Jerod Santo: Right. Because if they received a bunch of token for their package getting popular, and they went to go sell it and they were just dumping on the market, and the demand wasn’t there, then the price would crash, and you’d have your typical peaks and valleys of the crypto sphere. So you’re trying to like stabilize the coin, basically? Or what’s this tokenomics? You’re trying to stabilize the value of the token, or…?

Autumn Nash: The price of cars now is just wild. But I think that all this is gonna change the way that we have finances, too. Back in the day when I started art school you would buy a version of Adobe Photoshop, usually with like student loans or whatever, your first money, and then you’d use it until you had a feature you needed now, another one. Or your computer completely died, and you had to get a new computer, so you needed a new license. But now you have to pay Adobe Photoshop every month. And artists already don’t make a lot of money… You know what I mean? And then now, it’s just wild. We’re switching how people pay and own –

Jerod Santo: …there’s reasons to do that. It’s a compelling case, with price, and ownership, and the ability to modularly repair it over the years. Let’s take the tractor that you – I don’t know, is the tractor that you built back when you first started this different than the tractor that you’re running now? I assume you’ve got–

Daniel Whitenack: And looking back, I would tell myself to just go ahead and take that price hike… Because if that was the case, I wouldn’t constantly be moving models around between these two systems to do my testing… The way I understand it, the Threadripper, the AMD chip has more cores… So if yo’re doing a lot of multi-threaded stuff, it’s really good. But the single-core speed on the Intel processors is higher. So depending on what your workload is, then you could actually get a performance boost, even though you have fewer cores, but with a higher core speed with the Intels. All of that stuff - it seems to work a little bit better for me… And this is just my own personal experience.

Adam Jacob: Right, exactly. So the core problem is the same, which is the right-hand side of that equation sucks. “I need to raise my price. How do I raise my price?” And that’s how. Now, Hashi Corp – in the case of HashiCorp, they did it through saying “Well, the only builds that are available are the ones that come from us under our commercial terms. You’ll see those terms ratchet up over time.” So they didn’t want to freak everybody out. It’s not as soon as they did that, they also ratcheted the right-hand side intensely. But they absolutely will.

Guy Podjarny: I’m actually quite passionate about the freemium vs. free trial type model… So what I feel is that people conflate them. Snyk is a freemium model. And what does that mean? I’ll extrapolate beyond Snyk for us… Which is we actually serve a use case for users that would never pay us. And we serve that use case really well for a long period of time… Again, until they basically don’t qualify for that use case.

For us, for instance, we support individual developers who want to secure their applications, as well as small teams, within our freemium tier. And there’s no buts, there’s no exceptions. We also support open source users. That’s maybe even the first profile, or people protecting open source platforms… And those users should be - we do all we can - to make them really happy with the product, independent of whatever purchasing ability. There’s a use case there about an individual developer, about an open source developer, about a small team, where the product should within the free tier truly deliver success. And in general, I think in freemium that’s what you need. You need to define a use case that you will support for an extended period of time, or forever, presumably, that people would run.

[44:00] Free trial is really about a chance to explore the capabilities of your product. And what people do is they call it freemium because they think of free trial as something that is only limited by time. So they say “Well, no, if it’s not a time-limited thing, it’s not a trial.” But in practice, they don’t solve any use case… So they for instance – say you have a collaboration tool, and you allow collaboration for up to five users as something that’s within your free tier. But your product really only brings true value - you know, the product that you’ve designed is really around large group collaboration. And collaboration amidst five people – let’s assume for a second it was like a Slack style system… Well, the value is fairly dubious. It’s not really that significant. When you’re five people, you don’t really care about that much assisted collaboration between the five of you. I mean, you can communicate pretty well.

So you find yourself not really solving a problem, and in earnest, you don’t actually want people that are five-person teams to use it. What you want is you want the 100-person team or the 1,000-person team to start with five people. But for those people, you never started, you never solved an actual use case, an actual problem that they’ve had. And so this type of mistake typically happens when people think about freemium last. What they basically do is they say “Well, I’ll build a –” let’s say in our case I will build a security solution, I will price it by the number of developers building the application, and therefore now how do I keep people? I’ll give people something for free, but I will keep it from encroaching on my commercial; I want to keep enough value in my premium tier, so that there’s enough people that would move over to upgrade.

