Nicolas Martyanoff

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

Public Key

npub1dkg70vzk4vvfsc5cup9gkfu2tcrs5unysl7npfvegf78nkuswr5skh54dt

NIP-05 Address

Profile Code

nprofile1qqsxmy08kpt2kxycv2vwqj5ty799upc2wfjg0lfs5kv5ylremwg8p6gpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0ds2j5krc

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub1dkg70vzk4vvfsc5cup9gkfu2tcrs5unysl7npfvegf78nkuswr5skh54dt

Show more details

Last Notes

2026-05-25 12:42:29 GMT

Rust is a good example of how human behavior affects the perception of technology. While the language may be interesting (sort of: too many issues anyway), the constant insistence that everything must be rewritten with it and everything else is bad/stupid/vulnerable is a huge turn off. Lots of people don't want to deal with that toxicity.
If I cannot tolerate a GC and/or really need to use a lot of C libraries, C99 works just fine and it's going away any time soon.

2026-05-08 09:49:31 GMT

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 GMT

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 GMT

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

2026-02-22 13:02:14 GMT

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 GMT

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 GMT

- 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 GMT

[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?).

2026-02-10 18:04:46 GMT

- reply

@nprofile…6srv I don't trust _any_ large, centralized platform. But with the CA system, there's no solution: you depends on a handful of big actors, and this is how it is if you want your software to work with existing bundles.

2026-02-10 13:48:35 GMT

LetsEncrypt planning to remove support for TLS client certificates is puzzling. No, TLS is not just for web servers, but clearly they did not get the message.
I would love to know the real reason. And we need more ACME providers.

2026-02-05 10:51:59 GMT

The standard #Go HTTP client copies Authorization header fields when redirecting requests to different hosts. This breaks redirections to S3 URI when the original request already bears credentials. Error inducing and potentially a security issue, I cannot believe it is the default behaviour of a standard library.
Use CheckRedirect in http.Client to compare hosts and strip the Authorization field.

2026-01-25 21:03:02 GMT

A fundamental requirement for the success of a software project is someone who can keep the entire program in their head, mentally simulating how it would behave for every potential change being considered.
If the project is too large, split it in logical components and enforce strict interfaces. Then you need someone keeping each component in their head.
Without it, all software becomes a big ball of mud.

2026-01-25 15:19:06 GMT

Fun fact, on GitHub you still cannot control who gets notified when a workflow failed. It is the last person who touched the workflow and it's not configurable in any way.
Of course there's a discussion open since 2023, and of course GitHub does not care.

2026-01-21 18:31:54 GMT

Everytime I see someone complain about Gnome or KDE changing in a way they do not like, I can only wonder why they use it.
Switch to a window manager designed to actually get work done (e.g. i3, Sway, Hyprland…), you don't need training wheels.

2026-01-20 17:56:56 GMT

- reply

@nprofile…knk0 Because "filter" does not tell you if it returns what was selected or what was removed by the filter.

2026-01-20 17:39:56 GMT

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.

2026-01-18 17:07:45 GMT

I've started working on the UI system (C/OpenGL). Components, layouts, background/border rendering with padding. Here with a simple label.
Next step is to deal with sizing/positioning relative to the parent to build more complex components.
https://cdn.fosstodon.org/media_attachments/files/115/917/197/239/178/643/original/69e0a0012584b2a1.png

2026-01-09 13:29:59 GMT

The Go compiler fails if the project contains a go.work file containing a different Go version. Of course it politely tells the user to run "go work use". Yet another UX failure. There's no reason not to update go.work, just do it automatically!
It is particularly annoying because it makes gopls fail. But of course you won't see the error message without reading the logs in your editor.

