Heap Happens: True Crimes in Go 🧹🔪 Plus which YAML is the best for you?

So that's really cool to hear that we're making an impact. Not only are people going to go for camp because they heard of it on the show, but somebody's speaking there. Louis speaking on the topic of why your code deserves better and why you should give it a linter. So I'm a big linter fan. We had Eldaz on the show a couple of weeks ago to talk about Golangzhi Lint.

I sing the praises of Lunters a lot. So unfortunately I won't be able to attend this talk. Maybe there'll be a video online I can watch later. If you'll be in The Czech Republic, the boy should figure out the dates of that thing again.

And then you go do gomontidy anyways. Mhmm. A neat little time saver from, Gabriel. And if that convinced you, the Gopher Camp in The Czech Republic, which is the first Golan Conference in The Czech Republic, is in thirteen days and seven hours, 04/25/2025. Looks like a full day, really cool stuff, lots of speakers, lots of sessions. So if you're in the area, go sign up and even supported by Microsoft, Alban Labs, Odin. Gg. Very cool.

And then the wait function at the end waits until all of the started go routines are completed before it continues. But that just got simpler or will be getting simpler. They're adding a go function, which matches the error group. If you've ever used that, which is a similar concept matches the error group functionality. So instead of doing wg.

Add and then go and then defer wg done, you can just call waitgroup. Go and pass it a function that executes and it does the add and and the done for you automatically behind the scenes.

So don't don't even try it. Right. I wanna share a really, really cool blog post and I'm sort of on the fence. It's a seven minute read. So and it's like a bug hunt sort of blog post where there's a really specific, super interesting, hardcore quote bug.

Right? If you like reading these sorts of blog posts and trying to figure it out for yourself as you're running sort of like a murder mystery or a true crime podcast, I'm gonna spoil it, I think, as I sort of tell Jonathan about this blog post. It's called leak So spoiler warning, I guess.

So, you know, they have pprof enabled, so they went to slash debug slash pprof slash goroutine. If you don't know what that is, Go has super good built in profile stuff, which you can turn on by just adding a line, right? And it has almost no footprint on your production, so you should probably do it. Honestly, they grab the goroutine pprof. So are there hanging goroutines?

With debug equals one, which is like the level of a debug. This detail is important. Remember it for later. Okay. They found, you know, generated a graph and okay, it's related to the SQLite driver, something from a cloud resource and they show the graph here and, know, that beautiful graph where something's red and they're like, oh, let's look at the heap profile location trails to find where it leaks.

Okay. So they narrowed it down to a specific part of the code, which is okay. It feels like a dead end. Maybe the problem is Indigo runtime. Apparently, it was, like, in the Go runtime, they introduced another tool called GoRef, which I never used Mhmm.

Because I never got to a bug that's this hardcore. But it's a heap object reference analysis tool. So you could like look at, you know, the call stack, which created every single object at runtime. Like see the memory reference distribution and and of stuff that's reachable by the garbage collector. And from that, you can go back to the places in the code where the objects were created.

Blah, blah, blah, blah. Like, they're they're going through this research and they found that GoSMB two has been blocked for tons of minutes because of a race condition they introduced and that Go debug, like the pprof debug equals two hides it. And they actually opened the bug for the Go runtime, which at the point of your research, if you're opening bugs in the Go language itself, you know you've hit a real edge case. So yeah, you know, debug two should have included in any reasonable person should include stuff that is found in debug one, right? But it didn't, they walked through this entire thing and fixed it, it's resolved, super cool.

So not only they shared the blog post, they actually improved the language. So like Cyclo and they had super great stuff. They just implemented, like they had to find a fix even after they opened the three additional issues. So they just like generated a monitoring loop and saw how long that finalizer takes. So it like triggers them.

It gives them an event and they can restart the thing. They even share the link to that repo, which is called F I N Q, think. So if you have a slow running finalizer, which or a blocking finalizer, you can use this to track it. So opening issues, contributing back to open source, like, guys going above and beyond in this one, which I really liked, just like super hardcore research. You should go read the blog post, honestly, if you're into that bug hunt sort of stuff.

So I don't know what the sort of best approach to this is. I mean, you can continue using it obviously, but there are some other ones out there. I don't know if there's another one that everyone's rallying behind. I think there's a few different implementations. Honestly, I'll probably keep using this one. I haven't had problems with it. I don't know. What do you think, Shah? You're the one who wants it in the center library. And even though apparently nobody else does.

