan with a commitment as short as 30 days.
If you don't use all the cloud resources you've committed to, they will literally put the money back in your bank account to cover the difference. Other cost management tools may say they offer commitment insurance, but remember to ask will you actually give me my money back? Achero will click the link in the show notes to check them out on the AWS Marketplace.
[00:48:08] Speaker A: All right, moving on to Azure.
Azure FX V2 series are now powered by the 5th generation Intel Xeon Platinum processor are all generally available for compute intensive workloads like databases, analytics and eda. Integrates with Azure Boost for improved networking, storage, CPU performance and security and supports all Azure remote disk types including premium SSD v2 and those ultra disks offered a 50% better CPU performance versus previous FX v1 series with up to 96 VCPUs, 1,832 gigabytes of memory and enhanced AI capabilities with the Intel AMX capabilities competes favorably with similar compute optimizations from AWS, the C6i and the GCP C2 with a higher core count and memory. The targets customer is running SQL Server, Oracle databases, supply chain solutions and mission critical apps running high Iops and low latency.
[00:48:55] Speaker C: Cool.
[00:48:57] Speaker D: Hey. Yep Premium and ultra disk. How expensive can that be?
[00:49:01] Speaker A: You have access to them, you know, you don't have to use them.
[00:49:03] Speaker D: I'm in a sassy mood today. I apologize.
We run a few premium discs in my day job and you see that on the bill. They stick out your sore thumb.
So, like, you can. You gotta be careful with some of those things. But, you know, it's nice to see that they are continuing to improve. And if you really need 100.
Sorry, 1832 gigabytes of memory, you're doing.
[00:49:25] Speaker A: Your job wrong or you're doing it really right. One of the two.
All right. Red Hat OpenShift virtualization on Azure. Red Hat OpenShift is now available in public preview. This came up at the Red Hat Summit last week. Unifies management of VMS containers on a single platform, allowing organizations to modernize at their own pace while leveraging existing VM investments. Integrates with Azure services like Azure Hybrid Benefit for cost savings, Azure Security Tools and for enhanced protection and Azure. Red Hat OpenShift for managed OpenShift platform utilizes the KVM hypervisor and Red Hat Enterprise Linux for improved virtualization performance and security. Because it's all virtualization all the way down. And a differentiation from AWS and GCP by offering a fully managed, jointly engineered Red Hat OpenShift platform with native virtualization capabilities. Meaning you will never get this from anybody but Red Hat on Azure.
[00:50:12] Speaker C: Yeah.
[00:50:13] Speaker A: Similar to Redis. You're welcome, Matt.
[00:50:15] Speaker D: You people don't love me today. Yeah, that's what I'm learning.
[00:50:19] Speaker C: I mean, I'm glad that they still don't really know what OpenShift is for because they continue to sort of lose the message of this is a CICD platform for containers and now it's doing VMS like. Like it's VMware.
It sounds like they're just completely pivoting to deal with the Broadcom changes that are there. Capture that market. All right, Go nuts. Red Hat. You do you.
[00:50:41] Speaker A: You do you.
[00:50:42] Speaker D: 17 cents an hour, though. That's going to add up quickly.
[00:50:46] Speaker A: Yeah. It will get real pricey per worker. For every four CPU workers.
[00:50:50] Speaker D: Right.
CPU worker nodes. So you need at least two for ha.
Or three, depending on how much.
[00:50:58] Speaker A: Well, I thought it was for Every. For every four VCPU of worker nodes. If you have two worker nodes with four CPUs each, they would require two $0.17 per hour charges.
[00:51:08] Speaker D: Yeah, yeah.
[00:51:09] Speaker A: And then if you multiply it from there.
[00:51:10] Speaker D: Yeah. It's not going to be cheap.
[00:51:12] Speaker C: I mean, that can't be the workloads.
[00:51:14] Speaker A: I mean, OpenShift in general is not super cheap.
[00:51:16] Speaker C: Yeah, it can't be the hosted CPU like, because that would be a ridiculous price if you're hosting all your application cpu use at 17 cents an hour.
[00:51:26] Speaker D: I mean, it's the same thing as like they're trying to build a path to help people move. It's the whole VMware on AWS or on Azure.
[00:51:32] Speaker C: Right.
[00:51:33] Speaker D: Put your nodes here, let you expand here and then you know, if you need, great, modernize your application at your own will, you know, or forget and move there.
[00:51:42] Speaker C: I just hope the worker nodes are for orchestration, not for like. Because if that pricing is hosting the workload, then like you'll never, they'll never succeed. But if that's for the orchestration that's built into OpenShift, then okay, then it's just expensive.
[00:51:57] Speaker D: I think it's both. I mean you got to run the, the worker, the work, the end node somewhere that has the, the workload on it.
[00:52:04] Speaker C: Oh, sure, sorry.
[00:52:04] Speaker D: The word workload was really hard to.
[00:52:06] Speaker C: Get out of that right then I just.
[00:52:07] Speaker D: But you know. Yeah, and then in theory it should be the orchestration of it.
[00:52:11] Speaker C: Yeah, I know. I figure you're still paying for that. Yeah, yeah. Crazy.
[00:52:15] Speaker A: All right, moving on to key maintenance experience enhancements for Azure database for MySQL this provides you more control, visibility and predictability over how maintenance is orchestrated across Azure databases. For MySQL this includes the virtual canary and generally available allowing enrolling specific servers into an early maintenance ring to validate updates before broader rollout simplifying detecting potential compatibility issues. Earlier maintenance batches allowing you to explicitly assign servers to different execution batches within the same maintenance window. Ensuring maintenance proceeds in a predictable user defined order or not all your nodes are being maintained at the same time and then maintenance rollout status check preview centralized view of maintenance activities across servers Monitor roll out progress identify anomalies from the Azure Portal or programmally via Azure resource graph. This improves transparency, reliability and alignment with enterprise deployment strategies for Azure database for MySQL maintenance and targets customers running development workloads or managing complex multi environment MySQL rollouts on Azure.
