Nicolas Martyanoff

Contrarian software engineer. Hire me to solve your technical problems.

Public Key

npub1dkg70vzk4vvfsc5cup9gkfu2tcrs5unysl7npfvegf78nkuswr5skh54dt

NIP-05 Address

galdor@fosstodon-org.mostr.pub

Profile Code

nprofile1qqsxmy08kpt2kxycv2vwqj5ty799upc2wfjg0lfs5kv5ylremwg8p6gpz4mhxue69uhhyetvv9ujuerfw36x7tnsw43qqpm2aq

Publishing to

relay.ditto.pub

Author Public Key

npub1dkg70vzk4vvfsc5cup9gkfu2tcrs5unysl7npfvegf78nkuswr5skh54dt

Show more details

Last Notes

2026-05-08 09:49:31 UTC

An interesting aspect of having hand issues and switching to a column-staggered split keyboard (Corne) is that it made me realize how bad #Emacs key shortcuts are. Way too many key combinations forcing finger contortions all day long. Modal edition is clearly superior.

2026-04-29 10:10:11 UTC

I'd happily move out of GitHub for my own projects (my private ones are already somewhere else), but I'm not exactly keen on having various bots and AI agents abuse Git endpoints on my server if I put my repositories there.
Maybe just use GitHub as a read-only mirror.

2026-04-11 17:30:00 UTC

Lots of fun recently with text rendering, without OpenGL this time. Still Common Lisp, SDL3, FreeType2, Harfbuzz and FontConfig. Here rendering one of its own code file.
Rendering the cursor made me realize the need to split grapheme clusters myself. One day.
https://cdn.fosstodon.org/media_attachments/files/116/387/256/161/218/344/original/89309b81b8d8bdb6.png

2026-03-19 18:43:43 UTC

OpenAI buying Astral (Uv…) is exactly why I’m reluctant to use open source software written and maintained by companies. At some point they need to make money, and this kind of thing happens.
Also remember what happened to Terraform, Redis…

2026-02-22 13:02:14 UTC

The beauty of language standards is that I can ignore everything post C99 and have something that will always work, even in 10 or 20 years. Same thing for Common Lisp.
Without it, C will be ruined.

2026-02-21 18:23:33 UTC

My #CommonLisp reader is still in its infancy but it already supports conses, bigints, strings, characters, enough to play. So I started the printer; like the reader, it is very well specified by the standard. Going to take a lot of time to support all object types, but it's a start.
Then of course I'll see what can be done for evaluation.
A really good way to learn a lot about CL details, strongly recommended.

2026-02-19 11:35:33 UTC

- reply

[2/2]
Because no, you don't want to have to deal with creating accounts in advance, provisioning them to servers, etc. The very point of ACME is not to have to deal with that.

2026-02-19 11:35:22 UTC

[1/2] DNS-PERSIST-01 for the ACME protocol is disappointing: you still have to know the account id to provision the record, so you will still have to programmatically alter DNS record when each server automatically creates an account the first time they start.
Still the exact same problem as DNS-01, so back to HTTP-01 and no wildcard (but are wildcards really that useful with ACME anyway?).

2026-01-20 17:39:56 UTC

Any language or library that uses the term "filter" for a method is ill-designed: you have no way to know what it does! "select"/"reject" is good, #Ruby got it right. In #CommonLisp, REMOVE/DELETE are sort of ok, but there's no function for selection.

2025-12-22 11:55:01 UTC

After moving from Common Lisp to #C, I'm done updating my graphic programming experiment to #OpenGL 4.5 core. Pure DSA (Direct State Access) functions, VAO/VBO/IBO, GLSL shaders. Still no subpixel rendering but for the time being it's good enough.
And the screenshot made me realize that I need to adjust vertex positions for the bounding box.
https://cdn.fosstodon.org/media_attachments/files/115/763/084/655/538/689/original/2f661c16ece49894.png

2025-12-05 11:53:48 UTC

- reply

Selling commercial software is 100% fine. Pretending your non-free software is Open Source because you want to have your cake and eat it is not.