2026-01-09 11:31:54 GMT

- reply

@nprofile…drt7
While I certainly appreciate protocols that do not change every morning (e.g. everything Web, Cloud or AI), the email world (IMF, SMTP, IMAP) is a good example of how bad things become when they are not aggressively improved. IMAP in particular is an abomination.
Anyone who tried to implement these protocols realizes quickly how bad the situation is.
(same things for X.509. btw)

2025-12-29 16:18:45 GMT

First time I have to use DEFINE-SETF-EXPANDER. It took me some time to wrap my head around the 5 values returned, the building blocks for the expanded form of the SETF expression.
Still not 100% sure but it seems to work.
One of the most complex parts of Common Lisp.
https://cdn.fosstodon.org/media_attachments/files/115/803/758/294/991/658/original/0a5237bab0188697.png

2025-12-26 17:55:34 GMT

Fun SBCL experiment:
1/ create a type;
2/ create a class with a slot referencing this type;
3/ change the type;
4/ the slot accessor will still enforce the previous type.
No way to change it, have to restart SBCL.
So much for #CommonLisp dynamicity, especially when you just want to add a symbol to a :MEMBER type expression.

2025-12-23 16:39:14 GMT

AI usage today: I tried both Grok and Claude to diagnose a tricky memleak in a C program. Both were wrong a lot, misunderstanding the logic of fairly simple C code. But both pointed out an unrelated but very real array length bug. Grok was much better at pointing out paths to identify the real issue. I ended up finding the bug myself after following one of these paths.
Also used Grok to research vector/matrix representation in graphic engines. Much better than Google, it does not even compare.

2025-12-23 14:57:17 GMT

- reply

@nprofile…su48 I haven't, but I already had it in mind when I started.
I played a bit in Common Lisp first because it was a fun way to test my custom FFI layer, but it's just not a great choice for anything soft realtime like graphic applications: the GC is a real problem and ultimately you'll spend all your time binding hundreds of C functions in multiple libraries (OpenGL, SDL, Freetype and countless others).
I love CL, but it has limited added value here compared to C (and I won't touch C++).

2025-12-22 11:55:01 GMT

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

2025-12-06 23:59:04 GMT

- reply

[2/2]
Next step is to move my text rendering code from obsolete immediate calls to this new method. Then the whole codebase should be pure OpenGL 4.5. Fun.

2025-12-06 23:58:54 GMT

[1/2]
I spent the evening preparing utilities to deal with meshes using modern OpenGL DSA (Direct State Access) functions in C. There is not a lot of clear information about it, so I used Grok as a partner to study the topic (just for research/analysis, I wrote the code myself). Very productive. Learned a lot, ended up with a very clear abstraction for VBO/IBO/VAO and support for several vertex attribute combinations (position + color and/or texture coordinates).

2025-12-05 11:53:48 GMT

- 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 GMT

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 GMT

- 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 GMT

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.

2025-12-03 14:37:18 GMT

Quad9 is preparing to drop HTTP 1.1 support for DNS-over-HTTP, because they are "reaching the end of life for the libraries and code that support HTTP/1.1 in our production environment".
Imagine writing sensitive production systems in a tech stack so bleeding edge that you cannot even support one of the most common protocols on earth.
Everything that is wrong with modern software engineering in one line.

2025-11-28 15:59:42 GMT

TIL about the GL_ARB_debug_output which provides a user callback to have all #OpenGL debug/error messages be signaled without having to call glGetError for every single OpenGL call.
A bit chatty because errors tend to happen at every single frame, but rate limiting identical messages is easy with a hash table. Oh and it's core since OpenGL 4.3.

2025-11-27 19:17:55 GMT

Spent a couple hours adding signalfd support to my #CommonLisp event system only to realize that 1/ SBCL uses several of them (SIGUSR1/SIGUSR2) so you cannot use them for your own app and 2/ it messes with others (e.g. RT signals) so that you cannot reliably handle them. Also drops me in LDB because it is not happy I blocked SIGWINCH, while I must do that so that I can watch it with signalfd.
Lesson learn, Common Lisp and signals, not a good idea.

2025-11-26 17:06:38 GMT

- reply

@nprofile…tw26 @nprofile…9x5z
https://github.com/galdor/sysdef
I use it for several personal projects. Of course it does not do everything ASDF does, but it's kind of the point :)

2025-11-26 12:18:53 GMT

- 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 GMT

- 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 GMT

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 22:29:40 GMT

- reply

@nprofile…p7pa You can. But as much as I like Common Lisp, making it fast enough for graphic programming and non-consing requires a lot of efforts. Also for this use case it's all about calling C functions. Here CL is just more friction.
Also the default SBCL GC is not exactly incredible. There's a new GC in development that has very interesting characteristics, but I don't remember seeing any news past the first experimental version. Hopefully one day.