[00:53:09] Speaker D: And to answer your pre read question, this is better than what you get with Microsoft Manage SQL on there.
[00:53:17] Speaker A: So sad you can do maintenance Windows.
[00:53:20] Speaker D: I mean hopefully I'm wrong but like I don't know this feature in Hyperscale or any of the other Microsoft SQL platforms. Whether you're on, you know, Vcore or you know, GGU or any of Them business critical, but like you can control. And this is kind of cool if like your dev environment, you put in the early adopter ring and get that really rolled out early and make sure nothing crashes it, you know, and then let it auto upgrade the other environments afterwards. Like that's kind of nice to say, you know, even on rds, you can't say development environments roll out one week or two weeks before. You can just give it the date and time, like every Saturday from 10 to 2. But you can't say this one. This gets patched first, these ones get patched second.
This is actually, in theory letting you control, hey, these development environments roll out first and production roll out after.
It's a decently nice feature. It's just amazing. I came to MySQL first, so clearly their team is more advanced than the SQL team. I'm really getting in trouble by the.
[00:54:22] Speaker A: End of this episode.
[00:54:23] Speaker C: I think management of MySQL in general is from that hosting layer. It's just easier than Microsoft SQL Server.
But yeah, I would love to see this be a pattern that grows. Right, because you've been able to do this sort of on your own by like cobbling together maintenance windows the way you want and sort of managing those maintenance windows to sort of manage that.
Like making sure that your dev maintenance windows, you know, sooner than your PROD one and being able to sort of main or like sort of monitor that workflow and then put a stop to it if you have to. But.
[00:54:57] Speaker D: But you couldn't really. All you could do was days of the week. So you could say like once on Monday, once on Friday, but if they release a patch on Wednesday, you were sol.
[00:55:05] Speaker C: Oh, in Azure, I can pick a monthly, even aws. No, yeah, you're right. In aws, you couldn't. In rds, you're absolutely right.
[00:55:14] Speaker D: Yeah, you can just choose day of week and time. So, you know, and even on Azure you get less options. I feel like on aws, you were able to like set like days a week. On Azure, it's like weekend or weekdays between these times, like predetermined. So you only have like a couple options. You can't say, Hey, I went from 2:30 to 4:30. They don't. No, no, you have to like say like weekends from 8 to 5 or whatever the predefined ones were.
[00:55:41] Speaker C: Makes sense.
[00:55:43] Speaker D: No, it doesn't.
[00:55:44] Speaker A: I like that it does in some world.
All right, I'm a real time story that we're not going to have a digital twins, but if you need Digital twins. And you know what that is in fabric? You can have it now because I. No.
[00:55:57] Speaker C: Good call.
[00:55:59] Speaker A: Yeah, I don't know.
[00:56:00] Speaker C: Seconded.
[00:56:02] Speaker A: Seconded shows getting long.
[00:56:03] Speaker C: We got to start. Colin. Yeah.
[00:56:06] Speaker A: You can now get warehouse at the end.
Warehouse snapshots are available to you in Microsoft fabric and Preview. This allows you to guess what snapshot your Azure data warehouse. Because that's a lot of data you might need to keep for backup purposes. Snapshots can be seamlessly rolled forward to reflect the latest warehouse state or allowing consumers to access the same snapshot using consistent connection string. Feature integrates with Microsoft fabric ecosystem, including T SQL and the fabric API. So if you need snapshotting capability, you can now get that.
Very cool.
[00:56:37] Speaker C: Protects you from, you know, little Johnny drop table.
[00:56:41] Speaker A: Exactly.
[00:56:42] Speaker D: But this is also good for like we just talked about like dev environments. Right. If you do have a good data sanitization and you want to take your production, you know, data and move it to dev and you sanitize it before you come in. Comes in, it's not a bad way to do it.
[00:56:57] Speaker C: Naturally not.
[00:56:59] Speaker A: And a feature needed for all of your. Net teams in about 15 years. Azure App Service now supports.
Net Aspire's applications, enabling developers to host their distributed apps on Azure's fully managed platforms. The. Net Aspire is Microsoft's newest framework for building modern distributed applications, and this integration brings it to the broader ecosystem. Again, why you need about 10 to 15 years. Developers can use familiar tools like Visual Studio and Azure developer CLI to build, deploy and manage their Aspire apps on the app Service.
You have similar platforms on AWS and gcp. This preview targets. Net developers specifically on Azure and this is a great way to upgrade all of your app needs.
[00:57:34] Speaker D: I aspire to use this feature on my day job.
[00:57:38] Speaker A: You aspire to use it in 10 years or tomorrow.
[00:57:40] Speaker C: Yeah, we'll see.
I hope to see that every. NET app is running on this framework and it actually is new and modern. But I'm not real confident that's going to be the case.
[00:57:56] Speaker A: I'm not confident either.
[00:57:57] Speaker D: I mean, for them to add a feature like this isn't difficult. Or I'm going to say that even though I'm not the one doing it. You know, app services is just a container really, they run your app in. So it's just a matter of them supporting the container runtime and getting up there. So as long as it's available in theory on a Docker container, it should be available on this platform decently easily. My assumption.
[00:58:23] Speaker C: Yeah, I mean I imagine. I mean, I'm making a lot of guesses here that. Net Aspire is like a. Specific. NET framework and so your application has to be coded to that framework in order for this to work in the app service.
[00:58:38] Speaker A: Isn't it just the latest version. Net it's something really special. Is it?