2025-12-05 11:53:44 UTC

The only reason 37signals is hell-bent on proclaiming that Fizzy is Open Source software (it's not) is that they want the upsides: people working for them for free (also called "contributions"), exposure, the ability to pretend that this is a gift. But they do not want the downsides, i.e. anyone able to use it to compete with them. Ultimately the goal is to bring people to the hosted version which of course is not free.

2025-12-03 21:53:08 UTC

- reply

@nprofile…f3ll MinIO was free as in free beer, so I imagine you'd be right that users got exactly what they paid for, which is nothing (unless someone forks the last version available).

2025-12-03 20:03:56 UTC

MinIO moving their software to their new proprietary enterprise platform is the perfect example that no, the AGPL license does not protect users against rug pulling.

2025-11-26 12:18:53 UTC

- reply

@nprofile…9x5z Which ones? Pure curiosity, since I got fed up with ASDF and ended up writing my own system loader.

2025-11-24 14:50:49 UTC

- reply

@nprofile…c3qk I have Lisp System Implementation. I really appreciate the no-nonsense style: it's about building things, not pondering theory for hours. Really good work.

2025-11-19 15:44:59 UTC

While everyone is discussing about how Rust and unwrap brought down Cloudflare, I'm more interested about how a single mistake in a single program affected the entire infrastructure. This is not a software issue, it is an architecture one.

2025-11-16 17:10:39 UTC

I have started porting my Common Lisp graphic experiments to C because the GC is a deal breaker for soft real time programs. First time seriously using C in years, but it was less painful than I expected.
C99 + POSIX extensions, SDL 3, Freetype. GNU Make only (don't even try to mention CMake to me). ASan and MSan are a game changer, I do not even use Valgrind anymore.
Part of the fun is having to write a lot of low level code to be productive.

2025-11-12 08:53:20 UTC

One of the side effects of the ubiquitous nature of SSO is that you cannot have software supporting it without a web browser. Not exactly progress given the nightmare of complexity and lack of security they come with.

2025-11-11 18:13:58 UTC

If you are using clangd as LSP server, note that it does not support relative include directories. Unless you create an empty compile_flags.txt file along your .clangd configuration file. Because reason. Of course there is a bug report about it since 2022. No comment.

2025-11-10 19:21:35 UTC

- reply

@nprofile…mdnm Neither CLtL2 nor SBCL are the specification. CLtL2 is a pre-standard description of the language; it is very close to the standard, but not identical. SBCL is an implementation, meaning that its developers picked a behaviour. It's a perfectly good one (and I believe it depends on the safety level applied to your code).

2025-11-10 18:49:28 UTC

- reply

@nprofile…mdnm THE is not really about checking types at runtime, it is about specifying that the values returned by a form are of a specific type. And the consequences are undefined if they are not. Think "type declaration for the return values of a form", the same way as (DECLARE (TYPE …)) is about specifying types for variables in a lexical environment.
If you want safe runtime checks, use CHECK-TYPE.

2025-11-03 23:40:29 UTC

I'm really impressed by the way Ghostty is developed, with a deep attention to details and correctness. But if I install it on Linux and run it, it uses more than 300MB RES without doing anything. v1.2.3 based on Zig 0.14.1. So I'll stay on xterm.

2025-10-22 18:25:41 UTC

- reply

@nprofile…e4jd If you find one I'm interested. I just use it to simulate S3 buckets in Docker Compose stacks for development/tests.

2025-10-22 17:54:33 UTC

MinIO stopping distributing Docker images even though the last one is apparently affected by a security vulnerability is a funny example of open source dynamics. Yes, MinIO does not owe anyone anything. 100% true. But they just tanked their reputation in one move.

2025-10-19 19:56:18 UTC

- reply

@nprofile…efa6 This is a really small world! I was using with url-retrieve & co precisely to experiment with LLM APIs.
I may need to give gptel a try, though building it is 80% of the fun.