And so they start from that, and they reduce it back down. Or they now say “Okay, I will only allow up to five developers.” If I was to do that, and I would say “Well, it’s only teams of up to five developers”, then basically none of those is probably going to be that significant kind of a customer. If it’s organizations with no more than five developers, that’s not really the profile that I’m trying to reach.

So what you really want to do is you want to understand for a given user what is the use case that you are solving for them. You start with the users you want to reach, the people that you want to reach in the first place. Developers as a whole, and then developers within enterprises. Then you wanna say “Well, what’s the problem that you want to solve for them?” Let’s say you have a team over there, or you have an individual developer there - what is the problem that they want to solve? Well, they want to find out if there are vulnerabilities in their code. And now you build a solution that helps them address that, helps them find those problems and be very comfortable with the fact that they will never pay you.

And then as you expand it, when you think about like “Okay, another use case for us is developers in open source”, like an open source project developer. Well, what is the use case that I solve for that user? I help them integrate into their open source project an ability to identify and fix known vulnerabilities… And I want to do that indefinitely. What I want those two developers to do is very different. The developer in the enterprise - what I want to do is I want to make them successful, but then I want them to help introduce to other team members, I want to inform them about my commercial capabilities, I want them to engage sales when appropriate, so that they would go off to the purchase.

The open source developers - what do I want them to do? Well, above and beyond, again, like the first case, being successful and loving the solution, what I want them to do is tell the world. I want them to help the world know that open source security matters, and that Snyk is a good way to address this… So if they publicly talk about me or about Snyk, and if they help educate their peers, then that provides that value… And I shouldn’t bother them around purchasing. I can also consider saying “Well, if I think that this person might work at a job, work on also non-open source projects, maybe I want to have them consider using Snyk in that surrounding.”

[48:34] So these are all very Snyk-specific examples, but I think that the key distinction here is when you have a solution and you think about how do you bring it to market, there’s this aspiration that it’ll be a freemium-led go-to-market; so you want to get these masses of people using your product, and you want some portion of them naturally growing to buy your product. And what often happens is you will unleash a freemium product and then nobody would use it, or the products would not really get adopted. To avoid that problem, or if you see that that’s happening, you have to stop and think “Who are the people that this freemium product works for? What are the use cases that I’m truly solving for them, and what is it that I want them – that I as a business (a little bit more cynically) want to get out of them?”

And you know what - if you can’t figure those out, that’s entirely legit. You don’t have to have a freemium model. If you don’t have those solutions, if the minimum use of your product is one that really should be commercial, then it should be a free trial. But then you should adjust your methodologies, adjust your expectations that what you’re really doing is you’re offering them a free trial, and you should convey it accordingly to them, to say “This is a trial of our capabilities”, so they don’t have the misconception that your product is limited in the way that you’ve provided it.

Matt Jaffee: Yeah, that’s excellent, actually, and we’ve talked several times about the e-commerce use case, where you’re looking for a certain item - maybe it’s a TV, and you’re drilling down, you want a certain size and a certain resolution, in a certain price range… And that is really a slam dunk use case for a bitmap index, especially because it’s powering a UI that is very sensitive to latency. If you have a consumer who’s digging through something, they don’t wanna wait a full second or two seconds for their query to finish; it’s more engaging if they have to wait 50 milliseconds.