[00:58:42] Speaker D: No, that's what I thought it was.
[00:58:43] Speaker A: It's the next version of. NET Core. I thought probably.
NET Aspire 9.0. Yeah.
NET 8 was the prior version. This is NET 9, but they added Aspire to the name of it to make it sound cooler, which it does not.
And yeah, so if you're upgrading to net 8, switch that to net 9 and then you're up to date on.
[00:59:06] Speaker C: Aspire and then you can.
[00:59:07] Speaker D: Though it looks like from a very brief Google search, this was released in.
[00:59:11] Speaker A: It's been a couple of years now. Yeah, yeah.
[00:59:14] Speaker D: So I'm actually more concerned about why Aspire was not in app services.
[00:59:20] Speaker C: Hey man, have you ever tried to maintain multiple. NET ecosystems? It's hard.
Screw that noise.
[00:59:28] Speaker D: Yes.
[00:59:29] Speaker C: The app service team was like, no, I'm not updating. We just got to my net 8 probably yesterday.
[00:59:36] Speaker A: That bake first and then there's a bunch of new features in it and all those things.
Well, in news that's sort of shocking to all of us, Azure is retiring implicit outbound connectivity for VMs in September 2025. This default outbound access assigns public IPs that are insecure and hard to manage, which all the other clouds got rid of years ago.
The new private subnet feature in Preview prevents implicit outbound access. VMs in a private subnet require an explicit outbound method to connect to the Internet. This is where those pesky NAT gateways come into play. In aws, everyone complains, but are so expensive.
Azure is recommending that you also potentially use a NAT gateway or a public load balancer with outbound rules or a public IP on the VIM nic.
NAT gateways is the preferred option, of course, or firewall because. Because that is what most people do. Load balancers. Outbound rules also snap private IP addresses but require manual allocation of SNAP ports to each back end of VM, which sounds terrible. And public IPs on VM NICs give control over outbound IP but don't scale well for complex workloads needing many to one snaps but adjust to the traffic. I haven't talked about snaps as much since I did F5s.
These expensive methods integrate with Azure virtual network and follow a precedence order and voltage are configured so you can have multiple of these things turned on including the NAT gateway, the load balancer and the public ip. The shift to explicit outbound aligns with Azure secure by default approach. Well it wasn't secure by default before, let me tell you it matters. For security conscious customers running Internet facing workloads on Azure VM, the NAT Gateway cost will only cost you 4.5 cents per hour and additional 4 and a half cents per gigabyte processed. So not extortion but not cheap.
[01:01:11] Speaker D: No, but also the word there security conscious customers should not be correct. It should just be customers that want to operate. There is one other option which is using like the Azure firewall to run everything through it. It has a lower limit if you need more than the number of Snap ports running, you know so there's less so if you go to firewall versus the nat, but you know it also they made the announcement that they were retiring implicit outbound connectivity in like 2022 or 2023.
They're ending it in September and they're just gain this feature in May, really June, let's call it like this feature is just a basic in my opinion concept of the cloud and you know to me this is like Azure's running EC2 classic still, you know and to me they're finally moving into like hey let's actually use our VNETs and up PPCs and say not everything needs to be public. They did get rid of a lot of the hard limits that they had from the preview. Like you couldn't attach it, you had to use new subnets for everything and a lot of the like delegating permissions to services wasn't there and now it is so they have like fixed a lot of it which to me is shows a good sign of a GA product but you know it's definitely something that is far past needed.
I'm impressed it's taken them this long.
[01:02:35] Speaker A: I'm impressed it's taking this long as well. Just for sanity of all of us.
[01:02:40] Speaker C: Going to be second to the cloud hyperscaler marketplace, you think you'd learn.
[01:02:45] Speaker A: Yeah, I mean Google figured it out, why can't they?
[01:02:48] Speaker D: Yeah, there had to be something they did early on that hurt them probably.
[01:02:53] Speaker C: Yeah.
Network architecture.
[01:02:58] Speaker D: Yeah and changing network is always a hard thing.
[01:03:01] Speaker A: It is, it's always hard to pull. Well I mean that's why we had classic VPCs and non classic VPCs at Amazon because they figured out oh no one wants the classic VPC, they all want APCs and that takes major rearchitecture for Amazon even to do, but they did it when they were much younger, much less mature than Azure did.
Well, I have a cloud journey for you guys because I did a thing.
[01:03:22] Speaker C: Well, I've been waiting for this. I'm excited. I'm excited.
[01:03:26] Speaker A: You're going to ask me all kinds of fun questions, I'm sure.
So we have talked many times on the show about the idea of creating AI or a bot or something to help us with show notes.
[01:03:37] Speaker D: And we all started this project and never finished it.
[01:03:40] Speaker A: Ryan started it two years ago and gave up at the same point. Matt started it at least a year and a half ago and then gave up at the same point that Ryan gave up.
And I got there and was like, holy crap, this is bad.
And know why you guys both gave up. But AI helped me get there.
[01:03:55] Speaker C: Nice.
[01:03:56] Speaker A: You're welcome.
[01:03:57] Speaker D: Yeah, we do have AI. Okay, you cheated a little bit.
[01:04:00] Speaker A: So I. I got into the vibe coding thing because I was like, well, I'm going to try to see if I can solve this because I need a project to try out a bunch of AI that we're using for the day job to see if it's going to help developer activity. And I know enough of code that I can debug a bunch of stuff. And so I was like, I'm going to have vibe code. And so I set out to create finally the bottom. And we're calling him Bolt because the Cloud Pod and Lightning Bolts. It all makes sense.
[01:04:22] Speaker C: Oh, I didn't get that.
[01:04:24] Speaker D: I get it now.