2025-10-19 00:20:19 UTC

- reply

@nprofile…efa6 I imagine I could ask or even send a patch, the code is not that complicated. I'm just not a fan (most polite way I can put it) of emacs-devel debates. It might be stupid, but I'm already anticipating having to spend time and energy debating with dozens of nobodies defending the current code for ridiculous reasons, as it is too often the case.
Unrelated; are you the author of gptel?

2025-10-18 16:29:00 UTC

In #Emacs, json-serialize is supposed to be the modern, fast way to encode a JSON value. Infortunately for some ridiculous reason it does not support strings as object keys. So something as simple as (json-serialize '(("a" . 1) ("b" . 2))) fails with "Wrong type argument: symbolp, "a"".
So back to json.el.
And it's nothing compared to the insanity of the url structure or url-retrieve.

2025-10-12 15:14:13 UTC

- reply

@nprofile…dcu3 When digging into tech-related topics with Grok, I have identified far less that 10-30% of errors. Now there's always a change that I missed some of them, but I don't dig into subjects I know nothing about, so I'm reasonably confident.
So again, it's about the error percentage. Wikipedia itself contains errors, and everyone is fine with that.

2025-10-12 14:43:50 UTC

- reply

@nprofile…dcu3 So it's all about how often the information is correct; if it's often enough, you can tolerate that it sometimes makes mistake. And you can still verify when you have any doubt of course.

2025-10-12 13:57:20 UTC

- reply

@nprofile…dcu3 If you ask a friend about something you do not understand, how are you supposed to evaluate the accuracy of their answer?

2025-10-11 21:20:14 UTC

Anyone who finds AI useless should try using one as if it was a human assistant: simply ask about what you do not understand, read the response and iterate by going deeper and deeper until everything is clear. I have been doing that with Grok for some time, and the results have been mind blowing.
I still write code all by myself, but modern LLMs are a combination of Google on steroids and a partner to bounce ideas against. I'm now convinced that this is a productivity multiplier.

2025-09-06 23:58:54 UTC

Bad default setting example: on #FreeBSD, core files are written in the working directory. Many daemons use "/nonexistent", so they cannot generate core files. So if they segfault, you won't have any idea why.
Add kern.corefile=/var/coredumps/%U/%N.core to /etc/sysctl.conf.

2025-08-04 14:58:56 UTC

- reply

[2/2] I imagine that the argument is that the semantic of directory pathname component members is implementation defined. Is "foo/bar" a valid directory component member?
I still think both implementations are sloppy: both should raise an error indicating that "/" is not a valid filename character.
P.S. Thanks for forcing me to split my post in 2 messages #Mastodon . Because everyone knows that >500 character messages are useless right?

2025-08-04 14:58:05 UTC

[1/2] In #CommonLisp, SBCL accepts directory pathname components which are subpaths and not names:
(make-pathname :directory '(:absolute "tmp" "foo/bar")) returns #P"/tmp/foo/bar/".
On CCL, the same form yields #P"/tmp/foo\\/bar/" which of course results in errors in a lot of contexts because "/" is not a valid character in filenames.

2025-07-28 09:58:16 UTC

Crazy to still see people arguing in favour of soft deletion. In B2B as in B2C, if users delete data, you must actually delete them. Not doing so is a serious liability.

2025-07-05 13:50:45 UTC

To send the logs of a specific program to a separate file with syslogd on #FreeBSD:
!*
!boulevard
*.* /var/log/boulevard/boulevard.log
First reset program filtering, then set it up for your program, then and only then add rules (here send all facilities and levels to a file).
Also note that the file won't be created by syslogd if it does not exist.

2025-06-29 16:10:17 UTC

Still more #Go horrifying discoveries: net/url is non-conforming, it will happily parse a URL where the authority does not contain a host, e.g. "http://:80". No it is not valid, RFC 3986 3.2. is pretty clear about it.
Worse, the net/http client will happily accept this kind of URL, defaulting to a local address but with an absolutely invalid Host header field set to ":80" in this example (see RFC 9110 7.2., the host is not optional).
As usual, developers do not read specifications.

2025-06-26 10:55:31 UTC

- reply

@nprofile…yzv0 Given that the Plan9 clique lead to the Go language, I'm actually happy that it failed.
The Lisp machine model made so much sense, but history happened.