Bryan Boreham: Yeah, it could be. I mean, it depends how good your benchmark is. In some cases it’s really, really hard to emulate the true production conditions. There’s also a bunch of things to watch out for. You know how I was saying your processor can do a billion things in a second? As long as you don’t use more than like a few tens of kilobytes of RAM of memory. The minute you go up past your L1 cache, the whole thing is going to slow down by a factor of 10. When you go past your L2 cache, it’s going to slow down another factor of 10. So you need to be careful when you try and recreate the problem in your benchmark that you don’t make it too small. You make it so small that it’s unnaturally fitting in the really, really tight cash of the processor…

This is one of these mechanical sympathy things. It’s a huge amount of knowledge to kind of learn about processor architectures, and different layers of caching, and so on. I don’t think everyone has to learn that. But a little bit of – certainly just the fact that you don’t want your… You want your benchmark to be at a realistic size. You don’t want it to be so big it takes a day to run, and you don’t want it to be so small that it’s unrealistically fast.

Mihai Todor: Yeah, I can go… So one use case - let me give a bit of background. I was working at a company called Nitro, and we were ingesting a lot of click data. And I saw this thing called Benthos, and I was like “Man, I really want to use this somewhere.” And that was an interesting use case, where - you know, we get all these people who have a desktop app and they click on various things in the app and you send all those click events to the server… But you don’t really want them to end up being stuck there for a while, and then you do some Bash processing on them. You might want to have real data analytics, where you look at them in real time and you update a bunch of graphs, or… You know, you can do many things with them.

Maybe a more interesting use case was at another company where I was dealing with lots of audio data in real time, where you have a big call center where people are talking with customers, and you want to give them some hints on the screen. “Hey, you’re speaking too fast. You’re speaking too slow. Hey, you’re overlapping.” And there was a system in place that could analyze the audio and using some machine learning it could predict in real time this event happening. And then those events would have to be sent to the operator in the call center, and that was another use case for data streaming. This is a place where I ended up using NATS to kind of receive and send those events along. I could talk more about that a bit later.

Priyanka Sharma: I’d like to talk about documentation from a slightly different angle, which is that – we were just talking before we went on the air that the open source ecosystem today is a lot more than some code, some library that you pull in and throw into your own software. It’s a much more vibrant community, there’s a lot more value when things are going on. And part of that value is if you are building something - let’s say you’re being entrepreneurial and building an app, or something… It’s a community where you can discuss ideas, where you can find friends to test your stuff, to give you feedback, all of that… And documentation actually plays a really strong role in that.

As an example, I’ll tell you - the Jaeger project did this the first time, and I thought it was really cool… They don’t have any user telemetry baked into their project that’s open source, but they wanted to know who is using this, what’s going on, how can we improve… So they started an issue where it was like “Okay, anybody who is using Jaeger proactively, if you are open, just share what you’re doing, what you like, what you don’t like…” And I was like “Who’s gonna write that…?” No, it’s a long, long, long issue, with so many people writing in… And then that moved into, I think, a readme… But what that did was suddenly they had this great feedback from their community on how to iterate.

And then I actually – I sometimes contribute to the Jaeger project when I have time; I help them out… We picked people off of that list and asked them “Hey, would you please do case studies?” And people were already so engaged. Normally, what’s the case study process when you’re a product? People feel like they’re doing you a favor, and there’s a bit of a power dynamic, I think… But in this case, they were grateful we called.

Then suddenly we were writing all these case studies, and had this really deep detail of how to use Jaeger, what are the benefits etc. And all of that started because we used existing channels - issues, readmes - to write down that we’re looking for X, people come and respond… And you’re not just telling me, you’re not just telling an individual, you’re telling the whole community what you’re doing. The many-to-many communication, which is what documentation is, I think is really powerful. It’s about the how-to, but it’s also about “How did it go? Where should we go next?” And it changes how we do things, which is really cool, I think.

Cameron Seay: [00:24:10.21] I’d be glad to. I’m sure my colleagues, my business partners would love me to talk about that. [laughter] We have very flexible – first of all, I’ve been teaching bootcamps a lot. I’ve probably taught 10 or 15 bootcamps. They have varied anywhere from 2 days to 12 weeks. So there’s all kinds of gradations of it. These are like corporate engagement. So a company comes to us and they say – they’re iterations, but in this case, “We want you to train X number of people for us, and we’ll give you X number of dollars. And this is what we want them to learn.” And we’ll go through the process. In this case, they’re learning a week of Agile, before – everybody has a week of agile. Then mainframe basics; then something called vSAN, which is a file thing… There’s something called [unintelligible 00:24:52.17] which is a – think credit cards. Then COBOL. And then a little DB2. A little database in there, too.