[01:04:26] Speaker A: Wasn't obvious to me. I did use Claude to help me figure out like, what should I name a bot for a Cloud Pod website. It's like Thunderbolt. And I was like, oh, Bolt. Perfect. Because I'm not typing Thunderbolt every time.
[01:04:35] Speaker C: I'm not typing Thunderbolt.
[01:04:36] Speaker A: Yeah. And so that's like. There's also a movie called Bolt as well. It's a Disney movie. It's pretty cute if you have kids for Matt's needs. But, you know, so basically I started out with, you know, a blank GitHub repo. I called the Cloud Pod Bot at the time because I didn't know those name it. And I basically just got into a new application someone turned me on to called Brew Code. It was a port of Client. So I've talked about Klein many times on the show. Klein is basically a gentic AI capability built into Visual Studio code or through a plugin and you basically tie it into any API you want to use.
And so someone told me like, well, if you like Client, you should try out root code, which was originally a port of client. So it was root client originally. They rebranded to root code. But basically the big difference between client and root code is that root code has the whole idea of having different Personas. And so you basically have these different modes from basically a orchestrate mode, which is kind of like your main level where you live. There's an architect mode, there's an ask mode, there's a debug mode, and there's an orchestrator mode and each of the mode. Oh no, code mode, sorry. And each of those modes basically has certain rules that they have to follow. So like the Orchestrator isn't allowed to write code, it isn't allowed to write anything in the repo.
Architect's allowed to write markdown files, but it's not allowed to code. And so they have these different Personas and you can actually create additional ones as well. So if you don't like those or you want a DevOps one, or you want a security one, or you want a finops person, you can create all of those things and basically give it those contexts. And you can actually use root code to help write those so you don't have to write them from scratch.
You can actually use it for itself, which is interesting.
Net. Net. I said I'm going to go into Orchestrator and I describe what I wanted. I said I want to have a Slack bot that I basically can do it. Slash, command to or at to I'd like it to then take this URL. I'd like it to parse the content of the URL and then summarize it using Claude to basically provide the show notes. And then I want you to insert them into our Google Doc.
So very simple on paper, which is where both Ryan and Matt were like, this is super easy. We can do this because it was simple until you get to that Google part, which we'll get to in a second.
And so it basically produced a very detailed markdown write up architecture. It asked me a bunch of questions about what I wanted, what programming language did I want to use. And I chose Python for this one, which it had actually recommended Node, but I don't know Node as well and I hate Node as we talked about on the show. And so is Python for me or go. It's going to have to be one of the two or it could be Ruby but then I wouldn't get you guys to contribute on this code ever. And so.
[01:07:09] Speaker C: And we'd make fun of you, you.
[01:07:10] Speaker D: Don'T have to call Peter again.
[01:07:13] Speaker A: There was something like, well, you know, if they want to help improve this, because, you know, I've only done so much with it, I want to make sure I write something that they know as well. So Python was it. It's kind of the basic. Because I knew. I don't. I think. Pretty sure Matt knows Python. I know Ryan does know Jonathan.
[01:07:26] Speaker D: My bar in Python.
[01:07:27] Speaker A: Yeah. So, see, Perfect.
And so, yeah, so I also said originally I'm going to build this in Serverless because I am a masochist.
So the original architecture that I drawn up, I was like, okay, so we're going to have Dynamo, we're going to have Serverless, and we're going to have all these things.
And when I first did the first mock up of the code and I was trying to walk through the eventing and I was realized to myself, you know what, I don't want to do this. This is going to be bad. This is going to take way more time for me to troubleshoot and deal with it.
I don't have the best Amazon account where this lives because I don't have a full vpc. I don't have all the magic or some of those things. I realized that why as much fun as learning serverless was going to be, I was going to hate debugging it every day. I said to myself, scratch all of this. So I literally wiped out the entire repo.
[01:08:19] Speaker C: In true vibe coding fashion, Drew said.
[01:08:23] Speaker A: No, no, we're going to do the exact same thing, but we're going to do it in a container. Because I have an ECS container host and I can run this as a container and I know how to debug a container really well and it's going to make my life much easier. And so that's what it is running now as a container.
That's basically where I started out and where we ended up. That's the basic part of any questions on that. Before we get into the details of building this with Vive coding, trying to.
[01:08:44] Speaker C: Make sure my questions would actually be used.
Entertaining.
I mean, because it is just really, truly fascinating as you go through these projects. And I do love the. I love the fact that you went full circle on the serverless and then came back to running it based on just easiness to support. Because it's a true factor. Right. Like, it's fun to use all the new fancy things sometimes.
[01:09:09] Speaker A: But yeah, it was cool. Like, the architecture was badass. I was like reading it, I was like, oh, this is cool. We're gonna use all these serverless components and, you know, we're going to manage State and Dynamo. And I'm like, this is going to be neat. And then I'm like, thinking about the networking and I was like, ooh, that's going to be awkward. And then like, oh, but I have an event come into this, then go to that. And then.
So, yeah, I was.
And then I, you know, for another Slack team, I ended up building a bot as well. That one is serverless, but that one is a very simple, you know, simple web hook method. It's not very smart. But this one is much more complicated because it's got a bunch of components. And I have dreams of what I can do in the future as well that are beyond just what I've already created in the bot. But this was good. I say I did this over three days to get the prototype working, and I've been tweaking it since then, every day a little bit, enhancing it, fixing bugs, different things, debugging Google Docs, API.
But yeah, so the thing about AI is that it's awesome and it's really dumb.
So if you give it bad requirements, it gives you really bad architecture. Imagine that. So garbage in, lots of garbage out very quickly.
Which is part of the problem with the serverless code as well, was that I didn't know enough about serverless to actually give it good prompts. And so it made bad choices, which I figured out later.