2025-06-26 10:30:01 UTC

If Linux and BSD had their shit together and had built layers to deal with services, storage and networking in a distributed and programmable way, maybe we would not be stuck with Kubernetes and shitty cloud services. The UNIX paradigm stopped evolving a long time ago.

2025-06-23 13:15:41 UTC

As an engineer the most impactful resources in my career have been well articles and emails written by experimented people who took the time to really think about something and write about it.
I see a lot of engineers who do not read or write, balking in front of everything longer than a Slack message. NGMI.

2025-06-13 19:18:14 UTC

Apparently a couple developers have forked the Xorg display server under the name Xlibre. Good news: I'm not a huge fan of how Wayland is being shoved in people's throat, it reminds me too much of what happened with Systemd.
Having options is a good thing!

2025-06-11 08:33:47 UTC

I had to use a query builder recently (SQLAlchemy) and it confirmed my original opinion: they are fundamentally useless. You spend your time trying to guess the idiotic Python syntax for queries that would be trivial in SQL, and analyzing requests require copying them from PostgreSQL logs because you don't see them directly in the code.
All it does is prevent people from actually learning SQL. Terrible idea.

2025-06-10 13:10:01 UTC

- reply

@nprofile…fkdz What does it have to do with downtime or backward compatibility?

2025-06-10 11:40:01 UTC

Database migration rollbacks only make sense if you can atomically rollback your entire platform. If you can't, congrats, your database is indeed back to where it was but all past side effects remain. And you can never rollback an entire platform.

2025-04-26 22:22:25 UTC

I finally took the time to compile my notes on #CommonLisp pathnames into a proper article. I had to go back reading various sections of the standard, and I'm glad I did it: logical pathnames, among other things, are a fascinating feature, and now I want to use them more!
https://www.n16f.net/blog/working-with-common-lisp-pathnames/

2025-04-20 16:11:09 UTC

Fun Sunday project: what does it takes to define and load #CommonLisp systems? Turns out, not much. ASDF is 14k LoC, SYSDEF is 394 LoC.
Supports a minimal set of metadata, an extensible collection of components
(builtin support for static files, Common Lisp source files and component
groups), and two fundamental operations (build and load).
Works on Linux with SBCL, CCL and ECL.
That was fun!
https://cdn.fosstodon.org/media_attachments/files/114/371/163/154/824/905/original/3b6e717a8a372769.png

2025-04-20 10:27:54 UTC

- reply

@nprofile…z888 ok now I understand why you mentioned WAFs. How are you dealing with these bots right now ?

2025-04-07 00:15:32 UTC

- reply

@nprofile…mwvd your IDE might help you, but the web interface used to review code or the editor you’re using on a remote server trying to understand what’s actually deployed won’t.
Code should be meaningful on its own.

2025-04-06 13:50:30 UTC

A bad habit among #Python developers is to import names from packages ("from <pkg> import <name1>, <name2>"). It is useful for some well known and commonly used standard names, e.g. type names from the "typing" module, but it is easily abused.
It may look fine to you but it makes the code much harder to decipher because the reader cannot infer where the variable or function comes just by looking at it. And the reader can be future you, trying to debug the code at 2am on a Saturday.

2025-04-03 11:17:19 UTC

Saying in a software job interview you like no-SQL databases because you don't want to deal with a data schema actually sounds like "I have no clue about how to build quality software" sprinkled with "I'm so green I never had to deal with software where no one cared about data consistency".
Go back to the fundamentals.

2025-03-31 00:19:30 UTC

- reply

@nprofile…exnp why do you believe they are structurally equal ? :)

2025-03-30 23:20:28 UTC

#CommonLisp question: is anyone specifying dependencies for every single file instead of using :SERIAL? And if yes, how is this even used? I have never used anything but ASDF:RELOAD-SYSTEM which will reload all files anyway.

2025-03-23 23:43:50 UTC

- reply

@nprofile…qs0v No, you just made me realize that I messed up my example. The “http” block is of course a direct child of “server”, not part of a listener.
Thank you for noticing !

