monus

You may wonder whether I tried asking an LLM for help or not. Well, I did. In fact it was very helpful in some tasks like summarizing kernel logs [^13] and extracting the gist of them. But when it came to debugging based on all the clues that were available, it concluded that my code didn't have any bugs, and that the CPU hardware was faulty.

This matches my experience whenever I do an unconventional or deep work like the article mentions. The engineers comfortable with this type of work will multiply their worth.

Along the way I have developed a programming philosophy I now apply to everything: the best software for an agent is whatever is best for a programmer.

Not a plug but really that’s exactly why we’re building sandboxes for agents with local laptop quality. Starting with remote xcode+sim sandboxes for iOS, high mem sandbox with Android Emulator on GPU accel for Android.

No machine allocation but composable sandboxes that make up a developer persona’s laptop.

If interested, a quick demo here https://www.loom.com/share/c0c618ed756d46d39f0e20c7feec996d

muvaf[at]limrun[dot]com

Agreed. I don’t know if it will create or eliminate jobs but this is certainly another level from what we’ve seen before.

Since last 2 months, calling LLMs even internet-level invention is underserving.

You can see the sentiment shift happening last months from all prominent experienced devs to.

Well, we are serving latency sensitive remote control to <one of the biggest banks in US> via WebRTC which uses TLS over TURN so you get 443 HTTPS for the whole traffic.

No NAT, no UDP, just pure TURN traffic over Cloudflare TURN with TLS.

We’ve recently had a disk failure in the primary and CloudNativePG promoted another to be primary but it wasn’t zero downtime. During transition, several queries failed. So something like pgBouncer together with transactional queries (no prepared statements) is still needed which has performance penalty.

We’re building the Browserbase for mobile - Android & iOS at agentic scale with no concurrency limits.

We run them on bare metal without VM brittleness, fully GPU-accelerated with WebRTC streaming using hardware encoder. As good as it gets and it’s amazed every single person who tried it.

Still behind waitlist, give me a heads up at hello@limrun.com to try it out.

https://lim.run

The problem with old computers isn’t that they’re slow but fail randomly so they don’t need “smaller” Linux, they need more resiliency that can work with random RAM erros, corrupt disks, absurd CPU instruction failures.

The size was a 90s problem.

Agreed. Use LLM all you want to do the discovery and proof but do not use it to replace your voice. I literally can’t read, my brain just shuts off when I see LLM text.

The hard part is the content of isMalicious() function. The bots can crash but they’d be quick to restart anyway.

bringing HTTP/2 all the way to the Ruby app server is significantly complexifying your infrastructure for little benefit.

I think the author wrote it with encryption-is-a-must in the mind and after he corrected those parts, the article just ended up with these weird statements. What complexity is introduced apart from changing the serving library in your main file?

Well, yeah, “magic link” is a thing and one of the easiest form of authentication supported by many providers, like Supabase, Vercel and libraries like Next Auth.

Another great side effect is that your backend doesn’t have to store user passwords which means removal of a lot of compliance headaches.

If I would knew some killer feature, I'd go ahead and write it here. But that's the thing, I can't think of any.

Speed to open and general snappiness. Nothing comes close to Zed especially in larger codebases as most agree in the thread.

TCP connections are identified with source IP:port and target IP:port tuples. When a new pod is created, it gets a new IP so there is not much way to restore the TCP connections. So crik drops all TCP connections and lets the application handle the reconnection logic. There are some CNIs that can give a static IP to pod, but that’s rather unorthodox in k8s.

I built crik[1] to orchestrate CRIU operations inside a container running in Kubernetes so that you can migrate containers when spot node gets a shutdown signal. Presented it at KubeCon Paris 2024 [2] with a deep dive for those interested in the technical details.

[1] https://github.com/qawolf/crik

[2]: The Party Must Go On - Resume Pods After Spot Instance Shutdown, https://kccnceu2024.sched.com/event/1YeP3

They're conducting a Phase II trial now with 85 participants, the completion date is set to November 2022.

https://clinicaltrials.gov/ct2/show/NCT04375657

We have implemented Terrajet (https://github.com/crossplane/terrajet) that generates CRDs that's completely compliant with Kubernetes Resource Model; references, label selectors, spec/status, no TF state blob saved anywhere.

Check it out!

‪That’s not an architecture diagram though, so it doesn’t represent the complexity at all.