And then one of the things it does is it gets into. Sucks into loops. So one thing I was never good at when I coded a lot was writing tests and so Claude and using Orchestrator and all these things. It has a whole model just for building tests and doing testing. And so I have a lot more tests on this code than I've ever had on any code I've ever written as a hobby project in my life. Which is great.
But then that also means that you now get to troubleshoot a lot more test code, which is also annoying. And one of the things that AI gets really dumb about sometimes is that it goes, this test is failing because of a sort error. I'm going to fix that by implementing a static list. You're like, okay, well, that would work, I suppose. And so it implements it and then it fails. It goes, stack list didn't work. Let me implement a sort. You're like, wait, wait, we just undid that and it'll go back and forth. So then you had to jump in you to say, like, wait, wait, wait, I don't Think that's the problem? I think the problem is this or that. And so I was able, and that's where the coding, knowing a Python to be dangerous is helpful. I was able to basically tell it like, I think you should try this instead or you're on the wrong path, try harder. And so that was able to get through some of those things quite quickly. One of the other things that burned me so, so I going through parsing URLs, you know, you give it a. You tell Bolt, hey, Bolt, here's a URL. Originally I created a static list of topics sections because the cloud pod shows broken sections. And so I said, oh, aws. And then here's a topic. And then later on I realized that was dumb. I can create automation for that, which I'll get to in a second. But the problem with that was I would see it parse through the URL. It would give the syntax of it. It give the syntax basically the bot, basically it dumps it with hTPX all the content of a website into Claude, and then Claude basically summarizes that into the show note format that I like and gives us some show title recommendations, et cetera.
Well, some websites are behind bot control and so they 403 you right away. So I wasn't able to actually do that. So then I was like, well, how am I going to solve that issue? I'm like, well, Claude can do browsing, so why don't I just have Claude browse those websites? It'll be more expensive on the API tokens. But I only had to do it for ones that are 403, which isn't everything. It was only like Oracle and Cloudflare and OpenAI were the only ones who were using anything super sophisticated that was topping my bot.
And so I was like, well, I'll just use cloud browse. And so I sell, you know, try to think. I'm like, hey, I'd like you to use cloud browse. If the HPX gets a 403 error, you know, very standard thing, it goes, oh, well that sounds great, but Claude doesn't have browser.
Wait, what? Yes it does. It's had browse forever. I use it all the time. Like I use cloud client, it has browse capability. I've seen it do browsing of documentation when it doesn't know what to do. Like it has this capability.
So, you know, I'm using. At that point I was using Google's API, which I'll get to in a minute.
So Google's API was not aware that Claude had browsing, which is fun. So Gemini doesn't know about Claude and Claude doesn't really know about Gemini is what I kind of figured out. But I was able to tell it well, actually I think it is in the documentation. And so then you force it, basically. I switch back to anthropic and I said, hey, do you have browsing capability? And it goes, no, I don't. I'm like, yes, you do.
[01:13:29] Speaker D: I know more than you.
[01:13:30] Speaker A: Can you check your documentation? It goes, yes, I do have browsing capability, but you're using Claude Python version 0.5 and you need to have 0.21 or higher. I'm like, this is a brand new greenfield project.
Why would you choose an ancient version of the Python library by default? And that's because, again, thinking about learning model, it's just finding websites with instructions and tutorials and it's become glomerate and then together into something. And so whatever happened to be this particular use of the python one happens to be a 0.5 reference. And so when I force it to use new requirements or say like, hey, I want you to make sure you update all your requirements, latest version, it'll then do that and all that, but out of the box, it didn't do that out of the box. So that was sort of fun.
And it's super annoying too, because it's like, this is a bug I shouldn't have to troubleshoot, you should just be able to do this. But it does now work.
It now parses the thing, parses the URL to hbx. If it doesn't get that, it then dumps it into Claude, browse Claude browse, pulls it and then pulls the same content. It introduces another problem, which is that Claude doesn't understand that there's a URL topic, which is the show title page, which is how we put everything in the show notes. So then I was able to use just enough HTTPX to pull the header data to then put the header into the URL, but then still dump to Claude to pull the content to actually do the summarization of the content to put it into the show notes. There's lots of fun little wrinkles with parsing the web in 2025.
I remember the glory days when we used to scrape the Internet without any bot control, and that's no longer the case. So that's definitely a thing.
[01:15:02] Speaker C: I think we'll see it grow over time too, right?
[01:15:04] Speaker A: Oh, I think so too. Actually, I saw a bug today where certain Copilot URLs were not parsing with Claude and I was like, I bet they're blocking Claude.
That would make sense. So I might have to do another if in the code to say, hey, well, if you're using an OpenAI website, you have to use ChatGPT for that and then give it different API key for that. But one of the good things about building something from scratch is that you can use all the cool things like Secrets Manager. So I'm using secrets.
I'm using GitHub Actions to build it and test it, which is great. And that's my first experience using GitHub Actions. Death to Jenkins. Now I'm on the Death to Jenkins camp.
[01:15:39] Speaker D: Yay.
[01:15:40] Speaker A: GitHub actions. Awesome.
[01:15:43] Speaker D: Just are on this camp.
[01:15:45] Speaker A: I haven't coded since GitHub Actions came out, so it's not something.
[01:15:49] Speaker D: GitHub Actions is great.
[01:15:51] Speaker A: I've used code pipelines on Amazon and I liked it. Okay. But GitHub Actions, way better.
[01:15:57] Speaker D: It's a completely different ballgame.
[01:15:59] Speaker A: Yeah. So creating all of this up until this point, relatively easy. This took me a day and a half to get through all of this and debug it well enough to the point that I was happy with the exports that were coming out of Claude.