2025-03-23 17:06:48 UTC

I just added the simplest TLS setting ever to Boulevard: add "tls redirect", and all HTTP requests are automatically redirected to the first TLS listener. Bonus, add "hsts true" and it will do the right thing.
If you don't think it's useful, this is a comparison with NGINX:
https://cdn.fosstodon.org/media_attachments/files/114/212/838/157/765/479/original/a990a8e302d00de9.png

2025-03-23 17:03:07 UTC

I just added the simplest TLS setting ever to Boulevard: add "tls redirect", and all HTTP requests are automatically redirected to the first TLS listener. Bonus, add "hsts true", and it will do the right thing.
If you don't think it's useful, this is a comparison with NGINX:
https://cdn.fosstodon.org/media_attachments/files/114/212/823/096/562/346/original/9cc820111775230f.png

2025-03-19 21:12:42 UTC

If your lead engineer is saying "yes" more often than "no", you're cooked.

2025-03-15 23:50:38 UTC

- reply

@nprofile…utqm I honestly don’t know what to tell you, I spend forever trying to get Font Lock keywords to work with faces. Turns out the documentation is clear, you need a variable whose value is the name of the face. Fun fact, this is the case for all Font Lock default faces.

2025-03-15 22:34:14 UTC

- reply

@nprofile…qs0v I must confess than I'm used to having C-k bound to EVAL-BUFFER, same binding as the one ine SLIME for Common Lisp.
Having to move to the end of each expression to evaluate them is asinine.

2025-03-15 22:33:26 UTC

TIL that to use a face (e.g. created with DEFFACE) in FONT-LOCK-KEYWORDS (#Emacs), you need to declare a binding whose value is the name of the face. So for a face named FOO-FACE, just add (DEFVAR FOO-FACE 'FOO-FACE). This happens because Font Lock expects faces in FONT-LOCK-KEYWORDS lists to be symbol whose value is the face to use.
I would love to know why Font Lock cannot realize that the symbol is the name of a face and use it directly.

2025-03-15 19:37:08 UTC

- reply

@nprofile…qs0v EVAL-LAST-SEXP and EVAL-BUFFER behave differently. I stopped expecting consistency from Emacs a very long time ago.

2025-03-15 18:18:31 UTC

- reply

@nprofile…x0kw DEFCUSTOM is for values customizable by the user. In modules, you have a lot of values which are not supposed to be customizable (e.g. regexps used by font-locks) but must be easy to redefine when you're working on the module.

2025-03-15 17:28:47 UTC

Let us add DEFAPARAMETER to the long list of features #Emacs Lisp is missing compared to #CommonLisp. ELisp only has DEFVAR, which does not modify the binding if it already exists. Perfect to keep a state. But for a lot of top-level values, you want DEFPARAMETER, which creates or updates a binding.
Module development with DEFVAR only is really frustrating, forced to manually call SETQ each time you want to change a value.

2025-03-08 18:32:01 UTC

Generating FreeBSD packages is usually done with ports. OK if you're working as part of the port tree, not so much if you just want to provide a package.
So I've updated my Fpkg to make it easy to build packages. Compared to the standard Freebsd pkg program, it automatically generate file lists and hashes, makes it easy to specify file owners, groups and permissions, and automatically generates user and group creation scripts.
https://github.com/galdor/fpkg
https://cdn.fosstodon.org/media_attachments/files/114/128/238/297/876/978/original/24540bfe17239769.png

2025-03-08 17:53:50 UTC

- reply

@nprofile…qfh3 @nprofile…c4zu I'm curious about your experience in other languages.

2025-03-06 21:32:25 UTC

Yes #Python is a deeply unimpressive language, but in my experience it's tolerable for simple tasks as long as it's used correctly.
Most Python code out there is terrible because it's written by less experienced developers who don't read language and tool documentations and copy/paste what they find online (content itself written by junior developers), endlessly amplifying the problem.
This is the curse of languages perceived as easy for beginners.