2025-11-16 17:10:39 GMT

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 GMT

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 GMT

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 GMT

- 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 19:10:59 GMT

- reply

@nprofile…mdnm The spec says that? Where did you read it?
Because CL 3.8. is very clear: "The consequences are undefined if the values yielded by the form are not of the type specified by
value-type."
The purpose of THE is to act as an annotation so that the implementation may (or not) optimize the code.

2025-11-10 18:49:28 GMT

- 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 GMT

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.

2025-11-03 09:24:20 GMT

- reply

@nprofile…s7zg Good to know. I'll need to run a couple checks and fill a bug report.

2025-11-02 14:35:49 GMT

- reply

@nprofile…s7zg According to Calc's documentation "the Calculator stores integers to arbitrary precision". I can't see a good reason why dividing two integers using "\" should return anything else but a correct integer.

2025-11-02 14:31:00 GMT

- reply

@nprofile…ud8d As someone who also works in the industry, I spent way too much time having to deal with shitty software built by people who think that their math skills are a substitute for software engineering skills.

2025-11-01 16:39:05 GMT

With #Emacs Calc:
-999999999999999999999999999999 1000000000000000 \
("\" is the integer division)
Yields -1000000000000000 instead of 999999999999999, and the rest is 1 (instead of 999999999999999). Correct result checked with Bc and with my own big int implementation.
Am I the only one getting this incorrect result?

2025-11-01 15:43:42 GMT

Saying that programming is math is like saying that cooking, carpentry or playing music are math. Yes math are everywhere, we get it, but the world is more than numbers and formulas.
Programming is of course about using language to describe behaviour. Being good at maths has never been a good predictor of software engineering skills. Ask anyone working in this industry.

2025-10-31 14:22:02 GMT

Claude will generate #Python code with "from prefect import task as prefect_task" to avoid a conflict with an existing "task" function instead of simply "import prefect" followed by "prefect.task".
The Python world simply does not understand namespaces.

2025-10-28 15:38:24 GMT

With recent OpenSSH versions, you'll get a useless 3 lines warning about "not using a post-quantum key exchange algorithm".
No idea who thought it was a good idea, 99% of the time I have zero control on the algorithms supported by the server I'm connecting too.
Just set "WarnWeakCrypto no" for all hosts.

2025-10-23 16:41:46 GMT

To anyone thinking that ASN.1 is dead, I suggest getting more familiar with X.509 to understand how the entire world secures network connections.
Also, wait until you realize what is inside telecom infrastructures.
I would love an excuse to write an ASN.1 implementation.

2025-10-22 18:25:41 GMT

- 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 GMT

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 GMT

- 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 13:42:45 GMT

Common Lisp logical pathnames are perfect to abstract the physical location of your files.
Here I define that assets are in a specific directory of my system ("systems" are Common Lisp packages). Then I can reference asset files anywhere with logical paths, e.g. #P"NOX:ASSETS;FONTS;IOSEVKA-CUSTOM-REGULAR.TTF".
And for a production executable, all I have to do is to change the translation to the location expected in the final archive.
https://cdn.fosstodon.org/media_attachments/files/115/401/120/957/471/382/original/c6b207c75de64882.png

2025-10-19 13:35:02 GMT

The #CommonLisp standard should have included PARSE-NATIVE-NAMESTRING and NATIVE-NAMESTRING.
Also I'm curious why SB-EXT:NATIVE-NAMESTRING does not call TRANSLATE-LOGICAL-PATHNAME on logical pathnames instead of failing. The very point of the function is to obtain the native representation of a path.

2025-10-19 00:20:19 GMT

- 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 GMT

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 19:36:15 GMT

- reply

@nprofile…dcu3 Obviously there is no point in asking about things you already know. But not knowing about something does mean you do not have contextual knowledge to judge wether what the AI says makes sense or not. The exact same way you'd do with any human being telling you something.
And if you do not have any contextual knowledge at all, then of course you cannot judge. Exactly the same way as for a human being.

2025-10-12 15:14:13 GMT

- 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 GMT

- 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 GMT

- 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 GMT

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.