Then it was like, okay, this is where Ryan and Matt struggled. Exactly.
Now we have to put it into the Google Docs API.
There is now an MCP that has come out since I developed this two weeks ago for Google Docs, which I will not use because I have done so much work to make this work that I don't even want to go through the process again unless that MCP is really, really good.
But basically a Google Doc is a garbage API. Like the API. This thing runs like you have to know it inside and out to be really good. Is why both Matt and Ryan failed.
[01:16:46] Speaker D: Started at the authentication layer.
[01:16:48] Speaker C: Yeah. Just the client off is just.
[01:16:51] Speaker D: Just the client off took me and then I was messaging Ryan and Justin when I was doing this, being like, how do I do this? Because I outload Google. They're like, well, you need this full credential. And it's.
[01:17:02] Speaker A: Yeah. So the. The thing about Google Doc is that it's not just an API key, it's like a whole JSON payload. You have to provide back to the API for all of that, which is definitely unique. I had not seen before.
[01:17:15] Speaker C: It's just the. It's a bad OAuth 2 implementation is what it is. It's not really all that. It's. It's just they. I Don't know why they did it this way. It's like they. It's like they're trying to force the worst aspects of OAuth 2.
[01:17:30] Speaker A: Yeah, I mean, like, literally it, you know, you have to declare in this JSON thing, you know, the, the project id, the private key, the client email, which is some, you know, made up at the time of creation, the AUTH URL Token, your token URI, the auth provider x509.
[01:17:45] Speaker C: Like the full list of API scopes. Like there's a bunch.
[01:17:48] Speaker A: Yeah, like all these scopes have to all be done, like in every call to the API, which is just. Yeah, that is so that Claude took care of, no problem.
[01:17:57] Speaker C: Well, that's awesome.
[01:17:58] Speaker A: Claude ran through that like knife through butter. Like no problem. I got that. Where Claude fails is in dealing with space recognition of inside of a document. So the problem is that we have a document that's blank at the beginning of the week and it has sections, AWS, GCP, Azure. And then I would basically paste in the URLs we're going to talk about. And then I'd write show notes, which is basically quick summary of what's in there so we don't have to read the document. Cause we're live on the air, which is super awkward.
So to do that you have to find the AWS section and then you have to insert the summary from Claude. Then you have to do the next one and the next one and the next one. And they're supposed to happen in order. Well, let me tell you, making that correct is a massive amount of python code because it basically has to scan through the document, determine the set point, the set point then changes on it. So you can't keep that in cache. You had to devalidate cache for that stuff. You had to do a bunch of things. I thought at one point I'd create a SQLite database and I would just count up the number of rows I added. That was a disaster. It was easier to basically recalculate the location on every insert, which was just ridiculous. Then there's a couple of fun things about our Google Show Notes document. And at the end of it, basically we have a part where it says end of Show Notes.
Basically the low end of Show Notes is a bunch of sections that look very similar to the top sections. Aws, gcb, and these are areas where we put parking lot stories. So we don't want to talk about this today, or we have ideas that we're kind of bouncing around for Cloud Journey, or we Put our vacation schedules like so and so is going to be out from this time to this time. You name it, it's in this End of Show notes. It's basically garbage Ville. But the problem is when you use the same shownote headers, they also match.
So I had to basically create End of Show Notes as an artificial document header end in all the Python code. So it doesn't look beyond End of Show Notes as End of Page indicator, End of Document indicator, virtual End of page document indicator. Which blew my mind. This is even a possibility that you have to do this.
But the very last section of the Show Notes is the after show.
And so I had a really fun bug where it would detect that the topic I was giving it was after show, but it would say that End of show notes was below after show and say, oh, that's end of file. What's the line location for end of file? Oh, it's 4635, which is at the very bottom of the document because it wouldn't connect the dots. That End of Show Notes is an artificial virtual End of show notes. And that's where you actually need to be looking at to get a placement above that.
I can tell you that took a whole another day and a half just to troubleshoot all of the Google Docs API and read the documentation to help the AI. And switching to Gemini at this moment was the right call because I'd say it cut half of my pain down because Google Gemini actually understands Google Docs API pretty darn well.
[01:20:56] Speaker C: Magically, Gemini 1.0 did not, because I did circle back to my attempt at the Gemini one and I had the same problems.
It couldn't even figure out the client stuff.
[01:21:07] Speaker A: Yeah, but so that's a mess. And then after that I got confident and I was like, who? Well, I can create custom show topic detection because I'm smart, and then I can fall back on my manual list selection because I basically the bot, you either give it the tap the topic code. I made a bunch of three digit codes to equal out to the different spots. So you can be like app bolt, aws, blah.
And so that's how I would do it. And if you didn't give it the aws, it would prompt back to you and say which topic do you want to pick? Which is got kind of annoying when you're trying to do like a bunch of topics at one time because I did multi thread this. So I can submit, you know, 10 or 15 URLs at a time and it'll do it all at the same time, which is great.
So I was like, well I can train my own model on the fly using our shownote archive to figure out topic.
So basically.
So now, so now it has so basically Claude, which is cool. Basically takes the show, detects a bunch of keywords out of it and does a keyword density match to basically whatever is in the Shona archive document to basically give me a level of score of whatever score it thinks is highest. And if it's above 0.9 and 0 in accuracy or in confidence, it'll basically put it in that section. And if it's not confident, then it'll prompt you and say I'm not sure about this one. So today's after show topic it had no idea what to do with. Never talked about that before.