And so we have labs we’re gonna do… Today, right now, the professor’s lecturing. But certain days – like, the last time we had three hours lecture, three hours lab in the afternoon. Now we have two days of lab and a day and a half of lecture. So this one wanted MQ. The last one wanted to focus on vSAN. MQ is like a messaging service between applications. You guys understand the concept of messaging. Think Tomcat. Just a messaging server. So they wanted that. We can do whatever you want. I’ve got a worldwide network of guys that are credentialed to teach this stuff, and they want to teach it. So instructors are no problem.

Amal Hussein: Yeah, this is really fascinating, because I feel like I’ve had a very similar evolution, except for – I think for me it ties very much into the logging story, because I’m really a data freak, so I love monitoring and logging and classifying my errors, and making sure that if there’s something new that’s out of bounds, I’m notified right away and we’re able to hop on it. That’s kind of like that triage process.

So I think for me it’s very interesting – I’m not sure I’m fully on board with “Universally just don’t crash.” Ultimately, if there’s memory issues, or if there’s flushes that you didn’t do, the system is gonna crash anyway. It’s gonna crash on its own, whether you explicitly exit or not. But I think for me, especially during development and testing and staging - I don’t know, I would err on trying to collect as many of those data points and making sure that whatever logger you’re using, that you have event listeners set up on process exit and crash, so that you can capture that stack and that error, so that you’re able to send it somewhere, like put it on disk, and whatever… If you have a log rotate policy, or if you’re using a Fluentd, or something like that, you’re able to capture that data.

So where does that come into play for you? …you know, the logging story, and the monitoring story, and the triaging; how does that fit into that flow? Because I think for me without a rubric there, it’s very hard to implement any type of universal guideline. At least for me.

Ron Evans: So the real reason why you guys wanted me to come on this week, besides the fact that I’m a good punner is more importantly all about drones… Specifically, the latest and greatest drone from DJI, which is called the Tello.

[12:09] The Tello is a brand new drone - it was just announced at CES this year, catching everyone by surprise. For $99, it’s a collaboration between DJI and Intel. DJI is drone navigation and piloting - or flight control rather - technology combined with some very interesting image processing technology from Intel.

So built into the Tello is a chip that is from an Intel-owned company called Movidius. It’s what’s known as a VPU (video processing unit) and it’s a chip that’s specifically designed for doing computer vision applications. In the case of the drone, it’s used for what’s called optical flow. Optical flow is when you use the images that you’re seeing in order to stabilize in this case the drone; it’s very stable, it doesn’t tend to drift around or move because it’s able to use the high-quality HD video that it’s getting, match that to its current position, and then automatically reorient itself based on the visual information that it’s getting.

So as opposed to trying to do these things with less accurate ultrasound being the typical – if you look at the original most popular consumer drone which is the Parrot AR.Drone, on the bottom are these two little circles with screens over them; that’s the ultrasonic range finder that’s used for trying to calculate the distance with the ground, and then it’s got a little pinhole camera that it tries to use in order to stabilize itself… That way it doesn’t just retain the same height, but it also retains the same position. If the thing is six feet off the ground, but it’s moving left or right at random, it’s gonna knock into somebody or something, so to have a really good auto-stabilization algorithm you need to combine different sorts of data, and visual data is the most important data.

So the Tello, for this incredibly low price, has probably one of the most advanced chips used for deep neural network processing; in this case, it’s being used for the optical flow and stabilization algorithms that are part of the Tello. So kind of a long explanation, but it’s amazing how easy it is to use, because… Basically, when you try to fly radio-controlled helicopters in the last century, the first thing you did was crash your helicopter into a tree, because they’re very hard to keep stable… Whereas these days, lots and lots of people go out and buy a consumer drone, and within a very short time of taking it home or to the office, they’re able to fly and generally not crash into things, and that’s to a large extent because of the kind of computer vision and software and hardware that’s built into drones these days. So the Tello thing is I think down to $99 and you’ve got all this stuff built in.