I would go with that just because you want to flock towards the most popular one, even though the latest release was October and it's like April. I don't know, like, is that good? Is that bad?

And the ones that do get released, like two weeks ago, Go CC Y Go Yammer was released. You're like, wait, why do they still have like bugs? Generally not super happy with the situation. And whenever I'm going to like next time I'm going to do Yaml, I'm just going to try to make sure that I'm using only the simplest features just to not run into any edge case. That's I guess my best recommendation.

Don't try to do anything funky with YAML. If you find yourself in this situation where you're doing that, like, I don't know, reconsider. There's just so many edge cases and now that there's not a quote unquote standard library to do it, I don't know. Maybe it's just worth thinking about other languages that have all the features you're gonna want. Right?

We talked on the show about stuff like what was that configuration language? We had like two already on the show. Q maybe? Yeah. Q, c u I, c u e.

Sorry. C u e. Yeah. That's like, configure, unify, something, execute, which is, you know, a good way to define like schema, data validation, configuration, scripting, all these sorts of things, which like you might do with YAML and, you know, Q knows how to read and write YAML. Or maybe if you're like super hardcore, maybe you need a proto buff in reality and you've just been using YAML to hide it.

Like, I would seriously consider doing something else, which I wouldn't recommend if you were just using JSON. Right? Like, if you're just using JSON, that's fine. You're working with other web stuff and it just works for you. Fine. But if you're using YAML to configure stuff

That's not going to happen with TOML. So generally, like there are libraries I would recommend. I guess I would go with goccygoyaml as my first option. But the moment I do something slightly more complicated like validation or anchors, I would just stop and fix that because at this point, the YAML is like technical debt, but that might be a bit too much. If you're just using a normal file, goccoyyaml, I think is the best option.

Although it's now that I imagine it in my head, it's a bit too much like a a rat in a cage with their bottle hanging, you know? Anyway, my Lightning Run thing is super fast because it's the new release of FINE, which is faster than ever and passes goes race checks. Yesterday, Andy, friend of the show, released v 2.6, massive update to FINE, which is, like a GUI app development framework thing, influenced by a material design where you can develop like Go apps that run locally pretty easily. And now it's super, super fast without data races as well. So if you do stuff on the background, you can use find dot do to do that.

But other than that, it's a lot faster because it has a new threading model, which sounds like a big achievement for the fine team. Good stuff. And it's a super big update. It has, like, I don't know, maybe 50 bullet points on the Mhmm. On the change.

So I assume it's gonna be, you know, six point zero is gonna be slightly buggy. But if you're an early adopter, that's a great time to switch and up and upgrade. Cool. Let's move to our ad break, and then we have a sort of preview discussion about, MCPs. MCPs? I don't

You can join our Patreon and kick in $8 a month to help support the show. We need to pay for the show even though it's a hobby for us and a way to learn. Jonathan is gonna learn a little bit about MCPs or I learned about the new weight group stuff. But we do pay for hosting and editing and apparently refunding hoodies that went out, but didn't get to listeners. Luckily, they did get to listeners before I pressed refund because I was about to do that.

Anyway, we wanna shout out Kelsey Wolf. Kelsey Wolf, thanks a lot for joining our Patreon group. I don't know what this is called, audience, audience. We really appreciate it and it helps just make this hobby slightly more sustainable. So thanks a lot.

If you wanna support the show in other ways, or just talk to us, you can find everything in cupago.dev, that is cupago.dev, where you can find links to our Swag Store, our Slack channel, Cupago on the gopher Slack, which is growing. Like a lot of people are joining. I have a habit of doing like emoji response for everybody who's joining the channel just to like make it friendly. We're at five twenty eight people, which is a lot. And I don't know, Iska Ambraejak just joined.

Thank you for joining. I don't know if you wanted the shout out just for joining a Slack channel, but you were just the latest example. So yeah, cupago, cupopcase with hyphens on the go for Slack, or you can email us newscupago. Dev, that is newscupago. Dev.

Another way to support the show is to share the show with friends, coworkers, co students, anyone you think might be interested in Go. Maybe people who are interviewing for Go jobs right now. This show could give them like an edge, cause they could be like, oh, I'm listening to a show and I know about the latest proposals. I would be like hyped to hear about a candidate who's doing that extra effort. Someone gave me that advice recently.