2025-03-05 11:08:39 UTC

Small but useful improvement in SBCL 2.5.2, there is now a USE-VALUE restart when you call a structure constructor with a value having the wrong type. If you're using the REPL a lot, it's one of these small things that helps. Thanks!
#CommonLisp

2025-03-04 21:18:14 UTC

“OMG this is so cool how did you know how to make this software do that?”
I took 10min to glance through the documentation.
Happens all the time.
If you read documentation, you’ll get top 10% among developers, easy.

2025-02-28 20:20:40 UTC

The more people work on something at the same time, the more quality will gravitate towards mediocrity. Unavoidable since competence follows a gaussian distribution.

2025-02-21 16:58:57 UTC

- reply

@nprofile…8f02 What would you like to see in a Common Lisp book that has not been covered by existing ones?

2025-02-19 12:44:51 UTC

Go's encoding/json parser is fundamentally broken: it assumes that all numbers are floats (a very common mistake about JSON), meaning that if you decode without a concrete type and re-encode, integers become floats (e.g. {"a":42} -> {"a":42.0}).
The UseNumber() option helps with that, but now you have to deal with a custom json.Number type in all data processing functions.
All they had to do was to check for a '.' character in the token during parsing to identify the type of the number.

2025-02-18 11:26:00 UTC

- reply

@nprofile…8f02 If you're new to Git, yes absolutely. There's a moment when you understand how it works, and suddenly using it become second nature.

2025-02-18 11:16:03 UTC

5min in the Git history of a developer will tell you everything you need to know about their care and attention to details.
If the history is a mess, there's no need to dig further, and the conclusion is not good.

2025-02-16 14:14:44 UTC

I just finished a first version of the index view to show the content of directories with Boulevard. Useful when serving simple files.
It does not have to be complicated: add one single setting to your HTTP handler and let Boulevard deal with content negotiation. You get a nice view either in text, HTML for web browsers and even JSON because why not.
https://cdn.fosstodon.org/media_attachments/files/114/013/980/741/610/413/original/50bd03e924881461.png

2025-02-07 14:17:40 UTC

Looking at the open source software ecosystem, I'm astonished that so many developers believe it's their god-given right to have their contributions added to projects.
It get it, it can be frustrating, but unless you *own* the project, why would you think you have any right to decide anything? Discussions on the LKML are just insane.

2025-01-24 12:12:09 UTC

Obsessing over database schema rollbacks is a bad idea. They may seem useful on small systems, but the real world is distributed: you can rollback the schema but not the operations that affected other systems, so you end up with an inconsistent system.
Forget rollbacks. Make subsystems easy to quickly disable if they start acting out (e.g. feature flags), and learn to fix my moving forward.

2025-01-06 11:42:09 UTC

- reply

@nprofile…6r8y DSLs in htmx? Where?

2025-01-06 10:40:50 UTC

I do not have anything against htmx, it certainly is miles ahead from React/Vue/Angular & co. But I still think its main value is to get web developers on the path to the realization that no, they do not need a framework.

2024-12-11 19:16:59 UTC

Ok I'm fed up with YAML, it just does not work for large and complex configuration files (e.g. Boulevard: imagine an array of dozens of handlers…).
So I'm playing with a new custom format. Well not so custom, it's very similar to what NGINX has. Easier to read, without deeply nested objects, designed for programs needing more than a couple keys/values.
If I go all the way I'll write a specification and a reference library. Anyone interested?
https://cdn.fosstodon.org/media_attachments/files/113/635/792/349/072/760/original/e643085691c3f7b7.png

2024-12-06 09:27:32 UTC

I'm a bit disappointed that #CommonLisp functions such as COUNT, POSITION or FIND work for sequences but not for arrays. I would love to know the rational here, surely there is a reason.

2024-11-15 08:52:14 UTC

Fun fact, in #Go, you cannot hijack an http.Client or http.Transport connection. So if you're writing a reverse proxy and have to support connection upgrading, you have to implement the client part (to the upstream server) yourself with TCP sockets, http.Request.Write and http.ReadResponse. A bit annoying but it will give me more control anyway.