Gerhard Lazu: Yes. And the reason why we know this is from the documentation. They have certain, for instance – it’s like, one behavior that we had to work around… As you can see here, we have different instances running. So we have Pipely running, which is Varnish. That’s what – it’s like Varnish 7.7.0. But we have feeds, and feeds is - the TLS proxy, we talked about it in the last episode. The TLS proxy terminates TLS to backends. In this case, HTTPS traffic. Varnish itself cannot go to TLS backends. It doesn’t terminate SSL. Varnish Enterprise does. And the reason why I know that is because that’s what we use in the FastAVCL config. So Varnish in that case does terminate TLS. And that is a Varnish Enterprise feature only.

So that was like another thing that we had to solve somehow… And Nabil - again, thank you very much for helping out with that… Writing this very simple Go proxy, which uses little memories, is highly performant, and is able to terminate SSL, which then in this case Pipely connects to. And it’s all running locally. So feeds, assets, and app - they’re separate processes, and we can see this by… Let’s just do this. Look at that. This is like the whole process tree of what’s running in Pipely.

[01:15:59.25] So we have tmux… Obviously, that’s the session which I have opened here. Bash [unintelligible 01:16:04.07] It’s just a wrapper. It basically invokes Gorman, so it runs like all the various processes. And we have TLS exterminator, local port 5000, proxies to Changelog Flydev. We can see the process, we can see the memory usage, all of that. It’s using currently – what is it? Eight megabytes of memory. And that was asked of benchmarking. We ran a benchmark here.

TLS Exterminator, we’re going to feeds, and we’re going to… Which was the other one? There should be one more there. Changelog [unintelligible 01:16:37.26] the static assets. And then, eventually, we have Varnish.

So you have quite a few things running here just to get that experience that in Fastly’s case it’s just all part of Varnish. So we are bringing different components together, building what we’re missing, so that we get something similar. And ultimately, what we care is how the system behaves from the outside. Do the users get the experience that we want them to have, or that we expect for them?

Alright, so… I could do this live, but I think it’s easier. I can focus a bit better. So the tests - we can run them locally. Now, I did mention that we’re using Dagger. So if I do dagger login changelog, what that means is I’m going to authenticate to Dagger cloud. And then everything that runs locally will be sent the whole telemetry. Like, how the behavior of the various commands, how do they change? In this case, I’m running the acceptance test locally. And by connecting Dagger to Dagger Cloud, I’m able to see all the different things that run for those acceptance tests. All the commands that get installed, all the tools that get installed, all the commands that run… And in this case, I can even see the actual requests that go to the local instance of Varnish in great, great detail. It’s all real time, it’s all Wasm goodness… And the tests are hooked up, too. So when I run something locally, or even in CI, it all goes to the same place, and I can understand how these various components behave. How long do they take? That’s what we see here, like a trace, or the various steps. So when something is slow, or it misbehaves, I know where to look.

So the acceptance tests, they run locally in one minute and 26 seconds, and that’s pretty good. So what else is left? We’re nearing the end… What else is left before we can deliver this toy to Adam? That’s what we are working towards. So the first thing is the memory headroom. What does that mean? Varnish, we are configuring it to use a certain amount of memory, so that it can serve as many things as it can from memory. So it’s really, really fast. And I went through a couple of iterations basically, and we’ll see that in a minute… The value which I set initially was not the right one. Varnish kept crashing and I had to find out what the right value is, because it’s not very obvious.

Forwarding logs, that is the part which – I think it’s an important one, but it’s a smaller component compared to everything else. So we will have one more process running. In this case, it’ll be Vector. And Vector is going to consume all the Varnish logs, and it’s going to deliver them to different sinks. That’s what they’re called internally.

So one will go to Honeycomb, and we’ll be able to compare “Is the data the same format as we get from Fastly?” Because all the dashboards, and all the alerting and everything else should work the same. The SLOs and all of that. And are we able to send the same logs with the same format to S3, so that Jerod is able to process the metrics? And that is the important part. When you mentioned that the numbers went down - well, we’re not getting those metrics from the new instance.