And so when I put that one into prompt me I was like, where do you want this to go? Because this doesn't recognize in your document. So anyways, Bolt exists. I got some improvements coming for it today. So like you can live in our general chat rooms as well and talk to you like a chat bot does, but let you submit show notes. But he'll be in the chat rooms when he's ready to go. But yeah, Bolt is out there and this is what he does. So for the, you know, four co hosts, you know, for hosts, he'll take your show titles but everyone else does talk to you and say answer your question in a Bolty personality they're trying to create.
So that's, that's where I'm at with it. But yeah, that's, that was my project.
Three days worth of work.
I don't know how long you guys spent on your time, but I was.
[01:23:04] Speaker C: Successful and I spent a lot more than three days, man.
[01:23:07] Speaker D: Show off. Yeah, when I originally did it, it was a Lambda that was hitting Azure OpenAI when it just came out.
And then the goal was to have a hit to store it in Google Docs, to have that true multi cloud experience.
Just because I really hate my life and I was trying to make that do it that way and I got all the way through. It wasn't a bot the way you did it. It was just, you know, run on a Lambda and hit the RSS feed every day and count, you know, if we have it or not and then just add it to the shout outs. I got to the point when I authenticated, I was trying to figure out where the dock and I even remember having a conversation with Ryan being like, how do I use this API? And then I think we digressed after 45 minutes of yelling at it and moved on to, all right, I'll deal with this later. And then Justin did, I think, finally.
[01:23:55] Speaker C: Yeah, I mean, when you asked me about the Google Docs API, I do start twitching and the PTSD starts kicking in just because I did spend a lot of time trying to get it working. And I did get it as far as, like, being able to recognize headers, but it wasn't reliable. And like, all the same hard edges that Justin hit, he was able to resolve. And it just really does show you the power of using AI along with these things because it was. It's a lot slower iteration time when you're doing it as a single human.
[01:24:27] Speaker A: And.
[01:24:29] Speaker C: And, you know, the, the Google API Doc since my first or Google Doc API since my first attempt at this has changed a lot.
And so, like, trying to incorporate those in and it's probably, probably muddied the waters with the AI responses because the documentation is going to be or inclusive of both.
[01:24:49] Speaker A: But yeah, I mean, I will probably do a branch to see if the MCP is better, but I mean, the amount of debugging for over a day and a half to make this work with the current API, I'm like, I just don't know that I'm going to mess with it because it's working perfectly well enough. I mean, I've debugged a bunch of weird issues, like all the weird issues that you have a Google Docs just in general, you know, like, oh, I created a bulleted list. I don't want to be in a bulleted list anymore. And like, it. Yeah, they still keep trying to retain it or like, oh, I double return, but I'm still getting header two and all that. Those all exist in the API too. Yep. You had to be really, you get like, really specific. Very surprising.
[01:25:28] Speaker C: Yeah, it's very surprising how hard that stuff is. And it's, It's. I was, I was like, well, code is going to be way easier for this than the visual formatting things. And it is not.
[01:25:37] Speaker A: No, it is definitely not. I mean, again, like, you were trying to do something with anchors. I didn't go down the anchor path, but I mean, the anchor path. I definitely had my back pocket. I'm like, okay, Ryan had an idea of using anchors, so if this is failing for me, I'm going to try the anchor path. But I never got there because we ended up, you know, the bot and I ended up working it out how to, how to do the parsing and Parsing every time, you know, despite the.
[01:25:59] Speaker C: I mean, that's. The anchors were just a shortcut. Right. So that there was a specific thing to look for. So it doesn't have to read the whole doc. It just has to look at the anchors and list out the anchors. And then I had a lanker, an anchor for end of shore notes. Well, as, as well as a D mark, like nothing below this line kind of thing. Same type of logic. So that's cool.
[01:26:17] Speaker A: Yep. Well, that's my journey. So, yeah, take a look at Bolt when he's available in the general chat room. You can ask him questions. But yeah, I'm super excited about him. I have lots of ideas for what he's going to do for me in the future as well to continue to make my life easier, like package up the recording and getting it to our show author and a bunch of things. But yeah, super cool.
[01:26:36] Speaker C: I definitely have super. Some things too that I've been looking at because. And you've got. Talking through this, like, you know, the sneak preview I got, you know, a few days ago, like got me all excited and so now I have a couple ideas I'd like to add on to Bolt and.
[01:26:49] Speaker A: Yeah, so that's why. Yeah, it's internal open source for us, so we can, we can customize it and add functions. And the tests are really good. So you can. Don't worry about breaking my shit because you're gonna break my test and then you'll know you broke me.
[01:26:59] Speaker C: Oh, I'll break it. Don't worry. We know.
[01:27:02] Speaker D: Pretty easy for us to break. Don't worry.
[01:27:04] Speaker C: Very cool.
I mean, it's. It. And it really is just a.
I think it, it's night and day, like from the last time you developed code to this using AI. Like, it's just not the same experience. Right? Like, it's.
And it's. This is how I think that. I think this is how coding, all coding is going to be done in the future. Like, I really do. Like, it's.
[01:27:25] Speaker A: I mean, like, I. Again, like, I can, I can read the code, I can write Python a little bit. I'm. I'm not great at it, but I have a good architecture background. I know about DevOps, I know security. So one of the things I caught it doing was I was looking at this Docker build and I was like, wait, are you injecting the secrets into the container? I'm like, no, AI.
[01:27:48] Speaker C: Bad, bad AI. Yeah.
[01:27:50] Speaker A: I'm like, hey, I just noticed you're injecting the secrets directly into the container build. That's a bad security practice. He goes, oh my gosh, that's right. That's a terrible security risk. Let me fix that for you. I'm like, you're supposed to be using secrets.
[01:28:00] Speaker C: It is glorious how like, it is like having, you know, led a team of junior devs, like brand new college graduates like that are largely just taking source information from Stack overflow. How similar to that experience that I had previously, which is like, wait, what are you doing? Oh, don't do it like that with AI.
