Badabinski

it was a form from Google soliciting feedback on the thing.

Lovely, thank you for this. I’ve left my feedback, and I hope many, many other people do as well.

It also lacks any form of dependency management AFAICT. I don’t think there’s any way to say you depend on another service. I’m guessing you can probably order things lexically? But that’s, uh, shitty and bad.

I wrote and maintained a lot of sysvinit scripts and I fucking hated them. I wrote Upstart scripts and I fucking hated them. I wrote OpenRC scripts and I fucking hated them. Any init system that relies on one of the worst languages in common use nowadays can fuck right off. Systemd units are well documented, consistent, and reliable.

From my 30 seconds of looking, I actually like nitro a bit more than OpenRC or Upstart. It does seem like it’d struggle with daemons the way sysvinit scripts used to. Like, you have to write a process supervisor to track when your daemonized process dies so that it can then die and tell nitro (which is, ofc, a process supervisor), and it looks like the logging might be trickier in that case too. I fucking hate services that background themselves, but they do exist and systemd does a great job at handling those. It also doesn’t do any form of dependency management AFAICT, which is a more serious flaw.

Nitro seems like a good option for some use cases (although I cannot conceive why you’d want to run a service manager in a container when docker and k8s have robust service management built into them), but it’s never touching the disk on any of the tens of thousands of boxes I help administrate. systemd is just too good.

Just journalctl | grep and you’re good to go. The binary log files contain a lot of metadata per message that makes it easy to do more advanced filtering without breaking existing log file parsers.

I’ll agree that list comprehensions can be a bit annoying to write because your IDE can’t help you until the basic loop is done, but you solve that by just doing [thing for thing in things] and then add whatever conditions and attr access/function calls you need.

Anubis has worked if that’s happening. The point is to make it computationally expensive to access a webpage, because that’s a natural rate limiter. It kinda sounds like it needs to be made more computationally expensive, however.

Do you have any sources for the 10x memory thing? I’ve seen people who have made memory usage claims, but I haven’t seen benchmarks demonstrating this.

EDIT: glibc-based images wouldn’t be using service managers either. PID 1 is your application.

EDIT: In response to this:

There’s a reason a huge portion of docker images are alpine-based.

After months of research, my company pushed thousands and thousands of containers away from alpine for operational and performance reasons. You can get small images using glibc-based distros. Just look at chainguard if you want an example. We saved money (many many dollars a month) and had fewer tickets once we finished banning alpine containers. I haven’t seen a compelling reason to switch back, and I just don’t see much to recommend Alpine outside of embedded systems where disk space is actually a problem. I’m not going to tell you that you’re wrong for using it, but my experience has basically been a series of events telling me to avoid it. Also, I fucking hate the person that decided it wasn’t going to do search domains properly or DNS over TCP.

Debian is superior for server tasks. musl is designed to optimize for smaller binaries on disk. Memory is a secondary goal, and cpu time is a non-goal. musl isn’t meant to be fast, it’s meant to be small and easily embedded. Those are great things if you need to run in a network/disk constrained environment, but for a server? Why waste CPU cycles using a libc that is, by design, less time efficient?

EDIT: I had to fight this fight at my job. We had hundreds of thousands of Alpine containers running, and switching them to glibc-based containers resulted in quantifiable cloud spend savings. I’m not saying musl (or alpine) is bad, just that you have horses for courses.

Is it? I thought the thing that musl optimized for was disk usage, not memory usage or CPU time. It’s been my experience that alpine containers are worse than their glibc counterparts because glibc is damn good. It’s definitely faster in many cases. I think this is fixed now, but I remember when musl made the python interpreter run like 50-100x slower.

EDIT: musl is good at what it tries to be good at. It’s not trying to be the fastest, it’s trying to be small on disk or over the network.

The one where every LLM-generated shell script I read is another deep splinter in my fingernail quick that I have to rip out and destroy because it’s a godfucked mess of bad practices that we can never ever ever ever EVER train out of an LLM at this point.

Open source isn’t good enough, I want my software to use a strong copyleft license with no ability to relicense via a CLA (CLAs that don’t grant the ability to relicense software are rare, but acceptable). AGPL for servers, GPL for local software, LGPL for libraries when possible, and Apache, MIT, or BSD ONLY when LGPL doesn’t make sense.