Adam Stacoviak: I always felt like GitHub Auth was – anything I authorized with GitHub, it always felt like whatever was being asked of the authorization process was more than I thought was necessary in the authorization process. In almost every case.

Till Döhmen: I mean, one way to describe DuckDB is it’s the SQLite for analytics. So it is basically a very easy way, a very developer-friendly way to achieve what you just described. If I want to create a demo for my new text to SQL model, if I use DuckDB for it, I can even make a completely WASM-based demo out of it, for example. I don’t have any issues with CSV upload. There might be databases where I have to specify the limiter of the file that the user uploads… So I would have to show a dialogue to my user where it says “Oh, that’s comma-separated and it has a header row”, and so on. With DuckDB, it just works. So it takes away some of the edges you might have with other databases.

And on top of that, as you said, it integrates with different storage backends. It can read from S3, it can read from HTTP… When I see an interesting file on, let’s say Hugging Face or GitHub, I just run read CSV from this URL, and I have the dataset locally in my CLI or in my Python.

Furthermore, when I have, say a Python environment, I start a Colab notebook, and I create some data frames. Then with DuckDB, I can just read those data frames. I’ve seen very cool demos of people basically using Text-to-SQL for analytics on Pandas data frames. And under the hood, it’s just DuckDB sitting there, and basically reading straight from those Pandas data frames, which by the way is one of the other benefits of shared memory of in process.

[00:28:17.10] It’s not only for fetching results, it’s also for reading data straight from the process, so in that case from Pandas. That’s very exciting. I’m happy to talk more about Text-to-SQL. We have had a project about that at MotherDuck. But yeah.

Jakub Žádník: No, no. Like, something that outputs null instead of a value. So it unwraps the value, or returns if it’s a null, or something like that. I remember someone was doing [unintelligible 00:37:28.08] the ampersand you use in Bash for a background process. And we were trying to hack putting processes in the background using this ampersand sign… But we don’t have background processes in Nushell, so I don’t know how they did it, or how successful was it, but…

Josh Padnick: Yeah, that’s a great point. So one of the – there are a few kind of core tenants of our vision, and one of them is, of course, a clear, unambiguous open source license. Another one is backward compatibility. And what that specifically means for us is that any bug fixes that are reported, either to us, or that we see to TerraForm, we would like to implement fixes for. I should also say, it’s extremely important to the group to remain compliant with HashiCorp’s license. So we cannot just go to their repo and copy code improvements, just because they’re available. We are still working with our lawyers to figure out what information can we legally glean from what happens there, and where do we need to just not even look. So legal compliance is something that’s really important to us.

Anyway, to your point about compatibility… So that’s bug fixes. Then for new features that Hashi TerraForm releases, like I just said, we cannot just copy their implementation. But our understanding is that we can copy the interface, and we can do our own implementation, and make it interface-compatible. So to the greatest extent possible, we want to make OpenTF a drop-in replacement for Hashi TerraForm. And then once we feel like we’ve really achieved that expectation, where people feel like “Hey, this project is pretty stable. I can use this, I’m happy with it”, the other kind of big thread that we have is a community-driven RFC process for whatever improvements people want to see in OpenTF that could never make their way into HashiCorp. One example would be those encrypted state files, like I mentioned before. HashiCorp never wanted to merge that PR, and now this is an opportunity for OpenTF to figure out the right way to do encryption on state files, what all the potential pitfalls could be, go through a proper RFC process, get that merged, and now it’s a superset of what HashiTF would be.

Another example would be better importing from the Go codebase. So HashiCorp also made this sort of pivotal decision a while ago to lock down a lot of the internal libraries of TerraForm under a specific package in Go, so that they were difficult to access… And that meant that you couldn’t really use TerraForm as a library, you could really only use it as a CLI tool. And for those of us who are building things like pipeline solutions that operate TerraForm, or TerraGrunt, or whatever, we have to go through this awkwardness of using the CLI tool, when what we’d really like to do is just access the Go library directly.