I’m sure a troubleshooting map for bare linux server wouldn’t be less complicated than that.‬

I wonder if this will end up SSO providers selling this privilege to ad companies by allowing access.

Wouldn’t you need more than 12 patients to conduct such experiments?

In Crossplane, we have a similar notion for your apps deployed to Kubernetes. Platform builders would define the schema definitions for their API and then they would create several `Composition` objects, each representing a different way of satisfying the same API. Then app developers include a claim YAML together with their artifact. The `Composition` to be chosen is decided via selectors or by platform builder.

This way you can include generic `AcmeMySQLInstance` (think of it as interface in Golang or protocol in other languages) and then one of the Compositions is selected to satisfy this requirement: `AWSRDSInstance`, `AzureSQLServer` or `GCPCloudInstance` or some in-house configuration. The protocol between app dev and builder is the schema and the credential object (`Secret`) that is being published.

See details in docs: https://crossplane.io/docs/v1.0/getting-started/package-infr...

Disclosure: I'm one of Crossplane maintainers.

The thing I don't quite grasp is that QEMU does the translation in runtime, i.e. emulation, but why not do it once and use the translated binary?

Thanks for the explanation!

Different CPUs have different memory consistency models for parallel threads, leading to race conditions when translating

Does the race condition happen during the translation or execution? If it's former, we can somehow prevent it, right? The translation code is known.

Anything with a JIT compiler (including most web browsers) needs special handling.

I guess this is something Rosetta 2 does.

From what I can tell, the listed reasons are mostly alongside how hard or challenging it would be but there is no actual technical limitation. It's just no one really made the effort, I guess because of QEMU since it works for most cases albeit with a performance penalty since it's emulation rather than translation.

Do you think it's feasible to implement a translation software that similar to Rosetta 2 but open source and covers other ISAs like RISC-V? I understand that it's hard but it's also surprising that while we see the barrier for different ISAs to be the compatibility but at the same time we settled for only runtime emulation.

I've been using Firefox for more than a decade now. In fact, I switched from IE to Firefox and never made Chrome my main browser. At times, it was difficult as Firefox was really slower than Chrome before Quantum release but I think in recent years it's gotten better. Why did I keep using it? Privacy and open-source.

But after reading this, I think I will switch to Safari. It'll be a tough change after all these years but I believe Apple is more capable of enabling me to keep my privacy more than Mozilla Foundation. I don't regret the donations I made there but it's really sad to see that I paid that money to that CEO guy.

I agree. There are also cases where many Kubernetes clusters in different regions consume the same cloud service like RDS Cluster. So, tight coupling of cloud service (assumed to be stateful in most cases) and workloads doesn't work in some scenarios.

In Crossplane, our primary scenario is to have a dedicated small cluster for all the orchestration of infrastructure and many app clusters that will consume resources form that central cluster. So, clusters with workloads come and go but the one with your infra stays there and serves as control plane. See the latest design about this workflow: https://github.com/crossplane/crossplane/blob/master/design/... I'd be happy to hear your feedback on the design.

Disclaimer: I'm a maintainer at Crossplane.

I think “restartability” comes down to what you lose when you restart it. In case of pods, you lose logs and in case of StatefulSets or volume mounted pods, some kind of state. Apparently, Kubernetes decided that this is too much to lose if they automated it.

But I agree with OP that it should be optional to restart them automatically since in a lot of cases it’s simply a state problem. Is that a problem needs debugging so we need to keep some traces? Yes but the same goes for container restart, too. So there is a tradeoff and I can see that it’s not much different than the tradeoff we make for container restart. Write the logs somewhere before the restart for example.

I have resisted using anything other than Firefox since the IE days but I have been literally forced to switch to Chrome because of this issue on my work Macs and every time I checked, I was surprised to see that’s not implemented. Finally. Thank you Mozilla!

I’d be interested to see if the author will get as much exposure to new potential subscribers without a platform like Substack of Medium promoting it. IMO, it’s always been the community aspect of these platforms that keep the authors locked in rather than the technical difficulties of setting up your own website.

I have not encountered any other psychology concept that is so ingrained to almost all people.

Literally, think of every discussion you had related to politics.

Oh, thanks for letting me know!

I think that'd be a bit challenging because Config Connector is highly opinionated, for example, Kubernetes Namespace corresponds to GCP project. Though it might be enabled to be used as part of a Composition when we support namespaced CRs to be used as composition member.