2025-10-04 12:14:24 GMT

I was adding connection timeouts to my #CommonLisp client, and had to remember that there's no immediate way to do that with POSIX. You have to set the socket non-blocking with fcntl, call connect (which will fail with EINPROGRESS), poll until you get a POLLOUT event (or a timeout), then check for error with getsockopt and SOL_SOCKET/SO_ERROR and put the socket back to blocking mode. Could be worse.

2025-10-01 11:26:34 GMT

If you incorrectly use -H 'Foo=bar' instead of -H 'Foo: bar', #Curl will ignore it and not send the header field. Really unexpected and error inducing, I would expect it to validate the parameter and fail.

2025-09-29 20:37:21 GMT

Turns out Microsoft Entra does not follow the #OIDC specification either. It rejects authentication requests with a prompt set to "login consent" with "Unsupported 'prompt' value". Works with Google as it should be.

2025-09-09 19:01:10 GMT

- reply

@nprofile…uej8 I imagine I could. I did not yet because I foresee a long and frustrating debate with people who won't want to change the default behaviour. Long experience in the open source world made me wary of trying.

2025-09-06 23:58:54 GMT

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.

2025-08-25 17:05:57 GMT

I still love #CommonLisp but I spent too much time dealing with production issues to ship production software in a language that cannot be statically typechecked. Yes, type declarations sometimes help with specific implementations; but "sometimes" is just not good enough.

2025-08-20 13:38:15 GMT

HTTP SSE (Server Side Events) is a good idea with a terrible implementation: text based, three different line termination sequences (but why?!), inefficient representation, various absurd subtleties (guess what happens if an event id field contains a null byte…).
Nothing can surprise me anymore in the web ecosystem, but come on, how can you mess something that simple?

2025-08-18 14:13:15 GMT

- reply

@nprofile…etxc Please show me on this wonderful dialog where I can see individual cookies and their attributes :)

2025-08-18 09:57:18 GMT

I'm glad #Firefox exists, but the fact that it does not have a simple about:cookies page to visualize/edit/delete cookies _globally_ (and not just on the current domain with developer tools) is absurd.
To do anything, I have to stop Firefox (because it stupidely locks the cookies SQLite file), apply modifications and restart. They clearly are not even trying.

2025-08-13 11:28:06 GMT

#Go 1.25 introduces (*WaitGroup).Go, to spawn a goroutine and deal with the wait group counter automatically and correctly. Small but useful addition to the standard library.

2025-08-08 12:41:47 GMT

In software as in all crafts, a prototype is worth a hundred design documents.

2025-08-08 06:54:59 GMT

I just published a first draft of the BCL specification. BCL is a configuration language designed because I was fed up with YAML. It is already used by the Boulevard reverse proxy. More fun stuff to come in future versions such as support for units (e.g. writing `30KiB` instead of `30000`).
Go implementation and Emacs package available on my GitHub account.
And of course the specification is served by Boulevard :)

2025-08-04 16:49:54 GMT

- reply

@nprofile…6r8y Easier to get an audience by puking short, controversial statements all day. I should give it a try just to see if that works and write an article on the experiment.

2025-08-04 15:34:59 GMT

- reply

@nprofile…3flu Noted. However I already moved once (from the defunct emacs.cc server), and I'm not eager to do it again (impossible to move post history, lost followers in the move).
The only reason I still post to Mastodon is that I have interesting people (both followers and followed). But if I had to do it all over again, I would use X only.

2025-08-04 14:58:56 GMT

- 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 GMT

[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.

2025-08-03 13:33:57 GMT

While it's customary to avoid mixing the terms bytes and octets (because historically some CPUs have non-8 bit bytes), ISO/IEC 80000-13 clearly establishes the byte as a 8 bit unit.
One less thing to bikeshed about!

2025-07-28 10:10:33 GMT

- reply

@nprofile…fkdz GDPR already accounts for logs, backups, and any other system of the sort (including retention for legal reasons). If it's technically necessary (and you can show that it is), you're good.
It has nothing to do with soft deletion.

2025-07-28 10:09:14 GMT

- reply

@nprofile…fkdz And the day you get hacked (or when a developer forgets to filter a query), everything gets out. This is why it's a liability.