[38:20] Another example would be downloading providers from any registry. So right now, HashiCorp forces you to download providers from their registry. And if you’ve written a custom provider, like maybe you’ve got some custom API, and you want people to be able to write TerraForm code for your custom thing - this understanding I need to check, but my understanding is that you can’t host that provider except with HashiCorp’s paid services. I might be wrong on that, so I definitely need to check on that. But in either case, with a community-driven process, we can now allow users to download their providers from any registry. In fact, someone from our consortium released a proof of concept video just a couple days ago showing how they were able to download TerraForm providers from Docker Hub, or ECR, or GitHub Container Registry.

So our first priority is compatibility, so that you can just do a drop-in replacement for Hashi TerraForm. But once we feel like we’ve really met that expectation, then we want to start focusing on our own set of innovations around this community-driven RFC process.

Daniel J. Barrett: Yup. So now you have the concepts for that, and it’s one of 15 or so ways I mentioned of running commands: command substitution, process substitution, piping a Bash, pipelines in general, plain old commands, bash -c, and there’s seven or eight others that once you know them, you have that flexibility that I was mentioning, too.

Daniel Whitenack: In your experience - I mean, I know you’ve been likely talking to tons of different clients, different use cases that are really kind of thinking about how their workflows could adapt to the serverless workflow… I’m just thinking about my own workflows; we’re running a lot of models, but none of our models on my team are like receiving thousands of inferences per second, or something like that. It is very much in the zone where we kind of had a burst of activity, and then we’re kind of down for a bit, not getting that much, and then maybe another burst that we need to process. So in that case, I would probably be willing, in my own use cases, to put up with somewhat longer of a cold start, like response for the model when it comes up, and then subsequent ones during that burst being much faster. What have you noticed with clients? What is the tolerance there? Where are you trying to get, and where do you think is reasonable for most workflows, I guess?

Adam Stacoviak: So in like a pull request then with this kind of flow, would you - and Jerod, in our case… I’m looking at PR #400, “Move assets to S3”, the most recent one… So in this case – and you may not have this as a reference, but it’s just a pull request and you can kind of see the atomic commits over time from Jerod and Gerhard essentially committing code to our repository that moves our assets from a local disk to S3 as part of the process. So in this case, would you then just – with all those atomic commits over time, by going the reset route, would those be gone?

Adam Stacoviak: I’m calling myself out on it. That’s interesting… The predefined dig times is pretty interesting, too. I guess at this point it’s meant to be educational and exploratory, so a real mission might be actually having sensors for knowing if you hit bedrock, or certain objects that prevent you from actually digging. There might be some more smarts behind the digging process, whereas in this case it’s probably a little less smart; I was gonna say “dumb”, but that’s not cool… Less smart. It’s just gonna go and dig somewhere, wherever the X/Y coordinates lead you to, for a time, collect it, and drop it back into a hopper, which goes into maybe a – what did you call the thing before? The processing plant, or…?

Jason McGee: In the case of Istio, the process is documented; there’s links on the community page on how to contribute in different working groups, there’s links in there about the Technical Oversight Committee Working Group, which is the kind of overarching technical group… And then part of kind of moving forward is – and obviously, this has evolved over time, but… You know, how do we get to voting procedures, and what’s the formal governance rules, like the charter, if you will, and that’s been kind of evolving over time.

I think at Istio there’s no confusion about like “This thing is gonna run the right way.” We already use the CNCF code of conduct… It’s gonna have the right blend of who gets to participate. That’s pretty important to us at IBM, for sure, so we’re not gonna run a project where somehow we have undue influence.