Because it is, it's. I mean, it's. There's so many bad practices out in public Internet, right. That are documented. That's using that same Stack Overflow data. Just like a junior dev.
[01:28:33] Speaker A: Exactly. All right, guys, we'll talk to you next week here at the Cloud Pod.
[01:28:37] Speaker C: All right, bye everybody. Bye, everyone.
[01:28:43] Speaker B: And that's all for this week in Cloud. We'd like to thank our sponsor, Archera. Be sure to click the link in our show notes to learn more about their services.
While you're at it, head over to our
[email protected] where you can subscribe to our newsletter, join our Slack community, send us your feedback and ask any questions you might have. Thanks for listening and we'll catch you on the next episode.
[01:29:16] Speaker A: All right, I have a quick after show. I know we're kind of running late today, but a bunch of people have pinged this to me on Twitter and other places because I know I have a gaming computer and they're like, have you checked out this Silverstone case? And I'm like, yeah, I did see it and I have zero interest in it.
Mostly because I was a Mac guy back in the 486 era. So I have zero nostalgia for these beige cases like this. But I do see the idea of why, you know, the FLP02 case might be fun for people who were in the 486 or early pendium era.
And like the retro exterior of a, you know, slightly off color computer case with a turbo button. So, you know, it's kind of cool. I like the idea of it. I just.
Not what I would do because My first Windows PC was Windows 95 and it was a laptop my dad bought for I don't know how much money. And the CD sat underneath the keyboard and would spin. You could see it in a little window that sat under the touchpad and that was. It was a 486 and I think it was like 60 MHz and it had no RAM and it ran slow as molasses and I was like, my Mac is better because I was that guy.
But then I got intrigued by it and then that got me into Windows. Then I spent the next 20 years on Windows and then I came back to Mac OS after they added Linux back to the underneath it. Because then once you got into Windows and you got in there like well what else can I do with this hardware? Like oh Linux, let me get into that. And so that kind of led me into where I am today, a computer guy that I am so. But what do you guys think of this case?
[01:30:43] Speaker C: Well, secret time.
I have to admit something which is I've never really appreciated the industry for building your own computers. I've never understood the whole RGB led, like perfect cable routing, color coordinated everything like you know, picking specific color for your cooling fluid. I've never been that guy. I build computers for performance and I want them to be as small as possible and as black as possible and hidden behind some piece of furniture somewhere. Like I don't care about how they look.
[01:31:18] Speaker A: Oh, see I'm all on the RGB whammy rainbow vomit underneath my desk coming from my gaming tower all the time.
[01:31:25] Speaker C: So I had kids and that changed it all because they are all about it and so I've had to learn all that stuff. But yeah, I don't. There's no part of me at all that wants to build an intel inside beige PC case for that same reason. Like in my core I'm like, why? Like I built them, they looked like this.
[01:31:48] Speaker A: Yeah.
[01:31:48] Speaker C: And they sat under my desk.
[01:31:50] Speaker A: I supported some that look like this when I worked at the school district that I hated because they were always some ancient H vac 486 that I had to figure out how to make work in our modern ish network at the time.
Yeah, but I don't know. Matt, how about you? Do you have any solid for this case?
[01:32:06] Speaker D: Less for the case. I kind of, I dabbled in building my own PC doing stuff like that for years and then I just ended up moving to like laptops and stuff like that, you know, So I, I like it. I think it's kind of fun. Would I spend the money on it? Probably not.
[01:32:23] Speaker A: Yeah, I didn't even see the price of this. I just, I saw it, I was like I'm not interested and I didn't move on. Oh, 220 bucks. Not the cheapest case, but not terrible.
Yeah, I, Jonathan, have nostalgia for. I'll ask him when he gets back.
[01:32:36] Speaker C: I'm sure he does.
[01:32:38] Speaker A: I'm sure he does.
[01:32:38] Speaker C: Yeah.
[01:32:39] Speaker A: It feels like something he would have. Totally.
[01:32:40] Speaker C: I'm sort of annoyed. I mean like that you know, the Mini ATX is sort of the smallest form factor for building your own computers.
[01:32:47] Speaker A: Like I, I mean Micro ATX is still a thing.
[01:32:50] Speaker D: Yeah, it still exists. Is that, I mean I feel like that's more like business people do it and stuff like that.
[01:32:55] Speaker C: I think I'm saying the wrong thing because it's not particularly small. Right. Like super small.
[01:33:00] Speaker A: I mean it's, it's a smaller itx.
I remember. I remember Micro ATX or my, or ITX is cheaper or smaller. Sorry. Let's see if I can get quick Google search here for you.
[01:33:11] Speaker C: It's micro versus Mini. Is or am I just saying the wrong thing? And there's only ATX and Micro atx.
[01:33:16] Speaker A: No, there's, there's ATX and there's itx. There's are the two. So Mini itx okay. And atx.
[01:33:22] Speaker C: Yeah, yeah. So there's just, wasn't there like a.
[01:33:24] Speaker D: BTX at one point that came out that kind of failed?
[01:33:28] Speaker A: Maybe I don't remember that one. Yeah, I remember. I, I, I mean the problem is with the, with the video cards you need nowadays. No person should be running ITX.
Micro ATX is 244 by 244 and the mini ITX is 170 by 170 millimeter on the size. But yeah. Anyways, I don't like desktop computers of that size either. I could tower my desk. I can ignore it.
[01:33:54] Speaker D: I'm weird. I never really got into gaming or anything else along those lines. So I feel like, you know, the thrill and the design of building it and going all the way over there just never fully stuck with me.
[01:34:08] Speaker A: All right gentlemen, we'll see you next week.