2025-07-28 09:58:16 GMT

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.

2025-07-27 14:47:56 GMT

Boulevard load balancers now support health probes, TCP or HTTP as in this example. Servers that fail too much are ignored when routing new requests until they recover. It's such a simple feature, and yet NGINX cannot do that without paying for the enterprise version.
https://cdn.fosstodon.org/media_attachments/files/114/925/742/847/637/494/original/93a8a9533648c373.png

2025-07-20 14:39:20 GMT

It's hard not to smile when looking at engineers who snub the command line when they have to debug a production system for the first time.
Yes you have a remote shell and not much else, good luck. And yes using Linux instead of MacOS on a daily basis would have helped you there.

2025-07-20 13:49:34 GMT

- reply

@nprofile…er8p Same as with OVH, you get what you pay for. Both are very good for the price, but you cannot expect miracles either.

2025-07-19 12:56:17 GMT

What is wonderful about #Omarchy is that DHH is showing the world that yes, you can have a productive and good looking environment on #Linux that is not a carbon copy (e.g. Gnome) of the horrific Windows/Mac standard interfaces.
And once people see that it's not that scary, they may start peeking underneath and see what you can really do with a computer beyond clicking icons.

2025-07-13 13:31:21 GMT

HTTP load balancing just dropped in Boulevard! List servers, configure the reverse proxy, the end. It still is a bit bare-bone: soon I'll add health checks, DNS resolution control, balancing modes (round robin for the time being), per-server rate limiting…
Oh and soon the same load balancers will be supported in layer 4 TCP servers, of course.
https://cdn.fosstodon.org/media_attachments/files/114/846/169/562/647/228/original/a1a0fa46637193ee.png

2025-07-12 10:11:20 GMT

Software engineering requires discipline and focus. It is perfectly normal to lose a lot of juniors who thought is was all fun and vibe code then realized software is actually not for them. Companies that refuse to enforce this selection process because they want to be "nice" will never get the competent engineers they need.

2025-07-06 12:04:23 GMT

Boulevard is now available as a Docker image, exograd/boulevard! I originally wrote it as an HTTP server and reverse proxy for my dedicated servers, but it's also nice to add authentication to a service running in Kubernetes or to route requests to the right place.

2025-07-05 16:18:23 GMT

When you need a dentist, you get a dentist, not a cardiologist. When you need an M&A laywer, you don't get a criminal attorney. When you need a frontend engineer, you don't hire a backend engineer and try to force them to do your web work because "software is software stop complaining".
If you're an executive or recruiter thinking this way, you are unqualified to work in the tech industry.

2025-07-05 14:03:58 GMT

- reply

@nprofile…29fa In syslogd's case, it comes from the fact that this is very old software that is built upon even older syslog programs used by all kind of UNIX variants.
And no one ever stopped and thought "hey what about making it right instead of obsessing about half-assed backward compatibility with a badly designed configuration language".
This description extends to pretty much everything UNIX.

2025-07-05 13:50:45 GMT

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.

2025-07-04 13:16:17 GMT

- reply

@nprofile…zvzf Yes, moved from 27" 4k to 27" 5k. The difference is more than visible on text. The new monitor is also supposed to have much higher color fidelity.

2025-07-04 08:34:50 GMT

I switched to a single 27" 5k monitor (Benq PD2730S). The difference in visual comfort when reading text (i.e. all the time) is significant. Never going back below 220 DPI.

2025-07-03 17:51:03 GMT

The rumour that Meta was paying $100M bonuses spread everywhere online and a ton of people genuinely believed it.
If something that is so obviously bullshit can spread that fast and be actually believed by so many, imagine how many subtle lies go undetected.

2025-07-01 15:08:02 GMT

I can assure you that most business/product people care about technical quality when it impacts development velocity, support or the reputation of the company. They are just allergic to developers endlessly chasing the latest trends to pad their resumes instead of getting shit done.
Start figuring out what the business needs and conversations will become so much easier.

2025-06-29 17:59:15 GMT

- reply

@nprofile…4d9x frustrating sure, but mostly scary due to the quantity of software written in Go that may have subtle bugs due to these issues.