The challenge sometimes is getting projects over the hump, like getting them birthed into the world…

Adam Stacoviak: So that step there of unsealing it - I believe it’s step three in this interactive demo… It was just kind of scary language, like “Vault does not store the master key” as you mentioned, without at least one key, which is they give you the key1 - it’s one single key, because that’s what you said to do in the init process, which was “How many key shares you said equals one in this demo?” You could have said five or three, as you mentioned in your example there, but it was kind of scary when it said, “Your Vault remained permanently sealed.” I was just thinking, “Where do I keep these keys?” Sure, I can break them up into shares and give you one, give Jerod one, give other people one that maybe are part of my organization, to have access to Vault, to seal or unseal it. But I was thinking, so this primary key, since in this case in the init process was just one single share - there’s only one key, so where do I keep that?

Daniel Whitenack: [00:29:45.22] And I would say for those out there listening, in this episode we’ve talked a lot about parameter-efficient models, whether it be the Qualcomm AI models, or the 1-bit LLMs, or YOLO, and running these on device, and at the edge… It might be natural to think “Oh, the news cycle has totally switched to local models, running all the models locally, and that’ll solve all the problems.” And I think the reality is in the future it’s going to be kind of both/and. You’re not going to serve – let’s say that you integrate a model into some social media application, or whatever mobile application, or you’re serving a web app, and it’s got some AI integration, or something like that. It’s very unlikely, I think, that you’re gonna want to serve up millions and millions of requests using only local models… And in the same way, if you’ve got an enterprise batch use case, and you want to process 1.5 million documents through a large language model, you likely don’t want that running on your Mac M2, or something like that. That’s not the deployment strategy for that scenario. But yet, you will see a lot of models running at the edge, or locally. And I think the reality is that we’ll go into kind of a both/and sort of scenario, where yes, a lot of things you’ll be able to run locally… But the same as like – I mean, you can run a lot of software locally, but it doesn’t mean that you’re also not running software in the cloud. You know, AI is just a new layer in your kind of software stack. So we’re gonna run it locally, and we’re gonna run it in the cloud.

Sarah Drasner: For sure. Although I think yes and no… Like, just like in software engineering, there are some things that are true-ish, that could be applied to many things… There’s software that can be very brittle, that you probably don’t want to write. And I think that there’s stuff in engineering like that, too. There’s a lot that – and I think the book does try to generalize the things that are a little bit more – you know, the stuff that could be useful for more people. There’s a bunch of chapters where your mileage may vary. I talk about what it’s like if you try to prioritize between PM and Eng projects, and things like that; that’s very “Your mileage may vary.” That’s just in case you don’t have a process setup, or things of that nature… Like setting up your calendar. You can set up your calendar however you want, but here’s some things I do.

Grant Seltzer Richman: No, you could actually change things, and you could make responses. There’s certainly limitations, because safety is certainly a concern… You don’t wanna just be able to put anything into your running operating system, especially in production… But yeah, there’s a lot you can do. You can take action, you could prevent a process from occurring… In the case of routing networks, you can re-route packets as you please…

Jon Calhoun: Only because I’ve dealt with – early when I was founding a startup I used Heroku and Rails, and the issue we ran into was just that we had a lot of web requests coming in that basically we had to go ping other servers and wait for them, and Rails was just not good at that… And the whole way Rails works is there’s one server per request, and you can spin up multiple servers in a single Heroku Dyno… But it’s still really slow as a result. Whereas a single $5 could do what $60 or $80/month was doing on Heroku with a Rails setup. It was just one of those things where just our specific use case, because it wasn’t heavy processing, it was just waiting on stuff…

So there are some cases like that where I’ve definitely seen, at least in my side of it, I’ve seen that the Go stuff is much easier for me, scaling-wise. Josh, when you guys are talking about building, another case that I’ve seen a lot is on-premise stuff. I don’t know if any of you guys have to do on-premise, but I’ve seen several companies that when they’re like “When we have to go on-premise, it’s nice with Go.”

I’ve heard other startups in the past talking about like when they’re shipping something on-premise that’s in like Ruby or Python, they’re looking for ways to obfuscate their code, which all of a sudden is a whole other ball game… Whereas with Go, I don’t think you really have to worry about that as much.

Alright, so Adam - now that I’m done talking; sorry - when you guys were choosing Go, was performance a factor, or did hosting or any of that stuff come into play?