They said, oh, I'm telling people I'm listening to the GO podcast and I'm learning stuff about it and I'm sharing it at work. And the interviewers really like hearing about that. So maybe that's like a crowd we could tap into. So Share the Show, we don't pay to advertise and we're not doing a lot of social media work, we're just posting the episodes and hoping they're good enough. So sharing it or leaving a review on Apple Podcasts, Overcasts, Spotify, wherever you listen to your podcast helps us a lot pushing the show to other ears.

Finally, I want to mention our upcoming meetups. Well, yours is like the end of the year. Right?

Feel free to join. It would be super cool. Two hours, we're doing one, live podcast recording and a talk that's to be determined. Talk to me if you want to find that slot. It's kind of awkward because there is a Go San Francisco group, but their main like, their maintainers are not answering. So if anybody knows them and could, call them and connect us, that would be cool. Otherwise, Wednesday, May 28, possibly at WeWork or maybe at a cooler place if I can make it happen.

Super simple. Join us next week in LinkedIn Brain Rot, the show. So obviously LLMs have been a super big topic, right? Like I'm literally working at a copilot AI security startup right now. But even beyond the bubble of where I work, MCPs, which is model context protocols, have been making the rounds and have been super interesting recently.

Like they've just exploded in popularity. I don't know if it's going to die next month. It might, but there have been two big announcements. So I'm going to give you the like preview of the announcements, but I want to dive into a few projects, actually use them and maybe even develop an MCP myself before next week so I can be more informed about it. Also, I want to ask our community on Slack to share if they've been working with all these stuff like MCPs or A2As so we can highlight it next week instead of just randomly highlighting a few projects.

If they if someone's actually used it, that would be interesting. What is this? It's like a client server. MCP was introduced by Anthropic. It's basically a protocol, which is like a standard, but it's just by Anthropic. So it became standard just because they're popular, which gives your LLM a way to talk to data sources and tools. Right? And you expose specific capabilities through the MCP protocol. MCP protocol, that's stupid, right? That's like an ATM machine.

Through MCP servers to give access to the various services for your LLM. So you could, in theory, develop a model context protocol server for Jira and then, talk in natural language to your, MCP server or more realistically, the use cases your AI agent does that. And it provides like, if you want to list all the issues in Jira, you talk in natural language to that MCP and it translate it to the actual thing that talks to Jira, right? An API call or whatever. Talk to databases, local files.

Surprisingly, the most useful case for it, write and not read. So if you want your LLM to actually do stuff like create emails or write stuff or buy stock or do stuff for you. Let's say you want your, I don't know, AI agent to book you a flight or a meeting at a headrestaurant, right? Relevant for some of us, if not everybody on the call. But it has to go through a whole like it needs to open the web and click on reserve and then check the prices.

And like there are many things you could put in the prompt, but it actually has to do stuff, right? Call the place maybe, transcribe the call, get your confirmation if that works for you, check your calendar. So MCP could help the LLM connect to all these sorts of things. That's in a nutshell. Got it.

It was a huge announcement and everybody and their mother have been developing MCP servers for their thing because it makes sense. Right? You it's like an explosion of APIs. It's like people are making the bet that LLMs are gonna be the new Internet and this is gonna be the new web pages. Right?

Okay. Everybody's been jumping on it. And then April 9, just two days ago, Google announced basically the same thing, but with their protocol, which is called A2A protocol, agent to agent, which gives agent interoperability. And they announced it with contributions from a billion tech partners like Atlassian ServiceNow. So this is the enterprise e version of that basically, where it's shipped with McKinsey, KPMG, Atlassian, Box, MongoDB, PayPal, Salesforce, SAP, SAP integrations, ServiceNow.

And it basically, again, support for agentic capabilities. So let's allow AI agents to talk to tools and even multiple agents to talk to a tool. So you have a remote agent over in Atlassian side, then you have your local agent running locally, and then you can talk between these two agents and, like, share tasks. So it's sort of like another network layer on top.

Because there are quite a lot, actually. There are, like, seven Kubernetes ones, although I didn't know we needed so many. So I'm mostly interested to hear from our crowd. If you have used these technologies or developed a server just for fun or whatever, let us know. We'll talk about it, next week when I'll try to learn a bit more about that and share my thoughts on how it is to do it in Go. Looking forward to it. Yep.